Added parameter derivatives of perturbation solution up to 3 order
# Preliminary comments
I finished the identification toolbox at orders two and three using the pruned state space system, but before I merge request this, I decided to first merge the new functionality to compute parameter derivatives of perturbation solution matrices at higher orders. So after this is approved, I merge the identification toolbox.
I guess @rattoma, @sebastien, and @michel are best choices to review this.
I outline the main idea first and then provide some more detailed changes I made to the functions.
***
# Main idea
This merge request is concerned with the *analytical*computation of the parameter derivatives of first, second and third order perturbation solution matrices, i.e. using _closed-form_ expressions to efficiently compute the derivative of $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ *with respect to model parameters* $\theta$. Note that $\theta$ contains model parameters, stderr and corr parameters of shocks. stderr and corr parameters of measurement errors are not yet supported, (they can easily be included as exogenous shocks). The availability of such derivatives is beneficial in terms of more reliable analysis of model sensitivity and parameter identifiability as well as more efficient estimation methods, in particular for models solved up to third order, as it is well-known that numerical derivatives are a tricky business, especially for large models.
References for my approach are:
* Iskrev (2008, 2010) and Schmitt-Grohé and Uribe (2012, Appendix) who were the first to compute the parameter derivatives analytically at first order, however, using inefficient (sparse) Kronecker products.
* Mutschler (2015) who provides the expressions for a second-order, but again using inefficient (sparse) Kronecker products.
* Ratto and Iskrev (2012) who show how the first-order system can be solved accurately, fast and efficiently using existing numerical algorithms for generalized Sylvester equations by taking the parameter derivative with respect to each parameter separately.
* Julliard and Kamenik (2004) who provide the perturbation solution equation system in tensor notation at any order k.
* Levintal (2017) who introduces permutation matrices to express the perturbation solution equation system in matrix notation up to fifth order.
Note that @rattoma already implemented the parameter derivatives of $g_x$ and $g_u$ analytically (and numerically), and I rely heavily on his work in `get_first_order_solution_params_derivs.m` (previously `getH.m`). My additions are mainly to this function and thus it is renamed to `get_perturbation_params_derivs.m`.
The basic idea of this merge request is to take the second- and third-order perturbation solution systems in Julliard and Kamenik (2004), unfold these into an equivalent matrix representation using permutation matrices as in Levintal (2017). Then extending Ratto and Iskrev (2012) one takes the derivative with respect to each parameter separately and gets a computational problem that is linear, albeit large, as it involves either solving generalized Sylvester equations or taking inverses of highly sparse matrices. I will now briefly summarize the perturbation solution system at third order and the system that results when taking the derivative with respect to parameters.
## Perturbation Solution
The following systems arise at first, second, and third order:
$(ghx): f_{x} z_{x} = f_{y_{-}^*} + f_{y_0} g_{x} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{x}= A g_{x} + f_{y_{-}^*}=0$
$(ghu): f_{z} z_{u} = f_{y_0} g_{u} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{u} + f_{u}= A g_u + f_u = 0$
$(ghxx) : A g_{xx} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{x}\right) + f_{zz} \left( z_{x} \otimes z_{x} \right) = 0$
$(ghxu) : A g_{xu} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{x} \otimes z_{u} \right) = 0$
$(ghuu) : A g_{uu} + B g_{xx} \left(g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{u} \otimes z_{u} \right) = 0$
$(ghs2) : (A+B) g_{\sigma\sigma} + \left( f_{y^{**}_{+}y^{**}_{+}} \left(g^{**}_{u} \otimes g^{**}_{u}\right) + f_{y^{**}_{+}} g^{**}_{uu}\right)vec(\Sigma) = 0$
$(ghxxx) : A g_{xxx} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{x}\right) + f_{y_{+}}g^{**}_{xx} \left(g^{*}_x \otimes g^{*}_{xx}\right)P_{x\_xx} + f_{zz} \left( z_{x} \otimes z_{xx} \right)P_{x\_xx} + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{x} \right) = 0$
$(ghxxu) : A g_{xxu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{u} \right) + f_{zz} \left( \left( z_{x} \otimes z_{xu} \right)P_{x\_xu} + \left(z_{xx} \otimes z_{u}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{x} \otimes g^{*}_{xu}\right)P_{x\_xu} + \left(g^{*}_{xx} \otimes g^{*}_{u}\right) \right) = 0$
$(ghxuu) : A g_{xuu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( \left( z_{xu} \otimes z_{u} \right)P_{xu\_u} + \left(z_{x} \otimes z_{uu}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{xu} \otimes g^{*}_{u}\right)P_{xu\_u} + \left(g^{*}_{x} \otimes g^{*}_{uu}\right) \right) = 0$
$(ghuuu) : A g_{uuu} + B g_{xxx} \left(g^{*}_{u} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{u} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( z_{u} \otimes z_{uu} \right)P_{u\_uu} + f_{y_{+}}g^{**}_{xx} \left(g^{*}_{u} \otimes g^{*}_{uu}\right)P_{u\_uu} = 0$
$(ghx\sigma\sigma) : A g_{x\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_x + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{x} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{x} \otimes z_{\sigma\sigma}\right) + F_{xu_{+}u_{+}}\left(I_{n_x} \otimes vec(\Sigma)\right) = 0$
$F_{xu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_x^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{xu_{+}} \otimes z_{u_{+}} \right)P_{xu\_u} + \left(z_{x} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{x} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
$(ghu\sigma\sigma) : A g_{u\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_{u} + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{u} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{u} \otimes z_{\sigma\sigma}\right) + F_{uu_{+}u_{+}}\left(I_{n_u} \otimes vec(\Sigma_u)\right) = 0$
$F_{uu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_u^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{uu_{+}} \otimes z_{u_{+}} \right)P_{uu\_u} + \left(z_{u} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{u} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
A and B are the common perturbation matrices:
$A = f_{y_0} + \begin{pmatrix} \underbrace{0}_{n\times n_{static}} &\vdots& \underbrace{f_{y^{**}_{+}} \cdot g^{**}_{x}}_{n \times n_{spred}} &\vdots& \underbrace{0}_{n\times n_{frwd}} \end{pmatrix}$and $B = \begin{pmatrix} \underbrace{0}_{n \times n_{static}}&\vdots & \underbrace{0}_{n \times n_{pred}} & \vdots & \underbrace{f_{y^{**}_{+}}}_{n \times n_{sfwrd}} \end{pmatrix}$
and $z=(y_{-}^{\ast}; y; y_{+}^{\ast\ast}; u)$ denotes the dynamic model variables as in `M_.lead_lag_incidence`, $y^\ast$ denote state variables, $y^{\ast\ast}$ denote forward looking variables, $y_+$ denote the variables with a lead, $y_{-}$ denote variables with a lag, $y_0$ denote variables at period t, $f$ the model equations, and $f_z$ the first-order dynamic model derivatives, $f_{zz}$ the second-order dynamic derivatives, and $f_{zzz}$ the third-order dynamic model derivatives. Then:
$z_{x} = \begin{pmatrix}I\\g_{x}\\g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{u} =\begin{pmatrix}0\\g_{u}\\g^{**}_{x} \cdot g^{*}_{u}\\I\end{pmatrix}$, $z_{u_{+}} =\begin{pmatrix}0\\0\\g^{**}_{u}\\0\end{pmatrix}$
$z_{xx} = \begin{pmatrix} 0\\g_{xx}\\g^{**}_{x} \left( g^{*}_x \otimes g^{*}_{x} \right) + g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{xu} =\begin{pmatrix}0\\g_{xu}\\g^{**}_{xx} \left( g^{*}_x \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{xu}\\0\end{pmatrix}$, $z_{uu} =\begin{pmatrix}0\\g_{uu}\\g^{**}_{xx} \left( g^{*}_u \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{uu}\\0\end{pmatrix}$,
$z_{xu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_x \otimes I \right)\\0\end{pmatrix}$, $z_{uu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_{u} \otimes I \right)\\0\end{pmatrix}$, $z_{u_{+}u_{+}} =\begin{pmatrix}0\\0\\g^{\ast\ast}_{uu}\\0\end{pmatrix}$, $z_{\sigma\sigma} = \begin{pmatrix}0\\ g_{\sigma\sigma}\\ g^{\ast\ast}_{x}g^{\ast}_{\sigma\sigma} + g^{\ast\ast}_{\sigma\sigma}\\0 \end{pmatrix}$
$P$ are permutation matrices that can be computed using Matlab's `ipermute` function.
## Parameter derivatives of perturbation solutions
First, we need the parameter derivatives of first, second, third, and fourth derivatives of the dynamic model (i.e. g1,g2,g3,g4 in dynamic files), I make use of the implicit function theorem: Let $f_{z^k}$ denote the kth derivative (wrt all dynamic variables) of the dynamic model, then let $df_{z^k}$ denote the first-derivative (wrt all model parameters) of $f_{z^k}$ evaluated at the steady state. Note that $f_{z^k}$ is a function of both the model parameters $\theta$ and of the steady state of all dynamic variables $\bar{z}$, which also depend on the parameters. Hence, implicitly $f_{z^k}=f_{z^k}(\theta,\bar{z}(\theta))$ and $df_{z^k}$ consists of two parts:
1. direct derivative wrt to all model parameters given by the preprocessor in the `_params_derivs.m` files
2. contribution of derivative of steady state of dynamic variables (wrt all model parameters): $f_{z^{k+1}} \cdot d\bar{z}$
Note that we already have functionality to compute $d\bar{z}$ analytically.
Having this, the above perturbation systems are basically equations of the following types
$AX +BXC = RHS$ or $AX = RHS$
Now when taking the derivative (wrt to one single parameter $\theta_j$), we get
$A\mathrm{d}\{X\} + B\mathrm{d}\{X\}C = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X - \mathrm{d}\{B\}XC - BX\mathrm{d}\{C\}$
or
$A\mathrm{d}\{X\} = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X$
The first one is a Sylvester type equation, the second one can be solved by taking the inverse of $A$. The only diffculty and tedious work arrises in computing (the highly sparse) derivatives of $RHS$.
***
# New functions: `
## get_perturbation_params_derivs.m`and `get_perturbation_params_derivs_numerical_objective.m`
* The parameter derivatives up to third order are computed in the new function`get_perturbation_params_derivs.m` both analytically and numerically. For numerical derivatives `get_perturbation_params_derivs_numerical_objective.m` is the objective for `fjaco.m` or `hessian_sparse.m` or `hessian.m`.
* `get_perturbation_params_derivs.m` is basically an extended version of the previous `get_first_order_solution_params_derivs.m` function.
* * `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It is used for numerical derivatives, whenever `analytic_derivation_mode=-1|-2`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
* Detailed changes:
* Most important: notation of this function is now in accordance to the k_order_solver, i.e. we do not compute derivatives of Kalman transition matrices A and B, but rather the solution matrices ghx,ghu,ghxx,ghxu,ghuu,ghs2,ghxxx,ghxxu,ghxuu,ghuuu,ghxss,ghuss in accordance with notation used in `oo_.dr`. As a byproduct at first-order, focusing on ghx and ghu instead of Kalman transition matrices A and B makes the computations slightly faster for large models (e.g. for Quest the computations were faster by a couple of seconds, not much, but okay).
* Removed use of `kstate`, see also Dynare/dynare#1653 and Dynare/dynare!1656
* Output arguments are stored in a structure `DERIVS`, there is also a flag `d2flag` that computes parameter hessians needed only in `dsge_likelihood.m`.
* Removed `kronflag` as input. `options_.analytic_derivation_mode` is now used instead of `kronflag`.
* Removed `indvar`, an index that was used to selected specific variables in the derivatives. This is not needed, as we always compute the parameter derivatives for all variables first and then select a subset of variables. The selection now takes place in other functions, like `dsge_likelihood.m`.
* Introduced some checks: (i) deterministic exogenous variables are not supported, (ii) Kronecker method only compatible with first-order approximation so reset to sylvester method, (iii) for purely backward or forward models we need to be careful with the rows in `M_.lead_la g_incidence`, (iv) if `_params_derivs.m` files are missing an error is thrown.
* For numerical derivatives, if mod file does not contain an `estimated_params_block`, a temporary one with the most important parameter information is created.
## `unfold_g4.m`
* When evaluating g3 and g4 one needs to take into account that these do not contain symmetric elements, so one needs to use `unfold_g3.m` and the new function `unfold_g4.m`. This returns an unfolded version of the same matrix (i.e. with symmetric elements).
***
# New test models
`.gitignore` and `Makefile.am` are changed accordingly. Also now it is possible to run test suite on analytic_derivatives, i.e. run `make check m/analytic_derivatives`
## `analytic_derivatives/BrockMirman_PertParamsDerivs.mod`
* This is the Brock Mirman model, where we know the exact policy function $g$ for capital and consumption. As this does not imply a nonzero $g_{\sigma\sigma}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ I added some artificial equations to get nonzero solution matrices with respect to $\sigma$. The true perturbation solution matrices $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ are then computed analytically with Matlab's symbolic toolbox and saved in `nBrockMirmanSYM.mat`. There is a preprocessor flag that recreates these analytical computations if changes are needed (and to check whether I made some errors here ;-) )
* Then solution matrices up to third order and their parameter Jacobians are then compared to the ones computed by Dynare's `k_order_solver` and by `get_perturbation_params_derivs` for all `analytic_derivation_mode`'s. There will be an error if the maximum absolute deviation is too large, i.e. for numerical derivatives (`analytic_derivation_mode=-1|-2`) the tolerance is choosen lower (around 1e-5); for analytical methods we are stricter: around 1e-13 for first-order, 1e-12 for second order, and 1e-11 for third-order.
* As a side note, this mod file also checks Dynare's `k_order_solver` algorithm and throws an error if something is wrong.
* This test model shows that the new functionality works well. And analytical derivatives perform way better and accurate than numerical ones, even for this small model.
## `analytic_derivatives/burnside_3_order_PertParamsDerivs.mod`
* This builds upon `tests/k_order_perturbation/burnside_k_order.mod` and computes the true parameter derivatives analytically by hand.
* This test model also shows that the new functionality works well.
## `analytic_derivatives/LindeTrabandt2019.mod`
* Shows that the new functionality also works for medium-sized models, i.e. a SW type model solved at third order with 35 variables (11 states). 2 shocks and 20 parameters.
* This mod file can be used to tweak the speed of the computations in the future.
* Compares numerical versus analytical parameter derivatives (for first, second and third order). Note that this model clearly shows that numerical ones are quite different than analytical ones even at first order!
## `identification/LindeTrabandt2019_xfail.mod`
* This model is a check for issue Dynare/dynare#1595, see fjaco.m below, and will fail.
* Removed `analytic_derivatives/ls2003.mod` as this mod file is neither in the testsuite nor does it work.
***
# Detailed changes in other functions
## `get_first_order_solution_params_derivs.m`
* Deleted, or actually, renamed to `get_perturbation_params_derivs.m`, as this function now is able to compute the derivatives up to third order
## `identification_numerical_objective.m`
* `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
## `dsge_likelihood.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `T`matrix anymore, but instead of the dynare solution matrix `ghx`. So we recreate `T` here as this amounts to adding some zeros and focusing on selected variables only.
* Added some checks to make sure the first-order approximation is selected.
* Removed `kron_flag` as input, as `get_perturbation_params_derivs` looks into `options_.analytic_derivation_mode` for `kron_flag`.
## `dynare_identification.m`
* make sure that setting `analytic_derivation_mode` is set both in `options_ident` and `options_`. Note that at the end of the function we restore the `options_` structure, so all changes are local. In a next merge request, I will remove the global variables to make all variables local.
## `get_identification_jacobians.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `A` and `B` matrix anymore, but instead of the dynare solution matrix `ghx` and `ghu`. So we recreate these matrices here instead of in `get_perturbation_params_derivs.m`.
* Added `str2func` for better function handles in `fjaco.m`.
## `fjaco.m`
* make `tol`an option, which can be adjusted by changing `options_.dynatol.x`for identification and parameter derivatives purposes.
* include a check and an informative error message, if numerical derivatives (two-sided finite difference method) yield errors in `resol.m` for identification and parameter derivatives purposes. This closes issue Dynare/dynare#1595.
* Changed year of copyright to 2010-2017,2019
***
# Further suggestions and questions
* Ones this is merged, I will merge request an improvement of the identification toolbox, which will work up to third order using the pruned state space. This will also remove some issues and bugs, and also I will remove global variables in this request.
* The third-order derivatives can be further improved by taking sparsity into account and use mex versions for kronecker products etc. I leave this for further testing (and if anybody actually uses this ;-) )
2019-12-17 19:17:09 +01:00
function DERIVS = get_perturbation_params_derivs ( M, options, estim_params, oo, indpmodel, indpstderr, indpcorr, d2flag)
% DERIVS = get_perturbation_params_derivs(M, options, estim_params, oo, indpmodel, indpstderr, indpcorr, d2flag)
% previously getH.m in dynare 4.5
% -------------------------------------------------------------------------
% Computes derivatives (with respect to parameters) of
% (1) steady-state (oo_.dr.ys) and covariance matrix of shocks (M_.Sigma_e)
% (2) dynamic model (g1, g2, g3)
% (3) perturbation solution matrices up to third order oo_.dr.(ghx,ghu,ghxx,ghxu,ghuu,ghs2,ghxxx,ghxxu,ghxuu,ghuuu,ghxss,ghuss)
% Note that the order in the parameter Jacobians is the following:
% first stderr parameters, second corr parameters, third model parameters
%
% =========================================================================
% INPUTS
% M: [structure] storing the model information
% options: [structure] storing the options
% estim_params: [structure] storing the estimation information
% oo: [structure] storing the solution results
% indpmodel: [modparam_nbr by 1] index of estimated parameters in M.params;
% corresponds to model parameters (no stderr and no corr) in estimated_params block
% indpstderr: [stderrparam_nbr by 1] index of estimated standard errors,
% i.e. for all exogenous variables where "stderr" is given in the estimated_params block
% indpcorr: [corrparam_nbr by 2] matrix of estimated correlations,
% i.e. for all exogenous variables where "corr" is given in the estimated_params block
% d2flag: [boolean] flag to compute second-order parameter derivatives of steady state and first-order Kalman transition matrices
% -------------------------------------------------------------------------
% OUTPUTS
% DERIVS: Structure with the following fields:
% dYss: [endo_nbr by modparam_nbr] in DR order
% Jacobian (wrt model parameters only) of steady state oo_.dr.ys(oo_.dr.order_var,:)
% dSigma_e: [exo_nbr by exo_nbr by totparam_nbr] in declaration order
% Jacobian (wrt to all paramters) of covariance matrix of shocks M.Sigma_e
% dg1: [endo_nbr by yy0ex0_nbr by modparam_nbr] in DR order
% Parameter Jacobian of first derivative (wrt dynamic model variables) of dynamic model (wrt to model parameters only)
% dg2: [endo_nbr by yy0ex0_nbr^2*modparam_nbr] in DR order
% Parameter Jacobian of second derivative (wrt dynamic model variables) of dynamic model (wrt to model parameters only)
% Note that instead of tensors we use matrix notation with blocks: dg2 = [dg2_dp1 dg2_dp2 ...],
% where dg2_dpj is [endo_nbr by yy0ex0_nbr^2] and represents the derivative of g2 wrt parameter pj
% dg3: [endo_nbr by yy0ex0_nbr^3*modparam_nbr] in DR order
% Parameter Jacobian of third derivative (wrt dynamic model variables) of dynamic model (wrt to model parameters only)
% Note that instead of tensors we use matrix notation with blocks: dg3 = [dg3_dp1 dg3_dp2 ...],
% where dg3_dpj is [endo_nbr by yy0ex0_nbr^3] and represents the derivative of g3 wrt parameter pj
% dghx: [endo_nbr by states_nbr by totparam_nbr] in DR order
% Jacobian (wrt to all parameters) of first-order perturbation solution matrix oo_.dr.ghx
% dghu: [endo_nbr by exo_nbr by totparam_nbr] in DR order
% Jacobian (wrt to all parameters) of first-order perturbation solution matrix oo_.dr.ghu
% dOm: [endo_nbr by endo_nbr by totparam_nbr] in DR order
% Jacobian (wrt to all paramters) of Om = oo_.dr.ghu*M.Sigma_e*transpose(oo_.dr.ghu)
% dghxx [endo_nbr by states_nbr*states_nbr by totparam_nbr] in DR order
% Jacobian (wrt to all parameters) of second-order perturbation solution matrix oo_.dr.ghxx
% dghxu [endo_nbr by states_nbr*exo_nbr by totparam_nbr] in DR order
% Jacobian (wrt to all parameters) of second-order perturbation solution matrix oo_.dr.ghxu
% dghuu [endo_nbr by exo_nbr*exo_nbr by totparam_nbr] in DR order
% Jacobian (wrt to all parameters) of second-order perturbation solution matrix oo_.dr.ghuu
% dghs2 [endo_nbr by totparam_nbr] in DR order
% Jacobian (wrt to all parameters) of second-order perturbation solution matrix oo_.dr.ghs2
% dghxxx [endo_nbr by states_nbr*states_nbr*states_nbr by totparam_nbr] in DR order
% Jacobian (wrt to all parameters) of third-order perturbation solution matrix oo_.dr.ghxxx
% dghxxu [endo_nbr by states_nbr*states_nbr*exo_nbr by totparam_nbr] in DR order
% Jacobian (wrt to all parameters) of third-order perturbation solution matrix oo_.dr.ghxxu
% dghxuu [endo_nbr by states_nbr*exo_nbr*exo_nbr by totparam_nbr] in DR order
% Jacobian (wrt to all parameters) of third-order perturbation solution matrix oo_.dr.ghxuu
% dghuuu [endo_nbr by exo_nbr*exo_nbr*exo_nbr by totparam_nbr] in DR order
% Jacobian (wrt to all parameters) of third-order perturbation solution matrix oo_.dr.ghuuu
% dghxss [endo_nbr by states_nbr by totparam_nbr] in DR order
% Jacobian (wrt to all parameters) of third-order perturbation solution matrix oo_.dr.ghxss
% dghuss [endo_nbr by exo_nbr by totparam_nbr] in DR order
% Jacobian (wrt to all parameters) of third-order perturbation solution matrix oo_.dr.ghuss
% If d2flag==true, we additional output:
% d2KalmanA: [endo_nbr*endo_nbr by totparam_nbr*(totparam_nbr+1)/2] in DR order
% Unique entries of Hessian (wrt all parameters) of Kalman transition matrix A
% d2Om: [endo_nbr*(endo_nbr+1)/2 by totparam_nbr*(totparam_nbr+1)/2] in DR order
% Unique entries of Hessian (wrt all parameters) of Om=oo_.dr.ghu*M_.Sigma_e*transpose(oo_.dr.ghu)
% d2Yss: [endo_nbr by modparam_nbr by modparam_nbr] in DR order
% Unique entries of Hessian (wrt model parameters only) of steady state oo_.dr.ys(oo_.dr.order_var,:)
%
% -------------------------------------------------------------------------
% This function is called by
% * dsge_likelihood.m
% * get_identification_jacobians.m (previously getJJ.m in Dynare 4.5)
% -------------------------------------------------------------------------
% This function calls
% * [fname,'.dynamic']
% * [fname,'.dynamic_params_derivs']
% * [fname,'.static']
% * [fname,'.static_params_derivs']
% * commutation
% * dyn_vech
% * dyn_unvech
% * fjaco
% * get_2nd_deriv (embedded)
% * get_2nd_deriv_mat(embedded)
% * get_all_parameters
% * get_all_resid_2nd_derivs (embedded)
% * get_hess_deriv (embedded)
% * hessian_sparse
% * sylvester3
% * sylvester3a
% * get_perturbation_params_derivs_numerical_objective
% =========================================================================
2020-01-10 17:55:57 +01:00
% Copyright (C) 2019-2020 Dynare Team
Added parameter derivatives of perturbation solution up to 3 order
# Preliminary comments
I finished the identification toolbox at orders two and three using the pruned state space system, but before I merge request this, I decided to first merge the new functionality to compute parameter derivatives of perturbation solution matrices at higher orders. So after this is approved, I merge the identification toolbox.
I guess @rattoma, @sebastien, and @michel are best choices to review this.
I outline the main idea first and then provide some more detailed changes I made to the functions.
***
# Main idea
This merge request is concerned with the *analytical*computation of the parameter derivatives of first, second and third order perturbation solution matrices, i.e. using _closed-form_ expressions to efficiently compute the derivative of $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ *with respect to model parameters* $\theta$. Note that $\theta$ contains model parameters, stderr and corr parameters of shocks. stderr and corr parameters of measurement errors are not yet supported, (they can easily be included as exogenous shocks). The availability of such derivatives is beneficial in terms of more reliable analysis of model sensitivity and parameter identifiability as well as more efficient estimation methods, in particular for models solved up to third order, as it is well-known that numerical derivatives are a tricky business, especially for large models.
References for my approach are:
* Iskrev (2008, 2010) and Schmitt-Grohé and Uribe (2012, Appendix) who were the first to compute the parameter derivatives analytically at first order, however, using inefficient (sparse) Kronecker products.
* Mutschler (2015) who provides the expressions for a second-order, but again using inefficient (sparse) Kronecker products.
* Ratto and Iskrev (2012) who show how the first-order system can be solved accurately, fast and efficiently using existing numerical algorithms for generalized Sylvester equations by taking the parameter derivative with respect to each parameter separately.
* Julliard and Kamenik (2004) who provide the perturbation solution equation system in tensor notation at any order k.
* Levintal (2017) who introduces permutation matrices to express the perturbation solution equation system in matrix notation up to fifth order.
Note that @rattoma already implemented the parameter derivatives of $g_x$ and $g_u$ analytically (and numerically), and I rely heavily on his work in `get_first_order_solution_params_derivs.m` (previously `getH.m`). My additions are mainly to this function and thus it is renamed to `get_perturbation_params_derivs.m`.
The basic idea of this merge request is to take the second- and third-order perturbation solution systems in Julliard and Kamenik (2004), unfold these into an equivalent matrix representation using permutation matrices as in Levintal (2017). Then extending Ratto and Iskrev (2012) one takes the derivative with respect to each parameter separately and gets a computational problem that is linear, albeit large, as it involves either solving generalized Sylvester equations or taking inverses of highly sparse matrices. I will now briefly summarize the perturbation solution system at third order and the system that results when taking the derivative with respect to parameters.
## Perturbation Solution
The following systems arise at first, second, and third order:
$(ghx): f_{x} z_{x} = f_{y_{-}^*} + f_{y_0} g_{x} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{x}= A g_{x} + f_{y_{-}^*}=0$
$(ghu): f_{z} z_{u} = f_{y_0} g_{u} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{u} + f_{u}= A g_u + f_u = 0$
$(ghxx) : A g_{xx} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{x}\right) + f_{zz} \left( z_{x} \otimes z_{x} \right) = 0$
$(ghxu) : A g_{xu} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{x} \otimes z_{u} \right) = 0$
$(ghuu) : A g_{uu} + B g_{xx} \left(g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{u} \otimes z_{u} \right) = 0$
$(ghs2) : (A+B) g_{\sigma\sigma} + \left( f_{y^{**}_{+}y^{**}_{+}} \left(g^{**}_{u} \otimes g^{**}_{u}\right) + f_{y^{**}_{+}} g^{**}_{uu}\right)vec(\Sigma) = 0$
$(ghxxx) : A g_{xxx} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{x}\right) + f_{y_{+}}g^{**}_{xx} \left(g^{*}_x \otimes g^{*}_{xx}\right)P_{x\_xx} + f_{zz} \left( z_{x} \otimes z_{xx} \right)P_{x\_xx} + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{x} \right) = 0$
$(ghxxu) : A g_{xxu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{u} \right) + f_{zz} \left( \left( z_{x} \otimes z_{xu} \right)P_{x\_xu} + \left(z_{xx} \otimes z_{u}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{x} \otimes g^{*}_{xu}\right)P_{x\_xu} + \left(g^{*}_{xx} \otimes g^{*}_{u}\right) \right) = 0$
$(ghxuu) : A g_{xuu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( \left( z_{xu} \otimes z_{u} \right)P_{xu\_u} + \left(z_{x} \otimes z_{uu}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{xu} \otimes g^{*}_{u}\right)P_{xu\_u} + \left(g^{*}_{x} \otimes g^{*}_{uu}\right) \right) = 0$
$(ghuuu) : A g_{uuu} + B g_{xxx} \left(g^{*}_{u} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{u} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( z_{u} \otimes z_{uu} \right)P_{u\_uu} + f_{y_{+}}g^{**}_{xx} \left(g^{*}_{u} \otimes g^{*}_{uu}\right)P_{u\_uu} = 0$
$(ghx\sigma\sigma) : A g_{x\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_x + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{x} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{x} \otimes z_{\sigma\sigma}\right) + F_{xu_{+}u_{+}}\left(I_{n_x} \otimes vec(\Sigma)\right) = 0$
$F_{xu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_x^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{xu_{+}} \otimes z_{u_{+}} \right)P_{xu\_u} + \left(z_{x} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{x} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
$(ghu\sigma\sigma) : A g_{u\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_{u} + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{u} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{u} \otimes z_{\sigma\sigma}\right) + F_{uu_{+}u_{+}}\left(I_{n_u} \otimes vec(\Sigma_u)\right) = 0$
$F_{uu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_u^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{uu_{+}} \otimes z_{u_{+}} \right)P_{uu\_u} + \left(z_{u} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{u} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
A and B are the common perturbation matrices:
$A = f_{y_0} + \begin{pmatrix} \underbrace{0}_{n\times n_{static}} &\vdots& \underbrace{f_{y^{**}_{+}} \cdot g^{**}_{x}}_{n \times n_{spred}} &\vdots& \underbrace{0}_{n\times n_{frwd}} \end{pmatrix}$and $B = \begin{pmatrix} \underbrace{0}_{n \times n_{static}}&\vdots & \underbrace{0}_{n \times n_{pred}} & \vdots & \underbrace{f_{y^{**}_{+}}}_{n \times n_{sfwrd}} \end{pmatrix}$
and $z=(y_{-}^{\ast}; y; y_{+}^{\ast\ast}; u)$ denotes the dynamic model variables as in `M_.lead_lag_incidence`, $y^\ast$ denote state variables, $y^{\ast\ast}$ denote forward looking variables, $y_+$ denote the variables with a lead, $y_{-}$ denote variables with a lag, $y_0$ denote variables at period t, $f$ the model equations, and $f_z$ the first-order dynamic model derivatives, $f_{zz}$ the second-order dynamic derivatives, and $f_{zzz}$ the third-order dynamic model derivatives. Then:
$z_{x} = \begin{pmatrix}I\\g_{x}\\g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{u} =\begin{pmatrix}0\\g_{u}\\g^{**}_{x} \cdot g^{*}_{u}\\I\end{pmatrix}$, $z_{u_{+}} =\begin{pmatrix}0\\0\\g^{**}_{u}\\0\end{pmatrix}$
$z_{xx} = \begin{pmatrix} 0\\g_{xx}\\g^{**}_{x} \left( g^{*}_x \otimes g^{*}_{x} \right) + g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{xu} =\begin{pmatrix}0\\g_{xu}\\g^{**}_{xx} \left( g^{*}_x \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{xu}\\0\end{pmatrix}$, $z_{uu} =\begin{pmatrix}0\\g_{uu}\\g^{**}_{xx} \left( g^{*}_u \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{uu}\\0\end{pmatrix}$,
$z_{xu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_x \otimes I \right)\\0\end{pmatrix}$, $z_{uu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_{u} \otimes I \right)\\0\end{pmatrix}$, $z_{u_{+}u_{+}} =\begin{pmatrix}0\\0\\g^{\ast\ast}_{uu}\\0\end{pmatrix}$, $z_{\sigma\sigma} = \begin{pmatrix}0\\ g_{\sigma\sigma}\\ g^{\ast\ast}_{x}g^{\ast}_{\sigma\sigma} + g^{\ast\ast}_{\sigma\sigma}\\0 \end{pmatrix}$
$P$ are permutation matrices that can be computed using Matlab's `ipermute` function.
## Parameter derivatives of perturbation solutions
First, we need the parameter derivatives of first, second, third, and fourth derivatives of the dynamic model (i.e. g1,g2,g3,g4 in dynamic files), I make use of the implicit function theorem: Let $f_{z^k}$ denote the kth derivative (wrt all dynamic variables) of the dynamic model, then let $df_{z^k}$ denote the first-derivative (wrt all model parameters) of $f_{z^k}$ evaluated at the steady state. Note that $f_{z^k}$ is a function of both the model parameters $\theta$ and of the steady state of all dynamic variables $\bar{z}$, which also depend on the parameters. Hence, implicitly $f_{z^k}=f_{z^k}(\theta,\bar{z}(\theta))$ and $df_{z^k}$ consists of two parts:
1. direct derivative wrt to all model parameters given by the preprocessor in the `_params_derivs.m` files
2. contribution of derivative of steady state of dynamic variables (wrt all model parameters): $f_{z^{k+1}} \cdot d\bar{z}$
Note that we already have functionality to compute $d\bar{z}$ analytically.
Having this, the above perturbation systems are basically equations of the following types
$AX +BXC = RHS$ or $AX = RHS$
Now when taking the derivative (wrt to one single parameter $\theta_j$), we get
$A\mathrm{d}\{X\} + B\mathrm{d}\{X\}C = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X - \mathrm{d}\{B\}XC - BX\mathrm{d}\{C\}$
or
$A\mathrm{d}\{X\} = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X$
The first one is a Sylvester type equation, the second one can be solved by taking the inverse of $A$. The only diffculty and tedious work arrises in computing (the highly sparse) derivatives of $RHS$.
***
# New functions: `
## get_perturbation_params_derivs.m`and `get_perturbation_params_derivs_numerical_objective.m`
* The parameter derivatives up to third order are computed in the new function`get_perturbation_params_derivs.m` both analytically and numerically. For numerical derivatives `get_perturbation_params_derivs_numerical_objective.m` is the objective for `fjaco.m` or `hessian_sparse.m` or `hessian.m`.
* `get_perturbation_params_derivs.m` is basically an extended version of the previous `get_first_order_solution_params_derivs.m` function.
* * `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It is used for numerical derivatives, whenever `analytic_derivation_mode=-1|-2`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
* Detailed changes:
* Most important: notation of this function is now in accordance to the k_order_solver, i.e. we do not compute derivatives of Kalman transition matrices A and B, but rather the solution matrices ghx,ghu,ghxx,ghxu,ghuu,ghs2,ghxxx,ghxxu,ghxuu,ghuuu,ghxss,ghuss in accordance with notation used in `oo_.dr`. As a byproduct at first-order, focusing on ghx and ghu instead of Kalman transition matrices A and B makes the computations slightly faster for large models (e.g. for Quest the computations were faster by a couple of seconds, not much, but okay).
* Removed use of `kstate`, see also Dynare/dynare#1653 and Dynare/dynare!1656
* Output arguments are stored in a structure `DERIVS`, there is also a flag `d2flag` that computes parameter hessians needed only in `dsge_likelihood.m`.
* Removed `kronflag` as input. `options_.analytic_derivation_mode` is now used instead of `kronflag`.
* Removed `indvar`, an index that was used to selected specific variables in the derivatives. This is not needed, as we always compute the parameter derivatives for all variables first and then select a subset of variables. The selection now takes place in other functions, like `dsge_likelihood.m`.
* Introduced some checks: (i) deterministic exogenous variables are not supported, (ii) Kronecker method only compatible with first-order approximation so reset to sylvester method, (iii) for purely backward or forward models we need to be careful with the rows in `M_.lead_la g_incidence`, (iv) if `_params_derivs.m` files are missing an error is thrown.
* For numerical derivatives, if mod file does not contain an `estimated_params_block`, a temporary one with the most important parameter information is created.
## `unfold_g4.m`
* When evaluating g3 and g4 one needs to take into account that these do not contain symmetric elements, so one needs to use `unfold_g3.m` and the new function `unfold_g4.m`. This returns an unfolded version of the same matrix (i.e. with symmetric elements).
***
# New test models
`.gitignore` and `Makefile.am` are changed accordingly. Also now it is possible to run test suite on analytic_derivatives, i.e. run `make check m/analytic_derivatives`
## `analytic_derivatives/BrockMirman_PertParamsDerivs.mod`
* This is the Brock Mirman model, where we know the exact policy function $g$ for capital and consumption. As this does not imply a nonzero $g_{\sigma\sigma}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ I added some artificial equations to get nonzero solution matrices with respect to $\sigma$. The true perturbation solution matrices $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ are then computed analytically with Matlab's symbolic toolbox and saved in `nBrockMirmanSYM.mat`. There is a preprocessor flag that recreates these analytical computations if changes are needed (and to check whether I made some errors here ;-) )
* Then solution matrices up to third order and their parameter Jacobians are then compared to the ones computed by Dynare's `k_order_solver` and by `get_perturbation_params_derivs` for all `analytic_derivation_mode`'s. There will be an error if the maximum absolute deviation is too large, i.e. for numerical derivatives (`analytic_derivation_mode=-1|-2`) the tolerance is choosen lower (around 1e-5); for analytical methods we are stricter: around 1e-13 for first-order, 1e-12 for second order, and 1e-11 for third-order.
* As a side note, this mod file also checks Dynare's `k_order_solver` algorithm and throws an error if something is wrong.
* This test model shows that the new functionality works well. And analytical derivatives perform way better and accurate than numerical ones, even for this small model.
## `analytic_derivatives/burnside_3_order_PertParamsDerivs.mod`
* This builds upon `tests/k_order_perturbation/burnside_k_order.mod` and computes the true parameter derivatives analytically by hand.
* This test model also shows that the new functionality works well.
## `analytic_derivatives/LindeTrabandt2019.mod`
* Shows that the new functionality also works for medium-sized models, i.e. a SW type model solved at third order with 35 variables (11 states). 2 shocks and 20 parameters.
* This mod file can be used to tweak the speed of the computations in the future.
* Compares numerical versus analytical parameter derivatives (for first, second and third order). Note that this model clearly shows that numerical ones are quite different than analytical ones even at first order!
## `identification/LindeTrabandt2019_xfail.mod`
* This model is a check for issue Dynare/dynare#1595, see fjaco.m below, and will fail.
* Removed `analytic_derivatives/ls2003.mod` as this mod file is neither in the testsuite nor does it work.
***
# Detailed changes in other functions
## `get_first_order_solution_params_derivs.m`
* Deleted, or actually, renamed to `get_perturbation_params_derivs.m`, as this function now is able to compute the derivatives up to third order
## `identification_numerical_objective.m`
* `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
## `dsge_likelihood.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `T`matrix anymore, but instead of the dynare solution matrix `ghx`. So we recreate `T` here as this amounts to adding some zeros and focusing on selected variables only.
* Added some checks to make sure the first-order approximation is selected.
* Removed `kron_flag` as input, as `get_perturbation_params_derivs` looks into `options_.analytic_derivation_mode` for `kron_flag`.
## `dynare_identification.m`
* make sure that setting `analytic_derivation_mode` is set both in `options_ident` and `options_`. Note that at the end of the function we restore the `options_` structure, so all changes are local. In a next merge request, I will remove the global variables to make all variables local.
## `get_identification_jacobians.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `A` and `B` matrix anymore, but instead of the dynare solution matrix `ghx` and `ghu`. So we recreate these matrices here instead of in `get_perturbation_params_derivs.m`.
* Added `str2func` for better function handles in `fjaco.m`.
## `fjaco.m`
* make `tol`an option, which can be adjusted by changing `options_.dynatol.x`for identification and parameter derivatives purposes.
* include a check and an informative error message, if numerical derivatives (two-sided finite difference method) yield errors in `resol.m` for identification and parameter derivatives purposes. This closes issue Dynare/dynare#1595.
* Changed year of copyright to 2010-2017,2019
***
# Further suggestions and questions
* Ones this is merged, I will merge request an improvement of the identification toolbox, which will work up to third order using the pruned state space. This will also remove some issues and bugs, and also I will remove global variables in this request.
* The third-order derivatives can be further improved by taking sparsity into account and use mex versions for kronecker products etc. I leave this for further testing (and if anybody actually uses this ;-) )
2019-12-17 19:17:09 +01:00
%
% This file is part of Dynare.
%
% Dynare is free software: you can redistribute it and/or modify
% it under the terms of the GNU General Public License as published by
% the Free Software Foundation, either version 3 of the License, or
% (at your option) any later version.
%
% Dynare is distributed in the hope that it will be useful,
% but WITHOUT ANY WARRANTY; without even the implied warranty of
% MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
% GNU General Public License for more details.
%
% You should have received a copy of the GNU General Public License
% along with Dynare. If not, see <http://www.gnu.org/licenses/>.
% =========================================================================
% Some checks
if M . exo_det_nbr > 0
error ( ' ' ' get_perturbation_params_derivs' ' : not compatible with deterministic exogenous variables, please declare as endogenous.' )
end
if options . order > 1 && options . analytic_derivation_mode == 1
%analytic derivatives using Kronecker products is implemented only at first-order, at higher order we reset to analytic derivatives with sylvester equations
options . analytic_derivation_mode = 0 ; fprintf ( ' As order > 1, set ' ' options.analytic_derivation_mode' ' to 0\n' ) ;
end
% Get options from options structure
order = options . order ;
threads_BC = options . threads . kronecker . sparse_hessian_times_B_kronecker_C ;
qz_criterium = options . qz_criterium ;
analytic_derivation_mode = options . analytic_derivation_mode ;
% select method to compute Jacobians, default is 0
% * 0: efficient sylvester equation method to compute analytical derivatives as in Ratto & Iskrev (2012)
% * 1: kronecker products method to compute analytical derivatives as in Iskrev (2010), only for order=1
% * -1: numerical two-sided finite difference method to compute numerical derivatives of all output arguments using function get_perturbation_params_derivs_numerical_objective.m
% * -2: numerical two-sided finite difference method to compute numerically dYss, dg1, dg2, dg3, d2Yss and d2g1, the other output arguments are computed analytically as in kronflag=0
numerical_objective_fname = str2func ( ' get_perturbation_params_derivs_numerical_objective' ) ;
idx_states = M . nstatic + ( 1 : M . nspred ) ; %index for state variables, in DR order
modparam_nbr = length ( indpmodel ) ; %number of selected model parameters
stderrparam_nbr = length ( indpstderr ) ; %number of selected stderr parameters
corrparam_nbr = size ( indpcorr , 1 ) ; %number of selected corr parameters
totparam_nbr = modparam_nbr + stderrparam_nbr + corrparam_nbr ; %total number of selected parameters
[ I , ~ ] = find ( M . lead_lag_incidence ' ) ; %I is used to select nonzero columns of the Jacobian of endogenous variables in dynamic model files
yy0_nbr = length ( oo . dr . ys ( I ) ) ; %number of dynamic variables
yy0ex0_nbr = yy0_nbr + M . exo_nbr ; %number of dynamic variables + exogenous variables
kyy0 = nonzeros ( M . lead_lag_incidence ( : , oo . dr . order_var ) ' ) ; %index for nonzero entries in dynamic files at t-1,t,t+1 in DR order
kyy0ex0 = [ kyy0 ; length ( kyy0 ) + ( 1 : M . exo_nbr ) ' ] ; %dynamic files include derivatives wrt exogenous variables, note that exo_det is always 0
k2yy0ex0 = transpose ( reshape ( 1 : yy0ex0_nbr ^2 , yy0ex0_nbr , yy0ex0_nbr ) ) ; %index for the second dynamic derivatives, i.e. to evaluate the derivative of f wrt to yy0ex0(i) and yy0ex0(j), in DR order
k3yy0ex0 = permute ( reshape ( transpose ( reshape ( 1 : yy0ex0_nbr ^3 , yy0ex0_nbr , yy0ex0_nbr ^2 ) ) , yy0ex0_nbr , yy0ex0_nbr , yy0ex0_nbr ) , [ 2 1 3 ] ) ; %index for the third dynamic derivative, i.e. df/(dyyex0_i*dyyex0_j*dyyex0_k)
% Check for purely backward or forward looking models
if size ( M . lead_lag_incidence , 1 ) < 3
if M . nfwrd == 0 %purely backward models
klag = M . lead_lag_incidence ( 1 , oo . dr . order_var ) ; %indices of lagged (i.e. t-1) variables in dynamic files, columns are in DR order
kcurr = M . lead_lag_incidence ( 2 , oo . dr . order_var ) ; %indices of current (i.e. t) variables in dynamic files, columns are in DR order
klead = zeros ( 1 , size ( M . lead_lag_incidence , 2 ) ) ; %indices of lead (i.e. t+1) variables in dynamic files, columns are in DR order
elseif M . npred == 0 %purely forward models
klag = zeros ( 1 , size ( M . lead_lag_incidence , 2 ) ) ; %indices of lagged (i.e. t-1) variables in dynamic files, columns are in DR order
kcurr = M . lead_lag_incidence ( 1 , oo . dr . order_var ) ; %indices of current (i.e. t) variables in dynamic files, columns are in DR order
klead = M . lead_lag_incidence ( 2 , oo . dr . order_var ) ; %indices of lead (i.e. t+1) variables in dynamic files, columns are in DR order
end
else %normal models
klag = M . lead_lag_incidence ( 1 , oo . dr . order_var ) ; %indices of lagged (i.e. t-1) variables in dynamic files, columns are in DR order
kcurr = M . lead_lag_incidence ( 2 , oo . dr . order_var ) ; %indices of current (i.e. t) variables in dynamic files, columns are in DR order
klead = M . lead_lag_incidence ( 3 , oo . dr . order_var ) ; %indices of lead (i.e. t+1) variables in dynamic files, columns are in DR order
end
if analytic_derivation_mode < 0
%Create auxiliary estim_params blocks if not available for numerical derivatives, estim_params_model contains only model parameters
estim_params_model . np = length ( indpmodel ) ;
estim_params_model . param_vals ( : , 1 ) = indpmodel ;
estim_params_model . nvx = 0 ; estim_params_model . ncx = 0 ; estim_params_model . nvn = 0 ; estim_params_model . ncn = 0 ;
modparam1 = get_all_parameters ( estim_params_model , M ) ; %get all selected model parameters
if ~ isempty ( indpstderr ) && isempty ( estim_params . var_exo ) %if there are stderr parameters but no estimated_params_block
%provide temporary necessary information for stderr parameters
estim_params . nvx = length ( indpstderr ) ;
estim_params . var_exo ( : , 1 ) = indpstderr ;
end
if ~ isempty ( indpcorr ) && isempty ( estim_params . corrx ) %if there are corr parameters but no estimated_params_block
%provide temporary necessary information for stderr parameters
estim_params . ncx = size ( indpcorr , 1 ) ;
estim_params . corrx ( : , 1 : 2 ) = indpcorr ;
end
if ~ isfield ( estim_params , ' nvn' ) %stderr of measurement errors not yet
estim_params . nvn = 0 ;
estim_params . var_endo = [ ] ;
end
if ~ isfield ( estim_params , ' ncn' ) %corr of measurement errors not yet
estim_params . ncn = 0 ;
estim_params . corrn = [ ] ;
end
if ~ isempty ( indpmodel ) && isempty ( estim_params . param_vals ) %if there are model parameters but no estimated_params_block
%provide temporary necessary information for model parameters
estim_params . np = length ( indpmodel ) ;
estim_params . param_vals ( : , 1 ) = indpmodel ;
end
xparam1 = get_all_parameters ( estim_params , M ) ; %get all selected stderr, corr, and model parameters in estimated_params block, stderr and corr come first, then model parameters
end
if d2flag
modparam_nbr2 = modparam_nbr * ( modparam_nbr + 1 ) / 2 ; %number of unique entries of selected model parameters only in second-order derivative matrix
totparam_nbr2 = totparam_nbr * ( totparam_nbr + 1 ) / 2 ; %number of unique entries of all selected parameters in second-order derivative matrix
%get indices of elements in second derivatives of parameters
indp2tottot = reshape ( 1 : totparam_nbr ^2 , totparam_nbr , totparam_nbr ) ;
indp2stderrstderr = indp2tottot ( 1 : stderrparam_nbr , 1 : stderrparam_nbr ) ;
indp2stderrcorr = indp2tottot ( 1 : stderrparam_nbr , stderrparam_nbr + 1 : stderrparam_nbr + corrparam_nbr ) ;
indp2modmod = indp2tottot ( stderrparam_nbr + corrparam_nbr + 1 : stderrparam_nbr + corrparam_nbr + modparam_nbr , stderrparam_nbr + corrparam_nbr + 1 : stderrparam_nbr + corrparam_nbr + modparam_nbr ) ;
if totparam_nbr ~= 1
indp2tottot2 = dyn_vech ( indp2tottot ) ; %index of unique second-order derivatives
else
indp2tottot2 = indp2tottot ;
end
if modparam_nbr ~= 1
indp2modmod2 = dyn_vech ( indp2modmod ) ; %get rid of cross derivatives
else
indp2modmod2 = indp2modmod ;
end
%Kalman transition matrices, as in kalman_transition_matrix.m
KalmanA = zeros ( M . endo_nbr , M . endo_nbr ) ;
KalmanA ( : , idx_states ) = oo . dr . ghx ;
KalmanB = oo . dr . ghu ;
end
if analytic_derivation_mode == - 1
%% numerical two-sided finite difference method using function get_perturbation_params_derivs_numerical_objective.m (previously thet2tau.m in Dynare 4.5) for
% Jacobian (wrt selected stderr, corr and model parameters) of
% - dynamic model derivatives: dg1, dg2, dg3
% - steady state: dYss
% - covariance matrix of shocks: dSigma_e
% - perturbation solution matrices: dghx, dghu, dghxx, dghxu, dghuu, dghs2, dghxxx, dghxxu, dghxuu, dghuuu, dghxss, dghuss
%Parameter Jacobian of covariance matrix and solution matrices (wrt selected stderr, corr and model paramters)
dSig_gh = fjaco ( numerical_objective_fname , xparam1 , ' perturbation_solution' , estim_params , M , oo , options ) ;
ind_Sigma_e = ( 1 : M . exo_nbr ^2 ) ;
ind_ghx = ind_Sigma_e ( end ) + ( 1 : M . endo_nbr * M . nspred ) ;
ind_ghu = ind_ghx ( end ) + ( 1 : M . endo_nbr * M . exo_nbr ) ;
DERIVS . dSigma_e = reshape ( dSig_gh ( ind_Sigma_e , : ) , [ M . exo_nbr M . exo_nbr totparam_nbr ] ) ; %in tensor notation, wrt selected parameters
DERIVS . dghx = reshape ( dSig_gh ( ind_ghx , : ) , [ M . endo_nbr M . nspred totparam_nbr ] ) ; %in tensor notation, wrt selected parameters
DERIVS . dghu = reshape ( dSig_gh ( ind_ghu , : ) , [ M . endo_nbr M . exo_nbr totparam_nbr ] ) ; %in tensor notation, wrt selected parameters
if order > 1
ind_ghxx = ind_ghu ( end ) + ( 1 : M . endo_nbr * M . nspred ^2 ) ;
ind_ghxu = ind_ghxx ( end ) + ( 1 : M . endo_nbr * M . nspred * M . exo_nbr ) ;
ind_ghuu = ind_ghxu ( end ) + ( 1 : M . endo_nbr * M . exo_nbr * M . exo_nbr ) ;
ind_ghs2 = ind_ghuu ( end ) + ( 1 : M . endo_nbr ) ;
DERIVS . dghxx = reshape ( dSig_gh ( ind_ghxx , : ) , [ M . endo_nbr M . nspred ^2 totparam_nbr ] ) ; %in tensor notation, wrt selected parameters
DERIVS . dghxu = reshape ( dSig_gh ( ind_ghxu , : ) , [ M . endo_nbr M . nspred * M . exo_nbr totparam_nbr ] ) ; %in tensor notation, wrt selected parameters
DERIVS . dghuu = reshape ( dSig_gh ( ind_ghuu , : ) , [ M . endo_nbr M . exo_nbr * M . exo_nbr totparam_nbr ] ) ; %in tensor notation, wrt selected parameters
DERIVS . dghs2 = reshape ( dSig_gh ( ind_ghs2 , : ) , [ M . endo_nbr totparam_nbr ] ) ; %in tensor notation, wrt selected parameters
end
if order > 2
ind_ghxxx = ind_ghs2 ( end ) + ( 1 : M . endo_nbr * M . nspred ^3 ) ;
ind_ghxxu = ind_ghxxx ( end ) + ( 1 : M . endo_nbr * M . nspred ^2 * M . exo_nbr ) ;
ind_ghxuu = ind_ghxxu ( end ) + ( 1 : M . endo_nbr * M . nspred * M . exo_nbr ^2 ) ;
ind_ghuuu = ind_ghxuu ( end ) + ( 1 : M . endo_nbr * M . exo_nbr ^3 ) ;
ind_ghxss = ind_ghuuu ( end ) + ( 1 : M . endo_nbr * M . nspred ) ;
ind_ghuss = ind_ghxss ( end ) + ( 1 : M . endo_nbr * M . exo_nbr ) ;
DERIVS . dghxxx = reshape ( dSig_gh ( ind_ghxxx , : ) , [ M . endo_nbr M . nspred ^3 totparam_nbr ] ) ; %in tensor notation, wrt selected parameters
DERIVS . dghxxu = reshape ( dSig_gh ( ind_ghxxu , : ) , [ M . endo_nbr M . nspred ^2 * M . exo_nbr totparam_nbr ] ) ; %in tensor notation, wrt selected parameters
DERIVS . dghxuu = reshape ( dSig_gh ( ind_ghxuu , : ) , [ M . endo_nbr M . nspred * M . exo_nbr ^2 totparam_nbr ] ) ; %in tensor notation, wrt selected parameters
DERIVS . dghuuu = reshape ( dSig_gh ( ind_ghuuu , : ) , [ M . endo_nbr M . exo_nbr ^3 totparam_nbr ] ) ; %in tensor notation, wrt selected parameters
DERIVS . dghxss = reshape ( dSig_gh ( ind_ghxss , : ) , [ M . endo_nbr M . nspred totparam_nbr ] ) ; %in tensor notation, wrt selected parameters
DERIVS . dghuss = reshape ( dSig_gh ( ind_ghuss , : ) , [ M . endo_nbr M . exo_nbr totparam_nbr ] ) ; %in tensor notation, wrt selected parameters
end
2019-12-20 00:40:00 +01:00
% Parameter Jacobian of Om=ghu*Sigma_e*ghu' and Correlation_matrix (wrt selected stderr, corr and model paramters)
Added parameter derivatives of perturbation solution up to 3 order
# Preliminary comments
I finished the identification toolbox at orders two and three using the pruned state space system, but before I merge request this, I decided to first merge the new functionality to compute parameter derivatives of perturbation solution matrices at higher orders. So after this is approved, I merge the identification toolbox.
I guess @rattoma, @sebastien, and @michel are best choices to review this.
I outline the main idea first and then provide some more detailed changes I made to the functions.
***
# Main idea
This merge request is concerned with the *analytical*computation of the parameter derivatives of first, second and third order perturbation solution matrices, i.e. using _closed-form_ expressions to efficiently compute the derivative of $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ *with respect to model parameters* $\theta$. Note that $\theta$ contains model parameters, stderr and corr parameters of shocks. stderr and corr parameters of measurement errors are not yet supported, (they can easily be included as exogenous shocks). The availability of such derivatives is beneficial in terms of more reliable analysis of model sensitivity and parameter identifiability as well as more efficient estimation methods, in particular for models solved up to third order, as it is well-known that numerical derivatives are a tricky business, especially for large models.
References for my approach are:
* Iskrev (2008, 2010) and Schmitt-Grohé and Uribe (2012, Appendix) who were the first to compute the parameter derivatives analytically at first order, however, using inefficient (sparse) Kronecker products.
* Mutschler (2015) who provides the expressions for a second-order, but again using inefficient (sparse) Kronecker products.
* Ratto and Iskrev (2012) who show how the first-order system can be solved accurately, fast and efficiently using existing numerical algorithms for generalized Sylvester equations by taking the parameter derivative with respect to each parameter separately.
* Julliard and Kamenik (2004) who provide the perturbation solution equation system in tensor notation at any order k.
* Levintal (2017) who introduces permutation matrices to express the perturbation solution equation system in matrix notation up to fifth order.
Note that @rattoma already implemented the parameter derivatives of $g_x$ and $g_u$ analytically (and numerically), and I rely heavily on his work in `get_first_order_solution_params_derivs.m` (previously `getH.m`). My additions are mainly to this function and thus it is renamed to `get_perturbation_params_derivs.m`.
The basic idea of this merge request is to take the second- and third-order perturbation solution systems in Julliard and Kamenik (2004), unfold these into an equivalent matrix representation using permutation matrices as in Levintal (2017). Then extending Ratto and Iskrev (2012) one takes the derivative with respect to each parameter separately and gets a computational problem that is linear, albeit large, as it involves either solving generalized Sylvester equations or taking inverses of highly sparse matrices. I will now briefly summarize the perturbation solution system at third order and the system that results when taking the derivative with respect to parameters.
## Perturbation Solution
The following systems arise at first, second, and third order:
$(ghx): f_{x} z_{x} = f_{y_{-}^*} + f_{y_0} g_{x} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{x}= A g_{x} + f_{y_{-}^*}=0$
$(ghu): f_{z} z_{u} = f_{y_0} g_{u} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{u} + f_{u}= A g_u + f_u = 0$
$(ghxx) : A g_{xx} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{x}\right) + f_{zz} \left( z_{x} \otimes z_{x} \right) = 0$
$(ghxu) : A g_{xu} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{x} \otimes z_{u} \right) = 0$
$(ghuu) : A g_{uu} + B g_{xx} \left(g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{u} \otimes z_{u} \right) = 0$
$(ghs2) : (A+B) g_{\sigma\sigma} + \left( f_{y^{**}_{+}y^{**}_{+}} \left(g^{**}_{u} \otimes g^{**}_{u}\right) + f_{y^{**}_{+}} g^{**}_{uu}\right)vec(\Sigma) = 0$
$(ghxxx) : A g_{xxx} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{x}\right) + f_{y_{+}}g^{**}_{xx} \left(g^{*}_x \otimes g^{*}_{xx}\right)P_{x\_xx} + f_{zz} \left( z_{x} \otimes z_{xx} \right)P_{x\_xx} + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{x} \right) = 0$
$(ghxxu) : A g_{xxu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{u} \right) + f_{zz} \left( \left( z_{x} \otimes z_{xu} \right)P_{x\_xu} + \left(z_{xx} \otimes z_{u}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{x} \otimes g^{*}_{xu}\right)P_{x\_xu} + \left(g^{*}_{xx} \otimes g^{*}_{u}\right) \right) = 0$
$(ghxuu) : A g_{xuu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( \left( z_{xu} \otimes z_{u} \right)P_{xu\_u} + \left(z_{x} \otimes z_{uu}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{xu} \otimes g^{*}_{u}\right)P_{xu\_u} + \left(g^{*}_{x} \otimes g^{*}_{uu}\right) \right) = 0$
$(ghuuu) : A g_{uuu} + B g_{xxx} \left(g^{*}_{u} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{u} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( z_{u} \otimes z_{uu} \right)P_{u\_uu} + f_{y_{+}}g^{**}_{xx} \left(g^{*}_{u} \otimes g^{*}_{uu}\right)P_{u\_uu} = 0$
$(ghx\sigma\sigma) : A g_{x\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_x + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{x} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{x} \otimes z_{\sigma\sigma}\right) + F_{xu_{+}u_{+}}\left(I_{n_x} \otimes vec(\Sigma)\right) = 0$
$F_{xu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_x^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{xu_{+}} \otimes z_{u_{+}} \right)P_{xu\_u} + \left(z_{x} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{x} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
$(ghu\sigma\sigma) : A g_{u\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_{u} + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{u} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{u} \otimes z_{\sigma\sigma}\right) + F_{uu_{+}u_{+}}\left(I_{n_u} \otimes vec(\Sigma_u)\right) = 0$
$F_{uu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_u^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{uu_{+}} \otimes z_{u_{+}} \right)P_{uu\_u} + \left(z_{u} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{u} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
A and B are the common perturbation matrices:
$A = f_{y_0} + \begin{pmatrix} \underbrace{0}_{n\times n_{static}} &\vdots& \underbrace{f_{y^{**}_{+}} \cdot g^{**}_{x}}_{n \times n_{spred}} &\vdots& \underbrace{0}_{n\times n_{frwd}} \end{pmatrix}$and $B = \begin{pmatrix} \underbrace{0}_{n \times n_{static}}&\vdots & \underbrace{0}_{n \times n_{pred}} & \vdots & \underbrace{f_{y^{**}_{+}}}_{n \times n_{sfwrd}} \end{pmatrix}$
and $z=(y_{-}^{\ast}; y; y_{+}^{\ast\ast}; u)$ denotes the dynamic model variables as in `M_.lead_lag_incidence`, $y^\ast$ denote state variables, $y^{\ast\ast}$ denote forward looking variables, $y_+$ denote the variables with a lead, $y_{-}$ denote variables with a lag, $y_0$ denote variables at period t, $f$ the model equations, and $f_z$ the first-order dynamic model derivatives, $f_{zz}$ the second-order dynamic derivatives, and $f_{zzz}$ the third-order dynamic model derivatives. Then:
$z_{x} = \begin{pmatrix}I\\g_{x}\\g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{u} =\begin{pmatrix}0\\g_{u}\\g^{**}_{x} \cdot g^{*}_{u}\\I\end{pmatrix}$, $z_{u_{+}} =\begin{pmatrix}0\\0\\g^{**}_{u}\\0\end{pmatrix}$
$z_{xx} = \begin{pmatrix} 0\\g_{xx}\\g^{**}_{x} \left( g^{*}_x \otimes g^{*}_{x} \right) + g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{xu} =\begin{pmatrix}0\\g_{xu}\\g^{**}_{xx} \left( g^{*}_x \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{xu}\\0\end{pmatrix}$, $z_{uu} =\begin{pmatrix}0\\g_{uu}\\g^{**}_{xx} \left( g^{*}_u \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{uu}\\0\end{pmatrix}$,
$z_{xu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_x \otimes I \right)\\0\end{pmatrix}$, $z_{uu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_{u} \otimes I \right)\\0\end{pmatrix}$, $z_{u_{+}u_{+}} =\begin{pmatrix}0\\0\\g^{\ast\ast}_{uu}\\0\end{pmatrix}$, $z_{\sigma\sigma} = \begin{pmatrix}0\\ g_{\sigma\sigma}\\ g^{\ast\ast}_{x}g^{\ast}_{\sigma\sigma} + g^{\ast\ast}_{\sigma\sigma}\\0 \end{pmatrix}$
$P$ are permutation matrices that can be computed using Matlab's `ipermute` function.
## Parameter derivatives of perturbation solutions
First, we need the parameter derivatives of first, second, third, and fourth derivatives of the dynamic model (i.e. g1,g2,g3,g4 in dynamic files), I make use of the implicit function theorem: Let $f_{z^k}$ denote the kth derivative (wrt all dynamic variables) of the dynamic model, then let $df_{z^k}$ denote the first-derivative (wrt all model parameters) of $f_{z^k}$ evaluated at the steady state. Note that $f_{z^k}$ is a function of both the model parameters $\theta$ and of the steady state of all dynamic variables $\bar{z}$, which also depend on the parameters. Hence, implicitly $f_{z^k}=f_{z^k}(\theta,\bar{z}(\theta))$ and $df_{z^k}$ consists of two parts:
1. direct derivative wrt to all model parameters given by the preprocessor in the `_params_derivs.m` files
2. contribution of derivative of steady state of dynamic variables (wrt all model parameters): $f_{z^{k+1}} \cdot d\bar{z}$
Note that we already have functionality to compute $d\bar{z}$ analytically.
Having this, the above perturbation systems are basically equations of the following types
$AX +BXC = RHS$ or $AX = RHS$
Now when taking the derivative (wrt to one single parameter $\theta_j$), we get
$A\mathrm{d}\{X\} + B\mathrm{d}\{X\}C = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X - \mathrm{d}\{B\}XC - BX\mathrm{d}\{C\}$
or
$A\mathrm{d}\{X\} = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X$
The first one is a Sylvester type equation, the second one can be solved by taking the inverse of $A$. The only diffculty and tedious work arrises in computing (the highly sparse) derivatives of $RHS$.
***
# New functions: `
## get_perturbation_params_derivs.m`and `get_perturbation_params_derivs_numerical_objective.m`
* The parameter derivatives up to third order are computed in the new function`get_perturbation_params_derivs.m` both analytically and numerically. For numerical derivatives `get_perturbation_params_derivs_numerical_objective.m` is the objective for `fjaco.m` or `hessian_sparse.m` or `hessian.m`.
* `get_perturbation_params_derivs.m` is basically an extended version of the previous `get_first_order_solution_params_derivs.m` function.
* * `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It is used for numerical derivatives, whenever `analytic_derivation_mode=-1|-2`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
* Detailed changes:
* Most important: notation of this function is now in accordance to the k_order_solver, i.e. we do not compute derivatives of Kalman transition matrices A and B, but rather the solution matrices ghx,ghu,ghxx,ghxu,ghuu,ghs2,ghxxx,ghxxu,ghxuu,ghuuu,ghxss,ghuss in accordance with notation used in `oo_.dr`. As a byproduct at first-order, focusing on ghx and ghu instead of Kalman transition matrices A and B makes the computations slightly faster for large models (e.g. for Quest the computations were faster by a couple of seconds, not much, but okay).
* Removed use of `kstate`, see also Dynare/dynare#1653 and Dynare/dynare!1656
* Output arguments are stored in a structure `DERIVS`, there is also a flag `d2flag` that computes parameter hessians needed only in `dsge_likelihood.m`.
* Removed `kronflag` as input. `options_.analytic_derivation_mode` is now used instead of `kronflag`.
* Removed `indvar`, an index that was used to selected specific variables in the derivatives. This is not needed, as we always compute the parameter derivatives for all variables first and then select a subset of variables. The selection now takes place in other functions, like `dsge_likelihood.m`.
* Introduced some checks: (i) deterministic exogenous variables are not supported, (ii) Kronecker method only compatible with first-order approximation so reset to sylvester method, (iii) for purely backward or forward models we need to be careful with the rows in `M_.lead_la g_incidence`, (iv) if `_params_derivs.m` files are missing an error is thrown.
* For numerical derivatives, if mod file does not contain an `estimated_params_block`, a temporary one with the most important parameter information is created.
## `unfold_g4.m`
* When evaluating g3 and g4 one needs to take into account that these do not contain symmetric elements, so one needs to use `unfold_g3.m` and the new function `unfold_g4.m`. This returns an unfolded version of the same matrix (i.e. with symmetric elements).
***
# New test models
`.gitignore` and `Makefile.am` are changed accordingly. Also now it is possible to run test suite on analytic_derivatives, i.e. run `make check m/analytic_derivatives`
## `analytic_derivatives/BrockMirman_PertParamsDerivs.mod`
* This is the Brock Mirman model, where we know the exact policy function $g$ for capital and consumption. As this does not imply a nonzero $g_{\sigma\sigma}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ I added some artificial equations to get nonzero solution matrices with respect to $\sigma$. The true perturbation solution matrices $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ are then computed analytically with Matlab's symbolic toolbox and saved in `nBrockMirmanSYM.mat`. There is a preprocessor flag that recreates these analytical computations if changes are needed (and to check whether I made some errors here ;-) )
* Then solution matrices up to third order and their parameter Jacobians are then compared to the ones computed by Dynare's `k_order_solver` and by `get_perturbation_params_derivs` for all `analytic_derivation_mode`'s. There will be an error if the maximum absolute deviation is too large, i.e. for numerical derivatives (`analytic_derivation_mode=-1|-2`) the tolerance is choosen lower (around 1e-5); for analytical methods we are stricter: around 1e-13 for first-order, 1e-12 for second order, and 1e-11 for third-order.
* As a side note, this mod file also checks Dynare's `k_order_solver` algorithm and throws an error if something is wrong.
* This test model shows that the new functionality works well. And analytical derivatives perform way better and accurate than numerical ones, even for this small model.
## `analytic_derivatives/burnside_3_order_PertParamsDerivs.mod`
* This builds upon `tests/k_order_perturbation/burnside_k_order.mod` and computes the true parameter derivatives analytically by hand.
* This test model also shows that the new functionality works well.
## `analytic_derivatives/LindeTrabandt2019.mod`
* Shows that the new functionality also works for medium-sized models, i.e. a SW type model solved at third order with 35 variables (11 states). 2 shocks and 20 parameters.
* This mod file can be used to tweak the speed of the computations in the future.
* Compares numerical versus analytical parameter derivatives (for first, second and third order). Note that this model clearly shows that numerical ones are quite different than analytical ones even at first order!
## `identification/LindeTrabandt2019_xfail.mod`
* This model is a check for issue Dynare/dynare#1595, see fjaco.m below, and will fail.
* Removed `analytic_derivatives/ls2003.mod` as this mod file is neither in the testsuite nor does it work.
***
# Detailed changes in other functions
## `get_first_order_solution_params_derivs.m`
* Deleted, or actually, renamed to `get_perturbation_params_derivs.m`, as this function now is able to compute the derivatives up to third order
## `identification_numerical_objective.m`
* `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
## `dsge_likelihood.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `T`matrix anymore, but instead of the dynare solution matrix `ghx`. So we recreate `T` here as this amounts to adding some zeros and focusing on selected variables only.
* Added some checks to make sure the first-order approximation is selected.
* Removed `kron_flag` as input, as `get_perturbation_params_derivs` looks into `options_.analytic_derivation_mode` for `kron_flag`.
## `dynare_identification.m`
* make sure that setting `analytic_derivation_mode` is set both in `options_ident` and `options_`. Note that at the end of the function we restore the `options_` structure, so all changes are local. In a next merge request, I will remove the global variables to make all variables local.
## `get_identification_jacobians.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `A` and `B` matrix anymore, but instead of the dynare solution matrix `ghx` and `ghu`. So we recreate these matrices here instead of in `get_perturbation_params_derivs.m`.
* Added `str2func` for better function handles in `fjaco.m`.
## `fjaco.m`
* make `tol`an option, which can be adjusted by changing `options_.dynatol.x`for identification and parameter derivatives purposes.
* include a check and an informative error message, if numerical derivatives (two-sided finite difference method) yield errors in `resol.m` for identification and parameter derivatives purposes. This closes issue Dynare/dynare#1595.
* Changed year of copyright to 2010-2017,2019
***
# Further suggestions and questions
* Ones this is merged, I will merge request an improvement of the identification toolbox, which will work up to third order using the pruned state space. This will also remove some issues and bugs, and also I will remove global variables in this request.
* The third-order derivatives can be further improved by taking sparsity into account and use mex versions for kronecker products etc. I leave this for further testing (and if anybody actually uses this ;-) )
2019-12-17 19:17:09 +01:00
DERIVS . dOm = zeros ( M . endo_nbr , M . endo_nbr , totparam_nbr ) ; %initialize in tensor notation
2019-12-20 00:40:00 +01:00
DERIVS . dCorrelation_matrix = zeros ( M . exo_nbr , M . exo_nbr , totparam_nbr ) ; %initialize
Added parameter derivatives of perturbation solution up to 3 order
# Preliminary comments
I finished the identification toolbox at orders two and three using the pruned state space system, but before I merge request this, I decided to first merge the new functionality to compute parameter derivatives of perturbation solution matrices at higher orders. So after this is approved, I merge the identification toolbox.
I guess @rattoma, @sebastien, and @michel are best choices to review this.
I outline the main idea first and then provide some more detailed changes I made to the functions.
***
# Main idea
This merge request is concerned with the *analytical*computation of the parameter derivatives of first, second and third order perturbation solution matrices, i.e. using _closed-form_ expressions to efficiently compute the derivative of $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ *with respect to model parameters* $\theta$. Note that $\theta$ contains model parameters, stderr and corr parameters of shocks. stderr and corr parameters of measurement errors are not yet supported, (they can easily be included as exogenous shocks). The availability of such derivatives is beneficial in terms of more reliable analysis of model sensitivity and parameter identifiability as well as more efficient estimation methods, in particular for models solved up to third order, as it is well-known that numerical derivatives are a tricky business, especially for large models.
References for my approach are:
* Iskrev (2008, 2010) and Schmitt-Grohé and Uribe (2012, Appendix) who were the first to compute the parameter derivatives analytically at first order, however, using inefficient (sparse) Kronecker products.
* Mutschler (2015) who provides the expressions for a second-order, but again using inefficient (sparse) Kronecker products.
* Ratto and Iskrev (2012) who show how the first-order system can be solved accurately, fast and efficiently using existing numerical algorithms for generalized Sylvester equations by taking the parameter derivative with respect to each parameter separately.
* Julliard and Kamenik (2004) who provide the perturbation solution equation system in tensor notation at any order k.
* Levintal (2017) who introduces permutation matrices to express the perturbation solution equation system in matrix notation up to fifth order.
Note that @rattoma already implemented the parameter derivatives of $g_x$ and $g_u$ analytically (and numerically), and I rely heavily on his work in `get_first_order_solution_params_derivs.m` (previously `getH.m`). My additions are mainly to this function and thus it is renamed to `get_perturbation_params_derivs.m`.
The basic idea of this merge request is to take the second- and third-order perturbation solution systems in Julliard and Kamenik (2004), unfold these into an equivalent matrix representation using permutation matrices as in Levintal (2017). Then extending Ratto and Iskrev (2012) one takes the derivative with respect to each parameter separately and gets a computational problem that is linear, albeit large, as it involves either solving generalized Sylvester equations or taking inverses of highly sparse matrices. I will now briefly summarize the perturbation solution system at third order and the system that results when taking the derivative with respect to parameters.
## Perturbation Solution
The following systems arise at first, second, and third order:
$(ghx): f_{x} z_{x} = f_{y_{-}^*} + f_{y_0} g_{x} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{x}= A g_{x} + f_{y_{-}^*}=0$
$(ghu): f_{z} z_{u} = f_{y_0} g_{u} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{u} + f_{u}= A g_u + f_u = 0$
$(ghxx) : A g_{xx} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{x}\right) + f_{zz} \left( z_{x} \otimes z_{x} \right) = 0$
$(ghxu) : A g_{xu} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{x} \otimes z_{u} \right) = 0$
$(ghuu) : A g_{uu} + B g_{xx} \left(g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{u} \otimes z_{u} \right) = 0$
$(ghs2) : (A+B) g_{\sigma\sigma} + \left( f_{y^{**}_{+}y^{**}_{+}} \left(g^{**}_{u} \otimes g^{**}_{u}\right) + f_{y^{**}_{+}} g^{**}_{uu}\right)vec(\Sigma) = 0$
$(ghxxx) : A g_{xxx} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{x}\right) + f_{y_{+}}g^{**}_{xx} \left(g^{*}_x \otimes g^{*}_{xx}\right)P_{x\_xx} + f_{zz} \left( z_{x} \otimes z_{xx} \right)P_{x\_xx} + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{x} \right) = 0$
$(ghxxu) : A g_{xxu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{u} \right) + f_{zz} \left( \left( z_{x} \otimes z_{xu} \right)P_{x\_xu} + \left(z_{xx} \otimes z_{u}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{x} \otimes g^{*}_{xu}\right)P_{x\_xu} + \left(g^{*}_{xx} \otimes g^{*}_{u}\right) \right) = 0$
$(ghxuu) : A g_{xuu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( \left( z_{xu} \otimes z_{u} \right)P_{xu\_u} + \left(z_{x} \otimes z_{uu}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{xu} \otimes g^{*}_{u}\right)P_{xu\_u} + \left(g^{*}_{x} \otimes g^{*}_{uu}\right) \right) = 0$
$(ghuuu) : A g_{uuu} + B g_{xxx} \left(g^{*}_{u} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{u} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( z_{u} \otimes z_{uu} \right)P_{u\_uu} + f_{y_{+}}g^{**}_{xx} \left(g^{*}_{u} \otimes g^{*}_{uu}\right)P_{u\_uu} = 0$
$(ghx\sigma\sigma) : A g_{x\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_x + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{x} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{x} \otimes z_{\sigma\sigma}\right) + F_{xu_{+}u_{+}}\left(I_{n_x} \otimes vec(\Sigma)\right) = 0$
$F_{xu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_x^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{xu_{+}} \otimes z_{u_{+}} \right)P_{xu\_u} + \left(z_{x} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{x} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
$(ghu\sigma\sigma) : A g_{u\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_{u} + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{u} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{u} \otimes z_{\sigma\sigma}\right) + F_{uu_{+}u_{+}}\left(I_{n_u} \otimes vec(\Sigma_u)\right) = 0$
$F_{uu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_u^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{uu_{+}} \otimes z_{u_{+}} \right)P_{uu\_u} + \left(z_{u} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{u} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
A and B are the common perturbation matrices:
$A = f_{y_0} + \begin{pmatrix} \underbrace{0}_{n\times n_{static}} &\vdots& \underbrace{f_{y^{**}_{+}} \cdot g^{**}_{x}}_{n \times n_{spred}} &\vdots& \underbrace{0}_{n\times n_{frwd}} \end{pmatrix}$and $B = \begin{pmatrix} \underbrace{0}_{n \times n_{static}}&\vdots & \underbrace{0}_{n \times n_{pred}} & \vdots & \underbrace{f_{y^{**}_{+}}}_{n \times n_{sfwrd}} \end{pmatrix}$
and $z=(y_{-}^{\ast}; y; y_{+}^{\ast\ast}; u)$ denotes the dynamic model variables as in `M_.lead_lag_incidence`, $y^\ast$ denote state variables, $y^{\ast\ast}$ denote forward looking variables, $y_+$ denote the variables with a lead, $y_{-}$ denote variables with a lag, $y_0$ denote variables at period t, $f$ the model equations, and $f_z$ the first-order dynamic model derivatives, $f_{zz}$ the second-order dynamic derivatives, and $f_{zzz}$ the third-order dynamic model derivatives. Then:
$z_{x} = \begin{pmatrix}I\\g_{x}\\g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{u} =\begin{pmatrix}0\\g_{u}\\g^{**}_{x} \cdot g^{*}_{u}\\I\end{pmatrix}$, $z_{u_{+}} =\begin{pmatrix}0\\0\\g^{**}_{u}\\0\end{pmatrix}$
$z_{xx} = \begin{pmatrix} 0\\g_{xx}\\g^{**}_{x} \left( g^{*}_x \otimes g^{*}_{x} \right) + g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{xu} =\begin{pmatrix}0\\g_{xu}\\g^{**}_{xx} \left( g^{*}_x \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{xu}\\0\end{pmatrix}$, $z_{uu} =\begin{pmatrix}0\\g_{uu}\\g^{**}_{xx} \left( g^{*}_u \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{uu}\\0\end{pmatrix}$,
$z_{xu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_x \otimes I \right)\\0\end{pmatrix}$, $z_{uu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_{u} \otimes I \right)\\0\end{pmatrix}$, $z_{u_{+}u_{+}} =\begin{pmatrix}0\\0\\g^{\ast\ast}_{uu}\\0\end{pmatrix}$, $z_{\sigma\sigma} = \begin{pmatrix}0\\ g_{\sigma\sigma}\\ g^{\ast\ast}_{x}g^{\ast}_{\sigma\sigma} + g^{\ast\ast}_{\sigma\sigma}\\0 \end{pmatrix}$
$P$ are permutation matrices that can be computed using Matlab's `ipermute` function.
## Parameter derivatives of perturbation solutions
First, we need the parameter derivatives of first, second, third, and fourth derivatives of the dynamic model (i.e. g1,g2,g3,g4 in dynamic files), I make use of the implicit function theorem: Let $f_{z^k}$ denote the kth derivative (wrt all dynamic variables) of the dynamic model, then let $df_{z^k}$ denote the first-derivative (wrt all model parameters) of $f_{z^k}$ evaluated at the steady state. Note that $f_{z^k}$ is a function of both the model parameters $\theta$ and of the steady state of all dynamic variables $\bar{z}$, which also depend on the parameters. Hence, implicitly $f_{z^k}=f_{z^k}(\theta,\bar{z}(\theta))$ and $df_{z^k}$ consists of two parts:
1. direct derivative wrt to all model parameters given by the preprocessor in the `_params_derivs.m` files
2. contribution of derivative of steady state of dynamic variables (wrt all model parameters): $f_{z^{k+1}} \cdot d\bar{z}$
Note that we already have functionality to compute $d\bar{z}$ analytically.
Having this, the above perturbation systems are basically equations of the following types
$AX +BXC = RHS$ or $AX = RHS$
Now when taking the derivative (wrt to one single parameter $\theta_j$), we get
$A\mathrm{d}\{X\} + B\mathrm{d}\{X\}C = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X - \mathrm{d}\{B\}XC - BX\mathrm{d}\{C\}$
or
$A\mathrm{d}\{X\} = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X$
The first one is a Sylvester type equation, the second one can be solved by taking the inverse of $A$. The only diffculty and tedious work arrises in computing (the highly sparse) derivatives of $RHS$.
***
# New functions: `
## get_perturbation_params_derivs.m`and `get_perturbation_params_derivs_numerical_objective.m`
* The parameter derivatives up to third order are computed in the new function`get_perturbation_params_derivs.m` both analytically and numerically. For numerical derivatives `get_perturbation_params_derivs_numerical_objective.m` is the objective for `fjaco.m` or `hessian_sparse.m` or `hessian.m`.
* `get_perturbation_params_derivs.m` is basically an extended version of the previous `get_first_order_solution_params_derivs.m` function.
* * `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It is used for numerical derivatives, whenever `analytic_derivation_mode=-1|-2`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
* Detailed changes:
* Most important: notation of this function is now in accordance to the k_order_solver, i.e. we do not compute derivatives of Kalman transition matrices A and B, but rather the solution matrices ghx,ghu,ghxx,ghxu,ghuu,ghs2,ghxxx,ghxxu,ghxuu,ghuuu,ghxss,ghuss in accordance with notation used in `oo_.dr`. As a byproduct at first-order, focusing on ghx and ghu instead of Kalman transition matrices A and B makes the computations slightly faster for large models (e.g. for Quest the computations were faster by a couple of seconds, not much, but okay).
* Removed use of `kstate`, see also Dynare/dynare#1653 and Dynare/dynare!1656
* Output arguments are stored in a structure `DERIVS`, there is also a flag `d2flag` that computes parameter hessians needed only in `dsge_likelihood.m`.
* Removed `kronflag` as input. `options_.analytic_derivation_mode` is now used instead of `kronflag`.
* Removed `indvar`, an index that was used to selected specific variables in the derivatives. This is not needed, as we always compute the parameter derivatives for all variables first and then select a subset of variables. The selection now takes place in other functions, like `dsge_likelihood.m`.
* Introduced some checks: (i) deterministic exogenous variables are not supported, (ii) Kronecker method only compatible with first-order approximation so reset to sylvester method, (iii) for purely backward or forward models we need to be careful with the rows in `M_.lead_la g_incidence`, (iv) if `_params_derivs.m` files are missing an error is thrown.
* For numerical derivatives, if mod file does not contain an `estimated_params_block`, a temporary one with the most important parameter information is created.
## `unfold_g4.m`
* When evaluating g3 and g4 one needs to take into account that these do not contain symmetric elements, so one needs to use `unfold_g3.m` and the new function `unfold_g4.m`. This returns an unfolded version of the same matrix (i.e. with symmetric elements).
***
# New test models
`.gitignore` and `Makefile.am` are changed accordingly. Also now it is possible to run test suite on analytic_derivatives, i.e. run `make check m/analytic_derivatives`
## `analytic_derivatives/BrockMirman_PertParamsDerivs.mod`
* This is the Brock Mirman model, where we know the exact policy function $g$ for capital and consumption. As this does not imply a nonzero $g_{\sigma\sigma}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ I added some artificial equations to get nonzero solution matrices with respect to $\sigma$. The true perturbation solution matrices $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ are then computed analytically with Matlab's symbolic toolbox and saved in `nBrockMirmanSYM.mat`. There is a preprocessor flag that recreates these analytical computations if changes are needed (and to check whether I made some errors here ;-) )
* Then solution matrices up to third order and their parameter Jacobians are then compared to the ones computed by Dynare's `k_order_solver` and by `get_perturbation_params_derivs` for all `analytic_derivation_mode`'s. There will be an error if the maximum absolute deviation is too large, i.e. for numerical derivatives (`analytic_derivation_mode=-1|-2`) the tolerance is choosen lower (around 1e-5); for analytical methods we are stricter: around 1e-13 for first-order, 1e-12 for second order, and 1e-11 for third-order.
* As a side note, this mod file also checks Dynare's `k_order_solver` algorithm and throws an error if something is wrong.
* This test model shows that the new functionality works well. And analytical derivatives perform way better and accurate than numerical ones, even for this small model.
## `analytic_derivatives/burnside_3_order_PertParamsDerivs.mod`
* This builds upon `tests/k_order_perturbation/burnside_k_order.mod` and computes the true parameter derivatives analytically by hand.
* This test model also shows that the new functionality works well.
## `analytic_derivatives/LindeTrabandt2019.mod`
* Shows that the new functionality also works for medium-sized models, i.e. a SW type model solved at third order with 35 variables (11 states). 2 shocks and 20 parameters.
* This mod file can be used to tweak the speed of the computations in the future.
* Compares numerical versus analytical parameter derivatives (for first, second and third order). Note that this model clearly shows that numerical ones are quite different than analytical ones even at first order!
## `identification/LindeTrabandt2019_xfail.mod`
* This model is a check for issue Dynare/dynare#1595, see fjaco.m below, and will fail.
* Removed `analytic_derivatives/ls2003.mod` as this mod file is neither in the testsuite nor does it work.
***
# Detailed changes in other functions
## `get_first_order_solution_params_derivs.m`
* Deleted, or actually, renamed to `get_perturbation_params_derivs.m`, as this function now is able to compute the derivatives up to third order
## `identification_numerical_objective.m`
* `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
## `dsge_likelihood.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `T`matrix anymore, but instead of the dynare solution matrix `ghx`. So we recreate `T` here as this amounts to adding some zeros and focusing on selected variables only.
* Added some checks to make sure the first-order approximation is selected.
* Removed `kron_flag` as input, as `get_perturbation_params_derivs` looks into `options_.analytic_derivation_mode` for `kron_flag`.
## `dynare_identification.m`
* make sure that setting `analytic_derivation_mode` is set both in `options_ident` and `options_`. Note that at the end of the function we restore the `options_` structure, so all changes are local. In a next merge request, I will remove the global variables to make all variables local.
## `get_identification_jacobians.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `A` and `B` matrix anymore, but instead of the dynare solution matrix `ghx` and `ghu`. So we recreate these matrices here instead of in `get_perturbation_params_derivs.m`.
* Added `str2func` for better function handles in `fjaco.m`.
## `fjaco.m`
* make `tol`an option, which can be adjusted by changing `options_.dynatol.x`for identification and parameter derivatives purposes.
* include a check and an informative error message, if numerical derivatives (two-sided finite difference method) yield errors in `resol.m` for identification and parameter derivatives purposes. This closes issue Dynare/dynare#1595.
* Changed year of copyright to 2010-2017,2019
***
# Further suggestions and questions
* Ones this is merged, I will merge request an improvement of the identification toolbox, which will work up to third order using the pruned state space. This will also remove some issues and bugs, and also I will remove global variables in this request.
* The third-order derivatives can be further improved by taking sparsity into account and use mex versions for kronecker products etc. I leave this for further testing (and if anybody actually uses this ;-) )
2019-12-17 19:17:09 +01:00
if ~ isempty ( indpstderr ) %derivatives of ghu wrt stderr parameters are zero by construction
for jp = 1 : stderrparam_nbr
DERIVS . dOm ( : , : , jp ) = oo . dr . ghu * DERIVS . dSigma_e ( : , : , jp ) * oo . dr . ghu ' ;
end
end
if ~ isempty ( indpcorr ) %derivatives of ghu wrt corr parameters are zero by construction
for jp = 1 : corrparam_nbr
DERIVS . dOm ( : , : , stderrparam_nbr + jp ) = oo . dr . ghu * DERIVS . dSigma_e ( : , : , stderrparam_nbr + jp ) * oo . dr . ghu ' ;
2019-12-20 00:40:00 +01:00
DERIVS . dCorrelation_matrix ( indpcorr ( jp , 1 ) , indpcorr ( jp , 2 ) , stderrparam_nbr + jp ) = 1 ;
DERIVS . dCorrelation_matrix ( indpcorr ( jp , 2 ) , indpcorr ( jp , 1 ) , stderrparam_nbr + jp ) = 1 ; %symmetry
Added parameter derivatives of perturbation solution up to 3 order
# Preliminary comments
I finished the identification toolbox at orders two and three using the pruned state space system, but before I merge request this, I decided to first merge the new functionality to compute parameter derivatives of perturbation solution matrices at higher orders. So after this is approved, I merge the identification toolbox.
I guess @rattoma, @sebastien, and @michel are best choices to review this.
I outline the main idea first and then provide some more detailed changes I made to the functions.
***
# Main idea
This merge request is concerned with the *analytical*computation of the parameter derivatives of first, second and third order perturbation solution matrices, i.e. using _closed-form_ expressions to efficiently compute the derivative of $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ *with respect to model parameters* $\theta$. Note that $\theta$ contains model parameters, stderr and corr parameters of shocks. stderr and corr parameters of measurement errors are not yet supported, (they can easily be included as exogenous shocks). The availability of such derivatives is beneficial in terms of more reliable analysis of model sensitivity and parameter identifiability as well as more efficient estimation methods, in particular for models solved up to third order, as it is well-known that numerical derivatives are a tricky business, especially for large models.
References for my approach are:
* Iskrev (2008, 2010) and Schmitt-Grohé and Uribe (2012, Appendix) who were the first to compute the parameter derivatives analytically at first order, however, using inefficient (sparse) Kronecker products.
* Mutschler (2015) who provides the expressions for a second-order, but again using inefficient (sparse) Kronecker products.
* Ratto and Iskrev (2012) who show how the first-order system can be solved accurately, fast and efficiently using existing numerical algorithms for generalized Sylvester equations by taking the parameter derivative with respect to each parameter separately.
* Julliard and Kamenik (2004) who provide the perturbation solution equation system in tensor notation at any order k.
* Levintal (2017) who introduces permutation matrices to express the perturbation solution equation system in matrix notation up to fifth order.
Note that @rattoma already implemented the parameter derivatives of $g_x$ and $g_u$ analytically (and numerically), and I rely heavily on his work in `get_first_order_solution_params_derivs.m` (previously `getH.m`). My additions are mainly to this function and thus it is renamed to `get_perturbation_params_derivs.m`.
The basic idea of this merge request is to take the second- and third-order perturbation solution systems in Julliard and Kamenik (2004), unfold these into an equivalent matrix representation using permutation matrices as in Levintal (2017). Then extending Ratto and Iskrev (2012) one takes the derivative with respect to each parameter separately and gets a computational problem that is linear, albeit large, as it involves either solving generalized Sylvester equations or taking inverses of highly sparse matrices. I will now briefly summarize the perturbation solution system at third order and the system that results when taking the derivative with respect to parameters.
## Perturbation Solution
The following systems arise at first, second, and third order:
$(ghx): f_{x} z_{x} = f_{y_{-}^*} + f_{y_0} g_{x} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{x}= A g_{x} + f_{y_{-}^*}=0$
$(ghu): f_{z} z_{u} = f_{y_0} g_{u} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{u} + f_{u}= A g_u + f_u = 0$
$(ghxx) : A g_{xx} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{x}\right) + f_{zz} \left( z_{x} \otimes z_{x} \right) = 0$
$(ghxu) : A g_{xu} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{x} \otimes z_{u} \right) = 0$
$(ghuu) : A g_{uu} + B g_{xx} \left(g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{u} \otimes z_{u} \right) = 0$
$(ghs2) : (A+B) g_{\sigma\sigma} + \left( f_{y^{**}_{+}y^{**}_{+}} \left(g^{**}_{u} \otimes g^{**}_{u}\right) + f_{y^{**}_{+}} g^{**}_{uu}\right)vec(\Sigma) = 0$
$(ghxxx) : A g_{xxx} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{x}\right) + f_{y_{+}}g^{**}_{xx} \left(g^{*}_x \otimes g^{*}_{xx}\right)P_{x\_xx} + f_{zz} \left( z_{x} \otimes z_{xx} \right)P_{x\_xx} + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{x} \right) = 0$
$(ghxxu) : A g_{xxu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{u} \right) + f_{zz} \left( \left( z_{x} \otimes z_{xu} \right)P_{x\_xu} + \left(z_{xx} \otimes z_{u}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{x} \otimes g^{*}_{xu}\right)P_{x\_xu} + \left(g^{*}_{xx} \otimes g^{*}_{u}\right) \right) = 0$
$(ghxuu) : A g_{xuu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( \left( z_{xu} \otimes z_{u} \right)P_{xu\_u} + \left(z_{x} \otimes z_{uu}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{xu} \otimes g^{*}_{u}\right)P_{xu\_u} + \left(g^{*}_{x} \otimes g^{*}_{uu}\right) \right) = 0$
$(ghuuu) : A g_{uuu} + B g_{xxx} \left(g^{*}_{u} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{u} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( z_{u} \otimes z_{uu} \right)P_{u\_uu} + f_{y_{+}}g^{**}_{xx} \left(g^{*}_{u} \otimes g^{*}_{uu}\right)P_{u\_uu} = 0$
$(ghx\sigma\sigma) : A g_{x\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_x + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{x} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{x} \otimes z_{\sigma\sigma}\right) + F_{xu_{+}u_{+}}\left(I_{n_x} \otimes vec(\Sigma)\right) = 0$
$F_{xu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_x^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{xu_{+}} \otimes z_{u_{+}} \right)P_{xu\_u} + \left(z_{x} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{x} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
$(ghu\sigma\sigma) : A g_{u\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_{u} + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{u} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{u} \otimes z_{\sigma\sigma}\right) + F_{uu_{+}u_{+}}\left(I_{n_u} \otimes vec(\Sigma_u)\right) = 0$
$F_{uu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_u^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{uu_{+}} \otimes z_{u_{+}} \right)P_{uu\_u} + \left(z_{u} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{u} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
A and B are the common perturbation matrices:
$A = f_{y_0} + \begin{pmatrix} \underbrace{0}_{n\times n_{static}} &\vdots& \underbrace{f_{y^{**}_{+}} \cdot g^{**}_{x}}_{n \times n_{spred}} &\vdots& \underbrace{0}_{n\times n_{frwd}} \end{pmatrix}$and $B = \begin{pmatrix} \underbrace{0}_{n \times n_{static}}&\vdots & \underbrace{0}_{n \times n_{pred}} & \vdots & \underbrace{f_{y^{**}_{+}}}_{n \times n_{sfwrd}} \end{pmatrix}$
and $z=(y_{-}^{\ast}; y; y_{+}^{\ast\ast}; u)$ denotes the dynamic model variables as in `M_.lead_lag_incidence`, $y^\ast$ denote state variables, $y^{\ast\ast}$ denote forward looking variables, $y_+$ denote the variables with a lead, $y_{-}$ denote variables with a lag, $y_0$ denote variables at period t, $f$ the model equations, and $f_z$ the first-order dynamic model derivatives, $f_{zz}$ the second-order dynamic derivatives, and $f_{zzz}$ the third-order dynamic model derivatives. Then:
$z_{x} = \begin{pmatrix}I\\g_{x}\\g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{u} =\begin{pmatrix}0\\g_{u}\\g^{**}_{x} \cdot g^{*}_{u}\\I\end{pmatrix}$, $z_{u_{+}} =\begin{pmatrix}0\\0\\g^{**}_{u}\\0\end{pmatrix}$
$z_{xx} = \begin{pmatrix} 0\\g_{xx}\\g^{**}_{x} \left( g^{*}_x \otimes g^{*}_{x} \right) + g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{xu} =\begin{pmatrix}0\\g_{xu}\\g^{**}_{xx} \left( g^{*}_x \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{xu}\\0\end{pmatrix}$, $z_{uu} =\begin{pmatrix}0\\g_{uu}\\g^{**}_{xx} \left( g^{*}_u \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{uu}\\0\end{pmatrix}$,
$z_{xu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_x \otimes I \right)\\0\end{pmatrix}$, $z_{uu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_{u} \otimes I \right)\\0\end{pmatrix}$, $z_{u_{+}u_{+}} =\begin{pmatrix}0\\0\\g^{\ast\ast}_{uu}\\0\end{pmatrix}$, $z_{\sigma\sigma} = \begin{pmatrix}0\\ g_{\sigma\sigma}\\ g^{\ast\ast}_{x}g^{\ast}_{\sigma\sigma} + g^{\ast\ast}_{\sigma\sigma}\\0 \end{pmatrix}$
$P$ are permutation matrices that can be computed using Matlab's `ipermute` function.
## Parameter derivatives of perturbation solutions
First, we need the parameter derivatives of first, second, third, and fourth derivatives of the dynamic model (i.e. g1,g2,g3,g4 in dynamic files), I make use of the implicit function theorem: Let $f_{z^k}$ denote the kth derivative (wrt all dynamic variables) of the dynamic model, then let $df_{z^k}$ denote the first-derivative (wrt all model parameters) of $f_{z^k}$ evaluated at the steady state. Note that $f_{z^k}$ is a function of both the model parameters $\theta$ and of the steady state of all dynamic variables $\bar{z}$, which also depend on the parameters. Hence, implicitly $f_{z^k}=f_{z^k}(\theta,\bar{z}(\theta))$ and $df_{z^k}$ consists of two parts:
1. direct derivative wrt to all model parameters given by the preprocessor in the `_params_derivs.m` files
2. contribution of derivative of steady state of dynamic variables (wrt all model parameters): $f_{z^{k+1}} \cdot d\bar{z}$
Note that we already have functionality to compute $d\bar{z}$ analytically.
Having this, the above perturbation systems are basically equations of the following types
$AX +BXC = RHS$ or $AX = RHS$
Now when taking the derivative (wrt to one single parameter $\theta_j$), we get
$A\mathrm{d}\{X\} + B\mathrm{d}\{X\}C = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X - \mathrm{d}\{B\}XC - BX\mathrm{d}\{C\}$
or
$A\mathrm{d}\{X\} = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X$
The first one is a Sylvester type equation, the second one can be solved by taking the inverse of $A$. The only diffculty and tedious work arrises in computing (the highly sparse) derivatives of $RHS$.
***
# New functions: `
## get_perturbation_params_derivs.m`and `get_perturbation_params_derivs_numerical_objective.m`
* The parameter derivatives up to third order are computed in the new function`get_perturbation_params_derivs.m` both analytically and numerically. For numerical derivatives `get_perturbation_params_derivs_numerical_objective.m` is the objective for `fjaco.m` or `hessian_sparse.m` or `hessian.m`.
* `get_perturbation_params_derivs.m` is basically an extended version of the previous `get_first_order_solution_params_derivs.m` function.
* * `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It is used for numerical derivatives, whenever `analytic_derivation_mode=-1|-2`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
* Detailed changes:
* Most important: notation of this function is now in accordance to the k_order_solver, i.e. we do not compute derivatives of Kalman transition matrices A and B, but rather the solution matrices ghx,ghu,ghxx,ghxu,ghuu,ghs2,ghxxx,ghxxu,ghxuu,ghuuu,ghxss,ghuss in accordance with notation used in `oo_.dr`. As a byproduct at first-order, focusing on ghx and ghu instead of Kalman transition matrices A and B makes the computations slightly faster for large models (e.g. for Quest the computations were faster by a couple of seconds, not much, but okay).
* Removed use of `kstate`, see also Dynare/dynare#1653 and Dynare/dynare!1656
* Output arguments are stored in a structure `DERIVS`, there is also a flag `d2flag` that computes parameter hessians needed only in `dsge_likelihood.m`.
* Removed `kronflag` as input. `options_.analytic_derivation_mode` is now used instead of `kronflag`.
* Removed `indvar`, an index that was used to selected specific variables in the derivatives. This is not needed, as we always compute the parameter derivatives for all variables first and then select a subset of variables. The selection now takes place in other functions, like `dsge_likelihood.m`.
* Introduced some checks: (i) deterministic exogenous variables are not supported, (ii) Kronecker method only compatible with first-order approximation so reset to sylvester method, (iii) for purely backward or forward models we need to be careful with the rows in `M_.lead_la g_incidence`, (iv) if `_params_derivs.m` files are missing an error is thrown.
* For numerical derivatives, if mod file does not contain an `estimated_params_block`, a temporary one with the most important parameter information is created.
## `unfold_g4.m`
* When evaluating g3 and g4 one needs to take into account that these do not contain symmetric elements, so one needs to use `unfold_g3.m` and the new function `unfold_g4.m`. This returns an unfolded version of the same matrix (i.e. with symmetric elements).
***
# New test models
`.gitignore` and `Makefile.am` are changed accordingly. Also now it is possible to run test suite on analytic_derivatives, i.e. run `make check m/analytic_derivatives`
## `analytic_derivatives/BrockMirman_PertParamsDerivs.mod`
* This is the Brock Mirman model, where we know the exact policy function $g$ for capital and consumption. As this does not imply a nonzero $g_{\sigma\sigma}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ I added some artificial equations to get nonzero solution matrices with respect to $\sigma$. The true perturbation solution matrices $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ are then computed analytically with Matlab's symbolic toolbox and saved in `nBrockMirmanSYM.mat`. There is a preprocessor flag that recreates these analytical computations if changes are needed (and to check whether I made some errors here ;-) )
* Then solution matrices up to third order and their parameter Jacobians are then compared to the ones computed by Dynare's `k_order_solver` and by `get_perturbation_params_derivs` for all `analytic_derivation_mode`'s. There will be an error if the maximum absolute deviation is too large, i.e. for numerical derivatives (`analytic_derivation_mode=-1|-2`) the tolerance is choosen lower (around 1e-5); for analytical methods we are stricter: around 1e-13 for first-order, 1e-12 for second order, and 1e-11 for third-order.
* As a side note, this mod file also checks Dynare's `k_order_solver` algorithm and throws an error if something is wrong.
* This test model shows that the new functionality works well. And analytical derivatives perform way better and accurate than numerical ones, even for this small model.
## `analytic_derivatives/burnside_3_order_PertParamsDerivs.mod`
* This builds upon `tests/k_order_perturbation/burnside_k_order.mod` and computes the true parameter derivatives analytically by hand.
* This test model also shows that the new functionality works well.
## `analytic_derivatives/LindeTrabandt2019.mod`
* Shows that the new functionality also works for medium-sized models, i.e. a SW type model solved at third order with 35 variables (11 states). 2 shocks and 20 parameters.
* This mod file can be used to tweak the speed of the computations in the future.
* Compares numerical versus analytical parameter derivatives (for first, second and third order). Note that this model clearly shows that numerical ones are quite different than analytical ones even at first order!
## `identification/LindeTrabandt2019_xfail.mod`
* This model is a check for issue Dynare/dynare#1595, see fjaco.m below, and will fail.
* Removed `analytic_derivatives/ls2003.mod` as this mod file is neither in the testsuite nor does it work.
***
# Detailed changes in other functions
## `get_first_order_solution_params_derivs.m`
* Deleted, or actually, renamed to `get_perturbation_params_derivs.m`, as this function now is able to compute the derivatives up to third order
## `identification_numerical_objective.m`
* `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
## `dsge_likelihood.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `T`matrix anymore, but instead of the dynare solution matrix `ghx`. So we recreate `T` here as this amounts to adding some zeros and focusing on selected variables only.
* Added some checks to make sure the first-order approximation is selected.
* Removed `kron_flag` as input, as `get_perturbation_params_derivs` looks into `options_.analytic_derivation_mode` for `kron_flag`.
## `dynare_identification.m`
* make sure that setting `analytic_derivation_mode` is set both in `options_ident` and `options_`. Note that at the end of the function we restore the `options_` structure, so all changes are local. In a next merge request, I will remove the global variables to make all variables local.
## `get_identification_jacobians.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `A` and `B` matrix anymore, but instead of the dynare solution matrix `ghx` and `ghu`. So we recreate these matrices here instead of in `get_perturbation_params_derivs.m`.
* Added `str2func` for better function handles in `fjaco.m`.
## `fjaco.m`
* make `tol`an option, which can be adjusted by changing `options_.dynatol.x`for identification and parameter derivatives purposes.
* include a check and an informative error message, if numerical derivatives (two-sided finite difference method) yield errors in `resol.m` for identification and parameter derivatives purposes. This closes issue Dynare/dynare#1595.
* Changed year of copyright to 2010-2017,2019
***
# Further suggestions and questions
* Ones this is merged, I will merge request an improvement of the identification toolbox, which will work up to third order using the pruned state space. This will also remove some issues and bugs, and also I will remove global variables in this request.
* The third-order derivatives can be further improved by taking sparsity into account and use mex versions for kronecker products etc. I leave this for further testing (and if anybody actually uses this ;-) )
2019-12-17 19:17:09 +01:00
end
end
if ~ isempty ( indpmodel ) %derivatives of Sigma_e wrt model parameters are zero by construction
for jp = 1 : modparam_nbr
DERIVS . dOm ( : , : , stderrparam_nbr + corrparam_nbr + jp ) = DERIVS . dghu ( : , : , stderrparam_nbr + corrparam_nbr + jp ) * M . Sigma_e * oo . dr . ghu ' + oo . dr . ghu * M . Sigma_e * DERIVS . dghu ( : , : , stderrparam_nbr + corrparam_nbr + jp ) ' ;
end
end
%Parameter Jacobian of dynamic model derivatives (wrt selected model parameters only)
dYss_g = fjaco ( numerical_objective_fname , modparam1 , ' dynamic_model' , estim_params_model , M , oo , options ) ;
ind_Yss = 1 : M . endo_nbr ;
ind_g1 = ind_Yss ( end ) + ( 1 : M . endo_nbr * yy0ex0_nbr ) ;
DERIVS . dYss = dYss_g ( ind_Yss , : ) ; %in tensor notation, wrt selected model parameters only
DERIVS . dg1 = reshape ( dYss_g ( ind_g1 , : ) , [ M . endo_nbr , yy0ex0_nbr , modparam_nbr ] ) ; %in tensor notation
if order > 1
ind_g2 = ind_g1 ( end ) + ( 1 : M . endo_nbr * yy0ex0_nbr ^2 ) ;
DERIVS . dg2 = reshape ( sparse ( dYss_g ( ind_g2 , : ) ) , [ M . endo_nbr , yy0ex0_nbr ^2 * modparam_nbr ] ) ; %blockwise in matrix notation, i.e. [dg2_dp1 dg2_dp2 ...], where dg2_dpj has dimension M.endo_nbr by yy0ex0_nbr^2
end
if order > 2
ind_g3 = ind_g2 ( end ) + ( 1 : M . endo_nbr * yy0ex0_nbr ^3 ) ;
DERIVS . dg3 = reshape ( sparse ( dYss_g ( ind_g3 , : ) ) , [ M . endo_nbr , yy0ex0_nbr ^3 * modparam_nbr ] ) ; %blockwise in matrix notation, i.e. [dg3_dp1 dg3_dp2 ...], where dg3_dpj has dimension M.endo_nbr by yy0ex0_nbr^3
end
if d2flag
% Hessian (wrt paramters) of steady state and first-order solution matrices ghx and Om
% note that hessian_sparse.m (contrary to hessian.m) does not take symmetry into account, but focuses already on unique values
options . order = 1 ; %make sure only first order
d2Yss_KalmanA_Om = hessian_sparse ( numerical_objective_fname , xparam1 , options . gstep , ' Kalman_Transition' , estim_params , M , oo , options ) ;
options . order = order ; %make sure to set back
ind_KalmanA = ind_Yss ( end ) + ( 1 : M . endo_nbr ^2 ) ;
DERIVS . d2KalmanA = d2Yss_KalmanA_Om ( ind_KalmanA , indp2tottot2 ) ; %only unique elements
DERIVS . d2Om = d2Yss_KalmanA_Om ( ind_KalmanA ( end ) + 1 : end , indp2tottot2 ) ; %only unique elements
DERIVS . d2Yss = zeros ( M . endo_nbr , modparam_nbr , modparam_nbr ) ; %initialize
for j = 1 : M . endo_nbr
DERIVS . d2Yss ( j , : , : ) = dyn_unvech ( full ( d2Yss_KalmanA_Om ( j , indp2modmod2 ) ) ) ; %Hessian for d2Yss, but without cross derivatives
end
end
return %[END OF MAIN FUNCTION]!!!!!
end
if analytic_derivation_mode == - 2
%% Numerical two-sided finite difference method to compute parameter derivatives of steady state and dynamic model,
% i.e. dYss, dg1, dg2, dg3 as well as d2Yss, d2g1 numerically.
% The parameter derivatives of perturbation solution matrices are computed analytically below (analytic_derivation_mode=0)
if order == 3
[ ~ , g1 , g2 , g3 ] = feval ( [ M . fname , ' .dynamic' ] , oo . dr . ys ( I ) , oo . exo_steady_state ' , M . params , oo . dr . ys , 1 ) ;
g3 = unfold_g3 ( g3 , yy0ex0_nbr ) ;
elseif order == 2
[ ~ , g1 , g2 ] = feval ( [ M . fname , ' .dynamic' ] , oo . dr . ys ( I ) , oo . exo_steady_state ' , M . params , oo . dr . ys , 1 ) ;
elseif order == 1
[ ~ , g1 ] = feval ( [ M . fname , ' .dynamic' ] , oo . dr . ys ( I ) , oo . exo_steady_state ' , M . params , oo . dr . ys , 1 ) ;
end
if d2flag
% computation of d2Yss and d2g1
2019-12-20 16:28:06 +01:00
% note that hessian_sparse does not take symmetry into account, i.e. compare hessian_sparse.m to hessian.m, but focuses already on unique values, which are duplicated below
Added parameter derivatives of perturbation solution up to 3 order
# Preliminary comments
I finished the identification toolbox at orders two and three using the pruned state space system, but before I merge request this, I decided to first merge the new functionality to compute parameter derivatives of perturbation solution matrices at higher orders. So after this is approved, I merge the identification toolbox.
I guess @rattoma, @sebastien, and @michel are best choices to review this.
I outline the main idea first and then provide some more detailed changes I made to the functions.
***
# Main idea
This merge request is concerned with the *analytical*computation of the parameter derivatives of first, second and third order perturbation solution matrices, i.e. using _closed-form_ expressions to efficiently compute the derivative of $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ *with respect to model parameters* $\theta$. Note that $\theta$ contains model parameters, stderr and corr parameters of shocks. stderr and corr parameters of measurement errors are not yet supported, (they can easily be included as exogenous shocks). The availability of such derivatives is beneficial in terms of more reliable analysis of model sensitivity and parameter identifiability as well as more efficient estimation methods, in particular for models solved up to third order, as it is well-known that numerical derivatives are a tricky business, especially for large models.
References for my approach are:
* Iskrev (2008, 2010) and Schmitt-Grohé and Uribe (2012, Appendix) who were the first to compute the parameter derivatives analytically at first order, however, using inefficient (sparse) Kronecker products.
* Mutschler (2015) who provides the expressions for a second-order, but again using inefficient (sparse) Kronecker products.
* Ratto and Iskrev (2012) who show how the first-order system can be solved accurately, fast and efficiently using existing numerical algorithms for generalized Sylvester equations by taking the parameter derivative with respect to each parameter separately.
* Julliard and Kamenik (2004) who provide the perturbation solution equation system in tensor notation at any order k.
* Levintal (2017) who introduces permutation matrices to express the perturbation solution equation system in matrix notation up to fifth order.
Note that @rattoma already implemented the parameter derivatives of $g_x$ and $g_u$ analytically (and numerically), and I rely heavily on his work in `get_first_order_solution_params_derivs.m` (previously `getH.m`). My additions are mainly to this function and thus it is renamed to `get_perturbation_params_derivs.m`.
The basic idea of this merge request is to take the second- and third-order perturbation solution systems in Julliard and Kamenik (2004), unfold these into an equivalent matrix representation using permutation matrices as in Levintal (2017). Then extending Ratto and Iskrev (2012) one takes the derivative with respect to each parameter separately and gets a computational problem that is linear, albeit large, as it involves either solving generalized Sylvester equations or taking inverses of highly sparse matrices. I will now briefly summarize the perturbation solution system at third order and the system that results when taking the derivative with respect to parameters.
## Perturbation Solution
The following systems arise at first, second, and third order:
$(ghx): f_{x} z_{x} = f_{y_{-}^*} + f_{y_0} g_{x} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{x}= A g_{x} + f_{y_{-}^*}=0$
$(ghu): f_{z} z_{u} = f_{y_0} g_{u} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{u} + f_{u}= A g_u + f_u = 0$
$(ghxx) : A g_{xx} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{x}\right) + f_{zz} \left( z_{x} \otimes z_{x} \right) = 0$
$(ghxu) : A g_{xu} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{x} \otimes z_{u} \right) = 0$
$(ghuu) : A g_{uu} + B g_{xx} \left(g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{u} \otimes z_{u} \right) = 0$
$(ghs2) : (A+B) g_{\sigma\sigma} + \left( f_{y^{**}_{+}y^{**}_{+}} \left(g^{**}_{u} \otimes g^{**}_{u}\right) + f_{y^{**}_{+}} g^{**}_{uu}\right)vec(\Sigma) = 0$
$(ghxxx) : A g_{xxx} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{x}\right) + f_{y_{+}}g^{**}_{xx} \left(g^{*}_x \otimes g^{*}_{xx}\right)P_{x\_xx} + f_{zz} \left( z_{x} \otimes z_{xx} \right)P_{x\_xx} + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{x} \right) = 0$
$(ghxxu) : A g_{xxu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{u} \right) + f_{zz} \left( \left( z_{x} \otimes z_{xu} \right)P_{x\_xu} + \left(z_{xx} \otimes z_{u}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{x} \otimes g^{*}_{xu}\right)P_{x\_xu} + \left(g^{*}_{xx} \otimes g^{*}_{u}\right) \right) = 0$
$(ghxuu) : A g_{xuu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( \left( z_{xu} \otimes z_{u} \right)P_{xu\_u} + \left(z_{x} \otimes z_{uu}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{xu} \otimes g^{*}_{u}\right)P_{xu\_u} + \left(g^{*}_{x} \otimes g^{*}_{uu}\right) \right) = 0$
$(ghuuu) : A g_{uuu} + B g_{xxx} \left(g^{*}_{u} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{u} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( z_{u} \otimes z_{uu} \right)P_{u\_uu} + f_{y_{+}}g^{**}_{xx} \left(g^{*}_{u} \otimes g^{*}_{uu}\right)P_{u\_uu} = 0$
$(ghx\sigma\sigma) : A g_{x\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_x + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{x} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{x} \otimes z_{\sigma\sigma}\right) + F_{xu_{+}u_{+}}\left(I_{n_x} \otimes vec(\Sigma)\right) = 0$
$F_{xu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_x^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{xu_{+}} \otimes z_{u_{+}} \right)P_{xu\_u} + \left(z_{x} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{x} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
$(ghu\sigma\sigma) : A g_{u\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_{u} + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{u} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{u} \otimes z_{\sigma\sigma}\right) + F_{uu_{+}u_{+}}\left(I_{n_u} \otimes vec(\Sigma_u)\right) = 0$
$F_{uu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_u^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{uu_{+}} \otimes z_{u_{+}} \right)P_{uu\_u} + \left(z_{u} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{u} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
A and B are the common perturbation matrices:
$A = f_{y_0} + \begin{pmatrix} \underbrace{0}_{n\times n_{static}} &\vdots& \underbrace{f_{y^{**}_{+}} \cdot g^{**}_{x}}_{n \times n_{spred}} &\vdots& \underbrace{0}_{n\times n_{frwd}} \end{pmatrix}$and $B = \begin{pmatrix} \underbrace{0}_{n \times n_{static}}&\vdots & \underbrace{0}_{n \times n_{pred}} & \vdots & \underbrace{f_{y^{**}_{+}}}_{n \times n_{sfwrd}} \end{pmatrix}$
and $z=(y_{-}^{\ast}; y; y_{+}^{\ast\ast}; u)$ denotes the dynamic model variables as in `M_.lead_lag_incidence`, $y^\ast$ denote state variables, $y^{\ast\ast}$ denote forward looking variables, $y_+$ denote the variables with a lead, $y_{-}$ denote variables with a lag, $y_0$ denote variables at period t, $f$ the model equations, and $f_z$ the first-order dynamic model derivatives, $f_{zz}$ the second-order dynamic derivatives, and $f_{zzz}$ the third-order dynamic model derivatives. Then:
$z_{x} = \begin{pmatrix}I\\g_{x}\\g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{u} =\begin{pmatrix}0\\g_{u}\\g^{**}_{x} \cdot g^{*}_{u}\\I\end{pmatrix}$, $z_{u_{+}} =\begin{pmatrix}0\\0\\g^{**}_{u}\\0\end{pmatrix}$
$z_{xx} = \begin{pmatrix} 0\\g_{xx}\\g^{**}_{x} \left( g^{*}_x \otimes g^{*}_{x} \right) + g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{xu} =\begin{pmatrix}0\\g_{xu}\\g^{**}_{xx} \left( g^{*}_x \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{xu}\\0\end{pmatrix}$, $z_{uu} =\begin{pmatrix}0\\g_{uu}\\g^{**}_{xx} \left( g^{*}_u \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{uu}\\0\end{pmatrix}$,
$z_{xu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_x \otimes I \right)\\0\end{pmatrix}$, $z_{uu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_{u} \otimes I \right)\\0\end{pmatrix}$, $z_{u_{+}u_{+}} =\begin{pmatrix}0\\0\\g^{\ast\ast}_{uu}\\0\end{pmatrix}$, $z_{\sigma\sigma} = \begin{pmatrix}0\\ g_{\sigma\sigma}\\ g^{\ast\ast}_{x}g^{\ast}_{\sigma\sigma} + g^{\ast\ast}_{\sigma\sigma}\\0 \end{pmatrix}$
$P$ are permutation matrices that can be computed using Matlab's `ipermute` function.
## Parameter derivatives of perturbation solutions
First, we need the parameter derivatives of first, second, third, and fourth derivatives of the dynamic model (i.e. g1,g2,g3,g4 in dynamic files), I make use of the implicit function theorem: Let $f_{z^k}$ denote the kth derivative (wrt all dynamic variables) of the dynamic model, then let $df_{z^k}$ denote the first-derivative (wrt all model parameters) of $f_{z^k}$ evaluated at the steady state. Note that $f_{z^k}$ is a function of both the model parameters $\theta$ and of the steady state of all dynamic variables $\bar{z}$, which also depend on the parameters. Hence, implicitly $f_{z^k}=f_{z^k}(\theta,\bar{z}(\theta))$ and $df_{z^k}$ consists of two parts:
1. direct derivative wrt to all model parameters given by the preprocessor in the `_params_derivs.m` files
2. contribution of derivative of steady state of dynamic variables (wrt all model parameters): $f_{z^{k+1}} \cdot d\bar{z}$
Note that we already have functionality to compute $d\bar{z}$ analytically.
Having this, the above perturbation systems are basically equations of the following types
$AX +BXC = RHS$ or $AX = RHS$
Now when taking the derivative (wrt to one single parameter $\theta_j$), we get
$A\mathrm{d}\{X\} + B\mathrm{d}\{X\}C = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X - \mathrm{d}\{B\}XC - BX\mathrm{d}\{C\}$
or
$A\mathrm{d}\{X\} = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X$
The first one is a Sylvester type equation, the second one can be solved by taking the inverse of $A$. The only diffculty and tedious work arrises in computing (the highly sparse) derivatives of $RHS$.
***
# New functions: `
## get_perturbation_params_derivs.m`and `get_perturbation_params_derivs_numerical_objective.m`
* The parameter derivatives up to third order are computed in the new function`get_perturbation_params_derivs.m` both analytically and numerically. For numerical derivatives `get_perturbation_params_derivs_numerical_objective.m` is the objective for `fjaco.m` or `hessian_sparse.m` or `hessian.m`.
* `get_perturbation_params_derivs.m` is basically an extended version of the previous `get_first_order_solution_params_derivs.m` function.
* * `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It is used for numerical derivatives, whenever `analytic_derivation_mode=-1|-2`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
* Detailed changes:
* Most important: notation of this function is now in accordance to the k_order_solver, i.e. we do not compute derivatives of Kalman transition matrices A and B, but rather the solution matrices ghx,ghu,ghxx,ghxu,ghuu,ghs2,ghxxx,ghxxu,ghxuu,ghuuu,ghxss,ghuss in accordance with notation used in `oo_.dr`. As a byproduct at first-order, focusing on ghx and ghu instead of Kalman transition matrices A and B makes the computations slightly faster for large models (e.g. for Quest the computations were faster by a couple of seconds, not much, but okay).
* Removed use of `kstate`, see also Dynare/dynare#1653 and Dynare/dynare!1656
* Output arguments are stored in a structure `DERIVS`, there is also a flag `d2flag` that computes parameter hessians needed only in `dsge_likelihood.m`.
* Removed `kronflag` as input. `options_.analytic_derivation_mode` is now used instead of `kronflag`.
* Removed `indvar`, an index that was used to selected specific variables in the derivatives. This is not needed, as we always compute the parameter derivatives for all variables first and then select a subset of variables. The selection now takes place in other functions, like `dsge_likelihood.m`.
* Introduced some checks: (i) deterministic exogenous variables are not supported, (ii) Kronecker method only compatible with first-order approximation so reset to sylvester method, (iii) for purely backward or forward models we need to be careful with the rows in `M_.lead_la g_incidence`, (iv) if `_params_derivs.m` files are missing an error is thrown.
* For numerical derivatives, if mod file does not contain an `estimated_params_block`, a temporary one with the most important parameter information is created.
## `unfold_g4.m`
* When evaluating g3 and g4 one needs to take into account that these do not contain symmetric elements, so one needs to use `unfold_g3.m` and the new function `unfold_g4.m`. This returns an unfolded version of the same matrix (i.e. with symmetric elements).
***
# New test models
`.gitignore` and `Makefile.am` are changed accordingly. Also now it is possible to run test suite on analytic_derivatives, i.e. run `make check m/analytic_derivatives`
## `analytic_derivatives/BrockMirman_PertParamsDerivs.mod`
* This is the Brock Mirman model, where we know the exact policy function $g$ for capital and consumption. As this does not imply a nonzero $g_{\sigma\sigma}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ I added some artificial equations to get nonzero solution matrices with respect to $\sigma$. The true perturbation solution matrices $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ are then computed analytically with Matlab's symbolic toolbox and saved in `nBrockMirmanSYM.mat`. There is a preprocessor flag that recreates these analytical computations if changes are needed (and to check whether I made some errors here ;-) )
* Then solution matrices up to third order and their parameter Jacobians are then compared to the ones computed by Dynare's `k_order_solver` and by `get_perturbation_params_derivs` for all `analytic_derivation_mode`'s. There will be an error if the maximum absolute deviation is too large, i.e. for numerical derivatives (`analytic_derivation_mode=-1|-2`) the tolerance is choosen lower (around 1e-5); for analytical methods we are stricter: around 1e-13 for first-order, 1e-12 for second order, and 1e-11 for third-order.
* As a side note, this mod file also checks Dynare's `k_order_solver` algorithm and throws an error if something is wrong.
* This test model shows that the new functionality works well. And analytical derivatives perform way better and accurate than numerical ones, even for this small model.
## `analytic_derivatives/burnside_3_order_PertParamsDerivs.mod`
* This builds upon `tests/k_order_perturbation/burnside_k_order.mod` and computes the true parameter derivatives analytically by hand.
* This test model also shows that the new functionality works well.
## `analytic_derivatives/LindeTrabandt2019.mod`
* Shows that the new functionality also works for medium-sized models, i.e. a SW type model solved at third order with 35 variables (11 states). 2 shocks and 20 parameters.
* This mod file can be used to tweak the speed of the computations in the future.
* Compares numerical versus analytical parameter derivatives (for first, second and third order). Note that this model clearly shows that numerical ones are quite different than analytical ones even at first order!
## `identification/LindeTrabandt2019_xfail.mod`
* This model is a check for issue Dynare/dynare#1595, see fjaco.m below, and will fail.
* Removed `analytic_derivatives/ls2003.mod` as this mod file is neither in the testsuite nor does it work.
***
# Detailed changes in other functions
## `get_first_order_solution_params_derivs.m`
* Deleted, or actually, renamed to `get_perturbation_params_derivs.m`, as this function now is able to compute the derivatives up to third order
## `identification_numerical_objective.m`
* `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
## `dsge_likelihood.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `T`matrix anymore, but instead of the dynare solution matrix `ghx`. So we recreate `T` here as this amounts to adding some zeros and focusing on selected variables only.
* Added some checks to make sure the first-order approximation is selected.
* Removed `kron_flag` as input, as `get_perturbation_params_derivs` looks into `options_.analytic_derivation_mode` for `kron_flag`.
## `dynare_identification.m`
* make sure that setting `analytic_derivation_mode` is set both in `options_ident` and `options_`. Note that at the end of the function we restore the `options_` structure, so all changes are local. In a next merge request, I will remove the global variables to make all variables local.
## `get_identification_jacobians.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `A` and `B` matrix anymore, but instead of the dynare solution matrix `ghx` and `ghu`. So we recreate these matrices here instead of in `get_perturbation_params_derivs.m`.
* Added `str2func` for better function handles in `fjaco.m`.
## `fjaco.m`
* make `tol`an option, which can be adjusted by changing `options_.dynatol.x`for identification and parameter derivatives purposes.
* include a check and an informative error message, if numerical derivatives (two-sided finite difference method) yield errors in `resol.m` for identification and parameter derivatives purposes. This closes issue Dynare/dynare#1595.
* Changed year of copyright to 2010-2017,2019
***
# Further suggestions and questions
* Ones this is merged, I will merge request an improvement of the identification toolbox, which will work up to third order using the pruned state space. This will also remove some issues and bugs, and also I will remove global variables in this request.
* The third-order derivatives can be further improved by taking sparsity into account and use mex versions for kronecker products etc. I leave this for further testing (and if anybody actually uses this ;-) )
2019-12-17 19:17:09 +01:00
options . order = 1 ; %d2flag requires only first order
d2Yss_g1 = hessian_sparse ( numerical_objective_fname , modparam1 , options . gstep , ' dynamic_model' , estim_params_model , M , oo , options ) ; % d2flag requires only first-order
options . order = order ; %make sure to set back the order
d2Yss = reshape ( full ( d2Yss_g1 ( 1 : M . endo_nbr , : ) ) , [ M . endo_nbr modparam_nbr modparam_nbr ] ) ; %put into tensor notation
for j = 1 : M . endo_nbr
d2Yss ( j , : , : ) = dyn_unvech ( dyn_vech ( d2Yss ( j , : , : ) ) ) ; %add duplicate values to full hessian
end
d2g1_full = d2Yss_g1 ( M . endo_nbr + 1 : end , : ) ;
%store only nonzero unique entries and the corresponding indices of d2g1:
% rows: respective derivative term
% 1st column: equation number of the term appearing
% 2nd column: column number of variable in Jacobian of the dynamic model
% 3rd column: number of the first parameter in derivative
% 4th column: number of the second parameter in derivative
% 5th column: value of the Hessian term
ind_d2g1 = find ( d2g1_full ) ;
d2g1 = zeros ( length ( ind_d2g1 ) , 5 ) ;
for j = 1 : length ( ind_d2g1 )
[ i1 , i2 ] = ind2sub ( size ( d2g1_full ) , ind_d2g1 ( j ) ) ;
[ ig1 , ig2 ] = ind2sub ( size ( g1 ) , i1 ) ;
[ ip1 , ip2 ] = ind2sub ( [ modparam_nbr modparam_nbr ] , i2 ) ;
d2g1 ( j , : ) = [ ig1 ig2 ip1 ip2 d2g1_full ( ind_d2g1 ( j ) ) ] ;
end
clear d2g1_full d2Yss_g1 ;
end
%Parameter Jacobian of dynamic model derivatives (wrt selected model parameters only)
dYss_g = fjaco ( numerical_objective_fname , modparam1 , ' dynamic_model' , estim_params_model , M , oo , options ) ;
ind_Yss = 1 : M . endo_nbr ;
ind_g1 = ind_Yss ( end ) + ( 1 : M . endo_nbr * yy0ex0_nbr ) ;
dYss = dYss_g ( ind_Yss , : ) ; %in tensor notation, wrt selected model parameters only
dg1 = reshape ( dYss_g ( ind_g1 , : ) , [ M . endo_nbr , yy0ex0_nbr , modparam_nbr ] ) ; %in tensor notation
if order > 1
ind_g2 = ind_g1 ( end ) + ( 1 : M . endo_nbr * yy0ex0_nbr ^2 ) ;
dg2 = reshape ( sparse ( dYss_g ( ind_g2 , : ) ) , [ M . endo_nbr , yy0ex0_nbr ^2 * modparam_nbr ] ) ; %blockwise in matrix notation, i.e. [dg2_dp1 dg2_dp2 ...], where dg2_dpj has dimension M.endo_nbr by yy0ex0_nbr^2
end
if order > 2
ind_g3 = ind_g2 ( end ) + ( 1 : M . endo_nbr * yy0ex0_nbr ^3 ) ;
dg3 = reshape ( sparse ( dYss_g ( ind_g3 , : ) ) , [ M . endo_nbr , yy0ex0_nbr ^3 * modparam_nbr ] ) ; %blockwise in matrix notation, i.e. [dg3_dp1 dg3_dp2 ...], where dg3_dpj has dimension M.endo_nbr by yy0ex0_nbr^3
end
clear dYss_g
2019-12-20 16:28:06 +01:00
Added parameter derivatives of perturbation solution up to 3 order
# Preliminary comments
I finished the identification toolbox at orders two and three using the pruned state space system, but before I merge request this, I decided to first merge the new functionality to compute parameter derivatives of perturbation solution matrices at higher orders. So after this is approved, I merge the identification toolbox.
I guess @rattoma, @sebastien, and @michel are best choices to review this.
I outline the main idea first and then provide some more detailed changes I made to the functions.
***
# Main idea
This merge request is concerned with the *analytical*computation of the parameter derivatives of first, second and third order perturbation solution matrices, i.e. using _closed-form_ expressions to efficiently compute the derivative of $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ *with respect to model parameters* $\theta$. Note that $\theta$ contains model parameters, stderr and corr parameters of shocks. stderr and corr parameters of measurement errors are not yet supported, (they can easily be included as exogenous shocks). The availability of such derivatives is beneficial in terms of more reliable analysis of model sensitivity and parameter identifiability as well as more efficient estimation methods, in particular for models solved up to third order, as it is well-known that numerical derivatives are a tricky business, especially for large models.
References for my approach are:
* Iskrev (2008, 2010) and Schmitt-Grohé and Uribe (2012, Appendix) who were the first to compute the parameter derivatives analytically at first order, however, using inefficient (sparse) Kronecker products.
* Mutschler (2015) who provides the expressions for a second-order, but again using inefficient (sparse) Kronecker products.
* Ratto and Iskrev (2012) who show how the first-order system can be solved accurately, fast and efficiently using existing numerical algorithms for generalized Sylvester equations by taking the parameter derivative with respect to each parameter separately.
* Julliard and Kamenik (2004) who provide the perturbation solution equation system in tensor notation at any order k.
* Levintal (2017) who introduces permutation matrices to express the perturbation solution equation system in matrix notation up to fifth order.
Note that @rattoma already implemented the parameter derivatives of $g_x$ and $g_u$ analytically (and numerically), and I rely heavily on his work in `get_first_order_solution_params_derivs.m` (previously `getH.m`). My additions are mainly to this function and thus it is renamed to `get_perturbation_params_derivs.m`.
The basic idea of this merge request is to take the second- and third-order perturbation solution systems in Julliard and Kamenik (2004), unfold these into an equivalent matrix representation using permutation matrices as in Levintal (2017). Then extending Ratto and Iskrev (2012) one takes the derivative with respect to each parameter separately and gets a computational problem that is linear, albeit large, as it involves either solving generalized Sylvester equations or taking inverses of highly sparse matrices. I will now briefly summarize the perturbation solution system at third order and the system that results when taking the derivative with respect to parameters.
## Perturbation Solution
The following systems arise at first, second, and third order:
$(ghx): f_{x} z_{x} = f_{y_{-}^*} + f_{y_0} g_{x} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{x}= A g_{x} + f_{y_{-}^*}=0$
$(ghu): f_{z} z_{u} = f_{y_0} g_{u} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{u} + f_{u}= A g_u + f_u = 0$
$(ghxx) : A g_{xx} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{x}\right) + f_{zz} \left( z_{x} \otimes z_{x} \right) = 0$
$(ghxu) : A g_{xu} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{x} \otimes z_{u} \right) = 0$
$(ghuu) : A g_{uu} + B g_{xx} \left(g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{u} \otimes z_{u} \right) = 0$
$(ghs2) : (A+B) g_{\sigma\sigma} + \left( f_{y^{**}_{+}y^{**}_{+}} \left(g^{**}_{u} \otimes g^{**}_{u}\right) + f_{y^{**}_{+}} g^{**}_{uu}\right)vec(\Sigma) = 0$
$(ghxxx) : A g_{xxx} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{x}\right) + f_{y_{+}}g^{**}_{xx} \left(g^{*}_x \otimes g^{*}_{xx}\right)P_{x\_xx} + f_{zz} \left( z_{x} \otimes z_{xx} \right)P_{x\_xx} + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{x} \right) = 0$
$(ghxxu) : A g_{xxu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{u} \right) + f_{zz} \left( \left( z_{x} \otimes z_{xu} \right)P_{x\_xu} + \left(z_{xx} \otimes z_{u}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{x} \otimes g^{*}_{xu}\right)P_{x\_xu} + \left(g^{*}_{xx} \otimes g^{*}_{u}\right) \right) = 0$
$(ghxuu) : A g_{xuu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( \left( z_{xu} \otimes z_{u} \right)P_{xu\_u} + \left(z_{x} \otimes z_{uu}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{xu} \otimes g^{*}_{u}\right)P_{xu\_u} + \left(g^{*}_{x} \otimes g^{*}_{uu}\right) \right) = 0$
$(ghuuu) : A g_{uuu} + B g_{xxx} \left(g^{*}_{u} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{u} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( z_{u} \otimes z_{uu} \right)P_{u\_uu} + f_{y_{+}}g^{**}_{xx} \left(g^{*}_{u} \otimes g^{*}_{uu}\right)P_{u\_uu} = 0$
$(ghx\sigma\sigma) : A g_{x\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_x + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{x} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{x} \otimes z_{\sigma\sigma}\right) + F_{xu_{+}u_{+}}\left(I_{n_x} \otimes vec(\Sigma)\right) = 0$
$F_{xu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_x^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{xu_{+}} \otimes z_{u_{+}} \right)P_{xu\_u} + \left(z_{x} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{x} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
$(ghu\sigma\sigma) : A g_{u\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_{u} + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{u} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{u} \otimes z_{\sigma\sigma}\right) + F_{uu_{+}u_{+}}\left(I_{n_u} \otimes vec(\Sigma_u)\right) = 0$
$F_{uu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_u^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{uu_{+}} \otimes z_{u_{+}} \right)P_{uu\_u} + \left(z_{u} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{u} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
A and B are the common perturbation matrices:
$A = f_{y_0} + \begin{pmatrix} \underbrace{0}_{n\times n_{static}} &\vdots& \underbrace{f_{y^{**}_{+}} \cdot g^{**}_{x}}_{n \times n_{spred}} &\vdots& \underbrace{0}_{n\times n_{frwd}} \end{pmatrix}$and $B = \begin{pmatrix} \underbrace{0}_{n \times n_{static}}&\vdots & \underbrace{0}_{n \times n_{pred}} & \vdots & \underbrace{f_{y^{**}_{+}}}_{n \times n_{sfwrd}} \end{pmatrix}$
and $z=(y_{-}^{\ast}; y; y_{+}^{\ast\ast}; u)$ denotes the dynamic model variables as in `M_.lead_lag_incidence`, $y^\ast$ denote state variables, $y^{\ast\ast}$ denote forward looking variables, $y_+$ denote the variables with a lead, $y_{-}$ denote variables with a lag, $y_0$ denote variables at period t, $f$ the model equations, and $f_z$ the first-order dynamic model derivatives, $f_{zz}$ the second-order dynamic derivatives, and $f_{zzz}$ the third-order dynamic model derivatives. Then:
$z_{x} = \begin{pmatrix}I\\g_{x}\\g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{u} =\begin{pmatrix}0\\g_{u}\\g^{**}_{x} \cdot g^{*}_{u}\\I\end{pmatrix}$, $z_{u_{+}} =\begin{pmatrix}0\\0\\g^{**}_{u}\\0\end{pmatrix}$
$z_{xx} = \begin{pmatrix} 0\\g_{xx}\\g^{**}_{x} \left( g^{*}_x \otimes g^{*}_{x} \right) + g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{xu} =\begin{pmatrix}0\\g_{xu}\\g^{**}_{xx} \left( g^{*}_x \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{xu}\\0\end{pmatrix}$, $z_{uu} =\begin{pmatrix}0\\g_{uu}\\g^{**}_{xx} \left( g^{*}_u \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{uu}\\0\end{pmatrix}$,
$z_{xu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_x \otimes I \right)\\0\end{pmatrix}$, $z_{uu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_{u} \otimes I \right)\\0\end{pmatrix}$, $z_{u_{+}u_{+}} =\begin{pmatrix}0\\0\\g^{\ast\ast}_{uu}\\0\end{pmatrix}$, $z_{\sigma\sigma} = \begin{pmatrix}0\\ g_{\sigma\sigma}\\ g^{\ast\ast}_{x}g^{\ast}_{\sigma\sigma} + g^{\ast\ast}_{\sigma\sigma}\\0 \end{pmatrix}$
$P$ are permutation matrices that can be computed using Matlab's `ipermute` function.
## Parameter derivatives of perturbation solutions
First, we need the parameter derivatives of first, second, third, and fourth derivatives of the dynamic model (i.e. g1,g2,g3,g4 in dynamic files), I make use of the implicit function theorem: Let $f_{z^k}$ denote the kth derivative (wrt all dynamic variables) of the dynamic model, then let $df_{z^k}$ denote the first-derivative (wrt all model parameters) of $f_{z^k}$ evaluated at the steady state. Note that $f_{z^k}$ is a function of both the model parameters $\theta$ and of the steady state of all dynamic variables $\bar{z}$, which also depend on the parameters. Hence, implicitly $f_{z^k}=f_{z^k}(\theta,\bar{z}(\theta))$ and $df_{z^k}$ consists of two parts:
1. direct derivative wrt to all model parameters given by the preprocessor in the `_params_derivs.m` files
2. contribution of derivative of steady state of dynamic variables (wrt all model parameters): $f_{z^{k+1}} \cdot d\bar{z}$
Note that we already have functionality to compute $d\bar{z}$ analytically.
Having this, the above perturbation systems are basically equations of the following types
$AX +BXC = RHS$ or $AX = RHS$
Now when taking the derivative (wrt to one single parameter $\theta_j$), we get
$A\mathrm{d}\{X\} + B\mathrm{d}\{X\}C = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X - \mathrm{d}\{B\}XC - BX\mathrm{d}\{C\}$
or
$A\mathrm{d}\{X\} = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X$
The first one is a Sylvester type equation, the second one can be solved by taking the inverse of $A$. The only diffculty and tedious work arrises in computing (the highly sparse) derivatives of $RHS$.
***
# New functions: `
## get_perturbation_params_derivs.m`and `get_perturbation_params_derivs_numerical_objective.m`
* The parameter derivatives up to third order are computed in the new function`get_perturbation_params_derivs.m` both analytically and numerically. For numerical derivatives `get_perturbation_params_derivs_numerical_objective.m` is the objective for `fjaco.m` or `hessian_sparse.m` or `hessian.m`.
* `get_perturbation_params_derivs.m` is basically an extended version of the previous `get_first_order_solution_params_derivs.m` function.
* * `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It is used for numerical derivatives, whenever `analytic_derivation_mode=-1|-2`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
* Detailed changes:
* Most important: notation of this function is now in accordance to the k_order_solver, i.e. we do not compute derivatives of Kalman transition matrices A and B, but rather the solution matrices ghx,ghu,ghxx,ghxu,ghuu,ghs2,ghxxx,ghxxu,ghxuu,ghuuu,ghxss,ghuss in accordance with notation used in `oo_.dr`. As a byproduct at first-order, focusing on ghx and ghu instead of Kalman transition matrices A and B makes the computations slightly faster for large models (e.g. for Quest the computations were faster by a couple of seconds, not much, but okay).
* Removed use of `kstate`, see also Dynare/dynare#1653 and Dynare/dynare!1656
* Output arguments are stored in a structure `DERIVS`, there is also a flag `d2flag` that computes parameter hessians needed only in `dsge_likelihood.m`.
* Removed `kronflag` as input. `options_.analytic_derivation_mode` is now used instead of `kronflag`.
* Removed `indvar`, an index that was used to selected specific variables in the derivatives. This is not needed, as we always compute the parameter derivatives for all variables first and then select a subset of variables. The selection now takes place in other functions, like `dsge_likelihood.m`.
* Introduced some checks: (i) deterministic exogenous variables are not supported, (ii) Kronecker method only compatible with first-order approximation so reset to sylvester method, (iii) for purely backward or forward models we need to be careful with the rows in `M_.lead_la g_incidence`, (iv) if `_params_derivs.m` files are missing an error is thrown.
* For numerical derivatives, if mod file does not contain an `estimated_params_block`, a temporary one with the most important parameter information is created.
## `unfold_g4.m`
* When evaluating g3 and g4 one needs to take into account that these do not contain symmetric elements, so one needs to use `unfold_g3.m` and the new function `unfold_g4.m`. This returns an unfolded version of the same matrix (i.e. with symmetric elements).
***
# New test models
`.gitignore` and `Makefile.am` are changed accordingly. Also now it is possible to run test suite on analytic_derivatives, i.e. run `make check m/analytic_derivatives`
## `analytic_derivatives/BrockMirman_PertParamsDerivs.mod`
* This is the Brock Mirman model, where we know the exact policy function $g$ for capital and consumption. As this does not imply a nonzero $g_{\sigma\sigma}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ I added some artificial equations to get nonzero solution matrices with respect to $\sigma$. The true perturbation solution matrices $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ are then computed analytically with Matlab's symbolic toolbox and saved in `nBrockMirmanSYM.mat`. There is a preprocessor flag that recreates these analytical computations if changes are needed (and to check whether I made some errors here ;-) )
* Then solution matrices up to third order and their parameter Jacobians are then compared to the ones computed by Dynare's `k_order_solver` and by `get_perturbation_params_derivs` for all `analytic_derivation_mode`'s. There will be an error if the maximum absolute deviation is too large, i.e. for numerical derivatives (`analytic_derivation_mode=-1|-2`) the tolerance is choosen lower (around 1e-5); for analytical methods we are stricter: around 1e-13 for first-order, 1e-12 for second order, and 1e-11 for third-order.
* As a side note, this mod file also checks Dynare's `k_order_solver` algorithm and throws an error if something is wrong.
* This test model shows that the new functionality works well. And analytical derivatives perform way better and accurate than numerical ones, even for this small model.
## `analytic_derivatives/burnside_3_order_PertParamsDerivs.mod`
* This builds upon `tests/k_order_perturbation/burnside_k_order.mod` and computes the true parameter derivatives analytically by hand.
* This test model also shows that the new functionality works well.
## `analytic_derivatives/LindeTrabandt2019.mod`
* Shows that the new functionality also works for medium-sized models, i.e. a SW type model solved at third order with 35 variables (11 states). 2 shocks and 20 parameters.
* This mod file can be used to tweak the speed of the computations in the future.
* Compares numerical versus analytical parameter derivatives (for first, second and third order). Note that this model clearly shows that numerical ones are quite different than analytical ones even at first order!
## `identification/LindeTrabandt2019_xfail.mod`
* This model is a check for issue Dynare/dynare#1595, see fjaco.m below, and will fail.
* Removed `analytic_derivatives/ls2003.mod` as this mod file is neither in the testsuite nor does it work.
***
# Detailed changes in other functions
## `get_first_order_solution_params_derivs.m`
* Deleted, or actually, renamed to `get_perturbation_params_derivs.m`, as this function now is able to compute the derivatives up to third order
## `identification_numerical_objective.m`
* `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
## `dsge_likelihood.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `T`matrix anymore, but instead of the dynare solution matrix `ghx`. So we recreate `T` here as this amounts to adding some zeros and focusing on selected variables only.
* Added some checks to make sure the first-order approximation is selected.
* Removed `kron_flag` as input, as `get_perturbation_params_derivs` looks into `options_.analytic_derivation_mode` for `kron_flag`.
## `dynare_identification.m`
* make sure that setting `analytic_derivation_mode` is set both in `options_ident` and `options_`. Note that at the end of the function we restore the `options_` structure, so all changes are local. In a next merge request, I will remove the global variables to make all variables local.
## `get_identification_jacobians.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `A` and `B` matrix anymore, but instead of the dynare solution matrix `ghx` and `ghu`. So we recreate these matrices here instead of in `get_perturbation_params_derivs.m`.
* Added `str2func` for better function handles in `fjaco.m`.
## `fjaco.m`
* make `tol`an option, which can be adjusted by changing `options_.dynatol.x`for identification and parameter derivatives purposes.
* include a check and an informative error message, if numerical derivatives (two-sided finite difference method) yield errors in `resol.m` for identification and parameter derivatives purposes. This closes issue Dynare/dynare#1595.
* Changed year of copyright to 2010-2017,2019
***
# Further suggestions and questions
* Ones this is merged, I will merge request an improvement of the identification toolbox, which will work up to third order using the pruned state space. This will also remove some issues and bugs, and also I will remove global variables in this request.
* The third-order derivatives can be further improved by taking sparsity into account and use mex versions for kronecker products etc. I leave this for further testing (and if anybody actually uses this ;-) )
2019-12-17 19:17:09 +01:00
elseif ( analytic_derivation_mode == 0 || analytic_derivation_mode == 1 )
%% Analytical computation of Jacobian and Hessian (wrt selected model parameters) of steady state, i.e. dYss and d2Yss
[ ~ , g1_static ] = feval ( [ M . fname , ' .static' ] , oo . dr . ys , oo . exo_steady_state ' , M . params ) ; %g1_static is [M.endo_nbr by M.endo_nbr] first-derivative (wrt all endogenous variables) of static model equations f, i.e. df/dys, in declaration order
try
rp_static = feval ( [ M . fname , ' .static_params_derivs' ] , oo . dr . ys , oo . exo_steady_state ' , M . params ) ; %rp_static is [M.endo_nbr by M.param_nbr] first-derivative (wrt all model parameters) of static model equations f, i.e. df/dparams, in declaration order
catch
error ( ' For analytical parameter derivatives ' ' static_params_derivs.m' ' file is needed, this can be created by putting identification(order=%d) into your mod file.' , order )
end
dys = - g1_static \ rp_static ; %use implicit function theorem (equation 5 of Ratto and Iskrev (2012) to compute [M.endo_nbr by M.param_nbr] first-derivative (wrt all model parameters) of steady state for all endogenous variables analytically, note that dys is in declaration order
d2ys = zeros ( M . endo_nbr , M . param_nbr , M . param_nbr ) ; %initialize in tensor notation, note that d2ys is only needed for d2flag, i.e. for g1pp
if d2flag
[ ~ , ~ , g2_static ] = feval ( [ M . fname , ' .static' ] , oo . dr . ys , oo . exo_steady_state ' , M . params ) ; %g2_static is [M.endo_nbr by M.endo_nbr^2] second derivative (wrt all endogenous variables) of static model equations f, i.e. d(df/dys)/dys, in declaration order
if order < 3
2019-12-20 16:28:06 +01:00
[ ~ , g1 , g2 , g3 ] = feval ( [ M . fname , ' .dynamic' ] , oo . dr . ys ( I ) , oo . exo_steady_state ' , M . params , oo . dr . ys , 1 ) ; %note that g3 does not contain symmetric elements
Added parameter derivatives of perturbation solution up to 3 order
# Preliminary comments
I finished the identification toolbox at orders two and three using the pruned state space system, but before I merge request this, I decided to first merge the new functionality to compute parameter derivatives of perturbation solution matrices at higher orders. So after this is approved, I merge the identification toolbox.
I guess @rattoma, @sebastien, and @michel are best choices to review this.
I outline the main idea first and then provide some more detailed changes I made to the functions.
***
# Main idea
This merge request is concerned with the *analytical*computation of the parameter derivatives of first, second and third order perturbation solution matrices, i.e. using _closed-form_ expressions to efficiently compute the derivative of $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ *with respect to model parameters* $\theta$. Note that $\theta$ contains model parameters, stderr and corr parameters of shocks. stderr and corr parameters of measurement errors are not yet supported, (they can easily be included as exogenous shocks). The availability of such derivatives is beneficial in terms of more reliable analysis of model sensitivity and parameter identifiability as well as more efficient estimation methods, in particular for models solved up to third order, as it is well-known that numerical derivatives are a tricky business, especially for large models.
References for my approach are:
* Iskrev (2008, 2010) and Schmitt-Grohé and Uribe (2012, Appendix) who were the first to compute the parameter derivatives analytically at first order, however, using inefficient (sparse) Kronecker products.
* Mutschler (2015) who provides the expressions for a second-order, but again using inefficient (sparse) Kronecker products.
* Ratto and Iskrev (2012) who show how the first-order system can be solved accurately, fast and efficiently using existing numerical algorithms for generalized Sylvester equations by taking the parameter derivative with respect to each parameter separately.
* Julliard and Kamenik (2004) who provide the perturbation solution equation system in tensor notation at any order k.
* Levintal (2017) who introduces permutation matrices to express the perturbation solution equation system in matrix notation up to fifth order.
Note that @rattoma already implemented the parameter derivatives of $g_x$ and $g_u$ analytically (and numerically), and I rely heavily on his work in `get_first_order_solution_params_derivs.m` (previously `getH.m`). My additions are mainly to this function and thus it is renamed to `get_perturbation_params_derivs.m`.
The basic idea of this merge request is to take the second- and third-order perturbation solution systems in Julliard and Kamenik (2004), unfold these into an equivalent matrix representation using permutation matrices as in Levintal (2017). Then extending Ratto and Iskrev (2012) one takes the derivative with respect to each parameter separately and gets a computational problem that is linear, albeit large, as it involves either solving generalized Sylvester equations or taking inverses of highly sparse matrices. I will now briefly summarize the perturbation solution system at third order and the system that results when taking the derivative with respect to parameters.
## Perturbation Solution
The following systems arise at first, second, and third order:
$(ghx): f_{x} z_{x} = f_{y_{-}^*} + f_{y_0} g_{x} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{x}= A g_{x} + f_{y_{-}^*}=0$
$(ghu): f_{z} z_{u} = f_{y_0} g_{u} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{u} + f_{u}= A g_u + f_u = 0$
$(ghxx) : A g_{xx} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{x}\right) + f_{zz} \left( z_{x} \otimes z_{x} \right) = 0$
$(ghxu) : A g_{xu} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{x} \otimes z_{u} \right) = 0$
$(ghuu) : A g_{uu} + B g_{xx} \left(g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{u} \otimes z_{u} \right) = 0$
$(ghs2) : (A+B) g_{\sigma\sigma} + \left( f_{y^{**}_{+}y^{**}_{+}} \left(g^{**}_{u} \otimes g^{**}_{u}\right) + f_{y^{**}_{+}} g^{**}_{uu}\right)vec(\Sigma) = 0$
$(ghxxx) : A g_{xxx} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{x}\right) + f_{y_{+}}g^{**}_{xx} \left(g^{*}_x \otimes g^{*}_{xx}\right)P_{x\_xx} + f_{zz} \left( z_{x} \otimes z_{xx} \right)P_{x\_xx} + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{x} \right) = 0$
$(ghxxu) : A g_{xxu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{u} \right) + f_{zz} \left( \left( z_{x} \otimes z_{xu} \right)P_{x\_xu} + \left(z_{xx} \otimes z_{u}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{x} \otimes g^{*}_{xu}\right)P_{x\_xu} + \left(g^{*}_{xx} \otimes g^{*}_{u}\right) \right) = 0$
$(ghxuu) : A g_{xuu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( \left( z_{xu} \otimes z_{u} \right)P_{xu\_u} + \left(z_{x} \otimes z_{uu}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{xu} \otimes g^{*}_{u}\right)P_{xu\_u} + \left(g^{*}_{x} \otimes g^{*}_{uu}\right) \right) = 0$
$(ghuuu) : A g_{uuu} + B g_{xxx} \left(g^{*}_{u} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{u} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( z_{u} \otimes z_{uu} \right)P_{u\_uu} + f_{y_{+}}g^{**}_{xx} \left(g^{*}_{u} \otimes g^{*}_{uu}\right)P_{u\_uu} = 0$
$(ghx\sigma\sigma) : A g_{x\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_x + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{x} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{x} \otimes z_{\sigma\sigma}\right) + F_{xu_{+}u_{+}}\left(I_{n_x} \otimes vec(\Sigma)\right) = 0$
$F_{xu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_x^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{xu_{+}} \otimes z_{u_{+}} \right)P_{xu\_u} + \left(z_{x} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{x} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
$(ghu\sigma\sigma) : A g_{u\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_{u} + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{u} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{u} \otimes z_{\sigma\sigma}\right) + F_{uu_{+}u_{+}}\left(I_{n_u} \otimes vec(\Sigma_u)\right) = 0$
$F_{uu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_u^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{uu_{+}} \otimes z_{u_{+}} \right)P_{uu\_u} + \left(z_{u} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{u} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
A and B are the common perturbation matrices:
$A = f_{y_0} + \begin{pmatrix} \underbrace{0}_{n\times n_{static}} &\vdots& \underbrace{f_{y^{**}_{+}} \cdot g^{**}_{x}}_{n \times n_{spred}} &\vdots& \underbrace{0}_{n\times n_{frwd}} \end{pmatrix}$and $B = \begin{pmatrix} \underbrace{0}_{n \times n_{static}}&\vdots & \underbrace{0}_{n \times n_{pred}} & \vdots & \underbrace{f_{y^{**}_{+}}}_{n \times n_{sfwrd}} \end{pmatrix}$
and $z=(y_{-}^{\ast}; y; y_{+}^{\ast\ast}; u)$ denotes the dynamic model variables as in `M_.lead_lag_incidence`, $y^\ast$ denote state variables, $y^{\ast\ast}$ denote forward looking variables, $y_+$ denote the variables with a lead, $y_{-}$ denote variables with a lag, $y_0$ denote variables at period t, $f$ the model equations, and $f_z$ the first-order dynamic model derivatives, $f_{zz}$ the second-order dynamic derivatives, and $f_{zzz}$ the third-order dynamic model derivatives. Then:
$z_{x} = \begin{pmatrix}I\\g_{x}\\g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{u} =\begin{pmatrix}0\\g_{u}\\g^{**}_{x} \cdot g^{*}_{u}\\I\end{pmatrix}$, $z_{u_{+}} =\begin{pmatrix}0\\0\\g^{**}_{u}\\0\end{pmatrix}$
$z_{xx} = \begin{pmatrix} 0\\g_{xx}\\g^{**}_{x} \left( g^{*}_x \otimes g^{*}_{x} \right) + g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{xu} =\begin{pmatrix}0\\g_{xu}\\g^{**}_{xx} \left( g^{*}_x \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{xu}\\0\end{pmatrix}$, $z_{uu} =\begin{pmatrix}0\\g_{uu}\\g^{**}_{xx} \left( g^{*}_u \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{uu}\\0\end{pmatrix}$,
$z_{xu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_x \otimes I \right)\\0\end{pmatrix}$, $z_{uu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_{u} \otimes I \right)\\0\end{pmatrix}$, $z_{u_{+}u_{+}} =\begin{pmatrix}0\\0\\g^{\ast\ast}_{uu}\\0\end{pmatrix}$, $z_{\sigma\sigma} = \begin{pmatrix}0\\ g_{\sigma\sigma}\\ g^{\ast\ast}_{x}g^{\ast}_{\sigma\sigma} + g^{\ast\ast}_{\sigma\sigma}\\0 \end{pmatrix}$
$P$ are permutation matrices that can be computed using Matlab's `ipermute` function.
## Parameter derivatives of perturbation solutions
First, we need the parameter derivatives of first, second, third, and fourth derivatives of the dynamic model (i.e. g1,g2,g3,g4 in dynamic files), I make use of the implicit function theorem: Let $f_{z^k}$ denote the kth derivative (wrt all dynamic variables) of the dynamic model, then let $df_{z^k}$ denote the first-derivative (wrt all model parameters) of $f_{z^k}$ evaluated at the steady state. Note that $f_{z^k}$ is a function of both the model parameters $\theta$ and of the steady state of all dynamic variables $\bar{z}$, which also depend on the parameters. Hence, implicitly $f_{z^k}=f_{z^k}(\theta,\bar{z}(\theta))$ and $df_{z^k}$ consists of two parts:
1. direct derivative wrt to all model parameters given by the preprocessor in the `_params_derivs.m` files
2. contribution of derivative of steady state of dynamic variables (wrt all model parameters): $f_{z^{k+1}} \cdot d\bar{z}$
Note that we already have functionality to compute $d\bar{z}$ analytically.
Having this, the above perturbation systems are basically equations of the following types
$AX +BXC = RHS$ or $AX = RHS$
Now when taking the derivative (wrt to one single parameter $\theta_j$), we get
$A\mathrm{d}\{X\} + B\mathrm{d}\{X\}C = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X - \mathrm{d}\{B\}XC - BX\mathrm{d}\{C\}$
or
$A\mathrm{d}\{X\} = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X$
The first one is a Sylvester type equation, the second one can be solved by taking the inverse of $A$. The only diffculty and tedious work arrises in computing (the highly sparse) derivatives of $RHS$.
***
# New functions: `
## get_perturbation_params_derivs.m`and `get_perturbation_params_derivs_numerical_objective.m`
* The parameter derivatives up to third order are computed in the new function`get_perturbation_params_derivs.m` both analytically and numerically. For numerical derivatives `get_perturbation_params_derivs_numerical_objective.m` is the objective for `fjaco.m` or `hessian_sparse.m` or `hessian.m`.
* `get_perturbation_params_derivs.m` is basically an extended version of the previous `get_first_order_solution_params_derivs.m` function.
* * `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It is used for numerical derivatives, whenever `analytic_derivation_mode=-1|-2`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
* Detailed changes:
* Most important: notation of this function is now in accordance to the k_order_solver, i.e. we do not compute derivatives of Kalman transition matrices A and B, but rather the solution matrices ghx,ghu,ghxx,ghxu,ghuu,ghs2,ghxxx,ghxxu,ghxuu,ghuuu,ghxss,ghuss in accordance with notation used in `oo_.dr`. As a byproduct at first-order, focusing on ghx and ghu instead of Kalman transition matrices A and B makes the computations slightly faster for large models (e.g. for Quest the computations were faster by a couple of seconds, not much, but okay).
* Removed use of `kstate`, see also Dynare/dynare#1653 and Dynare/dynare!1656
* Output arguments are stored in a structure `DERIVS`, there is also a flag `d2flag` that computes parameter hessians needed only in `dsge_likelihood.m`.
* Removed `kronflag` as input. `options_.analytic_derivation_mode` is now used instead of `kronflag`.
* Removed `indvar`, an index that was used to selected specific variables in the derivatives. This is not needed, as we always compute the parameter derivatives for all variables first and then select a subset of variables. The selection now takes place in other functions, like `dsge_likelihood.m`.
* Introduced some checks: (i) deterministic exogenous variables are not supported, (ii) Kronecker method only compatible with first-order approximation so reset to sylvester method, (iii) for purely backward or forward models we need to be careful with the rows in `M_.lead_la g_incidence`, (iv) if `_params_derivs.m` files are missing an error is thrown.
* For numerical derivatives, if mod file does not contain an `estimated_params_block`, a temporary one with the most important parameter information is created.
## `unfold_g4.m`
* When evaluating g3 and g4 one needs to take into account that these do not contain symmetric elements, so one needs to use `unfold_g3.m` and the new function `unfold_g4.m`. This returns an unfolded version of the same matrix (i.e. with symmetric elements).
***
# New test models
`.gitignore` and `Makefile.am` are changed accordingly. Also now it is possible to run test suite on analytic_derivatives, i.e. run `make check m/analytic_derivatives`
## `analytic_derivatives/BrockMirman_PertParamsDerivs.mod`
* This is the Brock Mirman model, where we know the exact policy function $g$ for capital and consumption. As this does not imply a nonzero $g_{\sigma\sigma}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ I added some artificial equations to get nonzero solution matrices with respect to $\sigma$. The true perturbation solution matrices $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ are then computed analytically with Matlab's symbolic toolbox and saved in `nBrockMirmanSYM.mat`. There is a preprocessor flag that recreates these analytical computations if changes are needed (and to check whether I made some errors here ;-) )
* Then solution matrices up to third order and their parameter Jacobians are then compared to the ones computed by Dynare's `k_order_solver` and by `get_perturbation_params_derivs` for all `analytic_derivation_mode`'s. There will be an error if the maximum absolute deviation is too large, i.e. for numerical derivatives (`analytic_derivation_mode=-1|-2`) the tolerance is choosen lower (around 1e-5); for analytical methods we are stricter: around 1e-13 for first-order, 1e-12 for second order, and 1e-11 for third-order.
* As a side note, this mod file also checks Dynare's `k_order_solver` algorithm and throws an error if something is wrong.
* This test model shows that the new functionality works well. And analytical derivatives perform way better and accurate than numerical ones, even for this small model.
## `analytic_derivatives/burnside_3_order_PertParamsDerivs.mod`
* This builds upon `tests/k_order_perturbation/burnside_k_order.mod` and computes the true parameter derivatives analytically by hand.
* This test model also shows that the new functionality works well.
## `analytic_derivatives/LindeTrabandt2019.mod`
* Shows that the new functionality also works for medium-sized models, i.e. a SW type model solved at third order with 35 variables (11 states). 2 shocks and 20 parameters.
* This mod file can be used to tweak the speed of the computations in the future.
* Compares numerical versus analytical parameter derivatives (for first, second and third order). Note that this model clearly shows that numerical ones are quite different than analytical ones even at first order!
## `identification/LindeTrabandt2019_xfail.mod`
* This model is a check for issue Dynare/dynare#1595, see fjaco.m below, and will fail.
* Removed `analytic_derivatives/ls2003.mod` as this mod file is neither in the testsuite nor does it work.
***
# Detailed changes in other functions
## `get_first_order_solution_params_derivs.m`
* Deleted, or actually, renamed to `get_perturbation_params_derivs.m`, as this function now is able to compute the derivatives up to third order
## `identification_numerical_objective.m`
* `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
## `dsge_likelihood.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `T`matrix anymore, but instead of the dynare solution matrix `ghx`. So we recreate `T` here as this amounts to adding some zeros and focusing on selected variables only.
* Added some checks to make sure the first-order approximation is selected.
* Removed `kron_flag` as input, as `get_perturbation_params_derivs` looks into `options_.analytic_derivation_mode` for `kron_flag`.
## `dynare_identification.m`
* make sure that setting `analytic_derivation_mode` is set both in `options_ident` and `options_`. Note that at the end of the function we restore the `options_` structure, so all changes are local. In a next merge request, I will remove the global variables to make all variables local.
## `get_identification_jacobians.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `A` and `B` matrix anymore, but instead of the dynare solution matrix `ghx` and `ghu`. So we recreate these matrices here instead of in `get_perturbation_params_derivs.m`.
* Added `str2func` for better function handles in `fjaco.m`.
## `fjaco.m`
* make `tol`an option, which can be adjusted by changing `options_.dynatol.x`for identification and parameter derivatives purposes.
* include a check and an informative error message, if numerical derivatives (two-sided finite difference method) yield errors in `resol.m` for identification and parameter derivatives purposes. This closes issue Dynare/dynare#1595.
* Changed year of copyright to 2010-2017,2019
***
# Further suggestions and questions
* Ones this is merged, I will merge request an improvement of the identification toolbox, which will work up to third order using the pruned state space. This will also remove some issues and bugs, and also I will remove global variables in this request.
* The third-order derivatives can be further improved by taking sparsity into account and use mex versions for kronecker products etc. I leave this for further testing (and if anybody actually uses this ;-) )
2019-12-17 19:17:09 +01:00
g3 = unfold_g3 ( g3 , yy0ex0_nbr ) ; %add symmetric elements to g3
else
T = NaN ( sum ( M . dynamic_tmp_nbr ( 1 : 5 ) ) ) ;
T = feval ( [ M . fname , ' .dynamic_g4_tt' ] , T , oo . dr . ys ( I ) , oo . exo_steady_state ' , M . params , oo . dr . ys , 1 ) ;
g1 = feval ( [ M . fname , ' .dynamic_g1' ] , T , oo . dr . ys ( I ) , oo . exo_steady_state ' , M . params , oo . dr . ys , 1 , false ) ; %g1 is [M.endo_nbr by yy0ex0_nbr first derivative (wrt all dynamic variables) of dynamic model equations, i.e. df/dyy0ex0, rows are in declaration order, columns in lead_lag_incidence order
g2 = feval ( [ M . fname , ' .dynamic_g2' ] , T , oo . dr . ys ( I ) , oo . exo_steady_state ' , M . params , oo . dr . ys , 1 , false ) ; %g2 is [M.endo_nbr by yy0ex0_nbr^2] second derivative (wrt all dynamic variables) of dynamic model equations, i.e. d(df/dyy0ex0)/dyy0ex0, rows are in declaration order, columns in lead_lag_incidence order
g3 = feval ( [ M . fname , ' .dynamic_g3' ] , T , oo . dr . ys ( I ) , oo . exo_steady_state ' , M . params , oo . dr . ys , 1 , false ) ; %note that g3 does not contain symmetric elements
g4 = feval ( [ M . fname , ' .dynamic_g4' ] , T , oo . dr . ys ( I ) , oo . exo_steady_state ' , M . params , oo . dr . ys , 1 , false ) ; %note that g4 does not contain symmetric elements
g3 = unfold_g3 ( g3 , yy0ex0_nbr ) ; %add symmetric elements to g3, %g3 is [M.endo_nbr by yy0ex0_nbr^3] third-derivative (wrt all dynamic variables) of dynamic model equations, i.e. (d(df/dyy0ex0)/dyy0ex0)/dyy0ex0, rows are in declaration order, columns in lead_lag_incidence order
g4 = unfold_g4 ( g4 , yy0ex0_nbr ) ; %add symmetric elements to g4, %g4 is [M.endo_nbr by yy0ex0_nbr^4] fourth-derivative (wrt all dynamic variables) of dynamic model equations, i.e. ((d(df/dyy0ex0)/dyy0ex0)/dyy0ex0)/dyy0ex0, rows are in declaration order, columns in lead_lag_incidence order
end
%g1 is [M.endo_nbr by yy0ex0_nbr first derivative (wrt all dynamic variables) of dynamic model equations, i.e. df/dyy0ex0, rows are in declaration order, columns in lead_lag_incidence order
%g2 is [M.endo_nbr by yy0ex0_nbr^2] second derivative (wrt all dynamic variables) of dynamic model equations, i.e. d(df/dyy0ex0)/dyy0ex0, rows are in declaration order, columns in lead_lag_incidence order
%g3 is [M.endo_nbr by yy0ex0_nbr^3] third-derivative (wrt all dynamic variables) of dynamic model equations, i.e. (d(df/dyy0ex0)/dyy0ex0)/dyy0ex0, rows are in declaration order, columns in lead_lag_incidence order
try
[ ~ , g1p_static , rpp_static ] = feval ( [ M . fname , ' .static_params_derivs' ] , oo . dr . ys , oo . exo_steady_state ' , M . params ) ;
%g1p_static is [M.endo_nbr by M.endo_nbr by M.param_nbr] first derivative (wrt all model parameters) of first-derivative (wrt all endogenous variables) of static model equations f, i.e. (df/dys)/dparams, in declaration order
%rpp_static is [#second_order_residual_terms by 4] and contains nonzero values and corresponding indices of second derivatives (wrt all model parameters) of static model equations f, i.e. d(df/dparams)/dparams, in declaration order, where
% column 1 contains equation number; column 2 contains first parameter; column 3 contains second parameter; column 4 contains value of derivative
catch
error ( ' For analytical parameter derivatives ' ' static_params_derivs.m' ' file is needed, this can be created by putting identification(order=%d) into your mod file.' , order )
end
rpp_static = get_all_resid_2nd_derivs ( rpp_static , M . endo_nbr , M . param_nbr ) ; %make full matrix out of nonzero values and corresponding indices
%rpp_static is [M.endo_nbr by M.param_nbr by M.param_nbr] second derivatives (wrt all model parameters) of static model equations, i.e. d(df/dparams)/dparams, in declaration order
if isempty ( find ( g2_static ) )
%auxiliary expression on page 8 of Ratto and Iskrev (2012) is zero, i.e. gam = 0
for j = 1 : M . param_nbr
%using the implicit function theorem, equation 15 on page 7 of Ratto and Iskrev (2012)
d2ys ( : , : , j ) = - g1_static \ rpp_static ( : , : , j ) ;
%d2ys is [M.endo_nbr by M.param_nbr by M.param_nbr] second-derivative (wrt all model parameters) of steady state of all endogenous variables, i.e. d(dys/dparams)/dparams, in declaration order
end
else
gam = zeros ( M . endo_nbr , M . param_nbr , M . param_nbr ) ; %initialize auxiliary expression on page 8 of Ratto and Iskrev (2012)
for j = 1 : M . endo_nbr
tmp_g1p_static_dys = ( squeeze ( g1p_static ( j , : , : ) ) ' * dys ) ;
gam ( j , : , : ) = transpose ( reshape ( g2_static ( j , : ) , [ M . endo_nbr M . endo_nbr ] ) * dys ) * dys + tmp_g1p_static_dys + tmp_g1p_static_dys ' ;
end
for j = 1 : M . param_nbr
%using the implicit function theorem, equation 15 on page 7 of Ratto and Iskrev (2012)
d2ys ( : , : , j ) = - g1_static \ ( rpp_static ( : , : , j ) + gam ( : , : , j ) ) ;
%d2ys is [M.endo_nbr by M.param_nbr by M.param_nbr] second-derivative (wrt all model parameters) of steady state of all endogenous variables, i.e. d(dys/dparams)/dparams, in declaration order
end
clear g1p_static g2_static tmp_g1p_static_dys gam
end
end
%handling of steady state for nonstationary variables
if any ( any ( isnan ( dys ) ) )
[ U , T ] = schur ( g1_static ) ;
e1 = abs ( ordeig ( T ) ) < qz_criterium - 1 ;
k = sum ( e1 ) ; % Number of non stationary variables.
% Number of stationary variables: n = length(e1)-k
[ U , T ] = ordschur ( U , T , e1 ) ;
T = T ( k + 1 : end , k + 1 : end ) ;
%using implicit function theorem, equation 5 of Ratto and Iskrev (2012), in declaration order
dys = - U ( : , k + 1 : end ) * ( T \ U ( : , k + 1 : end ) ' ) * rp_static ;
if d2flag
disp ( ' Computation of d2ys for nonstationary variables is not yet correctly handled if g2_static is nonempty, but continue anyways...' )
for j = 1 : M . param_nbr
%using implicit function theorem, equation 15 of Ratto and Iskrev (2012), in declaration order
d2ys ( : , : , j ) = - U ( : , k + 1 : end ) * ( T \ U ( : , k + 1 : end ) ' ) * rpp_static ( : , : , j ) ; %THIS IS NOT CORRECT, IF g2_static IS NONEMPTY. WE NEED TO ADD GAM [willi]
end
end
end
if d2flag
try
if order < 3
[ ~ , g1p , ~ , g1pp , g2p ] = feval ( [ M . fname , ' .dynamic_params_derivs' ] , oo . dr . ys ( I ) , oo . exo_steady_state ' , M . params , oo . dr . ys , 1 , dys , d2ys ) ;
else
[ ~ , g1p , ~ , g1pp , g2p , g3p ] = feval ( [ M . fname , ' .dynamic_params_derivs' ] , oo . dr . ys ( I ) , oo . exo_steady_state ' , M . params , oo . dr . ys , 1 , dys , d2ys ) ;
end
catch
2019-12-20 16:28:06 +01:00
error ( ' For analytical parameter derivatives ' ' dynamic_params_derivs.m' ' file is needed, this can be created by putting identification(order=%d) into your mod file.' , order )
Added parameter derivatives of perturbation solution up to 3 order
# Preliminary comments
I finished the identification toolbox at orders two and three using the pruned state space system, but before I merge request this, I decided to first merge the new functionality to compute parameter derivatives of perturbation solution matrices at higher orders. So after this is approved, I merge the identification toolbox.
I guess @rattoma, @sebastien, and @michel are best choices to review this.
I outline the main idea first and then provide some more detailed changes I made to the functions.
***
# Main idea
This merge request is concerned with the *analytical*computation of the parameter derivatives of first, second and third order perturbation solution matrices, i.e. using _closed-form_ expressions to efficiently compute the derivative of $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ *with respect to model parameters* $\theta$. Note that $\theta$ contains model parameters, stderr and corr parameters of shocks. stderr and corr parameters of measurement errors are not yet supported, (they can easily be included as exogenous shocks). The availability of such derivatives is beneficial in terms of more reliable analysis of model sensitivity and parameter identifiability as well as more efficient estimation methods, in particular for models solved up to third order, as it is well-known that numerical derivatives are a tricky business, especially for large models.
References for my approach are:
* Iskrev (2008, 2010) and Schmitt-Grohé and Uribe (2012, Appendix) who were the first to compute the parameter derivatives analytically at first order, however, using inefficient (sparse) Kronecker products.
* Mutschler (2015) who provides the expressions for a second-order, but again using inefficient (sparse) Kronecker products.
* Ratto and Iskrev (2012) who show how the first-order system can be solved accurately, fast and efficiently using existing numerical algorithms for generalized Sylvester equations by taking the parameter derivative with respect to each parameter separately.
* Julliard and Kamenik (2004) who provide the perturbation solution equation system in tensor notation at any order k.
* Levintal (2017) who introduces permutation matrices to express the perturbation solution equation system in matrix notation up to fifth order.
Note that @rattoma already implemented the parameter derivatives of $g_x$ and $g_u$ analytically (and numerically), and I rely heavily on his work in `get_first_order_solution_params_derivs.m` (previously `getH.m`). My additions are mainly to this function and thus it is renamed to `get_perturbation_params_derivs.m`.
The basic idea of this merge request is to take the second- and third-order perturbation solution systems in Julliard and Kamenik (2004), unfold these into an equivalent matrix representation using permutation matrices as in Levintal (2017). Then extending Ratto and Iskrev (2012) one takes the derivative with respect to each parameter separately and gets a computational problem that is linear, albeit large, as it involves either solving generalized Sylvester equations or taking inverses of highly sparse matrices. I will now briefly summarize the perturbation solution system at third order and the system that results when taking the derivative with respect to parameters.
## Perturbation Solution
The following systems arise at first, second, and third order:
$(ghx): f_{x} z_{x} = f_{y_{-}^*} + f_{y_0} g_{x} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{x}= A g_{x} + f_{y_{-}^*}=0$
$(ghu): f_{z} z_{u} = f_{y_0} g_{u} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{u} + f_{u}= A g_u + f_u = 0$
$(ghxx) : A g_{xx} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{x}\right) + f_{zz} \left( z_{x} \otimes z_{x} \right) = 0$
$(ghxu) : A g_{xu} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{x} \otimes z_{u} \right) = 0$
$(ghuu) : A g_{uu} + B g_{xx} \left(g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{u} \otimes z_{u} \right) = 0$
$(ghs2) : (A+B) g_{\sigma\sigma} + \left( f_{y^{**}_{+}y^{**}_{+}} \left(g^{**}_{u} \otimes g^{**}_{u}\right) + f_{y^{**}_{+}} g^{**}_{uu}\right)vec(\Sigma) = 0$
$(ghxxx) : A g_{xxx} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{x}\right) + f_{y_{+}}g^{**}_{xx} \left(g^{*}_x \otimes g^{*}_{xx}\right)P_{x\_xx} + f_{zz} \left( z_{x} \otimes z_{xx} \right)P_{x\_xx} + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{x} \right) = 0$
$(ghxxu) : A g_{xxu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{u} \right) + f_{zz} \left( \left( z_{x} \otimes z_{xu} \right)P_{x\_xu} + \left(z_{xx} \otimes z_{u}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{x} \otimes g^{*}_{xu}\right)P_{x\_xu} + \left(g^{*}_{xx} \otimes g^{*}_{u}\right) \right) = 0$
$(ghxuu) : A g_{xuu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( \left( z_{xu} \otimes z_{u} \right)P_{xu\_u} + \left(z_{x} \otimes z_{uu}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{xu} \otimes g^{*}_{u}\right)P_{xu\_u} + \left(g^{*}_{x} \otimes g^{*}_{uu}\right) \right) = 0$
$(ghuuu) : A g_{uuu} + B g_{xxx} \left(g^{*}_{u} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{u} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( z_{u} \otimes z_{uu} \right)P_{u\_uu} + f_{y_{+}}g^{**}_{xx} \left(g^{*}_{u} \otimes g^{*}_{uu}\right)P_{u\_uu} = 0$
$(ghx\sigma\sigma) : A g_{x\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_x + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{x} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{x} \otimes z_{\sigma\sigma}\right) + F_{xu_{+}u_{+}}\left(I_{n_x} \otimes vec(\Sigma)\right) = 0$
$F_{xu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_x^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{xu_{+}} \otimes z_{u_{+}} \right)P_{xu\_u} + \left(z_{x} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{x} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
$(ghu\sigma\sigma) : A g_{u\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_{u} + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{u} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{u} \otimes z_{\sigma\sigma}\right) + F_{uu_{+}u_{+}}\left(I_{n_u} \otimes vec(\Sigma_u)\right) = 0$
$F_{uu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_u^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{uu_{+}} \otimes z_{u_{+}} \right)P_{uu\_u} + \left(z_{u} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{u} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
A and B are the common perturbation matrices:
$A = f_{y_0} + \begin{pmatrix} \underbrace{0}_{n\times n_{static}} &\vdots& \underbrace{f_{y^{**}_{+}} \cdot g^{**}_{x}}_{n \times n_{spred}} &\vdots& \underbrace{0}_{n\times n_{frwd}} \end{pmatrix}$and $B = \begin{pmatrix} \underbrace{0}_{n \times n_{static}}&\vdots & \underbrace{0}_{n \times n_{pred}} & \vdots & \underbrace{f_{y^{**}_{+}}}_{n \times n_{sfwrd}} \end{pmatrix}$
and $z=(y_{-}^{\ast}; y; y_{+}^{\ast\ast}; u)$ denotes the dynamic model variables as in `M_.lead_lag_incidence`, $y^\ast$ denote state variables, $y^{\ast\ast}$ denote forward looking variables, $y_+$ denote the variables with a lead, $y_{-}$ denote variables with a lag, $y_0$ denote variables at period t, $f$ the model equations, and $f_z$ the first-order dynamic model derivatives, $f_{zz}$ the second-order dynamic derivatives, and $f_{zzz}$ the third-order dynamic model derivatives. Then:
$z_{x} = \begin{pmatrix}I\\g_{x}\\g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{u} =\begin{pmatrix}0\\g_{u}\\g^{**}_{x} \cdot g^{*}_{u}\\I\end{pmatrix}$, $z_{u_{+}} =\begin{pmatrix}0\\0\\g^{**}_{u}\\0\end{pmatrix}$
$z_{xx} = \begin{pmatrix} 0\\g_{xx}\\g^{**}_{x} \left( g^{*}_x \otimes g^{*}_{x} \right) + g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{xu} =\begin{pmatrix}0\\g_{xu}\\g^{**}_{xx} \left( g^{*}_x \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{xu}\\0\end{pmatrix}$, $z_{uu} =\begin{pmatrix}0\\g_{uu}\\g^{**}_{xx} \left( g^{*}_u \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{uu}\\0\end{pmatrix}$,
$z_{xu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_x \otimes I \right)\\0\end{pmatrix}$, $z_{uu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_{u} \otimes I \right)\\0\end{pmatrix}$, $z_{u_{+}u_{+}} =\begin{pmatrix}0\\0\\g^{\ast\ast}_{uu}\\0\end{pmatrix}$, $z_{\sigma\sigma} = \begin{pmatrix}0\\ g_{\sigma\sigma}\\ g^{\ast\ast}_{x}g^{\ast}_{\sigma\sigma} + g^{\ast\ast}_{\sigma\sigma}\\0 \end{pmatrix}$
$P$ are permutation matrices that can be computed using Matlab's `ipermute` function.
## Parameter derivatives of perturbation solutions
First, we need the parameter derivatives of first, second, third, and fourth derivatives of the dynamic model (i.e. g1,g2,g3,g4 in dynamic files), I make use of the implicit function theorem: Let $f_{z^k}$ denote the kth derivative (wrt all dynamic variables) of the dynamic model, then let $df_{z^k}$ denote the first-derivative (wrt all model parameters) of $f_{z^k}$ evaluated at the steady state. Note that $f_{z^k}$ is a function of both the model parameters $\theta$ and of the steady state of all dynamic variables $\bar{z}$, which also depend on the parameters. Hence, implicitly $f_{z^k}=f_{z^k}(\theta,\bar{z}(\theta))$ and $df_{z^k}$ consists of two parts:
1. direct derivative wrt to all model parameters given by the preprocessor in the `_params_derivs.m` files
2. contribution of derivative of steady state of dynamic variables (wrt all model parameters): $f_{z^{k+1}} \cdot d\bar{z}$
Note that we already have functionality to compute $d\bar{z}$ analytically.
Having this, the above perturbation systems are basically equations of the following types
$AX +BXC = RHS$ or $AX = RHS$
Now when taking the derivative (wrt to one single parameter $\theta_j$), we get
$A\mathrm{d}\{X\} + B\mathrm{d}\{X\}C = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X - \mathrm{d}\{B\}XC - BX\mathrm{d}\{C\}$
or
$A\mathrm{d}\{X\} = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X$
The first one is a Sylvester type equation, the second one can be solved by taking the inverse of $A$. The only diffculty and tedious work arrises in computing (the highly sparse) derivatives of $RHS$.
***
# New functions: `
## get_perturbation_params_derivs.m`and `get_perturbation_params_derivs_numerical_objective.m`
* The parameter derivatives up to third order are computed in the new function`get_perturbation_params_derivs.m` both analytically and numerically. For numerical derivatives `get_perturbation_params_derivs_numerical_objective.m` is the objective for `fjaco.m` or `hessian_sparse.m` or `hessian.m`.
* `get_perturbation_params_derivs.m` is basically an extended version of the previous `get_first_order_solution_params_derivs.m` function.
* * `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It is used for numerical derivatives, whenever `analytic_derivation_mode=-1|-2`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
* Detailed changes:
* Most important: notation of this function is now in accordance to the k_order_solver, i.e. we do not compute derivatives of Kalman transition matrices A and B, but rather the solution matrices ghx,ghu,ghxx,ghxu,ghuu,ghs2,ghxxx,ghxxu,ghxuu,ghuuu,ghxss,ghuss in accordance with notation used in `oo_.dr`. As a byproduct at first-order, focusing on ghx and ghu instead of Kalman transition matrices A and B makes the computations slightly faster for large models (e.g. for Quest the computations were faster by a couple of seconds, not much, but okay).
* Removed use of `kstate`, see also Dynare/dynare#1653 and Dynare/dynare!1656
* Output arguments are stored in a structure `DERIVS`, there is also a flag `d2flag` that computes parameter hessians needed only in `dsge_likelihood.m`.
* Removed `kronflag` as input. `options_.analytic_derivation_mode` is now used instead of `kronflag`.
* Removed `indvar`, an index that was used to selected specific variables in the derivatives. This is not needed, as we always compute the parameter derivatives for all variables first and then select a subset of variables. The selection now takes place in other functions, like `dsge_likelihood.m`.
* Introduced some checks: (i) deterministic exogenous variables are not supported, (ii) Kronecker method only compatible with first-order approximation so reset to sylvester method, (iii) for purely backward or forward models we need to be careful with the rows in `M_.lead_la g_incidence`, (iv) if `_params_derivs.m` files are missing an error is thrown.
* For numerical derivatives, if mod file does not contain an `estimated_params_block`, a temporary one with the most important parameter information is created.
## `unfold_g4.m`
* When evaluating g3 and g4 one needs to take into account that these do not contain symmetric elements, so one needs to use `unfold_g3.m` and the new function `unfold_g4.m`. This returns an unfolded version of the same matrix (i.e. with symmetric elements).
***
# New test models
`.gitignore` and `Makefile.am` are changed accordingly. Also now it is possible to run test suite on analytic_derivatives, i.e. run `make check m/analytic_derivatives`
## `analytic_derivatives/BrockMirman_PertParamsDerivs.mod`
* This is the Brock Mirman model, where we know the exact policy function $g$ for capital and consumption. As this does not imply a nonzero $g_{\sigma\sigma}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ I added some artificial equations to get nonzero solution matrices with respect to $\sigma$. The true perturbation solution matrices $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ are then computed analytically with Matlab's symbolic toolbox and saved in `nBrockMirmanSYM.mat`. There is a preprocessor flag that recreates these analytical computations if changes are needed (and to check whether I made some errors here ;-) )
* Then solution matrices up to third order and their parameter Jacobians are then compared to the ones computed by Dynare's `k_order_solver` and by `get_perturbation_params_derivs` for all `analytic_derivation_mode`'s. There will be an error if the maximum absolute deviation is too large, i.e. for numerical derivatives (`analytic_derivation_mode=-1|-2`) the tolerance is choosen lower (around 1e-5); for analytical methods we are stricter: around 1e-13 for first-order, 1e-12 for second order, and 1e-11 for third-order.
* As a side note, this mod file also checks Dynare's `k_order_solver` algorithm and throws an error if something is wrong.
* This test model shows that the new functionality works well. And analytical derivatives perform way better and accurate than numerical ones, even for this small model.
## `analytic_derivatives/burnside_3_order_PertParamsDerivs.mod`
* This builds upon `tests/k_order_perturbation/burnside_k_order.mod` and computes the true parameter derivatives analytically by hand.
* This test model also shows that the new functionality works well.
## `analytic_derivatives/LindeTrabandt2019.mod`
* Shows that the new functionality also works for medium-sized models, i.e. a SW type model solved at third order with 35 variables (11 states). 2 shocks and 20 parameters.
* This mod file can be used to tweak the speed of the computations in the future.
* Compares numerical versus analytical parameter derivatives (for first, second and third order). Note that this model clearly shows that numerical ones are quite different than analytical ones even at first order!
## `identification/LindeTrabandt2019_xfail.mod`
* This model is a check for issue Dynare/dynare#1595, see fjaco.m below, and will fail.
* Removed `analytic_derivatives/ls2003.mod` as this mod file is neither in the testsuite nor does it work.
***
# Detailed changes in other functions
## `get_first_order_solution_params_derivs.m`
* Deleted, or actually, renamed to `get_perturbation_params_derivs.m`, as this function now is able to compute the derivatives up to third order
## `identification_numerical_objective.m`
* `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
## `dsge_likelihood.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `T`matrix anymore, but instead of the dynare solution matrix `ghx`. So we recreate `T` here as this amounts to adding some zeros and focusing on selected variables only.
* Added some checks to make sure the first-order approximation is selected.
* Removed `kron_flag` as input, as `get_perturbation_params_derivs` looks into `options_.analytic_derivation_mode` for `kron_flag`.
## `dynare_identification.m`
* make sure that setting `analytic_derivation_mode` is set both in `options_ident` and `options_`. Note that at the end of the function we restore the `options_` structure, so all changes are local. In a next merge request, I will remove the global variables to make all variables local.
## `get_identification_jacobians.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `A` and `B` matrix anymore, but instead of the dynare solution matrix `ghx` and `ghu`. So we recreate these matrices here instead of in `get_perturbation_params_derivs.m`.
* Added `str2func` for better function handles in `fjaco.m`.
## `fjaco.m`
* make `tol`an option, which can be adjusted by changing `options_.dynatol.x`for identification and parameter derivatives purposes.
* include a check and an informative error message, if numerical derivatives (two-sided finite difference method) yield errors in `resol.m` for identification and parameter derivatives purposes. This closes issue Dynare/dynare#1595.
* Changed year of copyright to 2010-2017,2019
***
# Further suggestions and questions
* Ones this is merged, I will merge request an improvement of the identification toolbox, which will work up to third order using the pruned state space. This will also remove some issues and bugs, and also I will remove global variables in this request.
* The third-order derivatives can be further improved by taking sparsity into account and use mex versions for kronecker products etc. I leave this for further testing (and if anybody actually uses this ;-) )
2019-12-17 19:17:09 +01:00
end
%g1pp are nonzero values and corresponding indices of second-derivatives (wrt all model parameters) of first-derivative (wrt all dynamic variables) of dynamic model equations, i.e. d(d(df/dyy0ex0)/dparam)/dparam, rows are in declaration order, first column in declaration order
d2Yss = d2ys ( oo . dr . order_var , indpmodel , indpmodel ) ; %[M.endo_nbr by mod_param_nbr by mod_param_nbr], put into DR order and focus only on selected model parameters
else
if order == 1
try
[ ~ , g1p ] = feval ( [ M . fname , ' .dynamic_params_derivs' ] , oo . dr . ys ( I ) , oo . exo_steady_state ' , M . params , oo . dr . ys , 1 , dys , d2ys ) ;
%g1p is [M.endo_nbr by yy0ex0_nbr by M.param_nbr] first-derivative (wrt all model parameters) of first-derivative (wrt all dynamic variables) of dynamic model equations, i.e. d(df/dyy0ex0)/dparam, rows are in declaration order, column in lead_lag_incidence order
catch
error ( ' For analytical parameter derivatives ' ' dynamic_params_derivs.m' ' file is needed, this can be created by putting identification(order=%d) into your mod file.' , order )
end
[ ~ , g1 , g2 ] = feval ( [ M . fname , ' .dynamic' ] , oo . dr . ys ( I ) , oo . exo_steady_state ' , M . params , oo . dr . ys , 1 ) ;
%g1 is [M.endo_nbr by yy0ex0_nbr first derivative (wrt all dynamic variables) of dynamic model equations, i.e. df/dyy0ex0, rows are in declaration order, columns in lead_lag_incidence order
%g2 is [M.endo_nbr by yy0ex0_nbr^2] second derivatives (wrt all dynamic variables) of dynamic model equations, i.e. d(df/dyy0ex0)/dyy0ex0, rows are in declaration order, columns in lead_lag_incidence order
elseif order == 2
try
[ ~ , g1p , ~ , ~ , g2p ] = feval ( [ M . fname , ' .dynamic_params_derivs' ] , oo . dr . ys ( I ) , oo . exo_steady_state ' , M . params , oo . dr . ys , 1 , dys , d2ys ) ;
%g1p is [M.endo_nbr by yy0ex0_nbr by M.param_nbr] first-derivative (wrt all model parameters) of first-derivative (wrt all dynamic variables) of dynamic model equations, i.e. d(df/dyy0ex0)/dparam, rows are in declaration order, column in lead_lag_incidence order
%g2p are nonzero values and corresponding indices of first-derivative (wrt all model parameters) of second-derivatives (wrt all dynamic variables) of dynamic model equations, i.e. d(d(df/dyy0ex0)/dyy0ex0)/dparam, rows are in declaration order, first and second column in declaration order
catch
error ( ' For analytical parameter derivatives ' ' dynamic_params_derivs.m' ' file is needed, this can be created by putting identification(order=%d) into your mod file.' , order )
end
[ ~ , g1 , g2 , g3 ] = feval ( [ M . fname , ' .dynamic' ] , oo . dr . ys ( I ) , oo . exo_steady_state ' , M . params , oo . dr . ys , 1 ) ; %note that g3 does not contain symmetric elements
g3 = unfold_g3 ( g3 , yy0ex0_nbr ) ; %add symmetric elements to g3
%g1 is [M.endo_nbr by yy0ex0_nbr first derivative (wrt all dynamic variables) of dynamic model equations, i.e. df/dyy0ex0, rows are in declaration order, columns in lead_lag_incidence order
%g2 is [M.endo_nbr by yy0ex0_nbr^2] second derivative (wrt all dynamic variables) of dynamic model equations, i.e. d(df/dyy0ex0)/dyy0ex0, rows are in declaration order, columns in lead_lag_incidence order
%g3 is [M.endo_nbr by yy0ex0_nbr^3] third-derivative (wrt all dynamic variables) of dynamic model equations, i.e. (d(df/dyy0ex0)/dyy0ex0)/dyy0ex0, rows are in declaration order, columns in lead_lag_incidence order
elseif order == 3
try
[ ~ , g1p , ~ , ~ , g2p , g3p ] = feval ( [ M . fname , ' .dynamic_params_derivs' ] , oo . dr . ys ( I ) , oo . exo_steady_state ' , M . params , oo . dr . ys , 1 , dys , d2ys ) ;
%g1p is [M.endo_nbr by yy0ex0_nbr by M.param_nbr] first-derivative (wrt all model parameters) of first-derivative (wrt all dynamic variables) of dynamic model equations, i.e. d(df/dyy0ex0)/dparam, rows are in declaration order, column in lead_lag_incidence order
%g2p are nonzero values and corresponding indices of first-derivative (wrt all model parameters) of second-derivatives (wrt all dynamic variables) of dynamic model equations, i.e. d(d(df/dyy0ex0)/dyy0ex0)/dparam, rows are in declaration order, first and second column in declaration order
%g3p are nonzero values and corresponding indices of first-derivative (wrt all model parameters) of third-derivatives (wrt all dynamic variables) of dynamic model equations, i.e. d(d(d(df/dyy0ex0)/dyy0ex0)/dyy0ex0)/dparam, rows are in declaration order, first, second and third column in declaration order
catch
error ( ' For analytical parameter derivatives ' ' dynamic_params_derivs.m' ' file is needed, this can be created by putting identification(order=%d) into your mod file.' , order )
end
T = NaN ( sum ( M . dynamic_tmp_nbr ( 1 : 5 ) ) ) ;
T = feval ( [ M . fname , ' .dynamic_g4_tt' ] , T , oo . dr . ys ( I ) , oo . exo_steady_state ' , M . params , oo . dr . ys , 1 ) ;
g1 = feval ( [ M . fname , ' .dynamic_g1' ] , T , oo . dr . ys ( I ) , oo . exo_steady_state ' , M . params , oo . dr . ys , 1 , false ) ; %g1 is [M.endo_nbr by yy0ex0_nbr first derivative (wrt all dynamic variables) of dynamic model equations, i.e. df/dyy0ex0, rows are in declaration order, columns in lead_lag_incidence order
g2 = feval ( [ M . fname , ' .dynamic_g2' ] , T , oo . dr . ys ( I ) , oo . exo_steady_state ' , M . params , oo . dr . ys , 1 , false ) ; %g2 is [M.endo_nbr by yy0ex0_nbr^2] second derivative (wrt all dynamic variables) of dynamic model equations, i.e. d(df/dyy0ex0)/dyy0ex0, rows are in declaration order, columns in lead_lag_incidence order
g3 = feval ( [ M . fname , ' .dynamic_g3' ] , T , oo . dr . ys ( I ) , oo . exo_steady_state ' , M . params , oo . dr . ys , 1 , false ) ; %note that g3 does not contain symmetric elements
g4 = feval ( [ M . fname , ' .dynamic_g4' ] , T , oo . dr . ys ( I ) , oo . exo_steady_state ' , M . params , oo . dr . ys , 1 , false ) ; %note that g4 does not contain symmetric elements
g3 = unfold_g3 ( g3 , yy0ex0_nbr ) ; %add symmetric elements to g3, %g3 is [M.endo_nbr by yy0ex0_nbr^3] third-derivative (wrt all dynamic variables) of dynamic model equations, i.e. (d(df/dyy0ex0)/dyy0ex0)/dyy0ex0, rows are in declaration order, columns in lead_lag_incidence order
g4 = unfold_g4 ( g4 , yy0ex0_nbr ) ; %add symmetric elements to g4, %g4 is [M.endo_nbr by yy0ex0_nbr^4] fourth-derivative (wrt all dynamic variables) of dynamic model equations, i.e. ((d(df/dyy0ex0)/dyy0ex0)/dyy0ex0)/dyy0ex0, rows are in declaration order, columns in lead_lag_incidence order
end
2019-12-20 16:28:06 +01:00
end
Added parameter derivatives of perturbation solution up to 3 order
# Preliminary comments
I finished the identification toolbox at orders two and three using the pruned state space system, but before I merge request this, I decided to first merge the new functionality to compute parameter derivatives of perturbation solution matrices at higher orders. So after this is approved, I merge the identification toolbox.
I guess @rattoma, @sebastien, and @michel are best choices to review this.
I outline the main idea first and then provide some more detailed changes I made to the functions.
***
# Main idea
This merge request is concerned with the *analytical*computation of the parameter derivatives of first, second and third order perturbation solution matrices, i.e. using _closed-form_ expressions to efficiently compute the derivative of $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ *with respect to model parameters* $\theta$. Note that $\theta$ contains model parameters, stderr and corr parameters of shocks. stderr and corr parameters of measurement errors are not yet supported, (they can easily be included as exogenous shocks). The availability of such derivatives is beneficial in terms of more reliable analysis of model sensitivity and parameter identifiability as well as more efficient estimation methods, in particular for models solved up to third order, as it is well-known that numerical derivatives are a tricky business, especially for large models.
References for my approach are:
* Iskrev (2008, 2010) and Schmitt-Grohé and Uribe (2012, Appendix) who were the first to compute the parameter derivatives analytically at first order, however, using inefficient (sparse) Kronecker products.
* Mutschler (2015) who provides the expressions for a second-order, but again using inefficient (sparse) Kronecker products.
* Ratto and Iskrev (2012) who show how the first-order system can be solved accurately, fast and efficiently using existing numerical algorithms for generalized Sylvester equations by taking the parameter derivative with respect to each parameter separately.
* Julliard and Kamenik (2004) who provide the perturbation solution equation system in tensor notation at any order k.
* Levintal (2017) who introduces permutation matrices to express the perturbation solution equation system in matrix notation up to fifth order.
Note that @rattoma already implemented the parameter derivatives of $g_x$ and $g_u$ analytically (and numerically), and I rely heavily on his work in `get_first_order_solution_params_derivs.m` (previously `getH.m`). My additions are mainly to this function and thus it is renamed to `get_perturbation_params_derivs.m`.
The basic idea of this merge request is to take the second- and third-order perturbation solution systems in Julliard and Kamenik (2004), unfold these into an equivalent matrix representation using permutation matrices as in Levintal (2017). Then extending Ratto and Iskrev (2012) one takes the derivative with respect to each parameter separately and gets a computational problem that is linear, albeit large, as it involves either solving generalized Sylvester equations or taking inverses of highly sparse matrices. I will now briefly summarize the perturbation solution system at third order and the system that results when taking the derivative with respect to parameters.
## Perturbation Solution
The following systems arise at first, second, and third order:
$(ghx): f_{x} z_{x} = f_{y_{-}^*} + f_{y_0} g_{x} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{x}= A g_{x} + f_{y_{-}^*}=0$
$(ghu): f_{z} z_{u} = f_{y_0} g_{u} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{u} + f_{u}= A g_u + f_u = 0$
$(ghxx) : A g_{xx} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{x}\right) + f_{zz} \left( z_{x} \otimes z_{x} \right) = 0$
$(ghxu) : A g_{xu} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{x} \otimes z_{u} \right) = 0$
$(ghuu) : A g_{uu} + B g_{xx} \left(g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{u} \otimes z_{u} \right) = 0$
$(ghs2) : (A+B) g_{\sigma\sigma} + \left( f_{y^{**}_{+}y^{**}_{+}} \left(g^{**}_{u} \otimes g^{**}_{u}\right) + f_{y^{**}_{+}} g^{**}_{uu}\right)vec(\Sigma) = 0$
$(ghxxx) : A g_{xxx} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{x}\right) + f_{y_{+}}g^{**}_{xx} \left(g^{*}_x \otimes g^{*}_{xx}\right)P_{x\_xx} + f_{zz} \left( z_{x} \otimes z_{xx} \right)P_{x\_xx} + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{x} \right) = 0$
$(ghxxu) : A g_{xxu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{u} \right) + f_{zz} \left( \left( z_{x} \otimes z_{xu} \right)P_{x\_xu} + \left(z_{xx} \otimes z_{u}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{x} \otimes g^{*}_{xu}\right)P_{x\_xu} + \left(g^{*}_{xx} \otimes g^{*}_{u}\right) \right) = 0$
$(ghxuu) : A g_{xuu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( \left( z_{xu} \otimes z_{u} \right)P_{xu\_u} + \left(z_{x} \otimes z_{uu}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{xu} \otimes g^{*}_{u}\right)P_{xu\_u} + \left(g^{*}_{x} \otimes g^{*}_{uu}\right) \right) = 0$
$(ghuuu) : A g_{uuu} + B g_{xxx} \left(g^{*}_{u} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{u} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( z_{u} \otimes z_{uu} \right)P_{u\_uu} + f_{y_{+}}g^{**}_{xx} \left(g^{*}_{u} \otimes g^{*}_{uu}\right)P_{u\_uu} = 0$
$(ghx\sigma\sigma) : A g_{x\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_x + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{x} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{x} \otimes z_{\sigma\sigma}\right) + F_{xu_{+}u_{+}}\left(I_{n_x} \otimes vec(\Sigma)\right) = 0$
$F_{xu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_x^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{xu_{+}} \otimes z_{u_{+}} \right)P_{xu\_u} + \left(z_{x} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{x} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
$(ghu\sigma\sigma) : A g_{u\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_{u} + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{u} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{u} \otimes z_{\sigma\sigma}\right) + F_{uu_{+}u_{+}}\left(I_{n_u} \otimes vec(\Sigma_u)\right) = 0$
$F_{uu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_u^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{uu_{+}} \otimes z_{u_{+}} \right)P_{uu\_u} + \left(z_{u} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{u} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
A and B are the common perturbation matrices:
$A = f_{y_0} + \begin{pmatrix} \underbrace{0}_{n\times n_{static}} &\vdots& \underbrace{f_{y^{**}_{+}} \cdot g^{**}_{x}}_{n \times n_{spred}} &\vdots& \underbrace{0}_{n\times n_{frwd}} \end{pmatrix}$and $B = \begin{pmatrix} \underbrace{0}_{n \times n_{static}}&\vdots & \underbrace{0}_{n \times n_{pred}} & \vdots & \underbrace{f_{y^{**}_{+}}}_{n \times n_{sfwrd}} \end{pmatrix}$
and $z=(y_{-}^{\ast}; y; y_{+}^{\ast\ast}; u)$ denotes the dynamic model variables as in `M_.lead_lag_incidence`, $y^\ast$ denote state variables, $y^{\ast\ast}$ denote forward looking variables, $y_+$ denote the variables with a lead, $y_{-}$ denote variables with a lag, $y_0$ denote variables at period t, $f$ the model equations, and $f_z$ the first-order dynamic model derivatives, $f_{zz}$ the second-order dynamic derivatives, and $f_{zzz}$ the third-order dynamic model derivatives. Then:
$z_{x} = \begin{pmatrix}I\\g_{x}\\g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{u} =\begin{pmatrix}0\\g_{u}\\g^{**}_{x} \cdot g^{*}_{u}\\I\end{pmatrix}$, $z_{u_{+}} =\begin{pmatrix}0\\0\\g^{**}_{u}\\0\end{pmatrix}$
$z_{xx} = \begin{pmatrix} 0\\g_{xx}\\g^{**}_{x} \left( g^{*}_x \otimes g^{*}_{x} \right) + g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{xu} =\begin{pmatrix}0\\g_{xu}\\g^{**}_{xx} \left( g^{*}_x \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{xu}\\0\end{pmatrix}$, $z_{uu} =\begin{pmatrix}0\\g_{uu}\\g^{**}_{xx} \left( g^{*}_u \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{uu}\\0\end{pmatrix}$,
$z_{xu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_x \otimes I \right)\\0\end{pmatrix}$, $z_{uu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_{u} \otimes I \right)\\0\end{pmatrix}$, $z_{u_{+}u_{+}} =\begin{pmatrix}0\\0\\g^{\ast\ast}_{uu}\\0\end{pmatrix}$, $z_{\sigma\sigma} = \begin{pmatrix}0\\ g_{\sigma\sigma}\\ g^{\ast\ast}_{x}g^{\ast}_{\sigma\sigma} + g^{\ast\ast}_{\sigma\sigma}\\0 \end{pmatrix}$
$P$ are permutation matrices that can be computed using Matlab's `ipermute` function.
## Parameter derivatives of perturbation solutions
First, we need the parameter derivatives of first, second, third, and fourth derivatives of the dynamic model (i.e. g1,g2,g3,g4 in dynamic files), I make use of the implicit function theorem: Let $f_{z^k}$ denote the kth derivative (wrt all dynamic variables) of the dynamic model, then let $df_{z^k}$ denote the first-derivative (wrt all model parameters) of $f_{z^k}$ evaluated at the steady state. Note that $f_{z^k}$ is a function of both the model parameters $\theta$ and of the steady state of all dynamic variables $\bar{z}$, which also depend on the parameters. Hence, implicitly $f_{z^k}=f_{z^k}(\theta,\bar{z}(\theta))$ and $df_{z^k}$ consists of two parts:
1. direct derivative wrt to all model parameters given by the preprocessor in the `_params_derivs.m` files
2. contribution of derivative of steady state of dynamic variables (wrt all model parameters): $f_{z^{k+1}} \cdot d\bar{z}$
Note that we already have functionality to compute $d\bar{z}$ analytically.
Having this, the above perturbation systems are basically equations of the following types
$AX +BXC = RHS$ or $AX = RHS$
Now when taking the derivative (wrt to one single parameter $\theta_j$), we get
$A\mathrm{d}\{X\} + B\mathrm{d}\{X\}C = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X - \mathrm{d}\{B\}XC - BX\mathrm{d}\{C\}$
or
$A\mathrm{d}\{X\} = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X$
The first one is a Sylvester type equation, the second one can be solved by taking the inverse of $A$. The only diffculty and tedious work arrises in computing (the highly sparse) derivatives of $RHS$.
***
# New functions: `
## get_perturbation_params_derivs.m`and `get_perturbation_params_derivs_numerical_objective.m`
* The parameter derivatives up to third order are computed in the new function`get_perturbation_params_derivs.m` both analytically and numerically. For numerical derivatives `get_perturbation_params_derivs_numerical_objective.m` is the objective for `fjaco.m` or `hessian_sparse.m` or `hessian.m`.
* `get_perturbation_params_derivs.m` is basically an extended version of the previous `get_first_order_solution_params_derivs.m` function.
* * `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It is used for numerical derivatives, whenever `analytic_derivation_mode=-1|-2`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
* Detailed changes:
* Most important: notation of this function is now in accordance to the k_order_solver, i.e. we do not compute derivatives of Kalman transition matrices A and B, but rather the solution matrices ghx,ghu,ghxx,ghxu,ghuu,ghs2,ghxxx,ghxxu,ghxuu,ghuuu,ghxss,ghuss in accordance with notation used in `oo_.dr`. As a byproduct at first-order, focusing on ghx and ghu instead of Kalman transition matrices A and B makes the computations slightly faster for large models (e.g. for Quest the computations were faster by a couple of seconds, not much, but okay).
* Removed use of `kstate`, see also Dynare/dynare#1653 and Dynare/dynare!1656
* Output arguments are stored in a structure `DERIVS`, there is also a flag `d2flag` that computes parameter hessians needed only in `dsge_likelihood.m`.
* Removed `kronflag` as input. `options_.analytic_derivation_mode` is now used instead of `kronflag`.
* Removed `indvar`, an index that was used to selected specific variables in the derivatives. This is not needed, as we always compute the parameter derivatives for all variables first and then select a subset of variables. The selection now takes place in other functions, like `dsge_likelihood.m`.
* Introduced some checks: (i) deterministic exogenous variables are not supported, (ii) Kronecker method only compatible with first-order approximation so reset to sylvester method, (iii) for purely backward or forward models we need to be careful with the rows in `M_.lead_la g_incidence`, (iv) if `_params_derivs.m` files are missing an error is thrown.
* For numerical derivatives, if mod file does not contain an `estimated_params_block`, a temporary one with the most important parameter information is created.
## `unfold_g4.m`
* When evaluating g3 and g4 one needs to take into account that these do not contain symmetric elements, so one needs to use `unfold_g3.m` and the new function `unfold_g4.m`. This returns an unfolded version of the same matrix (i.e. with symmetric elements).
***
# New test models
`.gitignore` and `Makefile.am` are changed accordingly. Also now it is possible to run test suite on analytic_derivatives, i.e. run `make check m/analytic_derivatives`
## `analytic_derivatives/BrockMirman_PertParamsDerivs.mod`
* This is the Brock Mirman model, where we know the exact policy function $g$ for capital and consumption. As this does not imply a nonzero $g_{\sigma\sigma}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ I added some artificial equations to get nonzero solution matrices with respect to $\sigma$. The true perturbation solution matrices $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ are then computed analytically with Matlab's symbolic toolbox and saved in `nBrockMirmanSYM.mat`. There is a preprocessor flag that recreates these analytical computations if changes are needed (and to check whether I made some errors here ;-) )
* Then solution matrices up to third order and their parameter Jacobians are then compared to the ones computed by Dynare's `k_order_solver` and by `get_perturbation_params_derivs` for all `analytic_derivation_mode`'s. There will be an error if the maximum absolute deviation is too large, i.e. for numerical derivatives (`analytic_derivation_mode=-1|-2`) the tolerance is choosen lower (around 1e-5); for analytical methods we are stricter: around 1e-13 for first-order, 1e-12 for second order, and 1e-11 for third-order.
* As a side note, this mod file also checks Dynare's `k_order_solver` algorithm and throws an error if something is wrong.
* This test model shows that the new functionality works well. And analytical derivatives perform way better and accurate than numerical ones, even for this small model.
## `analytic_derivatives/burnside_3_order_PertParamsDerivs.mod`
* This builds upon `tests/k_order_perturbation/burnside_k_order.mod` and computes the true parameter derivatives analytically by hand.
* This test model also shows that the new functionality works well.
## `analytic_derivatives/LindeTrabandt2019.mod`
* Shows that the new functionality also works for medium-sized models, i.e. a SW type model solved at third order with 35 variables (11 states). 2 shocks and 20 parameters.
* This mod file can be used to tweak the speed of the computations in the future.
* Compares numerical versus analytical parameter derivatives (for first, second and third order). Note that this model clearly shows that numerical ones are quite different than analytical ones even at first order!
## `identification/LindeTrabandt2019_xfail.mod`
* This model is a check for issue Dynare/dynare#1595, see fjaco.m below, and will fail.
* Removed `analytic_derivatives/ls2003.mod` as this mod file is neither in the testsuite nor does it work.
***
# Detailed changes in other functions
## `get_first_order_solution_params_derivs.m`
* Deleted, or actually, renamed to `get_perturbation_params_derivs.m`, as this function now is able to compute the derivatives up to third order
## `identification_numerical_objective.m`
* `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
## `dsge_likelihood.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `T`matrix anymore, but instead of the dynare solution matrix `ghx`. So we recreate `T` here as this amounts to adding some zeros and focusing on selected variables only.
* Added some checks to make sure the first-order approximation is selected.
* Removed `kron_flag` as input, as `get_perturbation_params_derivs` looks into `options_.analytic_derivation_mode` for `kron_flag`.
## `dynare_identification.m`
* make sure that setting `analytic_derivation_mode` is set both in `options_ident` and `options_`. Note that at the end of the function we restore the `options_` structure, so all changes are local. In a next merge request, I will remove the global variables to make all variables local.
## `get_identification_jacobians.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `A` and `B` matrix anymore, but instead of the dynare solution matrix `ghx` and `ghu`. So we recreate these matrices here instead of in `get_perturbation_params_derivs.m`.
* Added `str2func` for better function handles in `fjaco.m`.
## `fjaco.m`
* make `tol`an option, which can be adjusted by changing `options_.dynatol.x`for identification and parameter derivatives purposes.
* include a check and an informative error message, if numerical derivatives (two-sided finite difference method) yield errors in `resol.m` for identification and parameter derivatives purposes. This closes issue Dynare/dynare#1595.
* Changed year of copyright to 2010-2017,2019
***
# Further suggestions and questions
* Ones this is merged, I will merge request an improvement of the identification toolbox, which will work up to third order using the pruned state space. This will also remove some issues and bugs, and also I will remove global variables in this request.
* The third-order derivatives can be further improved by taking sparsity into account and use mex versions for kronecker products etc. I leave this for further testing (and if anybody actually uses this ;-) )
2019-12-17 19:17:09 +01:00
% Parameter Jacobian of steady state in different orderings, note dys is in declaration order
dYss = dys ( oo . dr . order_var , indpmodel ) ; %in DR-order, focus only on selected model parameters
dyy0 = dys ( I , : ) ; %in lead_lag_incidence order, Jacobian of dynamic (without exogenous) variables, focus on all model parameters
dyy0ex0 = sparse ( [ dyy0 ; zeros ( M . exo_nbr , M . param_nbr ) ] ) ; %in lead_lag_incidence order, Jacobian of dynamic (with exogenous) variables, focus on all model parameters
%% Analytical computation of Jacobian (wrt selected model parameters) of first-derivative of dynamic model, i.e. dg1
% Note that we use the implicit function theorem (see Ratto and Iskrev (2012) formula 7):
% Let g1 = df/dyy0ex0 be the first derivative (wrt all dynamic variables) of the dynamic model, then
% dg1 denotes the first-derivative (wrt model parameters) of g1 evaluated at the steady state.
% Note that g1 is a function of both the model parameters p and of the steady state of all dynamic variables, which also depend on the model parameters.
% Hence, implicitly g1=g1(p,yy0ex0(p)) and dg1 consists of two parts:
% (1) g1p: direct derivative (wrt to all model parameters) given by the preprocessor
% (2) g2*dyy0ex0: contribution of derivative of steady state of dynamic variables (wrt all model parameters)
% Note that in a stochastic context ex0 and dex0 is always zero and hence can be skipped in the computations
dg1 = zeros ( M . endo_nbr , yy0ex0_nbr , M . param_nbr ) ; %initialize part (2)
for j = 1 : M . endo_nbr
[ II , JJ ] = ind2sub ( [ yy0ex0_nbr yy0ex0_nbr ] , find ( g2 ( j , : ) ) ) ; %g2 is [M.endo_nbr by yy0ex0_nbr^2]
for i = 1 : yy0ex0_nbr
is = find ( II == i ) ;
is = is ( find ( JJ ( is ) < = yy0_nbr ) ) ; %focus only on oo.dr.ys(I) derivatives as exogenous variables are 0 in a stochastic context
if ~ isempty ( is )
tmp_g2 = full ( g2 ( j , find ( g2 ( j , : ) ) ) ) ;
dg1 ( j , i , : ) = tmp_g2 ( is ) * dyy0 ( JJ ( is ) , : ) ; %put into tensor notation
end
end
end
dg1 = g1p + dg1 ; %add part (1)
dg1 = dg1 ( : , : , indpmodel ) ; %focus only on selected model parameters
if order > 1
%% Analytical computation of Jacobian (wrt selected model parameters) of second derivative of dynamic model, i.e. dg2
% We use the implicit function theorem:
% Let g2 = d2f/(dyy0ex0*dyy0ex0) denote the second derivative (wrt all dynamic variables) of the dynamic model, then
% dg2 denotes the first-derivative (wrt all model parameters) of g2 evaluated at the steady state.
% Note that g2 is a function of both the model parameters p and of the steady state of all dynamic variables, which also depend on the parameters.
% Hence, implicitly g2=g2(p,yy0ex0(p)) and dg2 consists of two parts:
% (1) g2p: direct derivative wrt to all model parameters given by the preprocessor
% and
% (2) g3*dyy0ex0: contribution of derivative of steady state of dynamic variables (wrt all model parameters)
% Note that we focus on selected model parameters only, i.e. indpmodel
% Also note that we stack the parameter derivatives blockwise instead of in tensors
dg2 = reshape ( g3 , [ M . endo_nbr * yy0ex0_nbr ^2 yy0ex0_nbr ] ) * dyy0ex0 ( : , indpmodel ) ; %part (2)
dg2 = reshape ( dg2 , [ M . endo_nbr , yy0ex0_nbr ^2 * modparam_nbr ] ) ;
k2yy0ex0 = transpose ( reshape ( 1 : yy0ex0_nbr ^2 , yy0ex0_nbr , yy0ex0_nbr ) ) ; %index for the second dynamic derivatives, i.e. to evaluate the derivative of f wrt to yy0ex0(i) and yy0ex0(j), in DR order
for jj = 1 : size ( g2p , 1 )
jpos = find ( indpmodel == g2p ( jj , 4 ) ) ;
if jpos ~= 0
dg2 ( g2p ( jj , 1 ) , k2yy0ex0 ( g2p ( jj , 2 ) , g2p ( jj , 3 ) ) + ( jpos - 1 ) * yy0ex0_nbr ^2 ) = dg2 ( g2p ( jj , 1 ) , k2yy0ex0 ( g2p ( jj , 2 ) , g2p ( jj , 3 ) ) + ( jpos - 1 ) * yy0ex0_nbr ^2 ) + g2p ( jj , 5 ) ; %add part (1)
end
2019-12-20 16:28:06 +01:00
end
Added parameter derivatives of perturbation solution up to 3 order
# Preliminary comments
I finished the identification toolbox at orders two and three using the pruned state space system, but before I merge request this, I decided to first merge the new functionality to compute parameter derivatives of perturbation solution matrices at higher orders. So after this is approved, I merge the identification toolbox.
I guess @rattoma, @sebastien, and @michel are best choices to review this.
I outline the main idea first and then provide some more detailed changes I made to the functions.
***
# Main idea
This merge request is concerned with the *analytical*computation of the parameter derivatives of first, second and third order perturbation solution matrices, i.e. using _closed-form_ expressions to efficiently compute the derivative of $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ *with respect to model parameters* $\theta$. Note that $\theta$ contains model parameters, stderr and corr parameters of shocks. stderr and corr parameters of measurement errors are not yet supported, (they can easily be included as exogenous shocks). The availability of such derivatives is beneficial in terms of more reliable analysis of model sensitivity and parameter identifiability as well as more efficient estimation methods, in particular for models solved up to third order, as it is well-known that numerical derivatives are a tricky business, especially for large models.
References for my approach are:
* Iskrev (2008, 2010) and Schmitt-Grohé and Uribe (2012, Appendix) who were the first to compute the parameter derivatives analytically at first order, however, using inefficient (sparse) Kronecker products.
* Mutschler (2015) who provides the expressions for a second-order, but again using inefficient (sparse) Kronecker products.
* Ratto and Iskrev (2012) who show how the first-order system can be solved accurately, fast and efficiently using existing numerical algorithms for generalized Sylvester equations by taking the parameter derivative with respect to each parameter separately.
* Julliard and Kamenik (2004) who provide the perturbation solution equation system in tensor notation at any order k.
* Levintal (2017) who introduces permutation matrices to express the perturbation solution equation system in matrix notation up to fifth order.
Note that @rattoma already implemented the parameter derivatives of $g_x$ and $g_u$ analytically (and numerically), and I rely heavily on his work in `get_first_order_solution_params_derivs.m` (previously `getH.m`). My additions are mainly to this function and thus it is renamed to `get_perturbation_params_derivs.m`.
The basic idea of this merge request is to take the second- and third-order perturbation solution systems in Julliard and Kamenik (2004), unfold these into an equivalent matrix representation using permutation matrices as in Levintal (2017). Then extending Ratto and Iskrev (2012) one takes the derivative with respect to each parameter separately and gets a computational problem that is linear, albeit large, as it involves either solving generalized Sylvester equations or taking inverses of highly sparse matrices. I will now briefly summarize the perturbation solution system at third order and the system that results when taking the derivative with respect to parameters.
## Perturbation Solution
The following systems arise at first, second, and third order:
$(ghx): f_{x} z_{x} = f_{y_{-}^*} + f_{y_0} g_{x} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{x}= A g_{x} + f_{y_{-}^*}=0$
$(ghu): f_{z} z_{u} = f_{y_0} g_{u} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{u} + f_{u}= A g_u + f_u = 0$
$(ghxx) : A g_{xx} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{x}\right) + f_{zz} \left( z_{x} \otimes z_{x} \right) = 0$
$(ghxu) : A g_{xu} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{x} \otimes z_{u} \right) = 0$
$(ghuu) : A g_{uu} + B g_{xx} \left(g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{u} \otimes z_{u} \right) = 0$
$(ghs2) : (A+B) g_{\sigma\sigma} + \left( f_{y^{**}_{+}y^{**}_{+}} \left(g^{**}_{u} \otimes g^{**}_{u}\right) + f_{y^{**}_{+}} g^{**}_{uu}\right)vec(\Sigma) = 0$
$(ghxxx) : A g_{xxx} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{x}\right) + f_{y_{+}}g^{**}_{xx} \left(g^{*}_x \otimes g^{*}_{xx}\right)P_{x\_xx} + f_{zz} \left( z_{x} \otimes z_{xx} \right)P_{x\_xx} + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{x} \right) = 0$
$(ghxxu) : A g_{xxu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{u} \right) + f_{zz} \left( \left( z_{x} \otimes z_{xu} \right)P_{x\_xu} + \left(z_{xx} \otimes z_{u}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{x} \otimes g^{*}_{xu}\right)P_{x\_xu} + \left(g^{*}_{xx} \otimes g^{*}_{u}\right) \right) = 0$
$(ghxuu) : A g_{xuu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( \left( z_{xu} \otimes z_{u} \right)P_{xu\_u} + \left(z_{x} \otimes z_{uu}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{xu} \otimes g^{*}_{u}\right)P_{xu\_u} + \left(g^{*}_{x} \otimes g^{*}_{uu}\right) \right) = 0$
$(ghuuu) : A g_{uuu} + B g_{xxx} \left(g^{*}_{u} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{u} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( z_{u} \otimes z_{uu} \right)P_{u\_uu} + f_{y_{+}}g^{**}_{xx} \left(g^{*}_{u} \otimes g^{*}_{uu}\right)P_{u\_uu} = 0$
$(ghx\sigma\sigma) : A g_{x\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_x + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{x} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{x} \otimes z_{\sigma\sigma}\right) + F_{xu_{+}u_{+}}\left(I_{n_x} \otimes vec(\Sigma)\right) = 0$
$F_{xu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_x^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{xu_{+}} \otimes z_{u_{+}} \right)P_{xu\_u} + \left(z_{x} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{x} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
$(ghu\sigma\sigma) : A g_{u\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_{u} + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{u} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{u} \otimes z_{\sigma\sigma}\right) + F_{uu_{+}u_{+}}\left(I_{n_u} \otimes vec(\Sigma_u)\right) = 0$
$F_{uu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_u^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{uu_{+}} \otimes z_{u_{+}} \right)P_{uu\_u} + \left(z_{u} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{u} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
A and B are the common perturbation matrices:
$A = f_{y_0} + \begin{pmatrix} \underbrace{0}_{n\times n_{static}} &\vdots& \underbrace{f_{y^{**}_{+}} \cdot g^{**}_{x}}_{n \times n_{spred}} &\vdots& \underbrace{0}_{n\times n_{frwd}} \end{pmatrix}$and $B = \begin{pmatrix} \underbrace{0}_{n \times n_{static}}&\vdots & \underbrace{0}_{n \times n_{pred}} & \vdots & \underbrace{f_{y^{**}_{+}}}_{n \times n_{sfwrd}} \end{pmatrix}$
and $z=(y_{-}^{\ast}; y; y_{+}^{\ast\ast}; u)$ denotes the dynamic model variables as in `M_.lead_lag_incidence`, $y^\ast$ denote state variables, $y^{\ast\ast}$ denote forward looking variables, $y_+$ denote the variables with a lead, $y_{-}$ denote variables with a lag, $y_0$ denote variables at period t, $f$ the model equations, and $f_z$ the first-order dynamic model derivatives, $f_{zz}$ the second-order dynamic derivatives, and $f_{zzz}$ the third-order dynamic model derivatives. Then:
$z_{x} = \begin{pmatrix}I\\g_{x}\\g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{u} =\begin{pmatrix}0\\g_{u}\\g^{**}_{x} \cdot g^{*}_{u}\\I\end{pmatrix}$, $z_{u_{+}} =\begin{pmatrix}0\\0\\g^{**}_{u}\\0\end{pmatrix}$
$z_{xx} = \begin{pmatrix} 0\\g_{xx}\\g^{**}_{x} \left( g^{*}_x \otimes g^{*}_{x} \right) + g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{xu} =\begin{pmatrix}0\\g_{xu}\\g^{**}_{xx} \left( g^{*}_x \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{xu}\\0\end{pmatrix}$, $z_{uu} =\begin{pmatrix}0\\g_{uu}\\g^{**}_{xx} \left( g^{*}_u \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{uu}\\0\end{pmatrix}$,
$z_{xu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_x \otimes I \right)\\0\end{pmatrix}$, $z_{uu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_{u} \otimes I \right)\\0\end{pmatrix}$, $z_{u_{+}u_{+}} =\begin{pmatrix}0\\0\\g^{\ast\ast}_{uu}\\0\end{pmatrix}$, $z_{\sigma\sigma} = \begin{pmatrix}0\\ g_{\sigma\sigma}\\ g^{\ast\ast}_{x}g^{\ast}_{\sigma\sigma} + g^{\ast\ast}_{\sigma\sigma}\\0 \end{pmatrix}$
$P$ are permutation matrices that can be computed using Matlab's `ipermute` function.
## Parameter derivatives of perturbation solutions
First, we need the parameter derivatives of first, second, third, and fourth derivatives of the dynamic model (i.e. g1,g2,g3,g4 in dynamic files), I make use of the implicit function theorem: Let $f_{z^k}$ denote the kth derivative (wrt all dynamic variables) of the dynamic model, then let $df_{z^k}$ denote the first-derivative (wrt all model parameters) of $f_{z^k}$ evaluated at the steady state. Note that $f_{z^k}$ is a function of both the model parameters $\theta$ and of the steady state of all dynamic variables $\bar{z}$, which also depend on the parameters. Hence, implicitly $f_{z^k}=f_{z^k}(\theta,\bar{z}(\theta))$ and $df_{z^k}$ consists of two parts:
1. direct derivative wrt to all model parameters given by the preprocessor in the `_params_derivs.m` files
2. contribution of derivative of steady state of dynamic variables (wrt all model parameters): $f_{z^{k+1}} \cdot d\bar{z}$
Note that we already have functionality to compute $d\bar{z}$ analytically.
Having this, the above perturbation systems are basically equations of the following types
$AX +BXC = RHS$ or $AX = RHS$
Now when taking the derivative (wrt to one single parameter $\theta_j$), we get
$A\mathrm{d}\{X\} + B\mathrm{d}\{X\}C = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X - \mathrm{d}\{B\}XC - BX\mathrm{d}\{C\}$
or
$A\mathrm{d}\{X\} = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X$
The first one is a Sylvester type equation, the second one can be solved by taking the inverse of $A$. The only diffculty and tedious work arrises in computing (the highly sparse) derivatives of $RHS$.
***
# New functions: `
## get_perturbation_params_derivs.m`and `get_perturbation_params_derivs_numerical_objective.m`
* The parameter derivatives up to third order are computed in the new function`get_perturbation_params_derivs.m` both analytically and numerically. For numerical derivatives `get_perturbation_params_derivs_numerical_objective.m` is the objective for `fjaco.m` or `hessian_sparse.m` or `hessian.m`.
* `get_perturbation_params_derivs.m` is basically an extended version of the previous `get_first_order_solution_params_derivs.m` function.
* * `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It is used for numerical derivatives, whenever `analytic_derivation_mode=-1|-2`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
* Detailed changes:
* Most important: notation of this function is now in accordance to the k_order_solver, i.e. we do not compute derivatives of Kalman transition matrices A and B, but rather the solution matrices ghx,ghu,ghxx,ghxu,ghuu,ghs2,ghxxx,ghxxu,ghxuu,ghuuu,ghxss,ghuss in accordance with notation used in `oo_.dr`. As a byproduct at first-order, focusing on ghx and ghu instead of Kalman transition matrices A and B makes the computations slightly faster for large models (e.g. for Quest the computations were faster by a couple of seconds, not much, but okay).
* Removed use of `kstate`, see also Dynare/dynare#1653 and Dynare/dynare!1656
* Output arguments are stored in a structure `DERIVS`, there is also a flag `d2flag` that computes parameter hessians needed only in `dsge_likelihood.m`.
* Removed `kronflag` as input. `options_.analytic_derivation_mode` is now used instead of `kronflag`.
* Removed `indvar`, an index that was used to selected specific variables in the derivatives. This is not needed, as we always compute the parameter derivatives for all variables first and then select a subset of variables. The selection now takes place in other functions, like `dsge_likelihood.m`.
* Introduced some checks: (i) deterministic exogenous variables are not supported, (ii) Kronecker method only compatible with first-order approximation so reset to sylvester method, (iii) for purely backward or forward models we need to be careful with the rows in `M_.lead_la g_incidence`, (iv) if `_params_derivs.m` files are missing an error is thrown.
* For numerical derivatives, if mod file does not contain an `estimated_params_block`, a temporary one with the most important parameter information is created.
## `unfold_g4.m`
* When evaluating g3 and g4 one needs to take into account that these do not contain symmetric elements, so one needs to use `unfold_g3.m` and the new function `unfold_g4.m`. This returns an unfolded version of the same matrix (i.e. with symmetric elements).
***
# New test models
`.gitignore` and `Makefile.am` are changed accordingly. Also now it is possible to run test suite on analytic_derivatives, i.e. run `make check m/analytic_derivatives`
## `analytic_derivatives/BrockMirman_PertParamsDerivs.mod`
* This is the Brock Mirman model, where we know the exact policy function $g$ for capital and consumption. As this does not imply a nonzero $g_{\sigma\sigma}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ I added some artificial equations to get nonzero solution matrices with respect to $\sigma$. The true perturbation solution matrices $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ are then computed analytically with Matlab's symbolic toolbox and saved in `nBrockMirmanSYM.mat`. There is a preprocessor flag that recreates these analytical computations if changes are needed (and to check whether I made some errors here ;-) )
* Then solution matrices up to third order and their parameter Jacobians are then compared to the ones computed by Dynare's `k_order_solver` and by `get_perturbation_params_derivs` for all `analytic_derivation_mode`'s. There will be an error if the maximum absolute deviation is too large, i.e. for numerical derivatives (`analytic_derivation_mode=-1|-2`) the tolerance is choosen lower (around 1e-5); for analytical methods we are stricter: around 1e-13 for first-order, 1e-12 for second order, and 1e-11 for third-order.
* As a side note, this mod file also checks Dynare's `k_order_solver` algorithm and throws an error if something is wrong.
* This test model shows that the new functionality works well. And analytical derivatives perform way better and accurate than numerical ones, even for this small model.
## `analytic_derivatives/burnside_3_order_PertParamsDerivs.mod`
* This builds upon `tests/k_order_perturbation/burnside_k_order.mod` and computes the true parameter derivatives analytically by hand.
* This test model also shows that the new functionality works well.
## `analytic_derivatives/LindeTrabandt2019.mod`
* Shows that the new functionality also works for medium-sized models, i.e. a SW type model solved at third order with 35 variables (11 states). 2 shocks and 20 parameters.
* This mod file can be used to tweak the speed of the computations in the future.
* Compares numerical versus analytical parameter derivatives (for first, second and third order). Note that this model clearly shows that numerical ones are quite different than analytical ones even at first order!
## `identification/LindeTrabandt2019_xfail.mod`
* This model is a check for issue Dynare/dynare#1595, see fjaco.m below, and will fail.
* Removed `analytic_derivatives/ls2003.mod` as this mod file is neither in the testsuite nor does it work.
***
# Detailed changes in other functions
## `get_first_order_solution_params_derivs.m`
* Deleted, or actually, renamed to `get_perturbation_params_derivs.m`, as this function now is able to compute the derivatives up to third order
## `identification_numerical_objective.m`
* `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
## `dsge_likelihood.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `T`matrix anymore, but instead of the dynare solution matrix `ghx`. So we recreate `T` here as this amounts to adding some zeros and focusing on selected variables only.
* Added some checks to make sure the first-order approximation is selected.
* Removed `kron_flag` as input, as `get_perturbation_params_derivs` looks into `options_.analytic_derivation_mode` for `kron_flag`.
## `dynare_identification.m`
* make sure that setting `analytic_derivation_mode` is set both in `options_ident` and `options_`. Note that at the end of the function we restore the `options_` structure, so all changes are local. In a next merge request, I will remove the global variables to make all variables local.
## `get_identification_jacobians.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `A` and `B` matrix anymore, but instead of the dynare solution matrix `ghx` and `ghu`. So we recreate these matrices here instead of in `get_perturbation_params_derivs.m`.
* Added `str2func` for better function handles in `fjaco.m`.
## `fjaco.m`
* make `tol`an option, which can be adjusted by changing `options_.dynatol.x`for identification and parameter derivatives purposes.
* include a check and an informative error message, if numerical derivatives (two-sided finite difference method) yield errors in `resol.m` for identification and parameter derivatives purposes. This closes issue Dynare/dynare#1595.
* Changed year of copyright to 2010-2017,2019
***
# Further suggestions and questions
* Ones this is merged, I will merge request an improvement of the identification toolbox, which will work up to third order using the pruned state space. This will also remove some issues and bugs, and also I will remove global variables in this request.
* The third-order derivatives can be further improved by taking sparsity into account and use mex versions for kronecker products etc. I leave this for further testing (and if anybody actually uses this ;-) )
2019-12-17 19:17:09 +01:00
end
if order > 2
%% Analytical computation of Jacobian (wrt selected model parameters) of third derivative of dynamic model, i.e. dg3
% We use the implicit function theorem:
% Let g3 = d3f/(dyy0ex0*dyy0ex0*dyy0ex0) denote the third derivative (wrt all dynamic variables) of the dynamic model, then
% dg3 denotes the first-derivative (wrt all model parameters) of g3 evaluated at the steady state.
% Note that g3 is a function of both the model parameters p and of the steady state of all dynamic variables, which also depend on the parameters.
% Hence, implicitly g3=g3(p,yy0ex0(p)) and dg3 consists of two parts:
% (1) g3p: direct derivative wrt to all model parameters given by the preprocessor
% and
% (2) g4*dyy0ex0: contribution of derivative of steady state of dynamic variables (wrt all model parameters)
% Note that we focus on selected model parameters only, i.e. indpmodel
% Also note that we stack the parameter derivatives blockwise instead of in tensors
dg3 = reshape ( g4 , [ M . endo_nbr * yy0ex0_nbr ^3 yy0ex0_nbr ] ) * dyy0ex0 ( : , indpmodel ) ; %part (2)
dg3 = reshape ( dg3 , [ M . endo_nbr , yy0ex0_nbr ^3 * modparam_nbr ] ) ;
k3yy0ex0 = permute ( reshape ( transpose ( reshape ( 1 : yy0ex0_nbr ^3 , yy0ex0_nbr , yy0ex0_nbr ^2 ) ) , yy0ex0_nbr , yy0ex0_nbr , yy0ex0_nbr ) , [ 2 1 3 ] ) ; %index for the third dynamic derivative, i.e. df/(dyyex0_i*dyyex0_j*dyyex0_k)
for jj = 1 : size ( g3p , 1 )
jpos = find ( indpmodel == g3p ( jj , 5 ) ) ;
if jpos ~= 0
idyyy = unique ( perms ( [ g3p ( jj , 2 ) g3p ( jj , 3 ) g3p ( jj , 4 ) ] ) , ' rows' ) ; %note that g3p does not contain symmetric terms, so we use the perms and unique functions
for k = 1 : size ( idyyy , 1 )
dg3 ( g3p ( jj , 1 ) , k3yy0ex0 ( idyyy ( k , 1 ) , idyyy ( k , 2 ) , idyyy ( k , 3 ) ) + ( jpos - 1 ) * yy0ex0_nbr ^3 ) = dg3 ( g3p ( jj , 1 ) , k3yy0ex0 ( idyyy ( k , 1 ) , idyyy ( k , 2 ) , idyyy ( k , 3 ) ) + ( jpos - 1 ) * yy0ex0_nbr ^3 ) + g3p ( jj , 6 ) ; %add part (1)
end
end
2019-12-20 16:28:06 +01:00
end
Added parameter derivatives of perturbation solution up to 3 order
# Preliminary comments
I finished the identification toolbox at orders two and three using the pruned state space system, but before I merge request this, I decided to first merge the new functionality to compute parameter derivatives of perturbation solution matrices at higher orders. So after this is approved, I merge the identification toolbox.
I guess @rattoma, @sebastien, and @michel are best choices to review this.
I outline the main idea first and then provide some more detailed changes I made to the functions.
***
# Main idea
This merge request is concerned with the *analytical*computation of the parameter derivatives of first, second and third order perturbation solution matrices, i.e. using _closed-form_ expressions to efficiently compute the derivative of $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ *with respect to model parameters* $\theta$. Note that $\theta$ contains model parameters, stderr and corr parameters of shocks. stderr and corr parameters of measurement errors are not yet supported, (they can easily be included as exogenous shocks). The availability of such derivatives is beneficial in terms of more reliable analysis of model sensitivity and parameter identifiability as well as more efficient estimation methods, in particular for models solved up to third order, as it is well-known that numerical derivatives are a tricky business, especially for large models.
References for my approach are:
* Iskrev (2008, 2010) and Schmitt-Grohé and Uribe (2012, Appendix) who were the first to compute the parameter derivatives analytically at first order, however, using inefficient (sparse) Kronecker products.
* Mutschler (2015) who provides the expressions for a second-order, but again using inefficient (sparse) Kronecker products.
* Ratto and Iskrev (2012) who show how the first-order system can be solved accurately, fast and efficiently using existing numerical algorithms for generalized Sylvester equations by taking the parameter derivative with respect to each parameter separately.
* Julliard and Kamenik (2004) who provide the perturbation solution equation system in tensor notation at any order k.
* Levintal (2017) who introduces permutation matrices to express the perturbation solution equation system in matrix notation up to fifth order.
Note that @rattoma already implemented the parameter derivatives of $g_x$ and $g_u$ analytically (and numerically), and I rely heavily on his work in `get_first_order_solution_params_derivs.m` (previously `getH.m`). My additions are mainly to this function and thus it is renamed to `get_perturbation_params_derivs.m`.
The basic idea of this merge request is to take the second- and third-order perturbation solution systems in Julliard and Kamenik (2004), unfold these into an equivalent matrix representation using permutation matrices as in Levintal (2017). Then extending Ratto and Iskrev (2012) one takes the derivative with respect to each parameter separately and gets a computational problem that is linear, albeit large, as it involves either solving generalized Sylvester equations or taking inverses of highly sparse matrices. I will now briefly summarize the perturbation solution system at third order and the system that results when taking the derivative with respect to parameters.
## Perturbation Solution
The following systems arise at first, second, and third order:
$(ghx): f_{x} z_{x} = f_{y_{-}^*} + f_{y_0} g_{x} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{x}= A g_{x} + f_{y_{-}^*}=0$
$(ghu): f_{z} z_{u} = f_{y_0} g_{u} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{u} + f_{u}= A g_u + f_u = 0$
$(ghxx) : A g_{xx} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{x}\right) + f_{zz} \left( z_{x} \otimes z_{x} \right) = 0$
$(ghxu) : A g_{xu} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{x} \otimes z_{u} \right) = 0$
$(ghuu) : A g_{uu} + B g_{xx} \left(g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{u} \otimes z_{u} \right) = 0$
$(ghs2) : (A+B) g_{\sigma\sigma} + \left( f_{y^{**}_{+}y^{**}_{+}} \left(g^{**}_{u} \otimes g^{**}_{u}\right) + f_{y^{**}_{+}} g^{**}_{uu}\right)vec(\Sigma) = 0$
$(ghxxx) : A g_{xxx} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{x}\right) + f_{y_{+}}g^{**}_{xx} \left(g^{*}_x \otimes g^{*}_{xx}\right)P_{x\_xx} + f_{zz} \left( z_{x} \otimes z_{xx} \right)P_{x\_xx} + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{x} \right) = 0$
$(ghxxu) : A g_{xxu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{u} \right) + f_{zz} \left( \left( z_{x} \otimes z_{xu} \right)P_{x\_xu} + \left(z_{xx} \otimes z_{u}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{x} \otimes g^{*}_{xu}\right)P_{x\_xu} + \left(g^{*}_{xx} \otimes g^{*}_{u}\right) \right) = 0$
$(ghxuu) : A g_{xuu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( \left( z_{xu} \otimes z_{u} \right)P_{xu\_u} + \left(z_{x} \otimes z_{uu}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{xu} \otimes g^{*}_{u}\right)P_{xu\_u} + \left(g^{*}_{x} \otimes g^{*}_{uu}\right) \right) = 0$
$(ghuuu) : A g_{uuu} + B g_{xxx} \left(g^{*}_{u} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{u} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( z_{u} \otimes z_{uu} \right)P_{u\_uu} + f_{y_{+}}g^{**}_{xx} \left(g^{*}_{u} \otimes g^{*}_{uu}\right)P_{u\_uu} = 0$
$(ghx\sigma\sigma) : A g_{x\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_x + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{x} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{x} \otimes z_{\sigma\sigma}\right) + F_{xu_{+}u_{+}}\left(I_{n_x} \otimes vec(\Sigma)\right) = 0$
$F_{xu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_x^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{xu_{+}} \otimes z_{u_{+}} \right)P_{xu\_u} + \left(z_{x} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{x} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
$(ghu\sigma\sigma) : A g_{u\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_{u} + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{u} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{u} \otimes z_{\sigma\sigma}\right) + F_{uu_{+}u_{+}}\left(I_{n_u} \otimes vec(\Sigma_u)\right) = 0$
$F_{uu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_u^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{uu_{+}} \otimes z_{u_{+}} \right)P_{uu\_u} + \left(z_{u} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{u} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
A and B are the common perturbation matrices:
$A = f_{y_0} + \begin{pmatrix} \underbrace{0}_{n\times n_{static}} &\vdots& \underbrace{f_{y^{**}_{+}} \cdot g^{**}_{x}}_{n \times n_{spred}} &\vdots& \underbrace{0}_{n\times n_{frwd}} \end{pmatrix}$and $B = \begin{pmatrix} \underbrace{0}_{n \times n_{static}}&\vdots & \underbrace{0}_{n \times n_{pred}} & \vdots & \underbrace{f_{y^{**}_{+}}}_{n \times n_{sfwrd}} \end{pmatrix}$
and $z=(y_{-}^{\ast}; y; y_{+}^{\ast\ast}; u)$ denotes the dynamic model variables as in `M_.lead_lag_incidence`, $y^\ast$ denote state variables, $y^{\ast\ast}$ denote forward looking variables, $y_+$ denote the variables with a lead, $y_{-}$ denote variables with a lag, $y_0$ denote variables at period t, $f$ the model equations, and $f_z$ the first-order dynamic model derivatives, $f_{zz}$ the second-order dynamic derivatives, and $f_{zzz}$ the third-order dynamic model derivatives. Then:
$z_{x} = \begin{pmatrix}I\\g_{x}\\g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{u} =\begin{pmatrix}0\\g_{u}\\g^{**}_{x} \cdot g^{*}_{u}\\I\end{pmatrix}$, $z_{u_{+}} =\begin{pmatrix}0\\0\\g^{**}_{u}\\0\end{pmatrix}$
$z_{xx} = \begin{pmatrix} 0\\g_{xx}\\g^{**}_{x} \left( g^{*}_x \otimes g^{*}_{x} \right) + g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{xu} =\begin{pmatrix}0\\g_{xu}\\g^{**}_{xx} \left( g^{*}_x \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{xu}\\0\end{pmatrix}$, $z_{uu} =\begin{pmatrix}0\\g_{uu}\\g^{**}_{xx} \left( g^{*}_u \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{uu}\\0\end{pmatrix}$,
$z_{xu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_x \otimes I \right)\\0\end{pmatrix}$, $z_{uu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_{u} \otimes I \right)\\0\end{pmatrix}$, $z_{u_{+}u_{+}} =\begin{pmatrix}0\\0\\g^{\ast\ast}_{uu}\\0\end{pmatrix}$, $z_{\sigma\sigma} = \begin{pmatrix}0\\ g_{\sigma\sigma}\\ g^{\ast\ast}_{x}g^{\ast}_{\sigma\sigma} + g^{\ast\ast}_{\sigma\sigma}\\0 \end{pmatrix}$
$P$ are permutation matrices that can be computed using Matlab's `ipermute` function.
## Parameter derivatives of perturbation solutions
First, we need the parameter derivatives of first, second, third, and fourth derivatives of the dynamic model (i.e. g1,g2,g3,g4 in dynamic files), I make use of the implicit function theorem: Let $f_{z^k}$ denote the kth derivative (wrt all dynamic variables) of the dynamic model, then let $df_{z^k}$ denote the first-derivative (wrt all model parameters) of $f_{z^k}$ evaluated at the steady state. Note that $f_{z^k}$ is a function of both the model parameters $\theta$ and of the steady state of all dynamic variables $\bar{z}$, which also depend on the parameters. Hence, implicitly $f_{z^k}=f_{z^k}(\theta,\bar{z}(\theta))$ and $df_{z^k}$ consists of two parts:
1. direct derivative wrt to all model parameters given by the preprocessor in the `_params_derivs.m` files
2. contribution of derivative of steady state of dynamic variables (wrt all model parameters): $f_{z^{k+1}} \cdot d\bar{z}$
Note that we already have functionality to compute $d\bar{z}$ analytically.
Having this, the above perturbation systems are basically equations of the following types
$AX +BXC = RHS$ or $AX = RHS$
Now when taking the derivative (wrt to one single parameter $\theta_j$), we get
$A\mathrm{d}\{X\} + B\mathrm{d}\{X\}C = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X - \mathrm{d}\{B\}XC - BX\mathrm{d}\{C\}$
or
$A\mathrm{d}\{X\} = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X$
The first one is a Sylvester type equation, the second one can be solved by taking the inverse of $A$. The only diffculty and tedious work arrises in computing (the highly sparse) derivatives of $RHS$.
***
# New functions: `
## get_perturbation_params_derivs.m`and `get_perturbation_params_derivs_numerical_objective.m`
* The parameter derivatives up to third order are computed in the new function`get_perturbation_params_derivs.m` both analytically and numerically. For numerical derivatives `get_perturbation_params_derivs_numerical_objective.m` is the objective for `fjaco.m` or `hessian_sparse.m` or `hessian.m`.
* `get_perturbation_params_derivs.m` is basically an extended version of the previous `get_first_order_solution_params_derivs.m` function.
* * `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It is used for numerical derivatives, whenever `analytic_derivation_mode=-1|-2`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
* Detailed changes:
* Most important: notation of this function is now in accordance to the k_order_solver, i.e. we do not compute derivatives of Kalman transition matrices A and B, but rather the solution matrices ghx,ghu,ghxx,ghxu,ghuu,ghs2,ghxxx,ghxxu,ghxuu,ghuuu,ghxss,ghuss in accordance with notation used in `oo_.dr`. As a byproduct at first-order, focusing on ghx and ghu instead of Kalman transition matrices A and B makes the computations slightly faster for large models (e.g. for Quest the computations were faster by a couple of seconds, not much, but okay).
* Removed use of `kstate`, see also Dynare/dynare#1653 and Dynare/dynare!1656
* Output arguments are stored in a structure `DERIVS`, there is also a flag `d2flag` that computes parameter hessians needed only in `dsge_likelihood.m`.
* Removed `kronflag` as input. `options_.analytic_derivation_mode` is now used instead of `kronflag`.
* Removed `indvar`, an index that was used to selected specific variables in the derivatives. This is not needed, as we always compute the parameter derivatives for all variables first and then select a subset of variables. The selection now takes place in other functions, like `dsge_likelihood.m`.
* Introduced some checks: (i) deterministic exogenous variables are not supported, (ii) Kronecker method only compatible with first-order approximation so reset to sylvester method, (iii) for purely backward or forward models we need to be careful with the rows in `M_.lead_la g_incidence`, (iv) if `_params_derivs.m` files are missing an error is thrown.
* For numerical derivatives, if mod file does not contain an `estimated_params_block`, a temporary one with the most important parameter information is created.
## `unfold_g4.m`
* When evaluating g3 and g4 one needs to take into account that these do not contain symmetric elements, so one needs to use `unfold_g3.m` and the new function `unfold_g4.m`. This returns an unfolded version of the same matrix (i.e. with symmetric elements).
***
# New test models
`.gitignore` and `Makefile.am` are changed accordingly. Also now it is possible to run test suite on analytic_derivatives, i.e. run `make check m/analytic_derivatives`
## `analytic_derivatives/BrockMirman_PertParamsDerivs.mod`
* This is the Brock Mirman model, where we know the exact policy function $g$ for capital and consumption. As this does not imply a nonzero $g_{\sigma\sigma}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ I added some artificial equations to get nonzero solution matrices with respect to $\sigma$. The true perturbation solution matrices $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ are then computed analytically with Matlab's symbolic toolbox and saved in `nBrockMirmanSYM.mat`. There is a preprocessor flag that recreates these analytical computations if changes are needed (and to check whether I made some errors here ;-) )
* Then solution matrices up to third order and their parameter Jacobians are then compared to the ones computed by Dynare's `k_order_solver` and by `get_perturbation_params_derivs` for all `analytic_derivation_mode`'s. There will be an error if the maximum absolute deviation is too large, i.e. for numerical derivatives (`analytic_derivation_mode=-1|-2`) the tolerance is choosen lower (around 1e-5); for analytical methods we are stricter: around 1e-13 for first-order, 1e-12 for second order, and 1e-11 for third-order.
* As a side note, this mod file also checks Dynare's `k_order_solver` algorithm and throws an error if something is wrong.
* This test model shows that the new functionality works well. And analytical derivatives perform way better and accurate than numerical ones, even for this small model.
## `analytic_derivatives/burnside_3_order_PertParamsDerivs.mod`
* This builds upon `tests/k_order_perturbation/burnside_k_order.mod` and computes the true parameter derivatives analytically by hand.
* This test model also shows that the new functionality works well.
## `analytic_derivatives/LindeTrabandt2019.mod`
* Shows that the new functionality also works for medium-sized models, i.e. a SW type model solved at third order with 35 variables (11 states). 2 shocks and 20 parameters.
* This mod file can be used to tweak the speed of the computations in the future.
* Compares numerical versus analytical parameter derivatives (for first, second and third order). Note that this model clearly shows that numerical ones are quite different than analytical ones even at first order!
## `identification/LindeTrabandt2019_xfail.mod`
* This model is a check for issue Dynare/dynare#1595, see fjaco.m below, and will fail.
* Removed `analytic_derivatives/ls2003.mod` as this mod file is neither in the testsuite nor does it work.
***
# Detailed changes in other functions
## `get_first_order_solution_params_derivs.m`
* Deleted, or actually, renamed to `get_perturbation_params_derivs.m`, as this function now is able to compute the derivatives up to third order
## `identification_numerical_objective.m`
* `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
## `dsge_likelihood.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `T`matrix anymore, but instead of the dynare solution matrix `ghx`. So we recreate `T` here as this amounts to adding some zeros and focusing on selected variables only.
* Added some checks to make sure the first-order approximation is selected.
* Removed `kron_flag` as input, as `get_perturbation_params_derivs` looks into `options_.analytic_derivation_mode` for `kron_flag`.
## `dynare_identification.m`
* make sure that setting `analytic_derivation_mode` is set both in `options_ident` and `options_`. Note that at the end of the function we restore the `options_` structure, so all changes are local. In a next merge request, I will remove the global variables to make all variables local.
## `get_identification_jacobians.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `A` and `B` matrix anymore, but instead of the dynare solution matrix `ghx` and `ghu`. So we recreate these matrices here instead of in `get_perturbation_params_derivs.m`.
* Added `str2func` for better function handles in `fjaco.m`.
## `fjaco.m`
* make `tol`an option, which can be adjusted by changing `options_.dynatol.x`for identification and parameter derivatives purposes.
* include a check and an informative error message, if numerical derivatives (two-sided finite difference method) yield errors in `resol.m` for identification and parameter derivatives purposes. This closes issue Dynare/dynare#1595.
* Changed year of copyright to 2010-2017,2019
***
# Further suggestions and questions
* Ones this is merged, I will merge request an improvement of the identification toolbox, which will work up to third order using the pruned state space. This will also remove some issues and bugs, and also I will remove global variables in this request.
* The third-order derivatives can be further improved by taking sparsity into account and use mex versions for kronecker products etc. I leave this for further testing (and if anybody actually uses this ;-) )
2019-12-17 19:17:09 +01:00
end
if d2flag
%% Analytical computation of Hessian (wrt selected model parameters) of first-derivative of dynamic model, i.e. d2g1
% We use the implicit function theorem (above we already computed: dg1 = g1p + g2*dyy0ex0):
% Accordingly, d2g1, the second-derivative (wrt model parameters), consists of five parts (ignoring transposes, see Ratto and Iskrev (2012) formula 16)
% (1) d(g1p)/dp = g1pp
% (2) d(g1p)/dyy0ex0*d(yy0ex0)/dp = g2p * dyy0ex0
% (3) d(g2)/dp * dyy0ex0 = g2p * dyy0ex0
% (4) d(g2)/dyy0ex0*d(dyy0ex0)/dp * dyy0ex0 = g3 * dyy0ex0 * dyy0ex0
% (5) g2 * d(dyy0ex0)/dp = g2 * d2yy0ex0
% Note that part 2 and 3 are equivalent besides the use of transpose (see Ratto and Iskrev (2012) formula 16)
d2g1_full = sparse ( M . endo_nbr * yy0ex0_nbr , M . param_nbr * M . param_nbr ) ; %initialize
g3_tmp = reshape ( g3 , [ M . endo_nbr * yy0ex0_nbr * yy0ex0_nbr yy0ex0_nbr ] ) ;
d2g1_part4_left = sparse ( M . endo_nbr * yy0ex0_nbr * yy0ex0_nbr , M . param_nbr ) ;
for j = 1 : M . param_nbr
%compute first two terms of part 4
d2g1_part4_left ( : , j ) = g3_tmp * dyy0ex0 ( : , j ) ;
2019-12-20 16:28:06 +01:00
end
Added parameter derivatives of perturbation solution up to 3 order
# Preliminary comments
I finished the identification toolbox at orders two and three using the pruned state space system, but before I merge request this, I decided to first merge the new functionality to compute parameter derivatives of perturbation solution matrices at higher orders. So after this is approved, I merge the identification toolbox.
I guess @rattoma, @sebastien, and @michel are best choices to review this.
I outline the main idea first and then provide some more detailed changes I made to the functions.
***
# Main idea
This merge request is concerned with the *analytical*computation of the parameter derivatives of first, second and third order perturbation solution matrices, i.e. using _closed-form_ expressions to efficiently compute the derivative of $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ *with respect to model parameters* $\theta$. Note that $\theta$ contains model parameters, stderr and corr parameters of shocks. stderr and corr parameters of measurement errors are not yet supported, (they can easily be included as exogenous shocks). The availability of such derivatives is beneficial in terms of more reliable analysis of model sensitivity and parameter identifiability as well as more efficient estimation methods, in particular for models solved up to third order, as it is well-known that numerical derivatives are a tricky business, especially for large models.
References for my approach are:
* Iskrev (2008, 2010) and Schmitt-Grohé and Uribe (2012, Appendix) who were the first to compute the parameter derivatives analytically at first order, however, using inefficient (sparse) Kronecker products.
* Mutschler (2015) who provides the expressions for a second-order, but again using inefficient (sparse) Kronecker products.
* Ratto and Iskrev (2012) who show how the first-order system can be solved accurately, fast and efficiently using existing numerical algorithms for generalized Sylvester equations by taking the parameter derivative with respect to each parameter separately.
* Julliard and Kamenik (2004) who provide the perturbation solution equation system in tensor notation at any order k.
* Levintal (2017) who introduces permutation matrices to express the perturbation solution equation system in matrix notation up to fifth order.
Note that @rattoma already implemented the parameter derivatives of $g_x$ and $g_u$ analytically (and numerically), and I rely heavily on his work in `get_first_order_solution_params_derivs.m` (previously `getH.m`). My additions are mainly to this function and thus it is renamed to `get_perturbation_params_derivs.m`.
The basic idea of this merge request is to take the second- and third-order perturbation solution systems in Julliard and Kamenik (2004), unfold these into an equivalent matrix representation using permutation matrices as in Levintal (2017). Then extending Ratto and Iskrev (2012) one takes the derivative with respect to each parameter separately and gets a computational problem that is linear, albeit large, as it involves either solving generalized Sylvester equations or taking inverses of highly sparse matrices. I will now briefly summarize the perturbation solution system at third order and the system that results when taking the derivative with respect to parameters.
## Perturbation Solution
The following systems arise at first, second, and third order:
$(ghx): f_{x} z_{x} = f_{y_{-}^*} + f_{y_0} g_{x} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{x}= A g_{x} + f_{y_{-}^*}=0$
$(ghu): f_{z} z_{u} = f_{y_0} g_{u} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{u} + f_{u}= A g_u + f_u = 0$
$(ghxx) : A g_{xx} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{x}\right) + f_{zz} \left( z_{x} \otimes z_{x} \right) = 0$
$(ghxu) : A g_{xu} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{x} \otimes z_{u} \right) = 0$
$(ghuu) : A g_{uu} + B g_{xx} \left(g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{u} \otimes z_{u} \right) = 0$
$(ghs2) : (A+B) g_{\sigma\sigma} + \left( f_{y^{**}_{+}y^{**}_{+}} \left(g^{**}_{u} \otimes g^{**}_{u}\right) + f_{y^{**}_{+}} g^{**}_{uu}\right)vec(\Sigma) = 0$
$(ghxxx) : A g_{xxx} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{x}\right) + f_{y_{+}}g^{**}_{xx} \left(g^{*}_x \otimes g^{*}_{xx}\right)P_{x\_xx} + f_{zz} \left( z_{x} \otimes z_{xx} \right)P_{x\_xx} + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{x} \right) = 0$
$(ghxxu) : A g_{xxu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{u} \right) + f_{zz} \left( \left( z_{x} \otimes z_{xu} \right)P_{x\_xu} + \left(z_{xx} \otimes z_{u}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{x} \otimes g^{*}_{xu}\right)P_{x\_xu} + \left(g^{*}_{xx} \otimes g^{*}_{u}\right) \right) = 0$
$(ghxuu) : A g_{xuu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( \left( z_{xu} \otimes z_{u} \right)P_{xu\_u} + \left(z_{x} \otimes z_{uu}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{xu} \otimes g^{*}_{u}\right)P_{xu\_u} + \left(g^{*}_{x} \otimes g^{*}_{uu}\right) \right) = 0$
$(ghuuu) : A g_{uuu} + B g_{xxx} \left(g^{*}_{u} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{u} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( z_{u} \otimes z_{uu} \right)P_{u\_uu} + f_{y_{+}}g^{**}_{xx} \left(g^{*}_{u} \otimes g^{*}_{uu}\right)P_{u\_uu} = 0$
$(ghx\sigma\sigma) : A g_{x\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_x + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{x} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{x} \otimes z_{\sigma\sigma}\right) + F_{xu_{+}u_{+}}\left(I_{n_x} \otimes vec(\Sigma)\right) = 0$
$F_{xu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_x^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{xu_{+}} \otimes z_{u_{+}} \right)P_{xu\_u} + \left(z_{x} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{x} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
$(ghu\sigma\sigma) : A g_{u\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_{u} + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{u} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{u} \otimes z_{\sigma\sigma}\right) + F_{uu_{+}u_{+}}\left(I_{n_u} \otimes vec(\Sigma_u)\right) = 0$
$F_{uu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_u^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{uu_{+}} \otimes z_{u_{+}} \right)P_{uu\_u} + \left(z_{u} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{u} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
A and B are the common perturbation matrices:
$A = f_{y_0} + \begin{pmatrix} \underbrace{0}_{n\times n_{static}} &\vdots& \underbrace{f_{y^{**}_{+}} \cdot g^{**}_{x}}_{n \times n_{spred}} &\vdots& \underbrace{0}_{n\times n_{frwd}} \end{pmatrix}$and $B = \begin{pmatrix} \underbrace{0}_{n \times n_{static}}&\vdots & \underbrace{0}_{n \times n_{pred}} & \vdots & \underbrace{f_{y^{**}_{+}}}_{n \times n_{sfwrd}} \end{pmatrix}$
and $z=(y_{-}^{\ast}; y; y_{+}^{\ast\ast}; u)$ denotes the dynamic model variables as in `M_.lead_lag_incidence`, $y^\ast$ denote state variables, $y^{\ast\ast}$ denote forward looking variables, $y_+$ denote the variables with a lead, $y_{-}$ denote variables with a lag, $y_0$ denote variables at period t, $f$ the model equations, and $f_z$ the first-order dynamic model derivatives, $f_{zz}$ the second-order dynamic derivatives, and $f_{zzz}$ the third-order dynamic model derivatives. Then:
$z_{x} = \begin{pmatrix}I\\g_{x}\\g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{u} =\begin{pmatrix}0\\g_{u}\\g^{**}_{x} \cdot g^{*}_{u}\\I\end{pmatrix}$, $z_{u_{+}} =\begin{pmatrix}0\\0\\g^{**}_{u}\\0\end{pmatrix}$
$z_{xx} = \begin{pmatrix} 0\\g_{xx}\\g^{**}_{x} \left( g^{*}_x \otimes g^{*}_{x} \right) + g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{xu} =\begin{pmatrix}0\\g_{xu}\\g^{**}_{xx} \left( g^{*}_x \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{xu}\\0\end{pmatrix}$, $z_{uu} =\begin{pmatrix}0\\g_{uu}\\g^{**}_{xx} \left( g^{*}_u \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{uu}\\0\end{pmatrix}$,
$z_{xu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_x \otimes I \right)\\0\end{pmatrix}$, $z_{uu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_{u} \otimes I \right)\\0\end{pmatrix}$, $z_{u_{+}u_{+}} =\begin{pmatrix}0\\0\\g^{\ast\ast}_{uu}\\0\end{pmatrix}$, $z_{\sigma\sigma} = \begin{pmatrix}0\\ g_{\sigma\sigma}\\ g^{\ast\ast}_{x}g^{\ast}_{\sigma\sigma} + g^{\ast\ast}_{\sigma\sigma}\\0 \end{pmatrix}$
$P$ are permutation matrices that can be computed using Matlab's `ipermute` function.
## Parameter derivatives of perturbation solutions
First, we need the parameter derivatives of first, second, third, and fourth derivatives of the dynamic model (i.e. g1,g2,g3,g4 in dynamic files), I make use of the implicit function theorem: Let $f_{z^k}$ denote the kth derivative (wrt all dynamic variables) of the dynamic model, then let $df_{z^k}$ denote the first-derivative (wrt all model parameters) of $f_{z^k}$ evaluated at the steady state. Note that $f_{z^k}$ is a function of both the model parameters $\theta$ and of the steady state of all dynamic variables $\bar{z}$, which also depend on the parameters. Hence, implicitly $f_{z^k}=f_{z^k}(\theta,\bar{z}(\theta))$ and $df_{z^k}$ consists of two parts:
1. direct derivative wrt to all model parameters given by the preprocessor in the `_params_derivs.m` files
2. contribution of derivative of steady state of dynamic variables (wrt all model parameters): $f_{z^{k+1}} \cdot d\bar{z}$
Note that we already have functionality to compute $d\bar{z}$ analytically.
Having this, the above perturbation systems are basically equations of the following types
$AX +BXC = RHS$ or $AX = RHS$
Now when taking the derivative (wrt to one single parameter $\theta_j$), we get
$A\mathrm{d}\{X\} + B\mathrm{d}\{X\}C = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X - \mathrm{d}\{B\}XC - BX\mathrm{d}\{C\}$
or
$A\mathrm{d}\{X\} = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X$
The first one is a Sylvester type equation, the second one can be solved by taking the inverse of $A$. The only diffculty and tedious work arrises in computing (the highly sparse) derivatives of $RHS$.
***
# New functions: `
## get_perturbation_params_derivs.m`and `get_perturbation_params_derivs_numerical_objective.m`
* The parameter derivatives up to third order are computed in the new function`get_perturbation_params_derivs.m` both analytically and numerically. For numerical derivatives `get_perturbation_params_derivs_numerical_objective.m` is the objective for `fjaco.m` or `hessian_sparse.m` or `hessian.m`.
* `get_perturbation_params_derivs.m` is basically an extended version of the previous `get_first_order_solution_params_derivs.m` function.
* * `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It is used for numerical derivatives, whenever `analytic_derivation_mode=-1|-2`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
* Detailed changes:
* Most important: notation of this function is now in accordance to the k_order_solver, i.e. we do not compute derivatives of Kalman transition matrices A and B, but rather the solution matrices ghx,ghu,ghxx,ghxu,ghuu,ghs2,ghxxx,ghxxu,ghxuu,ghuuu,ghxss,ghuss in accordance with notation used in `oo_.dr`. As a byproduct at first-order, focusing on ghx and ghu instead of Kalman transition matrices A and B makes the computations slightly faster for large models (e.g. for Quest the computations were faster by a couple of seconds, not much, but okay).
* Removed use of `kstate`, see also Dynare/dynare#1653 and Dynare/dynare!1656
* Output arguments are stored in a structure `DERIVS`, there is also a flag `d2flag` that computes parameter hessians needed only in `dsge_likelihood.m`.
* Removed `kronflag` as input. `options_.analytic_derivation_mode` is now used instead of `kronflag`.
* Removed `indvar`, an index that was used to selected specific variables in the derivatives. This is not needed, as we always compute the parameter derivatives for all variables first and then select a subset of variables. The selection now takes place in other functions, like `dsge_likelihood.m`.
* Introduced some checks: (i) deterministic exogenous variables are not supported, (ii) Kronecker method only compatible with first-order approximation so reset to sylvester method, (iii) for purely backward or forward models we need to be careful with the rows in `M_.lead_la g_incidence`, (iv) if `_params_derivs.m` files are missing an error is thrown.
* For numerical derivatives, if mod file does not contain an `estimated_params_block`, a temporary one with the most important parameter information is created.
## `unfold_g4.m`
* When evaluating g3 and g4 one needs to take into account that these do not contain symmetric elements, so one needs to use `unfold_g3.m` and the new function `unfold_g4.m`. This returns an unfolded version of the same matrix (i.e. with symmetric elements).
***
# New test models
`.gitignore` and `Makefile.am` are changed accordingly. Also now it is possible to run test suite on analytic_derivatives, i.e. run `make check m/analytic_derivatives`
## `analytic_derivatives/BrockMirman_PertParamsDerivs.mod`
* This is the Brock Mirman model, where we know the exact policy function $g$ for capital and consumption. As this does not imply a nonzero $g_{\sigma\sigma}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ I added some artificial equations to get nonzero solution matrices with respect to $\sigma$. The true perturbation solution matrices $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ are then computed analytically with Matlab's symbolic toolbox and saved in `nBrockMirmanSYM.mat`. There is a preprocessor flag that recreates these analytical computations if changes are needed (and to check whether I made some errors here ;-) )
* Then solution matrices up to third order and their parameter Jacobians are then compared to the ones computed by Dynare's `k_order_solver` and by `get_perturbation_params_derivs` for all `analytic_derivation_mode`'s. There will be an error if the maximum absolute deviation is too large, i.e. for numerical derivatives (`analytic_derivation_mode=-1|-2`) the tolerance is choosen lower (around 1e-5); for analytical methods we are stricter: around 1e-13 for first-order, 1e-12 for second order, and 1e-11 for third-order.
* As a side note, this mod file also checks Dynare's `k_order_solver` algorithm and throws an error if something is wrong.
* This test model shows that the new functionality works well. And analytical derivatives perform way better and accurate than numerical ones, even for this small model.
## `analytic_derivatives/burnside_3_order_PertParamsDerivs.mod`
* This builds upon `tests/k_order_perturbation/burnside_k_order.mod` and computes the true parameter derivatives analytically by hand.
* This test model also shows that the new functionality works well.
## `analytic_derivatives/LindeTrabandt2019.mod`
* Shows that the new functionality also works for medium-sized models, i.e. a SW type model solved at third order with 35 variables (11 states). 2 shocks and 20 parameters.
* This mod file can be used to tweak the speed of the computations in the future.
* Compares numerical versus analytical parameter derivatives (for first, second and third order). Note that this model clearly shows that numerical ones are quite different than analytical ones even at first order!
## `identification/LindeTrabandt2019_xfail.mod`
* This model is a check for issue Dynare/dynare#1595, see fjaco.m below, and will fail.
* Removed `analytic_derivatives/ls2003.mod` as this mod file is neither in the testsuite nor does it work.
***
# Detailed changes in other functions
## `get_first_order_solution_params_derivs.m`
* Deleted, or actually, renamed to `get_perturbation_params_derivs.m`, as this function now is able to compute the derivatives up to third order
## `identification_numerical_objective.m`
* `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
## `dsge_likelihood.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `T`matrix anymore, but instead of the dynare solution matrix `ghx`. So we recreate `T` here as this amounts to adding some zeros and focusing on selected variables only.
* Added some checks to make sure the first-order approximation is selected.
* Removed `kron_flag` as input, as `get_perturbation_params_derivs` looks into `options_.analytic_derivation_mode` for `kron_flag`.
## `dynare_identification.m`
* make sure that setting `analytic_derivation_mode` is set both in `options_ident` and `options_`. Note that at the end of the function we restore the `options_` structure, so all changes are local. In a next merge request, I will remove the global variables to make all variables local.
## `get_identification_jacobians.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `A` and `B` matrix anymore, but instead of the dynare solution matrix `ghx` and `ghu`. So we recreate these matrices here instead of in `get_perturbation_params_derivs.m`.
* Added `str2func` for better function handles in `fjaco.m`.
## `fjaco.m`
* make `tol`an option, which can be adjusted by changing `options_.dynatol.x`for identification and parameter derivatives purposes.
* include a check and an informative error message, if numerical derivatives (two-sided finite difference method) yield errors in `resol.m` for identification and parameter derivatives purposes. This closes issue Dynare/dynare#1595.
* Changed year of copyright to 2010-2017,2019
***
# Further suggestions and questions
* Ones this is merged, I will merge request an improvement of the identification toolbox, which will work up to third order using the pruned state space. This will also remove some issues and bugs, and also I will remove global variables in this request.
* The third-order derivatives can be further improved by taking sparsity into account and use mex versions for kronecker products etc. I leave this for further testing (and if anybody actually uses this ;-) )
2019-12-17 19:17:09 +01:00
for j = 1 : M . endo_nbr
%Note that in the following we skip exogenous variables as these are 0 by construction in a stochastic setting
d2g1_part5 = reshape ( g2 ( j , : ) , [ yy0ex0_nbr yy0ex0_nbr ] ) ;
d2g1_part5 = d2g1_part5 ( : , 1 : yy0_nbr ) * reshape ( d2ys ( I , : , : ) , [ yy0_nbr , M . param_nbr * M . param_nbr ] ) ;
for i = 1 : yy0ex0_nbr
ind_part4 = sub2ind ( [ M . endo_nbr yy0ex0_nbr yy0ex0_nbr ] , ones ( yy0ex0_nbr , 1 ) * j , ones ( yy0ex0_nbr , 1 ) * i , ( 1 : yy0ex0_nbr ) ' ) ;
d2g1_part4 = ( d2g1_part4_left ( ind_part4 , : ) ) ' * dyy0ex0 ;
d2g1_part2_and_part3 = ( get_hess_deriv ( g2p , j , i , yy0ex0_nbr , M . param_nbr ) ) ' * dyy0ex0 ;
d2g1_part1 = get_2nd_deriv_mat ( g1pp , j , i , M . param_nbr ) ;
d2g1_tmp = d2g1_part1 + d2g1_part2_and_part3 + d2g1_part2_and_part3 ' + d2g1_part4 + reshape ( d2g1_part5 ( i , : , : ) , [ M . param_nbr M . param_nbr ] ) ;
d2g1_tmp = d2g1_tmp ( indpmodel , indpmodel ) ; %focus only on model parameters
if any ( any ( d2g1_tmp ) )
ind_d2g1_tmp = find ( triu ( d2g1_tmp ) ) ;
d2g1_full ( sub2ind ( [ M . endo_nbr yy0ex0_nbr ] , j , i ) , ind_d2g1_tmp ) = transpose ( d2g1_tmp ( ind_d2g1_tmp ) ) ;
end
end
end
clear d2g1_tmp d2g1_part1 d2g1_part2_and_part3 d2g1_part4 d2g1_part4_left d2g1_part5
%store only nonzero entries and the corresponding indices of d2g1:
% 1st column: equation number of the term appearing
% 2nd column: column number of variable in Jacobian of the dynamic model
% 3rd column: number of the first parameter in derivative
% 4th column: number of the second parameter in derivative
% 5th column: value of the Hessian term
ind_d2g1 = find ( d2g1_full ) ;
d2g1 = zeros ( length ( ind_d2g1 ) , 5 ) ;
for j = 1 : length ( ind_d2g1 )
[ i1 , i2 ] = ind2sub ( size ( d2g1_full ) , ind_d2g1 ( j ) ) ;
[ ig1 , ig2 ] = ind2sub ( size ( g1 ) , i1 ) ;
[ ip1 , ip2 ] = ind2sub ( [ modparam_nbr modparam_nbr ] , i2 ) ;
d2g1 ( j , : ) = [ ig1 ig2 ip1 ip2 d2g1_full ( ind_d2g1 ( j ) ) ] ;
end
clear d2g1_full ;
end
end
%% clear variables that are not used any more
clear dys dyy0 dyy0ex0 d2ys
2019-12-20 16:28:06 +01:00
clear rp_static rpp_static
Added parameter derivatives of perturbation solution up to 3 order
# Preliminary comments
I finished the identification toolbox at orders two and three using the pruned state space system, but before I merge request this, I decided to first merge the new functionality to compute parameter derivatives of perturbation solution matrices at higher orders. So after this is approved, I merge the identification toolbox.
I guess @rattoma, @sebastien, and @michel are best choices to review this.
I outline the main idea first and then provide some more detailed changes I made to the functions.
***
# Main idea
This merge request is concerned with the *analytical*computation of the parameter derivatives of first, second and third order perturbation solution matrices, i.e. using _closed-form_ expressions to efficiently compute the derivative of $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ *with respect to model parameters* $\theta$. Note that $\theta$ contains model parameters, stderr and corr parameters of shocks. stderr and corr parameters of measurement errors are not yet supported, (they can easily be included as exogenous shocks). The availability of such derivatives is beneficial in terms of more reliable analysis of model sensitivity and parameter identifiability as well as more efficient estimation methods, in particular for models solved up to third order, as it is well-known that numerical derivatives are a tricky business, especially for large models.
References for my approach are:
* Iskrev (2008, 2010) and Schmitt-Grohé and Uribe (2012, Appendix) who were the first to compute the parameter derivatives analytically at first order, however, using inefficient (sparse) Kronecker products.
* Mutschler (2015) who provides the expressions for a second-order, but again using inefficient (sparse) Kronecker products.
* Ratto and Iskrev (2012) who show how the first-order system can be solved accurately, fast and efficiently using existing numerical algorithms for generalized Sylvester equations by taking the parameter derivative with respect to each parameter separately.
* Julliard and Kamenik (2004) who provide the perturbation solution equation system in tensor notation at any order k.
* Levintal (2017) who introduces permutation matrices to express the perturbation solution equation system in matrix notation up to fifth order.
Note that @rattoma already implemented the parameter derivatives of $g_x$ and $g_u$ analytically (and numerically), and I rely heavily on his work in `get_first_order_solution_params_derivs.m` (previously `getH.m`). My additions are mainly to this function and thus it is renamed to `get_perturbation_params_derivs.m`.
The basic idea of this merge request is to take the second- and third-order perturbation solution systems in Julliard and Kamenik (2004), unfold these into an equivalent matrix representation using permutation matrices as in Levintal (2017). Then extending Ratto and Iskrev (2012) one takes the derivative with respect to each parameter separately and gets a computational problem that is linear, albeit large, as it involves either solving generalized Sylvester equations or taking inverses of highly sparse matrices. I will now briefly summarize the perturbation solution system at third order and the system that results when taking the derivative with respect to parameters.
## Perturbation Solution
The following systems arise at first, second, and third order:
$(ghx): f_{x} z_{x} = f_{y_{-}^*} + f_{y_0} g_{x} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{x}= A g_{x} + f_{y_{-}^*}=0$
$(ghu): f_{z} z_{u} = f_{y_0} g_{u} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{u} + f_{u}= A g_u + f_u = 0$
$(ghxx) : A g_{xx} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{x}\right) + f_{zz} \left( z_{x} \otimes z_{x} \right) = 0$
$(ghxu) : A g_{xu} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{x} \otimes z_{u} \right) = 0$
$(ghuu) : A g_{uu} + B g_{xx} \left(g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{u} \otimes z_{u} \right) = 0$
$(ghs2) : (A+B) g_{\sigma\sigma} + \left( f_{y^{**}_{+}y^{**}_{+}} \left(g^{**}_{u} \otimes g^{**}_{u}\right) + f_{y^{**}_{+}} g^{**}_{uu}\right)vec(\Sigma) = 0$
$(ghxxx) : A g_{xxx} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{x}\right) + f_{y_{+}}g^{**}_{xx} \left(g^{*}_x \otimes g^{*}_{xx}\right)P_{x\_xx} + f_{zz} \left( z_{x} \otimes z_{xx} \right)P_{x\_xx} + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{x} \right) = 0$
$(ghxxu) : A g_{xxu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{u} \right) + f_{zz} \left( \left( z_{x} \otimes z_{xu} \right)P_{x\_xu} + \left(z_{xx} \otimes z_{u}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{x} \otimes g^{*}_{xu}\right)P_{x\_xu} + \left(g^{*}_{xx} \otimes g^{*}_{u}\right) \right) = 0$
$(ghxuu) : A g_{xuu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( \left( z_{xu} \otimes z_{u} \right)P_{xu\_u} + \left(z_{x} \otimes z_{uu}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{xu} \otimes g^{*}_{u}\right)P_{xu\_u} + \left(g^{*}_{x} \otimes g^{*}_{uu}\right) \right) = 0$
$(ghuuu) : A g_{uuu} + B g_{xxx} \left(g^{*}_{u} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{u} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( z_{u} \otimes z_{uu} \right)P_{u\_uu} + f_{y_{+}}g^{**}_{xx} \left(g^{*}_{u} \otimes g^{*}_{uu}\right)P_{u\_uu} = 0$
$(ghx\sigma\sigma) : A g_{x\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_x + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{x} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{x} \otimes z_{\sigma\sigma}\right) + F_{xu_{+}u_{+}}\left(I_{n_x} \otimes vec(\Sigma)\right) = 0$
$F_{xu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_x^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{xu_{+}} \otimes z_{u_{+}} \right)P_{xu\_u} + \left(z_{x} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{x} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
$(ghu\sigma\sigma) : A g_{u\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_{u} + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{u} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{u} \otimes z_{\sigma\sigma}\right) + F_{uu_{+}u_{+}}\left(I_{n_u} \otimes vec(\Sigma_u)\right) = 0$
$F_{uu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_u^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{uu_{+}} \otimes z_{u_{+}} \right)P_{uu\_u} + \left(z_{u} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{u} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
A and B are the common perturbation matrices:
$A = f_{y_0} + \begin{pmatrix} \underbrace{0}_{n\times n_{static}} &\vdots& \underbrace{f_{y^{**}_{+}} \cdot g^{**}_{x}}_{n \times n_{spred}} &\vdots& \underbrace{0}_{n\times n_{frwd}} \end{pmatrix}$and $B = \begin{pmatrix} \underbrace{0}_{n \times n_{static}}&\vdots & \underbrace{0}_{n \times n_{pred}} & \vdots & \underbrace{f_{y^{**}_{+}}}_{n \times n_{sfwrd}} \end{pmatrix}$
and $z=(y_{-}^{\ast}; y; y_{+}^{\ast\ast}; u)$ denotes the dynamic model variables as in `M_.lead_lag_incidence`, $y^\ast$ denote state variables, $y^{\ast\ast}$ denote forward looking variables, $y_+$ denote the variables with a lead, $y_{-}$ denote variables with a lag, $y_0$ denote variables at period t, $f$ the model equations, and $f_z$ the first-order dynamic model derivatives, $f_{zz}$ the second-order dynamic derivatives, and $f_{zzz}$ the third-order dynamic model derivatives. Then:
$z_{x} = \begin{pmatrix}I\\g_{x}\\g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{u} =\begin{pmatrix}0\\g_{u}\\g^{**}_{x} \cdot g^{*}_{u}\\I\end{pmatrix}$, $z_{u_{+}} =\begin{pmatrix}0\\0\\g^{**}_{u}\\0\end{pmatrix}$
$z_{xx} = \begin{pmatrix} 0\\g_{xx}\\g^{**}_{x} \left( g^{*}_x \otimes g^{*}_{x} \right) + g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{xu} =\begin{pmatrix}0\\g_{xu}\\g^{**}_{xx} \left( g^{*}_x \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{xu}\\0\end{pmatrix}$, $z_{uu} =\begin{pmatrix}0\\g_{uu}\\g^{**}_{xx} \left( g^{*}_u \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{uu}\\0\end{pmatrix}$,
$z_{xu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_x \otimes I \right)\\0\end{pmatrix}$, $z_{uu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_{u} \otimes I \right)\\0\end{pmatrix}$, $z_{u_{+}u_{+}} =\begin{pmatrix}0\\0\\g^{\ast\ast}_{uu}\\0\end{pmatrix}$, $z_{\sigma\sigma} = \begin{pmatrix}0\\ g_{\sigma\sigma}\\ g^{\ast\ast}_{x}g^{\ast}_{\sigma\sigma} + g^{\ast\ast}_{\sigma\sigma}\\0 \end{pmatrix}$
$P$ are permutation matrices that can be computed using Matlab's `ipermute` function.
## Parameter derivatives of perturbation solutions
First, we need the parameter derivatives of first, second, third, and fourth derivatives of the dynamic model (i.e. g1,g2,g3,g4 in dynamic files), I make use of the implicit function theorem: Let $f_{z^k}$ denote the kth derivative (wrt all dynamic variables) of the dynamic model, then let $df_{z^k}$ denote the first-derivative (wrt all model parameters) of $f_{z^k}$ evaluated at the steady state. Note that $f_{z^k}$ is a function of both the model parameters $\theta$ and of the steady state of all dynamic variables $\bar{z}$, which also depend on the parameters. Hence, implicitly $f_{z^k}=f_{z^k}(\theta,\bar{z}(\theta))$ and $df_{z^k}$ consists of two parts:
1. direct derivative wrt to all model parameters given by the preprocessor in the `_params_derivs.m` files
2. contribution of derivative of steady state of dynamic variables (wrt all model parameters): $f_{z^{k+1}} \cdot d\bar{z}$
Note that we already have functionality to compute $d\bar{z}$ analytically.
Having this, the above perturbation systems are basically equations of the following types
$AX +BXC = RHS$ or $AX = RHS$
Now when taking the derivative (wrt to one single parameter $\theta_j$), we get
$A\mathrm{d}\{X\} + B\mathrm{d}\{X\}C = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X - \mathrm{d}\{B\}XC - BX\mathrm{d}\{C\}$
or
$A\mathrm{d}\{X\} = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X$
The first one is a Sylvester type equation, the second one can be solved by taking the inverse of $A$. The only diffculty and tedious work arrises in computing (the highly sparse) derivatives of $RHS$.
***
# New functions: `
## get_perturbation_params_derivs.m`and `get_perturbation_params_derivs_numerical_objective.m`
* The parameter derivatives up to third order are computed in the new function`get_perturbation_params_derivs.m` both analytically and numerically. For numerical derivatives `get_perturbation_params_derivs_numerical_objective.m` is the objective for `fjaco.m` or `hessian_sparse.m` or `hessian.m`.
* `get_perturbation_params_derivs.m` is basically an extended version of the previous `get_first_order_solution_params_derivs.m` function.
* * `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It is used for numerical derivatives, whenever `analytic_derivation_mode=-1|-2`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
* Detailed changes:
* Most important: notation of this function is now in accordance to the k_order_solver, i.e. we do not compute derivatives of Kalman transition matrices A and B, but rather the solution matrices ghx,ghu,ghxx,ghxu,ghuu,ghs2,ghxxx,ghxxu,ghxuu,ghuuu,ghxss,ghuss in accordance with notation used in `oo_.dr`. As a byproduct at first-order, focusing on ghx and ghu instead of Kalman transition matrices A and B makes the computations slightly faster for large models (e.g. for Quest the computations were faster by a couple of seconds, not much, but okay).
* Removed use of `kstate`, see also Dynare/dynare#1653 and Dynare/dynare!1656
* Output arguments are stored in a structure `DERIVS`, there is also a flag `d2flag` that computes parameter hessians needed only in `dsge_likelihood.m`.
* Removed `kronflag` as input. `options_.analytic_derivation_mode` is now used instead of `kronflag`.
* Removed `indvar`, an index that was used to selected specific variables in the derivatives. This is not needed, as we always compute the parameter derivatives for all variables first and then select a subset of variables. The selection now takes place in other functions, like `dsge_likelihood.m`.
* Introduced some checks: (i) deterministic exogenous variables are not supported, (ii) Kronecker method only compatible with first-order approximation so reset to sylvester method, (iii) for purely backward or forward models we need to be careful with the rows in `M_.lead_la g_incidence`, (iv) if `_params_derivs.m` files are missing an error is thrown.
* For numerical derivatives, if mod file does not contain an `estimated_params_block`, a temporary one with the most important parameter information is created.
## `unfold_g4.m`
* When evaluating g3 and g4 one needs to take into account that these do not contain symmetric elements, so one needs to use `unfold_g3.m` and the new function `unfold_g4.m`. This returns an unfolded version of the same matrix (i.e. with symmetric elements).
***
# New test models
`.gitignore` and `Makefile.am` are changed accordingly. Also now it is possible to run test suite on analytic_derivatives, i.e. run `make check m/analytic_derivatives`
## `analytic_derivatives/BrockMirman_PertParamsDerivs.mod`
* This is the Brock Mirman model, where we know the exact policy function $g$ for capital and consumption. As this does not imply a nonzero $g_{\sigma\sigma}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ I added some artificial equations to get nonzero solution matrices with respect to $\sigma$. The true perturbation solution matrices $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ are then computed analytically with Matlab's symbolic toolbox and saved in `nBrockMirmanSYM.mat`. There is a preprocessor flag that recreates these analytical computations if changes are needed (and to check whether I made some errors here ;-) )
* Then solution matrices up to third order and their parameter Jacobians are then compared to the ones computed by Dynare's `k_order_solver` and by `get_perturbation_params_derivs` for all `analytic_derivation_mode`'s. There will be an error if the maximum absolute deviation is too large, i.e. for numerical derivatives (`analytic_derivation_mode=-1|-2`) the tolerance is choosen lower (around 1e-5); for analytical methods we are stricter: around 1e-13 for first-order, 1e-12 for second order, and 1e-11 for third-order.
* As a side note, this mod file also checks Dynare's `k_order_solver` algorithm and throws an error if something is wrong.
* This test model shows that the new functionality works well. And analytical derivatives perform way better and accurate than numerical ones, even for this small model.
## `analytic_derivatives/burnside_3_order_PertParamsDerivs.mod`
* This builds upon `tests/k_order_perturbation/burnside_k_order.mod` and computes the true parameter derivatives analytically by hand.
* This test model also shows that the new functionality works well.
## `analytic_derivatives/LindeTrabandt2019.mod`
* Shows that the new functionality also works for medium-sized models, i.e. a SW type model solved at third order with 35 variables (11 states). 2 shocks and 20 parameters.
* This mod file can be used to tweak the speed of the computations in the future.
* Compares numerical versus analytical parameter derivatives (for first, second and third order). Note that this model clearly shows that numerical ones are quite different than analytical ones even at first order!
## `identification/LindeTrabandt2019_xfail.mod`
* This model is a check for issue Dynare/dynare#1595, see fjaco.m below, and will fail.
* Removed `analytic_derivatives/ls2003.mod` as this mod file is neither in the testsuite nor does it work.
***
# Detailed changes in other functions
## `get_first_order_solution_params_derivs.m`
* Deleted, or actually, renamed to `get_perturbation_params_derivs.m`, as this function now is able to compute the derivatives up to third order
## `identification_numerical_objective.m`
* `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
## `dsge_likelihood.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `T`matrix anymore, but instead of the dynare solution matrix `ghx`. So we recreate `T` here as this amounts to adding some zeros and focusing on selected variables only.
* Added some checks to make sure the first-order approximation is selected.
* Removed `kron_flag` as input, as `get_perturbation_params_derivs` looks into `options_.analytic_derivation_mode` for `kron_flag`.
## `dynare_identification.m`
* make sure that setting `analytic_derivation_mode` is set both in `options_ident` and `options_`. Note that at the end of the function we restore the `options_` structure, so all changes are local. In a next merge request, I will remove the global variables to make all variables local.
## `get_identification_jacobians.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `A` and `B` matrix anymore, but instead of the dynare solution matrix `ghx` and `ghu`. So we recreate these matrices here instead of in `get_perturbation_params_derivs.m`.
* Added `str2func` for better function handles in `fjaco.m`.
## `fjaco.m`
* make `tol`an option, which can be adjusted by changing `options_.dynatol.x`for identification and parameter derivatives purposes.
* include a check and an informative error message, if numerical derivatives (two-sided finite difference method) yield errors in `resol.m` for identification and parameter derivatives purposes. This closes issue Dynare/dynare#1595.
* Changed year of copyright to 2010-2017,2019
***
# Further suggestions and questions
* Ones this is merged, I will merge request an improvement of the identification toolbox, which will work up to third order using the pruned state space. This will also remove some issues and bugs, and also I will remove global variables in this request.
* The third-order derivatives can be further improved by taking sparsity into account and use mex versions for kronecker products etc. I leave this for further testing (and if anybody actually uses this ;-) )
2019-12-17 19:17:09 +01:00
clear g1_static g1p_static g1p g1pp
clear g2_static g2p tmp_g2 g3_tmp
clear ind_d2g1 ind_d2g1_tmp ind_part4 i j i1 i2 ig1 ig2 I II JJ ip1 ip2 is
if order == 1
clear g2 g3
elseif order == 2
clear g3
end
%% Construct Jacobian (wrt all selected parameters) of Sigma_e and Corr_e for Gaussian innovations
dSigma_e = zeros ( M . exo_nbr , M . exo_nbr , totparam_nbr ) ; %initialize
dCorrelation_matrix = zeros ( M . exo_nbr , M . exo_nbr , totparam_nbr ) ; %initialize
% Compute Jacobians wrt stderr parameters (these come first)
% note that derivatives wrt stderr parameters are zero by construction for Corr_e
if ~ isempty ( indpstderr )
for jp = 1 : stderrparam_nbr
dSigma_e ( indpstderr ( jp ) , indpstderr ( jp ) , jp ) = 2 * sqrt ( M . Sigma_e ( indpstderr ( jp ) , indpstderr ( jp ) ) ) ;
if isdiag ( M . Sigma_e ) == 0 % if there are correlated errors add cross derivatives
indotherex0 = 1 : M . exo_nbr ;
indotherex0 ( indpstderr ( jp ) ) = [ ] ;
for kk = indotherex0
dSigma_e ( indpstderr ( jp ) , kk , jp ) = M . Correlation_matrix ( indpstderr ( jp ) , kk ) * sqrt ( M . Sigma_e ( kk , kk ) ) ;
dSigma_e ( kk , indpstderr ( jp ) , jp ) = dSigma_e ( indpstderr ( jp ) , kk , jp ) ; %symmetry
end
end
end
end
% Compute Jacobians wrt corr parameters (these come second)
if ~ isempty ( indpcorr )
for jp = 1 : corrparam_nbr
dSigma_e ( indpcorr ( jp , 1 ) , indpcorr ( jp , 2 ) , stderrparam_nbr + jp ) = sqrt ( M . Sigma_e ( indpcorr ( jp , 1 ) , indpcorr ( jp , 1 ) ) ) * sqrt ( M . Sigma_e ( indpcorr ( jp , 2 ) , indpcorr ( jp , 2 ) ) ) ;
dSigma_e ( indpcorr ( jp , 2 ) , indpcorr ( jp , 1 ) , stderrparam_nbr + jp ) = dSigma_e ( indpcorr ( jp , 1 ) , indpcorr ( jp , 2 ) , stderrparam_nbr + jp ) ; %symmetry
dCorrelation_matrix ( indpcorr ( jp , 1 ) , indpcorr ( jp , 2 ) , stderrparam_nbr + jp ) = 1 ;
dCorrelation_matrix ( indpcorr ( jp , 2 ) , indpcorr ( jp , 1 ) , stderrparam_nbr + jp ) = 1 ; %symmetry
end
end
% note that derivatives wrt model parameters (these come third) are zero by construction for Sigma_e and Corr_e
%% Analytical computation of Jacobian (wrt selected model parameters) of first-order solution matrices ghx, ghu, and of Om=ghu*Sigma_e*ghu'
%Notation:
% fy_ = g1(:,nonzeros(klag)) in DR order
% fy0 = g1(:,nonzeros(kcurr)) in DR order
% fyp = g1(:,nonzeros(klead)) in DR order
if analytic_derivation_mode == 1
% The following derivations are based on Iskrev (2010) and its online appendix A.
% Basic idea is to make use of the implicit function theorem.
% Define Kalman transition matrix KalmanA = [0 ghx 0], where the first zero spans M.nstatic columns, and the second zero M.nfwrd columns
% At first order we have: F = GAM0*KalmanA - GAM1*KalmanA*KalmanA - GAM2 = 0, where GAM0=fy0, GAM1=-fyp, GAM2 = -fy_
% Note that F is a function of parameters p and KalmanA, which is also a function of p,therefore, F = F(p,KalmanA(p))=0, and hence,
% dF = Fp + dF_dKalmanA*dKalmanA = 0 or dKalmanA = - Fp/dF_dKalmanA
% Some auxiliary matrices
I_endo = speye ( M . endo_nbr ) ;
KalmanA = [ zeros ( M . endo_nbr , M . nstatic ) oo . dr . ghx zeros ( M . endo_nbr , M . nfwrd ) ] ;
% Reshape to write first dynamic derivatives in the Magnus and Neudecker style, i.e. dvec(X)/dp
GAM0 = zeros ( M . endo_nbr , M . endo_nbr ) ;
GAM0 ( : , kcurr ~= 0 , : ) = g1 ( : , nonzeros ( kcurr ) ) ;
dGAM0 = zeros ( M . endo_nbr , M . endo_nbr , modparam_nbr ) ;
dGAM0 ( : , kcurr ~= 0 , : ) = dg1 ( : , nonzeros ( kcurr ) , : ) ;
dGAM0 = reshape ( dGAM0 , M . endo_nbr * M . endo_nbr , modparam_nbr ) ;
GAM1 = zeros ( M . endo_nbr , M . endo_nbr ) ;
GAM1 ( : , klead ~= 0 , : ) = - g1 ( : , nonzeros ( klead ) ) ;
dGAM1 = zeros ( M . endo_nbr , M . endo_nbr , modparam_nbr ) ;
dGAM1 ( : , klead ~= 0 , : ) = - dg1 ( : , nonzeros ( klead ) , : ) ;
dGAM1 = reshape ( dGAM1 , M . endo_nbr * M . endo_nbr , modparam_nbr ) ;
dGAM2 = zeros ( M . endo_nbr , M . endo_nbr , modparam_nbr ) ;
dGAM2 ( : , klag ~= 0 , : ) = - dg1 ( : , nonzeros ( klag ) , : ) ;
dGAM2 = reshape ( dGAM2 , M . endo_nbr * M . endo_nbr , modparam_nbr ) ;
GAM3 = - g1 ( : , yy0_nbr + 1 : end ) ;
dGAM3 = reshape ( - dg1 ( : , yy0_nbr + 1 : end , : ) , M . endo_nbr * M . exo_nbr , modparam_nbr ) ;
% Compute dKalmanA via implicit function
dF_dKalmanAghx = kron ( I_endo , GAM0 ) - kron ( KalmanA ' , GAM1 ) - kron ( I_endo , GAM1 * KalmanA ) ; %equation 31 in Appendix A of Iskrev (2010)
Fp = kron ( KalmanA ' , I_endo ) * dGAM0 - kron ( ( KalmanA ' ) ^2 , I_endo ) * dGAM1 - dGAM2 ; %equation 32 in Appendix A of Iskrev (2010)
dKalmanA = - dF_dKalmanAghx \ Fp ;
2019-12-20 16:28:06 +01:00
Added parameter derivatives of perturbation solution up to 3 order
# Preliminary comments
I finished the identification toolbox at orders two and three using the pruned state space system, but before I merge request this, I decided to first merge the new functionality to compute parameter derivatives of perturbation solution matrices at higher orders. So after this is approved, I merge the identification toolbox.
I guess @rattoma, @sebastien, and @michel are best choices to review this.
I outline the main idea first and then provide some more detailed changes I made to the functions.
***
# Main idea
This merge request is concerned with the *analytical*computation of the parameter derivatives of first, second and third order perturbation solution matrices, i.e. using _closed-form_ expressions to efficiently compute the derivative of $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ *with respect to model parameters* $\theta$. Note that $\theta$ contains model parameters, stderr and corr parameters of shocks. stderr and corr parameters of measurement errors are not yet supported, (they can easily be included as exogenous shocks). The availability of such derivatives is beneficial in terms of more reliable analysis of model sensitivity and parameter identifiability as well as more efficient estimation methods, in particular for models solved up to third order, as it is well-known that numerical derivatives are a tricky business, especially for large models.
References for my approach are:
* Iskrev (2008, 2010) and Schmitt-Grohé and Uribe (2012, Appendix) who were the first to compute the parameter derivatives analytically at first order, however, using inefficient (sparse) Kronecker products.
* Mutschler (2015) who provides the expressions for a second-order, but again using inefficient (sparse) Kronecker products.
* Ratto and Iskrev (2012) who show how the first-order system can be solved accurately, fast and efficiently using existing numerical algorithms for generalized Sylvester equations by taking the parameter derivative with respect to each parameter separately.
* Julliard and Kamenik (2004) who provide the perturbation solution equation system in tensor notation at any order k.
* Levintal (2017) who introduces permutation matrices to express the perturbation solution equation system in matrix notation up to fifth order.
Note that @rattoma already implemented the parameter derivatives of $g_x$ and $g_u$ analytically (and numerically), and I rely heavily on his work in `get_first_order_solution_params_derivs.m` (previously `getH.m`). My additions are mainly to this function and thus it is renamed to `get_perturbation_params_derivs.m`.
The basic idea of this merge request is to take the second- and third-order perturbation solution systems in Julliard and Kamenik (2004), unfold these into an equivalent matrix representation using permutation matrices as in Levintal (2017). Then extending Ratto and Iskrev (2012) one takes the derivative with respect to each parameter separately and gets a computational problem that is linear, albeit large, as it involves either solving generalized Sylvester equations or taking inverses of highly sparse matrices. I will now briefly summarize the perturbation solution system at third order and the system that results when taking the derivative with respect to parameters.
## Perturbation Solution
The following systems arise at first, second, and third order:
$(ghx): f_{x} z_{x} = f_{y_{-}^*} + f_{y_0} g_{x} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{x}= A g_{x} + f_{y_{-}^*}=0$
$(ghu): f_{z} z_{u} = f_{y_0} g_{u} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{u} + f_{u}= A g_u + f_u = 0$
$(ghxx) : A g_{xx} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{x}\right) + f_{zz} \left( z_{x} \otimes z_{x} \right) = 0$
$(ghxu) : A g_{xu} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{x} \otimes z_{u} \right) = 0$
$(ghuu) : A g_{uu} + B g_{xx} \left(g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{u} \otimes z_{u} \right) = 0$
$(ghs2) : (A+B) g_{\sigma\sigma} + \left( f_{y^{**}_{+}y^{**}_{+}} \left(g^{**}_{u} \otimes g^{**}_{u}\right) + f_{y^{**}_{+}} g^{**}_{uu}\right)vec(\Sigma) = 0$
$(ghxxx) : A g_{xxx} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{x}\right) + f_{y_{+}}g^{**}_{xx} \left(g^{*}_x \otimes g^{*}_{xx}\right)P_{x\_xx} + f_{zz} \left( z_{x} \otimes z_{xx} \right)P_{x\_xx} + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{x} \right) = 0$
$(ghxxu) : A g_{xxu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{u} \right) + f_{zz} \left( \left( z_{x} \otimes z_{xu} \right)P_{x\_xu} + \left(z_{xx} \otimes z_{u}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{x} \otimes g^{*}_{xu}\right)P_{x\_xu} + \left(g^{*}_{xx} \otimes g^{*}_{u}\right) \right) = 0$
$(ghxuu) : A g_{xuu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( \left( z_{xu} \otimes z_{u} \right)P_{xu\_u} + \left(z_{x} \otimes z_{uu}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{xu} \otimes g^{*}_{u}\right)P_{xu\_u} + \left(g^{*}_{x} \otimes g^{*}_{uu}\right) \right) = 0$
$(ghuuu) : A g_{uuu} + B g_{xxx} \left(g^{*}_{u} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{u} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( z_{u} \otimes z_{uu} \right)P_{u\_uu} + f_{y_{+}}g^{**}_{xx} \left(g^{*}_{u} \otimes g^{*}_{uu}\right)P_{u\_uu} = 0$
$(ghx\sigma\sigma) : A g_{x\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_x + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{x} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{x} \otimes z_{\sigma\sigma}\right) + F_{xu_{+}u_{+}}\left(I_{n_x} \otimes vec(\Sigma)\right) = 0$
$F_{xu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_x^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{xu_{+}} \otimes z_{u_{+}} \right)P_{xu\_u} + \left(z_{x} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{x} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
$(ghu\sigma\sigma) : A g_{u\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_{u} + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{u} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{u} \otimes z_{\sigma\sigma}\right) + F_{uu_{+}u_{+}}\left(I_{n_u} \otimes vec(\Sigma_u)\right) = 0$
$F_{uu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_u^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{uu_{+}} \otimes z_{u_{+}} \right)P_{uu\_u} + \left(z_{u} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{u} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
A and B are the common perturbation matrices:
$A = f_{y_0} + \begin{pmatrix} \underbrace{0}_{n\times n_{static}} &\vdots& \underbrace{f_{y^{**}_{+}} \cdot g^{**}_{x}}_{n \times n_{spred}} &\vdots& \underbrace{0}_{n\times n_{frwd}} \end{pmatrix}$and $B = \begin{pmatrix} \underbrace{0}_{n \times n_{static}}&\vdots & \underbrace{0}_{n \times n_{pred}} & \vdots & \underbrace{f_{y^{**}_{+}}}_{n \times n_{sfwrd}} \end{pmatrix}$
and $z=(y_{-}^{\ast}; y; y_{+}^{\ast\ast}; u)$ denotes the dynamic model variables as in `M_.lead_lag_incidence`, $y^\ast$ denote state variables, $y^{\ast\ast}$ denote forward looking variables, $y_+$ denote the variables with a lead, $y_{-}$ denote variables with a lag, $y_0$ denote variables at period t, $f$ the model equations, and $f_z$ the first-order dynamic model derivatives, $f_{zz}$ the second-order dynamic derivatives, and $f_{zzz}$ the third-order dynamic model derivatives. Then:
$z_{x} = \begin{pmatrix}I\\g_{x}\\g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{u} =\begin{pmatrix}0\\g_{u}\\g^{**}_{x} \cdot g^{*}_{u}\\I\end{pmatrix}$, $z_{u_{+}} =\begin{pmatrix}0\\0\\g^{**}_{u}\\0\end{pmatrix}$
$z_{xx} = \begin{pmatrix} 0\\g_{xx}\\g^{**}_{x} \left( g^{*}_x \otimes g^{*}_{x} \right) + g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{xu} =\begin{pmatrix}0\\g_{xu}\\g^{**}_{xx} \left( g^{*}_x \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{xu}\\0\end{pmatrix}$, $z_{uu} =\begin{pmatrix}0\\g_{uu}\\g^{**}_{xx} \left( g^{*}_u \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{uu}\\0\end{pmatrix}$,
$z_{xu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_x \otimes I \right)\\0\end{pmatrix}$, $z_{uu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_{u} \otimes I \right)\\0\end{pmatrix}$, $z_{u_{+}u_{+}} =\begin{pmatrix}0\\0\\g^{\ast\ast}_{uu}\\0\end{pmatrix}$, $z_{\sigma\sigma} = \begin{pmatrix}0\\ g_{\sigma\sigma}\\ g^{\ast\ast}_{x}g^{\ast}_{\sigma\sigma} + g^{\ast\ast}_{\sigma\sigma}\\0 \end{pmatrix}$
$P$ are permutation matrices that can be computed using Matlab's `ipermute` function.
## Parameter derivatives of perturbation solutions
First, we need the parameter derivatives of first, second, third, and fourth derivatives of the dynamic model (i.e. g1,g2,g3,g4 in dynamic files), I make use of the implicit function theorem: Let $f_{z^k}$ denote the kth derivative (wrt all dynamic variables) of the dynamic model, then let $df_{z^k}$ denote the first-derivative (wrt all model parameters) of $f_{z^k}$ evaluated at the steady state. Note that $f_{z^k}$ is a function of both the model parameters $\theta$ and of the steady state of all dynamic variables $\bar{z}$, which also depend on the parameters. Hence, implicitly $f_{z^k}=f_{z^k}(\theta,\bar{z}(\theta))$ and $df_{z^k}$ consists of two parts:
1. direct derivative wrt to all model parameters given by the preprocessor in the `_params_derivs.m` files
2. contribution of derivative of steady state of dynamic variables (wrt all model parameters): $f_{z^{k+1}} \cdot d\bar{z}$
Note that we already have functionality to compute $d\bar{z}$ analytically.
Having this, the above perturbation systems are basically equations of the following types
$AX +BXC = RHS$ or $AX = RHS$
Now when taking the derivative (wrt to one single parameter $\theta_j$), we get
$A\mathrm{d}\{X\} + B\mathrm{d}\{X\}C = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X - \mathrm{d}\{B\}XC - BX\mathrm{d}\{C\}$
or
$A\mathrm{d}\{X\} = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X$
The first one is a Sylvester type equation, the second one can be solved by taking the inverse of $A$. The only diffculty and tedious work arrises in computing (the highly sparse) derivatives of $RHS$.
***
# New functions: `
## get_perturbation_params_derivs.m`and `get_perturbation_params_derivs_numerical_objective.m`
* The parameter derivatives up to third order are computed in the new function`get_perturbation_params_derivs.m` both analytically and numerically. For numerical derivatives `get_perturbation_params_derivs_numerical_objective.m` is the objective for `fjaco.m` or `hessian_sparse.m` or `hessian.m`.
* `get_perturbation_params_derivs.m` is basically an extended version of the previous `get_first_order_solution_params_derivs.m` function.
* * `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It is used for numerical derivatives, whenever `analytic_derivation_mode=-1|-2`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
* Detailed changes:
* Most important: notation of this function is now in accordance to the k_order_solver, i.e. we do not compute derivatives of Kalman transition matrices A and B, but rather the solution matrices ghx,ghu,ghxx,ghxu,ghuu,ghs2,ghxxx,ghxxu,ghxuu,ghuuu,ghxss,ghuss in accordance with notation used in `oo_.dr`. As a byproduct at first-order, focusing on ghx and ghu instead of Kalman transition matrices A and B makes the computations slightly faster for large models (e.g. for Quest the computations were faster by a couple of seconds, not much, but okay).
* Removed use of `kstate`, see also Dynare/dynare#1653 and Dynare/dynare!1656
* Output arguments are stored in a structure `DERIVS`, there is also a flag `d2flag` that computes parameter hessians needed only in `dsge_likelihood.m`.
* Removed `kronflag` as input. `options_.analytic_derivation_mode` is now used instead of `kronflag`.
* Removed `indvar`, an index that was used to selected specific variables in the derivatives. This is not needed, as we always compute the parameter derivatives for all variables first and then select a subset of variables. The selection now takes place in other functions, like `dsge_likelihood.m`.
* Introduced some checks: (i) deterministic exogenous variables are not supported, (ii) Kronecker method only compatible with first-order approximation so reset to sylvester method, (iii) for purely backward or forward models we need to be careful with the rows in `M_.lead_la g_incidence`, (iv) if `_params_derivs.m` files are missing an error is thrown.
* For numerical derivatives, if mod file does not contain an `estimated_params_block`, a temporary one with the most important parameter information is created.
## `unfold_g4.m`
* When evaluating g3 and g4 one needs to take into account that these do not contain symmetric elements, so one needs to use `unfold_g3.m` and the new function `unfold_g4.m`. This returns an unfolded version of the same matrix (i.e. with symmetric elements).
***
# New test models
`.gitignore` and `Makefile.am` are changed accordingly. Also now it is possible to run test suite on analytic_derivatives, i.e. run `make check m/analytic_derivatives`
## `analytic_derivatives/BrockMirman_PertParamsDerivs.mod`
* This is the Brock Mirman model, where we know the exact policy function $g$ for capital and consumption. As this does not imply a nonzero $g_{\sigma\sigma}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ I added some artificial equations to get nonzero solution matrices with respect to $\sigma$. The true perturbation solution matrices $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ are then computed analytically with Matlab's symbolic toolbox and saved in `nBrockMirmanSYM.mat`. There is a preprocessor flag that recreates these analytical computations if changes are needed (and to check whether I made some errors here ;-) )
* Then solution matrices up to third order and their parameter Jacobians are then compared to the ones computed by Dynare's `k_order_solver` and by `get_perturbation_params_derivs` for all `analytic_derivation_mode`'s. There will be an error if the maximum absolute deviation is too large, i.e. for numerical derivatives (`analytic_derivation_mode=-1|-2`) the tolerance is choosen lower (around 1e-5); for analytical methods we are stricter: around 1e-13 for first-order, 1e-12 for second order, and 1e-11 for third-order.
* As a side note, this mod file also checks Dynare's `k_order_solver` algorithm and throws an error if something is wrong.
* This test model shows that the new functionality works well. And analytical derivatives perform way better and accurate than numerical ones, even for this small model.
## `analytic_derivatives/burnside_3_order_PertParamsDerivs.mod`
* This builds upon `tests/k_order_perturbation/burnside_k_order.mod` and computes the true parameter derivatives analytically by hand.
* This test model also shows that the new functionality works well.
## `analytic_derivatives/LindeTrabandt2019.mod`
* Shows that the new functionality also works for medium-sized models, i.e. a SW type model solved at third order with 35 variables (11 states). 2 shocks and 20 parameters.
* This mod file can be used to tweak the speed of the computations in the future.
* Compares numerical versus analytical parameter derivatives (for first, second and third order). Note that this model clearly shows that numerical ones are quite different than analytical ones even at first order!
## `identification/LindeTrabandt2019_xfail.mod`
* This model is a check for issue Dynare/dynare#1595, see fjaco.m below, and will fail.
* Removed `analytic_derivatives/ls2003.mod` as this mod file is neither in the testsuite nor does it work.
***
# Detailed changes in other functions
## `get_first_order_solution_params_derivs.m`
* Deleted, or actually, renamed to `get_perturbation_params_derivs.m`, as this function now is able to compute the derivatives up to third order
## `identification_numerical_objective.m`
* `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
## `dsge_likelihood.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `T`matrix anymore, but instead of the dynare solution matrix `ghx`. So we recreate `T` here as this amounts to adding some zeros and focusing on selected variables only.
* Added some checks to make sure the first-order approximation is selected.
* Removed `kron_flag` as input, as `get_perturbation_params_derivs` looks into `options_.analytic_derivation_mode` for `kron_flag`.
## `dynare_identification.m`
* make sure that setting `analytic_derivation_mode` is set both in `options_ident` and `options_`. Note that at the end of the function we restore the `options_` structure, so all changes are local. In a next merge request, I will remove the global variables to make all variables local.
## `get_identification_jacobians.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `A` and `B` matrix anymore, but instead of the dynare solution matrix `ghx` and `ghu`. So we recreate these matrices here instead of in `get_perturbation_params_derivs.m`.
* Added `str2func` for better function handles in `fjaco.m`.
## `fjaco.m`
* make `tol`an option, which can be adjusted by changing `options_.dynatol.x`for identification and parameter derivatives purposes.
* include a check and an informative error message, if numerical derivatives (two-sided finite difference method) yield errors in `resol.m` for identification and parameter derivatives purposes. This closes issue Dynare/dynare#1595.
* Changed year of copyright to 2010-2017,2019
***
# Further suggestions and questions
* Ones this is merged, I will merge request an improvement of the identification toolbox, which will work up to third order using the pruned state space. This will also remove some issues and bugs, and also I will remove global variables in this request.
* The third-order derivatives can be further improved by taking sparsity into account and use mex versions for kronecker products etc. I leave this for further testing (and if anybody actually uses this ;-) )
2019-12-17 19:17:09 +01:00
% Compute dBB from expressions 33 in Iskrev (2010) Appendix A
MM = GAM0 - GAM1 * KalmanA ; %this corresponds to matrix M in Ratto and Iskrev (2012, page 6)
invMM = MM \ eye ( M . endo_nbr ) ;
dghu = - kron ( ( invMM * GAM3 ) ' , invMM ) * ( dGAM0 - kron ( KalmanA ' , I_endo ) * dGAM1 - kron ( I_endo , GAM1 ) * dKalmanA ) + kron ( speye ( M . exo_nbr ) , invMM ) * dGAM3 ;
2019-12-20 16:28:06 +01:00
Added parameter derivatives of perturbation solution up to 3 order
# Preliminary comments
I finished the identification toolbox at orders two and three using the pruned state space system, but before I merge request this, I decided to first merge the new functionality to compute parameter derivatives of perturbation solution matrices at higher orders. So after this is approved, I merge the identification toolbox.
I guess @rattoma, @sebastien, and @michel are best choices to review this.
I outline the main idea first and then provide some more detailed changes I made to the functions.
***
# Main idea
This merge request is concerned with the *analytical*computation of the parameter derivatives of first, second and third order perturbation solution matrices, i.e. using _closed-form_ expressions to efficiently compute the derivative of $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ *with respect to model parameters* $\theta$. Note that $\theta$ contains model parameters, stderr and corr parameters of shocks. stderr and corr parameters of measurement errors are not yet supported, (they can easily be included as exogenous shocks). The availability of such derivatives is beneficial in terms of more reliable analysis of model sensitivity and parameter identifiability as well as more efficient estimation methods, in particular for models solved up to third order, as it is well-known that numerical derivatives are a tricky business, especially for large models.
References for my approach are:
* Iskrev (2008, 2010) and Schmitt-Grohé and Uribe (2012, Appendix) who were the first to compute the parameter derivatives analytically at first order, however, using inefficient (sparse) Kronecker products.
* Mutschler (2015) who provides the expressions for a second-order, but again using inefficient (sparse) Kronecker products.
* Ratto and Iskrev (2012) who show how the first-order system can be solved accurately, fast and efficiently using existing numerical algorithms for generalized Sylvester equations by taking the parameter derivative with respect to each parameter separately.
* Julliard and Kamenik (2004) who provide the perturbation solution equation system in tensor notation at any order k.
* Levintal (2017) who introduces permutation matrices to express the perturbation solution equation system in matrix notation up to fifth order.
Note that @rattoma already implemented the parameter derivatives of $g_x$ and $g_u$ analytically (and numerically), and I rely heavily on his work in `get_first_order_solution_params_derivs.m` (previously `getH.m`). My additions are mainly to this function and thus it is renamed to `get_perturbation_params_derivs.m`.
The basic idea of this merge request is to take the second- and third-order perturbation solution systems in Julliard and Kamenik (2004), unfold these into an equivalent matrix representation using permutation matrices as in Levintal (2017). Then extending Ratto and Iskrev (2012) one takes the derivative with respect to each parameter separately and gets a computational problem that is linear, albeit large, as it involves either solving generalized Sylvester equations or taking inverses of highly sparse matrices. I will now briefly summarize the perturbation solution system at third order and the system that results when taking the derivative with respect to parameters.
## Perturbation Solution
The following systems arise at first, second, and third order:
$(ghx): f_{x} z_{x} = f_{y_{-}^*} + f_{y_0} g_{x} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{x}= A g_{x} + f_{y_{-}^*}=0$
$(ghu): f_{z} z_{u} = f_{y_0} g_{u} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{u} + f_{u}= A g_u + f_u = 0$
$(ghxx) : A g_{xx} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{x}\right) + f_{zz} \left( z_{x} \otimes z_{x} \right) = 0$
$(ghxu) : A g_{xu} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{x} \otimes z_{u} \right) = 0$
$(ghuu) : A g_{uu} + B g_{xx} \left(g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{u} \otimes z_{u} \right) = 0$
$(ghs2) : (A+B) g_{\sigma\sigma} + \left( f_{y^{**}_{+}y^{**}_{+}} \left(g^{**}_{u} \otimes g^{**}_{u}\right) + f_{y^{**}_{+}} g^{**}_{uu}\right)vec(\Sigma) = 0$
$(ghxxx) : A g_{xxx} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{x}\right) + f_{y_{+}}g^{**}_{xx} \left(g^{*}_x \otimes g^{*}_{xx}\right)P_{x\_xx} + f_{zz} \left( z_{x} \otimes z_{xx} \right)P_{x\_xx} + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{x} \right) = 0$
$(ghxxu) : A g_{xxu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{u} \right) + f_{zz} \left( \left( z_{x} \otimes z_{xu} \right)P_{x\_xu} + \left(z_{xx} \otimes z_{u}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{x} \otimes g^{*}_{xu}\right)P_{x\_xu} + \left(g^{*}_{xx} \otimes g^{*}_{u}\right) \right) = 0$
$(ghxuu) : A g_{xuu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( \left( z_{xu} \otimes z_{u} \right)P_{xu\_u} + \left(z_{x} \otimes z_{uu}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{xu} \otimes g^{*}_{u}\right)P_{xu\_u} + \left(g^{*}_{x} \otimes g^{*}_{uu}\right) \right) = 0$
$(ghuuu) : A g_{uuu} + B g_{xxx} \left(g^{*}_{u} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{u} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( z_{u} \otimes z_{uu} \right)P_{u\_uu} + f_{y_{+}}g^{**}_{xx} \left(g^{*}_{u} \otimes g^{*}_{uu}\right)P_{u\_uu} = 0$
$(ghx\sigma\sigma) : A g_{x\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_x + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{x} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{x} \otimes z_{\sigma\sigma}\right) + F_{xu_{+}u_{+}}\left(I_{n_x} \otimes vec(\Sigma)\right) = 0$
$F_{xu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_x^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{xu_{+}} \otimes z_{u_{+}} \right)P_{xu\_u} + \left(z_{x} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{x} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
$(ghu\sigma\sigma) : A g_{u\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_{u} + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{u} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{u} \otimes z_{\sigma\sigma}\right) + F_{uu_{+}u_{+}}\left(I_{n_u} \otimes vec(\Sigma_u)\right) = 0$
$F_{uu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_u^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{uu_{+}} \otimes z_{u_{+}} \right)P_{uu\_u} + \left(z_{u} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{u} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
A and B are the common perturbation matrices:
$A = f_{y_0} + \begin{pmatrix} \underbrace{0}_{n\times n_{static}} &\vdots& \underbrace{f_{y^{**}_{+}} \cdot g^{**}_{x}}_{n \times n_{spred}} &\vdots& \underbrace{0}_{n\times n_{frwd}} \end{pmatrix}$and $B = \begin{pmatrix} \underbrace{0}_{n \times n_{static}}&\vdots & \underbrace{0}_{n \times n_{pred}} & \vdots & \underbrace{f_{y^{**}_{+}}}_{n \times n_{sfwrd}} \end{pmatrix}$
and $z=(y_{-}^{\ast}; y; y_{+}^{\ast\ast}; u)$ denotes the dynamic model variables as in `M_.lead_lag_incidence`, $y^\ast$ denote state variables, $y^{\ast\ast}$ denote forward looking variables, $y_+$ denote the variables with a lead, $y_{-}$ denote variables with a lag, $y_0$ denote variables at period t, $f$ the model equations, and $f_z$ the first-order dynamic model derivatives, $f_{zz}$ the second-order dynamic derivatives, and $f_{zzz}$ the third-order dynamic model derivatives. Then:
$z_{x} = \begin{pmatrix}I\\g_{x}\\g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{u} =\begin{pmatrix}0\\g_{u}\\g^{**}_{x} \cdot g^{*}_{u}\\I\end{pmatrix}$, $z_{u_{+}} =\begin{pmatrix}0\\0\\g^{**}_{u}\\0\end{pmatrix}$
$z_{xx} = \begin{pmatrix} 0\\g_{xx}\\g^{**}_{x} \left( g^{*}_x \otimes g^{*}_{x} \right) + g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{xu} =\begin{pmatrix}0\\g_{xu}\\g^{**}_{xx} \left( g^{*}_x \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{xu}\\0\end{pmatrix}$, $z_{uu} =\begin{pmatrix}0\\g_{uu}\\g^{**}_{xx} \left( g^{*}_u \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{uu}\\0\end{pmatrix}$,
$z_{xu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_x \otimes I \right)\\0\end{pmatrix}$, $z_{uu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_{u} \otimes I \right)\\0\end{pmatrix}$, $z_{u_{+}u_{+}} =\begin{pmatrix}0\\0\\g^{\ast\ast}_{uu}\\0\end{pmatrix}$, $z_{\sigma\sigma} = \begin{pmatrix}0\\ g_{\sigma\sigma}\\ g^{\ast\ast}_{x}g^{\ast}_{\sigma\sigma} + g^{\ast\ast}_{\sigma\sigma}\\0 \end{pmatrix}$
$P$ are permutation matrices that can be computed using Matlab's `ipermute` function.
## Parameter derivatives of perturbation solutions
First, we need the parameter derivatives of first, second, third, and fourth derivatives of the dynamic model (i.e. g1,g2,g3,g4 in dynamic files), I make use of the implicit function theorem: Let $f_{z^k}$ denote the kth derivative (wrt all dynamic variables) of the dynamic model, then let $df_{z^k}$ denote the first-derivative (wrt all model parameters) of $f_{z^k}$ evaluated at the steady state. Note that $f_{z^k}$ is a function of both the model parameters $\theta$ and of the steady state of all dynamic variables $\bar{z}$, which also depend on the parameters. Hence, implicitly $f_{z^k}=f_{z^k}(\theta,\bar{z}(\theta))$ and $df_{z^k}$ consists of two parts:
1. direct derivative wrt to all model parameters given by the preprocessor in the `_params_derivs.m` files
2. contribution of derivative of steady state of dynamic variables (wrt all model parameters): $f_{z^{k+1}} \cdot d\bar{z}$
Note that we already have functionality to compute $d\bar{z}$ analytically.
Having this, the above perturbation systems are basically equations of the following types
$AX +BXC = RHS$ or $AX = RHS$
Now when taking the derivative (wrt to one single parameter $\theta_j$), we get
$A\mathrm{d}\{X\} + B\mathrm{d}\{X\}C = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X - \mathrm{d}\{B\}XC - BX\mathrm{d}\{C\}$
or
$A\mathrm{d}\{X\} = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X$
The first one is a Sylvester type equation, the second one can be solved by taking the inverse of $A$. The only diffculty and tedious work arrises in computing (the highly sparse) derivatives of $RHS$.
***
# New functions: `
## get_perturbation_params_derivs.m`and `get_perturbation_params_derivs_numerical_objective.m`
* The parameter derivatives up to third order are computed in the new function`get_perturbation_params_derivs.m` both analytically and numerically. For numerical derivatives `get_perturbation_params_derivs_numerical_objective.m` is the objective for `fjaco.m` or `hessian_sparse.m` or `hessian.m`.
* `get_perturbation_params_derivs.m` is basically an extended version of the previous `get_first_order_solution_params_derivs.m` function.
* * `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It is used for numerical derivatives, whenever `analytic_derivation_mode=-1|-2`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
* Detailed changes:
* Most important: notation of this function is now in accordance to the k_order_solver, i.e. we do not compute derivatives of Kalman transition matrices A and B, but rather the solution matrices ghx,ghu,ghxx,ghxu,ghuu,ghs2,ghxxx,ghxxu,ghxuu,ghuuu,ghxss,ghuss in accordance with notation used in `oo_.dr`. As a byproduct at first-order, focusing on ghx and ghu instead of Kalman transition matrices A and B makes the computations slightly faster for large models (e.g. for Quest the computations were faster by a couple of seconds, not much, but okay).
* Removed use of `kstate`, see also Dynare/dynare#1653 and Dynare/dynare!1656
* Output arguments are stored in a structure `DERIVS`, there is also a flag `d2flag` that computes parameter hessians needed only in `dsge_likelihood.m`.
* Removed `kronflag` as input. `options_.analytic_derivation_mode` is now used instead of `kronflag`.
* Removed `indvar`, an index that was used to selected specific variables in the derivatives. This is not needed, as we always compute the parameter derivatives for all variables first and then select a subset of variables. The selection now takes place in other functions, like `dsge_likelihood.m`.
* Introduced some checks: (i) deterministic exogenous variables are not supported, (ii) Kronecker method only compatible with first-order approximation so reset to sylvester method, (iii) for purely backward or forward models we need to be careful with the rows in `M_.lead_la g_incidence`, (iv) if `_params_derivs.m` files are missing an error is thrown.
* For numerical derivatives, if mod file does not contain an `estimated_params_block`, a temporary one with the most important parameter information is created.
## `unfold_g4.m`
* When evaluating g3 and g4 one needs to take into account that these do not contain symmetric elements, so one needs to use `unfold_g3.m` and the new function `unfold_g4.m`. This returns an unfolded version of the same matrix (i.e. with symmetric elements).
***
# New test models
`.gitignore` and `Makefile.am` are changed accordingly. Also now it is possible to run test suite on analytic_derivatives, i.e. run `make check m/analytic_derivatives`
## `analytic_derivatives/BrockMirman_PertParamsDerivs.mod`
* This is the Brock Mirman model, where we know the exact policy function $g$ for capital and consumption. As this does not imply a nonzero $g_{\sigma\sigma}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ I added some artificial equations to get nonzero solution matrices with respect to $\sigma$. The true perturbation solution matrices $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ are then computed analytically with Matlab's symbolic toolbox and saved in `nBrockMirmanSYM.mat`. There is a preprocessor flag that recreates these analytical computations if changes are needed (and to check whether I made some errors here ;-) )
* Then solution matrices up to third order and their parameter Jacobians are then compared to the ones computed by Dynare's `k_order_solver` and by `get_perturbation_params_derivs` for all `analytic_derivation_mode`'s. There will be an error if the maximum absolute deviation is too large, i.e. for numerical derivatives (`analytic_derivation_mode=-1|-2`) the tolerance is choosen lower (around 1e-5); for analytical methods we are stricter: around 1e-13 for first-order, 1e-12 for second order, and 1e-11 for third-order.
* As a side note, this mod file also checks Dynare's `k_order_solver` algorithm and throws an error if something is wrong.
* This test model shows that the new functionality works well. And analytical derivatives perform way better and accurate than numerical ones, even for this small model.
## `analytic_derivatives/burnside_3_order_PertParamsDerivs.mod`
* This builds upon `tests/k_order_perturbation/burnside_k_order.mod` and computes the true parameter derivatives analytically by hand.
* This test model also shows that the new functionality works well.
## `analytic_derivatives/LindeTrabandt2019.mod`
* Shows that the new functionality also works for medium-sized models, i.e. a SW type model solved at third order with 35 variables (11 states). 2 shocks and 20 parameters.
* This mod file can be used to tweak the speed of the computations in the future.
* Compares numerical versus analytical parameter derivatives (for first, second and third order). Note that this model clearly shows that numerical ones are quite different than analytical ones even at first order!
## `identification/LindeTrabandt2019_xfail.mod`
* This model is a check for issue Dynare/dynare#1595, see fjaco.m below, and will fail.
* Removed `analytic_derivatives/ls2003.mod` as this mod file is neither in the testsuite nor does it work.
***
# Detailed changes in other functions
## `get_first_order_solution_params_derivs.m`
* Deleted, or actually, renamed to `get_perturbation_params_derivs.m`, as this function now is able to compute the derivatives up to third order
## `identification_numerical_objective.m`
* `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
## `dsge_likelihood.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `T`matrix anymore, but instead of the dynare solution matrix `ghx`. So we recreate `T` here as this amounts to adding some zeros and focusing on selected variables only.
* Added some checks to make sure the first-order approximation is selected.
* Removed `kron_flag` as input, as `get_perturbation_params_derivs` looks into `options_.analytic_derivation_mode` for `kron_flag`.
## `dynare_identification.m`
* make sure that setting `analytic_derivation_mode` is set both in `options_ident` and `options_`. Note that at the end of the function we restore the `options_` structure, so all changes are local. In a next merge request, I will remove the global variables to make all variables local.
## `get_identification_jacobians.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `A` and `B` matrix anymore, but instead of the dynare solution matrix `ghx` and `ghu`. So we recreate these matrices here instead of in `get_perturbation_params_derivs.m`.
* Added `str2func` for better function handles in `fjaco.m`.
## `fjaco.m`
* make `tol`an option, which can be adjusted by changing `options_.dynatol.x`for identification and parameter derivatives purposes.
* include a check and an informative error message, if numerical derivatives (two-sided finite difference method) yield errors in `resol.m` for identification and parameter derivatives purposes. This closes issue Dynare/dynare#1595.
* Changed year of copyright to 2010-2017,2019
***
# Further suggestions and questions
* Ones this is merged, I will merge request an improvement of the identification toolbox, which will work up to third order using the pruned state space. This will also remove some issues and bugs, and also I will remove global variables in this request.
* The third-order derivatives can be further improved by taking sparsity into account and use mex versions for kronecker products etc. I leave this for further testing (and if anybody actually uses this ;-) )
2019-12-17 19:17:09 +01:00
% Add derivatives for stderr and corr parameters, which are zero by construction
dKalmanA = [ zeros ( M . endo_nbr ^2 , stderrparam_nbr + corrparam_nbr ) dKalmanA ] ;
dghu = [ zeros ( M . endo_nbr * M . exo_nbr , stderrparam_nbr + corrparam_nbr ) dghu ] ;
2019-12-20 16:28:06 +01:00
Added parameter derivatives of perturbation solution up to 3 order
# Preliminary comments
I finished the identification toolbox at orders two and three using the pruned state space system, but before I merge request this, I decided to first merge the new functionality to compute parameter derivatives of perturbation solution matrices at higher orders. So after this is approved, I merge the identification toolbox.
I guess @rattoma, @sebastien, and @michel are best choices to review this.
I outline the main idea first and then provide some more detailed changes I made to the functions.
***
# Main idea
This merge request is concerned with the *analytical*computation of the parameter derivatives of first, second and third order perturbation solution matrices, i.e. using _closed-form_ expressions to efficiently compute the derivative of $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ *with respect to model parameters* $\theta$. Note that $\theta$ contains model parameters, stderr and corr parameters of shocks. stderr and corr parameters of measurement errors are not yet supported, (they can easily be included as exogenous shocks). The availability of such derivatives is beneficial in terms of more reliable analysis of model sensitivity and parameter identifiability as well as more efficient estimation methods, in particular for models solved up to third order, as it is well-known that numerical derivatives are a tricky business, especially for large models.
References for my approach are:
* Iskrev (2008, 2010) and Schmitt-Grohé and Uribe (2012, Appendix) who were the first to compute the parameter derivatives analytically at first order, however, using inefficient (sparse) Kronecker products.
* Mutschler (2015) who provides the expressions for a second-order, but again using inefficient (sparse) Kronecker products.
* Ratto and Iskrev (2012) who show how the first-order system can be solved accurately, fast and efficiently using existing numerical algorithms for generalized Sylvester equations by taking the parameter derivative with respect to each parameter separately.
* Julliard and Kamenik (2004) who provide the perturbation solution equation system in tensor notation at any order k.
* Levintal (2017) who introduces permutation matrices to express the perturbation solution equation system in matrix notation up to fifth order.
Note that @rattoma already implemented the parameter derivatives of $g_x$ and $g_u$ analytically (and numerically), and I rely heavily on his work in `get_first_order_solution_params_derivs.m` (previously `getH.m`). My additions are mainly to this function and thus it is renamed to `get_perturbation_params_derivs.m`.
The basic idea of this merge request is to take the second- and third-order perturbation solution systems in Julliard and Kamenik (2004), unfold these into an equivalent matrix representation using permutation matrices as in Levintal (2017). Then extending Ratto and Iskrev (2012) one takes the derivative with respect to each parameter separately and gets a computational problem that is linear, albeit large, as it involves either solving generalized Sylvester equations or taking inverses of highly sparse matrices. I will now briefly summarize the perturbation solution system at third order and the system that results when taking the derivative with respect to parameters.
## Perturbation Solution
The following systems arise at first, second, and third order:
$(ghx): f_{x} z_{x} = f_{y_{-}^*} + f_{y_0} g_{x} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{x}= A g_{x} + f_{y_{-}^*}=0$
$(ghu): f_{z} z_{u} = f_{y_0} g_{u} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{u} + f_{u}= A g_u + f_u = 0$
$(ghxx) : A g_{xx} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{x}\right) + f_{zz} \left( z_{x} \otimes z_{x} \right) = 0$
$(ghxu) : A g_{xu} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{x} \otimes z_{u} \right) = 0$
$(ghuu) : A g_{uu} + B g_{xx} \left(g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{u} \otimes z_{u} \right) = 0$
$(ghs2) : (A+B) g_{\sigma\sigma} + \left( f_{y^{**}_{+}y^{**}_{+}} \left(g^{**}_{u} \otimes g^{**}_{u}\right) + f_{y^{**}_{+}} g^{**}_{uu}\right)vec(\Sigma) = 0$
$(ghxxx) : A g_{xxx} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{x}\right) + f_{y_{+}}g^{**}_{xx} \left(g^{*}_x \otimes g^{*}_{xx}\right)P_{x\_xx} + f_{zz} \left( z_{x} \otimes z_{xx} \right)P_{x\_xx} + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{x} \right) = 0$
$(ghxxu) : A g_{xxu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{u} \right) + f_{zz} \left( \left( z_{x} \otimes z_{xu} \right)P_{x\_xu} + \left(z_{xx} \otimes z_{u}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{x} \otimes g^{*}_{xu}\right)P_{x\_xu} + \left(g^{*}_{xx} \otimes g^{*}_{u}\right) \right) = 0$
$(ghxuu) : A g_{xuu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( \left( z_{xu} \otimes z_{u} \right)P_{xu\_u} + \left(z_{x} \otimes z_{uu}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{xu} \otimes g^{*}_{u}\right)P_{xu\_u} + \left(g^{*}_{x} \otimes g^{*}_{uu}\right) \right) = 0$
$(ghuuu) : A g_{uuu} + B g_{xxx} \left(g^{*}_{u} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{u} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( z_{u} \otimes z_{uu} \right)P_{u\_uu} + f_{y_{+}}g^{**}_{xx} \left(g^{*}_{u} \otimes g^{*}_{uu}\right)P_{u\_uu} = 0$
$(ghx\sigma\sigma) : A g_{x\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_x + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{x} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{x} \otimes z_{\sigma\sigma}\right) + F_{xu_{+}u_{+}}\left(I_{n_x} \otimes vec(\Sigma)\right) = 0$
$F_{xu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_x^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{xu_{+}} \otimes z_{u_{+}} \right)P_{xu\_u} + \left(z_{x} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{x} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
$(ghu\sigma\sigma) : A g_{u\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_{u} + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{u} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{u} \otimes z_{\sigma\sigma}\right) + F_{uu_{+}u_{+}}\left(I_{n_u} \otimes vec(\Sigma_u)\right) = 0$
$F_{uu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_u^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{uu_{+}} \otimes z_{u_{+}} \right)P_{uu\_u} + \left(z_{u} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{u} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
A and B are the common perturbation matrices:
$A = f_{y_0} + \begin{pmatrix} \underbrace{0}_{n\times n_{static}} &\vdots& \underbrace{f_{y^{**}_{+}} \cdot g^{**}_{x}}_{n \times n_{spred}} &\vdots& \underbrace{0}_{n\times n_{frwd}} \end{pmatrix}$and $B = \begin{pmatrix} \underbrace{0}_{n \times n_{static}}&\vdots & \underbrace{0}_{n \times n_{pred}} & \vdots & \underbrace{f_{y^{**}_{+}}}_{n \times n_{sfwrd}} \end{pmatrix}$
and $z=(y_{-}^{\ast}; y; y_{+}^{\ast\ast}; u)$ denotes the dynamic model variables as in `M_.lead_lag_incidence`, $y^\ast$ denote state variables, $y^{\ast\ast}$ denote forward looking variables, $y_+$ denote the variables with a lead, $y_{-}$ denote variables with a lag, $y_0$ denote variables at period t, $f$ the model equations, and $f_z$ the first-order dynamic model derivatives, $f_{zz}$ the second-order dynamic derivatives, and $f_{zzz}$ the third-order dynamic model derivatives. Then:
$z_{x} = \begin{pmatrix}I\\g_{x}\\g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{u} =\begin{pmatrix}0\\g_{u}\\g^{**}_{x} \cdot g^{*}_{u}\\I\end{pmatrix}$, $z_{u_{+}} =\begin{pmatrix}0\\0\\g^{**}_{u}\\0\end{pmatrix}$
$z_{xx} = \begin{pmatrix} 0\\g_{xx}\\g^{**}_{x} \left( g^{*}_x \otimes g^{*}_{x} \right) + g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{xu} =\begin{pmatrix}0\\g_{xu}\\g^{**}_{xx} \left( g^{*}_x \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{xu}\\0\end{pmatrix}$, $z_{uu} =\begin{pmatrix}0\\g_{uu}\\g^{**}_{xx} \left( g^{*}_u \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{uu}\\0\end{pmatrix}$,
$z_{xu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_x \otimes I \right)\\0\end{pmatrix}$, $z_{uu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_{u} \otimes I \right)\\0\end{pmatrix}$, $z_{u_{+}u_{+}} =\begin{pmatrix}0\\0\\g^{\ast\ast}_{uu}\\0\end{pmatrix}$, $z_{\sigma\sigma} = \begin{pmatrix}0\\ g_{\sigma\sigma}\\ g^{\ast\ast}_{x}g^{\ast}_{\sigma\sigma} + g^{\ast\ast}_{\sigma\sigma}\\0 \end{pmatrix}$
$P$ are permutation matrices that can be computed using Matlab's `ipermute` function.
## Parameter derivatives of perturbation solutions
First, we need the parameter derivatives of first, second, third, and fourth derivatives of the dynamic model (i.e. g1,g2,g3,g4 in dynamic files), I make use of the implicit function theorem: Let $f_{z^k}$ denote the kth derivative (wrt all dynamic variables) of the dynamic model, then let $df_{z^k}$ denote the first-derivative (wrt all model parameters) of $f_{z^k}$ evaluated at the steady state. Note that $f_{z^k}$ is a function of both the model parameters $\theta$ and of the steady state of all dynamic variables $\bar{z}$, which also depend on the parameters. Hence, implicitly $f_{z^k}=f_{z^k}(\theta,\bar{z}(\theta))$ and $df_{z^k}$ consists of two parts:
1. direct derivative wrt to all model parameters given by the preprocessor in the `_params_derivs.m` files
2. contribution of derivative of steady state of dynamic variables (wrt all model parameters): $f_{z^{k+1}} \cdot d\bar{z}$
Note that we already have functionality to compute $d\bar{z}$ analytically.
Having this, the above perturbation systems are basically equations of the following types
$AX +BXC = RHS$ or $AX = RHS$
Now when taking the derivative (wrt to one single parameter $\theta_j$), we get
$A\mathrm{d}\{X\} + B\mathrm{d}\{X\}C = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X - \mathrm{d}\{B\}XC - BX\mathrm{d}\{C\}$
or
$A\mathrm{d}\{X\} = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X$
The first one is a Sylvester type equation, the second one can be solved by taking the inverse of $A$. The only diffculty and tedious work arrises in computing (the highly sparse) derivatives of $RHS$.
***
# New functions: `
## get_perturbation_params_derivs.m`and `get_perturbation_params_derivs_numerical_objective.m`
* The parameter derivatives up to third order are computed in the new function`get_perturbation_params_derivs.m` both analytically and numerically. For numerical derivatives `get_perturbation_params_derivs_numerical_objective.m` is the objective for `fjaco.m` or `hessian_sparse.m` or `hessian.m`.
* `get_perturbation_params_derivs.m` is basically an extended version of the previous `get_first_order_solution_params_derivs.m` function.
* * `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It is used for numerical derivatives, whenever `analytic_derivation_mode=-1|-2`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
* Detailed changes:
* Most important: notation of this function is now in accordance to the k_order_solver, i.e. we do not compute derivatives of Kalman transition matrices A and B, but rather the solution matrices ghx,ghu,ghxx,ghxu,ghuu,ghs2,ghxxx,ghxxu,ghxuu,ghuuu,ghxss,ghuss in accordance with notation used in `oo_.dr`. As a byproduct at first-order, focusing on ghx and ghu instead of Kalman transition matrices A and B makes the computations slightly faster for large models (e.g. for Quest the computations were faster by a couple of seconds, not much, but okay).
* Removed use of `kstate`, see also Dynare/dynare#1653 and Dynare/dynare!1656
* Output arguments are stored in a structure `DERIVS`, there is also a flag `d2flag` that computes parameter hessians needed only in `dsge_likelihood.m`.
* Removed `kronflag` as input. `options_.analytic_derivation_mode` is now used instead of `kronflag`.
* Removed `indvar`, an index that was used to selected specific variables in the derivatives. This is not needed, as we always compute the parameter derivatives for all variables first and then select a subset of variables. The selection now takes place in other functions, like `dsge_likelihood.m`.
* Introduced some checks: (i) deterministic exogenous variables are not supported, (ii) Kronecker method only compatible with first-order approximation so reset to sylvester method, (iii) for purely backward or forward models we need to be careful with the rows in `M_.lead_la g_incidence`, (iv) if `_params_derivs.m` files are missing an error is thrown.
* For numerical derivatives, if mod file does not contain an `estimated_params_block`, a temporary one with the most important parameter information is created.
## `unfold_g4.m`
* When evaluating g3 and g4 one needs to take into account that these do not contain symmetric elements, so one needs to use `unfold_g3.m` and the new function `unfold_g4.m`. This returns an unfolded version of the same matrix (i.e. with symmetric elements).
***
# New test models
`.gitignore` and `Makefile.am` are changed accordingly. Also now it is possible to run test suite on analytic_derivatives, i.e. run `make check m/analytic_derivatives`
## `analytic_derivatives/BrockMirman_PertParamsDerivs.mod`
* This is the Brock Mirman model, where we know the exact policy function $g$ for capital and consumption. As this does not imply a nonzero $g_{\sigma\sigma}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ I added some artificial equations to get nonzero solution matrices with respect to $\sigma$. The true perturbation solution matrices $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ are then computed analytically with Matlab's symbolic toolbox and saved in `nBrockMirmanSYM.mat`. There is a preprocessor flag that recreates these analytical computations if changes are needed (and to check whether I made some errors here ;-) )
* Then solution matrices up to third order and their parameter Jacobians are then compared to the ones computed by Dynare's `k_order_solver` and by `get_perturbation_params_derivs` for all `analytic_derivation_mode`'s. There will be an error if the maximum absolute deviation is too large, i.e. for numerical derivatives (`analytic_derivation_mode=-1|-2`) the tolerance is choosen lower (around 1e-5); for analytical methods we are stricter: around 1e-13 for first-order, 1e-12 for second order, and 1e-11 for third-order.
* As a side note, this mod file also checks Dynare's `k_order_solver` algorithm and throws an error if something is wrong.
* This test model shows that the new functionality works well. And analytical derivatives perform way better and accurate than numerical ones, even for this small model.
## `analytic_derivatives/burnside_3_order_PertParamsDerivs.mod`
* This builds upon `tests/k_order_perturbation/burnside_k_order.mod` and computes the true parameter derivatives analytically by hand.
* This test model also shows that the new functionality works well.
## `analytic_derivatives/LindeTrabandt2019.mod`
* Shows that the new functionality also works for medium-sized models, i.e. a SW type model solved at third order with 35 variables (11 states). 2 shocks and 20 parameters.
* This mod file can be used to tweak the speed of the computations in the future.
* Compares numerical versus analytical parameter derivatives (for first, second and third order). Note that this model clearly shows that numerical ones are quite different than analytical ones even at first order!
## `identification/LindeTrabandt2019_xfail.mod`
* This model is a check for issue Dynare/dynare#1595, see fjaco.m below, and will fail.
* Removed `analytic_derivatives/ls2003.mod` as this mod file is neither in the testsuite nor does it work.
***
# Detailed changes in other functions
## `get_first_order_solution_params_derivs.m`
* Deleted, or actually, renamed to `get_perturbation_params_derivs.m`, as this function now is able to compute the derivatives up to third order
## `identification_numerical_objective.m`
* `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
## `dsge_likelihood.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `T`matrix anymore, but instead of the dynare solution matrix `ghx`. So we recreate `T` here as this amounts to adding some zeros and focusing on selected variables only.
* Added some checks to make sure the first-order approximation is selected.
* Removed `kron_flag` as input, as `get_perturbation_params_derivs` looks into `options_.analytic_derivation_mode` for `kron_flag`.
## `dynare_identification.m`
* make sure that setting `analytic_derivation_mode` is set both in `options_ident` and `options_`. Note that at the end of the function we restore the `options_` structure, so all changes are local. In a next merge request, I will remove the global variables to make all variables local.
## `get_identification_jacobians.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `A` and `B` matrix anymore, but instead of the dynare solution matrix `ghx` and `ghu`. So we recreate these matrices here instead of in `get_perturbation_params_derivs.m`.
* Added `str2func` for better function handles in `fjaco.m`.
## `fjaco.m`
* make `tol`an option, which can be adjusted by changing `options_.dynatol.x`for identification and parameter derivatives purposes.
* include a check and an informative error message, if numerical derivatives (two-sided finite difference method) yield errors in `resol.m` for identification and parameter derivatives purposes. This closes issue Dynare/dynare#1595.
* Changed year of copyright to 2010-2017,2019
***
# Further suggestions and questions
* Ones this is merged, I will merge request an improvement of the identification toolbox, which will work up to third order using the pruned state space. This will also remove some issues and bugs, and also I will remove global variables in this request.
* The third-order derivatives can be further improved by taking sparsity into account and use mex versions for kronecker products etc. I leave this for further testing (and if anybody actually uses this ;-) )
2019-12-17 19:17:09 +01:00
% Compute dOm = dvec(ghu*Sigma_e*ghu') from expressions 34 in Iskrev (2010) Appendix A
dOm = kron ( I_endo , oo . dr . ghu * M . Sigma_e ) * ( commutation ( M . endo_nbr , M . exo_nbr ) * dghu ) ...
+ kron ( oo . dr . ghu , oo . dr . ghu ) * reshape ( dSigma_e , M . exo_nbr ^2 , totparam_nbr ) + kron ( oo . dr . ghu * M . Sigma_e , I_endo ) * dghu ;
2019-12-20 16:28:06 +01:00
Added parameter derivatives of perturbation solution up to 3 order
# Preliminary comments
I finished the identification toolbox at orders two and three using the pruned state space system, but before I merge request this, I decided to first merge the new functionality to compute parameter derivatives of perturbation solution matrices at higher orders. So after this is approved, I merge the identification toolbox.
I guess @rattoma, @sebastien, and @michel are best choices to review this.
I outline the main idea first and then provide some more detailed changes I made to the functions.
***
# Main idea
This merge request is concerned with the *analytical*computation of the parameter derivatives of first, second and third order perturbation solution matrices, i.e. using _closed-form_ expressions to efficiently compute the derivative of $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ *with respect to model parameters* $\theta$. Note that $\theta$ contains model parameters, stderr and corr parameters of shocks. stderr and corr parameters of measurement errors are not yet supported, (they can easily be included as exogenous shocks). The availability of such derivatives is beneficial in terms of more reliable analysis of model sensitivity and parameter identifiability as well as more efficient estimation methods, in particular for models solved up to third order, as it is well-known that numerical derivatives are a tricky business, especially for large models.
References for my approach are:
* Iskrev (2008, 2010) and Schmitt-Grohé and Uribe (2012, Appendix) who were the first to compute the parameter derivatives analytically at first order, however, using inefficient (sparse) Kronecker products.
* Mutschler (2015) who provides the expressions for a second-order, but again using inefficient (sparse) Kronecker products.
* Ratto and Iskrev (2012) who show how the first-order system can be solved accurately, fast and efficiently using existing numerical algorithms for generalized Sylvester equations by taking the parameter derivative with respect to each parameter separately.
* Julliard and Kamenik (2004) who provide the perturbation solution equation system in tensor notation at any order k.
* Levintal (2017) who introduces permutation matrices to express the perturbation solution equation system in matrix notation up to fifth order.
Note that @rattoma already implemented the parameter derivatives of $g_x$ and $g_u$ analytically (and numerically), and I rely heavily on his work in `get_first_order_solution_params_derivs.m` (previously `getH.m`). My additions are mainly to this function and thus it is renamed to `get_perturbation_params_derivs.m`.
The basic idea of this merge request is to take the second- and third-order perturbation solution systems in Julliard and Kamenik (2004), unfold these into an equivalent matrix representation using permutation matrices as in Levintal (2017). Then extending Ratto and Iskrev (2012) one takes the derivative with respect to each parameter separately and gets a computational problem that is linear, albeit large, as it involves either solving generalized Sylvester equations or taking inverses of highly sparse matrices. I will now briefly summarize the perturbation solution system at third order and the system that results when taking the derivative with respect to parameters.
## Perturbation Solution
The following systems arise at first, second, and third order:
$(ghx): f_{x} z_{x} = f_{y_{-}^*} + f_{y_0} g_{x} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{x}= A g_{x} + f_{y_{-}^*}=0$
$(ghu): f_{z} z_{u} = f_{y_0} g_{u} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{u} + f_{u}= A g_u + f_u = 0$
$(ghxx) : A g_{xx} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{x}\right) + f_{zz} \left( z_{x} \otimes z_{x} \right) = 0$
$(ghxu) : A g_{xu} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{x} \otimes z_{u} \right) = 0$
$(ghuu) : A g_{uu} + B g_{xx} \left(g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{u} \otimes z_{u} \right) = 0$
$(ghs2) : (A+B) g_{\sigma\sigma} + \left( f_{y^{**}_{+}y^{**}_{+}} \left(g^{**}_{u} \otimes g^{**}_{u}\right) + f_{y^{**}_{+}} g^{**}_{uu}\right)vec(\Sigma) = 0$
$(ghxxx) : A g_{xxx} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{x}\right) + f_{y_{+}}g^{**}_{xx} \left(g^{*}_x \otimes g^{*}_{xx}\right)P_{x\_xx} + f_{zz} \left( z_{x} \otimes z_{xx} \right)P_{x\_xx} + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{x} \right) = 0$
$(ghxxu) : A g_{xxu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{u} \right) + f_{zz} \left( \left( z_{x} \otimes z_{xu} \right)P_{x\_xu} + \left(z_{xx} \otimes z_{u}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{x} \otimes g^{*}_{xu}\right)P_{x\_xu} + \left(g^{*}_{xx} \otimes g^{*}_{u}\right) \right) = 0$
$(ghxuu) : A g_{xuu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( \left( z_{xu} \otimes z_{u} \right)P_{xu\_u} + \left(z_{x} \otimes z_{uu}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{xu} \otimes g^{*}_{u}\right)P_{xu\_u} + \left(g^{*}_{x} \otimes g^{*}_{uu}\right) \right) = 0$
$(ghuuu) : A g_{uuu} + B g_{xxx} \left(g^{*}_{u} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{u} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( z_{u} \otimes z_{uu} \right)P_{u\_uu} + f_{y_{+}}g^{**}_{xx} \left(g^{*}_{u} \otimes g^{*}_{uu}\right)P_{u\_uu} = 0$
$(ghx\sigma\sigma) : A g_{x\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_x + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{x} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{x} \otimes z_{\sigma\sigma}\right) + F_{xu_{+}u_{+}}\left(I_{n_x} \otimes vec(\Sigma)\right) = 0$
$F_{xu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_x^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{xu_{+}} \otimes z_{u_{+}} \right)P_{xu\_u} + \left(z_{x} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{x} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
$(ghu\sigma\sigma) : A g_{u\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_{u} + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{u} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{u} \otimes z_{\sigma\sigma}\right) + F_{uu_{+}u_{+}}\left(I_{n_u} \otimes vec(\Sigma_u)\right) = 0$
$F_{uu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_u^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{uu_{+}} \otimes z_{u_{+}} \right)P_{uu\_u} + \left(z_{u} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{u} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
A and B are the common perturbation matrices:
$A = f_{y_0} + \begin{pmatrix} \underbrace{0}_{n\times n_{static}} &\vdots& \underbrace{f_{y^{**}_{+}} \cdot g^{**}_{x}}_{n \times n_{spred}} &\vdots& \underbrace{0}_{n\times n_{frwd}} \end{pmatrix}$and $B = \begin{pmatrix} \underbrace{0}_{n \times n_{static}}&\vdots & \underbrace{0}_{n \times n_{pred}} & \vdots & \underbrace{f_{y^{**}_{+}}}_{n \times n_{sfwrd}} \end{pmatrix}$
and $z=(y_{-}^{\ast}; y; y_{+}^{\ast\ast}; u)$ denotes the dynamic model variables as in `M_.lead_lag_incidence`, $y^\ast$ denote state variables, $y^{\ast\ast}$ denote forward looking variables, $y_+$ denote the variables with a lead, $y_{-}$ denote variables with a lag, $y_0$ denote variables at period t, $f$ the model equations, and $f_z$ the first-order dynamic model derivatives, $f_{zz}$ the second-order dynamic derivatives, and $f_{zzz}$ the third-order dynamic model derivatives. Then:
$z_{x} = \begin{pmatrix}I\\g_{x}\\g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{u} =\begin{pmatrix}0\\g_{u}\\g^{**}_{x} \cdot g^{*}_{u}\\I\end{pmatrix}$, $z_{u_{+}} =\begin{pmatrix}0\\0\\g^{**}_{u}\\0\end{pmatrix}$
$z_{xx} = \begin{pmatrix} 0\\g_{xx}\\g^{**}_{x} \left( g^{*}_x \otimes g^{*}_{x} \right) + g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{xu} =\begin{pmatrix}0\\g_{xu}\\g^{**}_{xx} \left( g^{*}_x \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{xu}\\0\end{pmatrix}$, $z_{uu} =\begin{pmatrix}0\\g_{uu}\\g^{**}_{xx} \left( g^{*}_u \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{uu}\\0\end{pmatrix}$,
$z_{xu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_x \otimes I \right)\\0\end{pmatrix}$, $z_{uu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_{u} \otimes I \right)\\0\end{pmatrix}$, $z_{u_{+}u_{+}} =\begin{pmatrix}0\\0\\g^{\ast\ast}_{uu}\\0\end{pmatrix}$, $z_{\sigma\sigma} = \begin{pmatrix}0\\ g_{\sigma\sigma}\\ g^{\ast\ast}_{x}g^{\ast}_{\sigma\sigma} + g^{\ast\ast}_{\sigma\sigma}\\0 \end{pmatrix}$
$P$ are permutation matrices that can be computed using Matlab's `ipermute` function.
## Parameter derivatives of perturbation solutions
First, we need the parameter derivatives of first, second, third, and fourth derivatives of the dynamic model (i.e. g1,g2,g3,g4 in dynamic files), I make use of the implicit function theorem: Let $f_{z^k}$ denote the kth derivative (wrt all dynamic variables) of the dynamic model, then let $df_{z^k}$ denote the first-derivative (wrt all model parameters) of $f_{z^k}$ evaluated at the steady state. Note that $f_{z^k}$ is a function of both the model parameters $\theta$ and of the steady state of all dynamic variables $\bar{z}$, which also depend on the parameters. Hence, implicitly $f_{z^k}=f_{z^k}(\theta,\bar{z}(\theta))$ and $df_{z^k}$ consists of two parts:
1. direct derivative wrt to all model parameters given by the preprocessor in the `_params_derivs.m` files
2. contribution of derivative of steady state of dynamic variables (wrt all model parameters): $f_{z^{k+1}} \cdot d\bar{z}$
Note that we already have functionality to compute $d\bar{z}$ analytically.
Having this, the above perturbation systems are basically equations of the following types
$AX +BXC = RHS$ or $AX = RHS$
Now when taking the derivative (wrt to one single parameter $\theta_j$), we get
$A\mathrm{d}\{X\} + B\mathrm{d}\{X\}C = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X - \mathrm{d}\{B\}XC - BX\mathrm{d}\{C\}$
or
$A\mathrm{d}\{X\} = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X$
The first one is a Sylvester type equation, the second one can be solved by taking the inverse of $A$. The only diffculty and tedious work arrises in computing (the highly sparse) derivatives of $RHS$.
***
# New functions: `
## get_perturbation_params_derivs.m`and `get_perturbation_params_derivs_numerical_objective.m`
* The parameter derivatives up to third order are computed in the new function`get_perturbation_params_derivs.m` both analytically and numerically. For numerical derivatives `get_perturbation_params_derivs_numerical_objective.m` is the objective for `fjaco.m` or `hessian_sparse.m` or `hessian.m`.
* `get_perturbation_params_derivs.m` is basically an extended version of the previous `get_first_order_solution_params_derivs.m` function.
* * `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It is used for numerical derivatives, whenever `analytic_derivation_mode=-1|-2`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
* Detailed changes:
* Most important: notation of this function is now in accordance to the k_order_solver, i.e. we do not compute derivatives of Kalman transition matrices A and B, but rather the solution matrices ghx,ghu,ghxx,ghxu,ghuu,ghs2,ghxxx,ghxxu,ghxuu,ghuuu,ghxss,ghuss in accordance with notation used in `oo_.dr`. As a byproduct at first-order, focusing on ghx and ghu instead of Kalman transition matrices A and B makes the computations slightly faster for large models (e.g. for Quest the computations were faster by a couple of seconds, not much, but okay).
* Removed use of `kstate`, see also Dynare/dynare#1653 and Dynare/dynare!1656
* Output arguments are stored in a structure `DERIVS`, there is also a flag `d2flag` that computes parameter hessians needed only in `dsge_likelihood.m`.
* Removed `kronflag` as input. `options_.analytic_derivation_mode` is now used instead of `kronflag`.
* Removed `indvar`, an index that was used to selected specific variables in the derivatives. This is not needed, as we always compute the parameter derivatives for all variables first and then select a subset of variables. The selection now takes place in other functions, like `dsge_likelihood.m`.
* Introduced some checks: (i) deterministic exogenous variables are not supported, (ii) Kronecker method only compatible with first-order approximation so reset to sylvester method, (iii) for purely backward or forward models we need to be careful with the rows in `M_.lead_la g_incidence`, (iv) if `_params_derivs.m` files are missing an error is thrown.
* For numerical derivatives, if mod file does not contain an `estimated_params_block`, a temporary one with the most important parameter information is created.
## `unfold_g4.m`
* When evaluating g3 and g4 one needs to take into account that these do not contain symmetric elements, so one needs to use `unfold_g3.m` and the new function `unfold_g4.m`. This returns an unfolded version of the same matrix (i.e. with symmetric elements).
***
# New test models
`.gitignore` and `Makefile.am` are changed accordingly. Also now it is possible to run test suite on analytic_derivatives, i.e. run `make check m/analytic_derivatives`
## `analytic_derivatives/BrockMirman_PertParamsDerivs.mod`
* This is the Brock Mirman model, where we know the exact policy function $g$ for capital and consumption. As this does not imply a nonzero $g_{\sigma\sigma}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ I added some artificial equations to get nonzero solution matrices with respect to $\sigma$. The true perturbation solution matrices $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ are then computed analytically with Matlab's symbolic toolbox and saved in `nBrockMirmanSYM.mat`. There is a preprocessor flag that recreates these analytical computations if changes are needed (and to check whether I made some errors here ;-) )
* Then solution matrices up to third order and their parameter Jacobians are then compared to the ones computed by Dynare's `k_order_solver` and by `get_perturbation_params_derivs` for all `analytic_derivation_mode`'s. There will be an error if the maximum absolute deviation is too large, i.e. for numerical derivatives (`analytic_derivation_mode=-1|-2`) the tolerance is choosen lower (around 1e-5); for analytical methods we are stricter: around 1e-13 for first-order, 1e-12 for second order, and 1e-11 for third-order.
* As a side note, this mod file also checks Dynare's `k_order_solver` algorithm and throws an error if something is wrong.
* This test model shows that the new functionality works well. And analytical derivatives perform way better and accurate than numerical ones, even for this small model.
## `analytic_derivatives/burnside_3_order_PertParamsDerivs.mod`
* This builds upon `tests/k_order_perturbation/burnside_k_order.mod` and computes the true parameter derivatives analytically by hand.
* This test model also shows that the new functionality works well.
## `analytic_derivatives/LindeTrabandt2019.mod`
* Shows that the new functionality also works for medium-sized models, i.e. a SW type model solved at third order with 35 variables (11 states). 2 shocks and 20 parameters.
* This mod file can be used to tweak the speed of the computations in the future.
* Compares numerical versus analytical parameter derivatives (for first, second and third order). Note that this model clearly shows that numerical ones are quite different than analytical ones even at first order!
## `identification/LindeTrabandt2019_xfail.mod`
* This model is a check for issue Dynare/dynare#1595, see fjaco.m below, and will fail.
* Removed `analytic_derivatives/ls2003.mod` as this mod file is neither in the testsuite nor does it work.
***
# Detailed changes in other functions
## `get_first_order_solution_params_derivs.m`
* Deleted, or actually, renamed to `get_perturbation_params_derivs.m`, as this function now is able to compute the derivatives up to third order
## `identification_numerical_objective.m`
* `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
## `dsge_likelihood.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `T`matrix anymore, but instead of the dynare solution matrix `ghx`. So we recreate `T` here as this amounts to adding some zeros and focusing on selected variables only.
* Added some checks to make sure the first-order approximation is selected.
* Removed `kron_flag` as input, as `get_perturbation_params_derivs` looks into `options_.analytic_derivation_mode` for `kron_flag`.
## `dynare_identification.m`
* make sure that setting `analytic_derivation_mode` is set both in `options_ident` and `options_`. Note that at the end of the function we restore the `options_` structure, so all changes are local. In a next merge request, I will remove the global variables to make all variables local.
## `get_identification_jacobians.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `A` and `B` matrix anymore, but instead of the dynare solution matrix `ghx` and `ghu`. So we recreate these matrices here instead of in `get_perturbation_params_derivs.m`.
* Added `str2func` for better function handles in `fjaco.m`.
## `fjaco.m`
* make `tol`an option, which can be adjusted by changing `options_.dynatol.x`for identification and parameter derivatives purposes.
* include a check and an informative error message, if numerical derivatives (two-sided finite difference method) yield errors in `resol.m` for identification and parameter derivatives purposes. This closes issue Dynare/dynare#1595.
* Changed year of copyright to 2010-2017,2019
***
# Further suggestions and questions
* Ones this is merged, I will merge request an improvement of the identification toolbox, which will work up to third order using the pruned state space. This will also remove some issues and bugs, and also I will remove global variables in this request.
* The third-order derivatives can be further improved by taking sparsity into account and use mex versions for kronecker products etc. I leave this for further testing (and if anybody actually uses this ;-) )
2019-12-17 19:17:09 +01:00
% Put into tensor notation
dKalmanA = reshape ( dKalmanA , M . endo_nbr , M . endo_nbr , totparam_nbr ) ;
dghx = dKalmanA ( : , M . nstatic + ( 1 : M . nspred ) , stderrparam_nbr + corrparam_nbr + 1 : end ) ; %get rid of zeros and focus on modparams only
dghu = reshape ( dghu , M . endo_nbr , M . exo_nbr , totparam_nbr ) ;
dghu = dghu ( : , : , stderrparam_nbr + corrparam_nbr + 1 : end ) ; %focus only on modparams
dOm = reshape ( dOm , M . endo_nbr , M . endo_nbr , totparam_nbr ) ;
clear dF_dKalmanAghx Fp dGAM0 dGAM1 dGAM2 dGAM3 MM invMM I_endo
2019-12-20 16:28:06 +01:00
Added parameter derivatives of perturbation solution up to 3 order
# Preliminary comments
I finished the identification toolbox at orders two and three using the pruned state space system, but before I merge request this, I decided to first merge the new functionality to compute parameter derivatives of perturbation solution matrices at higher orders. So after this is approved, I merge the identification toolbox.
I guess @rattoma, @sebastien, and @michel are best choices to review this.
I outline the main idea first and then provide some more detailed changes I made to the functions.
***
# Main idea
This merge request is concerned with the *analytical*computation of the parameter derivatives of first, second and third order perturbation solution matrices, i.e. using _closed-form_ expressions to efficiently compute the derivative of $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ *with respect to model parameters* $\theta$. Note that $\theta$ contains model parameters, stderr and corr parameters of shocks. stderr and corr parameters of measurement errors are not yet supported, (they can easily be included as exogenous shocks). The availability of such derivatives is beneficial in terms of more reliable analysis of model sensitivity and parameter identifiability as well as more efficient estimation methods, in particular for models solved up to third order, as it is well-known that numerical derivatives are a tricky business, especially for large models.
References for my approach are:
* Iskrev (2008, 2010) and Schmitt-Grohé and Uribe (2012, Appendix) who were the first to compute the parameter derivatives analytically at first order, however, using inefficient (sparse) Kronecker products.
* Mutschler (2015) who provides the expressions for a second-order, but again using inefficient (sparse) Kronecker products.
* Ratto and Iskrev (2012) who show how the first-order system can be solved accurately, fast and efficiently using existing numerical algorithms for generalized Sylvester equations by taking the parameter derivative with respect to each parameter separately.
* Julliard and Kamenik (2004) who provide the perturbation solution equation system in tensor notation at any order k.
* Levintal (2017) who introduces permutation matrices to express the perturbation solution equation system in matrix notation up to fifth order.
Note that @rattoma already implemented the parameter derivatives of $g_x$ and $g_u$ analytically (and numerically), and I rely heavily on his work in `get_first_order_solution_params_derivs.m` (previously `getH.m`). My additions are mainly to this function and thus it is renamed to `get_perturbation_params_derivs.m`.
The basic idea of this merge request is to take the second- and third-order perturbation solution systems in Julliard and Kamenik (2004), unfold these into an equivalent matrix representation using permutation matrices as in Levintal (2017). Then extending Ratto and Iskrev (2012) one takes the derivative with respect to each parameter separately and gets a computational problem that is linear, albeit large, as it involves either solving generalized Sylvester equations or taking inverses of highly sparse matrices. I will now briefly summarize the perturbation solution system at third order and the system that results when taking the derivative with respect to parameters.
## Perturbation Solution
The following systems arise at first, second, and third order:
$(ghx): f_{x} z_{x} = f_{y_{-}^*} + f_{y_0} g_{x} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{x}= A g_{x} + f_{y_{-}^*}=0$
$(ghu): f_{z} z_{u} = f_{y_0} g_{u} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{u} + f_{u}= A g_u + f_u = 0$
$(ghxx) : A g_{xx} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{x}\right) + f_{zz} \left( z_{x} \otimes z_{x} \right) = 0$
$(ghxu) : A g_{xu} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{x} \otimes z_{u} \right) = 0$
$(ghuu) : A g_{uu} + B g_{xx} \left(g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{u} \otimes z_{u} \right) = 0$
$(ghs2) : (A+B) g_{\sigma\sigma} + \left( f_{y^{**}_{+}y^{**}_{+}} \left(g^{**}_{u} \otimes g^{**}_{u}\right) + f_{y^{**}_{+}} g^{**}_{uu}\right)vec(\Sigma) = 0$
$(ghxxx) : A g_{xxx} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{x}\right) + f_{y_{+}}g^{**}_{xx} \left(g^{*}_x \otimes g^{*}_{xx}\right)P_{x\_xx} + f_{zz} \left( z_{x} \otimes z_{xx} \right)P_{x\_xx} + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{x} \right) = 0$
$(ghxxu) : A g_{xxu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{u} \right) + f_{zz} \left( \left( z_{x} \otimes z_{xu} \right)P_{x\_xu} + \left(z_{xx} \otimes z_{u}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{x} \otimes g^{*}_{xu}\right)P_{x\_xu} + \left(g^{*}_{xx} \otimes g^{*}_{u}\right) \right) = 0$
$(ghxuu) : A g_{xuu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( \left( z_{xu} \otimes z_{u} \right)P_{xu\_u} + \left(z_{x} \otimes z_{uu}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{xu} \otimes g^{*}_{u}\right)P_{xu\_u} + \left(g^{*}_{x} \otimes g^{*}_{uu}\right) \right) = 0$
$(ghuuu) : A g_{uuu} + B g_{xxx} \left(g^{*}_{u} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{u} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( z_{u} \otimes z_{uu} \right)P_{u\_uu} + f_{y_{+}}g^{**}_{xx} \left(g^{*}_{u} \otimes g^{*}_{uu}\right)P_{u\_uu} = 0$
$(ghx\sigma\sigma) : A g_{x\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_x + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{x} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{x} \otimes z_{\sigma\sigma}\right) + F_{xu_{+}u_{+}}\left(I_{n_x} \otimes vec(\Sigma)\right) = 0$
$F_{xu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_x^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{xu_{+}} \otimes z_{u_{+}} \right)P_{xu\_u} + \left(z_{x} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{x} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
$(ghu\sigma\sigma) : A g_{u\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_{u} + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{u} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{u} \otimes z_{\sigma\sigma}\right) + F_{uu_{+}u_{+}}\left(I_{n_u} \otimes vec(\Sigma_u)\right) = 0$
$F_{uu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_u^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{uu_{+}} \otimes z_{u_{+}} \right)P_{uu\_u} + \left(z_{u} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{u} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
A and B are the common perturbation matrices:
$A = f_{y_0} + \begin{pmatrix} \underbrace{0}_{n\times n_{static}} &\vdots& \underbrace{f_{y^{**}_{+}} \cdot g^{**}_{x}}_{n \times n_{spred}} &\vdots& \underbrace{0}_{n\times n_{frwd}} \end{pmatrix}$and $B = \begin{pmatrix} \underbrace{0}_{n \times n_{static}}&\vdots & \underbrace{0}_{n \times n_{pred}} & \vdots & \underbrace{f_{y^{**}_{+}}}_{n \times n_{sfwrd}} \end{pmatrix}$
and $z=(y_{-}^{\ast}; y; y_{+}^{\ast\ast}; u)$ denotes the dynamic model variables as in `M_.lead_lag_incidence`, $y^\ast$ denote state variables, $y^{\ast\ast}$ denote forward looking variables, $y_+$ denote the variables with a lead, $y_{-}$ denote variables with a lag, $y_0$ denote variables at period t, $f$ the model equations, and $f_z$ the first-order dynamic model derivatives, $f_{zz}$ the second-order dynamic derivatives, and $f_{zzz}$ the third-order dynamic model derivatives. Then:
$z_{x} = \begin{pmatrix}I\\g_{x}\\g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{u} =\begin{pmatrix}0\\g_{u}\\g^{**}_{x} \cdot g^{*}_{u}\\I\end{pmatrix}$, $z_{u_{+}} =\begin{pmatrix}0\\0\\g^{**}_{u}\\0\end{pmatrix}$
$z_{xx} = \begin{pmatrix} 0\\g_{xx}\\g^{**}_{x} \left( g^{*}_x \otimes g^{*}_{x} \right) + g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{xu} =\begin{pmatrix}0\\g_{xu}\\g^{**}_{xx} \left( g^{*}_x \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{xu}\\0\end{pmatrix}$, $z_{uu} =\begin{pmatrix}0\\g_{uu}\\g^{**}_{xx} \left( g^{*}_u \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{uu}\\0\end{pmatrix}$,
$z_{xu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_x \otimes I \right)\\0\end{pmatrix}$, $z_{uu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_{u} \otimes I \right)\\0\end{pmatrix}$, $z_{u_{+}u_{+}} =\begin{pmatrix}0\\0\\g^{\ast\ast}_{uu}\\0\end{pmatrix}$, $z_{\sigma\sigma} = \begin{pmatrix}0\\ g_{\sigma\sigma}\\ g^{\ast\ast}_{x}g^{\ast}_{\sigma\sigma} + g^{\ast\ast}_{\sigma\sigma}\\0 \end{pmatrix}$
$P$ are permutation matrices that can be computed using Matlab's `ipermute` function.
## Parameter derivatives of perturbation solutions
First, we need the parameter derivatives of first, second, third, and fourth derivatives of the dynamic model (i.e. g1,g2,g3,g4 in dynamic files), I make use of the implicit function theorem: Let $f_{z^k}$ denote the kth derivative (wrt all dynamic variables) of the dynamic model, then let $df_{z^k}$ denote the first-derivative (wrt all model parameters) of $f_{z^k}$ evaluated at the steady state. Note that $f_{z^k}$ is a function of both the model parameters $\theta$ and of the steady state of all dynamic variables $\bar{z}$, which also depend on the parameters. Hence, implicitly $f_{z^k}=f_{z^k}(\theta,\bar{z}(\theta))$ and $df_{z^k}$ consists of two parts:
1. direct derivative wrt to all model parameters given by the preprocessor in the `_params_derivs.m` files
2. contribution of derivative of steady state of dynamic variables (wrt all model parameters): $f_{z^{k+1}} \cdot d\bar{z}$
Note that we already have functionality to compute $d\bar{z}$ analytically.
Having this, the above perturbation systems are basically equations of the following types
$AX +BXC = RHS$ or $AX = RHS$
Now when taking the derivative (wrt to one single parameter $\theta_j$), we get
$A\mathrm{d}\{X\} + B\mathrm{d}\{X\}C = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X - \mathrm{d}\{B\}XC - BX\mathrm{d}\{C\}$
or
$A\mathrm{d}\{X\} = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X$
The first one is a Sylvester type equation, the second one can be solved by taking the inverse of $A$. The only diffculty and tedious work arrises in computing (the highly sparse) derivatives of $RHS$.
***
# New functions: `
## get_perturbation_params_derivs.m`and `get_perturbation_params_derivs_numerical_objective.m`
* The parameter derivatives up to third order are computed in the new function`get_perturbation_params_derivs.m` both analytically and numerically. For numerical derivatives `get_perturbation_params_derivs_numerical_objective.m` is the objective for `fjaco.m` or `hessian_sparse.m` or `hessian.m`.
* `get_perturbation_params_derivs.m` is basically an extended version of the previous `get_first_order_solution_params_derivs.m` function.
* * `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It is used for numerical derivatives, whenever `analytic_derivation_mode=-1|-2`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
* Detailed changes:
* Most important: notation of this function is now in accordance to the k_order_solver, i.e. we do not compute derivatives of Kalman transition matrices A and B, but rather the solution matrices ghx,ghu,ghxx,ghxu,ghuu,ghs2,ghxxx,ghxxu,ghxuu,ghuuu,ghxss,ghuss in accordance with notation used in `oo_.dr`. As a byproduct at first-order, focusing on ghx and ghu instead of Kalman transition matrices A and B makes the computations slightly faster for large models (e.g. for Quest the computations were faster by a couple of seconds, not much, but okay).
* Removed use of `kstate`, see also Dynare/dynare#1653 and Dynare/dynare!1656
* Output arguments are stored in a structure `DERIVS`, there is also a flag `d2flag` that computes parameter hessians needed only in `dsge_likelihood.m`.
* Removed `kronflag` as input. `options_.analytic_derivation_mode` is now used instead of `kronflag`.
* Removed `indvar`, an index that was used to selected specific variables in the derivatives. This is not needed, as we always compute the parameter derivatives for all variables first and then select a subset of variables. The selection now takes place in other functions, like `dsge_likelihood.m`.
* Introduced some checks: (i) deterministic exogenous variables are not supported, (ii) Kronecker method only compatible with first-order approximation so reset to sylvester method, (iii) for purely backward or forward models we need to be careful with the rows in `M_.lead_la g_incidence`, (iv) if `_params_derivs.m` files are missing an error is thrown.
* For numerical derivatives, if mod file does not contain an `estimated_params_block`, a temporary one with the most important parameter information is created.
## `unfold_g4.m`
* When evaluating g3 and g4 one needs to take into account that these do not contain symmetric elements, so one needs to use `unfold_g3.m` and the new function `unfold_g4.m`. This returns an unfolded version of the same matrix (i.e. with symmetric elements).
***
# New test models
`.gitignore` and `Makefile.am` are changed accordingly. Also now it is possible to run test suite on analytic_derivatives, i.e. run `make check m/analytic_derivatives`
## `analytic_derivatives/BrockMirman_PertParamsDerivs.mod`
* This is the Brock Mirman model, where we know the exact policy function $g$ for capital and consumption. As this does not imply a nonzero $g_{\sigma\sigma}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ I added some artificial equations to get nonzero solution matrices with respect to $\sigma$. The true perturbation solution matrices $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ are then computed analytically with Matlab's symbolic toolbox and saved in `nBrockMirmanSYM.mat`. There is a preprocessor flag that recreates these analytical computations if changes are needed (and to check whether I made some errors here ;-) )
* Then solution matrices up to third order and their parameter Jacobians are then compared to the ones computed by Dynare's `k_order_solver` and by `get_perturbation_params_derivs` for all `analytic_derivation_mode`'s. There will be an error if the maximum absolute deviation is too large, i.e. for numerical derivatives (`analytic_derivation_mode=-1|-2`) the tolerance is choosen lower (around 1e-5); for analytical methods we are stricter: around 1e-13 for first-order, 1e-12 for second order, and 1e-11 for third-order.
* As a side note, this mod file also checks Dynare's `k_order_solver` algorithm and throws an error if something is wrong.
* This test model shows that the new functionality works well. And analytical derivatives perform way better and accurate than numerical ones, even for this small model.
## `analytic_derivatives/burnside_3_order_PertParamsDerivs.mod`
* This builds upon `tests/k_order_perturbation/burnside_k_order.mod` and computes the true parameter derivatives analytically by hand.
* This test model also shows that the new functionality works well.
## `analytic_derivatives/LindeTrabandt2019.mod`
* Shows that the new functionality also works for medium-sized models, i.e. a SW type model solved at third order with 35 variables (11 states). 2 shocks and 20 parameters.
* This mod file can be used to tweak the speed of the computations in the future.
* Compares numerical versus analytical parameter derivatives (for first, second and third order). Note that this model clearly shows that numerical ones are quite different than analytical ones even at first order!
## `identification/LindeTrabandt2019_xfail.mod`
* This model is a check for issue Dynare/dynare#1595, see fjaco.m below, and will fail.
* Removed `analytic_derivatives/ls2003.mod` as this mod file is neither in the testsuite nor does it work.
***
# Detailed changes in other functions
## `get_first_order_solution_params_derivs.m`
* Deleted, or actually, renamed to `get_perturbation_params_derivs.m`, as this function now is able to compute the derivatives up to third order
## `identification_numerical_objective.m`
* `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
## `dsge_likelihood.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `T`matrix anymore, but instead of the dynare solution matrix `ghx`. So we recreate `T` here as this amounts to adding some zeros and focusing on selected variables only.
* Added some checks to make sure the first-order approximation is selected.
* Removed `kron_flag` as input, as `get_perturbation_params_derivs` looks into `options_.analytic_derivation_mode` for `kron_flag`.
## `dynare_identification.m`
* make sure that setting `analytic_derivation_mode` is set both in `options_ident` and `options_`. Note that at the end of the function we restore the `options_` structure, so all changes are local. In a next merge request, I will remove the global variables to make all variables local.
## `get_identification_jacobians.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `A` and `B` matrix anymore, but instead of the dynare solution matrix `ghx` and `ghu`. So we recreate these matrices here instead of in `get_perturbation_params_derivs.m`.
* Added `str2func` for better function handles in `fjaco.m`.
## `fjaco.m`
* make `tol`an option, which can be adjusted by changing `options_.dynatol.x`for identification and parameter derivatives purposes.
* include a check and an informative error message, if numerical derivatives (two-sided finite difference method) yield errors in `resol.m` for identification and parameter derivatives purposes. This closes issue Dynare/dynare#1595.
* Changed year of copyright to 2010-2017,2019
***
# Further suggestions and questions
* Ones this is merged, I will merge request an improvement of the identification toolbox, which will work up to third order using the pruned state space. This will also remove some issues and bugs, and also I will remove global variables in this request.
* The third-order derivatives can be further improved by taking sparsity into account and use mex versions for kronecker products etc. I leave this for further testing (and if anybody actually uses this ;-) )
2019-12-17 19:17:09 +01:00
elseif ( analytic_derivation_mode == 0 || analytic_derivation_mode == - 2 )
% Here we make use of more efficient generalized sylvester equations
% Notation: ghx_ = oo.dr.ghx(idx_states,:), ghx0 = oo.dr.ghx(kcurr~=0,:), ghxp = oo.dr.ghx(klead~=0,:)
% Note that at first-order we have the following expressions, which are (numerically) zero:
% * for ghx: g1*zx = fyp*ghxp*ghx_ + fy0*ghx0 + fy_ = A*ghx + fy_ = 0
% Taking the differential yields a generalized sylvester equation to get dghx: A*dghx + B*dghx*ghx_ = -dg1*zx
% * for ghu: g1*zu = A*ghu + fu = 0
% Taking the differential yields an invertible equation to get dghu: A*dghu = -(dfu + dA*ghu)
% INITIALIZATIONS
% Note that A and B are the perturbation matrices (NOT the Kalman transition matrices)!
A = zeros ( M . endo_nbr , M . endo_nbr ) ;
A ( : , kcurr ~= 0 ) = g1 ( : , nonzeros ( kcurr ) ) ;
A ( : , idx_states ) = A ( : , idx_states ) + g1 ( : , nonzeros ( klead ) ) * oo . dr . ghx ( klead ~= 0 , : ) ;
B = zeros ( M . endo_nbr , M . endo_nbr ) ;
B ( : , M . nstatic + M . npred + 1 : end ) = g1 ( : , nonzeros ( klead ) ) ;
zx = [ eye ( M . nspred ) ;
2019-12-20 16:28:06 +01:00
oo . dr . ghx ( kcurr ~= 0 , : ) ;
oo . dr . ghx ( klead ~= 0 , : ) * oo . dr . ghx ( idx_states , : ) ;
zeros ( M . exo_nbr , M . nspred ) ] ;
Added parameter derivatives of perturbation solution up to 3 order
# Preliminary comments
I finished the identification toolbox at orders two and three using the pruned state space system, but before I merge request this, I decided to first merge the new functionality to compute parameter derivatives of perturbation solution matrices at higher orders. So after this is approved, I merge the identification toolbox.
I guess @rattoma, @sebastien, and @michel are best choices to review this.
I outline the main idea first and then provide some more detailed changes I made to the functions.
***
# Main idea
This merge request is concerned with the *analytical*computation of the parameter derivatives of first, second and third order perturbation solution matrices, i.e. using _closed-form_ expressions to efficiently compute the derivative of $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ *with respect to model parameters* $\theta$. Note that $\theta$ contains model parameters, stderr and corr parameters of shocks. stderr and corr parameters of measurement errors are not yet supported, (they can easily be included as exogenous shocks). The availability of such derivatives is beneficial in terms of more reliable analysis of model sensitivity and parameter identifiability as well as more efficient estimation methods, in particular for models solved up to third order, as it is well-known that numerical derivatives are a tricky business, especially for large models.
References for my approach are:
* Iskrev (2008, 2010) and Schmitt-Grohé and Uribe (2012, Appendix) who were the first to compute the parameter derivatives analytically at first order, however, using inefficient (sparse) Kronecker products.
* Mutschler (2015) who provides the expressions for a second-order, but again using inefficient (sparse) Kronecker products.
* Ratto and Iskrev (2012) who show how the first-order system can be solved accurately, fast and efficiently using existing numerical algorithms for generalized Sylvester equations by taking the parameter derivative with respect to each parameter separately.
* Julliard and Kamenik (2004) who provide the perturbation solution equation system in tensor notation at any order k.
* Levintal (2017) who introduces permutation matrices to express the perturbation solution equation system in matrix notation up to fifth order.
Note that @rattoma already implemented the parameter derivatives of $g_x$ and $g_u$ analytically (and numerically), and I rely heavily on his work in `get_first_order_solution_params_derivs.m` (previously `getH.m`). My additions are mainly to this function and thus it is renamed to `get_perturbation_params_derivs.m`.
The basic idea of this merge request is to take the second- and third-order perturbation solution systems in Julliard and Kamenik (2004), unfold these into an equivalent matrix representation using permutation matrices as in Levintal (2017). Then extending Ratto and Iskrev (2012) one takes the derivative with respect to each parameter separately and gets a computational problem that is linear, albeit large, as it involves either solving generalized Sylvester equations or taking inverses of highly sparse matrices. I will now briefly summarize the perturbation solution system at third order and the system that results when taking the derivative with respect to parameters.
## Perturbation Solution
The following systems arise at first, second, and third order:
$(ghx): f_{x} z_{x} = f_{y_{-}^*} + f_{y_0} g_{x} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{x}= A g_{x} + f_{y_{-}^*}=0$
$(ghu): f_{z} z_{u} = f_{y_0} g_{u} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{u} + f_{u}= A g_u + f_u = 0$
$(ghxx) : A g_{xx} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{x}\right) + f_{zz} \left( z_{x} \otimes z_{x} \right) = 0$
$(ghxu) : A g_{xu} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{x} \otimes z_{u} \right) = 0$
$(ghuu) : A g_{uu} + B g_{xx} \left(g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{u} \otimes z_{u} \right) = 0$
$(ghs2) : (A+B) g_{\sigma\sigma} + \left( f_{y^{**}_{+}y^{**}_{+}} \left(g^{**}_{u} \otimes g^{**}_{u}\right) + f_{y^{**}_{+}} g^{**}_{uu}\right)vec(\Sigma) = 0$
$(ghxxx) : A g_{xxx} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{x}\right) + f_{y_{+}}g^{**}_{xx} \left(g^{*}_x \otimes g^{*}_{xx}\right)P_{x\_xx} + f_{zz} \left( z_{x} \otimes z_{xx} \right)P_{x\_xx} + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{x} \right) = 0$
$(ghxxu) : A g_{xxu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{u} \right) + f_{zz} \left( \left( z_{x} \otimes z_{xu} \right)P_{x\_xu} + \left(z_{xx} \otimes z_{u}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{x} \otimes g^{*}_{xu}\right)P_{x\_xu} + \left(g^{*}_{xx} \otimes g^{*}_{u}\right) \right) = 0$
$(ghxuu) : A g_{xuu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( \left( z_{xu} \otimes z_{u} \right)P_{xu\_u} + \left(z_{x} \otimes z_{uu}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{xu} \otimes g^{*}_{u}\right)P_{xu\_u} + \left(g^{*}_{x} \otimes g^{*}_{uu}\right) \right) = 0$
$(ghuuu) : A g_{uuu} + B g_{xxx} \left(g^{*}_{u} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{u} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( z_{u} \otimes z_{uu} \right)P_{u\_uu} + f_{y_{+}}g^{**}_{xx} \left(g^{*}_{u} \otimes g^{*}_{uu}\right)P_{u\_uu} = 0$
$(ghx\sigma\sigma) : A g_{x\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_x + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{x} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{x} \otimes z_{\sigma\sigma}\right) + F_{xu_{+}u_{+}}\left(I_{n_x} \otimes vec(\Sigma)\right) = 0$
$F_{xu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_x^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{xu_{+}} \otimes z_{u_{+}} \right)P_{xu\_u} + \left(z_{x} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{x} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
$(ghu\sigma\sigma) : A g_{u\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_{u} + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{u} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{u} \otimes z_{\sigma\sigma}\right) + F_{uu_{+}u_{+}}\left(I_{n_u} \otimes vec(\Sigma_u)\right) = 0$
$F_{uu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_u^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{uu_{+}} \otimes z_{u_{+}} \right)P_{uu\_u} + \left(z_{u} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{u} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
A and B are the common perturbation matrices:
$A = f_{y_0} + \begin{pmatrix} \underbrace{0}_{n\times n_{static}} &\vdots& \underbrace{f_{y^{**}_{+}} \cdot g^{**}_{x}}_{n \times n_{spred}} &\vdots& \underbrace{0}_{n\times n_{frwd}} \end{pmatrix}$and $B = \begin{pmatrix} \underbrace{0}_{n \times n_{static}}&\vdots & \underbrace{0}_{n \times n_{pred}} & \vdots & \underbrace{f_{y^{**}_{+}}}_{n \times n_{sfwrd}} \end{pmatrix}$
and $z=(y_{-}^{\ast}; y; y_{+}^{\ast\ast}; u)$ denotes the dynamic model variables as in `M_.lead_lag_incidence`, $y^\ast$ denote state variables, $y^{\ast\ast}$ denote forward looking variables, $y_+$ denote the variables with a lead, $y_{-}$ denote variables with a lag, $y_0$ denote variables at period t, $f$ the model equations, and $f_z$ the first-order dynamic model derivatives, $f_{zz}$ the second-order dynamic derivatives, and $f_{zzz}$ the third-order dynamic model derivatives. Then:
$z_{x} = \begin{pmatrix}I\\g_{x}\\g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{u} =\begin{pmatrix}0\\g_{u}\\g^{**}_{x} \cdot g^{*}_{u}\\I\end{pmatrix}$, $z_{u_{+}} =\begin{pmatrix}0\\0\\g^{**}_{u}\\0\end{pmatrix}$
$z_{xx} = \begin{pmatrix} 0\\g_{xx}\\g^{**}_{x} \left( g^{*}_x \otimes g^{*}_{x} \right) + g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{xu} =\begin{pmatrix}0\\g_{xu}\\g^{**}_{xx} \left( g^{*}_x \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{xu}\\0\end{pmatrix}$, $z_{uu} =\begin{pmatrix}0\\g_{uu}\\g^{**}_{xx} \left( g^{*}_u \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{uu}\\0\end{pmatrix}$,
$z_{xu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_x \otimes I \right)\\0\end{pmatrix}$, $z_{uu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_{u} \otimes I \right)\\0\end{pmatrix}$, $z_{u_{+}u_{+}} =\begin{pmatrix}0\\0\\g^{\ast\ast}_{uu}\\0\end{pmatrix}$, $z_{\sigma\sigma} = \begin{pmatrix}0\\ g_{\sigma\sigma}\\ g^{\ast\ast}_{x}g^{\ast}_{\sigma\sigma} + g^{\ast\ast}_{\sigma\sigma}\\0 \end{pmatrix}$
$P$ are permutation matrices that can be computed using Matlab's `ipermute` function.
## Parameter derivatives of perturbation solutions
First, we need the parameter derivatives of first, second, third, and fourth derivatives of the dynamic model (i.e. g1,g2,g3,g4 in dynamic files), I make use of the implicit function theorem: Let $f_{z^k}$ denote the kth derivative (wrt all dynamic variables) of the dynamic model, then let $df_{z^k}$ denote the first-derivative (wrt all model parameters) of $f_{z^k}$ evaluated at the steady state. Note that $f_{z^k}$ is a function of both the model parameters $\theta$ and of the steady state of all dynamic variables $\bar{z}$, which also depend on the parameters. Hence, implicitly $f_{z^k}=f_{z^k}(\theta,\bar{z}(\theta))$ and $df_{z^k}$ consists of two parts:
1. direct derivative wrt to all model parameters given by the preprocessor in the `_params_derivs.m` files
2. contribution of derivative of steady state of dynamic variables (wrt all model parameters): $f_{z^{k+1}} \cdot d\bar{z}$
Note that we already have functionality to compute $d\bar{z}$ analytically.
Having this, the above perturbation systems are basically equations of the following types
$AX +BXC = RHS$ or $AX = RHS$
Now when taking the derivative (wrt to one single parameter $\theta_j$), we get
$A\mathrm{d}\{X\} + B\mathrm{d}\{X\}C = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X - \mathrm{d}\{B\}XC - BX\mathrm{d}\{C\}$
or
$A\mathrm{d}\{X\} = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X$
The first one is a Sylvester type equation, the second one can be solved by taking the inverse of $A$. The only diffculty and tedious work arrises in computing (the highly sparse) derivatives of $RHS$.
***
# New functions: `
## get_perturbation_params_derivs.m`and `get_perturbation_params_derivs_numerical_objective.m`
* The parameter derivatives up to third order are computed in the new function`get_perturbation_params_derivs.m` both analytically and numerically. For numerical derivatives `get_perturbation_params_derivs_numerical_objective.m` is the objective for `fjaco.m` or `hessian_sparse.m` or `hessian.m`.
* `get_perturbation_params_derivs.m` is basically an extended version of the previous `get_first_order_solution_params_derivs.m` function.
* * `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It is used for numerical derivatives, whenever `analytic_derivation_mode=-1|-2`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
* Detailed changes:
* Most important: notation of this function is now in accordance to the k_order_solver, i.e. we do not compute derivatives of Kalman transition matrices A and B, but rather the solution matrices ghx,ghu,ghxx,ghxu,ghuu,ghs2,ghxxx,ghxxu,ghxuu,ghuuu,ghxss,ghuss in accordance with notation used in `oo_.dr`. As a byproduct at first-order, focusing on ghx and ghu instead of Kalman transition matrices A and B makes the computations slightly faster for large models (e.g. for Quest the computations were faster by a couple of seconds, not much, but okay).
* Removed use of `kstate`, see also Dynare/dynare#1653 and Dynare/dynare!1656
* Output arguments are stored in a structure `DERIVS`, there is also a flag `d2flag` that computes parameter hessians needed only in `dsge_likelihood.m`.
* Removed `kronflag` as input. `options_.analytic_derivation_mode` is now used instead of `kronflag`.
* Removed `indvar`, an index that was used to selected specific variables in the derivatives. This is not needed, as we always compute the parameter derivatives for all variables first and then select a subset of variables. The selection now takes place in other functions, like `dsge_likelihood.m`.
* Introduced some checks: (i) deterministic exogenous variables are not supported, (ii) Kronecker method only compatible with first-order approximation so reset to sylvester method, (iii) for purely backward or forward models we need to be careful with the rows in `M_.lead_la g_incidence`, (iv) if `_params_derivs.m` files are missing an error is thrown.
* For numerical derivatives, if mod file does not contain an `estimated_params_block`, a temporary one with the most important parameter information is created.
## `unfold_g4.m`
* When evaluating g3 and g4 one needs to take into account that these do not contain symmetric elements, so one needs to use `unfold_g3.m` and the new function `unfold_g4.m`. This returns an unfolded version of the same matrix (i.e. with symmetric elements).
***
# New test models
`.gitignore` and `Makefile.am` are changed accordingly. Also now it is possible to run test suite on analytic_derivatives, i.e. run `make check m/analytic_derivatives`
## `analytic_derivatives/BrockMirman_PertParamsDerivs.mod`
* This is the Brock Mirman model, where we know the exact policy function $g$ for capital and consumption. As this does not imply a nonzero $g_{\sigma\sigma}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ I added some artificial equations to get nonzero solution matrices with respect to $\sigma$. The true perturbation solution matrices $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ are then computed analytically with Matlab's symbolic toolbox and saved in `nBrockMirmanSYM.mat`. There is a preprocessor flag that recreates these analytical computations if changes are needed (and to check whether I made some errors here ;-) )
* Then solution matrices up to third order and their parameter Jacobians are then compared to the ones computed by Dynare's `k_order_solver` and by `get_perturbation_params_derivs` for all `analytic_derivation_mode`'s. There will be an error if the maximum absolute deviation is too large, i.e. for numerical derivatives (`analytic_derivation_mode=-1|-2`) the tolerance is choosen lower (around 1e-5); for analytical methods we are stricter: around 1e-13 for first-order, 1e-12 for second order, and 1e-11 for third-order.
* As a side note, this mod file also checks Dynare's `k_order_solver` algorithm and throws an error if something is wrong.
* This test model shows that the new functionality works well. And analytical derivatives perform way better and accurate than numerical ones, even for this small model.
## `analytic_derivatives/burnside_3_order_PertParamsDerivs.mod`
* This builds upon `tests/k_order_perturbation/burnside_k_order.mod` and computes the true parameter derivatives analytically by hand.
* This test model also shows that the new functionality works well.
## `analytic_derivatives/LindeTrabandt2019.mod`
* Shows that the new functionality also works for medium-sized models, i.e. a SW type model solved at third order with 35 variables (11 states). 2 shocks and 20 parameters.
* This mod file can be used to tweak the speed of the computations in the future.
* Compares numerical versus analytical parameter derivatives (for first, second and third order). Note that this model clearly shows that numerical ones are quite different than analytical ones even at first order!
## `identification/LindeTrabandt2019_xfail.mod`
* This model is a check for issue Dynare/dynare#1595, see fjaco.m below, and will fail.
* Removed `analytic_derivatives/ls2003.mod` as this mod file is neither in the testsuite nor does it work.
***
# Detailed changes in other functions
## `get_first_order_solution_params_derivs.m`
* Deleted, or actually, renamed to `get_perturbation_params_derivs.m`, as this function now is able to compute the derivatives up to third order
## `identification_numerical_objective.m`
* `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
## `dsge_likelihood.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `T`matrix anymore, but instead of the dynare solution matrix `ghx`. So we recreate `T` here as this amounts to adding some zeros and focusing on selected variables only.
* Added some checks to make sure the first-order approximation is selected.
* Removed `kron_flag` as input, as `get_perturbation_params_derivs` looks into `options_.analytic_derivation_mode` for `kron_flag`.
## `dynare_identification.m`
* make sure that setting `analytic_derivation_mode` is set both in `options_ident` and `options_`. Note that at the end of the function we restore the `options_` structure, so all changes are local. In a next merge request, I will remove the global variables to make all variables local.
## `get_identification_jacobians.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `A` and `B` matrix anymore, but instead of the dynare solution matrix `ghx` and `ghu`. So we recreate these matrices here instead of in `get_perturbation_params_derivs.m`.
* Added `str2func` for better function handles in `fjaco.m`.
## `fjaco.m`
* make `tol`an option, which can be adjusted by changing `options_.dynatol.x`for identification and parameter derivatives purposes.
* include a check and an informative error message, if numerical derivatives (two-sided finite difference method) yield errors in `resol.m` for identification and parameter derivatives purposes. This closes issue Dynare/dynare#1595.
* Changed year of copyright to 2010-2017,2019
***
# Further suggestions and questions
* Ones this is merged, I will merge request an improvement of the identification toolbox, which will work up to third order using the pruned state space. This will also remove some issues and bugs, and also I will remove global variables in this request.
* The third-order derivatives can be further improved by taking sparsity into account and use mex versions for kronecker products etc. I leave this for further testing (and if anybody actually uses this ;-) )
2019-12-17 19:17:09 +01:00
dRHSx = zeros ( M . endo_nbr , M . nspred , modparam_nbr ) ;
for jp = 1 : modparam_nbr
dRHSx ( : , : , jp ) = - dg1 ( : , kyy0 , jp ) * zx ( 1 : yy0_nbr , : ) ;
2019-12-20 16:28:06 +01:00
end
Added parameter derivatives of perturbation solution up to 3 order
# Preliminary comments
I finished the identification toolbox at orders two and three using the pruned state space system, but before I merge request this, I decided to first merge the new functionality to compute parameter derivatives of perturbation solution matrices at higher orders. So after this is approved, I merge the identification toolbox.
I guess @rattoma, @sebastien, and @michel are best choices to review this.
I outline the main idea first and then provide some more detailed changes I made to the functions.
***
# Main idea
This merge request is concerned with the *analytical*computation of the parameter derivatives of first, second and third order perturbation solution matrices, i.e. using _closed-form_ expressions to efficiently compute the derivative of $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ *with respect to model parameters* $\theta$. Note that $\theta$ contains model parameters, stderr and corr parameters of shocks. stderr and corr parameters of measurement errors are not yet supported, (they can easily be included as exogenous shocks). The availability of such derivatives is beneficial in terms of more reliable analysis of model sensitivity and parameter identifiability as well as more efficient estimation methods, in particular for models solved up to third order, as it is well-known that numerical derivatives are a tricky business, especially for large models.
References for my approach are:
* Iskrev (2008, 2010) and Schmitt-Grohé and Uribe (2012, Appendix) who were the first to compute the parameter derivatives analytically at first order, however, using inefficient (sparse) Kronecker products.
* Mutschler (2015) who provides the expressions for a second-order, but again using inefficient (sparse) Kronecker products.
* Ratto and Iskrev (2012) who show how the first-order system can be solved accurately, fast and efficiently using existing numerical algorithms for generalized Sylvester equations by taking the parameter derivative with respect to each parameter separately.
* Julliard and Kamenik (2004) who provide the perturbation solution equation system in tensor notation at any order k.
* Levintal (2017) who introduces permutation matrices to express the perturbation solution equation system in matrix notation up to fifth order.
Note that @rattoma already implemented the parameter derivatives of $g_x$ and $g_u$ analytically (and numerically), and I rely heavily on his work in `get_first_order_solution_params_derivs.m` (previously `getH.m`). My additions are mainly to this function and thus it is renamed to `get_perturbation_params_derivs.m`.
The basic idea of this merge request is to take the second- and third-order perturbation solution systems in Julliard and Kamenik (2004), unfold these into an equivalent matrix representation using permutation matrices as in Levintal (2017). Then extending Ratto and Iskrev (2012) one takes the derivative with respect to each parameter separately and gets a computational problem that is linear, albeit large, as it involves either solving generalized Sylvester equations or taking inverses of highly sparse matrices. I will now briefly summarize the perturbation solution system at third order and the system that results when taking the derivative with respect to parameters.
## Perturbation Solution
The following systems arise at first, second, and third order:
$(ghx): f_{x} z_{x} = f_{y_{-}^*} + f_{y_0} g_{x} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{x}= A g_{x} + f_{y_{-}^*}=0$
$(ghu): f_{z} z_{u} = f_{y_0} g_{u} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{u} + f_{u}= A g_u + f_u = 0$
$(ghxx) : A g_{xx} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{x}\right) + f_{zz} \left( z_{x} \otimes z_{x} \right) = 0$
$(ghxu) : A g_{xu} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{x} \otimes z_{u} \right) = 0$
$(ghuu) : A g_{uu} + B g_{xx} \left(g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{u} \otimes z_{u} \right) = 0$
$(ghs2) : (A+B) g_{\sigma\sigma} + \left( f_{y^{**}_{+}y^{**}_{+}} \left(g^{**}_{u} \otimes g^{**}_{u}\right) + f_{y^{**}_{+}} g^{**}_{uu}\right)vec(\Sigma) = 0$
$(ghxxx) : A g_{xxx} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{x}\right) + f_{y_{+}}g^{**}_{xx} \left(g^{*}_x \otimes g^{*}_{xx}\right)P_{x\_xx} + f_{zz} \left( z_{x} \otimes z_{xx} \right)P_{x\_xx} + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{x} \right) = 0$
$(ghxxu) : A g_{xxu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{u} \right) + f_{zz} \left( \left( z_{x} \otimes z_{xu} \right)P_{x\_xu} + \left(z_{xx} \otimes z_{u}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{x} \otimes g^{*}_{xu}\right)P_{x\_xu} + \left(g^{*}_{xx} \otimes g^{*}_{u}\right) \right) = 0$
$(ghxuu) : A g_{xuu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( \left( z_{xu} \otimes z_{u} \right)P_{xu\_u} + \left(z_{x} \otimes z_{uu}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{xu} \otimes g^{*}_{u}\right)P_{xu\_u} + \left(g^{*}_{x} \otimes g^{*}_{uu}\right) \right) = 0$
$(ghuuu) : A g_{uuu} + B g_{xxx} \left(g^{*}_{u} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{u} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( z_{u} \otimes z_{uu} \right)P_{u\_uu} + f_{y_{+}}g^{**}_{xx} \left(g^{*}_{u} \otimes g^{*}_{uu}\right)P_{u\_uu} = 0$
$(ghx\sigma\sigma) : A g_{x\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_x + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{x} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{x} \otimes z_{\sigma\sigma}\right) + F_{xu_{+}u_{+}}\left(I_{n_x} \otimes vec(\Sigma)\right) = 0$
$F_{xu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_x^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{xu_{+}} \otimes z_{u_{+}} \right)P_{xu\_u} + \left(z_{x} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{x} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
$(ghu\sigma\sigma) : A g_{u\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_{u} + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{u} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{u} \otimes z_{\sigma\sigma}\right) + F_{uu_{+}u_{+}}\left(I_{n_u} \otimes vec(\Sigma_u)\right) = 0$
$F_{uu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_u^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{uu_{+}} \otimes z_{u_{+}} \right)P_{uu\_u} + \left(z_{u} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{u} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
A and B are the common perturbation matrices:
$A = f_{y_0} + \begin{pmatrix} \underbrace{0}_{n\times n_{static}} &\vdots& \underbrace{f_{y^{**}_{+}} \cdot g^{**}_{x}}_{n \times n_{spred}} &\vdots& \underbrace{0}_{n\times n_{frwd}} \end{pmatrix}$and $B = \begin{pmatrix} \underbrace{0}_{n \times n_{static}}&\vdots & \underbrace{0}_{n \times n_{pred}} & \vdots & \underbrace{f_{y^{**}_{+}}}_{n \times n_{sfwrd}} \end{pmatrix}$
and $z=(y_{-}^{\ast}; y; y_{+}^{\ast\ast}; u)$ denotes the dynamic model variables as in `M_.lead_lag_incidence`, $y^\ast$ denote state variables, $y^{\ast\ast}$ denote forward looking variables, $y_+$ denote the variables with a lead, $y_{-}$ denote variables with a lag, $y_0$ denote variables at period t, $f$ the model equations, and $f_z$ the first-order dynamic model derivatives, $f_{zz}$ the second-order dynamic derivatives, and $f_{zzz}$ the third-order dynamic model derivatives. Then:
$z_{x} = \begin{pmatrix}I\\g_{x}\\g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{u} =\begin{pmatrix}0\\g_{u}\\g^{**}_{x} \cdot g^{*}_{u}\\I\end{pmatrix}$, $z_{u_{+}} =\begin{pmatrix}0\\0\\g^{**}_{u}\\0\end{pmatrix}$
$z_{xx} = \begin{pmatrix} 0\\g_{xx}\\g^{**}_{x} \left( g^{*}_x \otimes g^{*}_{x} \right) + g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{xu} =\begin{pmatrix}0\\g_{xu}\\g^{**}_{xx} \left( g^{*}_x \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{xu}\\0\end{pmatrix}$, $z_{uu} =\begin{pmatrix}0\\g_{uu}\\g^{**}_{xx} \left( g^{*}_u \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{uu}\\0\end{pmatrix}$,
$z_{xu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_x \otimes I \right)\\0\end{pmatrix}$, $z_{uu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_{u} \otimes I \right)\\0\end{pmatrix}$, $z_{u_{+}u_{+}} =\begin{pmatrix}0\\0\\g^{\ast\ast}_{uu}\\0\end{pmatrix}$, $z_{\sigma\sigma} = \begin{pmatrix}0\\ g_{\sigma\sigma}\\ g^{\ast\ast}_{x}g^{\ast}_{\sigma\sigma} + g^{\ast\ast}_{\sigma\sigma}\\0 \end{pmatrix}$
$P$ are permutation matrices that can be computed using Matlab's `ipermute` function.
## Parameter derivatives of perturbation solutions
First, we need the parameter derivatives of first, second, third, and fourth derivatives of the dynamic model (i.e. g1,g2,g3,g4 in dynamic files), I make use of the implicit function theorem: Let $f_{z^k}$ denote the kth derivative (wrt all dynamic variables) of the dynamic model, then let $df_{z^k}$ denote the first-derivative (wrt all model parameters) of $f_{z^k}$ evaluated at the steady state. Note that $f_{z^k}$ is a function of both the model parameters $\theta$ and of the steady state of all dynamic variables $\bar{z}$, which also depend on the parameters. Hence, implicitly $f_{z^k}=f_{z^k}(\theta,\bar{z}(\theta))$ and $df_{z^k}$ consists of two parts:
1. direct derivative wrt to all model parameters given by the preprocessor in the `_params_derivs.m` files
2. contribution of derivative of steady state of dynamic variables (wrt all model parameters): $f_{z^{k+1}} \cdot d\bar{z}$
Note that we already have functionality to compute $d\bar{z}$ analytically.
Having this, the above perturbation systems are basically equations of the following types
$AX +BXC = RHS$ or $AX = RHS$
Now when taking the derivative (wrt to one single parameter $\theta_j$), we get
$A\mathrm{d}\{X\} + B\mathrm{d}\{X\}C = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X - \mathrm{d}\{B\}XC - BX\mathrm{d}\{C\}$
or
$A\mathrm{d}\{X\} = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X$
The first one is a Sylvester type equation, the second one can be solved by taking the inverse of $A$. The only diffculty and tedious work arrises in computing (the highly sparse) derivatives of $RHS$.
***
# New functions: `
## get_perturbation_params_derivs.m`and `get_perturbation_params_derivs_numerical_objective.m`
* The parameter derivatives up to third order are computed in the new function`get_perturbation_params_derivs.m` both analytically and numerically. For numerical derivatives `get_perturbation_params_derivs_numerical_objective.m` is the objective for `fjaco.m` or `hessian_sparse.m` or `hessian.m`.
* `get_perturbation_params_derivs.m` is basically an extended version of the previous `get_first_order_solution_params_derivs.m` function.
* * `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It is used for numerical derivatives, whenever `analytic_derivation_mode=-1|-2`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
* Detailed changes:
* Most important: notation of this function is now in accordance to the k_order_solver, i.e. we do not compute derivatives of Kalman transition matrices A and B, but rather the solution matrices ghx,ghu,ghxx,ghxu,ghuu,ghs2,ghxxx,ghxxu,ghxuu,ghuuu,ghxss,ghuss in accordance with notation used in `oo_.dr`. As a byproduct at first-order, focusing on ghx and ghu instead of Kalman transition matrices A and B makes the computations slightly faster for large models (e.g. for Quest the computations were faster by a couple of seconds, not much, but okay).
* Removed use of `kstate`, see also Dynare/dynare#1653 and Dynare/dynare!1656
* Output arguments are stored in a structure `DERIVS`, there is also a flag `d2flag` that computes parameter hessians needed only in `dsge_likelihood.m`.
* Removed `kronflag` as input. `options_.analytic_derivation_mode` is now used instead of `kronflag`.
* Removed `indvar`, an index that was used to selected specific variables in the derivatives. This is not needed, as we always compute the parameter derivatives for all variables first and then select a subset of variables. The selection now takes place in other functions, like `dsge_likelihood.m`.
* Introduced some checks: (i) deterministic exogenous variables are not supported, (ii) Kronecker method only compatible with first-order approximation so reset to sylvester method, (iii) for purely backward or forward models we need to be careful with the rows in `M_.lead_la g_incidence`, (iv) if `_params_derivs.m` files are missing an error is thrown.
* For numerical derivatives, if mod file does not contain an `estimated_params_block`, a temporary one with the most important parameter information is created.
## `unfold_g4.m`
* When evaluating g3 and g4 one needs to take into account that these do not contain symmetric elements, so one needs to use `unfold_g3.m` and the new function `unfold_g4.m`. This returns an unfolded version of the same matrix (i.e. with symmetric elements).
***
# New test models
`.gitignore` and `Makefile.am` are changed accordingly. Also now it is possible to run test suite on analytic_derivatives, i.e. run `make check m/analytic_derivatives`
## `analytic_derivatives/BrockMirman_PertParamsDerivs.mod`
* This is the Brock Mirman model, where we know the exact policy function $g$ for capital and consumption. As this does not imply a nonzero $g_{\sigma\sigma}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ I added some artificial equations to get nonzero solution matrices with respect to $\sigma$. The true perturbation solution matrices $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ are then computed analytically with Matlab's symbolic toolbox and saved in `nBrockMirmanSYM.mat`. There is a preprocessor flag that recreates these analytical computations if changes are needed (and to check whether I made some errors here ;-) )
* Then solution matrices up to third order and their parameter Jacobians are then compared to the ones computed by Dynare's `k_order_solver` and by `get_perturbation_params_derivs` for all `analytic_derivation_mode`'s. There will be an error if the maximum absolute deviation is too large, i.e. for numerical derivatives (`analytic_derivation_mode=-1|-2`) the tolerance is choosen lower (around 1e-5); for analytical methods we are stricter: around 1e-13 for first-order, 1e-12 for second order, and 1e-11 for third-order.
* As a side note, this mod file also checks Dynare's `k_order_solver` algorithm and throws an error if something is wrong.
* This test model shows that the new functionality works well. And analytical derivatives perform way better and accurate than numerical ones, even for this small model.
## `analytic_derivatives/burnside_3_order_PertParamsDerivs.mod`
* This builds upon `tests/k_order_perturbation/burnside_k_order.mod` and computes the true parameter derivatives analytically by hand.
* This test model also shows that the new functionality works well.
## `analytic_derivatives/LindeTrabandt2019.mod`
* Shows that the new functionality also works for medium-sized models, i.e. a SW type model solved at third order with 35 variables (11 states). 2 shocks and 20 parameters.
* This mod file can be used to tweak the speed of the computations in the future.
* Compares numerical versus analytical parameter derivatives (for first, second and third order). Note that this model clearly shows that numerical ones are quite different than analytical ones even at first order!
## `identification/LindeTrabandt2019_xfail.mod`
* This model is a check for issue Dynare/dynare#1595, see fjaco.m below, and will fail.
* Removed `analytic_derivatives/ls2003.mod` as this mod file is neither in the testsuite nor does it work.
***
# Detailed changes in other functions
## `get_first_order_solution_params_derivs.m`
* Deleted, or actually, renamed to `get_perturbation_params_derivs.m`, as this function now is able to compute the derivatives up to third order
## `identification_numerical_objective.m`
* `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
## `dsge_likelihood.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `T`matrix anymore, but instead of the dynare solution matrix `ghx`. So we recreate `T` here as this amounts to adding some zeros and focusing on selected variables only.
* Added some checks to make sure the first-order approximation is selected.
* Removed `kron_flag` as input, as `get_perturbation_params_derivs` looks into `options_.analytic_derivation_mode` for `kron_flag`.
## `dynare_identification.m`
* make sure that setting `analytic_derivation_mode` is set both in `options_ident` and `options_`. Note that at the end of the function we restore the `options_` structure, so all changes are local. In a next merge request, I will remove the global variables to make all variables local.
## `get_identification_jacobians.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `A` and `B` matrix anymore, but instead of the dynare solution matrix `ghx` and `ghu`. So we recreate these matrices here instead of in `get_perturbation_params_derivs.m`.
* Added `str2func` for better function handles in `fjaco.m`.
## `fjaco.m`
* make `tol`an option, which can be adjusted by changing `options_.dynatol.x`for identification and parameter derivatives purposes.
* include a check and an informative error message, if numerical derivatives (two-sided finite difference method) yield errors in `resol.m` for identification and parameter derivatives purposes. This closes issue Dynare/dynare#1595.
* Changed year of copyright to 2010-2017,2019
***
# Further suggestions and questions
* Ones this is merged, I will merge request an improvement of the identification toolbox, which will work up to third order using the pruned state space. This will also remove some issues and bugs, and also I will remove global variables in this request.
* The third-order derivatives can be further improved by taking sparsity into account and use mex versions for kronecker products etc. I leave this for further testing (and if anybody actually uses this ;-) )
2019-12-17 19:17:09 +01:00
%use iterated generalized sylvester equation to compute dghx
dghx = sylvester3 ( A , B , oo . dr . ghx ( idx_states , : ) , dRHSx ) ;
flag = 1 ; icount = 0 ;
while flag && icount < 4
[ dghx , flag ] = sylvester3a ( dghx , A , B , oo . dr . ghx ( idx_states , : ) , dRHSx ) ;
icount = icount + 1 ;
end
2019-12-20 16:28:06 +01:00
Added parameter derivatives of perturbation solution up to 3 order
# Preliminary comments
I finished the identification toolbox at orders two and three using the pruned state space system, but before I merge request this, I decided to first merge the new functionality to compute parameter derivatives of perturbation solution matrices at higher orders. So after this is approved, I merge the identification toolbox.
I guess @rattoma, @sebastien, and @michel are best choices to review this.
I outline the main idea first and then provide some more detailed changes I made to the functions.
***
# Main idea
This merge request is concerned with the *analytical*computation of the parameter derivatives of first, second and third order perturbation solution matrices, i.e. using _closed-form_ expressions to efficiently compute the derivative of $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ *with respect to model parameters* $\theta$. Note that $\theta$ contains model parameters, stderr and corr parameters of shocks. stderr and corr parameters of measurement errors are not yet supported, (they can easily be included as exogenous shocks). The availability of such derivatives is beneficial in terms of more reliable analysis of model sensitivity and parameter identifiability as well as more efficient estimation methods, in particular for models solved up to third order, as it is well-known that numerical derivatives are a tricky business, especially for large models.
References for my approach are:
* Iskrev (2008, 2010) and Schmitt-Grohé and Uribe (2012, Appendix) who were the first to compute the parameter derivatives analytically at first order, however, using inefficient (sparse) Kronecker products.
* Mutschler (2015) who provides the expressions for a second-order, but again using inefficient (sparse) Kronecker products.
* Ratto and Iskrev (2012) who show how the first-order system can be solved accurately, fast and efficiently using existing numerical algorithms for generalized Sylvester equations by taking the parameter derivative with respect to each parameter separately.
* Julliard and Kamenik (2004) who provide the perturbation solution equation system in tensor notation at any order k.
* Levintal (2017) who introduces permutation matrices to express the perturbation solution equation system in matrix notation up to fifth order.
Note that @rattoma already implemented the parameter derivatives of $g_x$ and $g_u$ analytically (and numerically), and I rely heavily on his work in `get_first_order_solution_params_derivs.m` (previously `getH.m`). My additions are mainly to this function and thus it is renamed to `get_perturbation_params_derivs.m`.
The basic idea of this merge request is to take the second- and third-order perturbation solution systems in Julliard and Kamenik (2004), unfold these into an equivalent matrix representation using permutation matrices as in Levintal (2017). Then extending Ratto and Iskrev (2012) one takes the derivative with respect to each parameter separately and gets a computational problem that is linear, albeit large, as it involves either solving generalized Sylvester equations or taking inverses of highly sparse matrices. I will now briefly summarize the perturbation solution system at third order and the system that results when taking the derivative with respect to parameters.
## Perturbation Solution
The following systems arise at first, second, and third order:
$(ghx): f_{x} z_{x} = f_{y_{-}^*} + f_{y_0} g_{x} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{x}= A g_{x} + f_{y_{-}^*}=0$
$(ghu): f_{z} z_{u} = f_{y_0} g_{u} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{u} + f_{u}= A g_u + f_u = 0$
$(ghxx) : A g_{xx} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{x}\right) + f_{zz} \left( z_{x} \otimes z_{x} \right) = 0$
$(ghxu) : A g_{xu} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{x} \otimes z_{u} \right) = 0$
$(ghuu) : A g_{uu} + B g_{xx} \left(g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{u} \otimes z_{u} \right) = 0$
$(ghs2) : (A+B) g_{\sigma\sigma} + \left( f_{y^{**}_{+}y^{**}_{+}} \left(g^{**}_{u} \otimes g^{**}_{u}\right) + f_{y^{**}_{+}} g^{**}_{uu}\right)vec(\Sigma) = 0$
$(ghxxx) : A g_{xxx} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{x}\right) + f_{y_{+}}g^{**}_{xx} \left(g^{*}_x \otimes g^{*}_{xx}\right)P_{x\_xx} + f_{zz} \left( z_{x} \otimes z_{xx} \right)P_{x\_xx} + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{x} \right) = 0$
$(ghxxu) : A g_{xxu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{u} \right) + f_{zz} \left( \left( z_{x} \otimes z_{xu} \right)P_{x\_xu} + \left(z_{xx} \otimes z_{u}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{x} \otimes g^{*}_{xu}\right)P_{x\_xu} + \left(g^{*}_{xx} \otimes g^{*}_{u}\right) \right) = 0$
$(ghxuu) : A g_{xuu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( \left( z_{xu} \otimes z_{u} \right)P_{xu\_u} + \left(z_{x} \otimes z_{uu}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{xu} \otimes g^{*}_{u}\right)P_{xu\_u} + \left(g^{*}_{x} \otimes g^{*}_{uu}\right) \right) = 0$
$(ghuuu) : A g_{uuu} + B g_{xxx} \left(g^{*}_{u} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{u} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( z_{u} \otimes z_{uu} \right)P_{u\_uu} + f_{y_{+}}g^{**}_{xx} \left(g^{*}_{u} \otimes g^{*}_{uu}\right)P_{u\_uu} = 0$
$(ghx\sigma\sigma) : A g_{x\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_x + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{x} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{x} \otimes z_{\sigma\sigma}\right) + F_{xu_{+}u_{+}}\left(I_{n_x} \otimes vec(\Sigma)\right) = 0$
$F_{xu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_x^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{xu_{+}} \otimes z_{u_{+}} \right)P_{xu\_u} + \left(z_{x} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{x} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
$(ghu\sigma\sigma) : A g_{u\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_{u} + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{u} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{u} \otimes z_{\sigma\sigma}\right) + F_{uu_{+}u_{+}}\left(I_{n_u} \otimes vec(\Sigma_u)\right) = 0$
$F_{uu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_u^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{uu_{+}} \otimes z_{u_{+}} \right)P_{uu\_u} + \left(z_{u} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{u} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
A and B are the common perturbation matrices:
$A = f_{y_0} + \begin{pmatrix} \underbrace{0}_{n\times n_{static}} &\vdots& \underbrace{f_{y^{**}_{+}} \cdot g^{**}_{x}}_{n \times n_{spred}} &\vdots& \underbrace{0}_{n\times n_{frwd}} \end{pmatrix}$and $B = \begin{pmatrix} \underbrace{0}_{n \times n_{static}}&\vdots & \underbrace{0}_{n \times n_{pred}} & \vdots & \underbrace{f_{y^{**}_{+}}}_{n \times n_{sfwrd}} \end{pmatrix}$
and $z=(y_{-}^{\ast}; y; y_{+}^{\ast\ast}; u)$ denotes the dynamic model variables as in `M_.lead_lag_incidence`, $y^\ast$ denote state variables, $y^{\ast\ast}$ denote forward looking variables, $y_+$ denote the variables with a lead, $y_{-}$ denote variables with a lag, $y_0$ denote variables at period t, $f$ the model equations, and $f_z$ the first-order dynamic model derivatives, $f_{zz}$ the second-order dynamic derivatives, and $f_{zzz}$ the third-order dynamic model derivatives. Then:
$z_{x} = \begin{pmatrix}I\\g_{x}\\g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{u} =\begin{pmatrix}0\\g_{u}\\g^{**}_{x} \cdot g^{*}_{u}\\I\end{pmatrix}$, $z_{u_{+}} =\begin{pmatrix}0\\0\\g^{**}_{u}\\0\end{pmatrix}$
$z_{xx} = \begin{pmatrix} 0\\g_{xx}\\g^{**}_{x} \left( g^{*}_x \otimes g^{*}_{x} \right) + g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{xu} =\begin{pmatrix}0\\g_{xu}\\g^{**}_{xx} \left( g^{*}_x \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{xu}\\0\end{pmatrix}$, $z_{uu} =\begin{pmatrix}0\\g_{uu}\\g^{**}_{xx} \left( g^{*}_u \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{uu}\\0\end{pmatrix}$,
$z_{xu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_x \otimes I \right)\\0\end{pmatrix}$, $z_{uu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_{u} \otimes I \right)\\0\end{pmatrix}$, $z_{u_{+}u_{+}} =\begin{pmatrix}0\\0\\g^{\ast\ast}_{uu}\\0\end{pmatrix}$, $z_{\sigma\sigma} = \begin{pmatrix}0\\ g_{\sigma\sigma}\\ g^{\ast\ast}_{x}g^{\ast}_{\sigma\sigma} + g^{\ast\ast}_{\sigma\sigma}\\0 \end{pmatrix}$
$P$ are permutation matrices that can be computed using Matlab's `ipermute` function.
## Parameter derivatives of perturbation solutions
First, we need the parameter derivatives of first, second, third, and fourth derivatives of the dynamic model (i.e. g1,g2,g3,g4 in dynamic files), I make use of the implicit function theorem: Let $f_{z^k}$ denote the kth derivative (wrt all dynamic variables) of the dynamic model, then let $df_{z^k}$ denote the first-derivative (wrt all model parameters) of $f_{z^k}$ evaluated at the steady state. Note that $f_{z^k}$ is a function of both the model parameters $\theta$ and of the steady state of all dynamic variables $\bar{z}$, which also depend on the parameters. Hence, implicitly $f_{z^k}=f_{z^k}(\theta,\bar{z}(\theta))$ and $df_{z^k}$ consists of two parts:
1. direct derivative wrt to all model parameters given by the preprocessor in the `_params_derivs.m` files
2. contribution of derivative of steady state of dynamic variables (wrt all model parameters): $f_{z^{k+1}} \cdot d\bar{z}$
Note that we already have functionality to compute $d\bar{z}$ analytically.
Having this, the above perturbation systems are basically equations of the following types
$AX +BXC = RHS$ or $AX = RHS$
Now when taking the derivative (wrt to one single parameter $\theta_j$), we get
$A\mathrm{d}\{X\} + B\mathrm{d}\{X\}C = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X - \mathrm{d}\{B\}XC - BX\mathrm{d}\{C\}$
or
$A\mathrm{d}\{X\} = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X$
The first one is a Sylvester type equation, the second one can be solved by taking the inverse of $A$. The only diffculty and tedious work arrises in computing (the highly sparse) derivatives of $RHS$.
***
# New functions: `
## get_perturbation_params_derivs.m`and `get_perturbation_params_derivs_numerical_objective.m`
* The parameter derivatives up to third order are computed in the new function`get_perturbation_params_derivs.m` both analytically and numerically. For numerical derivatives `get_perturbation_params_derivs_numerical_objective.m` is the objective for `fjaco.m` or `hessian_sparse.m` or `hessian.m`.
* `get_perturbation_params_derivs.m` is basically an extended version of the previous `get_first_order_solution_params_derivs.m` function.
* * `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It is used for numerical derivatives, whenever `analytic_derivation_mode=-1|-2`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
* Detailed changes:
* Most important: notation of this function is now in accordance to the k_order_solver, i.e. we do not compute derivatives of Kalman transition matrices A and B, but rather the solution matrices ghx,ghu,ghxx,ghxu,ghuu,ghs2,ghxxx,ghxxu,ghxuu,ghuuu,ghxss,ghuss in accordance with notation used in `oo_.dr`. As a byproduct at first-order, focusing on ghx and ghu instead of Kalman transition matrices A and B makes the computations slightly faster for large models (e.g. for Quest the computations were faster by a couple of seconds, not much, but okay).
* Removed use of `kstate`, see also Dynare/dynare#1653 and Dynare/dynare!1656
* Output arguments are stored in a structure `DERIVS`, there is also a flag `d2flag` that computes parameter hessians needed only in `dsge_likelihood.m`.
* Removed `kronflag` as input. `options_.analytic_derivation_mode` is now used instead of `kronflag`.
* Removed `indvar`, an index that was used to selected specific variables in the derivatives. This is not needed, as we always compute the parameter derivatives for all variables first and then select a subset of variables. The selection now takes place in other functions, like `dsge_likelihood.m`.
* Introduced some checks: (i) deterministic exogenous variables are not supported, (ii) Kronecker method only compatible with first-order approximation so reset to sylvester method, (iii) for purely backward or forward models we need to be careful with the rows in `M_.lead_la g_incidence`, (iv) if `_params_derivs.m` files are missing an error is thrown.
* For numerical derivatives, if mod file does not contain an `estimated_params_block`, a temporary one with the most important parameter information is created.
## `unfold_g4.m`
* When evaluating g3 and g4 one needs to take into account that these do not contain symmetric elements, so one needs to use `unfold_g3.m` and the new function `unfold_g4.m`. This returns an unfolded version of the same matrix (i.e. with symmetric elements).
***
# New test models
`.gitignore` and `Makefile.am` are changed accordingly. Also now it is possible to run test suite on analytic_derivatives, i.e. run `make check m/analytic_derivatives`
## `analytic_derivatives/BrockMirman_PertParamsDerivs.mod`
* This is the Brock Mirman model, where we know the exact policy function $g$ for capital and consumption. As this does not imply a nonzero $g_{\sigma\sigma}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ I added some artificial equations to get nonzero solution matrices with respect to $\sigma$. The true perturbation solution matrices $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ are then computed analytically with Matlab's symbolic toolbox and saved in `nBrockMirmanSYM.mat`. There is a preprocessor flag that recreates these analytical computations if changes are needed (and to check whether I made some errors here ;-) )
* Then solution matrices up to third order and their parameter Jacobians are then compared to the ones computed by Dynare's `k_order_solver` and by `get_perturbation_params_derivs` for all `analytic_derivation_mode`'s. There will be an error if the maximum absolute deviation is too large, i.e. for numerical derivatives (`analytic_derivation_mode=-1|-2`) the tolerance is choosen lower (around 1e-5); for analytical methods we are stricter: around 1e-13 for first-order, 1e-12 for second order, and 1e-11 for third-order.
* As a side note, this mod file also checks Dynare's `k_order_solver` algorithm and throws an error if something is wrong.
* This test model shows that the new functionality works well. And analytical derivatives perform way better and accurate than numerical ones, even for this small model.
## `analytic_derivatives/burnside_3_order_PertParamsDerivs.mod`
* This builds upon `tests/k_order_perturbation/burnside_k_order.mod` and computes the true parameter derivatives analytically by hand.
* This test model also shows that the new functionality works well.
## `analytic_derivatives/LindeTrabandt2019.mod`
* Shows that the new functionality also works for medium-sized models, i.e. a SW type model solved at third order with 35 variables (11 states). 2 shocks and 20 parameters.
* This mod file can be used to tweak the speed of the computations in the future.
* Compares numerical versus analytical parameter derivatives (for first, second and third order). Note that this model clearly shows that numerical ones are quite different than analytical ones even at first order!
## `identification/LindeTrabandt2019_xfail.mod`
* This model is a check for issue Dynare/dynare#1595, see fjaco.m below, and will fail.
* Removed `analytic_derivatives/ls2003.mod` as this mod file is neither in the testsuite nor does it work.
***
# Detailed changes in other functions
## `get_first_order_solution_params_derivs.m`
* Deleted, or actually, renamed to `get_perturbation_params_derivs.m`, as this function now is able to compute the derivatives up to third order
## `identification_numerical_objective.m`
* `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
## `dsge_likelihood.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `T`matrix anymore, but instead of the dynare solution matrix `ghx`. So we recreate `T` here as this amounts to adding some zeros and focusing on selected variables only.
* Added some checks to make sure the first-order approximation is selected.
* Removed `kron_flag` as input, as `get_perturbation_params_derivs` looks into `options_.analytic_derivation_mode` for `kron_flag`.
## `dynare_identification.m`
* make sure that setting `analytic_derivation_mode` is set both in `options_ident` and `options_`. Note that at the end of the function we restore the `options_` structure, so all changes are local. In a next merge request, I will remove the global variables to make all variables local.
## `get_identification_jacobians.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `A` and `B` matrix anymore, but instead of the dynare solution matrix `ghx` and `ghu`. So we recreate these matrices here instead of in `get_perturbation_params_derivs.m`.
* Added `str2func` for better function handles in `fjaco.m`.
## `fjaco.m`
* make `tol`an option, which can be adjusted by changing `options_.dynatol.x`for identification and parameter derivatives purposes.
* include a check and an informative error message, if numerical derivatives (two-sided finite difference method) yield errors in `resol.m` for identification and parameter derivatives purposes. This closes issue Dynare/dynare#1595.
* Changed year of copyright to 2010-2017,2019
***
# Further suggestions and questions
* Ones this is merged, I will merge request an improvement of the identification toolbox, which will work up to third order using the pruned state space. This will also remove some issues and bugs, and also I will remove global variables in this request.
* The third-order derivatives can be further improved by taking sparsity into account and use mex versions for kronecker products etc. I leave this for further testing (and if anybody actually uses this ;-) )
2019-12-17 19:17:09 +01:00
%Compute dOm, dghu, dA, dB
dOm = zeros ( M . endo_nbr , M . endo_nbr , totparam_nbr ) ; %as Om=ghu*Sigma_e*ghu', we need to use totparam_nbr, because there is also a contribution from stderr and corr parameters, which we compute after modparams
dghu = zeros ( M . endo_nbr , M . exo_nbr , modparam_nbr ) ;
dA = zeros ( M . endo_nbr , M . endo_nbr , modparam_nbr ) ; %dA is also needed at higher orders
dA ( : , kcurr ~= 0 , : ) = dg1 ( : , nonzeros ( kcurr ) , : ) ;
invA = inv ( A ) ; %also needed at higher orders
for jp = 1 : modparam_nbr
dA ( : , idx_states , jp ) = dA ( : , idx_states , jp ) + dg1 ( : , nonzeros ( klead ) , jp ) * oo . dr . ghx ( klead ~= 0 , : ) + g1 ( : , nonzeros ( klead ) ) * dghx ( klead ~= 0 , : , jp ) ;
dghu ( : , : , jp ) = - invA * ( dg1 ( : , yy0_nbr + 1 : end , jp ) + dA ( : , : , jp ) * oo . dr . ghu ) ;
dOm ( : , : , stderrparam_nbr + corrparam_nbr + jp ) = dghu ( : , : , jp ) * M . Sigma_e * oo . dr . ghu ' + oo . dr . ghu * M . Sigma_e * dghu ( : , : , jp ) ' ;
2019-12-20 16:28:06 +01:00
end
Added parameter derivatives of perturbation solution up to 3 order
# Preliminary comments
I finished the identification toolbox at orders two and three using the pruned state space system, but before I merge request this, I decided to first merge the new functionality to compute parameter derivatives of perturbation solution matrices at higher orders. So after this is approved, I merge the identification toolbox.
I guess @rattoma, @sebastien, and @michel are best choices to review this.
I outline the main idea first and then provide some more detailed changes I made to the functions.
***
# Main idea
This merge request is concerned with the *analytical*computation of the parameter derivatives of first, second and third order perturbation solution matrices, i.e. using _closed-form_ expressions to efficiently compute the derivative of $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ *with respect to model parameters* $\theta$. Note that $\theta$ contains model parameters, stderr and corr parameters of shocks. stderr and corr parameters of measurement errors are not yet supported, (they can easily be included as exogenous shocks). The availability of such derivatives is beneficial in terms of more reliable analysis of model sensitivity and parameter identifiability as well as more efficient estimation methods, in particular for models solved up to third order, as it is well-known that numerical derivatives are a tricky business, especially for large models.
References for my approach are:
* Iskrev (2008, 2010) and Schmitt-Grohé and Uribe (2012, Appendix) who were the first to compute the parameter derivatives analytically at first order, however, using inefficient (sparse) Kronecker products.
* Mutschler (2015) who provides the expressions for a second-order, but again using inefficient (sparse) Kronecker products.
* Ratto and Iskrev (2012) who show how the first-order system can be solved accurately, fast and efficiently using existing numerical algorithms for generalized Sylvester equations by taking the parameter derivative with respect to each parameter separately.
* Julliard and Kamenik (2004) who provide the perturbation solution equation system in tensor notation at any order k.
* Levintal (2017) who introduces permutation matrices to express the perturbation solution equation system in matrix notation up to fifth order.
Note that @rattoma already implemented the parameter derivatives of $g_x$ and $g_u$ analytically (and numerically), and I rely heavily on his work in `get_first_order_solution_params_derivs.m` (previously `getH.m`). My additions are mainly to this function and thus it is renamed to `get_perturbation_params_derivs.m`.
The basic idea of this merge request is to take the second- and third-order perturbation solution systems in Julliard and Kamenik (2004), unfold these into an equivalent matrix representation using permutation matrices as in Levintal (2017). Then extending Ratto and Iskrev (2012) one takes the derivative with respect to each parameter separately and gets a computational problem that is linear, albeit large, as it involves either solving generalized Sylvester equations or taking inverses of highly sparse matrices. I will now briefly summarize the perturbation solution system at third order and the system that results when taking the derivative with respect to parameters.
## Perturbation Solution
The following systems arise at first, second, and third order:
$(ghx): f_{x} z_{x} = f_{y_{-}^*} + f_{y_0} g_{x} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{x}= A g_{x} + f_{y_{-}^*}=0$
$(ghu): f_{z} z_{u} = f_{y_0} g_{u} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{u} + f_{u}= A g_u + f_u = 0$
$(ghxx) : A g_{xx} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{x}\right) + f_{zz} \left( z_{x} \otimes z_{x} \right) = 0$
$(ghxu) : A g_{xu} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{x} \otimes z_{u} \right) = 0$
$(ghuu) : A g_{uu} + B g_{xx} \left(g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{u} \otimes z_{u} \right) = 0$
$(ghs2) : (A+B) g_{\sigma\sigma} + \left( f_{y^{**}_{+}y^{**}_{+}} \left(g^{**}_{u} \otimes g^{**}_{u}\right) + f_{y^{**}_{+}} g^{**}_{uu}\right)vec(\Sigma) = 0$
$(ghxxx) : A g_{xxx} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{x}\right) + f_{y_{+}}g^{**}_{xx} \left(g^{*}_x \otimes g^{*}_{xx}\right)P_{x\_xx} + f_{zz} \left( z_{x} \otimes z_{xx} \right)P_{x\_xx} + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{x} \right) = 0$
$(ghxxu) : A g_{xxu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{u} \right) + f_{zz} \left( \left( z_{x} \otimes z_{xu} \right)P_{x\_xu} + \left(z_{xx} \otimes z_{u}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{x} \otimes g^{*}_{xu}\right)P_{x\_xu} + \left(g^{*}_{xx} \otimes g^{*}_{u}\right) \right) = 0$
$(ghxuu) : A g_{xuu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( \left( z_{xu} \otimes z_{u} \right)P_{xu\_u} + \left(z_{x} \otimes z_{uu}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{xu} \otimes g^{*}_{u}\right)P_{xu\_u} + \left(g^{*}_{x} \otimes g^{*}_{uu}\right) \right) = 0$
$(ghuuu) : A g_{uuu} + B g_{xxx} \left(g^{*}_{u} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{u} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( z_{u} \otimes z_{uu} \right)P_{u\_uu} + f_{y_{+}}g^{**}_{xx} \left(g^{*}_{u} \otimes g^{*}_{uu}\right)P_{u\_uu} = 0$
$(ghx\sigma\sigma) : A g_{x\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_x + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{x} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{x} \otimes z_{\sigma\sigma}\right) + F_{xu_{+}u_{+}}\left(I_{n_x} \otimes vec(\Sigma)\right) = 0$
$F_{xu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_x^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{xu_{+}} \otimes z_{u_{+}} \right)P_{xu\_u} + \left(z_{x} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{x} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
$(ghu\sigma\sigma) : A g_{u\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_{u} + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{u} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{u} \otimes z_{\sigma\sigma}\right) + F_{uu_{+}u_{+}}\left(I_{n_u} \otimes vec(\Sigma_u)\right) = 0$
$F_{uu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_u^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{uu_{+}} \otimes z_{u_{+}} \right)P_{uu\_u} + \left(z_{u} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{u} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
A and B are the common perturbation matrices:
$A = f_{y_0} + \begin{pmatrix} \underbrace{0}_{n\times n_{static}} &\vdots& \underbrace{f_{y^{**}_{+}} \cdot g^{**}_{x}}_{n \times n_{spred}} &\vdots& \underbrace{0}_{n\times n_{frwd}} \end{pmatrix}$and $B = \begin{pmatrix} \underbrace{0}_{n \times n_{static}}&\vdots & \underbrace{0}_{n \times n_{pred}} & \vdots & \underbrace{f_{y^{**}_{+}}}_{n \times n_{sfwrd}} \end{pmatrix}$
and $z=(y_{-}^{\ast}; y; y_{+}^{\ast\ast}; u)$ denotes the dynamic model variables as in `M_.lead_lag_incidence`, $y^\ast$ denote state variables, $y^{\ast\ast}$ denote forward looking variables, $y_+$ denote the variables with a lead, $y_{-}$ denote variables with a lag, $y_0$ denote variables at period t, $f$ the model equations, and $f_z$ the first-order dynamic model derivatives, $f_{zz}$ the second-order dynamic derivatives, and $f_{zzz}$ the third-order dynamic model derivatives. Then:
$z_{x} = \begin{pmatrix}I\\g_{x}\\g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{u} =\begin{pmatrix}0\\g_{u}\\g^{**}_{x} \cdot g^{*}_{u}\\I\end{pmatrix}$, $z_{u_{+}} =\begin{pmatrix}0\\0\\g^{**}_{u}\\0\end{pmatrix}$
$z_{xx} = \begin{pmatrix} 0\\g_{xx}\\g^{**}_{x} \left( g^{*}_x \otimes g^{*}_{x} \right) + g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{xu} =\begin{pmatrix}0\\g_{xu}\\g^{**}_{xx} \left( g^{*}_x \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{xu}\\0\end{pmatrix}$, $z_{uu} =\begin{pmatrix}0\\g_{uu}\\g^{**}_{xx} \left( g^{*}_u \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{uu}\\0\end{pmatrix}$,
$z_{xu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_x \otimes I \right)\\0\end{pmatrix}$, $z_{uu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_{u} \otimes I \right)\\0\end{pmatrix}$, $z_{u_{+}u_{+}} =\begin{pmatrix}0\\0\\g^{\ast\ast}_{uu}\\0\end{pmatrix}$, $z_{\sigma\sigma} = \begin{pmatrix}0\\ g_{\sigma\sigma}\\ g^{\ast\ast}_{x}g^{\ast}_{\sigma\sigma} + g^{\ast\ast}_{\sigma\sigma}\\0 \end{pmatrix}$
$P$ are permutation matrices that can be computed using Matlab's `ipermute` function.
## Parameter derivatives of perturbation solutions
First, we need the parameter derivatives of first, second, third, and fourth derivatives of the dynamic model (i.e. g1,g2,g3,g4 in dynamic files), I make use of the implicit function theorem: Let $f_{z^k}$ denote the kth derivative (wrt all dynamic variables) of the dynamic model, then let $df_{z^k}$ denote the first-derivative (wrt all model parameters) of $f_{z^k}$ evaluated at the steady state. Note that $f_{z^k}$ is a function of both the model parameters $\theta$ and of the steady state of all dynamic variables $\bar{z}$, which also depend on the parameters. Hence, implicitly $f_{z^k}=f_{z^k}(\theta,\bar{z}(\theta))$ and $df_{z^k}$ consists of two parts:
1. direct derivative wrt to all model parameters given by the preprocessor in the `_params_derivs.m` files
2. contribution of derivative of steady state of dynamic variables (wrt all model parameters): $f_{z^{k+1}} \cdot d\bar{z}$
Note that we already have functionality to compute $d\bar{z}$ analytically.
Having this, the above perturbation systems are basically equations of the following types
$AX +BXC = RHS$ or $AX = RHS$
Now when taking the derivative (wrt to one single parameter $\theta_j$), we get
$A\mathrm{d}\{X\} + B\mathrm{d}\{X\}C = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X - \mathrm{d}\{B\}XC - BX\mathrm{d}\{C\}$
or
$A\mathrm{d}\{X\} = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X$
The first one is a Sylvester type equation, the second one can be solved by taking the inverse of $A$. The only diffculty and tedious work arrises in computing (the highly sparse) derivatives of $RHS$.
***
# New functions: `
## get_perturbation_params_derivs.m`and `get_perturbation_params_derivs_numerical_objective.m`
* The parameter derivatives up to third order are computed in the new function`get_perturbation_params_derivs.m` both analytically and numerically. For numerical derivatives `get_perturbation_params_derivs_numerical_objective.m` is the objective for `fjaco.m` or `hessian_sparse.m` or `hessian.m`.
* `get_perturbation_params_derivs.m` is basically an extended version of the previous `get_first_order_solution_params_derivs.m` function.
* * `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It is used for numerical derivatives, whenever `analytic_derivation_mode=-1|-2`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
* Detailed changes:
* Most important: notation of this function is now in accordance to the k_order_solver, i.e. we do not compute derivatives of Kalman transition matrices A and B, but rather the solution matrices ghx,ghu,ghxx,ghxu,ghuu,ghs2,ghxxx,ghxxu,ghxuu,ghuuu,ghxss,ghuss in accordance with notation used in `oo_.dr`. As a byproduct at first-order, focusing on ghx and ghu instead of Kalman transition matrices A and B makes the computations slightly faster for large models (e.g. for Quest the computations were faster by a couple of seconds, not much, but okay).
* Removed use of `kstate`, see also Dynare/dynare#1653 and Dynare/dynare!1656
* Output arguments are stored in a structure `DERIVS`, there is also a flag `d2flag` that computes parameter hessians needed only in `dsge_likelihood.m`.
* Removed `kronflag` as input. `options_.analytic_derivation_mode` is now used instead of `kronflag`.
* Removed `indvar`, an index that was used to selected specific variables in the derivatives. This is not needed, as we always compute the parameter derivatives for all variables first and then select a subset of variables. The selection now takes place in other functions, like `dsge_likelihood.m`.
* Introduced some checks: (i) deterministic exogenous variables are not supported, (ii) Kronecker method only compatible with first-order approximation so reset to sylvester method, (iii) for purely backward or forward models we need to be careful with the rows in `M_.lead_la g_incidence`, (iv) if `_params_derivs.m` files are missing an error is thrown.
* For numerical derivatives, if mod file does not contain an `estimated_params_block`, a temporary one with the most important parameter information is created.
## `unfold_g4.m`
* When evaluating g3 and g4 one needs to take into account that these do not contain symmetric elements, so one needs to use `unfold_g3.m` and the new function `unfold_g4.m`. This returns an unfolded version of the same matrix (i.e. with symmetric elements).
***
# New test models
`.gitignore` and `Makefile.am` are changed accordingly. Also now it is possible to run test suite on analytic_derivatives, i.e. run `make check m/analytic_derivatives`
## `analytic_derivatives/BrockMirman_PertParamsDerivs.mod`
* This is the Brock Mirman model, where we know the exact policy function $g$ for capital and consumption. As this does not imply a nonzero $g_{\sigma\sigma}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ I added some artificial equations to get nonzero solution matrices with respect to $\sigma$. The true perturbation solution matrices $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ are then computed analytically with Matlab's symbolic toolbox and saved in `nBrockMirmanSYM.mat`. There is a preprocessor flag that recreates these analytical computations if changes are needed (and to check whether I made some errors here ;-) )
* Then solution matrices up to third order and their parameter Jacobians are then compared to the ones computed by Dynare's `k_order_solver` and by `get_perturbation_params_derivs` for all `analytic_derivation_mode`'s. There will be an error if the maximum absolute deviation is too large, i.e. for numerical derivatives (`analytic_derivation_mode=-1|-2`) the tolerance is choosen lower (around 1e-5); for analytical methods we are stricter: around 1e-13 for first-order, 1e-12 for second order, and 1e-11 for third-order.
* As a side note, this mod file also checks Dynare's `k_order_solver` algorithm and throws an error if something is wrong.
* This test model shows that the new functionality works well. And analytical derivatives perform way better and accurate than numerical ones, even for this small model.
## `analytic_derivatives/burnside_3_order_PertParamsDerivs.mod`
* This builds upon `tests/k_order_perturbation/burnside_k_order.mod` and computes the true parameter derivatives analytically by hand.
* This test model also shows that the new functionality works well.
## `analytic_derivatives/LindeTrabandt2019.mod`
* Shows that the new functionality also works for medium-sized models, i.e. a SW type model solved at third order with 35 variables (11 states). 2 shocks and 20 parameters.
* This mod file can be used to tweak the speed of the computations in the future.
* Compares numerical versus analytical parameter derivatives (for first, second and third order). Note that this model clearly shows that numerical ones are quite different than analytical ones even at first order!
## `identification/LindeTrabandt2019_xfail.mod`
* This model is a check for issue Dynare/dynare#1595, see fjaco.m below, and will fail.
* Removed `analytic_derivatives/ls2003.mod` as this mod file is neither in the testsuite nor does it work.
***
# Detailed changes in other functions
## `get_first_order_solution_params_derivs.m`
* Deleted, or actually, renamed to `get_perturbation_params_derivs.m`, as this function now is able to compute the derivatives up to third order
## `identification_numerical_objective.m`
* `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
## `dsge_likelihood.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `T`matrix anymore, but instead of the dynare solution matrix `ghx`. So we recreate `T` here as this amounts to adding some zeros and focusing on selected variables only.
* Added some checks to make sure the first-order approximation is selected.
* Removed `kron_flag` as input, as `get_perturbation_params_derivs` looks into `options_.analytic_derivation_mode` for `kron_flag`.
## `dynare_identification.m`
* make sure that setting `analytic_derivation_mode` is set both in `options_ident` and `options_`. Note that at the end of the function we restore the `options_` structure, so all changes are local. In a next merge request, I will remove the global variables to make all variables local.
## `get_identification_jacobians.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `A` and `B` matrix anymore, but instead of the dynare solution matrix `ghx` and `ghu`. So we recreate these matrices here instead of in `get_perturbation_params_derivs.m`.
* Added `str2func` for better function handles in `fjaco.m`.
## `fjaco.m`
* make `tol`an option, which can be adjusted by changing `options_.dynatol.x`for identification and parameter derivatives purposes.
* include a check and an informative error message, if numerical derivatives (two-sided finite difference method) yield errors in `resol.m` for identification and parameter derivatives purposes. This closes issue Dynare/dynare#1595.
* Changed year of copyright to 2010-2017,2019
***
# Further suggestions and questions
* Ones this is merged, I will merge request an improvement of the identification toolbox, which will work up to third order using the pruned state space. This will also remove some issues and bugs, and also I will remove global variables in this request.
* The third-order derivatives can be further improved by taking sparsity into account and use mex versions for kronecker products etc. I leave this for further testing (and if anybody actually uses this ;-) )
2019-12-17 19:17:09 +01:00
%add stderr and corr derivatives to Om=ghu*Sigma_e*ghu'
if ~ isempty ( indpstderr )
for jp = 1 : stderrparam_nbr
dOm ( : , : , jp ) = oo . dr . ghu * dSigma_e ( : , : , jp ) * oo . dr . ghu ' ;
end
end
if ~ isempty ( indpcorr )
for jp = 1 : corrparam_nbr
dOm ( : , : , stderrparam_nbr + jp ) = oo . dr . ghu * dSigma_e ( : , : , stderrparam_nbr + jp ) * oo . dr . ghu ' ;
end
end
end
%% Analytical computation of Jacobian (wrt selected model parameters) of second-order solution matrices ghxx,ghxu,ghuu and ghs2
if order > 1
% Note that at second-order we have the following expressions, which are (numerically) zero:
% * for ghxx: A*ghxx + B*ghxx*kron(ghx_,ghx_) + g2*kron(zx,zx) = 0
% Taking the differential yields a generalized sylvester equation to get dghxx: A*dghxx + B*dghxx*kron(ghx_,ghx_) = RHSxx
% * for ghxu: A*ghxu + B*ghxx*kron(ghx_,ghu_) + g2*kron(zx,zu) = 0
% Taking the differential yields an invertible equation to get dghxu: A*dghxu = RHSxu
% * for ghuu: A*ghuu + B*ghxx*kron(ghu_,ghu_) + g2*kron(zu,zu) = 0
% Taking the differential yields an invertible equation to get dghuu: A*dghuu = RHSuu
% * for ghs2: Ahs2*ghs2 + (gg2_{++}*kron(ghup,ghup) + fyp*ghuup)*vec(Sigma_e) = 0
% Taking the differential yields an invertible equation to get dghs2: S*dghs2 = RHSs2
% * due to certainty equivalence and zero mean shocks, we note that ghxs and ghus are zero, and thus not computed
dB = zeros ( M . endo_nbr , M . endo_nbr , modparam_nbr ) ; %this matrix is also needed at higher orders
dB ( : , M . nstatic + M . npred + 1 : end , : ) = dg1 ( : , nonzeros ( klead ) , : ) ;
S = A + B ; %needed for dghs2, but also at higher orders
dS = dA + dB ;
invS = inv ( S ) ;
G_x_x = kron ( oo . dr . ghx ( idx_states , : ) , oo . dr . ghx ( idx_states , : ) ) ;
dG_x_x = zeros ( size ( G_x_x , 1 ) , size ( G_x_x , 2 ) , modparam_nbr ) ;
2019-12-20 16:28:06 +01:00
dzx = zeros ( size ( zx , 1 ) , size ( zx , 2 ) , modparam_nbr ) ;
Added parameter derivatives of perturbation solution up to 3 order
# Preliminary comments
I finished the identification toolbox at orders two and three using the pruned state space system, but before I merge request this, I decided to first merge the new functionality to compute parameter derivatives of perturbation solution matrices at higher orders. So after this is approved, I merge the identification toolbox.
I guess @rattoma, @sebastien, and @michel are best choices to review this.
I outline the main idea first and then provide some more detailed changes I made to the functions.
***
# Main idea
This merge request is concerned with the *analytical*computation of the parameter derivatives of first, second and third order perturbation solution matrices, i.e. using _closed-form_ expressions to efficiently compute the derivative of $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ *with respect to model parameters* $\theta$. Note that $\theta$ contains model parameters, stderr and corr parameters of shocks. stderr and corr parameters of measurement errors are not yet supported, (they can easily be included as exogenous shocks). The availability of such derivatives is beneficial in terms of more reliable analysis of model sensitivity and parameter identifiability as well as more efficient estimation methods, in particular for models solved up to third order, as it is well-known that numerical derivatives are a tricky business, especially for large models.
References for my approach are:
* Iskrev (2008, 2010) and Schmitt-Grohé and Uribe (2012, Appendix) who were the first to compute the parameter derivatives analytically at first order, however, using inefficient (sparse) Kronecker products.
* Mutschler (2015) who provides the expressions for a second-order, but again using inefficient (sparse) Kronecker products.
* Ratto and Iskrev (2012) who show how the first-order system can be solved accurately, fast and efficiently using existing numerical algorithms for generalized Sylvester equations by taking the parameter derivative with respect to each parameter separately.
* Julliard and Kamenik (2004) who provide the perturbation solution equation system in tensor notation at any order k.
* Levintal (2017) who introduces permutation matrices to express the perturbation solution equation system in matrix notation up to fifth order.
Note that @rattoma already implemented the parameter derivatives of $g_x$ and $g_u$ analytically (and numerically), and I rely heavily on his work in `get_first_order_solution_params_derivs.m` (previously `getH.m`). My additions are mainly to this function and thus it is renamed to `get_perturbation_params_derivs.m`.
The basic idea of this merge request is to take the second- and third-order perturbation solution systems in Julliard and Kamenik (2004), unfold these into an equivalent matrix representation using permutation matrices as in Levintal (2017). Then extending Ratto and Iskrev (2012) one takes the derivative with respect to each parameter separately and gets a computational problem that is linear, albeit large, as it involves either solving generalized Sylvester equations or taking inverses of highly sparse matrices. I will now briefly summarize the perturbation solution system at third order and the system that results when taking the derivative with respect to parameters.
## Perturbation Solution
The following systems arise at first, second, and third order:
$(ghx): f_{x} z_{x} = f_{y_{-}^*} + f_{y_0} g_{x} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{x}= A g_{x} + f_{y_{-}^*}=0$
$(ghu): f_{z} z_{u} = f_{y_0} g_{u} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{u} + f_{u}= A g_u + f_u = 0$
$(ghxx) : A g_{xx} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{x}\right) + f_{zz} \left( z_{x} \otimes z_{x} \right) = 0$
$(ghxu) : A g_{xu} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{x} \otimes z_{u} \right) = 0$
$(ghuu) : A g_{uu} + B g_{xx} \left(g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{u} \otimes z_{u} \right) = 0$
$(ghs2) : (A+B) g_{\sigma\sigma} + \left( f_{y^{**}_{+}y^{**}_{+}} \left(g^{**}_{u} \otimes g^{**}_{u}\right) + f_{y^{**}_{+}} g^{**}_{uu}\right)vec(\Sigma) = 0$
$(ghxxx) : A g_{xxx} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{x}\right) + f_{y_{+}}g^{**}_{xx} \left(g^{*}_x \otimes g^{*}_{xx}\right)P_{x\_xx} + f_{zz} \left( z_{x} \otimes z_{xx} \right)P_{x\_xx} + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{x} \right) = 0$
$(ghxxu) : A g_{xxu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{u} \right) + f_{zz} \left( \left( z_{x} \otimes z_{xu} \right)P_{x\_xu} + \left(z_{xx} \otimes z_{u}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{x} \otimes g^{*}_{xu}\right)P_{x\_xu} + \left(g^{*}_{xx} \otimes g^{*}_{u}\right) \right) = 0$
$(ghxuu) : A g_{xuu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( \left( z_{xu} \otimes z_{u} \right)P_{xu\_u} + \left(z_{x} \otimes z_{uu}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{xu} \otimes g^{*}_{u}\right)P_{xu\_u} + \left(g^{*}_{x} \otimes g^{*}_{uu}\right) \right) = 0$
$(ghuuu) : A g_{uuu} + B g_{xxx} \left(g^{*}_{u} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{u} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( z_{u} \otimes z_{uu} \right)P_{u\_uu} + f_{y_{+}}g^{**}_{xx} \left(g^{*}_{u} \otimes g^{*}_{uu}\right)P_{u\_uu} = 0$
$(ghx\sigma\sigma) : A g_{x\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_x + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{x} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{x} \otimes z_{\sigma\sigma}\right) + F_{xu_{+}u_{+}}\left(I_{n_x} \otimes vec(\Sigma)\right) = 0$
$F_{xu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_x^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{xu_{+}} \otimes z_{u_{+}} \right)P_{xu\_u} + \left(z_{x} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{x} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
$(ghu\sigma\sigma) : A g_{u\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_{u} + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{u} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{u} \otimes z_{\sigma\sigma}\right) + F_{uu_{+}u_{+}}\left(I_{n_u} \otimes vec(\Sigma_u)\right) = 0$
$F_{uu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_u^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{uu_{+}} \otimes z_{u_{+}} \right)P_{uu\_u} + \left(z_{u} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{u} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
A and B are the common perturbation matrices:
$A = f_{y_0} + \begin{pmatrix} \underbrace{0}_{n\times n_{static}} &\vdots& \underbrace{f_{y^{**}_{+}} \cdot g^{**}_{x}}_{n \times n_{spred}} &\vdots& \underbrace{0}_{n\times n_{frwd}} \end{pmatrix}$and $B = \begin{pmatrix} \underbrace{0}_{n \times n_{static}}&\vdots & \underbrace{0}_{n \times n_{pred}} & \vdots & \underbrace{f_{y^{**}_{+}}}_{n \times n_{sfwrd}} \end{pmatrix}$
and $z=(y_{-}^{\ast}; y; y_{+}^{\ast\ast}; u)$ denotes the dynamic model variables as in `M_.lead_lag_incidence`, $y^\ast$ denote state variables, $y^{\ast\ast}$ denote forward looking variables, $y_+$ denote the variables with a lead, $y_{-}$ denote variables with a lag, $y_0$ denote variables at period t, $f$ the model equations, and $f_z$ the first-order dynamic model derivatives, $f_{zz}$ the second-order dynamic derivatives, and $f_{zzz}$ the third-order dynamic model derivatives. Then:
$z_{x} = \begin{pmatrix}I\\g_{x}\\g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{u} =\begin{pmatrix}0\\g_{u}\\g^{**}_{x} \cdot g^{*}_{u}\\I\end{pmatrix}$, $z_{u_{+}} =\begin{pmatrix}0\\0\\g^{**}_{u}\\0\end{pmatrix}$
$z_{xx} = \begin{pmatrix} 0\\g_{xx}\\g^{**}_{x} \left( g^{*}_x \otimes g^{*}_{x} \right) + g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{xu} =\begin{pmatrix}0\\g_{xu}\\g^{**}_{xx} \left( g^{*}_x \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{xu}\\0\end{pmatrix}$, $z_{uu} =\begin{pmatrix}0\\g_{uu}\\g^{**}_{xx} \left( g^{*}_u \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{uu}\\0\end{pmatrix}$,
$z_{xu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_x \otimes I \right)\\0\end{pmatrix}$, $z_{uu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_{u} \otimes I \right)\\0\end{pmatrix}$, $z_{u_{+}u_{+}} =\begin{pmatrix}0\\0\\g^{\ast\ast}_{uu}\\0\end{pmatrix}$, $z_{\sigma\sigma} = \begin{pmatrix}0\\ g_{\sigma\sigma}\\ g^{\ast\ast}_{x}g^{\ast}_{\sigma\sigma} + g^{\ast\ast}_{\sigma\sigma}\\0 \end{pmatrix}$
$P$ are permutation matrices that can be computed using Matlab's `ipermute` function.
## Parameter derivatives of perturbation solutions
First, we need the parameter derivatives of first, second, third, and fourth derivatives of the dynamic model (i.e. g1,g2,g3,g4 in dynamic files), I make use of the implicit function theorem: Let $f_{z^k}$ denote the kth derivative (wrt all dynamic variables) of the dynamic model, then let $df_{z^k}$ denote the first-derivative (wrt all model parameters) of $f_{z^k}$ evaluated at the steady state. Note that $f_{z^k}$ is a function of both the model parameters $\theta$ and of the steady state of all dynamic variables $\bar{z}$, which also depend on the parameters. Hence, implicitly $f_{z^k}=f_{z^k}(\theta,\bar{z}(\theta))$ and $df_{z^k}$ consists of two parts:
1. direct derivative wrt to all model parameters given by the preprocessor in the `_params_derivs.m` files
2. contribution of derivative of steady state of dynamic variables (wrt all model parameters): $f_{z^{k+1}} \cdot d\bar{z}$
Note that we already have functionality to compute $d\bar{z}$ analytically.
Having this, the above perturbation systems are basically equations of the following types
$AX +BXC = RHS$ or $AX = RHS$
Now when taking the derivative (wrt to one single parameter $\theta_j$), we get
$A\mathrm{d}\{X\} + B\mathrm{d}\{X\}C = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X - \mathrm{d}\{B\}XC - BX\mathrm{d}\{C\}$
or
$A\mathrm{d}\{X\} = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X$
The first one is a Sylvester type equation, the second one can be solved by taking the inverse of $A$. The only diffculty and tedious work arrises in computing (the highly sparse) derivatives of $RHS$.
***
# New functions: `
## get_perturbation_params_derivs.m`and `get_perturbation_params_derivs_numerical_objective.m`
* The parameter derivatives up to third order are computed in the new function`get_perturbation_params_derivs.m` both analytically and numerically. For numerical derivatives `get_perturbation_params_derivs_numerical_objective.m` is the objective for `fjaco.m` or `hessian_sparse.m` or `hessian.m`.
* `get_perturbation_params_derivs.m` is basically an extended version of the previous `get_first_order_solution_params_derivs.m` function.
* * `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It is used for numerical derivatives, whenever `analytic_derivation_mode=-1|-2`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
* Detailed changes:
* Most important: notation of this function is now in accordance to the k_order_solver, i.e. we do not compute derivatives of Kalman transition matrices A and B, but rather the solution matrices ghx,ghu,ghxx,ghxu,ghuu,ghs2,ghxxx,ghxxu,ghxuu,ghuuu,ghxss,ghuss in accordance with notation used in `oo_.dr`. As a byproduct at first-order, focusing on ghx and ghu instead of Kalman transition matrices A and B makes the computations slightly faster for large models (e.g. for Quest the computations were faster by a couple of seconds, not much, but okay).
* Removed use of `kstate`, see also Dynare/dynare#1653 and Dynare/dynare!1656
* Output arguments are stored in a structure `DERIVS`, there is also a flag `d2flag` that computes parameter hessians needed only in `dsge_likelihood.m`.
* Removed `kronflag` as input. `options_.analytic_derivation_mode` is now used instead of `kronflag`.
* Removed `indvar`, an index that was used to selected specific variables in the derivatives. This is not needed, as we always compute the parameter derivatives for all variables first and then select a subset of variables. The selection now takes place in other functions, like `dsge_likelihood.m`.
* Introduced some checks: (i) deterministic exogenous variables are not supported, (ii) Kronecker method only compatible with first-order approximation so reset to sylvester method, (iii) for purely backward or forward models we need to be careful with the rows in `M_.lead_la g_incidence`, (iv) if `_params_derivs.m` files are missing an error is thrown.
* For numerical derivatives, if mod file does not contain an `estimated_params_block`, a temporary one with the most important parameter information is created.
## `unfold_g4.m`
* When evaluating g3 and g4 one needs to take into account that these do not contain symmetric elements, so one needs to use `unfold_g3.m` and the new function `unfold_g4.m`. This returns an unfolded version of the same matrix (i.e. with symmetric elements).
***
# New test models
`.gitignore` and `Makefile.am` are changed accordingly. Also now it is possible to run test suite on analytic_derivatives, i.e. run `make check m/analytic_derivatives`
## `analytic_derivatives/BrockMirman_PertParamsDerivs.mod`
* This is the Brock Mirman model, where we know the exact policy function $g$ for capital and consumption. As this does not imply a nonzero $g_{\sigma\sigma}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ I added some artificial equations to get nonzero solution matrices with respect to $\sigma$. The true perturbation solution matrices $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ are then computed analytically with Matlab's symbolic toolbox and saved in `nBrockMirmanSYM.mat`. There is a preprocessor flag that recreates these analytical computations if changes are needed (and to check whether I made some errors here ;-) )
* Then solution matrices up to third order and their parameter Jacobians are then compared to the ones computed by Dynare's `k_order_solver` and by `get_perturbation_params_derivs` for all `analytic_derivation_mode`'s. There will be an error if the maximum absolute deviation is too large, i.e. for numerical derivatives (`analytic_derivation_mode=-1|-2`) the tolerance is choosen lower (around 1e-5); for analytical methods we are stricter: around 1e-13 for first-order, 1e-12 for second order, and 1e-11 for third-order.
* As a side note, this mod file also checks Dynare's `k_order_solver` algorithm and throws an error if something is wrong.
* This test model shows that the new functionality works well. And analytical derivatives perform way better and accurate than numerical ones, even for this small model.
## `analytic_derivatives/burnside_3_order_PertParamsDerivs.mod`
* This builds upon `tests/k_order_perturbation/burnside_k_order.mod` and computes the true parameter derivatives analytically by hand.
* This test model also shows that the new functionality works well.
## `analytic_derivatives/LindeTrabandt2019.mod`
* Shows that the new functionality also works for medium-sized models, i.e. a SW type model solved at third order with 35 variables (11 states). 2 shocks and 20 parameters.
* This mod file can be used to tweak the speed of the computations in the future.
* Compares numerical versus analytical parameter derivatives (for first, second and third order). Note that this model clearly shows that numerical ones are quite different than analytical ones even at first order!
## `identification/LindeTrabandt2019_xfail.mod`
* This model is a check for issue Dynare/dynare#1595, see fjaco.m below, and will fail.
* Removed `analytic_derivatives/ls2003.mod` as this mod file is neither in the testsuite nor does it work.
***
# Detailed changes in other functions
## `get_first_order_solution_params_derivs.m`
* Deleted, or actually, renamed to `get_perturbation_params_derivs.m`, as this function now is able to compute the derivatives up to third order
## `identification_numerical_objective.m`
* `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
## `dsge_likelihood.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `T`matrix anymore, but instead of the dynare solution matrix `ghx`. So we recreate `T` here as this amounts to adding some zeros and focusing on selected variables only.
* Added some checks to make sure the first-order approximation is selected.
* Removed `kron_flag` as input, as `get_perturbation_params_derivs` looks into `options_.analytic_derivation_mode` for `kron_flag`.
## `dynare_identification.m`
* make sure that setting `analytic_derivation_mode` is set both in `options_ident` and `options_`. Note that at the end of the function we restore the `options_` structure, so all changes are local. In a next merge request, I will remove the global variables to make all variables local.
## `get_identification_jacobians.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `A` and `B` matrix anymore, but instead of the dynare solution matrix `ghx` and `ghu`. So we recreate these matrices here instead of in `get_perturbation_params_derivs.m`.
* Added `str2func` for better function handles in `fjaco.m`.
## `fjaco.m`
* make `tol`an option, which can be adjusted by changing `options_.dynatol.x`for identification and parameter derivatives purposes.
* include a check and an informative error message, if numerical derivatives (two-sided finite difference method) yield errors in `resol.m` for identification and parameter derivatives purposes. This closes issue Dynare/dynare#1595.
* Changed year of copyright to 2010-2017,2019
***
# Further suggestions and questions
* Ones this is merged, I will merge request an improvement of the identification toolbox, which will work up to third order using the pruned state space. This will also remove some issues and bugs, and also I will remove global variables in this request.
* The third-order derivatives can be further improved by taking sparsity into account and use mex versions for kronecker products etc. I leave this for further testing (and if anybody actually uses this ;-) )
2019-12-17 19:17:09 +01:00
dRHSghxx = zeros ( M . endo_nbr , M . nspred ^2 , modparam_nbr ) ;
for jp = 1 : modparam_nbr
dzx ( : , : , jp ) = [ zeros ( M . nspred , M . nspred ) ;
dghx ( kcurr ~= 0 , : , jp ) ;
dghx ( klead ~= 0 , : , jp ) * oo . dr . ghx ( idx_states , : ) + oo . dr . ghx ( klead ~= 0 , : ) * dghx ( idx_states , : , jp ) ;
zeros ( M . exo_nbr , M . nspred ) ] ;
2020-01-10 17:55:57 +01:00
dRHS_1 = sparse_hessian_times_B_kronecker_C ( dg2 ( : , k2yy0ex0 ( kyy0 , kyy0 ) + ( jp - 1 ) * yy0ex0_nbr ^2 ) , zx ( 1 : yy0_nbr , : ) , threads_BC ) ;
dRHS_2 = sparse_hessian_times_B_kronecker_C ( g2 ( : , k2yy0ex0 ( kyy0 , kyy0 ) ) , dzx ( 1 : yy0_nbr , : , jp ) , zx ( 1 : yy0_nbr , : ) , threads_BC ) ;
dRHS_3 = sparse_hessian_times_B_kronecker_C ( g2 ( : , k2yy0ex0 ( kyy0 , kyy0 ) ) , zx ( 1 : yy0_nbr , : ) , dzx ( 1 : yy0_nbr , : , jp ) , threads_BC ) ;
Added parameter derivatives of perturbation solution up to 3 order
# Preliminary comments
I finished the identification toolbox at orders two and three using the pruned state space system, but before I merge request this, I decided to first merge the new functionality to compute parameter derivatives of perturbation solution matrices at higher orders. So after this is approved, I merge the identification toolbox.
I guess @rattoma, @sebastien, and @michel are best choices to review this.
I outline the main idea first and then provide some more detailed changes I made to the functions.
***
# Main idea
This merge request is concerned with the *analytical*computation of the parameter derivatives of first, second and third order perturbation solution matrices, i.e. using _closed-form_ expressions to efficiently compute the derivative of $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ *with respect to model parameters* $\theta$. Note that $\theta$ contains model parameters, stderr and corr parameters of shocks. stderr and corr parameters of measurement errors are not yet supported, (they can easily be included as exogenous shocks). The availability of such derivatives is beneficial in terms of more reliable analysis of model sensitivity and parameter identifiability as well as more efficient estimation methods, in particular for models solved up to third order, as it is well-known that numerical derivatives are a tricky business, especially for large models.
References for my approach are:
* Iskrev (2008, 2010) and Schmitt-Grohé and Uribe (2012, Appendix) who were the first to compute the parameter derivatives analytically at first order, however, using inefficient (sparse) Kronecker products.
* Mutschler (2015) who provides the expressions for a second-order, but again using inefficient (sparse) Kronecker products.
* Ratto and Iskrev (2012) who show how the first-order system can be solved accurately, fast and efficiently using existing numerical algorithms for generalized Sylvester equations by taking the parameter derivative with respect to each parameter separately.
* Julliard and Kamenik (2004) who provide the perturbation solution equation system in tensor notation at any order k.
* Levintal (2017) who introduces permutation matrices to express the perturbation solution equation system in matrix notation up to fifth order.
Note that @rattoma already implemented the parameter derivatives of $g_x$ and $g_u$ analytically (and numerically), and I rely heavily on his work in `get_first_order_solution_params_derivs.m` (previously `getH.m`). My additions are mainly to this function and thus it is renamed to `get_perturbation_params_derivs.m`.
The basic idea of this merge request is to take the second- and third-order perturbation solution systems in Julliard and Kamenik (2004), unfold these into an equivalent matrix representation using permutation matrices as in Levintal (2017). Then extending Ratto and Iskrev (2012) one takes the derivative with respect to each parameter separately and gets a computational problem that is linear, albeit large, as it involves either solving generalized Sylvester equations or taking inverses of highly sparse matrices. I will now briefly summarize the perturbation solution system at third order and the system that results when taking the derivative with respect to parameters.
## Perturbation Solution
The following systems arise at first, second, and third order:
$(ghx): f_{x} z_{x} = f_{y_{-}^*} + f_{y_0} g_{x} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{x}= A g_{x} + f_{y_{-}^*}=0$
$(ghu): f_{z} z_{u} = f_{y_0} g_{u} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{u} + f_{u}= A g_u + f_u = 0$
$(ghxx) : A g_{xx} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{x}\right) + f_{zz} \left( z_{x} \otimes z_{x} \right) = 0$
$(ghxu) : A g_{xu} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{x} \otimes z_{u} \right) = 0$
$(ghuu) : A g_{uu} + B g_{xx} \left(g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{u} \otimes z_{u} \right) = 0$
$(ghs2) : (A+B) g_{\sigma\sigma} + \left( f_{y^{**}_{+}y^{**}_{+}} \left(g^{**}_{u} \otimes g^{**}_{u}\right) + f_{y^{**}_{+}} g^{**}_{uu}\right)vec(\Sigma) = 0$
$(ghxxx) : A g_{xxx} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{x}\right) + f_{y_{+}}g^{**}_{xx} \left(g^{*}_x \otimes g^{*}_{xx}\right)P_{x\_xx} + f_{zz} \left( z_{x} \otimes z_{xx} \right)P_{x\_xx} + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{x} \right) = 0$
$(ghxxu) : A g_{xxu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{u} \right) + f_{zz} \left( \left( z_{x} \otimes z_{xu} \right)P_{x\_xu} + \left(z_{xx} \otimes z_{u}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{x} \otimes g^{*}_{xu}\right)P_{x\_xu} + \left(g^{*}_{xx} \otimes g^{*}_{u}\right) \right) = 0$
$(ghxuu) : A g_{xuu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( \left( z_{xu} \otimes z_{u} \right)P_{xu\_u} + \left(z_{x} \otimes z_{uu}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{xu} \otimes g^{*}_{u}\right)P_{xu\_u} + \left(g^{*}_{x} \otimes g^{*}_{uu}\right) \right) = 0$
$(ghuuu) : A g_{uuu} + B g_{xxx} \left(g^{*}_{u} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{u} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( z_{u} \otimes z_{uu} \right)P_{u\_uu} + f_{y_{+}}g^{**}_{xx} \left(g^{*}_{u} \otimes g^{*}_{uu}\right)P_{u\_uu} = 0$
$(ghx\sigma\sigma) : A g_{x\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_x + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{x} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{x} \otimes z_{\sigma\sigma}\right) + F_{xu_{+}u_{+}}\left(I_{n_x} \otimes vec(\Sigma)\right) = 0$
$F_{xu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_x^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{xu_{+}} \otimes z_{u_{+}} \right)P_{xu\_u} + \left(z_{x} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{x} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
$(ghu\sigma\sigma) : A g_{u\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_{u} + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{u} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{u} \otimes z_{\sigma\sigma}\right) + F_{uu_{+}u_{+}}\left(I_{n_u} \otimes vec(\Sigma_u)\right) = 0$
$F_{uu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_u^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{uu_{+}} \otimes z_{u_{+}} \right)P_{uu\_u} + \left(z_{u} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{u} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
A and B are the common perturbation matrices:
$A = f_{y_0} + \begin{pmatrix} \underbrace{0}_{n\times n_{static}} &\vdots& \underbrace{f_{y^{**}_{+}} \cdot g^{**}_{x}}_{n \times n_{spred}} &\vdots& \underbrace{0}_{n\times n_{frwd}} \end{pmatrix}$and $B = \begin{pmatrix} \underbrace{0}_{n \times n_{static}}&\vdots & \underbrace{0}_{n \times n_{pred}} & \vdots & \underbrace{f_{y^{**}_{+}}}_{n \times n_{sfwrd}} \end{pmatrix}$
and $z=(y_{-}^{\ast}; y; y_{+}^{\ast\ast}; u)$ denotes the dynamic model variables as in `M_.lead_lag_incidence`, $y^\ast$ denote state variables, $y^{\ast\ast}$ denote forward looking variables, $y_+$ denote the variables with a lead, $y_{-}$ denote variables with a lag, $y_0$ denote variables at period t, $f$ the model equations, and $f_z$ the first-order dynamic model derivatives, $f_{zz}$ the second-order dynamic derivatives, and $f_{zzz}$ the third-order dynamic model derivatives. Then:
$z_{x} = \begin{pmatrix}I\\g_{x}\\g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{u} =\begin{pmatrix}0\\g_{u}\\g^{**}_{x} \cdot g^{*}_{u}\\I\end{pmatrix}$, $z_{u_{+}} =\begin{pmatrix}0\\0\\g^{**}_{u}\\0\end{pmatrix}$
$z_{xx} = \begin{pmatrix} 0\\g_{xx}\\g^{**}_{x} \left( g^{*}_x \otimes g^{*}_{x} \right) + g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{xu} =\begin{pmatrix}0\\g_{xu}\\g^{**}_{xx} \left( g^{*}_x \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{xu}\\0\end{pmatrix}$, $z_{uu} =\begin{pmatrix}0\\g_{uu}\\g^{**}_{xx} \left( g^{*}_u \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{uu}\\0\end{pmatrix}$,
$z_{xu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_x \otimes I \right)\\0\end{pmatrix}$, $z_{uu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_{u} \otimes I \right)\\0\end{pmatrix}$, $z_{u_{+}u_{+}} =\begin{pmatrix}0\\0\\g^{\ast\ast}_{uu}\\0\end{pmatrix}$, $z_{\sigma\sigma} = \begin{pmatrix}0\\ g_{\sigma\sigma}\\ g^{\ast\ast}_{x}g^{\ast}_{\sigma\sigma} + g^{\ast\ast}_{\sigma\sigma}\\0 \end{pmatrix}$
$P$ are permutation matrices that can be computed using Matlab's `ipermute` function.
## Parameter derivatives of perturbation solutions
First, we need the parameter derivatives of first, second, third, and fourth derivatives of the dynamic model (i.e. g1,g2,g3,g4 in dynamic files), I make use of the implicit function theorem: Let $f_{z^k}$ denote the kth derivative (wrt all dynamic variables) of the dynamic model, then let $df_{z^k}$ denote the first-derivative (wrt all model parameters) of $f_{z^k}$ evaluated at the steady state. Note that $f_{z^k}$ is a function of both the model parameters $\theta$ and of the steady state of all dynamic variables $\bar{z}$, which also depend on the parameters. Hence, implicitly $f_{z^k}=f_{z^k}(\theta,\bar{z}(\theta))$ and $df_{z^k}$ consists of two parts:
1. direct derivative wrt to all model parameters given by the preprocessor in the `_params_derivs.m` files
2. contribution of derivative of steady state of dynamic variables (wrt all model parameters): $f_{z^{k+1}} \cdot d\bar{z}$
Note that we already have functionality to compute $d\bar{z}$ analytically.
Having this, the above perturbation systems are basically equations of the following types
$AX +BXC = RHS$ or $AX = RHS$
Now when taking the derivative (wrt to one single parameter $\theta_j$), we get
$A\mathrm{d}\{X\} + B\mathrm{d}\{X\}C = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X - \mathrm{d}\{B\}XC - BX\mathrm{d}\{C\}$
or
$A\mathrm{d}\{X\} = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X$
The first one is a Sylvester type equation, the second one can be solved by taking the inverse of $A$. The only diffculty and tedious work arrises in computing (the highly sparse) derivatives of $RHS$.
***
# New functions: `
## get_perturbation_params_derivs.m`and `get_perturbation_params_derivs_numerical_objective.m`
* The parameter derivatives up to third order are computed in the new function`get_perturbation_params_derivs.m` both analytically and numerically. For numerical derivatives `get_perturbation_params_derivs_numerical_objective.m` is the objective for `fjaco.m` or `hessian_sparse.m` or `hessian.m`.
* `get_perturbation_params_derivs.m` is basically an extended version of the previous `get_first_order_solution_params_derivs.m` function.
* * `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It is used for numerical derivatives, whenever `analytic_derivation_mode=-1|-2`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
* Detailed changes:
* Most important: notation of this function is now in accordance to the k_order_solver, i.e. we do not compute derivatives of Kalman transition matrices A and B, but rather the solution matrices ghx,ghu,ghxx,ghxu,ghuu,ghs2,ghxxx,ghxxu,ghxuu,ghuuu,ghxss,ghuss in accordance with notation used in `oo_.dr`. As a byproduct at first-order, focusing on ghx and ghu instead of Kalman transition matrices A and B makes the computations slightly faster for large models (e.g. for Quest the computations were faster by a couple of seconds, not much, but okay).
* Removed use of `kstate`, see also Dynare/dynare#1653 and Dynare/dynare!1656
* Output arguments are stored in a structure `DERIVS`, there is also a flag `d2flag` that computes parameter hessians needed only in `dsge_likelihood.m`.
* Removed `kronflag` as input. `options_.analytic_derivation_mode` is now used instead of `kronflag`.
* Removed `indvar`, an index that was used to selected specific variables in the derivatives. This is not needed, as we always compute the parameter derivatives for all variables first and then select a subset of variables. The selection now takes place in other functions, like `dsge_likelihood.m`.
* Introduced some checks: (i) deterministic exogenous variables are not supported, (ii) Kronecker method only compatible with first-order approximation so reset to sylvester method, (iii) for purely backward or forward models we need to be careful with the rows in `M_.lead_la g_incidence`, (iv) if `_params_derivs.m` files are missing an error is thrown.
* For numerical derivatives, if mod file does not contain an `estimated_params_block`, a temporary one with the most important parameter information is created.
## `unfold_g4.m`
* When evaluating g3 and g4 one needs to take into account that these do not contain symmetric elements, so one needs to use `unfold_g3.m` and the new function `unfold_g4.m`. This returns an unfolded version of the same matrix (i.e. with symmetric elements).
***
# New test models
`.gitignore` and `Makefile.am` are changed accordingly. Also now it is possible to run test suite on analytic_derivatives, i.e. run `make check m/analytic_derivatives`
## `analytic_derivatives/BrockMirman_PertParamsDerivs.mod`
* This is the Brock Mirman model, where we know the exact policy function $g$ for capital and consumption. As this does not imply a nonzero $g_{\sigma\sigma}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ I added some artificial equations to get nonzero solution matrices with respect to $\sigma$. The true perturbation solution matrices $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ are then computed analytically with Matlab's symbolic toolbox and saved in `nBrockMirmanSYM.mat`. There is a preprocessor flag that recreates these analytical computations if changes are needed (and to check whether I made some errors here ;-) )
* Then solution matrices up to third order and their parameter Jacobians are then compared to the ones computed by Dynare's `k_order_solver` and by `get_perturbation_params_derivs` for all `analytic_derivation_mode`'s. There will be an error if the maximum absolute deviation is too large, i.e. for numerical derivatives (`analytic_derivation_mode=-1|-2`) the tolerance is choosen lower (around 1e-5); for analytical methods we are stricter: around 1e-13 for first-order, 1e-12 for second order, and 1e-11 for third-order.
* As a side note, this mod file also checks Dynare's `k_order_solver` algorithm and throws an error if something is wrong.
* This test model shows that the new functionality works well. And analytical derivatives perform way better and accurate than numerical ones, even for this small model.
## `analytic_derivatives/burnside_3_order_PertParamsDerivs.mod`
* This builds upon `tests/k_order_perturbation/burnside_k_order.mod` and computes the true parameter derivatives analytically by hand.
* This test model also shows that the new functionality works well.
## `analytic_derivatives/LindeTrabandt2019.mod`
* Shows that the new functionality also works for medium-sized models, i.e. a SW type model solved at third order with 35 variables (11 states). 2 shocks and 20 parameters.
* This mod file can be used to tweak the speed of the computations in the future.
* Compares numerical versus analytical parameter derivatives (for first, second and third order). Note that this model clearly shows that numerical ones are quite different than analytical ones even at first order!
## `identification/LindeTrabandt2019_xfail.mod`
* This model is a check for issue Dynare/dynare#1595, see fjaco.m below, and will fail.
* Removed `analytic_derivatives/ls2003.mod` as this mod file is neither in the testsuite nor does it work.
***
# Detailed changes in other functions
## `get_first_order_solution_params_derivs.m`
* Deleted, or actually, renamed to `get_perturbation_params_derivs.m`, as this function now is able to compute the derivatives up to third order
## `identification_numerical_objective.m`
* `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
## `dsge_likelihood.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `T`matrix anymore, but instead of the dynare solution matrix `ghx`. So we recreate `T` here as this amounts to adding some zeros and focusing on selected variables only.
* Added some checks to make sure the first-order approximation is selected.
* Removed `kron_flag` as input, as `get_perturbation_params_derivs` looks into `options_.analytic_derivation_mode` for `kron_flag`.
## `dynare_identification.m`
* make sure that setting `analytic_derivation_mode` is set both in `options_ident` and `options_`. Note that at the end of the function we restore the `options_` structure, so all changes are local. In a next merge request, I will remove the global variables to make all variables local.
## `get_identification_jacobians.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `A` and `B` matrix anymore, but instead of the dynare solution matrix `ghx` and `ghu`. So we recreate these matrices here instead of in `get_perturbation_params_derivs.m`.
* Added `str2func` for better function handles in `fjaco.m`.
## `fjaco.m`
* make `tol`an option, which can be adjusted by changing `options_.dynatol.x`for identification and parameter derivatives purposes.
* include a check and an informative error message, if numerical derivatives (two-sided finite difference method) yield errors in `resol.m` for identification and parameter derivatives purposes. This closes issue Dynare/dynare#1595.
* Changed year of copyright to 2010-2017,2019
***
# Further suggestions and questions
* Ones this is merged, I will merge request an improvement of the identification toolbox, which will work up to third order using the pruned state space. This will also remove some issues and bugs, and also I will remove global variables in this request.
* The third-order derivatives can be further improved by taking sparsity into account and use mex versions for kronecker products etc. I leave this for further testing (and if anybody actually uses this ;-) )
2019-12-17 19:17:09 +01:00
dG_x_x ( : , : , jp ) = kron ( dghx ( idx_states , : , jp ) , oo . dr . ghx ( idx_states , : ) ) + kron ( oo . dr . ghx ( idx_states , : ) , dghx ( idx_states , : , jp ) ) ;
dRHSghxx ( : , : , jp ) = - ( ( dRHS_1 + dRHS_2 + dRHS_3 ) + dA ( : , : , jp ) * oo . dr . ghxx + dB ( : , : , jp ) * oo . dr . ghxx * G_x_x + B * oo . dr . ghxx * dG_x_x ( : , : , jp ) ) ;
end
%use iterated generalized sylvester equation to compute dghxx
dghxx = sylvester3 ( A , B , G_x_x , dRHSghxx ) ;
flag = 1 ; icount = 0 ;
while flag && icount < 4
[ dghxx , flag ] = sylvester3a ( dghxx , A , B , G_x_x , dRHSghxx ) ;
icount = icount + 1 ;
end
zu = [ zeros ( M . nspred , M . exo_nbr ) ;
oo . dr . ghu ( kcurr ~= 0 , : ) ;
oo . dr . ghx ( klead ~= 0 , : ) * oo . dr . ghu ( idx_states , : ) ;
eye ( M . exo_nbr ) ] ;
2020-01-10 17:55:57 +01:00
abcOutxu = A_times_B_kronecker_C ( oo . dr . ghxx , oo . dr . ghx ( idx_states , : ) , oo . dr . ghu ( idx_states , : ) ) ; %auxiliary expressions for dghxu
abcOutuu = A_times_B_kronecker_C ( oo . dr . ghxx , oo . dr . ghu ( idx_states , : ) ) ; %auxiliary expressions for dghuu
RHSs2 = sparse_hessian_times_B_kronecker_C ( g2 ( : , k2yy0ex0 ( nonzeros ( klead ) , nonzeros ( klead ) ) ) , oo . dr . ghu ( klead ~= 0 , : ) , threads_BC ) ;
Added parameter derivatives of perturbation solution up to 3 order
# Preliminary comments
I finished the identification toolbox at orders two and three using the pruned state space system, but before I merge request this, I decided to first merge the new functionality to compute parameter derivatives of perturbation solution matrices at higher orders. So after this is approved, I merge the identification toolbox.
I guess @rattoma, @sebastien, and @michel are best choices to review this.
I outline the main idea first and then provide some more detailed changes I made to the functions.
***
# Main idea
This merge request is concerned with the *analytical*computation of the parameter derivatives of first, second and third order perturbation solution matrices, i.e. using _closed-form_ expressions to efficiently compute the derivative of $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ *with respect to model parameters* $\theta$. Note that $\theta$ contains model parameters, stderr and corr parameters of shocks. stderr and corr parameters of measurement errors are not yet supported, (they can easily be included as exogenous shocks). The availability of such derivatives is beneficial in terms of more reliable analysis of model sensitivity and parameter identifiability as well as more efficient estimation methods, in particular for models solved up to third order, as it is well-known that numerical derivatives are a tricky business, especially for large models.
References for my approach are:
* Iskrev (2008, 2010) and Schmitt-Grohé and Uribe (2012, Appendix) who were the first to compute the parameter derivatives analytically at first order, however, using inefficient (sparse) Kronecker products.
* Mutschler (2015) who provides the expressions for a second-order, but again using inefficient (sparse) Kronecker products.
* Ratto and Iskrev (2012) who show how the first-order system can be solved accurately, fast and efficiently using existing numerical algorithms for generalized Sylvester equations by taking the parameter derivative with respect to each parameter separately.
* Julliard and Kamenik (2004) who provide the perturbation solution equation system in tensor notation at any order k.
* Levintal (2017) who introduces permutation matrices to express the perturbation solution equation system in matrix notation up to fifth order.
Note that @rattoma already implemented the parameter derivatives of $g_x$ and $g_u$ analytically (and numerically), and I rely heavily on his work in `get_first_order_solution_params_derivs.m` (previously `getH.m`). My additions are mainly to this function and thus it is renamed to `get_perturbation_params_derivs.m`.
The basic idea of this merge request is to take the second- and third-order perturbation solution systems in Julliard and Kamenik (2004), unfold these into an equivalent matrix representation using permutation matrices as in Levintal (2017). Then extending Ratto and Iskrev (2012) one takes the derivative with respect to each parameter separately and gets a computational problem that is linear, albeit large, as it involves either solving generalized Sylvester equations or taking inverses of highly sparse matrices. I will now briefly summarize the perturbation solution system at third order and the system that results when taking the derivative with respect to parameters.
## Perturbation Solution
The following systems arise at first, second, and third order:
$(ghx): f_{x} z_{x} = f_{y_{-}^*} + f_{y_0} g_{x} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{x}= A g_{x} + f_{y_{-}^*}=0$
$(ghu): f_{z} z_{u} = f_{y_0} g_{u} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{u} + f_{u}= A g_u + f_u = 0$
$(ghxx) : A g_{xx} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{x}\right) + f_{zz} \left( z_{x} \otimes z_{x} \right) = 0$
$(ghxu) : A g_{xu} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{x} \otimes z_{u} \right) = 0$
$(ghuu) : A g_{uu} + B g_{xx} \left(g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{u} \otimes z_{u} \right) = 0$
$(ghs2) : (A+B) g_{\sigma\sigma} + \left( f_{y^{**}_{+}y^{**}_{+}} \left(g^{**}_{u} \otimes g^{**}_{u}\right) + f_{y^{**}_{+}} g^{**}_{uu}\right)vec(\Sigma) = 0$
$(ghxxx) : A g_{xxx} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{x}\right) + f_{y_{+}}g^{**}_{xx} \left(g^{*}_x \otimes g^{*}_{xx}\right)P_{x\_xx} + f_{zz} \left( z_{x} \otimes z_{xx} \right)P_{x\_xx} + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{x} \right) = 0$
$(ghxxu) : A g_{xxu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{u} \right) + f_{zz} \left( \left( z_{x} \otimes z_{xu} \right)P_{x\_xu} + \left(z_{xx} \otimes z_{u}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{x} \otimes g^{*}_{xu}\right)P_{x\_xu} + \left(g^{*}_{xx} \otimes g^{*}_{u}\right) \right) = 0$
$(ghxuu) : A g_{xuu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( \left( z_{xu} \otimes z_{u} \right)P_{xu\_u} + \left(z_{x} \otimes z_{uu}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{xu} \otimes g^{*}_{u}\right)P_{xu\_u} + \left(g^{*}_{x} \otimes g^{*}_{uu}\right) \right) = 0$
$(ghuuu) : A g_{uuu} + B g_{xxx} \left(g^{*}_{u} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{u} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( z_{u} \otimes z_{uu} \right)P_{u\_uu} + f_{y_{+}}g^{**}_{xx} \left(g^{*}_{u} \otimes g^{*}_{uu}\right)P_{u\_uu} = 0$
$(ghx\sigma\sigma) : A g_{x\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_x + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{x} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{x} \otimes z_{\sigma\sigma}\right) + F_{xu_{+}u_{+}}\left(I_{n_x} \otimes vec(\Sigma)\right) = 0$
$F_{xu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_x^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{xu_{+}} \otimes z_{u_{+}} \right)P_{xu\_u} + \left(z_{x} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{x} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
$(ghu\sigma\sigma) : A g_{u\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_{u} + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{u} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{u} \otimes z_{\sigma\sigma}\right) + F_{uu_{+}u_{+}}\left(I_{n_u} \otimes vec(\Sigma_u)\right) = 0$
$F_{uu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_u^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{uu_{+}} \otimes z_{u_{+}} \right)P_{uu\_u} + \left(z_{u} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{u} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
A and B are the common perturbation matrices:
$A = f_{y_0} + \begin{pmatrix} \underbrace{0}_{n\times n_{static}} &\vdots& \underbrace{f_{y^{**}_{+}} \cdot g^{**}_{x}}_{n \times n_{spred}} &\vdots& \underbrace{0}_{n\times n_{frwd}} \end{pmatrix}$and $B = \begin{pmatrix} \underbrace{0}_{n \times n_{static}}&\vdots & \underbrace{0}_{n \times n_{pred}} & \vdots & \underbrace{f_{y^{**}_{+}}}_{n \times n_{sfwrd}} \end{pmatrix}$
and $z=(y_{-}^{\ast}; y; y_{+}^{\ast\ast}; u)$ denotes the dynamic model variables as in `M_.lead_lag_incidence`, $y^\ast$ denote state variables, $y^{\ast\ast}$ denote forward looking variables, $y_+$ denote the variables with a lead, $y_{-}$ denote variables with a lag, $y_0$ denote variables at period t, $f$ the model equations, and $f_z$ the first-order dynamic model derivatives, $f_{zz}$ the second-order dynamic derivatives, and $f_{zzz}$ the third-order dynamic model derivatives. Then:
$z_{x} = \begin{pmatrix}I\\g_{x}\\g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{u} =\begin{pmatrix}0\\g_{u}\\g^{**}_{x} \cdot g^{*}_{u}\\I\end{pmatrix}$, $z_{u_{+}} =\begin{pmatrix}0\\0\\g^{**}_{u}\\0\end{pmatrix}$
$z_{xx} = \begin{pmatrix} 0\\g_{xx}\\g^{**}_{x} \left( g^{*}_x \otimes g^{*}_{x} \right) + g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{xu} =\begin{pmatrix}0\\g_{xu}\\g^{**}_{xx} \left( g^{*}_x \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{xu}\\0\end{pmatrix}$, $z_{uu} =\begin{pmatrix}0\\g_{uu}\\g^{**}_{xx} \left( g^{*}_u \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{uu}\\0\end{pmatrix}$,
$z_{xu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_x \otimes I \right)\\0\end{pmatrix}$, $z_{uu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_{u} \otimes I \right)\\0\end{pmatrix}$, $z_{u_{+}u_{+}} =\begin{pmatrix}0\\0\\g^{\ast\ast}_{uu}\\0\end{pmatrix}$, $z_{\sigma\sigma} = \begin{pmatrix}0\\ g_{\sigma\sigma}\\ g^{\ast\ast}_{x}g^{\ast}_{\sigma\sigma} + g^{\ast\ast}_{\sigma\sigma}\\0 \end{pmatrix}$
$P$ are permutation matrices that can be computed using Matlab's `ipermute` function.
## Parameter derivatives of perturbation solutions
First, we need the parameter derivatives of first, second, third, and fourth derivatives of the dynamic model (i.e. g1,g2,g3,g4 in dynamic files), I make use of the implicit function theorem: Let $f_{z^k}$ denote the kth derivative (wrt all dynamic variables) of the dynamic model, then let $df_{z^k}$ denote the first-derivative (wrt all model parameters) of $f_{z^k}$ evaluated at the steady state. Note that $f_{z^k}$ is a function of both the model parameters $\theta$ and of the steady state of all dynamic variables $\bar{z}$, which also depend on the parameters. Hence, implicitly $f_{z^k}=f_{z^k}(\theta,\bar{z}(\theta))$ and $df_{z^k}$ consists of two parts:
1. direct derivative wrt to all model parameters given by the preprocessor in the `_params_derivs.m` files
2. contribution of derivative of steady state of dynamic variables (wrt all model parameters): $f_{z^{k+1}} \cdot d\bar{z}$
Note that we already have functionality to compute $d\bar{z}$ analytically.
Having this, the above perturbation systems are basically equations of the following types
$AX +BXC = RHS$ or $AX = RHS$
Now when taking the derivative (wrt to one single parameter $\theta_j$), we get
$A\mathrm{d}\{X\} + B\mathrm{d}\{X\}C = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X - \mathrm{d}\{B\}XC - BX\mathrm{d}\{C\}$
or
$A\mathrm{d}\{X\} = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X$
The first one is a Sylvester type equation, the second one can be solved by taking the inverse of $A$. The only diffculty and tedious work arrises in computing (the highly sparse) derivatives of $RHS$.
***
# New functions: `
## get_perturbation_params_derivs.m`and `get_perturbation_params_derivs_numerical_objective.m`
* The parameter derivatives up to third order are computed in the new function`get_perturbation_params_derivs.m` both analytically and numerically. For numerical derivatives `get_perturbation_params_derivs_numerical_objective.m` is the objective for `fjaco.m` or `hessian_sparse.m` or `hessian.m`.
* `get_perturbation_params_derivs.m` is basically an extended version of the previous `get_first_order_solution_params_derivs.m` function.
* * `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It is used for numerical derivatives, whenever `analytic_derivation_mode=-1|-2`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
* Detailed changes:
* Most important: notation of this function is now in accordance to the k_order_solver, i.e. we do not compute derivatives of Kalman transition matrices A and B, but rather the solution matrices ghx,ghu,ghxx,ghxu,ghuu,ghs2,ghxxx,ghxxu,ghxuu,ghuuu,ghxss,ghuss in accordance with notation used in `oo_.dr`. As a byproduct at first-order, focusing on ghx and ghu instead of Kalman transition matrices A and B makes the computations slightly faster for large models (e.g. for Quest the computations were faster by a couple of seconds, not much, but okay).
* Removed use of `kstate`, see also Dynare/dynare#1653 and Dynare/dynare!1656
* Output arguments are stored in a structure `DERIVS`, there is also a flag `d2flag` that computes parameter hessians needed only in `dsge_likelihood.m`.
* Removed `kronflag` as input. `options_.analytic_derivation_mode` is now used instead of `kronflag`.
* Removed `indvar`, an index that was used to selected specific variables in the derivatives. This is not needed, as we always compute the parameter derivatives for all variables first and then select a subset of variables. The selection now takes place in other functions, like `dsge_likelihood.m`.
* Introduced some checks: (i) deterministic exogenous variables are not supported, (ii) Kronecker method only compatible with first-order approximation so reset to sylvester method, (iii) for purely backward or forward models we need to be careful with the rows in `M_.lead_la g_incidence`, (iv) if `_params_derivs.m` files are missing an error is thrown.
* For numerical derivatives, if mod file does not contain an `estimated_params_block`, a temporary one with the most important parameter information is created.
## `unfold_g4.m`
* When evaluating g3 and g4 one needs to take into account that these do not contain symmetric elements, so one needs to use `unfold_g3.m` and the new function `unfold_g4.m`. This returns an unfolded version of the same matrix (i.e. with symmetric elements).
***
# New test models
`.gitignore` and `Makefile.am` are changed accordingly. Also now it is possible to run test suite on analytic_derivatives, i.e. run `make check m/analytic_derivatives`
## `analytic_derivatives/BrockMirman_PertParamsDerivs.mod`
* This is the Brock Mirman model, where we know the exact policy function $g$ for capital and consumption. As this does not imply a nonzero $g_{\sigma\sigma}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ I added some artificial equations to get nonzero solution matrices with respect to $\sigma$. The true perturbation solution matrices $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ are then computed analytically with Matlab's symbolic toolbox and saved in `nBrockMirmanSYM.mat`. There is a preprocessor flag that recreates these analytical computations if changes are needed (and to check whether I made some errors here ;-) )
* Then solution matrices up to third order and their parameter Jacobians are then compared to the ones computed by Dynare's `k_order_solver` and by `get_perturbation_params_derivs` for all `analytic_derivation_mode`'s. There will be an error if the maximum absolute deviation is too large, i.e. for numerical derivatives (`analytic_derivation_mode=-1|-2`) the tolerance is choosen lower (around 1e-5); for analytical methods we are stricter: around 1e-13 for first-order, 1e-12 for second order, and 1e-11 for third-order.
* As a side note, this mod file also checks Dynare's `k_order_solver` algorithm and throws an error if something is wrong.
* This test model shows that the new functionality works well. And analytical derivatives perform way better and accurate than numerical ones, even for this small model.
## `analytic_derivatives/burnside_3_order_PertParamsDerivs.mod`
* This builds upon `tests/k_order_perturbation/burnside_k_order.mod` and computes the true parameter derivatives analytically by hand.
* This test model also shows that the new functionality works well.
## `analytic_derivatives/LindeTrabandt2019.mod`
* Shows that the new functionality also works for medium-sized models, i.e. a SW type model solved at third order with 35 variables (11 states). 2 shocks and 20 parameters.
* This mod file can be used to tweak the speed of the computations in the future.
* Compares numerical versus analytical parameter derivatives (for first, second and third order). Note that this model clearly shows that numerical ones are quite different than analytical ones even at first order!
## `identification/LindeTrabandt2019_xfail.mod`
* This model is a check for issue Dynare/dynare#1595, see fjaco.m below, and will fail.
* Removed `analytic_derivatives/ls2003.mod` as this mod file is neither in the testsuite nor does it work.
***
# Detailed changes in other functions
## `get_first_order_solution_params_derivs.m`
* Deleted, or actually, renamed to `get_perturbation_params_derivs.m`, as this function now is able to compute the derivatives up to third order
## `identification_numerical_objective.m`
* `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
## `dsge_likelihood.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `T`matrix anymore, but instead of the dynare solution matrix `ghx`. So we recreate `T` here as this amounts to adding some zeros and focusing on selected variables only.
* Added some checks to make sure the first-order approximation is selected.
* Removed `kron_flag` as input, as `get_perturbation_params_derivs` looks into `options_.analytic_derivation_mode` for `kron_flag`.
## `dynare_identification.m`
* make sure that setting `analytic_derivation_mode` is set both in `options_ident` and `options_`. Note that at the end of the function we restore the `options_` structure, so all changes are local. In a next merge request, I will remove the global variables to make all variables local.
## `get_identification_jacobians.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `A` and `B` matrix anymore, but instead of the dynare solution matrix `ghx` and `ghu`. So we recreate these matrices here instead of in `get_perturbation_params_derivs.m`.
* Added `str2func` for better function handles in `fjaco.m`.
## `fjaco.m`
* make `tol`an option, which can be adjusted by changing `options_.dynatol.x`for identification and parameter derivatives purposes.
* include a check and an informative error message, if numerical derivatives (two-sided finite difference method) yield errors in `resol.m` for identification and parameter derivatives purposes. This closes issue Dynare/dynare#1595.
* Changed year of copyright to 2010-2017,2019
***
# Further suggestions and questions
* Ones this is merged, I will merge request an improvement of the identification toolbox, which will work up to third order using the pruned state space. This will also remove some issues and bugs, and also I will remove global variables in this request.
* The third-order derivatives can be further improved by taking sparsity into account and use mex versions for kronecker products etc. I leave this for further testing (and if anybody actually uses this ;-) )
2019-12-17 19:17:09 +01:00
RHSs2 = RHSs2 + g1 ( : , nonzeros ( klead ) ) * oo . dr . ghuu ( klead ~= 0 , : ) ;
dzu = zeros ( size ( zu , 1 ) , size ( zu , 2 ) , modparam_nbr ) ;
dghxu = zeros ( M . endo_nbr , M . nspred * M . exo_nbr , modparam_nbr ) ;
dghuu = zeros ( M . endo_nbr , M . exo_nbr * M . exo_nbr , modparam_nbr ) ;
dghs2 = zeros ( M . endo_nbr , totparam_nbr ) ; %note that for modparam we ignore the contribution by dSigma_e and add it after the computations for stderr and corr
for jp = 1 : modparam_nbr
dzu ( : , : , jp ) = [ zeros ( M . nspred , M . exo_nbr ) ;
dghu ( kcurr ~= 0 , : , jp ) ;
dghx ( klead ~= 0 , : , jp ) * oo . dr . ghu ( idx_states , : ) + oo . dr . ghx ( klead ~= 0 , : ) * dghu ( idx_states , : , jp ) ;
zeros ( M . exo_nbr , M . exo_nbr ) ] ;
%compute dghxu
2020-01-10 17:55:57 +01:00
dRHS_1 = sparse_hessian_times_B_kronecker_C ( dg2 ( : , k2yy0ex0 ( kyy0ex0 , kyy0ex0 ) + ( jp - 1 ) * yy0ex0_nbr ^2 ) , zx , zu , threads_BC ) ;
dRHS_2 = sparse_hessian_times_B_kronecker_C ( g2 ( : , k2yy0ex0 ( kyy0ex0 , kyy0ex0 ) ) , dzx ( : , : , jp ) , zu , threads_BC ) ;
dRHS_3 = sparse_hessian_times_B_kronecker_C ( g2 ( : , k2yy0ex0 ( kyy0ex0 , kyy0ex0 ) ) , zx , dzu ( : , : , jp ) , threads_BC ) ;
dabcOut_1 = A_times_B_kronecker_C ( dghxx ( : , : , jp ) , oo . dr . ghx ( idx_states , : ) , oo . dr . ghu ( idx_states , : ) ) ;
dabcOut_2 = A_times_B_kronecker_C ( oo . dr . ghxx , dghx ( idx_states , : , jp ) , oo . dr . ghu ( idx_states , : ) ) ;
dabcOut_3 = A_times_B_kronecker_C ( oo . dr . ghxx , oo . dr . ghx ( idx_states , : ) , dghu ( idx_states , : , jp ) ) ;
Added parameter derivatives of perturbation solution up to 3 order
# Preliminary comments
I finished the identification toolbox at orders two and three using the pruned state space system, but before I merge request this, I decided to first merge the new functionality to compute parameter derivatives of perturbation solution matrices at higher orders. So after this is approved, I merge the identification toolbox.
I guess @rattoma, @sebastien, and @michel are best choices to review this.
I outline the main idea first and then provide some more detailed changes I made to the functions.
***
# Main idea
This merge request is concerned with the *analytical*computation of the parameter derivatives of first, second and third order perturbation solution matrices, i.e. using _closed-form_ expressions to efficiently compute the derivative of $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ *with respect to model parameters* $\theta$. Note that $\theta$ contains model parameters, stderr and corr parameters of shocks. stderr and corr parameters of measurement errors are not yet supported, (they can easily be included as exogenous shocks). The availability of such derivatives is beneficial in terms of more reliable analysis of model sensitivity and parameter identifiability as well as more efficient estimation methods, in particular for models solved up to third order, as it is well-known that numerical derivatives are a tricky business, especially for large models.
References for my approach are:
* Iskrev (2008, 2010) and Schmitt-Grohé and Uribe (2012, Appendix) who were the first to compute the parameter derivatives analytically at first order, however, using inefficient (sparse) Kronecker products.
* Mutschler (2015) who provides the expressions for a second-order, but again using inefficient (sparse) Kronecker products.
* Ratto and Iskrev (2012) who show how the first-order system can be solved accurately, fast and efficiently using existing numerical algorithms for generalized Sylvester equations by taking the parameter derivative with respect to each parameter separately.
* Julliard and Kamenik (2004) who provide the perturbation solution equation system in tensor notation at any order k.
* Levintal (2017) who introduces permutation matrices to express the perturbation solution equation system in matrix notation up to fifth order.
Note that @rattoma already implemented the parameter derivatives of $g_x$ and $g_u$ analytically (and numerically), and I rely heavily on his work in `get_first_order_solution_params_derivs.m` (previously `getH.m`). My additions are mainly to this function and thus it is renamed to `get_perturbation_params_derivs.m`.
The basic idea of this merge request is to take the second- and third-order perturbation solution systems in Julliard and Kamenik (2004), unfold these into an equivalent matrix representation using permutation matrices as in Levintal (2017). Then extending Ratto and Iskrev (2012) one takes the derivative with respect to each parameter separately and gets a computational problem that is linear, albeit large, as it involves either solving generalized Sylvester equations or taking inverses of highly sparse matrices. I will now briefly summarize the perturbation solution system at third order and the system that results when taking the derivative with respect to parameters.
## Perturbation Solution
The following systems arise at first, second, and third order:
$(ghx): f_{x} z_{x} = f_{y_{-}^*} + f_{y_0} g_{x} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{x}= A g_{x} + f_{y_{-}^*}=0$
$(ghu): f_{z} z_{u} = f_{y_0} g_{u} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{u} + f_{u}= A g_u + f_u = 0$
$(ghxx) : A g_{xx} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{x}\right) + f_{zz} \left( z_{x} \otimes z_{x} \right) = 0$
$(ghxu) : A g_{xu} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{x} \otimes z_{u} \right) = 0$
$(ghuu) : A g_{uu} + B g_{xx} \left(g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{u} \otimes z_{u} \right) = 0$
$(ghs2) : (A+B) g_{\sigma\sigma} + \left( f_{y^{**}_{+}y^{**}_{+}} \left(g^{**}_{u} \otimes g^{**}_{u}\right) + f_{y^{**}_{+}} g^{**}_{uu}\right)vec(\Sigma) = 0$
$(ghxxx) : A g_{xxx} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{x}\right) + f_{y_{+}}g^{**}_{xx} \left(g^{*}_x \otimes g^{*}_{xx}\right)P_{x\_xx} + f_{zz} \left( z_{x} \otimes z_{xx} \right)P_{x\_xx} + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{x} \right) = 0$
$(ghxxu) : A g_{xxu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{u} \right) + f_{zz} \left( \left( z_{x} \otimes z_{xu} \right)P_{x\_xu} + \left(z_{xx} \otimes z_{u}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{x} \otimes g^{*}_{xu}\right)P_{x\_xu} + \left(g^{*}_{xx} \otimes g^{*}_{u}\right) \right) = 0$
$(ghxuu) : A g_{xuu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( \left( z_{xu} \otimes z_{u} \right)P_{xu\_u} + \left(z_{x} \otimes z_{uu}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{xu} \otimes g^{*}_{u}\right)P_{xu\_u} + \left(g^{*}_{x} \otimes g^{*}_{uu}\right) \right) = 0$
$(ghuuu) : A g_{uuu} + B g_{xxx} \left(g^{*}_{u} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{u} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( z_{u} \otimes z_{uu} \right)P_{u\_uu} + f_{y_{+}}g^{**}_{xx} \left(g^{*}_{u} \otimes g^{*}_{uu}\right)P_{u\_uu} = 0$
$(ghx\sigma\sigma) : A g_{x\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_x + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{x} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{x} \otimes z_{\sigma\sigma}\right) + F_{xu_{+}u_{+}}\left(I_{n_x} \otimes vec(\Sigma)\right) = 0$
$F_{xu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_x^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{xu_{+}} \otimes z_{u_{+}} \right)P_{xu\_u} + \left(z_{x} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{x} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
$(ghu\sigma\sigma) : A g_{u\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_{u} + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{u} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{u} \otimes z_{\sigma\sigma}\right) + F_{uu_{+}u_{+}}\left(I_{n_u} \otimes vec(\Sigma_u)\right) = 0$
$F_{uu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_u^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{uu_{+}} \otimes z_{u_{+}} \right)P_{uu\_u} + \left(z_{u} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{u} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
A and B are the common perturbation matrices:
$A = f_{y_0} + \begin{pmatrix} \underbrace{0}_{n\times n_{static}} &\vdots& \underbrace{f_{y^{**}_{+}} \cdot g^{**}_{x}}_{n \times n_{spred}} &\vdots& \underbrace{0}_{n\times n_{frwd}} \end{pmatrix}$and $B = \begin{pmatrix} \underbrace{0}_{n \times n_{static}}&\vdots & \underbrace{0}_{n \times n_{pred}} & \vdots & \underbrace{f_{y^{**}_{+}}}_{n \times n_{sfwrd}} \end{pmatrix}$
and $z=(y_{-}^{\ast}; y; y_{+}^{\ast\ast}; u)$ denotes the dynamic model variables as in `M_.lead_lag_incidence`, $y^\ast$ denote state variables, $y^{\ast\ast}$ denote forward looking variables, $y_+$ denote the variables with a lead, $y_{-}$ denote variables with a lag, $y_0$ denote variables at period t, $f$ the model equations, and $f_z$ the first-order dynamic model derivatives, $f_{zz}$ the second-order dynamic derivatives, and $f_{zzz}$ the third-order dynamic model derivatives. Then:
$z_{x} = \begin{pmatrix}I\\g_{x}\\g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{u} =\begin{pmatrix}0\\g_{u}\\g^{**}_{x} \cdot g^{*}_{u}\\I\end{pmatrix}$, $z_{u_{+}} =\begin{pmatrix}0\\0\\g^{**}_{u}\\0\end{pmatrix}$
$z_{xx} = \begin{pmatrix} 0\\g_{xx}\\g^{**}_{x} \left( g^{*}_x \otimes g^{*}_{x} \right) + g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{xu} =\begin{pmatrix}0\\g_{xu}\\g^{**}_{xx} \left( g^{*}_x \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{xu}\\0\end{pmatrix}$, $z_{uu} =\begin{pmatrix}0\\g_{uu}\\g^{**}_{xx} \left( g^{*}_u \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{uu}\\0\end{pmatrix}$,
$z_{xu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_x \otimes I \right)\\0\end{pmatrix}$, $z_{uu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_{u} \otimes I \right)\\0\end{pmatrix}$, $z_{u_{+}u_{+}} =\begin{pmatrix}0\\0\\g^{\ast\ast}_{uu}\\0\end{pmatrix}$, $z_{\sigma\sigma} = \begin{pmatrix}0\\ g_{\sigma\sigma}\\ g^{\ast\ast}_{x}g^{\ast}_{\sigma\sigma} + g^{\ast\ast}_{\sigma\sigma}\\0 \end{pmatrix}$
$P$ are permutation matrices that can be computed using Matlab's `ipermute` function.
## Parameter derivatives of perturbation solutions
First, we need the parameter derivatives of first, second, third, and fourth derivatives of the dynamic model (i.e. g1,g2,g3,g4 in dynamic files), I make use of the implicit function theorem: Let $f_{z^k}$ denote the kth derivative (wrt all dynamic variables) of the dynamic model, then let $df_{z^k}$ denote the first-derivative (wrt all model parameters) of $f_{z^k}$ evaluated at the steady state. Note that $f_{z^k}$ is a function of both the model parameters $\theta$ and of the steady state of all dynamic variables $\bar{z}$, which also depend on the parameters. Hence, implicitly $f_{z^k}=f_{z^k}(\theta,\bar{z}(\theta))$ and $df_{z^k}$ consists of two parts:
1. direct derivative wrt to all model parameters given by the preprocessor in the `_params_derivs.m` files
2. contribution of derivative of steady state of dynamic variables (wrt all model parameters): $f_{z^{k+1}} \cdot d\bar{z}$
Note that we already have functionality to compute $d\bar{z}$ analytically.
Having this, the above perturbation systems are basically equations of the following types
$AX +BXC = RHS$ or $AX = RHS$
Now when taking the derivative (wrt to one single parameter $\theta_j$), we get
$A\mathrm{d}\{X\} + B\mathrm{d}\{X\}C = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X - \mathrm{d}\{B\}XC - BX\mathrm{d}\{C\}$
or
$A\mathrm{d}\{X\} = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X$
The first one is a Sylvester type equation, the second one can be solved by taking the inverse of $A$. The only diffculty and tedious work arrises in computing (the highly sparse) derivatives of $RHS$.
***
# New functions: `
## get_perturbation_params_derivs.m`and `get_perturbation_params_derivs_numerical_objective.m`
* The parameter derivatives up to third order are computed in the new function`get_perturbation_params_derivs.m` both analytically and numerically. For numerical derivatives `get_perturbation_params_derivs_numerical_objective.m` is the objective for `fjaco.m` or `hessian_sparse.m` or `hessian.m`.
* `get_perturbation_params_derivs.m` is basically an extended version of the previous `get_first_order_solution_params_derivs.m` function.
* * `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It is used for numerical derivatives, whenever `analytic_derivation_mode=-1|-2`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
* Detailed changes:
* Most important: notation of this function is now in accordance to the k_order_solver, i.e. we do not compute derivatives of Kalman transition matrices A and B, but rather the solution matrices ghx,ghu,ghxx,ghxu,ghuu,ghs2,ghxxx,ghxxu,ghxuu,ghuuu,ghxss,ghuss in accordance with notation used in `oo_.dr`. As a byproduct at first-order, focusing on ghx and ghu instead of Kalman transition matrices A and B makes the computations slightly faster for large models (e.g. for Quest the computations were faster by a couple of seconds, not much, but okay).
* Removed use of `kstate`, see also Dynare/dynare#1653 and Dynare/dynare!1656
* Output arguments are stored in a structure `DERIVS`, there is also a flag `d2flag` that computes parameter hessians needed only in `dsge_likelihood.m`.
* Removed `kronflag` as input. `options_.analytic_derivation_mode` is now used instead of `kronflag`.
* Removed `indvar`, an index that was used to selected specific variables in the derivatives. This is not needed, as we always compute the parameter derivatives for all variables first and then select a subset of variables. The selection now takes place in other functions, like `dsge_likelihood.m`.
* Introduced some checks: (i) deterministic exogenous variables are not supported, (ii) Kronecker method only compatible with first-order approximation so reset to sylvester method, (iii) for purely backward or forward models we need to be careful with the rows in `M_.lead_la g_incidence`, (iv) if `_params_derivs.m` files are missing an error is thrown.
* For numerical derivatives, if mod file does not contain an `estimated_params_block`, a temporary one with the most important parameter information is created.
## `unfold_g4.m`
* When evaluating g3 and g4 one needs to take into account that these do not contain symmetric elements, so one needs to use `unfold_g3.m` and the new function `unfold_g4.m`. This returns an unfolded version of the same matrix (i.e. with symmetric elements).
***
# New test models
`.gitignore` and `Makefile.am` are changed accordingly. Also now it is possible to run test suite on analytic_derivatives, i.e. run `make check m/analytic_derivatives`
## `analytic_derivatives/BrockMirman_PertParamsDerivs.mod`
* This is the Brock Mirman model, where we know the exact policy function $g$ for capital and consumption. As this does not imply a nonzero $g_{\sigma\sigma}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ I added some artificial equations to get nonzero solution matrices with respect to $\sigma$. The true perturbation solution matrices $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ are then computed analytically with Matlab's symbolic toolbox and saved in `nBrockMirmanSYM.mat`. There is a preprocessor flag that recreates these analytical computations if changes are needed (and to check whether I made some errors here ;-) )
* Then solution matrices up to third order and their parameter Jacobians are then compared to the ones computed by Dynare's `k_order_solver` and by `get_perturbation_params_derivs` for all `analytic_derivation_mode`'s. There will be an error if the maximum absolute deviation is too large, i.e. for numerical derivatives (`analytic_derivation_mode=-1|-2`) the tolerance is choosen lower (around 1e-5); for analytical methods we are stricter: around 1e-13 for first-order, 1e-12 for second order, and 1e-11 for third-order.
* As a side note, this mod file also checks Dynare's `k_order_solver` algorithm and throws an error if something is wrong.
* This test model shows that the new functionality works well. And analytical derivatives perform way better and accurate than numerical ones, even for this small model.
## `analytic_derivatives/burnside_3_order_PertParamsDerivs.mod`
* This builds upon `tests/k_order_perturbation/burnside_k_order.mod` and computes the true parameter derivatives analytically by hand.
* This test model also shows that the new functionality works well.
## `analytic_derivatives/LindeTrabandt2019.mod`
* Shows that the new functionality also works for medium-sized models, i.e. a SW type model solved at third order with 35 variables (11 states). 2 shocks and 20 parameters.
* This mod file can be used to tweak the speed of the computations in the future.
* Compares numerical versus analytical parameter derivatives (for first, second and third order). Note that this model clearly shows that numerical ones are quite different than analytical ones even at first order!
## `identification/LindeTrabandt2019_xfail.mod`
* This model is a check for issue Dynare/dynare#1595, see fjaco.m below, and will fail.
* Removed `analytic_derivatives/ls2003.mod` as this mod file is neither in the testsuite nor does it work.
***
# Detailed changes in other functions
## `get_first_order_solution_params_derivs.m`
* Deleted, or actually, renamed to `get_perturbation_params_derivs.m`, as this function now is able to compute the derivatives up to third order
## `identification_numerical_objective.m`
* `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
## `dsge_likelihood.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `T`matrix anymore, but instead of the dynare solution matrix `ghx`. So we recreate `T` here as this amounts to adding some zeros and focusing on selected variables only.
* Added some checks to make sure the first-order approximation is selected.
* Removed `kron_flag` as input, as `get_perturbation_params_derivs` looks into `options_.analytic_derivation_mode` for `kron_flag`.
## `dynare_identification.m`
* make sure that setting `analytic_derivation_mode` is set both in `options_ident` and `options_`. Note that at the end of the function we restore the `options_` structure, so all changes are local. In a next merge request, I will remove the global variables to make all variables local.
## `get_identification_jacobians.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `A` and `B` matrix anymore, but instead of the dynare solution matrix `ghx` and `ghu`. So we recreate these matrices here instead of in `get_perturbation_params_derivs.m`.
* Added `str2func` for better function handles in `fjaco.m`.
## `fjaco.m`
* make `tol`an option, which can be adjusted by changing `options_.dynatol.x`for identification and parameter derivatives purposes.
* include a check and an informative error message, if numerical derivatives (two-sided finite difference method) yield errors in `resol.m` for identification and parameter derivatives purposes. This closes issue Dynare/dynare#1595.
* Changed year of copyright to 2010-2017,2019
***
# Further suggestions and questions
* Ones this is merged, I will merge request an improvement of the identification toolbox, which will work up to third order using the pruned state space. This will also remove some issues and bugs, and also I will remove global variables in this request.
* The third-order derivatives can be further improved by taking sparsity into account and use mex versions for kronecker products etc. I leave this for further testing (and if anybody actually uses this ;-) )
2019-12-17 19:17:09 +01:00
dghxu ( : , : , jp ) = - invA * ( dA ( : , : , jp ) * oo . dr . ghxu + ( dRHS_1 + dRHS_2 + dRHS_3 ) + dB ( : , : , jp ) * abcOutxu + B * ( dabcOut_1 + dabcOut_2 + dabcOut_3 ) ) ;
%compute dghuu
2020-01-10 17:55:57 +01:00
dRHS_1 = sparse_hessian_times_B_kronecker_C ( dg2 ( : , k2yy0ex0 ( kyy0ex0 , kyy0ex0 ) + ( jp - 1 ) * yy0ex0_nbr ^2 ) , zu , threads_BC ) ;
dRHS_2 = sparse_hessian_times_B_kronecker_C ( g2 ( : , k2yy0ex0 ( kyy0ex0 , kyy0ex0 ) ) , dzu ( : , : , jp ) , zu , threads_BC ) ;
dRHS_3 = sparse_hessian_times_B_kronecker_C ( g2 ( : , k2yy0ex0 ( kyy0ex0 , kyy0ex0 ) ) , zu , dzu ( : , : , jp ) , threads_BC ) ;
dabcOut_1 = A_times_B_kronecker_C ( dghxx ( : , : , jp ) , oo . dr . ghu ( idx_states , : ) ) ;
dabcOut_2 = A_times_B_kronecker_C ( oo . dr . ghxx , dghu ( idx_states , : , jp ) , oo . dr . ghu ( idx_states , : ) ) ;
dabcOut_3 = A_times_B_kronecker_C ( oo . dr . ghxx , oo . dr . ghu ( idx_states , : ) , dghu ( idx_states , : , jp ) ) ;
Added parameter derivatives of perturbation solution up to 3 order
# Preliminary comments
I finished the identification toolbox at orders two and three using the pruned state space system, but before I merge request this, I decided to first merge the new functionality to compute parameter derivatives of perturbation solution matrices at higher orders. So after this is approved, I merge the identification toolbox.
I guess @rattoma, @sebastien, and @michel are best choices to review this.
I outline the main idea first and then provide some more detailed changes I made to the functions.
***
# Main idea
This merge request is concerned with the *analytical*computation of the parameter derivatives of first, second and third order perturbation solution matrices, i.e. using _closed-form_ expressions to efficiently compute the derivative of $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ *with respect to model parameters* $\theta$. Note that $\theta$ contains model parameters, stderr and corr parameters of shocks. stderr and corr parameters of measurement errors are not yet supported, (they can easily be included as exogenous shocks). The availability of such derivatives is beneficial in terms of more reliable analysis of model sensitivity and parameter identifiability as well as more efficient estimation methods, in particular for models solved up to third order, as it is well-known that numerical derivatives are a tricky business, especially for large models.
References for my approach are:
* Iskrev (2008, 2010) and Schmitt-Grohé and Uribe (2012, Appendix) who were the first to compute the parameter derivatives analytically at first order, however, using inefficient (sparse) Kronecker products.
* Mutschler (2015) who provides the expressions for a second-order, but again using inefficient (sparse) Kronecker products.
* Ratto and Iskrev (2012) who show how the first-order system can be solved accurately, fast and efficiently using existing numerical algorithms for generalized Sylvester equations by taking the parameter derivative with respect to each parameter separately.
* Julliard and Kamenik (2004) who provide the perturbation solution equation system in tensor notation at any order k.
* Levintal (2017) who introduces permutation matrices to express the perturbation solution equation system in matrix notation up to fifth order.
Note that @rattoma already implemented the parameter derivatives of $g_x$ and $g_u$ analytically (and numerically), and I rely heavily on his work in `get_first_order_solution_params_derivs.m` (previously `getH.m`). My additions are mainly to this function and thus it is renamed to `get_perturbation_params_derivs.m`.
The basic idea of this merge request is to take the second- and third-order perturbation solution systems in Julliard and Kamenik (2004), unfold these into an equivalent matrix representation using permutation matrices as in Levintal (2017). Then extending Ratto and Iskrev (2012) one takes the derivative with respect to each parameter separately and gets a computational problem that is linear, albeit large, as it involves either solving generalized Sylvester equations or taking inverses of highly sparse matrices. I will now briefly summarize the perturbation solution system at third order and the system that results when taking the derivative with respect to parameters.
## Perturbation Solution
The following systems arise at first, second, and third order:
$(ghx): f_{x} z_{x} = f_{y_{-}^*} + f_{y_0} g_{x} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{x}= A g_{x} + f_{y_{-}^*}=0$
$(ghu): f_{z} z_{u} = f_{y_0} g_{u} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{u} + f_{u}= A g_u + f_u = 0$
$(ghxx) : A g_{xx} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{x}\right) + f_{zz} \left( z_{x} \otimes z_{x} \right) = 0$
$(ghxu) : A g_{xu} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{x} \otimes z_{u} \right) = 0$
$(ghuu) : A g_{uu} + B g_{xx} \left(g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{u} \otimes z_{u} \right) = 0$
$(ghs2) : (A+B) g_{\sigma\sigma} + \left( f_{y^{**}_{+}y^{**}_{+}} \left(g^{**}_{u} \otimes g^{**}_{u}\right) + f_{y^{**}_{+}} g^{**}_{uu}\right)vec(\Sigma) = 0$
$(ghxxx) : A g_{xxx} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{x}\right) + f_{y_{+}}g^{**}_{xx} \left(g^{*}_x \otimes g^{*}_{xx}\right)P_{x\_xx} + f_{zz} \left( z_{x} \otimes z_{xx} \right)P_{x\_xx} + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{x} \right) = 0$
$(ghxxu) : A g_{xxu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{u} \right) + f_{zz} \left( \left( z_{x} \otimes z_{xu} \right)P_{x\_xu} + \left(z_{xx} \otimes z_{u}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{x} \otimes g^{*}_{xu}\right)P_{x\_xu} + \left(g^{*}_{xx} \otimes g^{*}_{u}\right) \right) = 0$
$(ghxuu) : A g_{xuu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( \left( z_{xu} \otimes z_{u} \right)P_{xu\_u} + \left(z_{x} \otimes z_{uu}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{xu} \otimes g^{*}_{u}\right)P_{xu\_u} + \left(g^{*}_{x} \otimes g^{*}_{uu}\right) \right) = 0$
$(ghuuu) : A g_{uuu} + B g_{xxx} \left(g^{*}_{u} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{u} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( z_{u} \otimes z_{uu} \right)P_{u\_uu} + f_{y_{+}}g^{**}_{xx} \left(g^{*}_{u} \otimes g^{*}_{uu}\right)P_{u\_uu} = 0$
$(ghx\sigma\sigma) : A g_{x\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_x + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{x} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{x} \otimes z_{\sigma\sigma}\right) + F_{xu_{+}u_{+}}\left(I_{n_x} \otimes vec(\Sigma)\right) = 0$
$F_{xu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_x^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{xu_{+}} \otimes z_{u_{+}} \right)P_{xu\_u} + \left(z_{x} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{x} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
$(ghu\sigma\sigma) : A g_{u\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_{u} + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{u} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{u} \otimes z_{\sigma\sigma}\right) + F_{uu_{+}u_{+}}\left(I_{n_u} \otimes vec(\Sigma_u)\right) = 0$
$F_{uu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_u^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{uu_{+}} \otimes z_{u_{+}} \right)P_{uu\_u} + \left(z_{u} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{u} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
A and B are the common perturbation matrices:
$A = f_{y_0} + \begin{pmatrix} \underbrace{0}_{n\times n_{static}} &\vdots& \underbrace{f_{y^{**}_{+}} \cdot g^{**}_{x}}_{n \times n_{spred}} &\vdots& \underbrace{0}_{n\times n_{frwd}} \end{pmatrix}$and $B = \begin{pmatrix} \underbrace{0}_{n \times n_{static}}&\vdots & \underbrace{0}_{n \times n_{pred}} & \vdots & \underbrace{f_{y^{**}_{+}}}_{n \times n_{sfwrd}} \end{pmatrix}$
and $z=(y_{-}^{\ast}; y; y_{+}^{\ast\ast}; u)$ denotes the dynamic model variables as in `M_.lead_lag_incidence`, $y^\ast$ denote state variables, $y^{\ast\ast}$ denote forward looking variables, $y_+$ denote the variables with a lead, $y_{-}$ denote variables with a lag, $y_0$ denote variables at period t, $f$ the model equations, and $f_z$ the first-order dynamic model derivatives, $f_{zz}$ the second-order dynamic derivatives, and $f_{zzz}$ the third-order dynamic model derivatives. Then:
$z_{x} = \begin{pmatrix}I\\g_{x}\\g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{u} =\begin{pmatrix}0\\g_{u}\\g^{**}_{x} \cdot g^{*}_{u}\\I\end{pmatrix}$, $z_{u_{+}} =\begin{pmatrix}0\\0\\g^{**}_{u}\\0\end{pmatrix}$
$z_{xx} = \begin{pmatrix} 0\\g_{xx}\\g^{**}_{x} \left( g^{*}_x \otimes g^{*}_{x} \right) + g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{xu} =\begin{pmatrix}0\\g_{xu}\\g^{**}_{xx} \left( g^{*}_x \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{xu}\\0\end{pmatrix}$, $z_{uu} =\begin{pmatrix}0\\g_{uu}\\g^{**}_{xx} \left( g^{*}_u \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{uu}\\0\end{pmatrix}$,
$z_{xu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_x \otimes I \right)\\0\end{pmatrix}$, $z_{uu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_{u} \otimes I \right)\\0\end{pmatrix}$, $z_{u_{+}u_{+}} =\begin{pmatrix}0\\0\\g^{\ast\ast}_{uu}\\0\end{pmatrix}$, $z_{\sigma\sigma} = \begin{pmatrix}0\\ g_{\sigma\sigma}\\ g^{\ast\ast}_{x}g^{\ast}_{\sigma\sigma} + g^{\ast\ast}_{\sigma\sigma}\\0 \end{pmatrix}$
$P$ are permutation matrices that can be computed using Matlab's `ipermute` function.
## Parameter derivatives of perturbation solutions
First, we need the parameter derivatives of first, second, third, and fourth derivatives of the dynamic model (i.e. g1,g2,g3,g4 in dynamic files), I make use of the implicit function theorem: Let $f_{z^k}$ denote the kth derivative (wrt all dynamic variables) of the dynamic model, then let $df_{z^k}$ denote the first-derivative (wrt all model parameters) of $f_{z^k}$ evaluated at the steady state. Note that $f_{z^k}$ is a function of both the model parameters $\theta$ and of the steady state of all dynamic variables $\bar{z}$, which also depend on the parameters. Hence, implicitly $f_{z^k}=f_{z^k}(\theta,\bar{z}(\theta))$ and $df_{z^k}$ consists of two parts:
1. direct derivative wrt to all model parameters given by the preprocessor in the `_params_derivs.m` files
2. contribution of derivative of steady state of dynamic variables (wrt all model parameters): $f_{z^{k+1}} \cdot d\bar{z}$
Note that we already have functionality to compute $d\bar{z}$ analytically.
Having this, the above perturbation systems are basically equations of the following types
$AX +BXC = RHS$ or $AX = RHS$
Now when taking the derivative (wrt to one single parameter $\theta_j$), we get
$A\mathrm{d}\{X\} + B\mathrm{d}\{X\}C = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X - \mathrm{d}\{B\}XC - BX\mathrm{d}\{C\}$
or
$A\mathrm{d}\{X\} = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X$
The first one is a Sylvester type equation, the second one can be solved by taking the inverse of $A$. The only diffculty and tedious work arrises in computing (the highly sparse) derivatives of $RHS$.
***
# New functions: `
## get_perturbation_params_derivs.m`and `get_perturbation_params_derivs_numerical_objective.m`
* The parameter derivatives up to third order are computed in the new function`get_perturbation_params_derivs.m` both analytically and numerically. For numerical derivatives `get_perturbation_params_derivs_numerical_objective.m` is the objective for `fjaco.m` or `hessian_sparse.m` or `hessian.m`.
* `get_perturbation_params_derivs.m` is basically an extended version of the previous `get_first_order_solution_params_derivs.m` function.
* * `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It is used for numerical derivatives, whenever `analytic_derivation_mode=-1|-2`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
* Detailed changes:
* Most important: notation of this function is now in accordance to the k_order_solver, i.e. we do not compute derivatives of Kalman transition matrices A and B, but rather the solution matrices ghx,ghu,ghxx,ghxu,ghuu,ghs2,ghxxx,ghxxu,ghxuu,ghuuu,ghxss,ghuss in accordance with notation used in `oo_.dr`. As a byproduct at first-order, focusing on ghx and ghu instead of Kalman transition matrices A and B makes the computations slightly faster for large models (e.g. for Quest the computations were faster by a couple of seconds, not much, but okay).
* Removed use of `kstate`, see also Dynare/dynare#1653 and Dynare/dynare!1656
* Output arguments are stored in a structure `DERIVS`, there is also a flag `d2flag` that computes parameter hessians needed only in `dsge_likelihood.m`.
* Removed `kronflag` as input. `options_.analytic_derivation_mode` is now used instead of `kronflag`.
* Removed `indvar`, an index that was used to selected specific variables in the derivatives. This is not needed, as we always compute the parameter derivatives for all variables first and then select a subset of variables. The selection now takes place in other functions, like `dsge_likelihood.m`.
* Introduced some checks: (i) deterministic exogenous variables are not supported, (ii) Kronecker method only compatible with first-order approximation so reset to sylvester method, (iii) for purely backward or forward models we need to be careful with the rows in `M_.lead_la g_incidence`, (iv) if `_params_derivs.m` files are missing an error is thrown.
* For numerical derivatives, if mod file does not contain an `estimated_params_block`, a temporary one with the most important parameter information is created.
## `unfold_g4.m`
* When evaluating g3 and g4 one needs to take into account that these do not contain symmetric elements, so one needs to use `unfold_g3.m` and the new function `unfold_g4.m`. This returns an unfolded version of the same matrix (i.e. with symmetric elements).
***
# New test models
`.gitignore` and `Makefile.am` are changed accordingly. Also now it is possible to run test suite on analytic_derivatives, i.e. run `make check m/analytic_derivatives`
## `analytic_derivatives/BrockMirman_PertParamsDerivs.mod`
* This is the Brock Mirman model, where we know the exact policy function $g$ for capital and consumption. As this does not imply a nonzero $g_{\sigma\sigma}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ I added some artificial equations to get nonzero solution matrices with respect to $\sigma$. The true perturbation solution matrices $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ are then computed analytically with Matlab's symbolic toolbox and saved in `nBrockMirmanSYM.mat`. There is a preprocessor flag that recreates these analytical computations if changes are needed (and to check whether I made some errors here ;-) )
* Then solution matrices up to third order and their parameter Jacobians are then compared to the ones computed by Dynare's `k_order_solver` and by `get_perturbation_params_derivs` for all `analytic_derivation_mode`'s. There will be an error if the maximum absolute deviation is too large, i.e. for numerical derivatives (`analytic_derivation_mode=-1|-2`) the tolerance is choosen lower (around 1e-5); for analytical methods we are stricter: around 1e-13 for first-order, 1e-12 for second order, and 1e-11 for third-order.
* As a side note, this mod file also checks Dynare's `k_order_solver` algorithm and throws an error if something is wrong.
* This test model shows that the new functionality works well. And analytical derivatives perform way better and accurate than numerical ones, even for this small model.
## `analytic_derivatives/burnside_3_order_PertParamsDerivs.mod`
* This builds upon `tests/k_order_perturbation/burnside_k_order.mod` and computes the true parameter derivatives analytically by hand.
* This test model also shows that the new functionality works well.
## `analytic_derivatives/LindeTrabandt2019.mod`
* Shows that the new functionality also works for medium-sized models, i.e. a SW type model solved at third order with 35 variables (11 states). 2 shocks and 20 parameters.
* This mod file can be used to tweak the speed of the computations in the future.
* Compares numerical versus analytical parameter derivatives (for first, second and third order). Note that this model clearly shows that numerical ones are quite different than analytical ones even at first order!
## `identification/LindeTrabandt2019_xfail.mod`
* This model is a check for issue Dynare/dynare#1595, see fjaco.m below, and will fail.
* Removed `analytic_derivatives/ls2003.mod` as this mod file is neither in the testsuite nor does it work.
***
# Detailed changes in other functions
## `get_first_order_solution_params_derivs.m`
* Deleted, or actually, renamed to `get_perturbation_params_derivs.m`, as this function now is able to compute the derivatives up to third order
## `identification_numerical_objective.m`
* `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
## `dsge_likelihood.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `T`matrix anymore, but instead of the dynare solution matrix `ghx`. So we recreate `T` here as this amounts to adding some zeros and focusing on selected variables only.
* Added some checks to make sure the first-order approximation is selected.
* Removed `kron_flag` as input, as `get_perturbation_params_derivs` looks into `options_.analytic_derivation_mode` for `kron_flag`.
## `dynare_identification.m`
* make sure that setting `analytic_derivation_mode` is set both in `options_ident` and `options_`. Note that at the end of the function we restore the `options_` structure, so all changes are local. In a next merge request, I will remove the global variables to make all variables local.
## `get_identification_jacobians.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `A` and `B` matrix anymore, but instead of the dynare solution matrix `ghx` and `ghu`. So we recreate these matrices here instead of in `get_perturbation_params_derivs.m`.
* Added `str2func` for better function handles in `fjaco.m`.
## `fjaco.m`
* make `tol`an option, which can be adjusted by changing `options_.dynatol.x`for identification and parameter derivatives purposes.
* include a check and an informative error message, if numerical derivatives (two-sided finite difference method) yield errors in `resol.m` for identification and parameter derivatives purposes. This closes issue Dynare/dynare#1595.
* Changed year of copyright to 2010-2017,2019
***
# Further suggestions and questions
* Ones this is merged, I will merge request an improvement of the identification toolbox, which will work up to third order using the pruned state space. This will also remove some issues and bugs, and also I will remove global variables in this request.
* The third-order derivatives can be further improved by taking sparsity into account and use mex versions for kronecker products etc. I leave this for further testing (and if anybody actually uses this ;-) )
2019-12-17 19:17:09 +01:00
dghuu ( : , : , jp ) = - invA * ( dA ( : , : , jp ) * oo . dr . ghuu + ( dRHS_1 + dRHS_2 + dRHS_3 ) + dB ( : , : , jp ) * abcOutuu + B * ( dabcOut_1 + dabcOut_2 + dabcOut_3 ) ) ;
%compute dghs2
2020-01-10 17:55:57 +01:00
dRHS_1 = sparse_hessian_times_B_kronecker_C ( dg2 ( : , k2yy0ex0 ( nonzeros ( klead ) , nonzeros ( klead ) ) + ( jp - 1 ) * yy0ex0_nbr ^2 ) , oo . dr . ghu ( klead ~= 0 , : ) , threads_BC ) ;
dRHS_2 = sparse_hessian_times_B_kronecker_C ( g2 ( : , k2yy0ex0 ( nonzeros ( klead ) , nonzeros ( klead ) ) ) , dghu ( klead ~= 0 , : , jp ) , oo . dr . ghu ( klead ~= 0 , : ) , threads_BC ) ;
dRHS_3 = sparse_hessian_times_B_kronecker_C ( g2 ( : , k2yy0ex0 ( nonzeros ( klead ) , nonzeros ( klead ) ) ) , oo . dr . ghu ( klead ~= 0 , : ) , dghu ( klead ~= 0 , : , jp ) , threads_BC ) ;
Added parameter derivatives of perturbation solution up to 3 order
# Preliminary comments
I finished the identification toolbox at orders two and three using the pruned state space system, but before I merge request this, I decided to first merge the new functionality to compute parameter derivatives of perturbation solution matrices at higher orders. So after this is approved, I merge the identification toolbox.
I guess @rattoma, @sebastien, and @michel are best choices to review this.
I outline the main idea first and then provide some more detailed changes I made to the functions.
***
# Main idea
This merge request is concerned with the *analytical*computation of the parameter derivatives of first, second and third order perturbation solution matrices, i.e. using _closed-form_ expressions to efficiently compute the derivative of $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ *with respect to model parameters* $\theta$. Note that $\theta$ contains model parameters, stderr and corr parameters of shocks. stderr and corr parameters of measurement errors are not yet supported, (they can easily be included as exogenous shocks). The availability of such derivatives is beneficial in terms of more reliable analysis of model sensitivity and parameter identifiability as well as more efficient estimation methods, in particular for models solved up to third order, as it is well-known that numerical derivatives are a tricky business, especially for large models.
References for my approach are:
* Iskrev (2008, 2010) and Schmitt-Grohé and Uribe (2012, Appendix) who were the first to compute the parameter derivatives analytically at first order, however, using inefficient (sparse) Kronecker products.
* Mutschler (2015) who provides the expressions for a second-order, but again using inefficient (sparse) Kronecker products.
* Ratto and Iskrev (2012) who show how the first-order system can be solved accurately, fast and efficiently using existing numerical algorithms for generalized Sylvester equations by taking the parameter derivative with respect to each parameter separately.
* Julliard and Kamenik (2004) who provide the perturbation solution equation system in tensor notation at any order k.
* Levintal (2017) who introduces permutation matrices to express the perturbation solution equation system in matrix notation up to fifth order.
Note that @rattoma already implemented the parameter derivatives of $g_x$ and $g_u$ analytically (and numerically), and I rely heavily on his work in `get_first_order_solution_params_derivs.m` (previously `getH.m`). My additions are mainly to this function and thus it is renamed to `get_perturbation_params_derivs.m`.
The basic idea of this merge request is to take the second- and third-order perturbation solution systems in Julliard and Kamenik (2004), unfold these into an equivalent matrix representation using permutation matrices as in Levintal (2017). Then extending Ratto and Iskrev (2012) one takes the derivative with respect to each parameter separately and gets a computational problem that is linear, albeit large, as it involves either solving generalized Sylvester equations or taking inverses of highly sparse matrices. I will now briefly summarize the perturbation solution system at third order and the system that results when taking the derivative with respect to parameters.
## Perturbation Solution
The following systems arise at first, second, and third order:
$(ghx): f_{x} z_{x} = f_{y_{-}^*} + f_{y_0} g_{x} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{x}= A g_{x} + f_{y_{-}^*}=0$
$(ghu): f_{z} z_{u} = f_{y_0} g_{u} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{u} + f_{u}= A g_u + f_u = 0$
$(ghxx) : A g_{xx} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{x}\right) + f_{zz} \left( z_{x} \otimes z_{x} \right) = 0$
$(ghxu) : A g_{xu} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{x} \otimes z_{u} \right) = 0$
$(ghuu) : A g_{uu} + B g_{xx} \left(g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{u} \otimes z_{u} \right) = 0$
$(ghs2) : (A+B) g_{\sigma\sigma} + \left( f_{y^{**}_{+}y^{**}_{+}} \left(g^{**}_{u} \otimes g^{**}_{u}\right) + f_{y^{**}_{+}} g^{**}_{uu}\right)vec(\Sigma) = 0$
$(ghxxx) : A g_{xxx} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{x}\right) + f_{y_{+}}g^{**}_{xx} \left(g^{*}_x \otimes g^{*}_{xx}\right)P_{x\_xx} + f_{zz} \left( z_{x} \otimes z_{xx} \right)P_{x\_xx} + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{x} \right) = 0$
$(ghxxu) : A g_{xxu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{u} \right) + f_{zz} \left( \left( z_{x} \otimes z_{xu} \right)P_{x\_xu} + \left(z_{xx} \otimes z_{u}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{x} \otimes g^{*}_{xu}\right)P_{x\_xu} + \left(g^{*}_{xx} \otimes g^{*}_{u}\right) \right) = 0$
$(ghxuu) : A g_{xuu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( \left( z_{xu} \otimes z_{u} \right)P_{xu\_u} + \left(z_{x} \otimes z_{uu}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{xu} \otimes g^{*}_{u}\right)P_{xu\_u} + \left(g^{*}_{x} \otimes g^{*}_{uu}\right) \right) = 0$
$(ghuuu) : A g_{uuu} + B g_{xxx} \left(g^{*}_{u} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{u} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( z_{u} \otimes z_{uu} \right)P_{u\_uu} + f_{y_{+}}g^{**}_{xx} \left(g^{*}_{u} \otimes g^{*}_{uu}\right)P_{u\_uu} = 0$
$(ghx\sigma\sigma) : A g_{x\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_x + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{x} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{x} \otimes z_{\sigma\sigma}\right) + F_{xu_{+}u_{+}}\left(I_{n_x} \otimes vec(\Sigma)\right) = 0$
$F_{xu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_x^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{xu_{+}} \otimes z_{u_{+}} \right)P_{xu\_u} + \left(z_{x} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{x} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
$(ghu\sigma\sigma) : A g_{u\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_{u} + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{u} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{u} \otimes z_{\sigma\sigma}\right) + F_{uu_{+}u_{+}}\left(I_{n_u} \otimes vec(\Sigma_u)\right) = 0$
$F_{uu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_u^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{uu_{+}} \otimes z_{u_{+}} \right)P_{uu\_u} + \left(z_{u} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{u} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
A and B are the common perturbation matrices:
$A = f_{y_0} + \begin{pmatrix} \underbrace{0}_{n\times n_{static}} &\vdots& \underbrace{f_{y^{**}_{+}} \cdot g^{**}_{x}}_{n \times n_{spred}} &\vdots& \underbrace{0}_{n\times n_{frwd}} \end{pmatrix}$and $B = \begin{pmatrix} \underbrace{0}_{n \times n_{static}}&\vdots & \underbrace{0}_{n \times n_{pred}} & \vdots & \underbrace{f_{y^{**}_{+}}}_{n \times n_{sfwrd}} \end{pmatrix}$
and $z=(y_{-}^{\ast}; y; y_{+}^{\ast\ast}; u)$ denotes the dynamic model variables as in `M_.lead_lag_incidence`, $y^\ast$ denote state variables, $y^{\ast\ast}$ denote forward looking variables, $y_+$ denote the variables with a lead, $y_{-}$ denote variables with a lag, $y_0$ denote variables at period t, $f$ the model equations, and $f_z$ the first-order dynamic model derivatives, $f_{zz}$ the second-order dynamic derivatives, and $f_{zzz}$ the third-order dynamic model derivatives. Then:
$z_{x} = \begin{pmatrix}I\\g_{x}\\g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{u} =\begin{pmatrix}0\\g_{u}\\g^{**}_{x} \cdot g^{*}_{u}\\I\end{pmatrix}$, $z_{u_{+}} =\begin{pmatrix}0\\0\\g^{**}_{u}\\0\end{pmatrix}$
$z_{xx} = \begin{pmatrix} 0\\g_{xx}\\g^{**}_{x} \left( g^{*}_x \otimes g^{*}_{x} \right) + g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{xu} =\begin{pmatrix}0\\g_{xu}\\g^{**}_{xx} \left( g^{*}_x \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{xu}\\0\end{pmatrix}$, $z_{uu} =\begin{pmatrix}0\\g_{uu}\\g^{**}_{xx} \left( g^{*}_u \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{uu}\\0\end{pmatrix}$,
$z_{xu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_x \otimes I \right)\\0\end{pmatrix}$, $z_{uu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_{u} \otimes I \right)\\0\end{pmatrix}$, $z_{u_{+}u_{+}} =\begin{pmatrix}0\\0\\g^{\ast\ast}_{uu}\\0\end{pmatrix}$, $z_{\sigma\sigma} = \begin{pmatrix}0\\ g_{\sigma\sigma}\\ g^{\ast\ast}_{x}g^{\ast}_{\sigma\sigma} + g^{\ast\ast}_{\sigma\sigma}\\0 \end{pmatrix}$
$P$ are permutation matrices that can be computed using Matlab's `ipermute` function.
## Parameter derivatives of perturbation solutions
First, we need the parameter derivatives of first, second, third, and fourth derivatives of the dynamic model (i.e. g1,g2,g3,g4 in dynamic files), I make use of the implicit function theorem: Let $f_{z^k}$ denote the kth derivative (wrt all dynamic variables) of the dynamic model, then let $df_{z^k}$ denote the first-derivative (wrt all model parameters) of $f_{z^k}$ evaluated at the steady state. Note that $f_{z^k}$ is a function of both the model parameters $\theta$ and of the steady state of all dynamic variables $\bar{z}$, which also depend on the parameters. Hence, implicitly $f_{z^k}=f_{z^k}(\theta,\bar{z}(\theta))$ and $df_{z^k}$ consists of two parts:
1. direct derivative wrt to all model parameters given by the preprocessor in the `_params_derivs.m` files
2. contribution of derivative of steady state of dynamic variables (wrt all model parameters): $f_{z^{k+1}} \cdot d\bar{z}$
Note that we already have functionality to compute $d\bar{z}$ analytically.
Having this, the above perturbation systems are basically equations of the following types
$AX +BXC = RHS$ or $AX = RHS$
Now when taking the derivative (wrt to one single parameter $\theta_j$), we get
$A\mathrm{d}\{X\} + B\mathrm{d}\{X\}C = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X - \mathrm{d}\{B\}XC - BX\mathrm{d}\{C\}$
or
$A\mathrm{d}\{X\} = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X$
The first one is a Sylvester type equation, the second one can be solved by taking the inverse of $A$. The only diffculty and tedious work arrises in computing (the highly sparse) derivatives of $RHS$.
***
# New functions: `
## get_perturbation_params_derivs.m`and `get_perturbation_params_derivs_numerical_objective.m`
* The parameter derivatives up to third order are computed in the new function`get_perturbation_params_derivs.m` both analytically and numerically. For numerical derivatives `get_perturbation_params_derivs_numerical_objective.m` is the objective for `fjaco.m` or `hessian_sparse.m` or `hessian.m`.
* `get_perturbation_params_derivs.m` is basically an extended version of the previous `get_first_order_solution_params_derivs.m` function.
* * `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It is used for numerical derivatives, whenever `analytic_derivation_mode=-1|-2`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
* Detailed changes:
* Most important: notation of this function is now in accordance to the k_order_solver, i.e. we do not compute derivatives of Kalman transition matrices A and B, but rather the solution matrices ghx,ghu,ghxx,ghxu,ghuu,ghs2,ghxxx,ghxxu,ghxuu,ghuuu,ghxss,ghuss in accordance with notation used in `oo_.dr`. As a byproduct at first-order, focusing on ghx and ghu instead of Kalman transition matrices A and B makes the computations slightly faster for large models (e.g. for Quest the computations were faster by a couple of seconds, not much, but okay).
* Removed use of `kstate`, see also Dynare/dynare#1653 and Dynare/dynare!1656
* Output arguments are stored in a structure `DERIVS`, there is also a flag `d2flag` that computes parameter hessians needed only in `dsge_likelihood.m`.
* Removed `kronflag` as input. `options_.analytic_derivation_mode` is now used instead of `kronflag`.
* Removed `indvar`, an index that was used to selected specific variables in the derivatives. This is not needed, as we always compute the parameter derivatives for all variables first and then select a subset of variables. The selection now takes place in other functions, like `dsge_likelihood.m`.
* Introduced some checks: (i) deterministic exogenous variables are not supported, (ii) Kronecker method only compatible with first-order approximation so reset to sylvester method, (iii) for purely backward or forward models we need to be careful with the rows in `M_.lead_la g_incidence`, (iv) if `_params_derivs.m` files are missing an error is thrown.
* For numerical derivatives, if mod file does not contain an `estimated_params_block`, a temporary one with the most important parameter information is created.
## `unfold_g4.m`
* When evaluating g3 and g4 one needs to take into account that these do not contain symmetric elements, so one needs to use `unfold_g3.m` and the new function `unfold_g4.m`. This returns an unfolded version of the same matrix (i.e. with symmetric elements).
***
# New test models
`.gitignore` and `Makefile.am` are changed accordingly. Also now it is possible to run test suite on analytic_derivatives, i.e. run `make check m/analytic_derivatives`
## `analytic_derivatives/BrockMirman_PertParamsDerivs.mod`
* This is the Brock Mirman model, where we know the exact policy function $g$ for capital and consumption. As this does not imply a nonzero $g_{\sigma\sigma}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ I added some artificial equations to get nonzero solution matrices with respect to $\sigma$. The true perturbation solution matrices $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ are then computed analytically with Matlab's symbolic toolbox and saved in `nBrockMirmanSYM.mat`. There is a preprocessor flag that recreates these analytical computations if changes are needed (and to check whether I made some errors here ;-) )
* Then solution matrices up to third order and their parameter Jacobians are then compared to the ones computed by Dynare's `k_order_solver` and by `get_perturbation_params_derivs` for all `analytic_derivation_mode`'s. There will be an error if the maximum absolute deviation is too large, i.e. for numerical derivatives (`analytic_derivation_mode=-1|-2`) the tolerance is choosen lower (around 1e-5); for analytical methods we are stricter: around 1e-13 for first-order, 1e-12 for second order, and 1e-11 for third-order.
* As a side note, this mod file also checks Dynare's `k_order_solver` algorithm and throws an error if something is wrong.
* This test model shows that the new functionality works well. And analytical derivatives perform way better and accurate than numerical ones, even for this small model.
## `analytic_derivatives/burnside_3_order_PertParamsDerivs.mod`
* This builds upon `tests/k_order_perturbation/burnside_k_order.mod` and computes the true parameter derivatives analytically by hand.
* This test model also shows that the new functionality works well.
## `analytic_derivatives/LindeTrabandt2019.mod`
* Shows that the new functionality also works for medium-sized models, i.e. a SW type model solved at third order with 35 variables (11 states). 2 shocks and 20 parameters.
* This mod file can be used to tweak the speed of the computations in the future.
* Compares numerical versus analytical parameter derivatives (for first, second and third order). Note that this model clearly shows that numerical ones are quite different than analytical ones even at first order!
## `identification/LindeTrabandt2019_xfail.mod`
* This model is a check for issue Dynare/dynare#1595, see fjaco.m below, and will fail.
* Removed `analytic_derivatives/ls2003.mod` as this mod file is neither in the testsuite nor does it work.
***
# Detailed changes in other functions
## `get_first_order_solution_params_derivs.m`
* Deleted, or actually, renamed to `get_perturbation_params_derivs.m`, as this function now is able to compute the derivatives up to third order
## `identification_numerical_objective.m`
* `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
## `dsge_likelihood.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `T`matrix anymore, but instead of the dynare solution matrix `ghx`. So we recreate `T` here as this amounts to adding some zeros and focusing on selected variables only.
* Added some checks to make sure the first-order approximation is selected.
* Removed `kron_flag` as input, as `get_perturbation_params_derivs` looks into `options_.analytic_derivation_mode` for `kron_flag`.
## `dynare_identification.m`
* make sure that setting `analytic_derivation_mode` is set both in `options_ident` and `options_`. Note that at the end of the function we restore the `options_` structure, so all changes are local. In a next merge request, I will remove the global variables to make all variables local.
## `get_identification_jacobians.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `A` and `B` matrix anymore, but instead of the dynare solution matrix `ghx` and `ghu`. So we recreate these matrices here instead of in `get_perturbation_params_derivs.m`.
* Added `str2func` for better function handles in `fjaco.m`.
## `fjaco.m`
* make `tol`an option, which can be adjusted by changing `options_.dynatol.x`for identification and parameter derivatives purposes.
* include a check and an informative error message, if numerical derivatives (two-sided finite difference method) yield errors in `resol.m` for identification and parameter derivatives purposes. This closes issue Dynare/dynare#1595.
* Changed year of copyright to 2010-2017,2019
***
# Further suggestions and questions
* Ones this is merged, I will merge request an improvement of the identification toolbox, which will work up to third order using the pruned state space. This will also remove some issues and bugs, and also I will remove global variables in this request.
* The third-order derivatives can be further improved by taking sparsity into account and use mex versions for kronecker products etc. I leave this for further testing (and if anybody actually uses this ;-) )
2019-12-17 19:17:09 +01:00
dghs2 ( : , stderrparam_nbr + corrparam_nbr + jp ) = - invS * ( dS ( : , : , jp ) * oo . dr . ghs2 + ( ( dRHS_1 + dRHS_2 + dRHS_3 ) + dg1 ( : , nonzeros ( klead ) , jp ) * oo . dr . ghuu ( klead ~= 0 , : ) + g1 ( : , nonzeros ( klead ) ) * dghuu ( klead ~= 0 , : , jp ) ) * M . Sigma_e ( : ) ) ;
end
%add contributions by dSigma_e to dghs2
if ~ isempty ( indpstderr )
for jp = 1 : stderrparam_nbr
dghs2 ( : , jp ) = - invS * RHSs2 * vec ( dSigma_e ( : , : , jp ) ) ;
end
end
if ~ isempty ( indpcorr )
for jp = 1 : corrparam_nbr
dghs2 ( : , stderrparam_nbr + jp ) = - invS * RHSs2 * vec ( dSigma_e ( : , : , stderrparam_nbr + jp ) ) ;
end
end
end
if order > 2
% Note that at third-order we have the following expressions, which are (numerically) zero, given suitable tensor-unfolding permuation matrices P:
% * for ghxxx: A*ghxxx + B*ghxxx*kron(ghx_,kron(ghx_,ghx_)) + g3*kron(kron(zx,zx),zx) + g2*kron(zx,zxx)*P_x_xx + fyp*ghxxp*kron(ghx_,ghxx_)*P_x_xx = 0
% Taking the differential yields a generalized sylvester equation to get dghxxx: A*dghxxx + B*dghxxx*kron(ghx_,kron(ghx_,ghx_)) = RHSxxx
% * for ghxxu: A*ghxxu + B*ghxxx*kron(ghx_,kron(ghx_,ghu_)) + gg3*kron(zx,kron(zx,zu)) + gg2*(kron(zx,zxu)*P_x_xu + kron(zxx,zu)) + fyp*ghxxp*(kron(ghx_,ghxu_)*P_x_xu + kron(ghxx_,ghu_)) = 0
% Taking the differential yields an invertible equation to get dghxxu: A*dghxxu = RHSxxu
% * for ghxuu: A*ghxuu + B*ghxxx*kron(ghx_,kron(ghu_,ghu_)) + gg3*kron(zx,kron(zu,zu)) + gg2*(kron(zxu,zu)*P_xu_u + kron(zx,zuu)) + fyp*ghxxp*(kron(ghxu_,ghu_)*Pxu_u + kron(ghx_,ghuu_)) = 0
% Taking the differential yields an invertible equation to get dghxuu: A*dghxuu = RHSxuu
% * for ghuuu: A*ghuuu + B*ghxxx*kron(ghu_,kron(ghu_,ghu_)) + gg3*kron(kron(zu,zu),zu) + gg2*kron(zu,zuu)*P_u_uu + fyp*ghxxp*kron(ghu_,ghuu_)*P_u_uu = 0
% Taking the differential yields an invertible equation to get dghuuu: A*dghuuu = RHSuuu
% * for ghxss: A*ghxss + B*ghxss*ghx_ + fyp*ghxxp*kron(ghx_,ghss_) + gg2*kron(zx,zss) + Fxupup*kron(Ix,Sigma_e(:)) = 0
% Taking the differential yields a generalized sylvester equation to get dghxss: A*dghxss + B*dghxss*ghx_ = RHSxss
% * for ghuss: A*ghuss + B*ghxss*ghu_ + gg2*kron(zu,zss) + fyp*ghxxp*kron(ghu_,ghss_) + Fuupup*kron(Iu,Sigma_e(:)) = 0
% Taking the differential yields an invertible equation to get dghuss: A*dghuss = RHSuss
% * due to certainty equivalence and Gaussian shocks, we note that ghxxs, ghxus, ghuus, and ghsss are zero and thus not computed
2019-12-20 16:28:06 +01:00
Added parameter derivatives of perturbation solution up to 3 order
# Preliminary comments
I finished the identification toolbox at orders two and three using the pruned state space system, but before I merge request this, I decided to first merge the new functionality to compute parameter derivatives of perturbation solution matrices at higher orders. So after this is approved, I merge the identification toolbox.
I guess @rattoma, @sebastien, and @michel are best choices to review this.
I outline the main idea first and then provide some more detailed changes I made to the functions.
***
# Main idea
This merge request is concerned with the *analytical*computation of the parameter derivatives of first, second and third order perturbation solution matrices, i.e. using _closed-form_ expressions to efficiently compute the derivative of $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ *with respect to model parameters* $\theta$. Note that $\theta$ contains model parameters, stderr and corr parameters of shocks. stderr and corr parameters of measurement errors are not yet supported, (they can easily be included as exogenous shocks). The availability of such derivatives is beneficial in terms of more reliable analysis of model sensitivity and parameter identifiability as well as more efficient estimation methods, in particular for models solved up to third order, as it is well-known that numerical derivatives are a tricky business, especially for large models.
References for my approach are:
* Iskrev (2008, 2010) and Schmitt-Grohé and Uribe (2012, Appendix) who were the first to compute the parameter derivatives analytically at first order, however, using inefficient (sparse) Kronecker products.
* Mutschler (2015) who provides the expressions for a second-order, but again using inefficient (sparse) Kronecker products.
* Ratto and Iskrev (2012) who show how the first-order system can be solved accurately, fast and efficiently using existing numerical algorithms for generalized Sylvester equations by taking the parameter derivative with respect to each parameter separately.
* Julliard and Kamenik (2004) who provide the perturbation solution equation system in tensor notation at any order k.
* Levintal (2017) who introduces permutation matrices to express the perturbation solution equation system in matrix notation up to fifth order.
Note that @rattoma already implemented the parameter derivatives of $g_x$ and $g_u$ analytically (and numerically), and I rely heavily on his work in `get_first_order_solution_params_derivs.m` (previously `getH.m`). My additions are mainly to this function and thus it is renamed to `get_perturbation_params_derivs.m`.
The basic idea of this merge request is to take the second- and third-order perturbation solution systems in Julliard and Kamenik (2004), unfold these into an equivalent matrix representation using permutation matrices as in Levintal (2017). Then extending Ratto and Iskrev (2012) one takes the derivative with respect to each parameter separately and gets a computational problem that is linear, albeit large, as it involves either solving generalized Sylvester equations or taking inverses of highly sparse matrices. I will now briefly summarize the perturbation solution system at third order and the system that results when taking the derivative with respect to parameters.
## Perturbation Solution
The following systems arise at first, second, and third order:
$(ghx): f_{x} z_{x} = f_{y_{-}^*} + f_{y_0} g_{x} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{x}= A g_{x} + f_{y_{-}^*}=0$
$(ghu): f_{z} z_{u} = f_{y_0} g_{u} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{u} + f_{u}= A g_u + f_u = 0$
$(ghxx) : A g_{xx} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{x}\right) + f_{zz} \left( z_{x} \otimes z_{x} \right) = 0$
$(ghxu) : A g_{xu} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{x} \otimes z_{u} \right) = 0$
$(ghuu) : A g_{uu} + B g_{xx} \left(g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{u} \otimes z_{u} \right) = 0$
$(ghs2) : (A+B) g_{\sigma\sigma} + \left( f_{y^{**}_{+}y^{**}_{+}} \left(g^{**}_{u} \otimes g^{**}_{u}\right) + f_{y^{**}_{+}} g^{**}_{uu}\right)vec(\Sigma) = 0$
$(ghxxx) : A g_{xxx} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{x}\right) + f_{y_{+}}g^{**}_{xx} \left(g^{*}_x \otimes g^{*}_{xx}\right)P_{x\_xx} + f_{zz} \left( z_{x} \otimes z_{xx} \right)P_{x\_xx} + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{x} \right) = 0$
$(ghxxu) : A g_{xxu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{u} \right) + f_{zz} \left( \left( z_{x} \otimes z_{xu} \right)P_{x\_xu} + \left(z_{xx} \otimes z_{u}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{x} \otimes g^{*}_{xu}\right)P_{x\_xu} + \left(g^{*}_{xx} \otimes g^{*}_{u}\right) \right) = 0$
$(ghxuu) : A g_{xuu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( \left( z_{xu} \otimes z_{u} \right)P_{xu\_u} + \left(z_{x} \otimes z_{uu}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{xu} \otimes g^{*}_{u}\right)P_{xu\_u} + \left(g^{*}_{x} \otimes g^{*}_{uu}\right) \right) = 0$
$(ghuuu) : A g_{uuu} + B g_{xxx} \left(g^{*}_{u} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{u} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( z_{u} \otimes z_{uu} \right)P_{u\_uu} + f_{y_{+}}g^{**}_{xx} \left(g^{*}_{u} \otimes g^{*}_{uu}\right)P_{u\_uu} = 0$
$(ghx\sigma\sigma) : A g_{x\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_x + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{x} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{x} \otimes z_{\sigma\sigma}\right) + F_{xu_{+}u_{+}}\left(I_{n_x} \otimes vec(\Sigma)\right) = 0$
$F_{xu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_x^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{xu_{+}} \otimes z_{u_{+}} \right)P_{xu\_u} + \left(z_{x} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{x} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
$(ghu\sigma\sigma) : A g_{u\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_{u} + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{u} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{u} \otimes z_{\sigma\sigma}\right) + F_{uu_{+}u_{+}}\left(I_{n_u} \otimes vec(\Sigma_u)\right) = 0$
$F_{uu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_u^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{uu_{+}} \otimes z_{u_{+}} \right)P_{uu\_u} + \left(z_{u} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{u} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
A and B are the common perturbation matrices:
$A = f_{y_0} + \begin{pmatrix} \underbrace{0}_{n\times n_{static}} &\vdots& \underbrace{f_{y^{**}_{+}} \cdot g^{**}_{x}}_{n \times n_{spred}} &\vdots& \underbrace{0}_{n\times n_{frwd}} \end{pmatrix}$and $B = \begin{pmatrix} \underbrace{0}_{n \times n_{static}}&\vdots & \underbrace{0}_{n \times n_{pred}} & \vdots & \underbrace{f_{y^{**}_{+}}}_{n \times n_{sfwrd}} \end{pmatrix}$
and $z=(y_{-}^{\ast}; y; y_{+}^{\ast\ast}; u)$ denotes the dynamic model variables as in `M_.lead_lag_incidence`, $y^\ast$ denote state variables, $y^{\ast\ast}$ denote forward looking variables, $y_+$ denote the variables with a lead, $y_{-}$ denote variables with a lag, $y_0$ denote variables at period t, $f$ the model equations, and $f_z$ the first-order dynamic model derivatives, $f_{zz}$ the second-order dynamic derivatives, and $f_{zzz}$ the third-order dynamic model derivatives. Then:
$z_{x} = \begin{pmatrix}I\\g_{x}\\g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{u} =\begin{pmatrix}0\\g_{u}\\g^{**}_{x} \cdot g^{*}_{u}\\I\end{pmatrix}$, $z_{u_{+}} =\begin{pmatrix}0\\0\\g^{**}_{u}\\0\end{pmatrix}$
$z_{xx} = \begin{pmatrix} 0\\g_{xx}\\g^{**}_{x} \left( g^{*}_x \otimes g^{*}_{x} \right) + g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{xu} =\begin{pmatrix}0\\g_{xu}\\g^{**}_{xx} \left( g^{*}_x \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{xu}\\0\end{pmatrix}$, $z_{uu} =\begin{pmatrix}0\\g_{uu}\\g^{**}_{xx} \left( g^{*}_u \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{uu}\\0\end{pmatrix}$,
$z_{xu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_x \otimes I \right)\\0\end{pmatrix}$, $z_{uu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_{u} \otimes I \right)\\0\end{pmatrix}$, $z_{u_{+}u_{+}} =\begin{pmatrix}0\\0\\g^{\ast\ast}_{uu}\\0\end{pmatrix}$, $z_{\sigma\sigma} = \begin{pmatrix}0\\ g_{\sigma\sigma}\\ g^{\ast\ast}_{x}g^{\ast}_{\sigma\sigma} + g^{\ast\ast}_{\sigma\sigma}\\0 \end{pmatrix}$
$P$ are permutation matrices that can be computed using Matlab's `ipermute` function.
## Parameter derivatives of perturbation solutions
First, we need the parameter derivatives of first, second, third, and fourth derivatives of the dynamic model (i.e. g1,g2,g3,g4 in dynamic files), I make use of the implicit function theorem: Let $f_{z^k}$ denote the kth derivative (wrt all dynamic variables) of the dynamic model, then let $df_{z^k}$ denote the first-derivative (wrt all model parameters) of $f_{z^k}$ evaluated at the steady state. Note that $f_{z^k}$ is a function of both the model parameters $\theta$ and of the steady state of all dynamic variables $\bar{z}$, which also depend on the parameters. Hence, implicitly $f_{z^k}=f_{z^k}(\theta,\bar{z}(\theta))$ and $df_{z^k}$ consists of two parts:
1. direct derivative wrt to all model parameters given by the preprocessor in the `_params_derivs.m` files
2. contribution of derivative of steady state of dynamic variables (wrt all model parameters): $f_{z^{k+1}} \cdot d\bar{z}$
Note that we already have functionality to compute $d\bar{z}$ analytically.
Having this, the above perturbation systems are basically equations of the following types
$AX +BXC = RHS$ or $AX = RHS$
Now when taking the derivative (wrt to one single parameter $\theta_j$), we get
$A\mathrm{d}\{X\} + B\mathrm{d}\{X\}C = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X - \mathrm{d}\{B\}XC - BX\mathrm{d}\{C\}$
or
$A\mathrm{d}\{X\} = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X$
The first one is a Sylvester type equation, the second one can be solved by taking the inverse of $A$. The only diffculty and tedious work arrises in computing (the highly sparse) derivatives of $RHS$.
***
# New functions: `
## get_perturbation_params_derivs.m`and `get_perturbation_params_derivs_numerical_objective.m`
* The parameter derivatives up to third order are computed in the new function`get_perturbation_params_derivs.m` both analytically and numerically. For numerical derivatives `get_perturbation_params_derivs_numerical_objective.m` is the objective for `fjaco.m` or `hessian_sparse.m` or `hessian.m`.
* `get_perturbation_params_derivs.m` is basically an extended version of the previous `get_first_order_solution_params_derivs.m` function.
* * `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It is used for numerical derivatives, whenever `analytic_derivation_mode=-1|-2`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
* Detailed changes:
* Most important: notation of this function is now in accordance to the k_order_solver, i.e. we do not compute derivatives of Kalman transition matrices A and B, but rather the solution matrices ghx,ghu,ghxx,ghxu,ghuu,ghs2,ghxxx,ghxxu,ghxuu,ghuuu,ghxss,ghuss in accordance with notation used in `oo_.dr`. As a byproduct at first-order, focusing on ghx and ghu instead of Kalman transition matrices A and B makes the computations slightly faster for large models (e.g. for Quest the computations were faster by a couple of seconds, not much, but okay).
* Removed use of `kstate`, see also Dynare/dynare#1653 and Dynare/dynare!1656
* Output arguments are stored in a structure `DERIVS`, there is also a flag `d2flag` that computes parameter hessians needed only in `dsge_likelihood.m`.
* Removed `kronflag` as input. `options_.analytic_derivation_mode` is now used instead of `kronflag`.
* Removed `indvar`, an index that was used to selected specific variables in the derivatives. This is not needed, as we always compute the parameter derivatives for all variables first and then select a subset of variables. The selection now takes place in other functions, like `dsge_likelihood.m`.
* Introduced some checks: (i) deterministic exogenous variables are not supported, (ii) Kronecker method only compatible with first-order approximation so reset to sylvester method, (iii) for purely backward or forward models we need to be careful with the rows in `M_.lead_la g_incidence`, (iv) if `_params_derivs.m` files are missing an error is thrown.
* For numerical derivatives, if mod file does not contain an `estimated_params_block`, a temporary one with the most important parameter information is created.
## `unfold_g4.m`
* When evaluating g3 and g4 one needs to take into account that these do not contain symmetric elements, so one needs to use `unfold_g3.m` and the new function `unfold_g4.m`. This returns an unfolded version of the same matrix (i.e. with symmetric elements).
***
# New test models
`.gitignore` and `Makefile.am` are changed accordingly. Also now it is possible to run test suite on analytic_derivatives, i.e. run `make check m/analytic_derivatives`
## `analytic_derivatives/BrockMirman_PertParamsDerivs.mod`
* This is the Brock Mirman model, where we know the exact policy function $g$ for capital and consumption. As this does not imply a nonzero $g_{\sigma\sigma}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ I added some artificial equations to get nonzero solution matrices with respect to $\sigma$. The true perturbation solution matrices $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ are then computed analytically with Matlab's symbolic toolbox and saved in `nBrockMirmanSYM.mat`. There is a preprocessor flag that recreates these analytical computations if changes are needed (and to check whether I made some errors here ;-) )
* Then solution matrices up to third order and their parameter Jacobians are then compared to the ones computed by Dynare's `k_order_solver` and by `get_perturbation_params_derivs` for all `analytic_derivation_mode`'s. There will be an error if the maximum absolute deviation is too large, i.e. for numerical derivatives (`analytic_derivation_mode=-1|-2`) the tolerance is choosen lower (around 1e-5); for analytical methods we are stricter: around 1e-13 for first-order, 1e-12 for second order, and 1e-11 for third-order.
* As a side note, this mod file also checks Dynare's `k_order_solver` algorithm and throws an error if something is wrong.
* This test model shows that the new functionality works well. And analytical derivatives perform way better and accurate than numerical ones, even for this small model.
## `analytic_derivatives/burnside_3_order_PertParamsDerivs.mod`
* This builds upon `tests/k_order_perturbation/burnside_k_order.mod` and computes the true parameter derivatives analytically by hand.
* This test model also shows that the new functionality works well.
## `analytic_derivatives/LindeTrabandt2019.mod`
* Shows that the new functionality also works for medium-sized models, i.e. a SW type model solved at third order with 35 variables (11 states). 2 shocks and 20 parameters.
* This mod file can be used to tweak the speed of the computations in the future.
* Compares numerical versus analytical parameter derivatives (for first, second and third order). Note that this model clearly shows that numerical ones are quite different than analytical ones even at first order!
## `identification/LindeTrabandt2019_xfail.mod`
* This model is a check for issue Dynare/dynare#1595, see fjaco.m below, and will fail.
* Removed `analytic_derivatives/ls2003.mod` as this mod file is neither in the testsuite nor does it work.
***
# Detailed changes in other functions
## `get_first_order_solution_params_derivs.m`
* Deleted, or actually, renamed to `get_perturbation_params_derivs.m`, as this function now is able to compute the derivatives up to third order
## `identification_numerical_objective.m`
* `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
## `dsge_likelihood.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `T`matrix anymore, but instead of the dynare solution matrix `ghx`. So we recreate `T` here as this amounts to adding some zeros and focusing on selected variables only.
* Added some checks to make sure the first-order approximation is selected.
* Removed `kron_flag` as input, as `get_perturbation_params_derivs` looks into `options_.analytic_derivation_mode` for `kron_flag`.
## `dynare_identification.m`
* make sure that setting `analytic_derivation_mode` is set both in `options_ident` and `options_`. Note that at the end of the function we restore the `options_` structure, so all changes are local. In a next merge request, I will remove the global variables to make all variables local.
## `get_identification_jacobians.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `A` and `B` matrix anymore, but instead of the dynare solution matrix `ghx` and `ghu`. So we recreate these matrices here instead of in `get_perturbation_params_derivs.m`.
* Added `str2func` for better function handles in `fjaco.m`.
## `fjaco.m`
* make `tol`an option, which can be adjusted by changing `options_.dynatol.x`for identification and parameter derivatives purposes.
* include a check and an informative error message, if numerical derivatives (two-sided finite difference method) yield errors in `resol.m` for identification and parameter derivatives purposes. This closes issue Dynare/dynare#1595.
* Changed year of copyright to 2010-2017,2019
***
# Further suggestions and questions
* Ones this is merged, I will merge request an improvement of the identification toolbox, which will work up to third order using the pruned state space. This will also remove some issues and bugs, and also I will remove global variables in this request.
* The third-order derivatives can be further improved by taking sparsity into account and use mex versions for kronecker products etc. I leave this for further testing (and if anybody actually uses this ;-) )
2019-12-17 19:17:09 +01:00
% permutation matrices
id_xxx = reshape ( 1 : M . nspred ^3 , 1 , M . nspred , M . nspred , M . nspred ) ;
id_uux = reshape ( 1 : M . nspred * M . exo_nbr ^2 , 1 , M . exo_nbr , M . exo_nbr , M . nspred ) ;
id_uxx = reshape ( 1 : M . nspred ^2 * M . exo_nbr , 1 , M . exo_nbr , M . nspred , M . nspred ) ;
2019-12-20 16:28:06 +01:00
id_uuu = reshape ( 1 : M . exo_nbr ^3 , 1 , M . exo_nbr , M . exo_nbr , M . exo_nbr ) ;
Added parameter derivatives of perturbation solution up to 3 order
# Preliminary comments
I finished the identification toolbox at orders two and three using the pruned state space system, but before I merge request this, I decided to first merge the new functionality to compute parameter derivatives of perturbation solution matrices at higher orders. So after this is approved, I merge the identification toolbox.
I guess @rattoma, @sebastien, and @michel are best choices to review this.
I outline the main idea first and then provide some more detailed changes I made to the functions.
***
# Main idea
This merge request is concerned with the *analytical*computation of the parameter derivatives of first, second and third order perturbation solution matrices, i.e. using _closed-form_ expressions to efficiently compute the derivative of $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ *with respect to model parameters* $\theta$. Note that $\theta$ contains model parameters, stderr and corr parameters of shocks. stderr and corr parameters of measurement errors are not yet supported, (they can easily be included as exogenous shocks). The availability of such derivatives is beneficial in terms of more reliable analysis of model sensitivity and parameter identifiability as well as more efficient estimation methods, in particular for models solved up to third order, as it is well-known that numerical derivatives are a tricky business, especially for large models.
References for my approach are:
* Iskrev (2008, 2010) and Schmitt-Grohé and Uribe (2012, Appendix) who were the first to compute the parameter derivatives analytically at first order, however, using inefficient (sparse) Kronecker products.
* Mutschler (2015) who provides the expressions for a second-order, but again using inefficient (sparse) Kronecker products.
* Ratto and Iskrev (2012) who show how the first-order system can be solved accurately, fast and efficiently using existing numerical algorithms for generalized Sylvester equations by taking the parameter derivative with respect to each parameter separately.
* Julliard and Kamenik (2004) who provide the perturbation solution equation system in tensor notation at any order k.
* Levintal (2017) who introduces permutation matrices to express the perturbation solution equation system in matrix notation up to fifth order.
Note that @rattoma already implemented the parameter derivatives of $g_x$ and $g_u$ analytically (and numerically), and I rely heavily on his work in `get_first_order_solution_params_derivs.m` (previously `getH.m`). My additions are mainly to this function and thus it is renamed to `get_perturbation_params_derivs.m`.
The basic idea of this merge request is to take the second- and third-order perturbation solution systems in Julliard and Kamenik (2004), unfold these into an equivalent matrix representation using permutation matrices as in Levintal (2017). Then extending Ratto and Iskrev (2012) one takes the derivative with respect to each parameter separately and gets a computational problem that is linear, albeit large, as it involves either solving generalized Sylvester equations or taking inverses of highly sparse matrices. I will now briefly summarize the perturbation solution system at third order and the system that results when taking the derivative with respect to parameters.
## Perturbation Solution
The following systems arise at first, second, and third order:
$(ghx): f_{x} z_{x} = f_{y_{-}^*} + f_{y_0} g_{x} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{x}= A g_{x} + f_{y_{-}^*}=0$
$(ghu): f_{z} z_{u} = f_{y_0} g_{u} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{u} + f_{u}= A g_u + f_u = 0$
$(ghxx) : A g_{xx} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{x}\right) + f_{zz} \left( z_{x} \otimes z_{x} \right) = 0$
$(ghxu) : A g_{xu} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{x} \otimes z_{u} \right) = 0$
$(ghuu) : A g_{uu} + B g_{xx} \left(g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{u} \otimes z_{u} \right) = 0$
$(ghs2) : (A+B) g_{\sigma\sigma} + \left( f_{y^{**}_{+}y^{**}_{+}} \left(g^{**}_{u} \otimes g^{**}_{u}\right) + f_{y^{**}_{+}} g^{**}_{uu}\right)vec(\Sigma) = 0$
$(ghxxx) : A g_{xxx} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{x}\right) + f_{y_{+}}g^{**}_{xx} \left(g^{*}_x \otimes g^{*}_{xx}\right)P_{x\_xx} + f_{zz} \left( z_{x} \otimes z_{xx} \right)P_{x\_xx} + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{x} \right) = 0$
$(ghxxu) : A g_{xxu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{u} \right) + f_{zz} \left( \left( z_{x} \otimes z_{xu} \right)P_{x\_xu} + \left(z_{xx} \otimes z_{u}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{x} \otimes g^{*}_{xu}\right)P_{x\_xu} + \left(g^{*}_{xx} \otimes g^{*}_{u}\right) \right) = 0$
$(ghxuu) : A g_{xuu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( \left( z_{xu} \otimes z_{u} \right)P_{xu\_u} + \left(z_{x} \otimes z_{uu}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{xu} \otimes g^{*}_{u}\right)P_{xu\_u} + \left(g^{*}_{x} \otimes g^{*}_{uu}\right) \right) = 0$
$(ghuuu) : A g_{uuu} + B g_{xxx} \left(g^{*}_{u} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{u} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( z_{u} \otimes z_{uu} \right)P_{u\_uu} + f_{y_{+}}g^{**}_{xx} \left(g^{*}_{u} \otimes g^{*}_{uu}\right)P_{u\_uu} = 0$
$(ghx\sigma\sigma) : A g_{x\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_x + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{x} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{x} \otimes z_{\sigma\sigma}\right) + F_{xu_{+}u_{+}}\left(I_{n_x} \otimes vec(\Sigma)\right) = 0$
$F_{xu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_x^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{xu_{+}} \otimes z_{u_{+}} \right)P_{xu\_u} + \left(z_{x} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{x} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
$(ghu\sigma\sigma) : A g_{u\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_{u} + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{u} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{u} \otimes z_{\sigma\sigma}\right) + F_{uu_{+}u_{+}}\left(I_{n_u} \otimes vec(\Sigma_u)\right) = 0$
$F_{uu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_u^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{uu_{+}} \otimes z_{u_{+}} \right)P_{uu\_u} + \left(z_{u} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{u} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
A and B are the common perturbation matrices:
$A = f_{y_0} + \begin{pmatrix} \underbrace{0}_{n\times n_{static}} &\vdots& \underbrace{f_{y^{**}_{+}} \cdot g^{**}_{x}}_{n \times n_{spred}} &\vdots& \underbrace{0}_{n\times n_{frwd}} \end{pmatrix}$and $B = \begin{pmatrix} \underbrace{0}_{n \times n_{static}}&\vdots & \underbrace{0}_{n \times n_{pred}} & \vdots & \underbrace{f_{y^{**}_{+}}}_{n \times n_{sfwrd}} \end{pmatrix}$
and $z=(y_{-}^{\ast}; y; y_{+}^{\ast\ast}; u)$ denotes the dynamic model variables as in `M_.lead_lag_incidence`, $y^\ast$ denote state variables, $y^{\ast\ast}$ denote forward looking variables, $y_+$ denote the variables with a lead, $y_{-}$ denote variables with a lag, $y_0$ denote variables at period t, $f$ the model equations, and $f_z$ the first-order dynamic model derivatives, $f_{zz}$ the second-order dynamic derivatives, and $f_{zzz}$ the third-order dynamic model derivatives. Then:
$z_{x} = \begin{pmatrix}I\\g_{x}\\g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{u} =\begin{pmatrix}0\\g_{u}\\g^{**}_{x} \cdot g^{*}_{u}\\I\end{pmatrix}$, $z_{u_{+}} =\begin{pmatrix}0\\0\\g^{**}_{u}\\0\end{pmatrix}$
$z_{xx} = \begin{pmatrix} 0\\g_{xx}\\g^{**}_{x} \left( g^{*}_x \otimes g^{*}_{x} \right) + g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{xu} =\begin{pmatrix}0\\g_{xu}\\g^{**}_{xx} \left( g^{*}_x \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{xu}\\0\end{pmatrix}$, $z_{uu} =\begin{pmatrix}0\\g_{uu}\\g^{**}_{xx} \left( g^{*}_u \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{uu}\\0\end{pmatrix}$,
$z_{xu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_x \otimes I \right)\\0\end{pmatrix}$, $z_{uu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_{u} \otimes I \right)\\0\end{pmatrix}$, $z_{u_{+}u_{+}} =\begin{pmatrix}0\\0\\g^{\ast\ast}_{uu}\\0\end{pmatrix}$, $z_{\sigma\sigma} = \begin{pmatrix}0\\ g_{\sigma\sigma}\\ g^{\ast\ast}_{x}g^{\ast}_{\sigma\sigma} + g^{\ast\ast}_{\sigma\sigma}\\0 \end{pmatrix}$
$P$ are permutation matrices that can be computed using Matlab's `ipermute` function.
## Parameter derivatives of perturbation solutions
First, we need the parameter derivatives of first, second, third, and fourth derivatives of the dynamic model (i.e. g1,g2,g3,g4 in dynamic files), I make use of the implicit function theorem: Let $f_{z^k}$ denote the kth derivative (wrt all dynamic variables) of the dynamic model, then let $df_{z^k}$ denote the first-derivative (wrt all model parameters) of $f_{z^k}$ evaluated at the steady state. Note that $f_{z^k}$ is a function of both the model parameters $\theta$ and of the steady state of all dynamic variables $\bar{z}$, which also depend on the parameters. Hence, implicitly $f_{z^k}=f_{z^k}(\theta,\bar{z}(\theta))$ and $df_{z^k}$ consists of two parts:
1. direct derivative wrt to all model parameters given by the preprocessor in the `_params_derivs.m` files
2. contribution of derivative of steady state of dynamic variables (wrt all model parameters): $f_{z^{k+1}} \cdot d\bar{z}$
Note that we already have functionality to compute $d\bar{z}$ analytically.
Having this, the above perturbation systems are basically equations of the following types
$AX +BXC = RHS$ or $AX = RHS$
Now when taking the derivative (wrt to one single parameter $\theta_j$), we get
$A\mathrm{d}\{X\} + B\mathrm{d}\{X\}C = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X - \mathrm{d}\{B\}XC - BX\mathrm{d}\{C\}$
or
$A\mathrm{d}\{X\} = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X$
The first one is a Sylvester type equation, the second one can be solved by taking the inverse of $A$. The only diffculty and tedious work arrises in computing (the highly sparse) derivatives of $RHS$.
***
# New functions: `
## get_perturbation_params_derivs.m`and `get_perturbation_params_derivs_numerical_objective.m`
* The parameter derivatives up to third order are computed in the new function`get_perturbation_params_derivs.m` both analytically and numerically. For numerical derivatives `get_perturbation_params_derivs_numerical_objective.m` is the objective for `fjaco.m` or `hessian_sparse.m` or `hessian.m`.
* `get_perturbation_params_derivs.m` is basically an extended version of the previous `get_first_order_solution_params_derivs.m` function.
* * `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It is used for numerical derivatives, whenever `analytic_derivation_mode=-1|-2`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
* Detailed changes:
* Most important: notation of this function is now in accordance to the k_order_solver, i.e. we do not compute derivatives of Kalman transition matrices A and B, but rather the solution matrices ghx,ghu,ghxx,ghxu,ghuu,ghs2,ghxxx,ghxxu,ghxuu,ghuuu,ghxss,ghuss in accordance with notation used in `oo_.dr`. As a byproduct at first-order, focusing on ghx and ghu instead of Kalman transition matrices A and B makes the computations slightly faster for large models (e.g. for Quest the computations were faster by a couple of seconds, not much, but okay).
* Removed use of `kstate`, see also Dynare/dynare#1653 and Dynare/dynare!1656
* Output arguments are stored in a structure `DERIVS`, there is also a flag `d2flag` that computes parameter hessians needed only in `dsge_likelihood.m`.
* Removed `kronflag` as input. `options_.analytic_derivation_mode` is now used instead of `kronflag`.
* Removed `indvar`, an index that was used to selected specific variables in the derivatives. This is not needed, as we always compute the parameter derivatives for all variables first and then select a subset of variables. The selection now takes place in other functions, like `dsge_likelihood.m`.
* Introduced some checks: (i) deterministic exogenous variables are not supported, (ii) Kronecker method only compatible with first-order approximation so reset to sylvester method, (iii) for purely backward or forward models we need to be careful with the rows in `M_.lead_la g_incidence`, (iv) if `_params_derivs.m` files are missing an error is thrown.
* For numerical derivatives, if mod file does not contain an `estimated_params_block`, a temporary one with the most important parameter information is created.
## `unfold_g4.m`
* When evaluating g3 and g4 one needs to take into account that these do not contain symmetric elements, so one needs to use `unfold_g3.m` and the new function `unfold_g4.m`. This returns an unfolded version of the same matrix (i.e. with symmetric elements).
***
# New test models
`.gitignore` and `Makefile.am` are changed accordingly. Also now it is possible to run test suite on analytic_derivatives, i.e. run `make check m/analytic_derivatives`
## `analytic_derivatives/BrockMirman_PertParamsDerivs.mod`
* This is the Brock Mirman model, where we know the exact policy function $g$ for capital and consumption. As this does not imply a nonzero $g_{\sigma\sigma}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ I added some artificial equations to get nonzero solution matrices with respect to $\sigma$. The true perturbation solution matrices $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ are then computed analytically with Matlab's symbolic toolbox and saved in `nBrockMirmanSYM.mat`. There is a preprocessor flag that recreates these analytical computations if changes are needed (and to check whether I made some errors here ;-) )
* Then solution matrices up to third order and their parameter Jacobians are then compared to the ones computed by Dynare's `k_order_solver` and by `get_perturbation_params_derivs` for all `analytic_derivation_mode`'s. There will be an error if the maximum absolute deviation is too large, i.e. for numerical derivatives (`analytic_derivation_mode=-1|-2`) the tolerance is choosen lower (around 1e-5); for analytical methods we are stricter: around 1e-13 for first-order, 1e-12 for second order, and 1e-11 for third-order.
* As a side note, this mod file also checks Dynare's `k_order_solver` algorithm and throws an error if something is wrong.
* This test model shows that the new functionality works well. And analytical derivatives perform way better and accurate than numerical ones, even for this small model.
## `analytic_derivatives/burnside_3_order_PertParamsDerivs.mod`
* This builds upon `tests/k_order_perturbation/burnside_k_order.mod` and computes the true parameter derivatives analytically by hand.
* This test model also shows that the new functionality works well.
## `analytic_derivatives/LindeTrabandt2019.mod`
* Shows that the new functionality also works for medium-sized models, i.e. a SW type model solved at third order with 35 variables (11 states). 2 shocks and 20 parameters.
* This mod file can be used to tweak the speed of the computations in the future.
* Compares numerical versus analytical parameter derivatives (for first, second and third order). Note that this model clearly shows that numerical ones are quite different than analytical ones even at first order!
## `identification/LindeTrabandt2019_xfail.mod`
* This model is a check for issue Dynare/dynare#1595, see fjaco.m below, and will fail.
* Removed `analytic_derivatives/ls2003.mod` as this mod file is neither in the testsuite nor does it work.
***
# Detailed changes in other functions
## `get_first_order_solution_params_derivs.m`
* Deleted, or actually, renamed to `get_perturbation_params_derivs.m`, as this function now is able to compute the derivatives up to third order
## `identification_numerical_objective.m`
* `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
## `dsge_likelihood.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `T`matrix anymore, but instead of the dynare solution matrix `ghx`. So we recreate `T` here as this amounts to adding some zeros and focusing on selected variables only.
* Added some checks to make sure the first-order approximation is selected.
* Removed `kron_flag` as input, as `get_perturbation_params_derivs` looks into `options_.analytic_derivation_mode` for `kron_flag`.
## `dynare_identification.m`
* make sure that setting `analytic_derivation_mode` is set both in `options_ident` and `options_`. Note that at the end of the function we restore the `options_` structure, so all changes are local. In a next merge request, I will remove the global variables to make all variables local.
## `get_identification_jacobians.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `A` and `B` matrix anymore, but instead of the dynare solution matrix `ghx` and `ghu`. So we recreate these matrices here instead of in `get_perturbation_params_derivs.m`.
* Added `str2func` for better function handles in `fjaco.m`.
## `fjaco.m`
* make `tol`an option, which can be adjusted by changing `options_.dynatol.x`for identification and parameter derivatives purposes.
* include a check and an informative error message, if numerical derivatives (two-sided finite difference method) yield errors in `resol.m` for identification and parameter derivatives purposes. This closes issue Dynare/dynare#1595.
* Changed year of copyright to 2010-2017,2019
***
# Further suggestions and questions
* Ones this is merged, I will merge request an improvement of the identification toolbox, which will work up to third order using the pruned state space. This will also remove some issues and bugs, and also I will remove global variables in this request.
* The third-order derivatives can be further improved by taking sparsity into account and use mex versions for kronecker products etc. I leave this for further testing (and if anybody actually uses this ;-) )
2019-12-17 19:17:09 +01:00
I_xxx = speye ( M . nspred ^3 ) ;
I_xxu = speye ( M . nspred ^2 * M . exo_nbr ) ;
I_xuu = speye ( M . nspred * M . exo_nbr ^2 ) ;
2019-12-20 16:28:06 +01:00
I_uuu = speye ( M . exo_nbr ^3 ) ;
Added parameter derivatives of perturbation solution up to 3 order
# Preliminary comments
I finished the identification toolbox at orders two and three using the pruned state space system, but before I merge request this, I decided to first merge the new functionality to compute parameter derivatives of perturbation solution matrices at higher orders. So after this is approved, I merge the identification toolbox.
I guess @rattoma, @sebastien, and @michel are best choices to review this.
I outline the main idea first and then provide some more detailed changes I made to the functions.
***
# Main idea
This merge request is concerned with the *analytical*computation of the parameter derivatives of first, second and third order perturbation solution matrices, i.e. using _closed-form_ expressions to efficiently compute the derivative of $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ *with respect to model parameters* $\theta$. Note that $\theta$ contains model parameters, stderr and corr parameters of shocks. stderr and corr parameters of measurement errors are not yet supported, (they can easily be included as exogenous shocks). The availability of such derivatives is beneficial in terms of more reliable analysis of model sensitivity and parameter identifiability as well as more efficient estimation methods, in particular for models solved up to third order, as it is well-known that numerical derivatives are a tricky business, especially for large models.
References for my approach are:
* Iskrev (2008, 2010) and Schmitt-Grohé and Uribe (2012, Appendix) who were the first to compute the parameter derivatives analytically at first order, however, using inefficient (sparse) Kronecker products.
* Mutschler (2015) who provides the expressions for a second-order, but again using inefficient (sparse) Kronecker products.
* Ratto and Iskrev (2012) who show how the first-order system can be solved accurately, fast and efficiently using existing numerical algorithms for generalized Sylvester equations by taking the parameter derivative with respect to each parameter separately.
* Julliard and Kamenik (2004) who provide the perturbation solution equation system in tensor notation at any order k.
* Levintal (2017) who introduces permutation matrices to express the perturbation solution equation system in matrix notation up to fifth order.
Note that @rattoma already implemented the parameter derivatives of $g_x$ and $g_u$ analytically (and numerically), and I rely heavily on his work in `get_first_order_solution_params_derivs.m` (previously `getH.m`). My additions are mainly to this function and thus it is renamed to `get_perturbation_params_derivs.m`.
The basic idea of this merge request is to take the second- and third-order perturbation solution systems in Julliard and Kamenik (2004), unfold these into an equivalent matrix representation using permutation matrices as in Levintal (2017). Then extending Ratto and Iskrev (2012) one takes the derivative with respect to each parameter separately and gets a computational problem that is linear, albeit large, as it involves either solving generalized Sylvester equations or taking inverses of highly sparse matrices. I will now briefly summarize the perturbation solution system at third order and the system that results when taking the derivative with respect to parameters.
## Perturbation Solution
The following systems arise at first, second, and third order:
$(ghx): f_{x} z_{x} = f_{y_{-}^*} + f_{y_0} g_{x} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{x}= A g_{x} + f_{y_{-}^*}=0$
$(ghu): f_{z} z_{u} = f_{y_0} g_{u} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{u} + f_{u}= A g_u + f_u = 0$
$(ghxx) : A g_{xx} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{x}\right) + f_{zz} \left( z_{x} \otimes z_{x} \right) = 0$
$(ghxu) : A g_{xu} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{x} \otimes z_{u} \right) = 0$
$(ghuu) : A g_{uu} + B g_{xx} \left(g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{u} \otimes z_{u} \right) = 0$
$(ghs2) : (A+B) g_{\sigma\sigma} + \left( f_{y^{**}_{+}y^{**}_{+}} \left(g^{**}_{u} \otimes g^{**}_{u}\right) + f_{y^{**}_{+}} g^{**}_{uu}\right)vec(\Sigma) = 0$
$(ghxxx) : A g_{xxx} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{x}\right) + f_{y_{+}}g^{**}_{xx} \left(g^{*}_x \otimes g^{*}_{xx}\right)P_{x\_xx} + f_{zz} \left( z_{x} \otimes z_{xx} \right)P_{x\_xx} + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{x} \right) = 0$
$(ghxxu) : A g_{xxu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{u} \right) + f_{zz} \left( \left( z_{x} \otimes z_{xu} \right)P_{x\_xu} + \left(z_{xx} \otimes z_{u}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{x} \otimes g^{*}_{xu}\right)P_{x\_xu} + \left(g^{*}_{xx} \otimes g^{*}_{u}\right) \right) = 0$
$(ghxuu) : A g_{xuu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( \left( z_{xu} \otimes z_{u} \right)P_{xu\_u} + \left(z_{x} \otimes z_{uu}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{xu} \otimes g^{*}_{u}\right)P_{xu\_u} + \left(g^{*}_{x} \otimes g^{*}_{uu}\right) \right) = 0$
$(ghuuu) : A g_{uuu} + B g_{xxx} \left(g^{*}_{u} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{u} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( z_{u} \otimes z_{uu} \right)P_{u\_uu} + f_{y_{+}}g^{**}_{xx} \left(g^{*}_{u} \otimes g^{*}_{uu}\right)P_{u\_uu} = 0$
$(ghx\sigma\sigma) : A g_{x\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_x + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{x} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{x} \otimes z_{\sigma\sigma}\right) + F_{xu_{+}u_{+}}\left(I_{n_x} \otimes vec(\Sigma)\right) = 0$
$F_{xu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_x^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{xu_{+}} \otimes z_{u_{+}} \right)P_{xu\_u} + \left(z_{x} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{x} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
$(ghu\sigma\sigma) : A g_{u\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_{u} + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{u} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{u} \otimes z_{\sigma\sigma}\right) + F_{uu_{+}u_{+}}\left(I_{n_u} \otimes vec(\Sigma_u)\right) = 0$
$F_{uu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_u^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{uu_{+}} \otimes z_{u_{+}} \right)P_{uu\_u} + \left(z_{u} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{u} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
A and B are the common perturbation matrices:
$A = f_{y_0} + \begin{pmatrix} \underbrace{0}_{n\times n_{static}} &\vdots& \underbrace{f_{y^{**}_{+}} \cdot g^{**}_{x}}_{n \times n_{spred}} &\vdots& \underbrace{0}_{n\times n_{frwd}} \end{pmatrix}$and $B = \begin{pmatrix} \underbrace{0}_{n \times n_{static}}&\vdots & \underbrace{0}_{n \times n_{pred}} & \vdots & \underbrace{f_{y^{**}_{+}}}_{n \times n_{sfwrd}} \end{pmatrix}$
and $z=(y_{-}^{\ast}; y; y_{+}^{\ast\ast}; u)$ denotes the dynamic model variables as in `M_.lead_lag_incidence`, $y^\ast$ denote state variables, $y^{\ast\ast}$ denote forward looking variables, $y_+$ denote the variables with a lead, $y_{-}$ denote variables with a lag, $y_0$ denote variables at period t, $f$ the model equations, and $f_z$ the first-order dynamic model derivatives, $f_{zz}$ the second-order dynamic derivatives, and $f_{zzz}$ the third-order dynamic model derivatives. Then:
$z_{x} = \begin{pmatrix}I\\g_{x}\\g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{u} =\begin{pmatrix}0\\g_{u}\\g^{**}_{x} \cdot g^{*}_{u}\\I\end{pmatrix}$, $z_{u_{+}} =\begin{pmatrix}0\\0\\g^{**}_{u}\\0\end{pmatrix}$
$z_{xx} = \begin{pmatrix} 0\\g_{xx}\\g^{**}_{x} \left( g^{*}_x \otimes g^{*}_{x} \right) + g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{xu} =\begin{pmatrix}0\\g_{xu}\\g^{**}_{xx} \left( g^{*}_x \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{xu}\\0\end{pmatrix}$, $z_{uu} =\begin{pmatrix}0\\g_{uu}\\g^{**}_{xx} \left( g^{*}_u \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{uu}\\0\end{pmatrix}$,
$z_{xu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_x \otimes I \right)\\0\end{pmatrix}$, $z_{uu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_{u} \otimes I \right)\\0\end{pmatrix}$, $z_{u_{+}u_{+}} =\begin{pmatrix}0\\0\\g^{\ast\ast}_{uu}\\0\end{pmatrix}$, $z_{\sigma\sigma} = \begin{pmatrix}0\\ g_{\sigma\sigma}\\ g^{\ast\ast}_{x}g^{\ast}_{\sigma\sigma} + g^{\ast\ast}_{\sigma\sigma}\\0 \end{pmatrix}$
$P$ are permutation matrices that can be computed using Matlab's `ipermute` function.
## Parameter derivatives of perturbation solutions
First, we need the parameter derivatives of first, second, third, and fourth derivatives of the dynamic model (i.e. g1,g2,g3,g4 in dynamic files), I make use of the implicit function theorem: Let $f_{z^k}$ denote the kth derivative (wrt all dynamic variables) of the dynamic model, then let $df_{z^k}$ denote the first-derivative (wrt all model parameters) of $f_{z^k}$ evaluated at the steady state. Note that $f_{z^k}$ is a function of both the model parameters $\theta$ and of the steady state of all dynamic variables $\bar{z}$, which also depend on the parameters. Hence, implicitly $f_{z^k}=f_{z^k}(\theta,\bar{z}(\theta))$ and $df_{z^k}$ consists of two parts:
1. direct derivative wrt to all model parameters given by the preprocessor in the `_params_derivs.m` files
2. contribution of derivative of steady state of dynamic variables (wrt all model parameters): $f_{z^{k+1}} \cdot d\bar{z}$
Note that we already have functionality to compute $d\bar{z}$ analytically.
Having this, the above perturbation systems are basically equations of the following types
$AX +BXC = RHS$ or $AX = RHS$
Now when taking the derivative (wrt to one single parameter $\theta_j$), we get
$A\mathrm{d}\{X\} + B\mathrm{d}\{X\}C = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X - \mathrm{d}\{B\}XC - BX\mathrm{d}\{C\}$
or
$A\mathrm{d}\{X\} = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X$
The first one is a Sylvester type equation, the second one can be solved by taking the inverse of $A$. The only diffculty and tedious work arrises in computing (the highly sparse) derivatives of $RHS$.
***
# New functions: `
## get_perturbation_params_derivs.m`and `get_perturbation_params_derivs_numerical_objective.m`
* The parameter derivatives up to third order are computed in the new function`get_perturbation_params_derivs.m` both analytically and numerically. For numerical derivatives `get_perturbation_params_derivs_numerical_objective.m` is the objective for `fjaco.m` or `hessian_sparse.m` or `hessian.m`.
* `get_perturbation_params_derivs.m` is basically an extended version of the previous `get_first_order_solution_params_derivs.m` function.
* * `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It is used for numerical derivatives, whenever `analytic_derivation_mode=-1|-2`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
* Detailed changes:
* Most important: notation of this function is now in accordance to the k_order_solver, i.e. we do not compute derivatives of Kalman transition matrices A and B, but rather the solution matrices ghx,ghu,ghxx,ghxu,ghuu,ghs2,ghxxx,ghxxu,ghxuu,ghuuu,ghxss,ghuss in accordance with notation used in `oo_.dr`. As a byproduct at first-order, focusing on ghx and ghu instead of Kalman transition matrices A and B makes the computations slightly faster for large models (e.g. for Quest the computations were faster by a couple of seconds, not much, but okay).
* Removed use of `kstate`, see also Dynare/dynare#1653 and Dynare/dynare!1656
* Output arguments are stored in a structure `DERIVS`, there is also a flag `d2flag` that computes parameter hessians needed only in `dsge_likelihood.m`.
* Removed `kronflag` as input. `options_.analytic_derivation_mode` is now used instead of `kronflag`.
* Removed `indvar`, an index that was used to selected specific variables in the derivatives. This is not needed, as we always compute the parameter derivatives for all variables first and then select a subset of variables. The selection now takes place in other functions, like `dsge_likelihood.m`.
* Introduced some checks: (i) deterministic exogenous variables are not supported, (ii) Kronecker method only compatible with first-order approximation so reset to sylvester method, (iii) for purely backward or forward models we need to be careful with the rows in `M_.lead_la g_incidence`, (iv) if `_params_derivs.m` files are missing an error is thrown.
* For numerical derivatives, if mod file does not contain an `estimated_params_block`, a temporary one with the most important parameter information is created.
## `unfold_g4.m`
* When evaluating g3 and g4 one needs to take into account that these do not contain symmetric elements, so one needs to use `unfold_g3.m` and the new function `unfold_g4.m`. This returns an unfolded version of the same matrix (i.e. with symmetric elements).
***
# New test models
`.gitignore` and `Makefile.am` are changed accordingly. Also now it is possible to run test suite on analytic_derivatives, i.e. run `make check m/analytic_derivatives`
## `analytic_derivatives/BrockMirman_PertParamsDerivs.mod`
* This is the Brock Mirman model, where we know the exact policy function $g$ for capital and consumption. As this does not imply a nonzero $g_{\sigma\sigma}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ I added some artificial equations to get nonzero solution matrices with respect to $\sigma$. The true perturbation solution matrices $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ are then computed analytically with Matlab's symbolic toolbox and saved in `nBrockMirmanSYM.mat`. There is a preprocessor flag that recreates these analytical computations if changes are needed (and to check whether I made some errors here ;-) )
* Then solution matrices up to third order and their parameter Jacobians are then compared to the ones computed by Dynare's `k_order_solver` and by `get_perturbation_params_derivs` for all `analytic_derivation_mode`'s. There will be an error if the maximum absolute deviation is too large, i.e. for numerical derivatives (`analytic_derivation_mode=-1|-2`) the tolerance is choosen lower (around 1e-5); for analytical methods we are stricter: around 1e-13 for first-order, 1e-12 for second order, and 1e-11 for third-order.
* As a side note, this mod file also checks Dynare's `k_order_solver` algorithm and throws an error if something is wrong.
* This test model shows that the new functionality works well. And analytical derivatives perform way better and accurate than numerical ones, even for this small model.
## `analytic_derivatives/burnside_3_order_PertParamsDerivs.mod`
* This builds upon `tests/k_order_perturbation/burnside_k_order.mod` and computes the true parameter derivatives analytically by hand.
* This test model also shows that the new functionality works well.
## `analytic_derivatives/LindeTrabandt2019.mod`
* Shows that the new functionality also works for medium-sized models, i.e. a SW type model solved at third order with 35 variables (11 states). 2 shocks and 20 parameters.
* This mod file can be used to tweak the speed of the computations in the future.
* Compares numerical versus analytical parameter derivatives (for first, second and third order). Note that this model clearly shows that numerical ones are quite different than analytical ones even at first order!
## `identification/LindeTrabandt2019_xfail.mod`
* This model is a check for issue Dynare/dynare#1595, see fjaco.m below, and will fail.
* Removed `analytic_derivatives/ls2003.mod` as this mod file is neither in the testsuite nor does it work.
***
# Detailed changes in other functions
## `get_first_order_solution_params_derivs.m`
* Deleted, or actually, renamed to `get_perturbation_params_derivs.m`, as this function now is able to compute the derivatives up to third order
## `identification_numerical_objective.m`
* `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
## `dsge_likelihood.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `T`matrix anymore, but instead of the dynare solution matrix `ghx`. So we recreate `T` here as this amounts to adding some zeros and focusing on selected variables only.
* Added some checks to make sure the first-order approximation is selected.
* Removed `kron_flag` as input, as `get_perturbation_params_derivs` looks into `options_.analytic_derivation_mode` for `kron_flag`.
## `dynare_identification.m`
* make sure that setting `analytic_derivation_mode` is set both in `options_ident` and `options_`. Note that at the end of the function we restore the `options_` structure, so all changes are local. In a next merge request, I will remove the global variables to make all variables local.
## `get_identification_jacobians.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `A` and `B` matrix anymore, but instead of the dynare solution matrix `ghx` and `ghu`. So we recreate these matrices here instead of in `get_perturbation_params_derivs.m`.
* Added `str2func` for better function handles in `fjaco.m`.
## `fjaco.m`
* make `tol`an option, which can be adjusted by changing `options_.dynatol.x`for identification and parameter derivatives purposes.
* include a check and an informative error message, if numerical derivatives (two-sided finite difference method) yield errors in `resol.m` for identification and parameter derivatives purposes. This closes issue Dynare/dynare#1595.
* Changed year of copyright to 2010-2017,2019
***
# Further suggestions and questions
* Ones this is merged, I will merge request an improvement of the identification toolbox, which will work up to third order using the pruned state space. This will also remove some issues and bugs, and also I will remove global variables in this request.
* The third-order derivatives can be further improved by taking sparsity into account and use mex versions for kronecker products etc. I leave this for further testing (and if anybody actually uses this ;-) )
2019-12-17 19:17:09 +01:00
P_x_xx = I_xxx ( : , ipermute ( id_xxx , [ 1 , 3 , 4 , 2 ] ) ) + I_xxx ( : , ipermute ( id_xxx , [ 1 , 2 , 4 , 3 ] ) ) + I_xxx ( : , ipermute ( id_xxx , [ 1 , 2 , 3 , 4 ] ) ) ;
P_x_xu = I_xxu ( : , ipermute ( id_uxx , [ 1 , 2 , 3 , 4 ] ) ) + I_xxu ( : , ipermute ( id_uxx , [ 1 , 2 , 4 , 3 ] ) ) ;
P_xu_u = I_xuu ( : , ipermute ( id_uux , [ 1 , 2 , 3 , 4 ] ) ) + I_xuu ( : , ipermute ( id_uux , [ 1 , 3 , 2 , 4 ] ) ) ;
P_u_uu = I_uuu ( : , ipermute ( id_uuu , [ 1 , 3 , 4 , 2 ] ) ) + I_uuu ( : , ipermute ( id_uuu , [ 1 , 2 , 4 , 3 ] ) ) + I_uuu ( : , ipermute ( id_uuu , [ 1 , 2 , 3 , 4 ] ) ) ;
P_uu_u = I_uuu ( : , ipermute ( id_uuu , [ 1 , 2 , 3 , 4 ] ) ) + I_uuu ( : , ipermute ( id_uuu , [ 1 , 3 , 4 , 2 ] ) ) ;
2019-12-20 16:28:06 +01:00
Added parameter derivatives of perturbation solution up to 3 order
# Preliminary comments
I finished the identification toolbox at orders two and three using the pruned state space system, but before I merge request this, I decided to first merge the new functionality to compute parameter derivatives of perturbation solution matrices at higher orders. So after this is approved, I merge the identification toolbox.
I guess @rattoma, @sebastien, and @michel are best choices to review this.
I outline the main idea first and then provide some more detailed changes I made to the functions.
***
# Main idea
This merge request is concerned with the *analytical*computation of the parameter derivatives of first, second and third order perturbation solution matrices, i.e. using _closed-form_ expressions to efficiently compute the derivative of $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ *with respect to model parameters* $\theta$. Note that $\theta$ contains model parameters, stderr and corr parameters of shocks. stderr and corr parameters of measurement errors are not yet supported, (they can easily be included as exogenous shocks). The availability of such derivatives is beneficial in terms of more reliable analysis of model sensitivity and parameter identifiability as well as more efficient estimation methods, in particular for models solved up to third order, as it is well-known that numerical derivatives are a tricky business, especially for large models.
References for my approach are:
* Iskrev (2008, 2010) and Schmitt-Grohé and Uribe (2012, Appendix) who were the first to compute the parameter derivatives analytically at first order, however, using inefficient (sparse) Kronecker products.
* Mutschler (2015) who provides the expressions for a second-order, but again using inefficient (sparse) Kronecker products.
* Ratto and Iskrev (2012) who show how the first-order system can be solved accurately, fast and efficiently using existing numerical algorithms for generalized Sylvester equations by taking the parameter derivative with respect to each parameter separately.
* Julliard and Kamenik (2004) who provide the perturbation solution equation system in tensor notation at any order k.
* Levintal (2017) who introduces permutation matrices to express the perturbation solution equation system in matrix notation up to fifth order.
Note that @rattoma already implemented the parameter derivatives of $g_x$ and $g_u$ analytically (and numerically), and I rely heavily on his work in `get_first_order_solution_params_derivs.m` (previously `getH.m`). My additions are mainly to this function and thus it is renamed to `get_perturbation_params_derivs.m`.
The basic idea of this merge request is to take the second- and third-order perturbation solution systems in Julliard and Kamenik (2004), unfold these into an equivalent matrix representation using permutation matrices as in Levintal (2017). Then extending Ratto and Iskrev (2012) one takes the derivative with respect to each parameter separately and gets a computational problem that is linear, albeit large, as it involves either solving generalized Sylvester equations or taking inverses of highly sparse matrices. I will now briefly summarize the perturbation solution system at third order and the system that results when taking the derivative with respect to parameters.
## Perturbation Solution
The following systems arise at first, second, and third order:
$(ghx): f_{x} z_{x} = f_{y_{-}^*} + f_{y_0} g_{x} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{x}= A g_{x} + f_{y_{-}^*}=0$
$(ghu): f_{z} z_{u} = f_{y_0} g_{u} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{u} + f_{u}= A g_u + f_u = 0$
$(ghxx) : A g_{xx} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{x}\right) + f_{zz} \left( z_{x} \otimes z_{x} \right) = 0$
$(ghxu) : A g_{xu} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{x} \otimes z_{u} \right) = 0$
$(ghuu) : A g_{uu} + B g_{xx} \left(g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{u} \otimes z_{u} \right) = 0$
$(ghs2) : (A+B) g_{\sigma\sigma} + \left( f_{y^{**}_{+}y^{**}_{+}} \left(g^{**}_{u} \otimes g^{**}_{u}\right) + f_{y^{**}_{+}} g^{**}_{uu}\right)vec(\Sigma) = 0$
$(ghxxx) : A g_{xxx} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{x}\right) + f_{y_{+}}g^{**}_{xx} \left(g^{*}_x \otimes g^{*}_{xx}\right)P_{x\_xx} + f_{zz} \left( z_{x} \otimes z_{xx} \right)P_{x\_xx} + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{x} \right) = 0$
$(ghxxu) : A g_{xxu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{u} \right) + f_{zz} \left( \left( z_{x} \otimes z_{xu} \right)P_{x\_xu} + \left(z_{xx} \otimes z_{u}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{x} \otimes g^{*}_{xu}\right)P_{x\_xu} + \left(g^{*}_{xx} \otimes g^{*}_{u}\right) \right) = 0$
$(ghxuu) : A g_{xuu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( \left( z_{xu} \otimes z_{u} \right)P_{xu\_u} + \left(z_{x} \otimes z_{uu}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{xu} \otimes g^{*}_{u}\right)P_{xu\_u} + \left(g^{*}_{x} \otimes g^{*}_{uu}\right) \right) = 0$
$(ghuuu) : A g_{uuu} + B g_{xxx} \left(g^{*}_{u} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{u} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( z_{u} \otimes z_{uu} \right)P_{u\_uu} + f_{y_{+}}g^{**}_{xx} \left(g^{*}_{u} \otimes g^{*}_{uu}\right)P_{u\_uu} = 0$
$(ghx\sigma\sigma) : A g_{x\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_x + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{x} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{x} \otimes z_{\sigma\sigma}\right) + F_{xu_{+}u_{+}}\left(I_{n_x} \otimes vec(\Sigma)\right) = 0$
$F_{xu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_x^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{xu_{+}} \otimes z_{u_{+}} \right)P_{xu\_u} + \left(z_{x} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{x} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
$(ghu\sigma\sigma) : A g_{u\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_{u} + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{u} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{u} \otimes z_{\sigma\sigma}\right) + F_{uu_{+}u_{+}}\left(I_{n_u} \otimes vec(\Sigma_u)\right) = 0$
$F_{uu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_u^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{uu_{+}} \otimes z_{u_{+}} \right)P_{uu\_u} + \left(z_{u} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{u} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
A and B are the common perturbation matrices:
$A = f_{y_0} + \begin{pmatrix} \underbrace{0}_{n\times n_{static}} &\vdots& \underbrace{f_{y^{**}_{+}} \cdot g^{**}_{x}}_{n \times n_{spred}} &\vdots& \underbrace{0}_{n\times n_{frwd}} \end{pmatrix}$and $B = \begin{pmatrix} \underbrace{0}_{n \times n_{static}}&\vdots & \underbrace{0}_{n \times n_{pred}} & \vdots & \underbrace{f_{y^{**}_{+}}}_{n \times n_{sfwrd}} \end{pmatrix}$
and $z=(y_{-}^{\ast}; y; y_{+}^{\ast\ast}; u)$ denotes the dynamic model variables as in `M_.lead_lag_incidence`, $y^\ast$ denote state variables, $y^{\ast\ast}$ denote forward looking variables, $y_+$ denote the variables with a lead, $y_{-}$ denote variables with a lag, $y_0$ denote variables at period t, $f$ the model equations, and $f_z$ the first-order dynamic model derivatives, $f_{zz}$ the second-order dynamic derivatives, and $f_{zzz}$ the third-order dynamic model derivatives. Then:
$z_{x} = \begin{pmatrix}I\\g_{x}\\g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{u} =\begin{pmatrix}0\\g_{u}\\g^{**}_{x} \cdot g^{*}_{u}\\I\end{pmatrix}$, $z_{u_{+}} =\begin{pmatrix}0\\0\\g^{**}_{u}\\0\end{pmatrix}$
$z_{xx} = \begin{pmatrix} 0\\g_{xx}\\g^{**}_{x} \left( g^{*}_x \otimes g^{*}_{x} \right) + g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{xu} =\begin{pmatrix}0\\g_{xu}\\g^{**}_{xx} \left( g^{*}_x \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{xu}\\0\end{pmatrix}$, $z_{uu} =\begin{pmatrix}0\\g_{uu}\\g^{**}_{xx} \left( g^{*}_u \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{uu}\\0\end{pmatrix}$,
$z_{xu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_x \otimes I \right)\\0\end{pmatrix}$, $z_{uu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_{u} \otimes I \right)\\0\end{pmatrix}$, $z_{u_{+}u_{+}} =\begin{pmatrix}0\\0\\g^{\ast\ast}_{uu}\\0\end{pmatrix}$, $z_{\sigma\sigma} = \begin{pmatrix}0\\ g_{\sigma\sigma}\\ g^{\ast\ast}_{x}g^{\ast}_{\sigma\sigma} + g^{\ast\ast}_{\sigma\sigma}\\0 \end{pmatrix}$
$P$ are permutation matrices that can be computed using Matlab's `ipermute` function.
## Parameter derivatives of perturbation solutions
First, we need the parameter derivatives of first, second, third, and fourth derivatives of the dynamic model (i.e. g1,g2,g3,g4 in dynamic files), I make use of the implicit function theorem: Let $f_{z^k}$ denote the kth derivative (wrt all dynamic variables) of the dynamic model, then let $df_{z^k}$ denote the first-derivative (wrt all model parameters) of $f_{z^k}$ evaluated at the steady state. Note that $f_{z^k}$ is a function of both the model parameters $\theta$ and of the steady state of all dynamic variables $\bar{z}$, which also depend on the parameters. Hence, implicitly $f_{z^k}=f_{z^k}(\theta,\bar{z}(\theta))$ and $df_{z^k}$ consists of two parts:
1. direct derivative wrt to all model parameters given by the preprocessor in the `_params_derivs.m` files
2. contribution of derivative of steady state of dynamic variables (wrt all model parameters): $f_{z^{k+1}} \cdot d\bar{z}$
Note that we already have functionality to compute $d\bar{z}$ analytically.
Having this, the above perturbation systems are basically equations of the following types
$AX +BXC = RHS$ or $AX = RHS$
Now when taking the derivative (wrt to one single parameter $\theta_j$), we get
$A\mathrm{d}\{X\} + B\mathrm{d}\{X\}C = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X - \mathrm{d}\{B\}XC - BX\mathrm{d}\{C\}$
or
$A\mathrm{d}\{X\} = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X$
The first one is a Sylvester type equation, the second one can be solved by taking the inverse of $A$. The only diffculty and tedious work arrises in computing (the highly sparse) derivatives of $RHS$.
***
# New functions: `
## get_perturbation_params_derivs.m`and `get_perturbation_params_derivs_numerical_objective.m`
* The parameter derivatives up to third order are computed in the new function`get_perturbation_params_derivs.m` both analytically and numerically. For numerical derivatives `get_perturbation_params_derivs_numerical_objective.m` is the objective for `fjaco.m` or `hessian_sparse.m` or `hessian.m`.
* `get_perturbation_params_derivs.m` is basically an extended version of the previous `get_first_order_solution_params_derivs.m` function.
* * `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It is used for numerical derivatives, whenever `analytic_derivation_mode=-1|-2`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
* Detailed changes:
* Most important: notation of this function is now in accordance to the k_order_solver, i.e. we do not compute derivatives of Kalman transition matrices A and B, but rather the solution matrices ghx,ghu,ghxx,ghxu,ghuu,ghs2,ghxxx,ghxxu,ghxuu,ghuuu,ghxss,ghuss in accordance with notation used in `oo_.dr`. As a byproduct at first-order, focusing on ghx and ghu instead of Kalman transition matrices A and B makes the computations slightly faster for large models (e.g. for Quest the computations were faster by a couple of seconds, not much, but okay).
* Removed use of `kstate`, see also Dynare/dynare#1653 and Dynare/dynare!1656
* Output arguments are stored in a structure `DERIVS`, there is also a flag `d2flag` that computes parameter hessians needed only in `dsge_likelihood.m`.
* Removed `kronflag` as input. `options_.analytic_derivation_mode` is now used instead of `kronflag`.
* Removed `indvar`, an index that was used to selected specific variables in the derivatives. This is not needed, as we always compute the parameter derivatives for all variables first and then select a subset of variables. The selection now takes place in other functions, like `dsge_likelihood.m`.
* Introduced some checks: (i) deterministic exogenous variables are not supported, (ii) Kronecker method only compatible with first-order approximation so reset to sylvester method, (iii) for purely backward or forward models we need to be careful with the rows in `M_.lead_la g_incidence`, (iv) if `_params_derivs.m` files are missing an error is thrown.
* For numerical derivatives, if mod file does not contain an `estimated_params_block`, a temporary one with the most important parameter information is created.
## `unfold_g4.m`
* When evaluating g3 and g4 one needs to take into account that these do not contain symmetric elements, so one needs to use `unfold_g3.m` and the new function `unfold_g4.m`. This returns an unfolded version of the same matrix (i.e. with symmetric elements).
***
# New test models
`.gitignore` and `Makefile.am` are changed accordingly. Also now it is possible to run test suite on analytic_derivatives, i.e. run `make check m/analytic_derivatives`
## `analytic_derivatives/BrockMirman_PertParamsDerivs.mod`
* This is the Brock Mirman model, where we know the exact policy function $g$ for capital and consumption. As this does not imply a nonzero $g_{\sigma\sigma}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ I added some artificial equations to get nonzero solution matrices with respect to $\sigma$. The true perturbation solution matrices $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ are then computed analytically with Matlab's symbolic toolbox and saved in `nBrockMirmanSYM.mat`. There is a preprocessor flag that recreates these analytical computations if changes are needed (and to check whether I made some errors here ;-) )
* Then solution matrices up to third order and their parameter Jacobians are then compared to the ones computed by Dynare's `k_order_solver` and by `get_perturbation_params_derivs` for all `analytic_derivation_mode`'s. There will be an error if the maximum absolute deviation is too large, i.e. for numerical derivatives (`analytic_derivation_mode=-1|-2`) the tolerance is choosen lower (around 1e-5); for analytical methods we are stricter: around 1e-13 for first-order, 1e-12 for second order, and 1e-11 for third-order.
* As a side note, this mod file also checks Dynare's `k_order_solver` algorithm and throws an error if something is wrong.
* This test model shows that the new functionality works well. And analytical derivatives perform way better and accurate than numerical ones, even for this small model.
## `analytic_derivatives/burnside_3_order_PertParamsDerivs.mod`
* This builds upon `tests/k_order_perturbation/burnside_k_order.mod` and computes the true parameter derivatives analytically by hand.
* This test model also shows that the new functionality works well.
## `analytic_derivatives/LindeTrabandt2019.mod`
* Shows that the new functionality also works for medium-sized models, i.e. a SW type model solved at third order with 35 variables (11 states). 2 shocks and 20 parameters.
* This mod file can be used to tweak the speed of the computations in the future.
* Compares numerical versus analytical parameter derivatives (for first, second and third order). Note that this model clearly shows that numerical ones are quite different than analytical ones even at first order!
## `identification/LindeTrabandt2019_xfail.mod`
* This model is a check for issue Dynare/dynare#1595, see fjaco.m below, and will fail.
* Removed `analytic_derivatives/ls2003.mod` as this mod file is neither in the testsuite nor does it work.
***
# Detailed changes in other functions
## `get_first_order_solution_params_derivs.m`
* Deleted, or actually, renamed to `get_perturbation_params_derivs.m`, as this function now is able to compute the derivatives up to third order
## `identification_numerical_objective.m`
* `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
## `dsge_likelihood.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `T`matrix anymore, but instead of the dynare solution matrix `ghx`. So we recreate `T` here as this amounts to adding some zeros and focusing on selected variables only.
* Added some checks to make sure the first-order approximation is selected.
* Removed `kron_flag` as input, as `get_perturbation_params_derivs` looks into `options_.analytic_derivation_mode` for `kron_flag`.
## `dynare_identification.m`
* make sure that setting `analytic_derivation_mode` is set both in `options_ident` and `options_`. Note that at the end of the function we restore the `options_` structure, so all changes are local. In a next merge request, I will remove the global variables to make all variables local.
## `get_identification_jacobians.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `A` and `B` matrix anymore, but instead of the dynare solution matrix `ghx` and `ghu`. So we recreate these matrices here instead of in `get_perturbation_params_derivs.m`.
* Added `str2func` for better function handles in `fjaco.m`.
## `fjaco.m`
* make `tol`an option, which can be adjusted by changing `options_.dynatol.x`for identification and parameter derivatives purposes.
* include a check and an informative error message, if numerical derivatives (two-sided finite difference method) yield errors in `resol.m` for identification and parameter derivatives purposes. This closes issue Dynare/dynare#1595.
* Changed year of copyright to 2010-2017,2019
***
# Further suggestions and questions
* Ones this is merged, I will merge request an improvement of the identification toolbox, which will work up to third order using the pruned state space. This will also remove some issues and bugs, and also I will remove global variables in this request.
* The third-order derivatives can be further improved by taking sparsity into account and use mex versions for kronecker products etc. I leave this for further testing (and if anybody actually uses this ;-) )
2019-12-17 19:17:09 +01:00
zxx = [ spalloc ( M . nspred , M . nspred ^2 , 0 ) ;
oo . dr . ghxx ( kcurr ~= 0 , : ) ;
oo . dr . ghxx ( klead ~= 0 , : ) * G_x_x + oo . dr . ghx ( klead ~= 0 , : ) * oo . dr . ghxx ( idx_states , : ) ;
spalloc ( M . exo_nbr , M . nspred ^2 , 0 ) ] ;
G_x_x_x = kron ( oo . dr . ghx ( idx_states , : ) , G_x_x ) ;
G_x_xx = kron ( oo . dr . ghx ( idx_states , : ) , oo . dr . ghxx ( idx_states , : ) ) ;
Z_x_x = kron ( zx , zx ) ;
Z_x_x_x = kron ( zx , Z_x_x ) ;
Z_x_xx = kron ( zx , zxx ) ;
fyp_ghxxp = sparse ( g1 ( : , nonzeros ( klead ) ) * oo . dr . ghxx ( klead ~= 0 , : ) ) ;
2019-12-20 16:28:06 +01:00
B_ghxxx = B * oo . dr . ghxxx ;
Added parameter derivatives of perturbation solution up to 3 order
# Preliminary comments
I finished the identification toolbox at orders two and three using the pruned state space system, but before I merge request this, I decided to first merge the new functionality to compute parameter derivatives of perturbation solution matrices at higher orders. So after this is approved, I merge the identification toolbox.
I guess @rattoma, @sebastien, and @michel are best choices to review this.
I outline the main idea first and then provide some more detailed changes I made to the functions.
***
# Main idea
This merge request is concerned with the *analytical*computation of the parameter derivatives of first, second and third order perturbation solution matrices, i.e. using _closed-form_ expressions to efficiently compute the derivative of $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ *with respect to model parameters* $\theta$. Note that $\theta$ contains model parameters, stderr and corr parameters of shocks. stderr and corr parameters of measurement errors are not yet supported, (they can easily be included as exogenous shocks). The availability of such derivatives is beneficial in terms of more reliable analysis of model sensitivity and parameter identifiability as well as more efficient estimation methods, in particular for models solved up to third order, as it is well-known that numerical derivatives are a tricky business, especially for large models.
References for my approach are:
* Iskrev (2008, 2010) and Schmitt-Grohé and Uribe (2012, Appendix) who were the first to compute the parameter derivatives analytically at first order, however, using inefficient (sparse) Kronecker products.
* Mutschler (2015) who provides the expressions for a second-order, but again using inefficient (sparse) Kronecker products.
* Ratto and Iskrev (2012) who show how the first-order system can be solved accurately, fast and efficiently using existing numerical algorithms for generalized Sylvester equations by taking the parameter derivative with respect to each parameter separately.
* Julliard and Kamenik (2004) who provide the perturbation solution equation system in tensor notation at any order k.
* Levintal (2017) who introduces permutation matrices to express the perturbation solution equation system in matrix notation up to fifth order.
Note that @rattoma already implemented the parameter derivatives of $g_x$ and $g_u$ analytically (and numerically), and I rely heavily on his work in `get_first_order_solution_params_derivs.m` (previously `getH.m`). My additions are mainly to this function and thus it is renamed to `get_perturbation_params_derivs.m`.
The basic idea of this merge request is to take the second- and third-order perturbation solution systems in Julliard and Kamenik (2004), unfold these into an equivalent matrix representation using permutation matrices as in Levintal (2017). Then extending Ratto and Iskrev (2012) one takes the derivative with respect to each parameter separately and gets a computational problem that is linear, albeit large, as it involves either solving generalized Sylvester equations or taking inverses of highly sparse matrices. I will now briefly summarize the perturbation solution system at third order and the system that results when taking the derivative with respect to parameters.
## Perturbation Solution
The following systems arise at first, second, and third order:
$(ghx): f_{x} z_{x} = f_{y_{-}^*} + f_{y_0} g_{x} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{x}= A g_{x} + f_{y_{-}^*}=0$
$(ghu): f_{z} z_{u} = f_{y_0} g_{u} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{u} + f_{u}= A g_u + f_u = 0$
$(ghxx) : A g_{xx} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{x}\right) + f_{zz} \left( z_{x} \otimes z_{x} \right) = 0$
$(ghxu) : A g_{xu} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{x} \otimes z_{u} \right) = 0$
$(ghuu) : A g_{uu} + B g_{xx} \left(g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{u} \otimes z_{u} \right) = 0$
$(ghs2) : (A+B) g_{\sigma\sigma} + \left( f_{y^{**}_{+}y^{**}_{+}} \left(g^{**}_{u} \otimes g^{**}_{u}\right) + f_{y^{**}_{+}} g^{**}_{uu}\right)vec(\Sigma) = 0$
$(ghxxx) : A g_{xxx} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{x}\right) + f_{y_{+}}g^{**}_{xx} \left(g^{*}_x \otimes g^{*}_{xx}\right)P_{x\_xx} + f_{zz} \left( z_{x} \otimes z_{xx} \right)P_{x\_xx} + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{x} \right) = 0$
$(ghxxu) : A g_{xxu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{u} \right) + f_{zz} \left( \left( z_{x} \otimes z_{xu} \right)P_{x\_xu} + \left(z_{xx} \otimes z_{u}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{x} \otimes g^{*}_{xu}\right)P_{x\_xu} + \left(g^{*}_{xx} \otimes g^{*}_{u}\right) \right) = 0$
$(ghxuu) : A g_{xuu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( \left( z_{xu} \otimes z_{u} \right)P_{xu\_u} + \left(z_{x} \otimes z_{uu}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{xu} \otimes g^{*}_{u}\right)P_{xu\_u} + \left(g^{*}_{x} \otimes g^{*}_{uu}\right) \right) = 0$
$(ghuuu) : A g_{uuu} + B g_{xxx} \left(g^{*}_{u} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{u} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( z_{u} \otimes z_{uu} \right)P_{u\_uu} + f_{y_{+}}g^{**}_{xx} \left(g^{*}_{u} \otimes g^{*}_{uu}\right)P_{u\_uu} = 0$
$(ghx\sigma\sigma) : A g_{x\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_x + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{x} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{x} \otimes z_{\sigma\sigma}\right) + F_{xu_{+}u_{+}}\left(I_{n_x} \otimes vec(\Sigma)\right) = 0$
$F_{xu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_x^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{xu_{+}} \otimes z_{u_{+}} \right)P_{xu\_u} + \left(z_{x} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{x} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
$(ghu\sigma\sigma) : A g_{u\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_{u} + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{u} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{u} \otimes z_{\sigma\sigma}\right) + F_{uu_{+}u_{+}}\left(I_{n_u} \otimes vec(\Sigma_u)\right) = 0$
$F_{uu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_u^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{uu_{+}} \otimes z_{u_{+}} \right)P_{uu\_u} + \left(z_{u} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{u} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
A and B are the common perturbation matrices:
$A = f_{y_0} + \begin{pmatrix} \underbrace{0}_{n\times n_{static}} &\vdots& \underbrace{f_{y^{**}_{+}} \cdot g^{**}_{x}}_{n \times n_{spred}} &\vdots& \underbrace{0}_{n\times n_{frwd}} \end{pmatrix}$and $B = \begin{pmatrix} \underbrace{0}_{n \times n_{static}}&\vdots & \underbrace{0}_{n \times n_{pred}} & \vdots & \underbrace{f_{y^{**}_{+}}}_{n \times n_{sfwrd}} \end{pmatrix}$
and $z=(y_{-}^{\ast}; y; y_{+}^{\ast\ast}; u)$ denotes the dynamic model variables as in `M_.lead_lag_incidence`, $y^\ast$ denote state variables, $y^{\ast\ast}$ denote forward looking variables, $y_+$ denote the variables with a lead, $y_{-}$ denote variables with a lag, $y_0$ denote variables at period t, $f$ the model equations, and $f_z$ the first-order dynamic model derivatives, $f_{zz}$ the second-order dynamic derivatives, and $f_{zzz}$ the third-order dynamic model derivatives. Then:
$z_{x} = \begin{pmatrix}I\\g_{x}\\g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{u} =\begin{pmatrix}0\\g_{u}\\g^{**}_{x} \cdot g^{*}_{u}\\I\end{pmatrix}$, $z_{u_{+}} =\begin{pmatrix}0\\0\\g^{**}_{u}\\0\end{pmatrix}$
$z_{xx} = \begin{pmatrix} 0\\g_{xx}\\g^{**}_{x} \left( g^{*}_x \otimes g^{*}_{x} \right) + g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{xu} =\begin{pmatrix}0\\g_{xu}\\g^{**}_{xx} \left( g^{*}_x \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{xu}\\0\end{pmatrix}$, $z_{uu} =\begin{pmatrix}0\\g_{uu}\\g^{**}_{xx} \left( g^{*}_u \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{uu}\\0\end{pmatrix}$,
$z_{xu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_x \otimes I \right)\\0\end{pmatrix}$, $z_{uu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_{u} \otimes I \right)\\0\end{pmatrix}$, $z_{u_{+}u_{+}} =\begin{pmatrix}0\\0\\g^{\ast\ast}_{uu}\\0\end{pmatrix}$, $z_{\sigma\sigma} = \begin{pmatrix}0\\ g_{\sigma\sigma}\\ g^{\ast\ast}_{x}g^{\ast}_{\sigma\sigma} + g^{\ast\ast}_{\sigma\sigma}\\0 \end{pmatrix}$
$P$ are permutation matrices that can be computed using Matlab's `ipermute` function.
## Parameter derivatives of perturbation solutions
First, we need the parameter derivatives of first, second, third, and fourth derivatives of the dynamic model (i.e. g1,g2,g3,g4 in dynamic files), I make use of the implicit function theorem: Let $f_{z^k}$ denote the kth derivative (wrt all dynamic variables) of the dynamic model, then let $df_{z^k}$ denote the first-derivative (wrt all model parameters) of $f_{z^k}$ evaluated at the steady state. Note that $f_{z^k}$ is a function of both the model parameters $\theta$ and of the steady state of all dynamic variables $\bar{z}$, which also depend on the parameters. Hence, implicitly $f_{z^k}=f_{z^k}(\theta,\bar{z}(\theta))$ and $df_{z^k}$ consists of two parts:
1. direct derivative wrt to all model parameters given by the preprocessor in the `_params_derivs.m` files
2. contribution of derivative of steady state of dynamic variables (wrt all model parameters): $f_{z^{k+1}} \cdot d\bar{z}$
Note that we already have functionality to compute $d\bar{z}$ analytically.
Having this, the above perturbation systems are basically equations of the following types
$AX +BXC = RHS$ or $AX = RHS$
Now when taking the derivative (wrt to one single parameter $\theta_j$), we get
$A\mathrm{d}\{X\} + B\mathrm{d}\{X\}C = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X - \mathrm{d}\{B\}XC - BX\mathrm{d}\{C\}$
or
$A\mathrm{d}\{X\} = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X$
The first one is a Sylvester type equation, the second one can be solved by taking the inverse of $A$. The only diffculty and tedious work arrises in computing (the highly sparse) derivatives of $RHS$.
***
# New functions: `
## get_perturbation_params_derivs.m`and `get_perturbation_params_derivs_numerical_objective.m`
* The parameter derivatives up to third order are computed in the new function`get_perturbation_params_derivs.m` both analytically and numerically. For numerical derivatives `get_perturbation_params_derivs_numerical_objective.m` is the objective for `fjaco.m` or `hessian_sparse.m` or `hessian.m`.
* `get_perturbation_params_derivs.m` is basically an extended version of the previous `get_first_order_solution_params_derivs.m` function.
* * `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It is used for numerical derivatives, whenever `analytic_derivation_mode=-1|-2`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
* Detailed changes:
* Most important: notation of this function is now in accordance to the k_order_solver, i.e. we do not compute derivatives of Kalman transition matrices A and B, but rather the solution matrices ghx,ghu,ghxx,ghxu,ghuu,ghs2,ghxxx,ghxxu,ghxuu,ghuuu,ghxss,ghuss in accordance with notation used in `oo_.dr`. As a byproduct at first-order, focusing on ghx and ghu instead of Kalman transition matrices A and B makes the computations slightly faster for large models (e.g. for Quest the computations were faster by a couple of seconds, not much, but okay).
* Removed use of `kstate`, see also Dynare/dynare#1653 and Dynare/dynare!1656
* Output arguments are stored in a structure `DERIVS`, there is also a flag `d2flag` that computes parameter hessians needed only in `dsge_likelihood.m`.
* Removed `kronflag` as input. `options_.analytic_derivation_mode` is now used instead of `kronflag`.
* Removed `indvar`, an index that was used to selected specific variables in the derivatives. This is not needed, as we always compute the parameter derivatives for all variables first and then select a subset of variables. The selection now takes place in other functions, like `dsge_likelihood.m`.
* Introduced some checks: (i) deterministic exogenous variables are not supported, (ii) Kronecker method only compatible with first-order approximation so reset to sylvester method, (iii) for purely backward or forward models we need to be careful with the rows in `M_.lead_la g_incidence`, (iv) if `_params_derivs.m` files are missing an error is thrown.
* For numerical derivatives, if mod file does not contain an `estimated_params_block`, a temporary one with the most important parameter information is created.
## `unfold_g4.m`
* When evaluating g3 and g4 one needs to take into account that these do not contain symmetric elements, so one needs to use `unfold_g3.m` and the new function `unfold_g4.m`. This returns an unfolded version of the same matrix (i.e. with symmetric elements).
***
# New test models
`.gitignore` and `Makefile.am` are changed accordingly. Also now it is possible to run test suite on analytic_derivatives, i.e. run `make check m/analytic_derivatives`
## `analytic_derivatives/BrockMirman_PertParamsDerivs.mod`
* This is the Brock Mirman model, where we know the exact policy function $g$ for capital and consumption. As this does not imply a nonzero $g_{\sigma\sigma}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ I added some artificial equations to get nonzero solution matrices with respect to $\sigma$. The true perturbation solution matrices $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ are then computed analytically with Matlab's symbolic toolbox and saved in `nBrockMirmanSYM.mat`. There is a preprocessor flag that recreates these analytical computations if changes are needed (and to check whether I made some errors here ;-) )
* Then solution matrices up to third order and their parameter Jacobians are then compared to the ones computed by Dynare's `k_order_solver` and by `get_perturbation_params_derivs` for all `analytic_derivation_mode`'s. There will be an error if the maximum absolute deviation is too large, i.e. for numerical derivatives (`analytic_derivation_mode=-1|-2`) the tolerance is choosen lower (around 1e-5); for analytical methods we are stricter: around 1e-13 for first-order, 1e-12 for second order, and 1e-11 for third-order.
* As a side note, this mod file also checks Dynare's `k_order_solver` algorithm and throws an error if something is wrong.
* This test model shows that the new functionality works well. And analytical derivatives perform way better and accurate than numerical ones, even for this small model.
## `analytic_derivatives/burnside_3_order_PertParamsDerivs.mod`
* This builds upon `tests/k_order_perturbation/burnside_k_order.mod` and computes the true parameter derivatives analytically by hand.
* This test model also shows that the new functionality works well.
## `analytic_derivatives/LindeTrabandt2019.mod`
* Shows that the new functionality also works for medium-sized models, i.e. a SW type model solved at third order with 35 variables (11 states). 2 shocks and 20 parameters.
* This mod file can be used to tweak the speed of the computations in the future.
* Compares numerical versus analytical parameter derivatives (for first, second and third order). Note that this model clearly shows that numerical ones are quite different than analytical ones even at first order!
## `identification/LindeTrabandt2019_xfail.mod`
* This model is a check for issue Dynare/dynare#1595, see fjaco.m below, and will fail.
* Removed `analytic_derivatives/ls2003.mod` as this mod file is neither in the testsuite nor does it work.
***
# Detailed changes in other functions
## `get_first_order_solution_params_derivs.m`
* Deleted, or actually, renamed to `get_perturbation_params_derivs.m`, as this function now is able to compute the derivatives up to third order
## `identification_numerical_objective.m`
* `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
## `dsge_likelihood.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `T`matrix anymore, but instead of the dynare solution matrix `ghx`. So we recreate `T` here as this amounts to adding some zeros and focusing on selected variables only.
* Added some checks to make sure the first-order approximation is selected.
* Removed `kron_flag` as input, as `get_perturbation_params_derivs` looks into `options_.analytic_derivation_mode` for `kron_flag`.
## `dynare_identification.m`
* make sure that setting `analytic_derivation_mode` is set both in `options_ident` and `options_`. Note that at the end of the function we restore the `options_` structure, so all changes are local. In a next merge request, I will remove the global variables to make all variables local.
## `get_identification_jacobians.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `A` and `B` matrix anymore, but instead of the dynare solution matrix `ghx` and `ghu`. So we recreate these matrices here instead of in `get_perturbation_params_derivs.m`.
* Added `str2func` for better function handles in `fjaco.m`.
## `fjaco.m`
* make `tol`an option, which can be adjusted by changing `options_.dynatol.x`for identification and parameter derivatives purposes.
* include a check and an informative error message, if numerical derivatives (two-sided finite difference method) yield errors in `resol.m` for identification and parameter derivatives purposes. This closes issue Dynare/dynare#1595.
* Changed year of copyright to 2010-2017,2019
***
# Further suggestions and questions
* Ones this is merged, I will merge request an improvement of the identification toolbox, which will work up to third order using the pruned state space. This will also remove some issues and bugs, and also I will remove global variables in this request.
* The third-order derivatives can be further improved by taking sparsity into account and use mex versions for kronecker products etc. I leave this for further testing (and if anybody actually uses this ;-) )
2019-12-17 19:17:09 +01:00
dzxx = zeros ( size ( zxx , 1 ) , size ( zxx , 2 ) , modparam_nbr ) ;
dfyp_ghxxp = zeros ( size ( fyp_ghxxp , 1 ) , size ( fyp_ghxxp , 2 ) , modparam_nbr ) ;
2019-12-20 16:28:06 +01:00
Added parameter derivatives of perturbation solution up to 3 order
# Preliminary comments
I finished the identification toolbox at orders two and three using the pruned state space system, but before I merge request this, I decided to first merge the new functionality to compute parameter derivatives of perturbation solution matrices at higher orders. So after this is approved, I merge the identification toolbox.
I guess @rattoma, @sebastien, and @michel are best choices to review this.
I outline the main idea first and then provide some more detailed changes I made to the functions.
***
# Main idea
This merge request is concerned with the *analytical*computation of the parameter derivatives of first, second and third order perturbation solution matrices, i.e. using _closed-form_ expressions to efficiently compute the derivative of $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ *with respect to model parameters* $\theta$. Note that $\theta$ contains model parameters, stderr and corr parameters of shocks. stderr and corr parameters of measurement errors are not yet supported, (they can easily be included as exogenous shocks). The availability of such derivatives is beneficial in terms of more reliable analysis of model sensitivity and parameter identifiability as well as more efficient estimation methods, in particular for models solved up to third order, as it is well-known that numerical derivatives are a tricky business, especially for large models.
References for my approach are:
* Iskrev (2008, 2010) and Schmitt-Grohé and Uribe (2012, Appendix) who were the first to compute the parameter derivatives analytically at first order, however, using inefficient (sparse) Kronecker products.
* Mutschler (2015) who provides the expressions for a second-order, but again using inefficient (sparse) Kronecker products.
* Ratto and Iskrev (2012) who show how the first-order system can be solved accurately, fast and efficiently using existing numerical algorithms for generalized Sylvester equations by taking the parameter derivative with respect to each parameter separately.
* Julliard and Kamenik (2004) who provide the perturbation solution equation system in tensor notation at any order k.
* Levintal (2017) who introduces permutation matrices to express the perturbation solution equation system in matrix notation up to fifth order.
Note that @rattoma already implemented the parameter derivatives of $g_x$ and $g_u$ analytically (and numerically), and I rely heavily on his work in `get_first_order_solution_params_derivs.m` (previously `getH.m`). My additions are mainly to this function and thus it is renamed to `get_perturbation_params_derivs.m`.
The basic idea of this merge request is to take the second- and third-order perturbation solution systems in Julliard and Kamenik (2004), unfold these into an equivalent matrix representation using permutation matrices as in Levintal (2017). Then extending Ratto and Iskrev (2012) one takes the derivative with respect to each parameter separately and gets a computational problem that is linear, albeit large, as it involves either solving generalized Sylvester equations or taking inverses of highly sparse matrices. I will now briefly summarize the perturbation solution system at third order and the system that results when taking the derivative with respect to parameters.
## Perturbation Solution
The following systems arise at first, second, and third order:
$(ghx): f_{x} z_{x} = f_{y_{-}^*} + f_{y_0} g_{x} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{x}= A g_{x} + f_{y_{-}^*}=0$
$(ghu): f_{z} z_{u} = f_{y_0} g_{u} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{u} + f_{u}= A g_u + f_u = 0$
$(ghxx) : A g_{xx} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{x}\right) + f_{zz} \left( z_{x} \otimes z_{x} \right) = 0$
$(ghxu) : A g_{xu} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{x} \otimes z_{u} \right) = 0$
$(ghuu) : A g_{uu} + B g_{xx} \left(g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{u} \otimes z_{u} \right) = 0$
$(ghs2) : (A+B) g_{\sigma\sigma} + \left( f_{y^{**}_{+}y^{**}_{+}} \left(g^{**}_{u} \otimes g^{**}_{u}\right) + f_{y^{**}_{+}} g^{**}_{uu}\right)vec(\Sigma) = 0$
$(ghxxx) : A g_{xxx} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{x}\right) + f_{y_{+}}g^{**}_{xx} \left(g^{*}_x \otimes g^{*}_{xx}\right)P_{x\_xx} + f_{zz} \left( z_{x} \otimes z_{xx} \right)P_{x\_xx} + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{x} \right) = 0$
$(ghxxu) : A g_{xxu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{u} \right) + f_{zz} \left( \left( z_{x} \otimes z_{xu} \right)P_{x\_xu} + \left(z_{xx} \otimes z_{u}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{x} \otimes g^{*}_{xu}\right)P_{x\_xu} + \left(g^{*}_{xx} \otimes g^{*}_{u}\right) \right) = 0$
$(ghxuu) : A g_{xuu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( \left( z_{xu} \otimes z_{u} \right)P_{xu\_u} + \left(z_{x} \otimes z_{uu}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{xu} \otimes g^{*}_{u}\right)P_{xu\_u} + \left(g^{*}_{x} \otimes g^{*}_{uu}\right) \right) = 0$
$(ghuuu) : A g_{uuu} + B g_{xxx} \left(g^{*}_{u} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{u} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( z_{u} \otimes z_{uu} \right)P_{u\_uu} + f_{y_{+}}g^{**}_{xx} \left(g^{*}_{u} \otimes g^{*}_{uu}\right)P_{u\_uu} = 0$
$(ghx\sigma\sigma) : A g_{x\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_x + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{x} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{x} \otimes z_{\sigma\sigma}\right) + F_{xu_{+}u_{+}}\left(I_{n_x} \otimes vec(\Sigma)\right) = 0$
$F_{xu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_x^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{xu_{+}} \otimes z_{u_{+}} \right)P_{xu\_u} + \left(z_{x} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{x} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
$(ghu\sigma\sigma) : A g_{u\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_{u} + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{u} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{u} \otimes z_{\sigma\sigma}\right) + F_{uu_{+}u_{+}}\left(I_{n_u} \otimes vec(\Sigma_u)\right) = 0$
$F_{uu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_u^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{uu_{+}} \otimes z_{u_{+}} \right)P_{uu\_u} + \left(z_{u} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{u} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
A and B are the common perturbation matrices:
$A = f_{y_0} + \begin{pmatrix} \underbrace{0}_{n\times n_{static}} &\vdots& \underbrace{f_{y^{**}_{+}} \cdot g^{**}_{x}}_{n \times n_{spred}} &\vdots& \underbrace{0}_{n\times n_{frwd}} \end{pmatrix}$and $B = \begin{pmatrix} \underbrace{0}_{n \times n_{static}}&\vdots & \underbrace{0}_{n \times n_{pred}} & \vdots & \underbrace{f_{y^{**}_{+}}}_{n \times n_{sfwrd}} \end{pmatrix}$
and $z=(y_{-}^{\ast}; y; y_{+}^{\ast\ast}; u)$ denotes the dynamic model variables as in `M_.lead_lag_incidence`, $y^\ast$ denote state variables, $y^{\ast\ast}$ denote forward looking variables, $y_+$ denote the variables with a lead, $y_{-}$ denote variables with a lag, $y_0$ denote variables at period t, $f$ the model equations, and $f_z$ the first-order dynamic model derivatives, $f_{zz}$ the second-order dynamic derivatives, and $f_{zzz}$ the third-order dynamic model derivatives. Then:
$z_{x} = \begin{pmatrix}I\\g_{x}\\g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{u} =\begin{pmatrix}0\\g_{u}\\g^{**}_{x} \cdot g^{*}_{u}\\I\end{pmatrix}$, $z_{u_{+}} =\begin{pmatrix}0\\0\\g^{**}_{u}\\0\end{pmatrix}$
$z_{xx} = \begin{pmatrix} 0\\g_{xx}\\g^{**}_{x} \left( g^{*}_x \otimes g^{*}_{x} \right) + g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{xu} =\begin{pmatrix}0\\g_{xu}\\g^{**}_{xx} \left( g^{*}_x \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{xu}\\0\end{pmatrix}$, $z_{uu} =\begin{pmatrix}0\\g_{uu}\\g^{**}_{xx} \left( g^{*}_u \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{uu}\\0\end{pmatrix}$,
$z_{xu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_x \otimes I \right)\\0\end{pmatrix}$, $z_{uu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_{u} \otimes I \right)\\0\end{pmatrix}$, $z_{u_{+}u_{+}} =\begin{pmatrix}0\\0\\g^{\ast\ast}_{uu}\\0\end{pmatrix}$, $z_{\sigma\sigma} = \begin{pmatrix}0\\ g_{\sigma\sigma}\\ g^{\ast\ast}_{x}g^{\ast}_{\sigma\sigma} + g^{\ast\ast}_{\sigma\sigma}\\0 \end{pmatrix}$
$P$ are permutation matrices that can be computed using Matlab's `ipermute` function.
## Parameter derivatives of perturbation solutions
First, we need the parameter derivatives of first, second, third, and fourth derivatives of the dynamic model (i.e. g1,g2,g3,g4 in dynamic files), I make use of the implicit function theorem: Let $f_{z^k}$ denote the kth derivative (wrt all dynamic variables) of the dynamic model, then let $df_{z^k}$ denote the first-derivative (wrt all model parameters) of $f_{z^k}$ evaluated at the steady state. Note that $f_{z^k}$ is a function of both the model parameters $\theta$ and of the steady state of all dynamic variables $\bar{z}$, which also depend on the parameters. Hence, implicitly $f_{z^k}=f_{z^k}(\theta,\bar{z}(\theta))$ and $df_{z^k}$ consists of two parts:
1. direct derivative wrt to all model parameters given by the preprocessor in the `_params_derivs.m` files
2. contribution of derivative of steady state of dynamic variables (wrt all model parameters): $f_{z^{k+1}} \cdot d\bar{z}$
Note that we already have functionality to compute $d\bar{z}$ analytically.
Having this, the above perturbation systems are basically equations of the following types
$AX +BXC = RHS$ or $AX = RHS$
Now when taking the derivative (wrt to one single parameter $\theta_j$), we get
$A\mathrm{d}\{X\} + B\mathrm{d}\{X\}C = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X - \mathrm{d}\{B\}XC - BX\mathrm{d}\{C\}$
or
$A\mathrm{d}\{X\} = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X$
The first one is a Sylvester type equation, the second one can be solved by taking the inverse of $A$. The only diffculty and tedious work arrises in computing (the highly sparse) derivatives of $RHS$.
***
# New functions: `
## get_perturbation_params_derivs.m`and `get_perturbation_params_derivs_numerical_objective.m`
* The parameter derivatives up to third order are computed in the new function`get_perturbation_params_derivs.m` both analytically and numerically. For numerical derivatives `get_perturbation_params_derivs_numerical_objective.m` is the objective for `fjaco.m` or `hessian_sparse.m` or `hessian.m`.
* `get_perturbation_params_derivs.m` is basically an extended version of the previous `get_first_order_solution_params_derivs.m` function.
* * `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It is used for numerical derivatives, whenever `analytic_derivation_mode=-1|-2`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
* Detailed changes:
* Most important: notation of this function is now in accordance to the k_order_solver, i.e. we do not compute derivatives of Kalman transition matrices A and B, but rather the solution matrices ghx,ghu,ghxx,ghxu,ghuu,ghs2,ghxxx,ghxxu,ghxuu,ghuuu,ghxss,ghuss in accordance with notation used in `oo_.dr`. As a byproduct at first-order, focusing on ghx and ghu instead of Kalman transition matrices A and B makes the computations slightly faster for large models (e.g. for Quest the computations were faster by a couple of seconds, not much, but okay).
* Removed use of `kstate`, see also Dynare/dynare#1653 and Dynare/dynare!1656
* Output arguments are stored in a structure `DERIVS`, there is also a flag `d2flag` that computes parameter hessians needed only in `dsge_likelihood.m`.
* Removed `kronflag` as input. `options_.analytic_derivation_mode` is now used instead of `kronflag`.
* Removed `indvar`, an index that was used to selected specific variables in the derivatives. This is not needed, as we always compute the parameter derivatives for all variables first and then select a subset of variables. The selection now takes place in other functions, like `dsge_likelihood.m`.
* Introduced some checks: (i) deterministic exogenous variables are not supported, (ii) Kronecker method only compatible with first-order approximation so reset to sylvester method, (iii) for purely backward or forward models we need to be careful with the rows in `M_.lead_la g_incidence`, (iv) if `_params_derivs.m` files are missing an error is thrown.
* For numerical derivatives, if mod file does not contain an `estimated_params_block`, a temporary one with the most important parameter information is created.
## `unfold_g4.m`
* When evaluating g3 and g4 one needs to take into account that these do not contain symmetric elements, so one needs to use `unfold_g3.m` and the new function `unfold_g4.m`. This returns an unfolded version of the same matrix (i.e. with symmetric elements).
***
# New test models
`.gitignore` and `Makefile.am` are changed accordingly. Also now it is possible to run test suite on analytic_derivatives, i.e. run `make check m/analytic_derivatives`
## `analytic_derivatives/BrockMirman_PertParamsDerivs.mod`
* This is the Brock Mirman model, where we know the exact policy function $g$ for capital and consumption. As this does not imply a nonzero $g_{\sigma\sigma}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ I added some artificial equations to get nonzero solution matrices with respect to $\sigma$. The true perturbation solution matrices $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ are then computed analytically with Matlab's symbolic toolbox and saved in `nBrockMirmanSYM.mat`. There is a preprocessor flag that recreates these analytical computations if changes are needed (and to check whether I made some errors here ;-) )
* Then solution matrices up to third order and their parameter Jacobians are then compared to the ones computed by Dynare's `k_order_solver` and by `get_perturbation_params_derivs` for all `analytic_derivation_mode`'s. There will be an error if the maximum absolute deviation is too large, i.e. for numerical derivatives (`analytic_derivation_mode=-1|-2`) the tolerance is choosen lower (around 1e-5); for analytical methods we are stricter: around 1e-13 for first-order, 1e-12 for second order, and 1e-11 for third-order.
* As a side note, this mod file also checks Dynare's `k_order_solver` algorithm and throws an error if something is wrong.
* This test model shows that the new functionality works well. And analytical derivatives perform way better and accurate than numerical ones, even for this small model.
## `analytic_derivatives/burnside_3_order_PertParamsDerivs.mod`
* This builds upon `tests/k_order_perturbation/burnside_k_order.mod` and computes the true parameter derivatives analytically by hand.
* This test model also shows that the new functionality works well.
## `analytic_derivatives/LindeTrabandt2019.mod`
* Shows that the new functionality also works for medium-sized models, i.e. a SW type model solved at third order with 35 variables (11 states). 2 shocks and 20 parameters.
* This mod file can be used to tweak the speed of the computations in the future.
* Compares numerical versus analytical parameter derivatives (for first, second and third order). Note that this model clearly shows that numerical ones are quite different than analytical ones even at first order!
## `identification/LindeTrabandt2019_xfail.mod`
* This model is a check for issue Dynare/dynare#1595, see fjaco.m below, and will fail.
* Removed `analytic_derivatives/ls2003.mod` as this mod file is neither in the testsuite nor does it work.
***
# Detailed changes in other functions
## `get_first_order_solution_params_derivs.m`
* Deleted, or actually, renamed to `get_perturbation_params_derivs.m`, as this function now is able to compute the derivatives up to third order
## `identification_numerical_objective.m`
* `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
## `dsge_likelihood.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `T`matrix anymore, but instead of the dynare solution matrix `ghx`. So we recreate `T` here as this amounts to adding some zeros and focusing on selected variables only.
* Added some checks to make sure the first-order approximation is selected.
* Removed `kron_flag` as input, as `get_perturbation_params_derivs` looks into `options_.analytic_derivation_mode` for `kron_flag`.
## `dynare_identification.m`
* make sure that setting `analytic_derivation_mode` is set both in `options_ident` and `options_`. Note that at the end of the function we restore the `options_` structure, so all changes are local. In a next merge request, I will remove the global variables to make all variables local.
## `get_identification_jacobians.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `A` and `B` matrix anymore, but instead of the dynare solution matrix `ghx` and `ghu`. So we recreate these matrices here instead of in `get_perturbation_params_derivs.m`.
* Added `str2func` for better function handles in `fjaco.m`.
## `fjaco.m`
* make `tol`an option, which can be adjusted by changing `options_.dynatol.x`for identification and parameter derivatives purposes.
* include a check and an informative error message, if numerical derivatives (two-sided finite difference method) yield errors in `resol.m` for identification and parameter derivatives purposes. This closes issue Dynare/dynare#1595.
* Changed year of copyright to 2010-2017,2019
***
# Further suggestions and questions
* Ones this is merged, I will merge request an improvement of the identification toolbox, which will work up to third order using the pruned state space. This will also remove some issues and bugs, and also I will remove global variables in this request.
* The third-order derivatives can be further improved by taking sparsity into account and use mex versions for kronecker products etc. I leave this for further testing (and if anybody actually uses this ;-) )
2019-12-17 19:17:09 +01:00
dRHSghxxx = zeros ( M . endo_nbr , M . nspred ^3 , modparam_nbr ) ;
for jp = 1 : modparam_nbr
dzxx ( : , : , jp ) = [ zeros ( M . nspred , M . nspred ^2 ) ;
dghxx ( kcurr ~= 0 , : , jp ) ;
dghxx ( klead ~= 0 , : , jp ) * G_x_x + oo . dr . ghxx ( klead ~= 0 , : ) * dG_x_x ( : , : , jp ) + dghx ( klead ~= 0 , : , jp ) * oo . dr . ghxx ( idx_states , : ) + oo . dr . ghx ( klead ~= 0 , : ) * dghxx ( idx_states , : , jp ) ;
zeros ( M . exo_nbr , M . nspred ^2 ) ] ;
dG_x_x_x = kron ( dghx ( idx_states , : , jp ) , G_x_x ) + kron ( oo . dr . ghx ( idx_states , : ) , dG_x_x ( : , : , jp ) ) ;
dG_x_xx = kron ( dghx ( idx_states , : , jp ) , oo . dr . ghxx ( idx_states , : ) ) + kron ( oo . dr . ghx ( idx_states , : ) , dghxx ( idx_states , : , jp ) ) ;
dZ_x_x = kron ( dzx ( : , : , jp ) , zx ) + kron ( zx , dzx ( : , : , jp ) ) ;
dZ_x_x_x = kron ( dzx ( : , : , jp ) , Z_x_x ) + kron ( zx , dZ_x_x ) ;
dZ_x_xx = kron ( dzx ( : , : , jp ) , zxx ) + kron ( zx , dzxx ( : , : , jp ) ) ;
2019-12-20 16:28:06 +01:00
dfyp_ghxxp ( : , : , jp ) = dg1 ( : , nonzeros ( klead ) , jp ) * oo . dr . ghxx ( klead ~= 0 , : ) + g1 ( : , nonzeros ( klead ) ) * dghxx ( klead ~= 0 , : , jp ) ;
Added parameter derivatives of perturbation solution up to 3 order
# Preliminary comments
I finished the identification toolbox at orders two and three using the pruned state space system, but before I merge request this, I decided to first merge the new functionality to compute parameter derivatives of perturbation solution matrices at higher orders. So after this is approved, I merge the identification toolbox.
I guess @rattoma, @sebastien, and @michel are best choices to review this.
I outline the main idea first and then provide some more detailed changes I made to the functions.
***
# Main idea
This merge request is concerned with the *analytical*computation of the parameter derivatives of first, second and third order perturbation solution matrices, i.e. using _closed-form_ expressions to efficiently compute the derivative of $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ *with respect to model parameters* $\theta$. Note that $\theta$ contains model parameters, stderr and corr parameters of shocks. stderr and corr parameters of measurement errors are not yet supported, (they can easily be included as exogenous shocks). The availability of such derivatives is beneficial in terms of more reliable analysis of model sensitivity and parameter identifiability as well as more efficient estimation methods, in particular for models solved up to third order, as it is well-known that numerical derivatives are a tricky business, especially for large models.
References for my approach are:
* Iskrev (2008, 2010) and Schmitt-Grohé and Uribe (2012, Appendix) who were the first to compute the parameter derivatives analytically at first order, however, using inefficient (sparse) Kronecker products.
* Mutschler (2015) who provides the expressions for a second-order, but again using inefficient (sparse) Kronecker products.
* Ratto and Iskrev (2012) who show how the first-order system can be solved accurately, fast and efficiently using existing numerical algorithms for generalized Sylvester equations by taking the parameter derivative with respect to each parameter separately.
* Julliard and Kamenik (2004) who provide the perturbation solution equation system in tensor notation at any order k.
* Levintal (2017) who introduces permutation matrices to express the perturbation solution equation system in matrix notation up to fifth order.
Note that @rattoma already implemented the parameter derivatives of $g_x$ and $g_u$ analytically (and numerically), and I rely heavily on his work in `get_first_order_solution_params_derivs.m` (previously `getH.m`). My additions are mainly to this function and thus it is renamed to `get_perturbation_params_derivs.m`.
The basic idea of this merge request is to take the second- and third-order perturbation solution systems in Julliard and Kamenik (2004), unfold these into an equivalent matrix representation using permutation matrices as in Levintal (2017). Then extending Ratto and Iskrev (2012) one takes the derivative with respect to each parameter separately and gets a computational problem that is linear, albeit large, as it involves either solving generalized Sylvester equations or taking inverses of highly sparse matrices. I will now briefly summarize the perturbation solution system at third order and the system that results when taking the derivative with respect to parameters.
## Perturbation Solution
The following systems arise at first, second, and third order:
$(ghx): f_{x} z_{x} = f_{y_{-}^*} + f_{y_0} g_{x} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{x}= A g_{x} + f_{y_{-}^*}=0$
$(ghu): f_{z} z_{u} = f_{y_0} g_{u} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{u} + f_{u}= A g_u + f_u = 0$
$(ghxx) : A g_{xx} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{x}\right) + f_{zz} \left( z_{x} \otimes z_{x} \right) = 0$
$(ghxu) : A g_{xu} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{x} \otimes z_{u} \right) = 0$
$(ghuu) : A g_{uu} + B g_{xx} \left(g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{u} \otimes z_{u} \right) = 0$
$(ghs2) : (A+B) g_{\sigma\sigma} + \left( f_{y^{**}_{+}y^{**}_{+}} \left(g^{**}_{u} \otimes g^{**}_{u}\right) + f_{y^{**}_{+}} g^{**}_{uu}\right)vec(\Sigma) = 0$
$(ghxxx) : A g_{xxx} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{x}\right) + f_{y_{+}}g^{**}_{xx} \left(g^{*}_x \otimes g^{*}_{xx}\right)P_{x\_xx} + f_{zz} \left( z_{x} \otimes z_{xx} \right)P_{x\_xx} + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{x} \right) = 0$
$(ghxxu) : A g_{xxu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{u} \right) + f_{zz} \left( \left( z_{x} \otimes z_{xu} \right)P_{x\_xu} + \left(z_{xx} \otimes z_{u}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{x} \otimes g^{*}_{xu}\right)P_{x\_xu} + \left(g^{*}_{xx} \otimes g^{*}_{u}\right) \right) = 0$
$(ghxuu) : A g_{xuu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( \left( z_{xu} \otimes z_{u} \right)P_{xu\_u} + \left(z_{x} \otimes z_{uu}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{xu} \otimes g^{*}_{u}\right)P_{xu\_u} + \left(g^{*}_{x} \otimes g^{*}_{uu}\right) \right) = 0$
$(ghuuu) : A g_{uuu} + B g_{xxx} \left(g^{*}_{u} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{u} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( z_{u} \otimes z_{uu} \right)P_{u\_uu} + f_{y_{+}}g^{**}_{xx} \left(g^{*}_{u} \otimes g^{*}_{uu}\right)P_{u\_uu} = 0$
$(ghx\sigma\sigma) : A g_{x\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_x + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{x} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{x} \otimes z_{\sigma\sigma}\right) + F_{xu_{+}u_{+}}\left(I_{n_x} \otimes vec(\Sigma)\right) = 0$
$F_{xu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_x^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{xu_{+}} \otimes z_{u_{+}} \right)P_{xu\_u} + \left(z_{x} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{x} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
$(ghu\sigma\sigma) : A g_{u\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_{u} + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{u} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{u} \otimes z_{\sigma\sigma}\right) + F_{uu_{+}u_{+}}\left(I_{n_u} \otimes vec(\Sigma_u)\right) = 0$
$F_{uu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_u^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{uu_{+}} \otimes z_{u_{+}} \right)P_{uu\_u} + \left(z_{u} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{u} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
A and B are the common perturbation matrices:
$A = f_{y_0} + \begin{pmatrix} \underbrace{0}_{n\times n_{static}} &\vdots& \underbrace{f_{y^{**}_{+}} \cdot g^{**}_{x}}_{n \times n_{spred}} &\vdots& \underbrace{0}_{n\times n_{frwd}} \end{pmatrix}$and $B = \begin{pmatrix} \underbrace{0}_{n \times n_{static}}&\vdots & \underbrace{0}_{n \times n_{pred}} & \vdots & \underbrace{f_{y^{**}_{+}}}_{n \times n_{sfwrd}} \end{pmatrix}$
and $z=(y_{-}^{\ast}; y; y_{+}^{\ast\ast}; u)$ denotes the dynamic model variables as in `M_.lead_lag_incidence`, $y^\ast$ denote state variables, $y^{\ast\ast}$ denote forward looking variables, $y_+$ denote the variables with a lead, $y_{-}$ denote variables with a lag, $y_0$ denote variables at period t, $f$ the model equations, and $f_z$ the first-order dynamic model derivatives, $f_{zz}$ the second-order dynamic derivatives, and $f_{zzz}$ the third-order dynamic model derivatives. Then:
$z_{x} = \begin{pmatrix}I\\g_{x}\\g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{u} =\begin{pmatrix}0\\g_{u}\\g^{**}_{x} \cdot g^{*}_{u}\\I\end{pmatrix}$, $z_{u_{+}} =\begin{pmatrix}0\\0\\g^{**}_{u}\\0\end{pmatrix}$
$z_{xx} = \begin{pmatrix} 0\\g_{xx}\\g^{**}_{x} \left( g^{*}_x \otimes g^{*}_{x} \right) + g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{xu} =\begin{pmatrix}0\\g_{xu}\\g^{**}_{xx} \left( g^{*}_x \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{xu}\\0\end{pmatrix}$, $z_{uu} =\begin{pmatrix}0\\g_{uu}\\g^{**}_{xx} \left( g^{*}_u \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{uu}\\0\end{pmatrix}$,
$z_{xu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_x \otimes I \right)\\0\end{pmatrix}$, $z_{uu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_{u} \otimes I \right)\\0\end{pmatrix}$, $z_{u_{+}u_{+}} =\begin{pmatrix}0\\0\\g^{\ast\ast}_{uu}\\0\end{pmatrix}$, $z_{\sigma\sigma} = \begin{pmatrix}0\\ g_{\sigma\sigma}\\ g^{\ast\ast}_{x}g^{\ast}_{\sigma\sigma} + g^{\ast\ast}_{\sigma\sigma}\\0 \end{pmatrix}$
$P$ are permutation matrices that can be computed using Matlab's `ipermute` function.
## Parameter derivatives of perturbation solutions
First, we need the parameter derivatives of first, second, third, and fourth derivatives of the dynamic model (i.e. g1,g2,g3,g4 in dynamic files), I make use of the implicit function theorem: Let $f_{z^k}$ denote the kth derivative (wrt all dynamic variables) of the dynamic model, then let $df_{z^k}$ denote the first-derivative (wrt all model parameters) of $f_{z^k}$ evaluated at the steady state. Note that $f_{z^k}$ is a function of both the model parameters $\theta$ and of the steady state of all dynamic variables $\bar{z}$, which also depend on the parameters. Hence, implicitly $f_{z^k}=f_{z^k}(\theta,\bar{z}(\theta))$ and $df_{z^k}$ consists of two parts:
1. direct derivative wrt to all model parameters given by the preprocessor in the `_params_derivs.m` files
2. contribution of derivative of steady state of dynamic variables (wrt all model parameters): $f_{z^{k+1}} \cdot d\bar{z}$
Note that we already have functionality to compute $d\bar{z}$ analytically.
Having this, the above perturbation systems are basically equations of the following types
$AX +BXC = RHS$ or $AX = RHS$
Now when taking the derivative (wrt to one single parameter $\theta_j$), we get
$A\mathrm{d}\{X\} + B\mathrm{d}\{X\}C = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X - \mathrm{d}\{B\}XC - BX\mathrm{d}\{C\}$
or
$A\mathrm{d}\{X\} = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X$
The first one is a Sylvester type equation, the second one can be solved by taking the inverse of $A$. The only diffculty and tedious work arrises in computing (the highly sparse) derivatives of $RHS$.
***
# New functions: `
## get_perturbation_params_derivs.m`and `get_perturbation_params_derivs_numerical_objective.m`
* The parameter derivatives up to third order are computed in the new function`get_perturbation_params_derivs.m` both analytically and numerically. For numerical derivatives `get_perturbation_params_derivs_numerical_objective.m` is the objective for `fjaco.m` or `hessian_sparse.m` or `hessian.m`.
* `get_perturbation_params_derivs.m` is basically an extended version of the previous `get_first_order_solution_params_derivs.m` function.
* * `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It is used for numerical derivatives, whenever `analytic_derivation_mode=-1|-2`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
* Detailed changes:
* Most important: notation of this function is now in accordance to the k_order_solver, i.e. we do not compute derivatives of Kalman transition matrices A and B, but rather the solution matrices ghx,ghu,ghxx,ghxu,ghuu,ghs2,ghxxx,ghxxu,ghxuu,ghuuu,ghxss,ghuss in accordance with notation used in `oo_.dr`. As a byproduct at first-order, focusing on ghx and ghu instead of Kalman transition matrices A and B makes the computations slightly faster for large models (e.g. for Quest the computations were faster by a couple of seconds, not much, but okay).
* Removed use of `kstate`, see also Dynare/dynare#1653 and Dynare/dynare!1656
* Output arguments are stored in a structure `DERIVS`, there is also a flag `d2flag` that computes parameter hessians needed only in `dsge_likelihood.m`.
* Removed `kronflag` as input. `options_.analytic_derivation_mode` is now used instead of `kronflag`.
* Removed `indvar`, an index that was used to selected specific variables in the derivatives. This is not needed, as we always compute the parameter derivatives for all variables first and then select a subset of variables. The selection now takes place in other functions, like `dsge_likelihood.m`.
* Introduced some checks: (i) deterministic exogenous variables are not supported, (ii) Kronecker method only compatible with first-order approximation so reset to sylvester method, (iii) for purely backward or forward models we need to be careful with the rows in `M_.lead_la g_incidence`, (iv) if `_params_derivs.m` files are missing an error is thrown.
* For numerical derivatives, if mod file does not contain an `estimated_params_block`, a temporary one with the most important parameter information is created.
## `unfold_g4.m`
* When evaluating g3 and g4 one needs to take into account that these do not contain symmetric elements, so one needs to use `unfold_g3.m` and the new function `unfold_g4.m`. This returns an unfolded version of the same matrix (i.e. with symmetric elements).
***
# New test models
`.gitignore` and `Makefile.am` are changed accordingly. Also now it is possible to run test suite on analytic_derivatives, i.e. run `make check m/analytic_derivatives`
## `analytic_derivatives/BrockMirman_PertParamsDerivs.mod`
* This is the Brock Mirman model, where we know the exact policy function $g$ for capital and consumption. As this does not imply a nonzero $g_{\sigma\sigma}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ I added some artificial equations to get nonzero solution matrices with respect to $\sigma$. The true perturbation solution matrices $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ are then computed analytically with Matlab's symbolic toolbox and saved in `nBrockMirmanSYM.mat`. There is a preprocessor flag that recreates these analytical computations if changes are needed (and to check whether I made some errors here ;-) )
* Then solution matrices up to third order and their parameter Jacobians are then compared to the ones computed by Dynare's `k_order_solver` and by `get_perturbation_params_derivs` for all `analytic_derivation_mode`'s. There will be an error if the maximum absolute deviation is too large, i.e. for numerical derivatives (`analytic_derivation_mode=-1|-2`) the tolerance is choosen lower (around 1e-5); for analytical methods we are stricter: around 1e-13 for first-order, 1e-12 for second order, and 1e-11 for third-order.
* As a side note, this mod file also checks Dynare's `k_order_solver` algorithm and throws an error if something is wrong.
* This test model shows that the new functionality works well. And analytical derivatives perform way better and accurate than numerical ones, even for this small model.
## `analytic_derivatives/burnside_3_order_PertParamsDerivs.mod`
* This builds upon `tests/k_order_perturbation/burnside_k_order.mod` and computes the true parameter derivatives analytically by hand.
* This test model also shows that the new functionality works well.
## `analytic_derivatives/LindeTrabandt2019.mod`
* Shows that the new functionality also works for medium-sized models, i.e. a SW type model solved at third order with 35 variables (11 states). 2 shocks and 20 parameters.
* This mod file can be used to tweak the speed of the computations in the future.
* Compares numerical versus analytical parameter derivatives (for first, second and third order). Note that this model clearly shows that numerical ones are quite different than analytical ones even at first order!
## `identification/LindeTrabandt2019_xfail.mod`
* This model is a check for issue Dynare/dynare#1595, see fjaco.m below, and will fail.
* Removed `analytic_derivatives/ls2003.mod` as this mod file is neither in the testsuite nor does it work.
***
# Detailed changes in other functions
## `get_first_order_solution_params_derivs.m`
* Deleted, or actually, renamed to `get_perturbation_params_derivs.m`, as this function now is able to compute the derivatives up to third order
## `identification_numerical_objective.m`
* `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
## `dsge_likelihood.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `T`matrix anymore, but instead of the dynare solution matrix `ghx`. So we recreate `T` here as this amounts to adding some zeros and focusing on selected variables only.
* Added some checks to make sure the first-order approximation is selected.
* Removed `kron_flag` as input, as `get_perturbation_params_derivs` looks into `options_.analytic_derivation_mode` for `kron_flag`.
## `dynare_identification.m`
* make sure that setting `analytic_derivation_mode` is set both in `options_ident` and `options_`. Note that at the end of the function we restore the `options_` structure, so all changes are local. In a next merge request, I will remove the global variables to make all variables local.
## `get_identification_jacobians.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `A` and `B` matrix anymore, but instead of the dynare solution matrix `ghx` and `ghu`. So we recreate these matrices here instead of in `get_perturbation_params_derivs.m`.
* Added `str2func` for better function handles in `fjaco.m`.
## `fjaco.m`
* make `tol`an option, which can be adjusted by changing `options_.dynatol.x`for identification and parameter derivatives purposes.
* include a check and an informative error message, if numerical derivatives (two-sided finite difference method) yield errors in `resol.m` for identification and parameter derivatives purposes. This closes issue Dynare/dynare#1595.
* Changed year of copyright to 2010-2017,2019
***
# Further suggestions and questions
* Ones this is merged, I will merge request an improvement of the identification toolbox, which will work up to third order using the pruned state space. This will also remove some issues and bugs, and also I will remove global variables in this request.
* The third-order derivatives can be further improved by taking sparsity into account and use mex versions for kronecker products etc. I leave this for further testing (and if anybody actually uses this ;-) )
2019-12-17 19:17:09 +01:00
dRHSghxxx ( : , : , jp ) = dA ( : , : , jp ) * oo . dr . ghxxx + dB ( : , : , jp ) * oo . dr . ghxxx * G_x_x_x + B_ghxxx * dG_x_x_x ;
dRHSghxxx ( : , : , jp ) = dRHSghxxx ( : , : , jp ) + dg3 ( : , k3yy0ex0 ( kyy0ex0 , kyy0ex0 , kyy0ex0 ) + ( jp - 1 ) * yy0ex0_nbr ^3 ) * Z_x_x_x + g3 ( : , k3yy0ex0 ( kyy0ex0 , kyy0ex0 , kyy0ex0 ) ) * dZ_x_x_x ;
dRHSghxxx ( : , : , jp ) = dRHSghxxx ( : , : , jp ) + dg2 ( : , k2yy0ex0 ( kyy0ex0 , kyy0ex0 ) + ( jp - 1 ) * yy0ex0_nbr ^2 ) * Z_x_xx * P_x_xx + g2 ( : , k2yy0ex0 ( kyy0ex0 , kyy0ex0 ) ) * dZ_x_xx * P_x_xx ;
dRHSghxxx ( : , : , jp ) = dRHSghxxx ( : , : , jp ) + dfyp_ghxxp ( : , : , jp ) * G_x_xx * P_x_xx + fyp_ghxxp * dG_x_xx * P_x_xx ;
end
dRHSghxxx = - dRHSghxxx ;
%use iterated generalized sylvester equation to compute dghxxx
dghxxx = sylvester3 ( A , B , G_x_x_x , dRHSghxxx ) ;
flag = 1 ; icount = 0 ;
while flag && icount < 4
[ dghxxx , flag ] = sylvester3a ( dghxxx , A , B , G_x_x_x , dRHSghxxx ) ;
icount = icount + 1 ;
end
%Auxiliary expressions for dghxxu, dghxuu, dghuuu, dghxss, dghuss
G_x_u = kron ( oo . dr . ghx ( idx_states , : ) , oo . dr . ghu ( idx_states , : ) ) ;
G_u_u = kron ( oo . dr . ghu ( idx_states , : ) , oo . dr . ghu ( idx_states , : ) ) ;
zxu = [ zeros ( M . nspred , M . nspred * M . exo_nbr ) ;
oo . dr . ghxu ( kcurr ~= 0 , : ) ;
oo . dr . ghxx ( klead ~= 0 , : ) * G_x_u + oo . dr . ghx ( klead ~= 0 , : ) * oo . dr . ghxu ( idx_states , : ) ;
zeros ( M . exo_nbr , M . exo_nbr * M . nspred ) ] ;
zuu = [ zeros ( M . nspred , M . exo_nbr ^2 ) ;
oo . dr . ghuu ( kcurr ~= 0 , : ) ;
oo . dr . ghxx ( klead ~= 0 , : ) * G_u_u + oo . dr . ghx ( klead ~= 0 , : ) * oo . dr . ghuu ( idx_states , : ) ;
2019-12-20 16:28:06 +01:00
zeros ( M . exo_nbr , M . exo_nbr ^2 ) ] ;
Added parameter derivatives of perturbation solution up to 3 order
# Preliminary comments
I finished the identification toolbox at orders two and three using the pruned state space system, but before I merge request this, I decided to first merge the new functionality to compute parameter derivatives of perturbation solution matrices at higher orders. So after this is approved, I merge the identification toolbox.
I guess @rattoma, @sebastien, and @michel are best choices to review this.
I outline the main idea first and then provide some more detailed changes I made to the functions.
***
# Main idea
This merge request is concerned with the *analytical*computation of the parameter derivatives of first, second and third order perturbation solution matrices, i.e. using _closed-form_ expressions to efficiently compute the derivative of $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ *with respect to model parameters* $\theta$. Note that $\theta$ contains model parameters, stderr and corr parameters of shocks. stderr and corr parameters of measurement errors are not yet supported, (they can easily be included as exogenous shocks). The availability of such derivatives is beneficial in terms of more reliable analysis of model sensitivity and parameter identifiability as well as more efficient estimation methods, in particular for models solved up to third order, as it is well-known that numerical derivatives are a tricky business, especially for large models.
References for my approach are:
* Iskrev (2008, 2010) and Schmitt-Grohé and Uribe (2012, Appendix) who were the first to compute the parameter derivatives analytically at first order, however, using inefficient (sparse) Kronecker products.
* Mutschler (2015) who provides the expressions for a second-order, but again using inefficient (sparse) Kronecker products.
* Ratto and Iskrev (2012) who show how the first-order system can be solved accurately, fast and efficiently using existing numerical algorithms for generalized Sylvester equations by taking the parameter derivative with respect to each parameter separately.
* Julliard and Kamenik (2004) who provide the perturbation solution equation system in tensor notation at any order k.
* Levintal (2017) who introduces permutation matrices to express the perturbation solution equation system in matrix notation up to fifth order.
Note that @rattoma already implemented the parameter derivatives of $g_x$ and $g_u$ analytically (and numerically), and I rely heavily on his work in `get_first_order_solution_params_derivs.m` (previously `getH.m`). My additions are mainly to this function and thus it is renamed to `get_perturbation_params_derivs.m`.
The basic idea of this merge request is to take the second- and third-order perturbation solution systems in Julliard and Kamenik (2004), unfold these into an equivalent matrix representation using permutation matrices as in Levintal (2017). Then extending Ratto and Iskrev (2012) one takes the derivative with respect to each parameter separately and gets a computational problem that is linear, albeit large, as it involves either solving generalized Sylvester equations or taking inverses of highly sparse matrices. I will now briefly summarize the perturbation solution system at third order and the system that results when taking the derivative with respect to parameters.
## Perturbation Solution
The following systems arise at first, second, and third order:
$(ghx): f_{x} z_{x} = f_{y_{-}^*} + f_{y_0} g_{x} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{x}= A g_{x} + f_{y_{-}^*}=0$
$(ghu): f_{z} z_{u} = f_{y_0} g_{u} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{u} + f_{u}= A g_u + f_u = 0$
$(ghxx) : A g_{xx} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{x}\right) + f_{zz} \left( z_{x} \otimes z_{x} \right) = 0$
$(ghxu) : A g_{xu} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{x} \otimes z_{u} \right) = 0$
$(ghuu) : A g_{uu} + B g_{xx} \left(g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{u} \otimes z_{u} \right) = 0$
$(ghs2) : (A+B) g_{\sigma\sigma} + \left( f_{y^{**}_{+}y^{**}_{+}} \left(g^{**}_{u} \otimes g^{**}_{u}\right) + f_{y^{**}_{+}} g^{**}_{uu}\right)vec(\Sigma) = 0$
$(ghxxx) : A g_{xxx} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{x}\right) + f_{y_{+}}g^{**}_{xx} \left(g^{*}_x \otimes g^{*}_{xx}\right)P_{x\_xx} + f_{zz} \left( z_{x} \otimes z_{xx} \right)P_{x\_xx} + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{x} \right) = 0$
$(ghxxu) : A g_{xxu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{u} \right) + f_{zz} \left( \left( z_{x} \otimes z_{xu} \right)P_{x\_xu} + \left(z_{xx} \otimes z_{u}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{x} \otimes g^{*}_{xu}\right)P_{x\_xu} + \left(g^{*}_{xx} \otimes g^{*}_{u}\right) \right) = 0$
$(ghxuu) : A g_{xuu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( \left( z_{xu} \otimes z_{u} \right)P_{xu\_u} + \left(z_{x} \otimes z_{uu}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{xu} \otimes g^{*}_{u}\right)P_{xu\_u} + \left(g^{*}_{x} \otimes g^{*}_{uu}\right) \right) = 0$
$(ghuuu) : A g_{uuu} + B g_{xxx} \left(g^{*}_{u} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{u} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( z_{u} \otimes z_{uu} \right)P_{u\_uu} + f_{y_{+}}g^{**}_{xx} \left(g^{*}_{u} \otimes g^{*}_{uu}\right)P_{u\_uu} = 0$
$(ghx\sigma\sigma) : A g_{x\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_x + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{x} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{x} \otimes z_{\sigma\sigma}\right) + F_{xu_{+}u_{+}}\left(I_{n_x} \otimes vec(\Sigma)\right) = 0$
$F_{xu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_x^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{xu_{+}} \otimes z_{u_{+}} \right)P_{xu\_u} + \left(z_{x} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{x} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
$(ghu\sigma\sigma) : A g_{u\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_{u} + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{u} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{u} \otimes z_{\sigma\sigma}\right) + F_{uu_{+}u_{+}}\left(I_{n_u} \otimes vec(\Sigma_u)\right) = 0$
$F_{uu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_u^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{uu_{+}} \otimes z_{u_{+}} \right)P_{uu\_u} + \left(z_{u} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{u} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
A and B are the common perturbation matrices:
$A = f_{y_0} + \begin{pmatrix} \underbrace{0}_{n\times n_{static}} &\vdots& \underbrace{f_{y^{**}_{+}} \cdot g^{**}_{x}}_{n \times n_{spred}} &\vdots& \underbrace{0}_{n\times n_{frwd}} \end{pmatrix}$and $B = \begin{pmatrix} \underbrace{0}_{n \times n_{static}}&\vdots & \underbrace{0}_{n \times n_{pred}} & \vdots & \underbrace{f_{y^{**}_{+}}}_{n \times n_{sfwrd}} \end{pmatrix}$
and $z=(y_{-}^{\ast}; y; y_{+}^{\ast\ast}; u)$ denotes the dynamic model variables as in `M_.lead_lag_incidence`, $y^\ast$ denote state variables, $y^{\ast\ast}$ denote forward looking variables, $y_+$ denote the variables with a lead, $y_{-}$ denote variables with a lag, $y_0$ denote variables at period t, $f$ the model equations, and $f_z$ the first-order dynamic model derivatives, $f_{zz}$ the second-order dynamic derivatives, and $f_{zzz}$ the third-order dynamic model derivatives. Then:
$z_{x} = \begin{pmatrix}I\\g_{x}\\g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{u} =\begin{pmatrix}0\\g_{u}\\g^{**}_{x} \cdot g^{*}_{u}\\I\end{pmatrix}$, $z_{u_{+}} =\begin{pmatrix}0\\0\\g^{**}_{u}\\0\end{pmatrix}$
$z_{xx} = \begin{pmatrix} 0\\g_{xx}\\g^{**}_{x} \left( g^{*}_x \otimes g^{*}_{x} \right) + g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{xu} =\begin{pmatrix}0\\g_{xu}\\g^{**}_{xx} \left( g^{*}_x \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{xu}\\0\end{pmatrix}$, $z_{uu} =\begin{pmatrix}0\\g_{uu}\\g^{**}_{xx} \left( g^{*}_u \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{uu}\\0\end{pmatrix}$,
$z_{xu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_x \otimes I \right)\\0\end{pmatrix}$, $z_{uu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_{u} \otimes I \right)\\0\end{pmatrix}$, $z_{u_{+}u_{+}} =\begin{pmatrix}0\\0\\g^{\ast\ast}_{uu}\\0\end{pmatrix}$, $z_{\sigma\sigma} = \begin{pmatrix}0\\ g_{\sigma\sigma}\\ g^{\ast\ast}_{x}g^{\ast}_{\sigma\sigma} + g^{\ast\ast}_{\sigma\sigma}\\0 \end{pmatrix}$
$P$ are permutation matrices that can be computed using Matlab's `ipermute` function.
## Parameter derivatives of perturbation solutions
First, we need the parameter derivatives of first, second, third, and fourth derivatives of the dynamic model (i.e. g1,g2,g3,g4 in dynamic files), I make use of the implicit function theorem: Let $f_{z^k}$ denote the kth derivative (wrt all dynamic variables) of the dynamic model, then let $df_{z^k}$ denote the first-derivative (wrt all model parameters) of $f_{z^k}$ evaluated at the steady state. Note that $f_{z^k}$ is a function of both the model parameters $\theta$ and of the steady state of all dynamic variables $\bar{z}$, which also depend on the parameters. Hence, implicitly $f_{z^k}=f_{z^k}(\theta,\bar{z}(\theta))$ and $df_{z^k}$ consists of two parts:
1. direct derivative wrt to all model parameters given by the preprocessor in the `_params_derivs.m` files
2. contribution of derivative of steady state of dynamic variables (wrt all model parameters): $f_{z^{k+1}} \cdot d\bar{z}$
Note that we already have functionality to compute $d\bar{z}$ analytically.
Having this, the above perturbation systems are basically equations of the following types
$AX +BXC = RHS$ or $AX = RHS$
Now when taking the derivative (wrt to one single parameter $\theta_j$), we get
$A\mathrm{d}\{X\} + B\mathrm{d}\{X\}C = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X - \mathrm{d}\{B\}XC - BX\mathrm{d}\{C\}$
or
$A\mathrm{d}\{X\} = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X$
The first one is a Sylvester type equation, the second one can be solved by taking the inverse of $A$. The only diffculty and tedious work arrises in computing (the highly sparse) derivatives of $RHS$.
***
# New functions: `
## get_perturbation_params_derivs.m`and `get_perturbation_params_derivs_numerical_objective.m`
* The parameter derivatives up to third order are computed in the new function`get_perturbation_params_derivs.m` both analytically and numerically. For numerical derivatives `get_perturbation_params_derivs_numerical_objective.m` is the objective for `fjaco.m` or `hessian_sparse.m` or `hessian.m`.
* `get_perturbation_params_derivs.m` is basically an extended version of the previous `get_first_order_solution_params_derivs.m` function.
* * `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It is used for numerical derivatives, whenever `analytic_derivation_mode=-1|-2`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
* Detailed changes:
* Most important: notation of this function is now in accordance to the k_order_solver, i.e. we do not compute derivatives of Kalman transition matrices A and B, but rather the solution matrices ghx,ghu,ghxx,ghxu,ghuu,ghs2,ghxxx,ghxxu,ghxuu,ghuuu,ghxss,ghuss in accordance with notation used in `oo_.dr`. As a byproduct at first-order, focusing on ghx and ghu instead of Kalman transition matrices A and B makes the computations slightly faster for large models (e.g. for Quest the computations were faster by a couple of seconds, not much, but okay).
* Removed use of `kstate`, see also Dynare/dynare#1653 and Dynare/dynare!1656
* Output arguments are stored in a structure `DERIVS`, there is also a flag `d2flag` that computes parameter hessians needed only in `dsge_likelihood.m`.
* Removed `kronflag` as input. `options_.analytic_derivation_mode` is now used instead of `kronflag`.
* Removed `indvar`, an index that was used to selected specific variables in the derivatives. This is not needed, as we always compute the parameter derivatives for all variables first and then select a subset of variables. The selection now takes place in other functions, like `dsge_likelihood.m`.
* Introduced some checks: (i) deterministic exogenous variables are not supported, (ii) Kronecker method only compatible with first-order approximation so reset to sylvester method, (iii) for purely backward or forward models we need to be careful with the rows in `M_.lead_la g_incidence`, (iv) if `_params_derivs.m` files are missing an error is thrown.
* For numerical derivatives, if mod file does not contain an `estimated_params_block`, a temporary one with the most important parameter information is created.
## `unfold_g4.m`
* When evaluating g3 and g4 one needs to take into account that these do not contain symmetric elements, so one needs to use `unfold_g3.m` and the new function `unfold_g4.m`. This returns an unfolded version of the same matrix (i.e. with symmetric elements).
***
# New test models
`.gitignore` and `Makefile.am` are changed accordingly. Also now it is possible to run test suite on analytic_derivatives, i.e. run `make check m/analytic_derivatives`
## `analytic_derivatives/BrockMirman_PertParamsDerivs.mod`
* This is the Brock Mirman model, where we know the exact policy function $g$ for capital and consumption. As this does not imply a nonzero $g_{\sigma\sigma}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ I added some artificial equations to get nonzero solution matrices with respect to $\sigma$. The true perturbation solution matrices $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ are then computed analytically with Matlab's symbolic toolbox and saved in `nBrockMirmanSYM.mat`. There is a preprocessor flag that recreates these analytical computations if changes are needed (and to check whether I made some errors here ;-) )
* Then solution matrices up to third order and their parameter Jacobians are then compared to the ones computed by Dynare's `k_order_solver` and by `get_perturbation_params_derivs` for all `analytic_derivation_mode`'s. There will be an error if the maximum absolute deviation is too large, i.e. for numerical derivatives (`analytic_derivation_mode=-1|-2`) the tolerance is choosen lower (around 1e-5); for analytical methods we are stricter: around 1e-13 for first-order, 1e-12 for second order, and 1e-11 for third-order.
* As a side note, this mod file also checks Dynare's `k_order_solver` algorithm and throws an error if something is wrong.
* This test model shows that the new functionality works well. And analytical derivatives perform way better and accurate than numerical ones, even for this small model.
## `analytic_derivatives/burnside_3_order_PertParamsDerivs.mod`
* This builds upon `tests/k_order_perturbation/burnside_k_order.mod` and computes the true parameter derivatives analytically by hand.
* This test model also shows that the new functionality works well.
## `analytic_derivatives/LindeTrabandt2019.mod`
* Shows that the new functionality also works for medium-sized models, i.e. a SW type model solved at third order with 35 variables (11 states). 2 shocks and 20 parameters.
* This mod file can be used to tweak the speed of the computations in the future.
* Compares numerical versus analytical parameter derivatives (for first, second and third order). Note that this model clearly shows that numerical ones are quite different than analytical ones even at first order!
## `identification/LindeTrabandt2019_xfail.mod`
* This model is a check for issue Dynare/dynare#1595, see fjaco.m below, and will fail.
* Removed `analytic_derivatives/ls2003.mod` as this mod file is neither in the testsuite nor does it work.
***
# Detailed changes in other functions
## `get_first_order_solution_params_derivs.m`
* Deleted, or actually, renamed to `get_perturbation_params_derivs.m`, as this function now is able to compute the derivatives up to third order
## `identification_numerical_objective.m`
* `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
## `dsge_likelihood.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `T`matrix anymore, but instead of the dynare solution matrix `ghx`. So we recreate `T` here as this amounts to adding some zeros and focusing on selected variables only.
* Added some checks to make sure the first-order approximation is selected.
* Removed `kron_flag` as input, as `get_perturbation_params_derivs` looks into `options_.analytic_derivation_mode` for `kron_flag`.
## `dynare_identification.m`
* make sure that setting `analytic_derivation_mode` is set both in `options_ident` and `options_`. Note that at the end of the function we restore the `options_` structure, so all changes are local. In a next merge request, I will remove the global variables to make all variables local.
## `get_identification_jacobians.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `A` and `B` matrix anymore, but instead of the dynare solution matrix `ghx` and `ghu`. So we recreate these matrices here instead of in `get_perturbation_params_derivs.m`.
* Added `str2func` for better function handles in `fjaco.m`.
## `fjaco.m`
* make `tol`an option, which can be adjusted by changing `options_.dynatol.x`for identification and parameter derivatives purposes.
* include a check and an informative error message, if numerical derivatives (two-sided finite difference method) yield errors in `resol.m` for identification and parameter derivatives purposes. This closes issue Dynare/dynare#1595.
* Changed year of copyright to 2010-2017,2019
***
# Further suggestions and questions
* Ones this is merged, I will merge request an improvement of the identification toolbox, which will work up to third order using the pruned state space. This will also remove some issues and bugs, and also I will remove global variables in this request.
* The third-order derivatives can be further improved by taking sparsity into account and use mex versions for kronecker products etc. I leave this for further testing (and if anybody actually uses this ;-) )
2019-12-17 19:17:09 +01:00
Z_x_u = kron ( zx , zu ) ;
Z_u_u = kron ( zu , zu ) ;
Z_x_xu = kron ( zx , zxu ) ;
Z_xx_u = kron ( zxx , zu ) ;
Z_xu_u = kron ( zxu , zu ) ;
Z_x_uu = kron ( zx , zuu ) ;
Z_u_uu = kron ( zu , zuu ) ;
Z_x_x_u = kron ( Z_x_x , zu ) ;
Z_x_u_u = kron ( Z_x_u , zu ) ;
Z_u_u_u = kron ( Z_u_u , zu ) ;
G_x_xu = kron ( oo . dr . ghx ( idx_states , : ) , oo . dr . ghxu ( idx_states , : ) ) ;
G_xx_u = kron ( oo . dr . ghxx ( idx_states , : ) , oo . dr . ghu ( idx_states , : ) ) ;
G_xu_u = kron ( oo . dr . ghxu ( idx_states , : ) , oo . dr . ghu ( idx_states , : ) ) ;
G_x_uu = kron ( oo . dr . ghx ( idx_states , : ) , oo . dr . ghuu ( idx_states , : ) ) ;
G_u_uu = kron ( oo . dr . ghu ( idx_states , : ) , oo . dr . ghuu ( idx_states , : ) ) ;
G_x_x_u = kron ( G_x_x , oo . dr . ghu ( idx_states , : ) ) ;
G_x_u_u = kron ( G_x_u , oo . dr . ghu ( idx_states , : ) ) ;
G_u_u_u = kron ( G_u_u , oo . dr . ghu ( idx_states , : ) ) ;
aux_ZP_x_xu_Z_xx_u = Z_x_xu * P_x_xu + Z_xx_u ;
aux_ZP_xu_u_Z_x_uu = Z_xu_u * P_xu_u + Z_x_uu ;
aux_GP_x_xu_G_xx_u = G_x_xu * P_x_xu + G_xx_u ;
2019-12-20 16:28:06 +01:00
aux_GP_xu_u_G_x_uu = G_xu_u * P_xu_u + G_x_uu ;
Added parameter derivatives of perturbation solution up to 3 order
# Preliminary comments
I finished the identification toolbox at orders two and three using the pruned state space system, but before I merge request this, I decided to first merge the new functionality to compute parameter derivatives of perturbation solution matrices at higher orders. So after this is approved, I merge the identification toolbox.
I guess @rattoma, @sebastien, and @michel are best choices to review this.
I outline the main idea first and then provide some more detailed changes I made to the functions.
***
# Main idea
This merge request is concerned with the *analytical*computation of the parameter derivatives of first, second and third order perturbation solution matrices, i.e. using _closed-form_ expressions to efficiently compute the derivative of $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ *with respect to model parameters* $\theta$. Note that $\theta$ contains model parameters, stderr and corr parameters of shocks. stderr and corr parameters of measurement errors are not yet supported, (they can easily be included as exogenous shocks). The availability of such derivatives is beneficial in terms of more reliable analysis of model sensitivity and parameter identifiability as well as more efficient estimation methods, in particular for models solved up to third order, as it is well-known that numerical derivatives are a tricky business, especially for large models.
References for my approach are:
* Iskrev (2008, 2010) and Schmitt-Grohé and Uribe (2012, Appendix) who were the first to compute the parameter derivatives analytically at first order, however, using inefficient (sparse) Kronecker products.
* Mutschler (2015) who provides the expressions for a second-order, but again using inefficient (sparse) Kronecker products.
* Ratto and Iskrev (2012) who show how the first-order system can be solved accurately, fast and efficiently using existing numerical algorithms for generalized Sylvester equations by taking the parameter derivative with respect to each parameter separately.
* Julliard and Kamenik (2004) who provide the perturbation solution equation system in tensor notation at any order k.
* Levintal (2017) who introduces permutation matrices to express the perturbation solution equation system in matrix notation up to fifth order.
Note that @rattoma already implemented the parameter derivatives of $g_x$ and $g_u$ analytically (and numerically), and I rely heavily on his work in `get_first_order_solution_params_derivs.m` (previously `getH.m`). My additions are mainly to this function and thus it is renamed to `get_perturbation_params_derivs.m`.
The basic idea of this merge request is to take the second- and third-order perturbation solution systems in Julliard and Kamenik (2004), unfold these into an equivalent matrix representation using permutation matrices as in Levintal (2017). Then extending Ratto and Iskrev (2012) one takes the derivative with respect to each parameter separately and gets a computational problem that is linear, albeit large, as it involves either solving generalized Sylvester equations or taking inverses of highly sparse matrices. I will now briefly summarize the perturbation solution system at third order and the system that results when taking the derivative with respect to parameters.
## Perturbation Solution
The following systems arise at first, second, and third order:
$(ghx): f_{x} z_{x} = f_{y_{-}^*} + f_{y_0} g_{x} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{x}= A g_{x} + f_{y_{-}^*}=0$
$(ghu): f_{z} z_{u} = f_{y_0} g_{u} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{u} + f_{u}= A g_u + f_u = 0$
$(ghxx) : A g_{xx} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{x}\right) + f_{zz} \left( z_{x} \otimes z_{x} \right) = 0$
$(ghxu) : A g_{xu} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{x} \otimes z_{u} \right) = 0$
$(ghuu) : A g_{uu} + B g_{xx} \left(g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{u} \otimes z_{u} \right) = 0$
$(ghs2) : (A+B) g_{\sigma\sigma} + \left( f_{y^{**}_{+}y^{**}_{+}} \left(g^{**}_{u} \otimes g^{**}_{u}\right) + f_{y^{**}_{+}} g^{**}_{uu}\right)vec(\Sigma) = 0$
$(ghxxx) : A g_{xxx} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{x}\right) + f_{y_{+}}g^{**}_{xx} \left(g^{*}_x \otimes g^{*}_{xx}\right)P_{x\_xx} + f_{zz} \left( z_{x} \otimes z_{xx} \right)P_{x\_xx} + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{x} \right) = 0$
$(ghxxu) : A g_{xxu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{u} \right) + f_{zz} \left( \left( z_{x} \otimes z_{xu} \right)P_{x\_xu} + \left(z_{xx} \otimes z_{u}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{x} \otimes g^{*}_{xu}\right)P_{x\_xu} + \left(g^{*}_{xx} \otimes g^{*}_{u}\right) \right) = 0$
$(ghxuu) : A g_{xuu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( \left( z_{xu} \otimes z_{u} \right)P_{xu\_u} + \left(z_{x} \otimes z_{uu}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{xu} \otimes g^{*}_{u}\right)P_{xu\_u} + \left(g^{*}_{x} \otimes g^{*}_{uu}\right) \right) = 0$
$(ghuuu) : A g_{uuu} + B g_{xxx} \left(g^{*}_{u} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{u} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( z_{u} \otimes z_{uu} \right)P_{u\_uu} + f_{y_{+}}g^{**}_{xx} \left(g^{*}_{u} \otimes g^{*}_{uu}\right)P_{u\_uu} = 0$
$(ghx\sigma\sigma) : A g_{x\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_x + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{x} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{x} \otimes z_{\sigma\sigma}\right) + F_{xu_{+}u_{+}}\left(I_{n_x} \otimes vec(\Sigma)\right) = 0$
$F_{xu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_x^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{xu_{+}} \otimes z_{u_{+}} \right)P_{xu\_u} + \left(z_{x} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{x} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
$(ghu\sigma\sigma) : A g_{u\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_{u} + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{u} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{u} \otimes z_{\sigma\sigma}\right) + F_{uu_{+}u_{+}}\left(I_{n_u} \otimes vec(\Sigma_u)\right) = 0$
$F_{uu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_u^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{uu_{+}} \otimes z_{u_{+}} \right)P_{uu\_u} + \left(z_{u} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{u} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
A and B are the common perturbation matrices:
$A = f_{y_0} + \begin{pmatrix} \underbrace{0}_{n\times n_{static}} &\vdots& \underbrace{f_{y^{**}_{+}} \cdot g^{**}_{x}}_{n \times n_{spred}} &\vdots& \underbrace{0}_{n\times n_{frwd}} \end{pmatrix}$and $B = \begin{pmatrix} \underbrace{0}_{n \times n_{static}}&\vdots & \underbrace{0}_{n \times n_{pred}} & \vdots & \underbrace{f_{y^{**}_{+}}}_{n \times n_{sfwrd}} \end{pmatrix}$
and $z=(y_{-}^{\ast}; y; y_{+}^{\ast\ast}; u)$ denotes the dynamic model variables as in `M_.lead_lag_incidence`, $y^\ast$ denote state variables, $y^{\ast\ast}$ denote forward looking variables, $y_+$ denote the variables with a lead, $y_{-}$ denote variables with a lag, $y_0$ denote variables at period t, $f$ the model equations, and $f_z$ the first-order dynamic model derivatives, $f_{zz}$ the second-order dynamic derivatives, and $f_{zzz}$ the third-order dynamic model derivatives. Then:
$z_{x} = \begin{pmatrix}I\\g_{x}\\g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{u} =\begin{pmatrix}0\\g_{u}\\g^{**}_{x} \cdot g^{*}_{u}\\I\end{pmatrix}$, $z_{u_{+}} =\begin{pmatrix}0\\0\\g^{**}_{u}\\0\end{pmatrix}$
$z_{xx} = \begin{pmatrix} 0\\g_{xx}\\g^{**}_{x} \left( g^{*}_x \otimes g^{*}_{x} \right) + g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{xu} =\begin{pmatrix}0\\g_{xu}\\g^{**}_{xx} \left( g^{*}_x \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{xu}\\0\end{pmatrix}$, $z_{uu} =\begin{pmatrix}0\\g_{uu}\\g^{**}_{xx} \left( g^{*}_u \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{uu}\\0\end{pmatrix}$,
$z_{xu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_x \otimes I \right)\\0\end{pmatrix}$, $z_{uu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_{u} \otimes I \right)\\0\end{pmatrix}$, $z_{u_{+}u_{+}} =\begin{pmatrix}0\\0\\g^{\ast\ast}_{uu}\\0\end{pmatrix}$, $z_{\sigma\sigma} = \begin{pmatrix}0\\ g_{\sigma\sigma}\\ g^{\ast\ast}_{x}g^{\ast}_{\sigma\sigma} + g^{\ast\ast}_{\sigma\sigma}\\0 \end{pmatrix}$
$P$ are permutation matrices that can be computed using Matlab's `ipermute` function.
## Parameter derivatives of perturbation solutions
First, we need the parameter derivatives of first, second, third, and fourth derivatives of the dynamic model (i.e. g1,g2,g3,g4 in dynamic files), I make use of the implicit function theorem: Let $f_{z^k}$ denote the kth derivative (wrt all dynamic variables) of the dynamic model, then let $df_{z^k}$ denote the first-derivative (wrt all model parameters) of $f_{z^k}$ evaluated at the steady state. Note that $f_{z^k}$ is a function of both the model parameters $\theta$ and of the steady state of all dynamic variables $\bar{z}$, which also depend on the parameters. Hence, implicitly $f_{z^k}=f_{z^k}(\theta,\bar{z}(\theta))$ and $df_{z^k}$ consists of two parts:
1. direct derivative wrt to all model parameters given by the preprocessor in the `_params_derivs.m` files
2. contribution of derivative of steady state of dynamic variables (wrt all model parameters): $f_{z^{k+1}} \cdot d\bar{z}$
Note that we already have functionality to compute $d\bar{z}$ analytically.
Having this, the above perturbation systems are basically equations of the following types
$AX +BXC = RHS$ or $AX = RHS$
Now when taking the derivative (wrt to one single parameter $\theta_j$), we get
$A\mathrm{d}\{X\} + B\mathrm{d}\{X\}C = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X - \mathrm{d}\{B\}XC - BX\mathrm{d}\{C\}$
or
$A\mathrm{d}\{X\} = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X$
The first one is a Sylvester type equation, the second one can be solved by taking the inverse of $A$. The only diffculty and tedious work arrises in computing (the highly sparse) derivatives of $RHS$.
***
# New functions: `
## get_perturbation_params_derivs.m`and `get_perturbation_params_derivs_numerical_objective.m`
* The parameter derivatives up to third order are computed in the new function`get_perturbation_params_derivs.m` both analytically and numerically. For numerical derivatives `get_perturbation_params_derivs_numerical_objective.m` is the objective for `fjaco.m` or `hessian_sparse.m` or `hessian.m`.
* `get_perturbation_params_derivs.m` is basically an extended version of the previous `get_first_order_solution_params_derivs.m` function.
* * `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It is used for numerical derivatives, whenever `analytic_derivation_mode=-1|-2`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
* Detailed changes:
* Most important: notation of this function is now in accordance to the k_order_solver, i.e. we do not compute derivatives of Kalman transition matrices A and B, but rather the solution matrices ghx,ghu,ghxx,ghxu,ghuu,ghs2,ghxxx,ghxxu,ghxuu,ghuuu,ghxss,ghuss in accordance with notation used in `oo_.dr`. As a byproduct at first-order, focusing on ghx and ghu instead of Kalman transition matrices A and B makes the computations slightly faster for large models (e.g. for Quest the computations were faster by a couple of seconds, not much, but okay).
* Removed use of `kstate`, see also Dynare/dynare#1653 and Dynare/dynare!1656
* Output arguments are stored in a structure `DERIVS`, there is also a flag `d2flag` that computes parameter hessians needed only in `dsge_likelihood.m`.
* Removed `kronflag` as input. `options_.analytic_derivation_mode` is now used instead of `kronflag`.
* Removed `indvar`, an index that was used to selected specific variables in the derivatives. This is not needed, as we always compute the parameter derivatives for all variables first and then select a subset of variables. The selection now takes place in other functions, like `dsge_likelihood.m`.
* Introduced some checks: (i) deterministic exogenous variables are not supported, (ii) Kronecker method only compatible with first-order approximation so reset to sylvester method, (iii) for purely backward or forward models we need to be careful with the rows in `M_.lead_la g_incidence`, (iv) if `_params_derivs.m` files are missing an error is thrown.
* For numerical derivatives, if mod file does not contain an `estimated_params_block`, a temporary one with the most important parameter information is created.
## `unfold_g4.m`
* When evaluating g3 and g4 one needs to take into account that these do not contain symmetric elements, so one needs to use `unfold_g3.m` and the new function `unfold_g4.m`. This returns an unfolded version of the same matrix (i.e. with symmetric elements).
***
# New test models
`.gitignore` and `Makefile.am` are changed accordingly. Also now it is possible to run test suite on analytic_derivatives, i.e. run `make check m/analytic_derivatives`
## `analytic_derivatives/BrockMirman_PertParamsDerivs.mod`
* This is the Brock Mirman model, where we know the exact policy function $g$ for capital and consumption. As this does not imply a nonzero $g_{\sigma\sigma}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ I added some artificial equations to get nonzero solution matrices with respect to $\sigma$. The true perturbation solution matrices $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ are then computed analytically with Matlab's symbolic toolbox and saved in `nBrockMirmanSYM.mat`. There is a preprocessor flag that recreates these analytical computations if changes are needed (and to check whether I made some errors here ;-) )
* Then solution matrices up to third order and their parameter Jacobians are then compared to the ones computed by Dynare's `k_order_solver` and by `get_perturbation_params_derivs` for all `analytic_derivation_mode`'s. There will be an error if the maximum absolute deviation is too large, i.e. for numerical derivatives (`analytic_derivation_mode=-1|-2`) the tolerance is choosen lower (around 1e-5); for analytical methods we are stricter: around 1e-13 for first-order, 1e-12 for second order, and 1e-11 for third-order.
* As a side note, this mod file also checks Dynare's `k_order_solver` algorithm and throws an error if something is wrong.
* This test model shows that the new functionality works well. And analytical derivatives perform way better and accurate than numerical ones, even for this small model.
## `analytic_derivatives/burnside_3_order_PertParamsDerivs.mod`
* This builds upon `tests/k_order_perturbation/burnside_k_order.mod` and computes the true parameter derivatives analytically by hand.
* This test model also shows that the new functionality works well.
## `analytic_derivatives/LindeTrabandt2019.mod`
* Shows that the new functionality also works for medium-sized models, i.e. a SW type model solved at third order with 35 variables (11 states). 2 shocks and 20 parameters.
* This mod file can be used to tweak the speed of the computations in the future.
* Compares numerical versus analytical parameter derivatives (for first, second and third order). Note that this model clearly shows that numerical ones are quite different than analytical ones even at first order!
## `identification/LindeTrabandt2019_xfail.mod`
* This model is a check for issue Dynare/dynare#1595, see fjaco.m below, and will fail.
* Removed `analytic_derivatives/ls2003.mod` as this mod file is neither in the testsuite nor does it work.
***
# Detailed changes in other functions
## `get_first_order_solution_params_derivs.m`
* Deleted, or actually, renamed to `get_perturbation_params_derivs.m`, as this function now is able to compute the derivatives up to third order
## `identification_numerical_objective.m`
* `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
## `dsge_likelihood.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `T`matrix anymore, but instead of the dynare solution matrix `ghx`. So we recreate `T` here as this amounts to adding some zeros and focusing on selected variables only.
* Added some checks to make sure the first-order approximation is selected.
* Removed `kron_flag` as input, as `get_perturbation_params_derivs` looks into `options_.analytic_derivation_mode` for `kron_flag`.
## `dynare_identification.m`
* make sure that setting `analytic_derivation_mode` is set both in `options_ident` and `options_`. Note that at the end of the function we restore the `options_` structure, so all changes are local. In a next merge request, I will remove the global variables to make all variables local.
## `get_identification_jacobians.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `A` and `B` matrix anymore, but instead of the dynare solution matrix `ghx` and `ghu`. So we recreate these matrices here instead of in `get_perturbation_params_derivs.m`.
* Added `str2func` for better function handles in `fjaco.m`.
## `fjaco.m`
* make `tol`an option, which can be adjusted by changing `options_.dynatol.x`for identification and parameter derivatives purposes.
* include a check and an informative error message, if numerical derivatives (two-sided finite difference method) yield errors in `resol.m` for identification and parameter derivatives purposes. This closes issue Dynare/dynare#1595.
* Changed year of copyright to 2010-2017,2019
***
# Further suggestions and questions
* Ones this is merged, I will merge request an improvement of the identification toolbox, which will work up to third order using the pruned state space. This will also remove some issues and bugs, and also I will remove global variables in this request.
* The third-order derivatives can be further improved by taking sparsity into account and use mex versions for kronecker products etc. I leave this for further testing (and if anybody actually uses this ;-) )
2019-12-17 19:17:09 +01:00
dghxxu = zeros ( M . endo_nbr , M . nspred ^2 * M . exo_nbr , modparam_nbr ) ;
dghxuu = zeros ( M . endo_nbr , M . nspred * M . exo_nbr ^2 , modparam_nbr ) ;
dghuuu = zeros ( M . endo_nbr , M . exo_nbr ^3 , modparam_nbr ) ;
2019-12-20 16:28:06 +01:00
Added parameter derivatives of perturbation solution up to 3 order
# Preliminary comments
I finished the identification toolbox at orders two and three using the pruned state space system, but before I merge request this, I decided to first merge the new functionality to compute parameter derivatives of perturbation solution matrices at higher orders. So after this is approved, I merge the identification toolbox.
I guess @rattoma, @sebastien, and @michel are best choices to review this.
I outline the main idea first and then provide some more detailed changes I made to the functions.
***
# Main idea
This merge request is concerned with the *analytical*computation of the parameter derivatives of first, second and third order perturbation solution matrices, i.e. using _closed-form_ expressions to efficiently compute the derivative of $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ *with respect to model parameters* $\theta$. Note that $\theta$ contains model parameters, stderr and corr parameters of shocks. stderr and corr parameters of measurement errors are not yet supported, (they can easily be included as exogenous shocks). The availability of such derivatives is beneficial in terms of more reliable analysis of model sensitivity and parameter identifiability as well as more efficient estimation methods, in particular for models solved up to third order, as it is well-known that numerical derivatives are a tricky business, especially for large models.
References for my approach are:
* Iskrev (2008, 2010) and Schmitt-Grohé and Uribe (2012, Appendix) who were the first to compute the parameter derivatives analytically at first order, however, using inefficient (sparse) Kronecker products.
* Mutschler (2015) who provides the expressions for a second-order, but again using inefficient (sparse) Kronecker products.
* Ratto and Iskrev (2012) who show how the first-order system can be solved accurately, fast and efficiently using existing numerical algorithms for generalized Sylvester equations by taking the parameter derivative with respect to each parameter separately.
* Julliard and Kamenik (2004) who provide the perturbation solution equation system in tensor notation at any order k.
* Levintal (2017) who introduces permutation matrices to express the perturbation solution equation system in matrix notation up to fifth order.
Note that @rattoma already implemented the parameter derivatives of $g_x$ and $g_u$ analytically (and numerically), and I rely heavily on his work in `get_first_order_solution_params_derivs.m` (previously `getH.m`). My additions are mainly to this function and thus it is renamed to `get_perturbation_params_derivs.m`.
The basic idea of this merge request is to take the second- and third-order perturbation solution systems in Julliard and Kamenik (2004), unfold these into an equivalent matrix representation using permutation matrices as in Levintal (2017). Then extending Ratto and Iskrev (2012) one takes the derivative with respect to each parameter separately and gets a computational problem that is linear, albeit large, as it involves either solving generalized Sylvester equations or taking inverses of highly sparse matrices. I will now briefly summarize the perturbation solution system at third order and the system that results when taking the derivative with respect to parameters.
## Perturbation Solution
The following systems arise at first, second, and third order:
$(ghx): f_{x} z_{x} = f_{y_{-}^*} + f_{y_0} g_{x} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{x}= A g_{x} + f_{y_{-}^*}=0$
$(ghu): f_{z} z_{u} = f_{y_0} g_{u} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{u} + f_{u}= A g_u + f_u = 0$
$(ghxx) : A g_{xx} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{x}\right) + f_{zz} \left( z_{x} \otimes z_{x} \right) = 0$
$(ghxu) : A g_{xu} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{x} \otimes z_{u} \right) = 0$
$(ghuu) : A g_{uu} + B g_{xx} \left(g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{u} \otimes z_{u} \right) = 0$
$(ghs2) : (A+B) g_{\sigma\sigma} + \left( f_{y^{**}_{+}y^{**}_{+}} \left(g^{**}_{u} \otimes g^{**}_{u}\right) + f_{y^{**}_{+}} g^{**}_{uu}\right)vec(\Sigma) = 0$
$(ghxxx) : A g_{xxx} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{x}\right) + f_{y_{+}}g^{**}_{xx} \left(g^{*}_x \otimes g^{*}_{xx}\right)P_{x\_xx} + f_{zz} \left( z_{x} \otimes z_{xx} \right)P_{x\_xx} + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{x} \right) = 0$
$(ghxxu) : A g_{xxu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{u} \right) + f_{zz} \left( \left( z_{x} \otimes z_{xu} \right)P_{x\_xu} + \left(z_{xx} \otimes z_{u}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{x} \otimes g^{*}_{xu}\right)P_{x\_xu} + \left(g^{*}_{xx} \otimes g^{*}_{u}\right) \right) = 0$
$(ghxuu) : A g_{xuu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( \left( z_{xu} \otimes z_{u} \right)P_{xu\_u} + \left(z_{x} \otimes z_{uu}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{xu} \otimes g^{*}_{u}\right)P_{xu\_u} + \left(g^{*}_{x} \otimes g^{*}_{uu}\right) \right) = 0$
$(ghuuu) : A g_{uuu} + B g_{xxx} \left(g^{*}_{u} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{u} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( z_{u} \otimes z_{uu} \right)P_{u\_uu} + f_{y_{+}}g^{**}_{xx} \left(g^{*}_{u} \otimes g^{*}_{uu}\right)P_{u\_uu} = 0$
$(ghx\sigma\sigma) : A g_{x\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_x + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{x} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{x} \otimes z_{\sigma\sigma}\right) + F_{xu_{+}u_{+}}\left(I_{n_x} \otimes vec(\Sigma)\right) = 0$
$F_{xu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_x^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{xu_{+}} \otimes z_{u_{+}} \right)P_{xu\_u} + \left(z_{x} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{x} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
$(ghu\sigma\sigma) : A g_{u\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_{u} + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{u} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{u} \otimes z_{\sigma\sigma}\right) + F_{uu_{+}u_{+}}\left(I_{n_u} \otimes vec(\Sigma_u)\right) = 0$
$F_{uu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_u^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{uu_{+}} \otimes z_{u_{+}} \right)P_{uu\_u} + \left(z_{u} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{u} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
A and B are the common perturbation matrices:
$A = f_{y_0} + \begin{pmatrix} \underbrace{0}_{n\times n_{static}} &\vdots& \underbrace{f_{y^{**}_{+}} \cdot g^{**}_{x}}_{n \times n_{spred}} &\vdots& \underbrace{0}_{n\times n_{frwd}} \end{pmatrix}$and $B = \begin{pmatrix} \underbrace{0}_{n \times n_{static}}&\vdots & \underbrace{0}_{n \times n_{pred}} & \vdots & \underbrace{f_{y^{**}_{+}}}_{n \times n_{sfwrd}} \end{pmatrix}$
and $z=(y_{-}^{\ast}; y; y_{+}^{\ast\ast}; u)$ denotes the dynamic model variables as in `M_.lead_lag_incidence`, $y^\ast$ denote state variables, $y^{\ast\ast}$ denote forward looking variables, $y_+$ denote the variables with a lead, $y_{-}$ denote variables with a lag, $y_0$ denote variables at period t, $f$ the model equations, and $f_z$ the first-order dynamic model derivatives, $f_{zz}$ the second-order dynamic derivatives, and $f_{zzz}$ the third-order dynamic model derivatives. Then:
$z_{x} = \begin{pmatrix}I\\g_{x}\\g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{u} =\begin{pmatrix}0\\g_{u}\\g^{**}_{x} \cdot g^{*}_{u}\\I\end{pmatrix}$, $z_{u_{+}} =\begin{pmatrix}0\\0\\g^{**}_{u}\\0\end{pmatrix}$
$z_{xx} = \begin{pmatrix} 0\\g_{xx}\\g^{**}_{x} \left( g^{*}_x \otimes g^{*}_{x} \right) + g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{xu} =\begin{pmatrix}0\\g_{xu}\\g^{**}_{xx} \left( g^{*}_x \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{xu}\\0\end{pmatrix}$, $z_{uu} =\begin{pmatrix}0\\g_{uu}\\g^{**}_{xx} \left( g^{*}_u \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{uu}\\0\end{pmatrix}$,
$z_{xu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_x \otimes I \right)\\0\end{pmatrix}$, $z_{uu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_{u} \otimes I \right)\\0\end{pmatrix}$, $z_{u_{+}u_{+}} =\begin{pmatrix}0\\0\\g^{\ast\ast}_{uu}\\0\end{pmatrix}$, $z_{\sigma\sigma} = \begin{pmatrix}0\\ g_{\sigma\sigma}\\ g^{\ast\ast}_{x}g^{\ast}_{\sigma\sigma} + g^{\ast\ast}_{\sigma\sigma}\\0 \end{pmatrix}$
$P$ are permutation matrices that can be computed using Matlab's `ipermute` function.
## Parameter derivatives of perturbation solutions
First, we need the parameter derivatives of first, second, third, and fourth derivatives of the dynamic model (i.e. g1,g2,g3,g4 in dynamic files), I make use of the implicit function theorem: Let $f_{z^k}$ denote the kth derivative (wrt all dynamic variables) of the dynamic model, then let $df_{z^k}$ denote the first-derivative (wrt all model parameters) of $f_{z^k}$ evaluated at the steady state. Note that $f_{z^k}$ is a function of both the model parameters $\theta$ and of the steady state of all dynamic variables $\bar{z}$, which also depend on the parameters. Hence, implicitly $f_{z^k}=f_{z^k}(\theta,\bar{z}(\theta))$ and $df_{z^k}$ consists of two parts:
1. direct derivative wrt to all model parameters given by the preprocessor in the `_params_derivs.m` files
2. contribution of derivative of steady state of dynamic variables (wrt all model parameters): $f_{z^{k+1}} \cdot d\bar{z}$
Note that we already have functionality to compute $d\bar{z}$ analytically.
Having this, the above perturbation systems are basically equations of the following types
$AX +BXC = RHS$ or $AX = RHS$
Now when taking the derivative (wrt to one single parameter $\theta_j$), we get
$A\mathrm{d}\{X\} + B\mathrm{d}\{X\}C = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X - \mathrm{d}\{B\}XC - BX\mathrm{d}\{C\}$
or
$A\mathrm{d}\{X\} = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X$
The first one is a Sylvester type equation, the second one can be solved by taking the inverse of $A$. The only diffculty and tedious work arrises in computing (the highly sparse) derivatives of $RHS$.
***
# New functions: `
## get_perturbation_params_derivs.m`and `get_perturbation_params_derivs_numerical_objective.m`
* The parameter derivatives up to third order are computed in the new function`get_perturbation_params_derivs.m` both analytically and numerically. For numerical derivatives `get_perturbation_params_derivs_numerical_objective.m` is the objective for `fjaco.m` or `hessian_sparse.m` or `hessian.m`.
* `get_perturbation_params_derivs.m` is basically an extended version of the previous `get_first_order_solution_params_derivs.m` function.
* * `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It is used for numerical derivatives, whenever `analytic_derivation_mode=-1|-2`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
* Detailed changes:
* Most important: notation of this function is now in accordance to the k_order_solver, i.e. we do not compute derivatives of Kalman transition matrices A and B, but rather the solution matrices ghx,ghu,ghxx,ghxu,ghuu,ghs2,ghxxx,ghxxu,ghxuu,ghuuu,ghxss,ghuss in accordance with notation used in `oo_.dr`. As a byproduct at first-order, focusing on ghx and ghu instead of Kalman transition matrices A and B makes the computations slightly faster for large models (e.g. for Quest the computations were faster by a couple of seconds, not much, but okay).
* Removed use of `kstate`, see also Dynare/dynare#1653 and Dynare/dynare!1656
* Output arguments are stored in a structure `DERIVS`, there is also a flag `d2flag` that computes parameter hessians needed only in `dsge_likelihood.m`.
* Removed `kronflag` as input. `options_.analytic_derivation_mode` is now used instead of `kronflag`.
* Removed `indvar`, an index that was used to selected specific variables in the derivatives. This is not needed, as we always compute the parameter derivatives for all variables first and then select a subset of variables. The selection now takes place in other functions, like `dsge_likelihood.m`.
* Introduced some checks: (i) deterministic exogenous variables are not supported, (ii) Kronecker method only compatible with first-order approximation so reset to sylvester method, (iii) for purely backward or forward models we need to be careful with the rows in `M_.lead_la g_incidence`, (iv) if `_params_derivs.m` files are missing an error is thrown.
* For numerical derivatives, if mod file does not contain an `estimated_params_block`, a temporary one with the most important parameter information is created.
## `unfold_g4.m`
* When evaluating g3 and g4 one needs to take into account that these do not contain symmetric elements, so one needs to use `unfold_g3.m` and the new function `unfold_g4.m`. This returns an unfolded version of the same matrix (i.e. with symmetric elements).
***
# New test models
`.gitignore` and `Makefile.am` are changed accordingly. Also now it is possible to run test suite on analytic_derivatives, i.e. run `make check m/analytic_derivatives`
## `analytic_derivatives/BrockMirman_PertParamsDerivs.mod`
* This is the Brock Mirman model, where we know the exact policy function $g$ for capital and consumption. As this does not imply a nonzero $g_{\sigma\sigma}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ I added some artificial equations to get nonzero solution matrices with respect to $\sigma$. The true perturbation solution matrices $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ are then computed analytically with Matlab's symbolic toolbox and saved in `nBrockMirmanSYM.mat`. There is a preprocessor flag that recreates these analytical computations if changes are needed (and to check whether I made some errors here ;-) )
* Then solution matrices up to third order and their parameter Jacobians are then compared to the ones computed by Dynare's `k_order_solver` and by `get_perturbation_params_derivs` for all `analytic_derivation_mode`'s. There will be an error if the maximum absolute deviation is too large, i.e. for numerical derivatives (`analytic_derivation_mode=-1|-2`) the tolerance is choosen lower (around 1e-5); for analytical methods we are stricter: around 1e-13 for first-order, 1e-12 for second order, and 1e-11 for third-order.
* As a side note, this mod file also checks Dynare's `k_order_solver` algorithm and throws an error if something is wrong.
* This test model shows that the new functionality works well. And analytical derivatives perform way better and accurate than numerical ones, even for this small model.
## `analytic_derivatives/burnside_3_order_PertParamsDerivs.mod`
* This builds upon `tests/k_order_perturbation/burnside_k_order.mod` and computes the true parameter derivatives analytically by hand.
* This test model also shows that the new functionality works well.
## `analytic_derivatives/LindeTrabandt2019.mod`
* Shows that the new functionality also works for medium-sized models, i.e. a SW type model solved at third order with 35 variables (11 states). 2 shocks and 20 parameters.
* This mod file can be used to tweak the speed of the computations in the future.
* Compares numerical versus analytical parameter derivatives (for first, second and third order). Note that this model clearly shows that numerical ones are quite different than analytical ones even at first order!
## `identification/LindeTrabandt2019_xfail.mod`
* This model is a check for issue Dynare/dynare#1595, see fjaco.m below, and will fail.
* Removed `analytic_derivatives/ls2003.mod` as this mod file is neither in the testsuite nor does it work.
***
# Detailed changes in other functions
## `get_first_order_solution_params_derivs.m`
* Deleted, or actually, renamed to `get_perturbation_params_derivs.m`, as this function now is able to compute the derivatives up to third order
## `identification_numerical_objective.m`
* `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
## `dsge_likelihood.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `T`matrix anymore, but instead of the dynare solution matrix `ghx`. So we recreate `T` here as this amounts to adding some zeros and focusing on selected variables only.
* Added some checks to make sure the first-order approximation is selected.
* Removed `kron_flag` as input, as `get_perturbation_params_derivs` looks into `options_.analytic_derivation_mode` for `kron_flag`.
## `dynare_identification.m`
* make sure that setting `analytic_derivation_mode` is set both in `options_ident` and `options_`. Note that at the end of the function we restore the `options_` structure, so all changes are local. In a next merge request, I will remove the global variables to make all variables local.
## `get_identification_jacobians.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `A` and `B` matrix anymore, but instead of the dynare solution matrix `ghx` and `ghu`. So we recreate these matrices here instead of in `get_perturbation_params_derivs.m`.
* Added `str2func` for better function handles in `fjaco.m`.
## `fjaco.m`
* make `tol`an option, which can be adjusted by changing `options_.dynatol.x`for identification and parameter derivatives purposes.
* include a check and an informative error message, if numerical derivatives (two-sided finite difference method) yield errors in `resol.m` for identification and parameter derivatives purposes. This closes issue Dynare/dynare#1595.
* Changed year of copyright to 2010-2017,2019
***
# Further suggestions and questions
* Ones this is merged, I will merge request an improvement of the identification toolbox, which will work up to third order using the pruned state space. This will also remove some issues and bugs, and also I will remove global variables in this request.
* The third-order derivatives can be further improved by taking sparsity into account and use mex versions for kronecker products etc. I leave this for further testing (and if anybody actually uses this ;-) )
2019-12-17 19:17:09 +01:00
%stuff for ghxss
zup = [ zeros ( M . nspred , M . exo_nbr ) ;
zeros ( length ( nonzeros ( kcurr ) ) , M . exo_nbr ) ;
oo . dr . ghu ( klead ~= 0 , : ) ;
zeros ( M . exo_nbr , M . exo_nbr ) ] ;
zss = [ zeros ( M . nspred , 1 ) ;
oo . dr . ghs2 ( kcurr ~= 0 , : ) ;
oo . dr . ghs2 ( klead ~= 0 , : ) + oo . dr . ghx ( klead ~= 0 , : ) * oo . dr . ghs2 ( idx_states , : ) ;
zeros ( M . exo_nbr , 1 ) ] ;
zxup = [ zeros ( M . nspred , M . nspred * M . exo_nbr ) ;
zeros ( length ( nonzeros ( kcurr ) ) , M . nspred * M . exo_nbr ) ;
2019-12-20 16:28:06 +01:00
oo . dr . ghxu ( klead ~= 0 , : ) * kron ( oo . dr . ghx ( idx_states , : ) , eye ( M . exo_nbr ) ) ;
Added parameter derivatives of perturbation solution up to 3 order
# Preliminary comments
I finished the identification toolbox at orders two and three using the pruned state space system, but before I merge request this, I decided to first merge the new functionality to compute parameter derivatives of perturbation solution matrices at higher orders. So after this is approved, I merge the identification toolbox.
I guess @rattoma, @sebastien, and @michel are best choices to review this.
I outline the main idea first and then provide some more detailed changes I made to the functions.
***
# Main idea
This merge request is concerned with the *analytical*computation of the parameter derivatives of first, second and third order perturbation solution matrices, i.e. using _closed-form_ expressions to efficiently compute the derivative of $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ *with respect to model parameters* $\theta$. Note that $\theta$ contains model parameters, stderr and corr parameters of shocks. stderr and corr parameters of measurement errors are not yet supported, (they can easily be included as exogenous shocks). The availability of such derivatives is beneficial in terms of more reliable analysis of model sensitivity and parameter identifiability as well as more efficient estimation methods, in particular for models solved up to third order, as it is well-known that numerical derivatives are a tricky business, especially for large models.
References for my approach are:
* Iskrev (2008, 2010) and Schmitt-Grohé and Uribe (2012, Appendix) who were the first to compute the parameter derivatives analytically at first order, however, using inefficient (sparse) Kronecker products.
* Mutschler (2015) who provides the expressions for a second-order, but again using inefficient (sparse) Kronecker products.
* Ratto and Iskrev (2012) who show how the first-order system can be solved accurately, fast and efficiently using existing numerical algorithms for generalized Sylvester equations by taking the parameter derivative with respect to each parameter separately.
* Julliard and Kamenik (2004) who provide the perturbation solution equation system in tensor notation at any order k.
* Levintal (2017) who introduces permutation matrices to express the perturbation solution equation system in matrix notation up to fifth order.
Note that @rattoma already implemented the parameter derivatives of $g_x$ and $g_u$ analytically (and numerically), and I rely heavily on his work in `get_first_order_solution_params_derivs.m` (previously `getH.m`). My additions are mainly to this function and thus it is renamed to `get_perturbation_params_derivs.m`.
The basic idea of this merge request is to take the second- and third-order perturbation solution systems in Julliard and Kamenik (2004), unfold these into an equivalent matrix representation using permutation matrices as in Levintal (2017). Then extending Ratto and Iskrev (2012) one takes the derivative with respect to each parameter separately and gets a computational problem that is linear, albeit large, as it involves either solving generalized Sylvester equations or taking inverses of highly sparse matrices. I will now briefly summarize the perturbation solution system at third order and the system that results when taking the derivative with respect to parameters.
## Perturbation Solution
The following systems arise at first, second, and third order:
$(ghx): f_{x} z_{x} = f_{y_{-}^*} + f_{y_0} g_{x} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{x}= A g_{x} + f_{y_{-}^*}=0$
$(ghu): f_{z} z_{u} = f_{y_0} g_{u} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{u} + f_{u}= A g_u + f_u = 0$
$(ghxx) : A g_{xx} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{x}\right) + f_{zz} \left( z_{x} \otimes z_{x} \right) = 0$
$(ghxu) : A g_{xu} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{x} \otimes z_{u} \right) = 0$
$(ghuu) : A g_{uu} + B g_{xx} \left(g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{u} \otimes z_{u} \right) = 0$
$(ghs2) : (A+B) g_{\sigma\sigma} + \left( f_{y^{**}_{+}y^{**}_{+}} \left(g^{**}_{u} \otimes g^{**}_{u}\right) + f_{y^{**}_{+}} g^{**}_{uu}\right)vec(\Sigma) = 0$
$(ghxxx) : A g_{xxx} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{x}\right) + f_{y_{+}}g^{**}_{xx} \left(g^{*}_x \otimes g^{*}_{xx}\right)P_{x\_xx} + f_{zz} \left( z_{x} \otimes z_{xx} \right)P_{x\_xx} + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{x} \right) = 0$
$(ghxxu) : A g_{xxu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{u} \right) + f_{zz} \left( \left( z_{x} \otimes z_{xu} \right)P_{x\_xu} + \left(z_{xx} \otimes z_{u}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{x} \otimes g^{*}_{xu}\right)P_{x\_xu} + \left(g^{*}_{xx} \otimes g^{*}_{u}\right) \right) = 0$
$(ghxuu) : A g_{xuu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( \left( z_{xu} \otimes z_{u} \right)P_{xu\_u} + \left(z_{x} \otimes z_{uu}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{xu} \otimes g^{*}_{u}\right)P_{xu\_u} + \left(g^{*}_{x} \otimes g^{*}_{uu}\right) \right) = 0$
$(ghuuu) : A g_{uuu} + B g_{xxx} \left(g^{*}_{u} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{u} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( z_{u} \otimes z_{uu} \right)P_{u\_uu} + f_{y_{+}}g^{**}_{xx} \left(g^{*}_{u} \otimes g^{*}_{uu}\right)P_{u\_uu} = 0$
$(ghx\sigma\sigma) : A g_{x\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_x + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{x} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{x} \otimes z_{\sigma\sigma}\right) + F_{xu_{+}u_{+}}\left(I_{n_x} \otimes vec(\Sigma)\right) = 0$
$F_{xu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_x^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{xu_{+}} \otimes z_{u_{+}} \right)P_{xu\_u} + \left(z_{x} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{x} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
$(ghu\sigma\sigma) : A g_{u\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_{u} + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{u} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{u} \otimes z_{\sigma\sigma}\right) + F_{uu_{+}u_{+}}\left(I_{n_u} \otimes vec(\Sigma_u)\right) = 0$
$F_{uu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_u^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{uu_{+}} \otimes z_{u_{+}} \right)P_{uu\_u} + \left(z_{u} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{u} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
A and B are the common perturbation matrices:
$A = f_{y_0} + \begin{pmatrix} \underbrace{0}_{n\times n_{static}} &\vdots& \underbrace{f_{y^{**}_{+}} \cdot g^{**}_{x}}_{n \times n_{spred}} &\vdots& \underbrace{0}_{n\times n_{frwd}} \end{pmatrix}$and $B = \begin{pmatrix} \underbrace{0}_{n \times n_{static}}&\vdots & \underbrace{0}_{n \times n_{pred}} & \vdots & \underbrace{f_{y^{**}_{+}}}_{n \times n_{sfwrd}} \end{pmatrix}$
and $z=(y_{-}^{\ast}; y; y_{+}^{\ast\ast}; u)$ denotes the dynamic model variables as in `M_.lead_lag_incidence`, $y^\ast$ denote state variables, $y^{\ast\ast}$ denote forward looking variables, $y_+$ denote the variables with a lead, $y_{-}$ denote variables with a lag, $y_0$ denote variables at period t, $f$ the model equations, and $f_z$ the first-order dynamic model derivatives, $f_{zz}$ the second-order dynamic derivatives, and $f_{zzz}$ the third-order dynamic model derivatives. Then:
$z_{x} = \begin{pmatrix}I\\g_{x}\\g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{u} =\begin{pmatrix}0\\g_{u}\\g^{**}_{x} \cdot g^{*}_{u}\\I\end{pmatrix}$, $z_{u_{+}} =\begin{pmatrix}0\\0\\g^{**}_{u}\\0\end{pmatrix}$
$z_{xx} = \begin{pmatrix} 0\\g_{xx}\\g^{**}_{x} \left( g^{*}_x \otimes g^{*}_{x} \right) + g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{xu} =\begin{pmatrix}0\\g_{xu}\\g^{**}_{xx} \left( g^{*}_x \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{xu}\\0\end{pmatrix}$, $z_{uu} =\begin{pmatrix}0\\g_{uu}\\g^{**}_{xx} \left( g^{*}_u \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{uu}\\0\end{pmatrix}$,
$z_{xu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_x \otimes I \right)\\0\end{pmatrix}$, $z_{uu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_{u} \otimes I \right)\\0\end{pmatrix}$, $z_{u_{+}u_{+}} =\begin{pmatrix}0\\0\\g^{\ast\ast}_{uu}\\0\end{pmatrix}$, $z_{\sigma\sigma} = \begin{pmatrix}0\\ g_{\sigma\sigma}\\ g^{\ast\ast}_{x}g^{\ast}_{\sigma\sigma} + g^{\ast\ast}_{\sigma\sigma}\\0 \end{pmatrix}$
$P$ are permutation matrices that can be computed using Matlab's `ipermute` function.
## Parameter derivatives of perturbation solutions
First, we need the parameter derivatives of first, second, third, and fourth derivatives of the dynamic model (i.e. g1,g2,g3,g4 in dynamic files), I make use of the implicit function theorem: Let $f_{z^k}$ denote the kth derivative (wrt all dynamic variables) of the dynamic model, then let $df_{z^k}$ denote the first-derivative (wrt all model parameters) of $f_{z^k}$ evaluated at the steady state. Note that $f_{z^k}$ is a function of both the model parameters $\theta$ and of the steady state of all dynamic variables $\bar{z}$, which also depend on the parameters. Hence, implicitly $f_{z^k}=f_{z^k}(\theta,\bar{z}(\theta))$ and $df_{z^k}$ consists of two parts:
1. direct derivative wrt to all model parameters given by the preprocessor in the `_params_derivs.m` files
2. contribution of derivative of steady state of dynamic variables (wrt all model parameters): $f_{z^{k+1}} \cdot d\bar{z}$
Note that we already have functionality to compute $d\bar{z}$ analytically.
Having this, the above perturbation systems are basically equations of the following types
$AX +BXC = RHS$ or $AX = RHS$
Now when taking the derivative (wrt to one single parameter $\theta_j$), we get
$A\mathrm{d}\{X\} + B\mathrm{d}\{X\}C = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X - \mathrm{d}\{B\}XC - BX\mathrm{d}\{C\}$
or
$A\mathrm{d}\{X\} = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X$
The first one is a Sylvester type equation, the second one can be solved by taking the inverse of $A$. The only diffculty and tedious work arrises in computing (the highly sparse) derivatives of $RHS$.
***
# New functions: `
## get_perturbation_params_derivs.m`and `get_perturbation_params_derivs_numerical_objective.m`
* The parameter derivatives up to third order are computed in the new function`get_perturbation_params_derivs.m` both analytically and numerically. For numerical derivatives `get_perturbation_params_derivs_numerical_objective.m` is the objective for `fjaco.m` or `hessian_sparse.m` or `hessian.m`.
* `get_perturbation_params_derivs.m` is basically an extended version of the previous `get_first_order_solution_params_derivs.m` function.
* * `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It is used for numerical derivatives, whenever `analytic_derivation_mode=-1|-2`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
* Detailed changes:
* Most important: notation of this function is now in accordance to the k_order_solver, i.e. we do not compute derivatives of Kalman transition matrices A and B, but rather the solution matrices ghx,ghu,ghxx,ghxu,ghuu,ghs2,ghxxx,ghxxu,ghxuu,ghuuu,ghxss,ghuss in accordance with notation used in `oo_.dr`. As a byproduct at first-order, focusing on ghx and ghu instead of Kalman transition matrices A and B makes the computations slightly faster for large models (e.g. for Quest the computations were faster by a couple of seconds, not much, but okay).
* Removed use of `kstate`, see also Dynare/dynare#1653 and Dynare/dynare!1656
* Output arguments are stored in a structure `DERIVS`, there is also a flag `d2flag` that computes parameter hessians needed only in `dsge_likelihood.m`.
* Removed `kronflag` as input. `options_.analytic_derivation_mode` is now used instead of `kronflag`.
* Removed `indvar`, an index that was used to selected specific variables in the derivatives. This is not needed, as we always compute the parameter derivatives for all variables first and then select a subset of variables. The selection now takes place in other functions, like `dsge_likelihood.m`.
* Introduced some checks: (i) deterministic exogenous variables are not supported, (ii) Kronecker method only compatible with first-order approximation so reset to sylvester method, (iii) for purely backward or forward models we need to be careful with the rows in `M_.lead_la g_incidence`, (iv) if `_params_derivs.m` files are missing an error is thrown.
* For numerical derivatives, if mod file does not contain an `estimated_params_block`, a temporary one with the most important parameter information is created.
## `unfold_g4.m`
* When evaluating g3 and g4 one needs to take into account that these do not contain symmetric elements, so one needs to use `unfold_g3.m` and the new function `unfold_g4.m`. This returns an unfolded version of the same matrix (i.e. with symmetric elements).
***
# New test models
`.gitignore` and `Makefile.am` are changed accordingly. Also now it is possible to run test suite on analytic_derivatives, i.e. run `make check m/analytic_derivatives`
## `analytic_derivatives/BrockMirman_PertParamsDerivs.mod`
* This is the Brock Mirman model, where we know the exact policy function $g$ for capital and consumption. As this does not imply a nonzero $g_{\sigma\sigma}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ I added some artificial equations to get nonzero solution matrices with respect to $\sigma$. The true perturbation solution matrices $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ are then computed analytically with Matlab's symbolic toolbox and saved in `nBrockMirmanSYM.mat`. There is a preprocessor flag that recreates these analytical computations if changes are needed (and to check whether I made some errors here ;-) )
* Then solution matrices up to third order and their parameter Jacobians are then compared to the ones computed by Dynare's `k_order_solver` and by `get_perturbation_params_derivs` for all `analytic_derivation_mode`'s. There will be an error if the maximum absolute deviation is too large, i.e. for numerical derivatives (`analytic_derivation_mode=-1|-2`) the tolerance is choosen lower (around 1e-5); for analytical methods we are stricter: around 1e-13 for first-order, 1e-12 for second order, and 1e-11 for third-order.
* As a side note, this mod file also checks Dynare's `k_order_solver` algorithm and throws an error if something is wrong.
* This test model shows that the new functionality works well. And analytical derivatives perform way better and accurate than numerical ones, even for this small model.
## `analytic_derivatives/burnside_3_order_PertParamsDerivs.mod`
* This builds upon `tests/k_order_perturbation/burnside_k_order.mod` and computes the true parameter derivatives analytically by hand.
* This test model also shows that the new functionality works well.
## `analytic_derivatives/LindeTrabandt2019.mod`
* Shows that the new functionality also works for medium-sized models, i.e. a SW type model solved at third order with 35 variables (11 states). 2 shocks and 20 parameters.
* This mod file can be used to tweak the speed of the computations in the future.
* Compares numerical versus analytical parameter derivatives (for first, second and third order). Note that this model clearly shows that numerical ones are quite different than analytical ones even at first order!
## `identification/LindeTrabandt2019_xfail.mod`
* This model is a check for issue Dynare/dynare#1595, see fjaco.m below, and will fail.
* Removed `analytic_derivatives/ls2003.mod` as this mod file is neither in the testsuite nor does it work.
***
# Detailed changes in other functions
## `get_first_order_solution_params_derivs.m`
* Deleted, or actually, renamed to `get_perturbation_params_derivs.m`, as this function now is able to compute the derivatives up to third order
## `identification_numerical_objective.m`
* `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
## `dsge_likelihood.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `T`matrix anymore, but instead of the dynare solution matrix `ghx`. So we recreate `T` here as this amounts to adding some zeros and focusing on selected variables only.
* Added some checks to make sure the first-order approximation is selected.
* Removed `kron_flag` as input, as `get_perturbation_params_derivs` looks into `options_.analytic_derivation_mode` for `kron_flag`.
## `dynare_identification.m`
* make sure that setting `analytic_derivation_mode` is set both in `options_ident` and `options_`. Note that at the end of the function we restore the `options_` structure, so all changes are local. In a next merge request, I will remove the global variables to make all variables local.
## `get_identification_jacobians.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `A` and `B` matrix anymore, but instead of the dynare solution matrix `ghx` and `ghu`. So we recreate these matrices here instead of in `get_perturbation_params_derivs.m`.
* Added `str2func` for better function handles in `fjaco.m`.
## `fjaco.m`
* make `tol`an option, which can be adjusted by changing `options_.dynatol.x`for identification and parameter derivatives purposes.
* include a check and an informative error message, if numerical derivatives (two-sided finite difference method) yield errors in `resol.m` for identification and parameter derivatives purposes. This closes issue Dynare/dynare#1595.
* Changed year of copyright to 2010-2017,2019
***
# Further suggestions and questions
* Ones this is merged, I will merge request an improvement of the identification toolbox, which will work up to third order using the pruned state space. This will also remove some issues and bugs, and also I will remove global variables in this request.
* The third-order derivatives can be further improved by taking sparsity into account and use mex versions for kronecker products etc. I leave this for further testing (and if anybody actually uses this ;-) )
2019-12-17 19:17:09 +01:00
zeros ( M . exo_nbr , M . nspred * M . exo_nbr ) ] ;
zupup = [ zeros ( M . nspred , M . exo_nbr ^2 ) ;
zeros ( length ( nonzeros ( kcurr ) ) , M . exo_nbr ^2 ) ;
oo . dr . ghuu ( klead ~= 0 , : ) ;
zeros ( M . exo_nbr , M . exo_nbr ^2 ) ] ;
G_x_ss = kron ( oo . dr . ghx ( idx_states , : ) , oo . dr . ghs2 ( idx_states , : ) ) ;
Z_x_ss = kron ( zx , zss ) ;
Z_up_up = kron ( zup , zup ) ;
Z_xup_up = kron ( zxup , zup ) ;
Z_x_upup = kron ( zx , zupup ) ;
2019-12-20 16:28:06 +01:00
Z_x_up_up = kron ( zx , Z_up_up ) ;
aux_ZP_xup_up_Z_x_upup = Z_xup_up * P_xu_u + Z_x_upup ;
Added parameter derivatives of perturbation solution up to 3 order
# Preliminary comments
I finished the identification toolbox at orders two and three using the pruned state space system, but before I merge request this, I decided to first merge the new functionality to compute parameter derivatives of perturbation solution matrices at higher orders. So after this is approved, I merge the identification toolbox.
I guess @rattoma, @sebastien, and @michel are best choices to review this.
I outline the main idea first and then provide some more detailed changes I made to the functions.
***
# Main idea
This merge request is concerned with the *analytical*computation of the parameter derivatives of first, second and third order perturbation solution matrices, i.e. using _closed-form_ expressions to efficiently compute the derivative of $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ *with respect to model parameters* $\theta$. Note that $\theta$ contains model parameters, stderr and corr parameters of shocks. stderr and corr parameters of measurement errors are not yet supported, (they can easily be included as exogenous shocks). The availability of such derivatives is beneficial in terms of more reliable analysis of model sensitivity and parameter identifiability as well as more efficient estimation methods, in particular for models solved up to third order, as it is well-known that numerical derivatives are a tricky business, especially for large models.
References for my approach are:
* Iskrev (2008, 2010) and Schmitt-Grohé and Uribe (2012, Appendix) who were the first to compute the parameter derivatives analytically at first order, however, using inefficient (sparse) Kronecker products.
* Mutschler (2015) who provides the expressions for a second-order, but again using inefficient (sparse) Kronecker products.
* Ratto and Iskrev (2012) who show how the first-order system can be solved accurately, fast and efficiently using existing numerical algorithms for generalized Sylvester equations by taking the parameter derivative with respect to each parameter separately.
* Julliard and Kamenik (2004) who provide the perturbation solution equation system in tensor notation at any order k.
* Levintal (2017) who introduces permutation matrices to express the perturbation solution equation system in matrix notation up to fifth order.
Note that @rattoma already implemented the parameter derivatives of $g_x$ and $g_u$ analytically (and numerically), and I rely heavily on his work in `get_first_order_solution_params_derivs.m` (previously `getH.m`). My additions are mainly to this function and thus it is renamed to `get_perturbation_params_derivs.m`.
The basic idea of this merge request is to take the second- and third-order perturbation solution systems in Julliard and Kamenik (2004), unfold these into an equivalent matrix representation using permutation matrices as in Levintal (2017). Then extending Ratto and Iskrev (2012) one takes the derivative with respect to each parameter separately and gets a computational problem that is linear, albeit large, as it involves either solving generalized Sylvester equations or taking inverses of highly sparse matrices. I will now briefly summarize the perturbation solution system at third order and the system that results when taking the derivative with respect to parameters.
## Perturbation Solution
The following systems arise at first, second, and third order:
$(ghx): f_{x} z_{x} = f_{y_{-}^*} + f_{y_0} g_{x} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{x}= A g_{x} + f_{y_{-}^*}=0$
$(ghu): f_{z} z_{u} = f_{y_0} g_{u} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{u} + f_{u}= A g_u + f_u = 0$
$(ghxx) : A g_{xx} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{x}\right) + f_{zz} \left( z_{x} \otimes z_{x} \right) = 0$
$(ghxu) : A g_{xu} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{x} \otimes z_{u} \right) = 0$
$(ghuu) : A g_{uu} + B g_{xx} \left(g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{u} \otimes z_{u} \right) = 0$
$(ghs2) : (A+B) g_{\sigma\sigma} + \left( f_{y^{**}_{+}y^{**}_{+}} \left(g^{**}_{u} \otimes g^{**}_{u}\right) + f_{y^{**}_{+}} g^{**}_{uu}\right)vec(\Sigma) = 0$
$(ghxxx) : A g_{xxx} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{x}\right) + f_{y_{+}}g^{**}_{xx} \left(g^{*}_x \otimes g^{*}_{xx}\right)P_{x\_xx} + f_{zz} \left( z_{x} \otimes z_{xx} \right)P_{x\_xx} + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{x} \right) = 0$
$(ghxxu) : A g_{xxu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{u} \right) + f_{zz} \left( \left( z_{x} \otimes z_{xu} \right)P_{x\_xu} + \left(z_{xx} \otimes z_{u}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{x} \otimes g^{*}_{xu}\right)P_{x\_xu} + \left(g^{*}_{xx} \otimes g^{*}_{u}\right) \right) = 0$
$(ghxuu) : A g_{xuu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( \left( z_{xu} \otimes z_{u} \right)P_{xu\_u} + \left(z_{x} \otimes z_{uu}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{xu} \otimes g^{*}_{u}\right)P_{xu\_u} + \left(g^{*}_{x} \otimes g^{*}_{uu}\right) \right) = 0$
$(ghuuu) : A g_{uuu} + B g_{xxx} \left(g^{*}_{u} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{u} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( z_{u} \otimes z_{uu} \right)P_{u\_uu} + f_{y_{+}}g^{**}_{xx} \left(g^{*}_{u} \otimes g^{*}_{uu}\right)P_{u\_uu} = 0$
$(ghx\sigma\sigma) : A g_{x\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_x + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{x} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{x} \otimes z_{\sigma\sigma}\right) + F_{xu_{+}u_{+}}\left(I_{n_x} \otimes vec(\Sigma)\right) = 0$
$F_{xu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_x^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{xu_{+}} \otimes z_{u_{+}} \right)P_{xu\_u} + \left(z_{x} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{x} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
$(ghu\sigma\sigma) : A g_{u\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_{u} + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{u} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{u} \otimes z_{\sigma\sigma}\right) + F_{uu_{+}u_{+}}\left(I_{n_u} \otimes vec(\Sigma_u)\right) = 0$
$F_{uu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_u^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{uu_{+}} \otimes z_{u_{+}} \right)P_{uu\_u} + \left(z_{u} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{u} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
A and B are the common perturbation matrices:
$A = f_{y_0} + \begin{pmatrix} \underbrace{0}_{n\times n_{static}} &\vdots& \underbrace{f_{y^{**}_{+}} \cdot g^{**}_{x}}_{n \times n_{spred}} &\vdots& \underbrace{0}_{n\times n_{frwd}} \end{pmatrix}$and $B = \begin{pmatrix} \underbrace{0}_{n \times n_{static}}&\vdots & \underbrace{0}_{n \times n_{pred}} & \vdots & \underbrace{f_{y^{**}_{+}}}_{n \times n_{sfwrd}} \end{pmatrix}$
and $z=(y_{-}^{\ast}; y; y_{+}^{\ast\ast}; u)$ denotes the dynamic model variables as in `M_.lead_lag_incidence`, $y^\ast$ denote state variables, $y^{\ast\ast}$ denote forward looking variables, $y_+$ denote the variables with a lead, $y_{-}$ denote variables with a lag, $y_0$ denote variables at period t, $f$ the model equations, and $f_z$ the first-order dynamic model derivatives, $f_{zz}$ the second-order dynamic derivatives, and $f_{zzz}$ the third-order dynamic model derivatives. Then:
$z_{x} = \begin{pmatrix}I\\g_{x}\\g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{u} =\begin{pmatrix}0\\g_{u}\\g^{**}_{x} \cdot g^{*}_{u}\\I\end{pmatrix}$, $z_{u_{+}} =\begin{pmatrix}0\\0\\g^{**}_{u}\\0\end{pmatrix}$
$z_{xx} = \begin{pmatrix} 0\\g_{xx}\\g^{**}_{x} \left( g^{*}_x \otimes g^{*}_{x} \right) + g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{xu} =\begin{pmatrix}0\\g_{xu}\\g^{**}_{xx} \left( g^{*}_x \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{xu}\\0\end{pmatrix}$, $z_{uu} =\begin{pmatrix}0\\g_{uu}\\g^{**}_{xx} \left( g^{*}_u \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{uu}\\0\end{pmatrix}$,
$z_{xu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_x \otimes I \right)\\0\end{pmatrix}$, $z_{uu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_{u} \otimes I \right)\\0\end{pmatrix}$, $z_{u_{+}u_{+}} =\begin{pmatrix}0\\0\\g^{\ast\ast}_{uu}\\0\end{pmatrix}$, $z_{\sigma\sigma} = \begin{pmatrix}0\\ g_{\sigma\sigma}\\ g^{\ast\ast}_{x}g^{\ast}_{\sigma\sigma} + g^{\ast\ast}_{\sigma\sigma}\\0 \end{pmatrix}$
$P$ are permutation matrices that can be computed using Matlab's `ipermute` function.
## Parameter derivatives of perturbation solutions
First, we need the parameter derivatives of first, second, third, and fourth derivatives of the dynamic model (i.e. g1,g2,g3,g4 in dynamic files), I make use of the implicit function theorem: Let $f_{z^k}$ denote the kth derivative (wrt all dynamic variables) of the dynamic model, then let $df_{z^k}$ denote the first-derivative (wrt all model parameters) of $f_{z^k}$ evaluated at the steady state. Note that $f_{z^k}$ is a function of both the model parameters $\theta$ and of the steady state of all dynamic variables $\bar{z}$, which also depend on the parameters. Hence, implicitly $f_{z^k}=f_{z^k}(\theta,\bar{z}(\theta))$ and $df_{z^k}$ consists of two parts:
1. direct derivative wrt to all model parameters given by the preprocessor in the `_params_derivs.m` files
2. contribution of derivative of steady state of dynamic variables (wrt all model parameters): $f_{z^{k+1}} \cdot d\bar{z}$
Note that we already have functionality to compute $d\bar{z}$ analytically.
Having this, the above perturbation systems are basically equations of the following types
$AX +BXC = RHS$ or $AX = RHS$
Now when taking the derivative (wrt to one single parameter $\theta_j$), we get
$A\mathrm{d}\{X\} + B\mathrm{d}\{X\}C = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X - \mathrm{d}\{B\}XC - BX\mathrm{d}\{C\}$
or
$A\mathrm{d}\{X\} = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X$
The first one is a Sylvester type equation, the second one can be solved by taking the inverse of $A$. The only diffculty and tedious work arrises in computing (the highly sparse) derivatives of $RHS$.
***
# New functions: `
## get_perturbation_params_derivs.m`and `get_perturbation_params_derivs_numerical_objective.m`
* The parameter derivatives up to third order are computed in the new function`get_perturbation_params_derivs.m` both analytically and numerically. For numerical derivatives `get_perturbation_params_derivs_numerical_objective.m` is the objective for `fjaco.m` or `hessian_sparse.m` or `hessian.m`.
* `get_perturbation_params_derivs.m` is basically an extended version of the previous `get_first_order_solution_params_derivs.m` function.
* * `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It is used for numerical derivatives, whenever `analytic_derivation_mode=-1|-2`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
* Detailed changes:
* Most important: notation of this function is now in accordance to the k_order_solver, i.e. we do not compute derivatives of Kalman transition matrices A and B, but rather the solution matrices ghx,ghu,ghxx,ghxu,ghuu,ghs2,ghxxx,ghxxu,ghxuu,ghuuu,ghxss,ghuss in accordance with notation used in `oo_.dr`. As a byproduct at first-order, focusing on ghx and ghu instead of Kalman transition matrices A and B makes the computations slightly faster for large models (e.g. for Quest the computations were faster by a couple of seconds, not much, but okay).
* Removed use of `kstate`, see also Dynare/dynare#1653 and Dynare/dynare!1656
* Output arguments are stored in a structure `DERIVS`, there is also a flag `d2flag` that computes parameter hessians needed only in `dsge_likelihood.m`.
* Removed `kronflag` as input. `options_.analytic_derivation_mode` is now used instead of `kronflag`.
* Removed `indvar`, an index that was used to selected specific variables in the derivatives. This is not needed, as we always compute the parameter derivatives for all variables first and then select a subset of variables. The selection now takes place in other functions, like `dsge_likelihood.m`.
* Introduced some checks: (i) deterministic exogenous variables are not supported, (ii) Kronecker method only compatible with first-order approximation so reset to sylvester method, (iii) for purely backward or forward models we need to be careful with the rows in `M_.lead_la g_incidence`, (iv) if `_params_derivs.m` files are missing an error is thrown.
* For numerical derivatives, if mod file does not contain an `estimated_params_block`, a temporary one with the most important parameter information is created.
## `unfold_g4.m`
* When evaluating g3 and g4 one needs to take into account that these do not contain symmetric elements, so one needs to use `unfold_g3.m` and the new function `unfold_g4.m`. This returns an unfolded version of the same matrix (i.e. with symmetric elements).
***
# New test models
`.gitignore` and `Makefile.am` are changed accordingly. Also now it is possible to run test suite on analytic_derivatives, i.e. run `make check m/analytic_derivatives`
## `analytic_derivatives/BrockMirman_PertParamsDerivs.mod`
* This is the Brock Mirman model, where we know the exact policy function $g$ for capital and consumption. As this does not imply a nonzero $g_{\sigma\sigma}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ I added some artificial equations to get nonzero solution matrices with respect to $\sigma$. The true perturbation solution matrices $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ are then computed analytically with Matlab's symbolic toolbox and saved in `nBrockMirmanSYM.mat`. There is a preprocessor flag that recreates these analytical computations if changes are needed (and to check whether I made some errors here ;-) )
* Then solution matrices up to third order and their parameter Jacobians are then compared to the ones computed by Dynare's `k_order_solver` and by `get_perturbation_params_derivs` for all `analytic_derivation_mode`'s. There will be an error if the maximum absolute deviation is too large, i.e. for numerical derivatives (`analytic_derivation_mode=-1|-2`) the tolerance is choosen lower (around 1e-5); for analytical methods we are stricter: around 1e-13 for first-order, 1e-12 for second order, and 1e-11 for third-order.
* As a side note, this mod file also checks Dynare's `k_order_solver` algorithm and throws an error if something is wrong.
* This test model shows that the new functionality works well. And analytical derivatives perform way better and accurate than numerical ones, even for this small model.
## `analytic_derivatives/burnside_3_order_PertParamsDerivs.mod`
* This builds upon `tests/k_order_perturbation/burnside_k_order.mod` and computes the true parameter derivatives analytically by hand.
* This test model also shows that the new functionality works well.
## `analytic_derivatives/LindeTrabandt2019.mod`
* Shows that the new functionality also works for medium-sized models, i.e. a SW type model solved at third order with 35 variables (11 states). 2 shocks and 20 parameters.
* This mod file can be used to tweak the speed of the computations in the future.
* Compares numerical versus analytical parameter derivatives (for first, second and third order). Note that this model clearly shows that numerical ones are quite different than analytical ones even at first order!
## `identification/LindeTrabandt2019_xfail.mod`
* This model is a check for issue Dynare/dynare#1595, see fjaco.m below, and will fail.
* Removed `analytic_derivatives/ls2003.mod` as this mod file is neither in the testsuite nor does it work.
***
# Detailed changes in other functions
## `get_first_order_solution_params_derivs.m`
* Deleted, or actually, renamed to `get_perturbation_params_derivs.m`, as this function now is able to compute the derivatives up to third order
## `identification_numerical_objective.m`
* `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
## `dsge_likelihood.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `T`matrix anymore, but instead of the dynare solution matrix `ghx`. So we recreate `T` here as this amounts to adding some zeros and focusing on selected variables only.
* Added some checks to make sure the first-order approximation is selected.
* Removed `kron_flag` as input, as `get_perturbation_params_derivs` looks into `options_.analytic_derivation_mode` for `kron_flag`.
## `dynare_identification.m`
* make sure that setting `analytic_derivation_mode` is set both in `options_ident` and `options_`. Note that at the end of the function we restore the `options_` structure, so all changes are local. In a next merge request, I will remove the global variables to make all variables local.
## `get_identification_jacobians.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `A` and `B` matrix anymore, but instead of the dynare solution matrix `ghx` and `ghu`. So we recreate these matrices here instead of in `get_perturbation_params_derivs.m`.
* Added `str2func` for better function handles in `fjaco.m`.
## `fjaco.m`
* make `tol`an option, which can be adjusted by changing `options_.dynatol.x`for identification and parameter derivatives purposes.
* include a check and an informative error message, if numerical derivatives (two-sided finite difference method) yield errors in `resol.m` for identification and parameter derivatives purposes. This closes issue Dynare/dynare#1595.
* Changed year of copyright to 2010-2017,2019
***
# Further suggestions and questions
* Ones this is merged, I will merge request an improvement of the identification toolbox, which will work up to third order using the pruned state space. This will also remove some issues and bugs, and also I will remove global variables in this request.
* The third-order derivatives can be further improved by taking sparsity into account and use mex versions for kronecker products etc. I leave this for further testing (and if anybody actually uses this ;-) )
2019-12-17 19:17:09 +01:00
Fxupup = g3 ( : , k3yy0ex0 ( kyy0ex0 , kyy0ex0 , kyy0ex0 ) ) * Z_x_up_up + g2 ( : , k2yy0ex0 ( kyy0ex0 , kyy0ex0 ) ) * aux_ZP_xup_up_Z_x_upup + g1 ( : , nonzeros ( klead ) ) * oo . dr . ghxuu ( klead ~= 0 , : ) * kron ( oo . dr . ghx ( idx_states , : ) , eye ( M . exo_nbr ^2 ) ) ;
Ix_vecSig_e = kron ( speye ( M . nspred ) , M . Sigma_e ( : ) ) ;
dRHSxss = zeros ( M . endo_nbr , M . nspred , totparam_nbr ) ;
%stuff for ghuss
zuup = [ zeros ( M . nspred , M . exo_nbr ^2 ) ;
zeros ( length ( nonzeros ( kcurr ) ) , M . exo_nbr ^2 ) ;
oo . dr . ghxu ( klead ~= 0 , : ) * kron ( oo . dr . ghu ( idx_states , : ) , eye ( M . exo_nbr ) ) ;
zeros ( M . exo_nbr , M . exo_nbr ^2 ) ] ;
G_u_ss = kron ( oo . dr . ghu ( idx_states , : ) , oo . dr . ghs2 ( idx_states , : ) ) ;
Z_u_ss = kron ( zu , zss ) ;
Z_u_upup = kron ( zu , zupup ) ;
Z_uup_up = kron ( zuup , zup ) ;
2019-12-20 16:28:06 +01:00
Z_u_up_up = kron ( zu , Z_up_up ) ;
Added parameter derivatives of perturbation solution up to 3 order
# Preliminary comments
I finished the identification toolbox at orders two and three using the pruned state space system, but before I merge request this, I decided to first merge the new functionality to compute parameter derivatives of perturbation solution matrices at higher orders. So after this is approved, I merge the identification toolbox.
I guess @rattoma, @sebastien, and @michel are best choices to review this.
I outline the main idea first and then provide some more detailed changes I made to the functions.
***
# Main idea
This merge request is concerned with the *analytical*computation of the parameter derivatives of first, second and third order perturbation solution matrices, i.e. using _closed-form_ expressions to efficiently compute the derivative of $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ *with respect to model parameters* $\theta$. Note that $\theta$ contains model parameters, stderr and corr parameters of shocks. stderr and corr parameters of measurement errors are not yet supported, (they can easily be included as exogenous shocks). The availability of such derivatives is beneficial in terms of more reliable analysis of model sensitivity and parameter identifiability as well as more efficient estimation methods, in particular for models solved up to third order, as it is well-known that numerical derivatives are a tricky business, especially for large models.
References for my approach are:
* Iskrev (2008, 2010) and Schmitt-Grohé and Uribe (2012, Appendix) who were the first to compute the parameter derivatives analytically at first order, however, using inefficient (sparse) Kronecker products.
* Mutschler (2015) who provides the expressions for a second-order, but again using inefficient (sparse) Kronecker products.
* Ratto and Iskrev (2012) who show how the first-order system can be solved accurately, fast and efficiently using existing numerical algorithms for generalized Sylvester equations by taking the parameter derivative with respect to each parameter separately.
* Julliard and Kamenik (2004) who provide the perturbation solution equation system in tensor notation at any order k.
* Levintal (2017) who introduces permutation matrices to express the perturbation solution equation system in matrix notation up to fifth order.
Note that @rattoma already implemented the parameter derivatives of $g_x$ and $g_u$ analytically (and numerically), and I rely heavily on his work in `get_first_order_solution_params_derivs.m` (previously `getH.m`). My additions are mainly to this function and thus it is renamed to `get_perturbation_params_derivs.m`.
The basic idea of this merge request is to take the second- and third-order perturbation solution systems in Julliard and Kamenik (2004), unfold these into an equivalent matrix representation using permutation matrices as in Levintal (2017). Then extending Ratto and Iskrev (2012) one takes the derivative with respect to each parameter separately and gets a computational problem that is linear, albeit large, as it involves either solving generalized Sylvester equations or taking inverses of highly sparse matrices. I will now briefly summarize the perturbation solution system at third order and the system that results when taking the derivative with respect to parameters.
## Perturbation Solution
The following systems arise at first, second, and third order:
$(ghx): f_{x} z_{x} = f_{y_{-}^*} + f_{y_0} g_{x} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{x}= A g_{x} + f_{y_{-}^*}=0$
$(ghu): f_{z} z_{u} = f_{y_0} g_{u} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{u} + f_{u}= A g_u + f_u = 0$
$(ghxx) : A g_{xx} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{x}\right) + f_{zz} \left( z_{x} \otimes z_{x} \right) = 0$
$(ghxu) : A g_{xu} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{x} \otimes z_{u} \right) = 0$
$(ghuu) : A g_{uu} + B g_{xx} \left(g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{u} \otimes z_{u} \right) = 0$
$(ghs2) : (A+B) g_{\sigma\sigma} + \left( f_{y^{**}_{+}y^{**}_{+}} \left(g^{**}_{u} \otimes g^{**}_{u}\right) + f_{y^{**}_{+}} g^{**}_{uu}\right)vec(\Sigma) = 0$
$(ghxxx) : A g_{xxx} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{x}\right) + f_{y_{+}}g^{**}_{xx} \left(g^{*}_x \otimes g^{*}_{xx}\right)P_{x\_xx} + f_{zz} \left( z_{x} \otimes z_{xx} \right)P_{x\_xx} + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{x} \right) = 0$
$(ghxxu) : A g_{xxu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{u} \right) + f_{zz} \left( \left( z_{x} \otimes z_{xu} \right)P_{x\_xu} + \left(z_{xx} \otimes z_{u}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{x} \otimes g^{*}_{xu}\right)P_{x\_xu} + \left(g^{*}_{xx} \otimes g^{*}_{u}\right) \right) = 0$
$(ghxuu) : A g_{xuu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( \left( z_{xu} \otimes z_{u} \right)P_{xu\_u} + \left(z_{x} \otimes z_{uu}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{xu} \otimes g^{*}_{u}\right)P_{xu\_u} + \left(g^{*}_{x} \otimes g^{*}_{uu}\right) \right) = 0$
$(ghuuu) : A g_{uuu} + B g_{xxx} \left(g^{*}_{u} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{u} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( z_{u} \otimes z_{uu} \right)P_{u\_uu} + f_{y_{+}}g^{**}_{xx} \left(g^{*}_{u} \otimes g^{*}_{uu}\right)P_{u\_uu} = 0$
$(ghx\sigma\sigma) : A g_{x\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_x + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{x} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{x} \otimes z_{\sigma\sigma}\right) + F_{xu_{+}u_{+}}\left(I_{n_x} \otimes vec(\Sigma)\right) = 0$
$F_{xu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_x^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{xu_{+}} \otimes z_{u_{+}} \right)P_{xu\_u} + \left(z_{x} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{x} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
$(ghu\sigma\sigma) : A g_{u\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_{u} + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{u} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{u} \otimes z_{\sigma\sigma}\right) + F_{uu_{+}u_{+}}\left(I_{n_u} \otimes vec(\Sigma_u)\right) = 0$
$F_{uu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_u^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{uu_{+}} \otimes z_{u_{+}} \right)P_{uu\_u} + \left(z_{u} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{u} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
A and B are the common perturbation matrices:
$A = f_{y_0} + \begin{pmatrix} \underbrace{0}_{n\times n_{static}} &\vdots& \underbrace{f_{y^{**}_{+}} \cdot g^{**}_{x}}_{n \times n_{spred}} &\vdots& \underbrace{0}_{n\times n_{frwd}} \end{pmatrix}$and $B = \begin{pmatrix} \underbrace{0}_{n \times n_{static}}&\vdots & \underbrace{0}_{n \times n_{pred}} & \vdots & \underbrace{f_{y^{**}_{+}}}_{n \times n_{sfwrd}} \end{pmatrix}$
and $z=(y_{-}^{\ast}; y; y_{+}^{\ast\ast}; u)$ denotes the dynamic model variables as in `M_.lead_lag_incidence`, $y^\ast$ denote state variables, $y^{\ast\ast}$ denote forward looking variables, $y_+$ denote the variables with a lead, $y_{-}$ denote variables with a lag, $y_0$ denote variables at period t, $f$ the model equations, and $f_z$ the first-order dynamic model derivatives, $f_{zz}$ the second-order dynamic derivatives, and $f_{zzz}$ the third-order dynamic model derivatives. Then:
$z_{x} = \begin{pmatrix}I\\g_{x}\\g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{u} =\begin{pmatrix}0\\g_{u}\\g^{**}_{x} \cdot g^{*}_{u}\\I\end{pmatrix}$, $z_{u_{+}} =\begin{pmatrix}0\\0\\g^{**}_{u}\\0\end{pmatrix}$
$z_{xx} = \begin{pmatrix} 0\\g_{xx}\\g^{**}_{x} \left( g^{*}_x \otimes g^{*}_{x} \right) + g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{xu} =\begin{pmatrix}0\\g_{xu}\\g^{**}_{xx} \left( g^{*}_x \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{xu}\\0\end{pmatrix}$, $z_{uu} =\begin{pmatrix}0\\g_{uu}\\g^{**}_{xx} \left( g^{*}_u \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{uu}\\0\end{pmatrix}$,
$z_{xu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_x \otimes I \right)\\0\end{pmatrix}$, $z_{uu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_{u} \otimes I \right)\\0\end{pmatrix}$, $z_{u_{+}u_{+}} =\begin{pmatrix}0\\0\\g^{\ast\ast}_{uu}\\0\end{pmatrix}$, $z_{\sigma\sigma} = \begin{pmatrix}0\\ g_{\sigma\sigma}\\ g^{\ast\ast}_{x}g^{\ast}_{\sigma\sigma} + g^{\ast\ast}_{\sigma\sigma}\\0 \end{pmatrix}$
$P$ are permutation matrices that can be computed using Matlab's `ipermute` function.
## Parameter derivatives of perturbation solutions
First, we need the parameter derivatives of first, second, third, and fourth derivatives of the dynamic model (i.e. g1,g2,g3,g4 in dynamic files), I make use of the implicit function theorem: Let $f_{z^k}$ denote the kth derivative (wrt all dynamic variables) of the dynamic model, then let $df_{z^k}$ denote the first-derivative (wrt all model parameters) of $f_{z^k}$ evaluated at the steady state. Note that $f_{z^k}$ is a function of both the model parameters $\theta$ and of the steady state of all dynamic variables $\bar{z}$, which also depend on the parameters. Hence, implicitly $f_{z^k}=f_{z^k}(\theta,\bar{z}(\theta))$ and $df_{z^k}$ consists of two parts:
1. direct derivative wrt to all model parameters given by the preprocessor in the `_params_derivs.m` files
2. contribution of derivative of steady state of dynamic variables (wrt all model parameters): $f_{z^{k+1}} \cdot d\bar{z}$
Note that we already have functionality to compute $d\bar{z}$ analytically.
Having this, the above perturbation systems are basically equations of the following types
$AX +BXC = RHS$ or $AX = RHS$
Now when taking the derivative (wrt to one single parameter $\theta_j$), we get
$A\mathrm{d}\{X\} + B\mathrm{d}\{X\}C = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X - \mathrm{d}\{B\}XC - BX\mathrm{d}\{C\}$
or
$A\mathrm{d}\{X\} = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X$
The first one is a Sylvester type equation, the second one can be solved by taking the inverse of $A$. The only diffculty and tedious work arrises in computing (the highly sparse) derivatives of $RHS$.
***
# New functions: `
## get_perturbation_params_derivs.m`and `get_perturbation_params_derivs_numerical_objective.m`
* The parameter derivatives up to third order are computed in the new function`get_perturbation_params_derivs.m` both analytically and numerically. For numerical derivatives `get_perturbation_params_derivs_numerical_objective.m` is the objective for `fjaco.m` or `hessian_sparse.m` or `hessian.m`.
* `get_perturbation_params_derivs.m` is basically an extended version of the previous `get_first_order_solution_params_derivs.m` function.
* * `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It is used for numerical derivatives, whenever `analytic_derivation_mode=-1|-2`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
* Detailed changes:
* Most important: notation of this function is now in accordance to the k_order_solver, i.e. we do not compute derivatives of Kalman transition matrices A and B, but rather the solution matrices ghx,ghu,ghxx,ghxu,ghuu,ghs2,ghxxx,ghxxu,ghxuu,ghuuu,ghxss,ghuss in accordance with notation used in `oo_.dr`. As a byproduct at first-order, focusing on ghx and ghu instead of Kalman transition matrices A and B makes the computations slightly faster for large models (e.g. for Quest the computations were faster by a couple of seconds, not much, but okay).
* Removed use of `kstate`, see also Dynare/dynare#1653 and Dynare/dynare!1656
* Output arguments are stored in a structure `DERIVS`, there is also a flag `d2flag` that computes parameter hessians needed only in `dsge_likelihood.m`.
* Removed `kronflag` as input. `options_.analytic_derivation_mode` is now used instead of `kronflag`.
* Removed `indvar`, an index that was used to selected specific variables in the derivatives. This is not needed, as we always compute the parameter derivatives for all variables first and then select a subset of variables. The selection now takes place in other functions, like `dsge_likelihood.m`.
* Introduced some checks: (i) deterministic exogenous variables are not supported, (ii) Kronecker method only compatible with first-order approximation so reset to sylvester method, (iii) for purely backward or forward models we need to be careful with the rows in `M_.lead_la g_incidence`, (iv) if `_params_derivs.m` files are missing an error is thrown.
* For numerical derivatives, if mod file does not contain an `estimated_params_block`, a temporary one with the most important parameter information is created.
## `unfold_g4.m`
* When evaluating g3 and g4 one needs to take into account that these do not contain symmetric elements, so one needs to use `unfold_g3.m` and the new function `unfold_g4.m`. This returns an unfolded version of the same matrix (i.e. with symmetric elements).
***
# New test models
`.gitignore` and `Makefile.am` are changed accordingly. Also now it is possible to run test suite on analytic_derivatives, i.e. run `make check m/analytic_derivatives`
## `analytic_derivatives/BrockMirman_PertParamsDerivs.mod`
* This is the Brock Mirman model, where we know the exact policy function $g$ for capital and consumption. As this does not imply a nonzero $g_{\sigma\sigma}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ I added some artificial equations to get nonzero solution matrices with respect to $\sigma$. The true perturbation solution matrices $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ are then computed analytically with Matlab's symbolic toolbox and saved in `nBrockMirmanSYM.mat`. There is a preprocessor flag that recreates these analytical computations if changes are needed (and to check whether I made some errors here ;-) )
* Then solution matrices up to third order and their parameter Jacobians are then compared to the ones computed by Dynare's `k_order_solver` and by `get_perturbation_params_derivs` for all `analytic_derivation_mode`'s. There will be an error if the maximum absolute deviation is too large, i.e. for numerical derivatives (`analytic_derivation_mode=-1|-2`) the tolerance is choosen lower (around 1e-5); for analytical methods we are stricter: around 1e-13 for first-order, 1e-12 for second order, and 1e-11 for third-order.
* As a side note, this mod file also checks Dynare's `k_order_solver` algorithm and throws an error if something is wrong.
* This test model shows that the new functionality works well. And analytical derivatives perform way better and accurate than numerical ones, even for this small model.
## `analytic_derivatives/burnside_3_order_PertParamsDerivs.mod`
* This builds upon `tests/k_order_perturbation/burnside_k_order.mod` and computes the true parameter derivatives analytically by hand.
* This test model also shows that the new functionality works well.
## `analytic_derivatives/LindeTrabandt2019.mod`
* Shows that the new functionality also works for medium-sized models, i.e. a SW type model solved at third order with 35 variables (11 states). 2 shocks and 20 parameters.
* This mod file can be used to tweak the speed of the computations in the future.
* Compares numerical versus analytical parameter derivatives (for first, second and third order). Note that this model clearly shows that numerical ones are quite different than analytical ones even at first order!
## `identification/LindeTrabandt2019_xfail.mod`
* This model is a check for issue Dynare/dynare#1595, see fjaco.m below, and will fail.
* Removed `analytic_derivatives/ls2003.mod` as this mod file is neither in the testsuite nor does it work.
***
# Detailed changes in other functions
## `get_first_order_solution_params_derivs.m`
* Deleted, or actually, renamed to `get_perturbation_params_derivs.m`, as this function now is able to compute the derivatives up to third order
## `identification_numerical_objective.m`
* `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
## `dsge_likelihood.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `T`matrix anymore, but instead of the dynare solution matrix `ghx`. So we recreate `T` here as this amounts to adding some zeros and focusing on selected variables only.
* Added some checks to make sure the first-order approximation is selected.
* Removed `kron_flag` as input, as `get_perturbation_params_derivs` looks into `options_.analytic_derivation_mode` for `kron_flag`.
## `dynare_identification.m`
* make sure that setting `analytic_derivation_mode` is set both in `options_ident` and `options_`. Note that at the end of the function we restore the `options_` structure, so all changes are local. In a next merge request, I will remove the global variables to make all variables local.
## `get_identification_jacobians.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `A` and `B` matrix anymore, but instead of the dynare solution matrix `ghx` and `ghu`. So we recreate these matrices here instead of in `get_perturbation_params_derivs.m`.
* Added `str2func` for better function handles in `fjaco.m`.
## `fjaco.m`
* make `tol`an option, which can be adjusted by changing `options_.dynatol.x`for identification and parameter derivatives purposes.
* include a check and an informative error message, if numerical derivatives (two-sided finite difference method) yield errors in `resol.m` for identification and parameter derivatives purposes. This closes issue Dynare/dynare#1595.
* Changed year of copyright to 2010-2017,2019
***
# Further suggestions and questions
* Ones this is merged, I will merge request an improvement of the identification toolbox, which will work up to third order using the pruned state space. This will also remove some issues and bugs, and also I will remove global variables in this request.
* The third-order derivatives can be further improved by taking sparsity into account and use mex versions for kronecker products etc. I leave this for further testing (and if anybody actually uses this ;-) )
2019-12-17 19:17:09 +01:00
aux_ZP_uup_up_Z_u_upup = Z_uup_up * P_uu_u + Z_u_upup ;
Fuupup = g3 ( : , k3yy0ex0 ( kyy0ex0 , kyy0ex0 , kyy0ex0 ) ) * Z_u_up_up + g2 ( : , k2yy0ex0 ( kyy0ex0 , kyy0ex0 ) ) * aux_ZP_uup_up_Z_u_upup + g1 ( : , nonzeros ( klead ) ) * oo . dr . ghxuu ( klead ~= 0 , : ) * kron ( oo . dr . ghu ( idx_states , : ) , eye ( M . exo_nbr ^2 ) ) ;
2019-12-20 16:28:06 +01:00
Iu_vecSig_e = kron ( speye ( M . exo_nbr ) , M . Sigma_e ( : ) ) ;
Added parameter derivatives of perturbation solution up to 3 order
# Preliminary comments
I finished the identification toolbox at orders two and three using the pruned state space system, but before I merge request this, I decided to first merge the new functionality to compute parameter derivatives of perturbation solution matrices at higher orders. So after this is approved, I merge the identification toolbox.
I guess @rattoma, @sebastien, and @michel are best choices to review this.
I outline the main idea first and then provide some more detailed changes I made to the functions.
***
# Main idea
This merge request is concerned with the *analytical*computation of the parameter derivatives of first, second and third order perturbation solution matrices, i.e. using _closed-form_ expressions to efficiently compute the derivative of $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ *with respect to model parameters* $\theta$. Note that $\theta$ contains model parameters, stderr and corr parameters of shocks. stderr and corr parameters of measurement errors are not yet supported, (they can easily be included as exogenous shocks). The availability of such derivatives is beneficial in terms of more reliable analysis of model sensitivity and parameter identifiability as well as more efficient estimation methods, in particular for models solved up to third order, as it is well-known that numerical derivatives are a tricky business, especially for large models.
References for my approach are:
* Iskrev (2008, 2010) and Schmitt-Grohé and Uribe (2012, Appendix) who were the first to compute the parameter derivatives analytically at first order, however, using inefficient (sparse) Kronecker products.
* Mutschler (2015) who provides the expressions for a second-order, but again using inefficient (sparse) Kronecker products.
* Ratto and Iskrev (2012) who show how the first-order system can be solved accurately, fast and efficiently using existing numerical algorithms for generalized Sylvester equations by taking the parameter derivative with respect to each parameter separately.
* Julliard and Kamenik (2004) who provide the perturbation solution equation system in tensor notation at any order k.
* Levintal (2017) who introduces permutation matrices to express the perturbation solution equation system in matrix notation up to fifth order.
Note that @rattoma already implemented the parameter derivatives of $g_x$ and $g_u$ analytically (and numerically), and I rely heavily on his work in `get_first_order_solution_params_derivs.m` (previously `getH.m`). My additions are mainly to this function and thus it is renamed to `get_perturbation_params_derivs.m`.
The basic idea of this merge request is to take the second- and third-order perturbation solution systems in Julliard and Kamenik (2004), unfold these into an equivalent matrix representation using permutation matrices as in Levintal (2017). Then extending Ratto and Iskrev (2012) one takes the derivative with respect to each parameter separately and gets a computational problem that is linear, albeit large, as it involves either solving generalized Sylvester equations or taking inverses of highly sparse matrices. I will now briefly summarize the perturbation solution system at third order and the system that results when taking the derivative with respect to parameters.
## Perturbation Solution
The following systems arise at first, second, and third order:
$(ghx): f_{x} z_{x} = f_{y_{-}^*} + f_{y_0} g_{x} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{x}= A g_{x} + f_{y_{-}^*}=0$
$(ghu): f_{z} z_{u} = f_{y_0} g_{u} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{u} + f_{u}= A g_u + f_u = 0$
$(ghxx) : A g_{xx} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{x}\right) + f_{zz} \left( z_{x} \otimes z_{x} \right) = 0$
$(ghxu) : A g_{xu} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{x} \otimes z_{u} \right) = 0$
$(ghuu) : A g_{uu} + B g_{xx} \left(g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{u} \otimes z_{u} \right) = 0$
$(ghs2) : (A+B) g_{\sigma\sigma} + \left( f_{y^{**}_{+}y^{**}_{+}} \left(g^{**}_{u} \otimes g^{**}_{u}\right) + f_{y^{**}_{+}} g^{**}_{uu}\right)vec(\Sigma) = 0$
$(ghxxx) : A g_{xxx} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{x}\right) + f_{y_{+}}g^{**}_{xx} \left(g^{*}_x \otimes g^{*}_{xx}\right)P_{x\_xx} + f_{zz} \left( z_{x} \otimes z_{xx} \right)P_{x\_xx} + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{x} \right) = 0$
$(ghxxu) : A g_{xxu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{u} \right) + f_{zz} \left( \left( z_{x} \otimes z_{xu} \right)P_{x\_xu} + \left(z_{xx} \otimes z_{u}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{x} \otimes g^{*}_{xu}\right)P_{x\_xu} + \left(g^{*}_{xx} \otimes g^{*}_{u}\right) \right) = 0$
$(ghxuu) : A g_{xuu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( \left( z_{xu} \otimes z_{u} \right)P_{xu\_u} + \left(z_{x} \otimes z_{uu}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{xu} \otimes g^{*}_{u}\right)P_{xu\_u} + \left(g^{*}_{x} \otimes g^{*}_{uu}\right) \right) = 0$
$(ghuuu) : A g_{uuu} + B g_{xxx} \left(g^{*}_{u} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{u} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( z_{u} \otimes z_{uu} \right)P_{u\_uu} + f_{y_{+}}g^{**}_{xx} \left(g^{*}_{u} \otimes g^{*}_{uu}\right)P_{u\_uu} = 0$
$(ghx\sigma\sigma) : A g_{x\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_x + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{x} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{x} \otimes z_{\sigma\sigma}\right) + F_{xu_{+}u_{+}}\left(I_{n_x} \otimes vec(\Sigma)\right) = 0$
$F_{xu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_x^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{xu_{+}} \otimes z_{u_{+}} \right)P_{xu\_u} + \left(z_{x} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{x} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
$(ghu\sigma\sigma) : A g_{u\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_{u} + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{u} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{u} \otimes z_{\sigma\sigma}\right) + F_{uu_{+}u_{+}}\left(I_{n_u} \otimes vec(\Sigma_u)\right) = 0$
$F_{uu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_u^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{uu_{+}} \otimes z_{u_{+}} \right)P_{uu\_u} + \left(z_{u} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{u} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
A and B are the common perturbation matrices:
$A = f_{y_0} + \begin{pmatrix} \underbrace{0}_{n\times n_{static}} &\vdots& \underbrace{f_{y^{**}_{+}} \cdot g^{**}_{x}}_{n \times n_{spred}} &\vdots& \underbrace{0}_{n\times n_{frwd}} \end{pmatrix}$and $B = \begin{pmatrix} \underbrace{0}_{n \times n_{static}}&\vdots & \underbrace{0}_{n \times n_{pred}} & \vdots & \underbrace{f_{y^{**}_{+}}}_{n \times n_{sfwrd}} \end{pmatrix}$
and $z=(y_{-}^{\ast}; y; y_{+}^{\ast\ast}; u)$ denotes the dynamic model variables as in `M_.lead_lag_incidence`, $y^\ast$ denote state variables, $y^{\ast\ast}$ denote forward looking variables, $y_+$ denote the variables with a lead, $y_{-}$ denote variables with a lag, $y_0$ denote variables at period t, $f$ the model equations, and $f_z$ the first-order dynamic model derivatives, $f_{zz}$ the second-order dynamic derivatives, and $f_{zzz}$ the third-order dynamic model derivatives. Then:
$z_{x} = \begin{pmatrix}I\\g_{x}\\g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{u} =\begin{pmatrix}0\\g_{u}\\g^{**}_{x} \cdot g^{*}_{u}\\I\end{pmatrix}$, $z_{u_{+}} =\begin{pmatrix}0\\0\\g^{**}_{u}\\0\end{pmatrix}$
$z_{xx} = \begin{pmatrix} 0\\g_{xx}\\g^{**}_{x} \left( g^{*}_x \otimes g^{*}_{x} \right) + g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{xu} =\begin{pmatrix}0\\g_{xu}\\g^{**}_{xx} \left( g^{*}_x \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{xu}\\0\end{pmatrix}$, $z_{uu} =\begin{pmatrix}0\\g_{uu}\\g^{**}_{xx} \left( g^{*}_u \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{uu}\\0\end{pmatrix}$,
$z_{xu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_x \otimes I \right)\\0\end{pmatrix}$, $z_{uu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_{u} \otimes I \right)\\0\end{pmatrix}$, $z_{u_{+}u_{+}} =\begin{pmatrix}0\\0\\g^{\ast\ast}_{uu}\\0\end{pmatrix}$, $z_{\sigma\sigma} = \begin{pmatrix}0\\ g_{\sigma\sigma}\\ g^{\ast\ast}_{x}g^{\ast}_{\sigma\sigma} + g^{\ast\ast}_{\sigma\sigma}\\0 \end{pmatrix}$
$P$ are permutation matrices that can be computed using Matlab's `ipermute` function.
## Parameter derivatives of perturbation solutions
First, we need the parameter derivatives of first, second, third, and fourth derivatives of the dynamic model (i.e. g1,g2,g3,g4 in dynamic files), I make use of the implicit function theorem: Let $f_{z^k}$ denote the kth derivative (wrt all dynamic variables) of the dynamic model, then let $df_{z^k}$ denote the first-derivative (wrt all model parameters) of $f_{z^k}$ evaluated at the steady state. Note that $f_{z^k}$ is a function of both the model parameters $\theta$ and of the steady state of all dynamic variables $\bar{z}$, which also depend on the parameters. Hence, implicitly $f_{z^k}=f_{z^k}(\theta,\bar{z}(\theta))$ and $df_{z^k}$ consists of two parts:
1. direct derivative wrt to all model parameters given by the preprocessor in the `_params_derivs.m` files
2. contribution of derivative of steady state of dynamic variables (wrt all model parameters): $f_{z^{k+1}} \cdot d\bar{z}$
Note that we already have functionality to compute $d\bar{z}$ analytically.
Having this, the above perturbation systems are basically equations of the following types
$AX +BXC = RHS$ or $AX = RHS$
Now when taking the derivative (wrt to one single parameter $\theta_j$), we get
$A\mathrm{d}\{X\} + B\mathrm{d}\{X\}C = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X - \mathrm{d}\{B\}XC - BX\mathrm{d}\{C\}$
or
$A\mathrm{d}\{X\} = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X$
The first one is a Sylvester type equation, the second one can be solved by taking the inverse of $A$. The only diffculty and tedious work arrises in computing (the highly sparse) derivatives of $RHS$.
***
# New functions: `
## get_perturbation_params_derivs.m`and `get_perturbation_params_derivs_numerical_objective.m`
* The parameter derivatives up to third order are computed in the new function`get_perturbation_params_derivs.m` both analytically and numerically. For numerical derivatives `get_perturbation_params_derivs_numerical_objective.m` is the objective for `fjaco.m` or `hessian_sparse.m` or `hessian.m`.
* `get_perturbation_params_derivs.m` is basically an extended version of the previous `get_first_order_solution_params_derivs.m` function.
* * `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It is used for numerical derivatives, whenever `analytic_derivation_mode=-1|-2`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
* Detailed changes:
* Most important: notation of this function is now in accordance to the k_order_solver, i.e. we do not compute derivatives of Kalman transition matrices A and B, but rather the solution matrices ghx,ghu,ghxx,ghxu,ghuu,ghs2,ghxxx,ghxxu,ghxuu,ghuuu,ghxss,ghuss in accordance with notation used in `oo_.dr`. As a byproduct at first-order, focusing on ghx and ghu instead of Kalman transition matrices A and B makes the computations slightly faster for large models (e.g. for Quest the computations were faster by a couple of seconds, not much, but okay).
* Removed use of `kstate`, see also Dynare/dynare#1653 and Dynare/dynare!1656
* Output arguments are stored in a structure `DERIVS`, there is also a flag `d2flag` that computes parameter hessians needed only in `dsge_likelihood.m`.
* Removed `kronflag` as input. `options_.analytic_derivation_mode` is now used instead of `kronflag`.
* Removed `indvar`, an index that was used to selected specific variables in the derivatives. This is not needed, as we always compute the parameter derivatives for all variables first and then select a subset of variables. The selection now takes place in other functions, like `dsge_likelihood.m`.
* Introduced some checks: (i) deterministic exogenous variables are not supported, (ii) Kronecker method only compatible with first-order approximation so reset to sylvester method, (iii) for purely backward or forward models we need to be careful with the rows in `M_.lead_la g_incidence`, (iv) if `_params_derivs.m` files are missing an error is thrown.
* For numerical derivatives, if mod file does not contain an `estimated_params_block`, a temporary one with the most important parameter information is created.
## `unfold_g4.m`
* When evaluating g3 and g4 one needs to take into account that these do not contain symmetric elements, so one needs to use `unfold_g3.m` and the new function `unfold_g4.m`. This returns an unfolded version of the same matrix (i.e. with symmetric elements).
***
# New test models
`.gitignore` and `Makefile.am` are changed accordingly. Also now it is possible to run test suite on analytic_derivatives, i.e. run `make check m/analytic_derivatives`
## `analytic_derivatives/BrockMirman_PertParamsDerivs.mod`
* This is the Brock Mirman model, where we know the exact policy function $g$ for capital and consumption. As this does not imply a nonzero $g_{\sigma\sigma}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ I added some artificial equations to get nonzero solution matrices with respect to $\sigma$. The true perturbation solution matrices $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ are then computed analytically with Matlab's symbolic toolbox and saved in `nBrockMirmanSYM.mat`. There is a preprocessor flag that recreates these analytical computations if changes are needed (and to check whether I made some errors here ;-) )
* Then solution matrices up to third order and their parameter Jacobians are then compared to the ones computed by Dynare's `k_order_solver` and by `get_perturbation_params_derivs` for all `analytic_derivation_mode`'s. There will be an error if the maximum absolute deviation is too large, i.e. for numerical derivatives (`analytic_derivation_mode=-1|-2`) the tolerance is choosen lower (around 1e-5); for analytical methods we are stricter: around 1e-13 for first-order, 1e-12 for second order, and 1e-11 for third-order.
* As a side note, this mod file also checks Dynare's `k_order_solver` algorithm and throws an error if something is wrong.
* This test model shows that the new functionality works well. And analytical derivatives perform way better and accurate than numerical ones, even for this small model.
## `analytic_derivatives/burnside_3_order_PertParamsDerivs.mod`
* This builds upon `tests/k_order_perturbation/burnside_k_order.mod` and computes the true parameter derivatives analytically by hand.
* This test model also shows that the new functionality works well.
## `analytic_derivatives/LindeTrabandt2019.mod`
* Shows that the new functionality also works for medium-sized models, i.e. a SW type model solved at third order with 35 variables (11 states). 2 shocks and 20 parameters.
* This mod file can be used to tweak the speed of the computations in the future.
* Compares numerical versus analytical parameter derivatives (for first, second and third order). Note that this model clearly shows that numerical ones are quite different than analytical ones even at first order!
## `identification/LindeTrabandt2019_xfail.mod`
* This model is a check for issue Dynare/dynare#1595, see fjaco.m below, and will fail.
* Removed `analytic_derivatives/ls2003.mod` as this mod file is neither in the testsuite nor does it work.
***
# Detailed changes in other functions
## `get_first_order_solution_params_derivs.m`
* Deleted, or actually, renamed to `get_perturbation_params_derivs.m`, as this function now is able to compute the derivatives up to third order
## `identification_numerical_objective.m`
* `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
## `dsge_likelihood.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `T`matrix anymore, but instead of the dynare solution matrix `ghx`. So we recreate `T` here as this amounts to adding some zeros and focusing on selected variables only.
* Added some checks to make sure the first-order approximation is selected.
* Removed `kron_flag` as input, as `get_perturbation_params_derivs` looks into `options_.analytic_derivation_mode` for `kron_flag`.
## `dynare_identification.m`
* make sure that setting `analytic_derivation_mode` is set both in `options_ident` and `options_`. Note that at the end of the function we restore the `options_` structure, so all changes are local. In a next merge request, I will remove the global variables to make all variables local.
## `get_identification_jacobians.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `A` and `B` matrix anymore, but instead of the dynare solution matrix `ghx` and `ghu`. So we recreate these matrices here instead of in `get_perturbation_params_derivs.m`.
* Added `str2func` for better function handles in `fjaco.m`.
## `fjaco.m`
* make `tol`an option, which can be adjusted by changing `options_.dynatol.x`for identification and parameter derivatives purposes.
* include a check and an informative error message, if numerical derivatives (two-sided finite difference method) yield errors in `resol.m` for identification and parameter derivatives purposes. This closes issue Dynare/dynare#1595.
* Changed year of copyright to 2010-2017,2019
***
# Further suggestions and questions
* Ones this is merged, I will merge request an improvement of the identification toolbox, which will work up to third order using the pruned state space. This will also remove some issues and bugs, and also I will remove global variables in this request.
* The third-order derivatives can be further improved by taking sparsity into account and use mex versions for kronecker products etc. I leave this for further testing (and if anybody actually uses this ;-) )
2019-12-17 19:17:09 +01:00
dRHSuss = zeros ( M . endo_nbr , M . exo_nbr , totparam_nbr ) ;
for jp = 1 : modparam_nbr
dG_x_u = kron ( dghx ( idx_states , : , jp ) , oo . dr . ghu ( idx_states , : ) ) + kron ( oo . dr . ghx ( idx_states , : ) , dghu ( idx_states , : , jp ) ) ;
dG_u_u = kron ( dghu ( idx_states , : , jp ) , oo . dr . ghu ( idx_states , : ) ) + kron ( oo . dr . ghu ( idx_states , : ) , dghu ( idx_states , : , jp ) ) ;
dzxu = [ zeros ( M . nspred , M . nspred * M . exo_nbr ) ;
dghxu ( kcurr ~= 0 , : , jp ) ;
dghxx ( klead ~= 0 , : , jp ) * G_x_u + oo . dr . ghxx ( klead ~= 0 , : ) * dG_x_u + dghx ( klead ~= 0 , : , jp ) * oo . dr . ghxu ( idx_states , : ) + oo . dr . ghx ( klead ~= 0 , : ) * dghxu ( idx_states , : , jp ) ;
zeros ( M . exo_nbr , M . nspred * M . exo_nbr ) ] ;
dzuu = [ zeros ( M . nspred , M . exo_nbr ^2 ) ;
dghuu ( kcurr ~= 0 , : , jp ) ;
dghxx ( klead ~= 0 , : , jp ) * G_u_u + oo . dr . ghxx ( klead ~= 0 , : ) * dG_u_u + dghx ( klead ~= 0 , : , jp ) * oo . dr . ghuu ( idx_states , : ) + oo . dr . ghx ( klead ~= 0 , : ) * dghuu ( idx_states , : , jp ) ;
zeros ( M . exo_nbr , M . exo_nbr ^2 ) ] ;
dG_x_xu = kron ( dghx ( idx_states , : , jp ) , oo . dr . ghxu ( idx_states , : ) ) + kron ( oo . dr . ghx ( idx_states , : ) , dghxu ( idx_states , : , jp ) ) ;
dG_x_uu = kron ( dghx ( idx_states , : , jp ) , oo . dr . ghuu ( idx_states , : ) ) + kron ( oo . dr . ghx ( idx_states , : ) , dghuu ( idx_states , : , jp ) ) ;
dG_u_uu = kron ( dghu ( idx_states , : , jp ) , oo . dr . ghuu ( idx_states , : ) ) + kron ( oo . dr . ghu ( idx_states , : ) , dghuu ( idx_states , : , jp ) ) ;
dG_xx_u = kron ( dghxx ( idx_states , : , jp ) , oo . dr . ghu ( idx_states , : ) ) + kron ( oo . dr . ghxx ( idx_states , : ) , dghu ( idx_states , : , jp ) ) ;
dG_xu_u = kron ( dghxu ( idx_states , : , jp ) , oo . dr . ghu ( idx_states , : ) ) + kron ( oo . dr . ghxu ( idx_states , : ) , dghu ( idx_states , : , jp ) ) ;
dG_x_x_u = kron ( G_x_x , dghu ( idx_states , : , jp ) ) + kron ( dG_x_x ( : , : , jp ) , oo . dr . ghu ( idx_states , : ) ) ;
dG_x_u_u = kron ( G_x_u , dghu ( idx_states , : , jp ) ) + kron ( dG_x_u , oo . dr . ghu ( idx_states , : ) ) ;
dG_u_u_u = kron ( G_u_u , dghu ( idx_states , : , jp ) ) + kron ( dG_u_u , oo . dr . ghu ( idx_states , : ) ) ;
dZ_x_u = kron ( dzx ( : , : , jp ) , zu ) + kron ( zx , dzu ( : , : , jp ) ) ;
dZ_u_u = kron ( dzu ( : , : , jp ) , zu ) + kron ( zu , dzu ( : , : , jp ) ) ;
dZ_x_x_u = kron ( dzx ( : , : , jp ) , Z_x_u ) + kron ( zx , dZ_x_u ) ;
dZ_x_u_u = kron ( dZ_x_u , zu ) + kron ( Z_x_u , dzu ( : , : , jp ) ) ;
dZ_u_u_u = kron ( dZ_u_u , zu ) + kron ( Z_u_u , dzu ( : , : , jp ) ) ;
dZ_xx_u = kron ( dzxx ( : , : , jp ) , zu ) + kron ( zxx , dzu ( : , : , jp ) ) ;
dZ_xu_u = kron ( dzxu , zu ) + kron ( zxu , dzu ( : , : , jp ) ) ;
dZ_x_xu = kron ( dzx ( : , : , jp ) , zxu ) + kron ( zx , dzxu ) ;
dZ_x_uu = kron ( dzx ( : , : , jp ) , zuu ) + kron ( zx , dzuu ) ;
dZ_u_uu = kron ( dzu ( : , : , jp ) , zuu ) + kron ( zu , dzuu ) ;
dB_ghxxx = dB ( : , : , jp ) * oo . dr . ghxxx + B * dghxxx ( : , : , jp ) ;
%Compute dghxxu
dRHS = dA ( : , : , jp ) * oo . dr . ghxxu + dB_ghxxx * G_x_x_u + B_ghxxx * dG_x_x_u ;
dRHS = dRHS + dg3 ( : , k3yy0ex0 ( kyy0ex0 , kyy0ex0 , kyy0ex0 ) + ( jp - 1 ) * yy0ex0_nbr ^3 ) * Z_x_x_u + g3 ( : , k3yy0ex0 ( kyy0ex0 , kyy0ex0 , kyy0ex0 ) ) * dZ_x_x_u ;
dRHS = dRHS + dg2 ( : , k2yy0ex0 ( kyy0ex0 , kyy0ex0 ) + ( jp - 1 ) * yy0ex0_nbr ^2 ) * aux_ZP_x_xu_Z_xx_u + g2 ( : , k2yy0ex0 ( kyy0ex0 , kyy0ex0 ) ) * ( dZ_x_xu * P_x_xu + dZ_xx_u ) ;
dRHS = dRHS + dfyp_ghxxp ( : , : , jp ) * aux_GP_x_xu_G_xx_u + fyp_ghxxp * ( dG_x_xu * P_x_xu + dG_xx_u ) ;
2019-12-20 16:28:06 +01:00
dghxxu ( : , : , jp ) = invA * ( - dRHS ) ;
Added parameter derivatives of perturbation solution up to 3 order
# Preliminary comments
I finished the identification toolbox at orders two and three using the pruned state space system, but before I merge request this, I decided to first merge the new functionality to compute parameter derivatives of perturbation solution matrices at higher orders. So after this is approved, I merge the identification toolbox.
I guess @rattoma, @sebastien, and @michel are best choices to review this.
I outline the main idea first and then provide some more detailed changes I made to the functions.
***
# Main idea
This merge request is concerned with the *analytical*computation of the parameter derivatives of first, second and third order perturbation solution matrices, i.e. using _closed-form_ expressions to efficiently compute the derivative of $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ *with respect to model parameters* $\theta$. Note that $\theta$ contains model parameters, stderr and corr parameters of shocks. stderr and corr parameters of measurement errors are not yet supported, (they can easily be included as exogenous shocks). The availability of such derivatives is beneficial in terms of more reliable analysis of model sensitivity and parameter identifiability as well as more efficient estimation methods, in particular for models solved up to third order, as it is well-known that numerical derivatives are a tricky business, especially for large models.
References for my approach are:
* Iskrev (2008, 2010) and Schmitt-Grohé and Uribe (2012, Appendix) who were the first to compute the parameter derivatives analytically at first order, however, using inefficient (sparse) Kronecker products.
* Mutschler (2015) who provides the expressions for a second-order, but again using inefficient (sparse) Kronecker products.
* Ratto and Iskrev (2012) who show how the first-order system can be solved accurately, fast and efficiently using existing numerical algorithms for generalized Sylvester equations by taking the parameter derivative with respect to each parameter separately.
* Julliard and Kamenik (2004) who provide the perturbation solution equation system in tensor notation at any order k.
* Levintal (2017) who introduces permutation matrices to express the perturbation solution equation system in matrix notation up to fifth order.
Note that @rattoma already implemented the parameter derivatives of $g_x$ and $g_u$ analytically (and numerically), and I rely heavily on his work in `get_first_order_solution_params_derivs.m` (previously `getH.m`). My additions are mainly to this function and thus it is renamed to `get_perturbation_params_derivs.m`.
The basic idea of this merge request is to take the second- and third-order perturbation solution systems in Julliard and Kamenik (2004), unfold these into an equivalent matrix representation using permutation matrices as in Levintal (2017). Then extending Ratto and Iskrev (2012) one takes the derivative with respect to each parameter separately and gets a computational problem that is linear, albeit large, as it involves either solving generalized Sylvester equations or taking inverses of highly sparse matrices. I will now briefly summarize the perturbation solution system at third order and the system that results when taking the derivative with respect to parameters.
## Perturbation Solution
The following systems arise at first, second, and third order:
$(ghx): f_{x} z_{x} = f_{y_{-}^*} + f_{y_0} g_{x} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{x}= A g_{x} + f_{y_{-}^*}=0$
$(ghu): f_{z} z_{u} = f_{y_0} g_{u} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{u} + f_{u}= A g_u + f_u = 0$
$(ghxx) : A g_{xx} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{x}\right) + f_{zz} \left( z_{x} \otimes z_{x} \right) = 0$
$(ghxu) : A g_{xu} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{x} \otimes z_{u} \right) = 0$
$(ghuu) : A g_{uu} + B g_{xx} \left(g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{u} \otimes z_{u} \right) = 0$
$(ghs2) : (A+B) g_{\sigma\sigma} + \left( f_{y^{**}_{+}y^{**}_{+}} \left(g^{**}_{u} \otimes g^{**}_{u}\right) + f_{y^{**}_{+}} g^{**}_{uu}\right)vec(\Sigma) = 0$
$(ghxxx) : A g_{xxx} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{x}\right) + f_{y_{+}}g^{**}_{xx} \left(g^{*}_x \otimes g^{*}_{xx}\right)P_{x\_xx} + f_{zz} \left( z_{x} \otimes z_{xx} \right)P_{x\_xx} + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{x} \right) = 0$
$(ghxxu) : A g_{xxu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{u} \right) + f_{zz} \left( \left( z_{x} \otimes z_{xu} \right)P_{x\_xu} + \left(z_{xx} \otimes z_{u}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{x} \otimes g^{*}_{xu}\right)P_{x\_xu} + \left(g^{*}_{xx} \otimes g^{*}_{u}\right) \right) = 0$
$(ghxuu) : A g_{xuu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( \left( z_{xu} \otimes z_{u} \right)P_{xu\_u} + \left(z_{x} \otimes z_{uu}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{xu} \otimes g^{*}_{u}\right)P_{xu\_u} + \left(g^{*}_{x} \otimes g^{*}_{uu}\right) \right) = 0$
$(ghuuu) : A g_{uuu} + B g_{xxx} \left(g^{*}_{u} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{u} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( z_{u} \otimes z_{uu} \right)P_{u\_uu} + f_{y_{+}}g^{**}_{xx} \left(g^{*}_{u} \otimes g^{*}_{uu}\right)P_{u\_uu} = 0$
$(ghx\sigma\sigma) : A g_{x\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_x + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{x} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{x} \otimes z_{\sigma\sigma}\right) + F_{xu_{+}u_{+}}\left(I_{n_x} \otimes vec(\Sigma)\right) = 0$
$F_{xu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_x^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{xu_{+}} \otimes z_{u_{+}} \right)P_{xu\_u} + \left(z_{x} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{x} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
$(ghu\sigma\sigma) : A g_{u\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_{u} + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{u} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{u} \otimes z_{\sigma\sigma}\right) + F_{uu_{+}u_{+}}\left(I_{n_u} \otimes vec(\Sigma_u)\right) = 0$
$F_{uu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_u^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{uu_{+}} \otimes z_{u_{+}} \right)P_{uu\_u} + \left(z_{u} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{u} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
A and B are the common perturbation matrices:
$A = f_{y_0} + \begin{pmatrix} \underbrace{0}_{n\times n_{static}} &\vdots& \underbrace{f_{y^{**}_{+}} \cdot g^{**}_{x}}_{n \times n_{spred}} &\vdots& \underbrace{0}_{n\times n_{frwd}} \end{pmatrix}$and $B = \begin{pmatrix} \underbrace{0}_{n \times n_{static}}&\vdots & \underbrace{0}_{n \times n_{pred}} & \vdots & \underbrace{f_{y^{**}_{+}}}_{n \times n_{sfwrd}} \end{pmatrix}$
and $z=(y_{-}^{\ast}; y; y_{+}^{\ast\ast}; u)$ denotes the dynamic model variables as in `M_.lead_lag_incidence`, $y^\ast$ denote state variables, $y^{\ast\ast}$ denote forward looking variables, $y_+$ denote the variables with a lead, $y_{-}$ denote variables with a lag, $y_0$ denote variables at period t, $f$ the model equations, and $f_z$ the first-order dynamic model derivatives, $f_{zz}$ the second-order dynamic derivatives, and $f_{zzz}$ the third-order dynamic model derivatives. Then:
$z_{x} = \begin{pmatrix}I\\g_{x}\\g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{u} =\begin{pmatrix}0\\g_{u}\\g^{**}_{x} \cdot g^{*}_{u}\\I\end{pmatrix}$, $z_{u_{+}} =\begin{pmatrix}0\\0\\g^{**}_{u}\\0\end{pmatrix}$
$z_{xx} = \begin{pmatrix} 0\\g_{xx}\\g^{**}_{x} \left( g^{*}_x \otimes g^{*}_{x} \right) + g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{xu} =\begin{pmatrix}0\\g_{xu}\\g^{**}_{xx} \left( g^{*}_x \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{xu}\\0\end{pmatrix}$, $z_{uu} =\begin{pmatrix}0\\g_{uu}\\g^{**}_{xx} \left( g^{*}_u \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{uu}\\0\end{pmatrix}$,
$z_{xu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_x \otimes I \right)\\0\end{pmatrix}$, $z_{uu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_{u} \otimes I \right)\\0\end{pmatrix}$, $z_{u_{+}u_{+}} =\begin{pmatrix}0\\0\\g^{\ast\ast}_{uu}\\0\end{pmatrix}$, $z_{\sigma\sigma} = \begin{pmatrix}0\\ g_{\sigma\sigma}\\ g^{\ast\ast}_{x}g^{\ast}_{\sigma\sigma} + g^{\ast\ast}_{\sigma\sigma}\\0 \end{pmatrix}$
$P$ are permutation matrices that can be computed using Matlab's `ipermute` function.
## Parameter derivatives of perturbation solutions
First, we need the parameter derivatives of first, second, third, and fourth derivatives of the dynamic model (i.e. g1,g2,g3,g4 in dynamic files), I make use of the implicit function theorem: Let $f_{z^k}$ denote the kth derivative (wrt all dynamic variables) of the dynamic model, then let $df_{z^k}$ denote the first-derivative (wrt all model parameters) of $f_{z^k}$ evaluated at the steady state. Note that $f_{z^k}$ is a function of both the model parameters $\theta$ and of the steady state of all dynamic variables $\bar{z}$, which also depend on the parameters. Hence, implicitly $f_{z^k}=f_{z^k}(\theta,\bar{z}(\theta))$ and $df_{z^k}$ consists of two parts:
1. direct derivative wrt to all model parameters given by the preprocessor in the `_params_derivs.m` files
2. contribution of derivative of steady state of dynamic variables (wrt all model parameters): $f_{z^{k+1}} \cdot d\bar{z}$
Note that we already have functionality to compute $d\bar{z}$ analytically.
Having this, the above perturbation systems are basically equations of the following types
$AX +BXC = RHS$ or $AX = RHS$
Now when taking the derivative (wrt to one single parameter $\theta_j$), we get
$A\mathrm{d}\{X\} + B\mathrm{d}\{X\}C = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X - \mathrm{d}\{B\}XC - BX\mathrm{d}\{C\}$
or
$A\mathrm{d}\{X\} = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X$
The first one is a Sylvester type equation, the second one can be solved by taking the inverse of $A$. The only diffculty and tedious work arrises in computing (the highly sparse) derivatives of $RHS$.
***
# New functions: `
## get_perturbation_params_derivs.m`and `get_perturbation_params_derivs_numerical_objective.m`
* The parameter derivatives up to third order are computed in the new function`get_perturbation_params_derivs.m` both analytically and numerically. For numerical derivatives `get_perturbation_params_derivs_numerical_objective.m` is the objective for `fjaco.m` or `hessian_sparse.m` or `hessian.m`.
* `get_perturbation_params_derivs.m` is basically an extended version of the previous `get_first_order_solution_params_derivs.m` function.
* * `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It is used for numerical derivatives, whenever `analytic_derivation_mode=-1|-2`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
* Detailed changes:
* Most important: notation of this function is now in accordance to the k_order_solver, i.e. we do not compute derivatives of Kalman transition matrices A and B, but rather the solution matrices ghx,ghu,ghxx,ghxu,ghuu,ghs2,ghxxx,ghxxu,ghxuu,ghuuu,ghxss,ghuss in accordance with notation used in `oo_.dr`. As a byproduct at first-order, focusing on ghx and ghu instead of Kalman transition matrices A and B makes the computations slightly faster for large models (e.g. for Quest the computations were faster by a couple of seconds, not much, but okay).
* Removed use of `kstate`, see also Dynare/dynare#1653 and Dynare/dynare!1656
* Output arguments are stored in a structure `DERIVS`, there is also a flag `d2flag` that computes parameter hessians needed only in `dsge_likelihood.m`.
* Removed `kronflag` as input. `options_.analytic_derivation_mode` is now used instead of `kronflag`.
* Removed `indvar`, an index that was used to selected specific variables in the derivatives. This is not needed, as we always compute the parameter derivatives for all variables first and then select a subset of variables. The selection now takes place in other functions, like `dsge_likelihood.m`.
* Introduced some checks: (i) deterministic exogenous variables are not supported, (ii) Kronecker method only compatible with first-order approximation so reset to sylvester method, (iii) for purely backward or forward models we need to be careful with the rows in `M_.lead_la g_incidence`, (iv) if `_params_derivs.m` files are missing an error is thrown.
* For numerical derivatives, if mod file does not contain an `estimated_params_block`, a temporary one with the most important parameter information is created.
## `unfold_g4.m`
* When evaluating g3 and g4 one needs to take into account that these do not contain symmetric elements, so one needs to use `unfold_g3.m` and the new function `unfold_g4.m`. This returns an unfolded version of the same matrix (i.e. with symmetric elements).
***
# New test models
`.gitignore` and `Makefile.am` are changed accordingly. Also now it is possible to run test suite on analytic_derivatives, i.e. run `make check m/analytic_derivatives`
## `analytic_derivatives/BrockMirman_PertParamsDerivs.mod`
* This is the Brock Mirman model, where we know the exact policy function $g$ for capital and consumption. As this does not imply a nonzero $g_{\sigma\sigma}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ I added some artificial equations to get nonzero solution matrices with respect to $\sigma$. The true perturbation solution matrices $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ are then computed analytically with Matlab's symbolic toolbox and saved in `nBrockMirmanSYM.mat`. There is a preprocessor flag that recreates these analytical computations if changes are needed (and to check whether I made some errors here ;-) )
* Then solution matrices up to third order and their parameter Jacobians are then compared to the ones computed by Dynare's `k_order_solver` and by `get_perturbation_params_derivs` for all `analytic_derivation_mode`'s. There will be an error if the maximum absolute deviation is too large, i.e. for numerical derivatives (`analytic_derivation_mode=-1|-2`) the tolerance is choosen lower (around 1e-5); for analytical methods we are stricter: around 1e-13 for first-order, 1e-12 for second order, and 1e-11 for third-order.
* As a side note, this mod file also checks Dynare's `k_order_solver` algorithm and throws an error if something is wrong.
* This test model shows that the new functionality works well. And analytical derivatives perform way better and accurate than numerical ones, even for this small model.
## `analytic_derivatives/burnside_3_order_PertParamsDerivs.mod`
* This builds upon `tests/k_order_perturbation/burnside_k_order.mod` and computes the true parameter derivatives analytically by hand.
* This test model also shows that the new functionality works well.
## `analytic_derivatives/LindeTrabandt2019.mod`
* Shows that the new functionality also works for medium-sized models, i.e. a SW type model solved at third order with 35 variables (11 states). 2 shocks and 20 parameters.
* This mod file can be used to tweak the speed of the computations in the future.
* Compares numerical versus analytical parameter derivatives (for first, second and third order). Note that this model clearly shows that numerical ones are quite different than analytical ones even at first order!
## `identification/LindeTrabandt2019_xfail.mod`
* This model is a check for issue Dynare/dynare#1595, see fjaco.m below, and will fail.
* Removed `analytic_derivatives/ls2003.mod` as this mod file is neither in the testsuite nor does it work.
***
# Detailed changes in other functions
## `get_first_order_solution_params_derivs.m`
* Deleted, or actually, renamed to `get_perturbation_params_derivs.m`, as this function now is able to compute the derivatives up to third order
## `identification_numerical_objective.m`
* `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
## `dsge_likelihood.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `T`matrix anymore, but instead of the dynare solution matrix `ghx`. So we recreate `T` here as this amounts to adding some zeros and focusing on selected variables only.
* Added some checks to make sure the first-order approximation is selected.
* Removed `kron_flag` as input, as `get_perturbation_params_derivs` looks into `options_.analytic_derivation_mode` for `kron_flag`.
## `dynare_identification.m`
* make sure that setting `analytic_derivation_mode` is set both in `options_ident` and `options_`. Note that at the end of the function we restore the `options_` structure, so all changes are local. In a next merge request, I will remove the global variables to make all variables local.
## `get_identification_jacobians.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `A` and `B` matrix anymore, but instead of the dynare solution matrix `ghx` and `ghu`. So we recreate these matrices here instead of in `get_perturbation_params_derivs.m`.
* Added `str2func` for better function handles in `fjaco.m`.
## `fjaco.m`
* make `tol`an option, which can be adjusted by changing `options_.dynatol.x`for identification and parameter derivatives purposes.
* include a check and an informative error message, if numerical derivatives (two-sided finite difference method) yield errors in `resol.m` for identification and parameter derivatives purposes. This closes issue Dynare/dynare#1595.
* Changed year of copyright to 2010-2017,2019
***
# Further suggestions and questions
* Ones this is merged, I will merge request an improvement of the identification toolbox, which will work up to third order using the pruned state space. This will also remove some issues and bugs, and also I will remove global variables in this request.
* The third-order derivatives can be further improved by taking sparsity into account and use mex versions for kronecker products etc. I leave this for further testing (and if anybody actually uses this ;-) )
2019-12-17 19:17:09 +01:00
%Compute dghxuu
dRHS = dA ( : , : , jp ) * oo . dr . ghxuu + dB_ghxxx * G_x_u_u + B_ghxxx * dG_x_u_u ;
dRHS = dRHS + dg3 ( : , k3yy0ex0 ( kyy0ex0 , kyy0ex0 , kyy0ex0 ) + ( jp - 1 ) * yy0ex0_nbr ^3 ) * Z_x_u_u + g3 ( : , k3yy0ex0 ( kyy0ex0 , kyy0ex0 , kyy0ex0 ) ) * dZ_x_u_u ;
dRHS = dRHS + dg2 ( : , k2yy0ex0 ( kyy0ex0 , kyy0ex0 ) + ( jp - 1 ) * yy0ex0_nbr ^2 ) * aux_ZP_xu_u_Z_x_uu + g2 ( : , k2yy0ex0 ( kyy0ex0 , kyy0ex0 ) ) * ( dZ_xu_u * P_xu_u + dZ_x_uu ) ;
dRHS = dRHS + dfyp_ghxxp ( : , : , jp ) * aux_GP_xu_u_G_x_uu + fyp_ghxxp * ( dG_xu_u * P_xu_u + dG_x_uu ) ;
dghxuu ( : , : , jp ) = invA * ( - dRHS ) ;
%Compute dghuuu
dRHS = dA ( : , : , jp ) * oo . dr . ghuuu + dB_ghxxx * G_u_u_u + B_ghxxx * dG_u_u_u ;
dRHS = dRHS + dg3 ( : , k3yy0ex0 ( kyy0ex0 , kyy0ex0 , kyy0ex0 ) + ( jp - 1 ) * yy0ex0_nbr ^3 ) * Z_u_u_u + g3 ( : , k3yy0ex0 ( kyy0ex0 , kyy0ex0 , kyy0ex0 ) ) * dZ_u_u_u ;
dRHS = dRHS + dg2 ( : , k2yy0ex0 ( kyy0ex0 , kyy0ex0 ) + ( jp - 1 ) * yy0ex0_nbr ^2 ) * Z_u_uu * P_u_uu + g2 ( : , k2yy0ex0 ( kyy0ex0 , kyy0ex0 ) ) * dZ_u_uu * P_u_uu ;
dRHS = dRHS + dfyp_ghxxp ( : , : , jp ) * G_u_uu * P_u_uu + fyp_ghxxp * dG_u_uu * P_u_uu ;
dghuuu ( : , : , jp ) = invA * ( - dRHS ) ;
%Compute dRHSxss
dzup = [ zeros ( M . nspred , M . exo_nbr ) ;
zeros ( length ( nonzeros ( kcurr ) ) , M . exo_nbr ) ;
dghu ( klead ~= 0 , : , jp ) ;
zeros ( M . exo_nbr , M . exo_nbr ) ] ;
dzss = [ zeros ( M . nspred , 1 ) ;
dghs2 ( kcurr ~= 0 , stderrparam_nbr + corrparam_nbr + jp ) ;
dghs2 ( klead ~= 0 , stderrparam_nbr + corrparam_nbr + jp ) + dghx ( klead ~= 0 , : , jp ) * oo . dr . ghs2 ( idx_states , : ) + oo . dr . ghx ( klead ~= 0 , : ) * dghs2 ( idx_states , stderrparam_nbr + corrparam_nbr + jp ) ;
zeros ( M . exo_nbr , 1 ) ] ;
dzxup = [ zeros ( M . nspred , M . nspred * M . exo_nbr ) ;
zeros ( length ( nonzeros ( kcurr ) ) , M . nspred * M . exo_nbr ) ;
dghxu ( klead ~= 0 , : , jp ) * kron ( oo . dr . ghx ( idx_states , : ) , eye ( M . exo_nbr ) ) + oo . dr . ghxu ( klead ~= 0 , : ) * kron ( dghx ( idx_states , : , jp ) , eye ( M . exo_nbr ) ) ;
zeros ( M . exo_nbr , M . nspred * M . exo_nbr ) ] ;
dzupup = [ zeros ( M . nspred , M . exo_nbr ^2 ) ;
zeros ( length ( nonzeros ( kcurr ) ) , M . exo_nbr ^2 ) ;
dghuu ( klead ~= 0 , : , jp ) ;
zeros ( M . exo_nbr , M . exo_nbr ^2 ) ] ;
dG_x_ss = kron ( dghx ( idx_states , : , jp ) , oo . dr . ghs2 ( idx_states , : ) ) + kron ( oo . dr . ghx ( idx_states , : ) , dghs2 ( idx_states , stderrparam_nbr + corrparam_nbr + jp ) ) ;
dZ_x_ss = kron ( dzx ( : , : , jp ) , zss ) + kron ( zx , dzss ) ;
dZ_up_up = kron ( dzup , zup ) + kron ( zup , dzup ) ;
dZ_xup_up = kron ( dzxup , zup ) + kron ( zxup , dzup ) ;
dZ_x_upup = kron ( dzx ( : , : , jp ) , zupup ) + kron ( zx , dzupup ) ;
dZ_x_up_up = kron ( dzx ( : , : , jp ) , Z_up_up ) + kron ( zx , dZ_up_up ) ;
daux_ZP_xup_up_Z_x_upup = dZ_xup_up * P_xu_u + dZ_x_upup ;
dFxupup = dg3 ( : , k3yy0ex0 ( kyy0ex0 , kyy0ex0 , kyy0ex0 ) + ( jp - 1 ) * yy0ex0_nbr ^3 ) * Z_x_up_up + g3 ( : , k3yy0ex0 ( kyy0ex0 , kyy0ex0 , kyy0ex0 ) ) * dZ_x_up_up ...
2019-12-20 16:28:06 +01:00
+ dg2 ( : , k2yy0ex0 ( kyy0ex0 , kyy0ex0 ) + ( jp - 1 ) * yy0ex0_nbr ^2 ) * aux_ZP_xup_up_Z_x_upup + g2 ( : , k2yy0ex0 ( kyy0ex0 , kyy0ex0 ) ) * daux_ZP_xup_up_Z_x_upup ...
+ dg1 ( : , nonzeros ( klead ) , jp ) * oo . dr . ghxuu ( klead ~= 0 , : ) * kron ( oo . dr . ghx ( idx_states , : ) , eye ( M . exo_nbr ^2 ) ) + g1 ( : , nonzeros ( klead ) ) * dghxuu ( klead ~= 0 , : , jp ) * kron ( oo . dr . ghx ( idx_states , : ) , eye ( M . exo_nbr ^2 ) ) + g1 ( : , nonzeros ( klead ) ) * oo . dr . ghxuu ( klead ~= 0 , : ) * kron ( dghx ( idx_states , : , jp ) , eye ( M . exo_nbr ^2 ) ) ;
Added parameter derivatives of perturbation solution up to 3 order
# Preliminary comments
I finished the identification toolbox at orders two and three using the pruned state space system, but before I merge request this, I decided to first merge the new functionality to compute parameter derivatives of perturbation solution matrices at higher orders. So after this is approved, I merge the identification toolbox.
I guess @rattoma, @sebastien, and @michel are best choices to review this.
I outline the main idea first and then provide some more detailed changes I made to the functions.
***
# Main idea
This merge request is concerned with the *analytical*computation of the parameter derivatives of first, second and third order perturbation solution matrices, i.e. using _closed-form_ expressions to efficiently compute the derivative of $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ *with respect to model parameters* $\theta$. Note that $\theta$ contains model parameters, stderr and corr parameters of shocks. stderr and corr parameters of measurement errors are not yet supported, (they can easily be included as exogenous shocks). The availability of such derivatives is beneficial in terms of more reliable analysis of model sensitivity and parameter identifiability as well as more efficient estimation methods, in particular for models solved up to third order, as it is well-known that numerical derivatives are a tricky business, especially for large models.
References for my approach are:
* Iskrev (2008, 2010) and Schmitt-Grohé and Uribe (2012, Appendix) who were the first to compute the parameter derivatives analytically at first order, however, using inefficient (sparse) Kronecker products.
* Mutschler (2015) who provides the expressions for a second-order, but again using inefficient (sparse) Kronecker products.
* Ratto and Iskrev (2012) who show how the first-order system can be solved accurately, fast and efficiently using existing numerical algorithms for generalized Sylvester equations by taking the parameter derivative with respect to each parameter separately.
* Julliard and Kamenik (2004) who provide the perturbation solution equation system in tensor notation at any order k.
* Levintal (2017) who introduces permutation matrices to express the perturbation solution equation system in matrix notation up to fifth order.
Note that @rattoma already implemented the parameter derivatives of $g_x$ and $g_u$ analytically (and numerically), and I rely heavily on his work in `get_first_order_solution_params_derivs.m` (previously `getH.m`). My additions are mainly to this function and thus it is renamed to `get_perturbation_params_derivs.m`.
The basic idea of this merge request is to take the second- and third-order perturbation solution systems in Julliard and Kamenik (2004), unfold these into an equivalent matrix representation using permutation matrices as in Levintal (2017). Then extending Ratto and Iskrev (2012) one takes the derivative with respect to each parameter separately and gets a computational problem that is linear, albeit large, as it involves either solving generalized Sylvester equations or taking inverses of highly sparse matrices. I will now briefly summarize the perturbation solution system at third order and the system that results when taking the derivative with respect to parameters.
## Perturbation Solution
The following systems arise at first, second, and third order:
$(ghx): f_{x} z_{x} = f_{y_{-}^*} + f_{y_0} g_{x} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{x}= A g_{x} + f_{y_{-}^*}=0$
$(ghu): f_{z} z_{u} = f_{y_0} g_{u} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{u} + f_{u}= A g_u + f_u = 0$
$(ghxx) : A g_{xx} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{x}\right) + f_{zz} \left( z_{x} \otimes z_{x} \right) = 0$
$(ghxu) : A g_{xu} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{x} \otimes z_{u} \right) = 0$
$(ghuu) : A g_{uu} + B g_{xx} \left(g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{u} \otimes z_{u} \right) = 0$
$(ghs2) : (A+B) g_{\sigma\sigma} + \left( f_{y^{**}_{+}y^{**}_{+}} \left(g^{**}_{u} \otimes g^{**}_{u}\right) + f_{y^{**}_{+}} g^{**}_{uu}\right)vec(\Sigma) = 0$
$(ghxxx) : A g_{xxx} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{x}\right) + f_{y_{+}}g^{**}_{xx} \left(g^{*}_x \otimes g^{*}_{xx}\right)P_{x\_xx} + f_{zz} \left( z_{x} \otimes z_{xx} \right)P_{x\_xx} + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{x} \right) = 0$
$(ghxxu) : A g_{xxu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{u} \right) + f_{zz} \left( \left( z_{x} \otimes z_{xu} \right)P_{x\_xu} + \left(z_{xx} \otimes z_{u}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{x} \otimes g^{*}_{xu}\right)P_{x\_xu} + \left(g^{*}_{xx} \otimes g^{*}_{u}\right) \right) = 0$
$(ghxuu) : A g_{xuu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( \left( z_{xu} \otimes z_{u} \right)P_{xu\_u} + \left(z_{x} \otimes z_{uu}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{xu} \otimes g^{*}_{u}\right)P_{xu\_u} + \left(g^{*}_{x} \otimes g^{*}_{uu}\right) \right) = 0$
$(ghuuu) : A g_{uuu} + B g_{xxx} \left(g^{*}_{u} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{u} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( z_{u} \otimes z_{uu} \right)P_{u\_uu} + f_{y_{+}}g^{**}_{xx} \left(g^{*}_{u} \otimes g^{*}_{uu}\right)P_{u\_uu} = 0$
$(ghx\sigma\sigma) : A g_{x\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_x + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{x} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{x} \otimes z_{\sigma\sigma}\right) + F_{xu_{+}u_{+}}\left(I_{n_x} \otimes vec(\Sigma)\right) = 0$
$F_{xu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_x^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{xu_{+}} \otimes z_{u_{+}} \right)P_{xu\_u} + \left(z_{x} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{x} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
$(ghu\sigma\sigma) : A g_{u\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_{u} + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{u} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{u} \otimes z_{\sigma\sigma}\right) + F_{uu_{+}u_{+}}\left(I_{n_u} \otimes vec(\Sigma_u)\right) = 0$
$F_{uu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_u^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{uu_{+}} \otimes z_{u_{+}} \right)P_{uu\_u} + \left(z_{u} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{u} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
A and B are the common perturbation matrices:
$A = f_{y_0} + \begin{pmatrix} \underbrace{0}_{n\times n_{static}} &\vdots& \underbrace{f_{y^{**}_{+}} \cdot g^{**}_{x}}_{n \times n_{spred}} &\vdots& \underbrace{0}_{n\times n_{frwd}} \end{pmatrix}$and $B = \begin{pmatrix} \underbrace{0}_{n \times n_{static}}&\vdots & \underbrace{0}_{n \times n_{pred}} & \vdots & \underbrace{f_{y^{**}_{+}}}_{n \times n_{sfwrd}} \end{pmatrix}$
and $z=(y_{-}^{\ast}; y; y_{+}^{\ast\ast}; u)$ denotes the dynamic model variables as in `M_.lead_lag_incidence`, $y^\ast$ denote state variables, $y^{\ast\ast}$ denote forward looking variables, $y_+$ denote the variables with a lead, $y_{-}$ denote variables with a lag, $y_0$ denote variables at period t, $f$ the model equations, and $f_z$ the first-order dynamic model derivatives, $f_{zz}$ the second-order dynamic derivatives, and $f_{zzz}$ the third-order dynamic model derivatives. Then:
$z_{x} = \begin{pmatrix}I\\g_{x}\\g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{u} =\begin{pmatrix}0\\g_{u}\\g^{**}_{x} \cdot g^{*}_{u}\\I\end{pmatrix}$, $z_{u_{+}} =\begin{pmatrix}0\\0\\g^{**}_{u}\\0\end{pmatrix}$
$z_{xx} = \begin{pmatrix} 0\\g_{xx}\\g^{**}_{x} \left( g^{*}_x \otimes g^{*}_{x} \right) + g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{xu} =\begin{pmatrix}0\\g_{xu}\\g^{**}_{xx} \left( g^{*}_x \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{xu}\\0\end{pmatrix}$, $z_{uu} =\begin{pmatrix}0\\g_{uu}\\g^{**}_{xx} \left( g^{*}_u \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{uu}\\0\end{pmatrix}$,
$z_{xu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_x \otimes I \right)\\0\end{pmatrix}$, $z_{uu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_{u} \otimes I \right)\\0\end{pmatrix}$, $z_{u_{+}u_{+}} =\begin{pmatrix}0\\0\\g^{\ast\ast}_{uu}\\0\end{pmatrix}$, $z_{\sigma\sigma} = \begin{pmatrix}0\\ g_{\sigma\sigma}\\ g^{\ast\ast}_{x}g^{\ast}_{\sigma\sigma} + g^{\ast\ast}_{\sigma\sigma}\\0 \end{pmatrix}$
$P$ are permutation matrices that can be computed using Matlab's `ipermute` function.
## Parameter derivatives of perturbation solutions
First, we need the parameter derivatives of first, second, third, and fourth derivatives of the dynamic model (i.e. g1,g2,g3,g4 in dynamic files), I make use of the implicit function theorem: Let $f_{z^k}$ denote the kth derivative (wrt all dynamic variables) of the dynamic model, then let $df_{z^k}$ denote the first-derivative (wrt all model parameters) of $f_{z^k}$ evaluated at the steady state. Note that $f_{z^k}$ is a function of both the model parameters $\theta$ and of the steady state of all dynamic variables $\bar{z}$, which also depend on the parameters. Hence, implicitly $f_{z^k}=f_{z^k}(\theta,\bar{z}(\theta))$ and $df_{z^k}$ consists of two parts:
1. direct derivative wrt to all model parameters given by the preprocessor in the `_params_derivs.m` files
2. contribution of derivative of steady state of dynamic variables (wrt all model parameters): $f_{z^{k+1}} \cdot d\bar{z}$
Note that we already have functionality to compute $d\bar{z}$ analytically.
Having this, the above perturbation systems are basically equations of the following types
$AX +BXC = RHS$ or $AX = RHS$
Now when taking the derivative (wrt to one single parameter $\theta_j$), we get
$A\mathrm{d}\{X\} + B\mathrm{d}\{X\}C = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X - \mathrm{d}\{B\}XC - BX\mathrm{d}\{C\}$
or
$A\mathrm{d}\{X\} = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X$
The first one is a Sylvester type equation, the second one can be solved by taking the inverse of $A$. The only diffculty and tedious work arrises in computing (the highly sparse) derivatives of $RHS$.
***
# New functions: `
## get_perturbation_params_derivs.m`and `get_perturbation_params_derivs_numerical_objective.m`
* The parameter derivatives up to third order are computed in the new function`get_perturbation_params_derivs.m` both analytically and numerically. For numerical derivatives `get_perturbation_params_derivs_numerical_objective.m` is the objective for `fjaco.m` or `hessian_sparse.m` or `hessian.m`.
* `get_perturbation_params_derivs.m` is basically an extended version of the previous `get_first_order_solution_params_derivs.m` function.
* * `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It is used for numerical derivatives, whenever `analytic_derivation_mode=-1|-2`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
* Detailed changes:
* Most important: notation of this function is now in accordance to the k_order_solver, i.e. we do not compute derivatives of Kalman transition matrices A and B, but rather the solution matrices ghx,ghu,ghxx,ghxu,ghuu,ghs2,ghxxx,ghxxu,ghxuu,ghuuu,ghxss,ghuss in accordance with notation used in `oo_.dr`. As a byproduct at first-order, focusing on ghx and ghu instead of Kalman transition matrices A and B makes the computations slightly faster for large models (e.g. for Quest the computations were faster by a couple of seconds, not much, but okay).
* Removed use of `kstate`, see also Dynare/dynare#1653 and Dynare/dynare!1656
* Output arguments are stored in a structure `DERIVS`, there is also a flag `d2flag` that computes parameter hessians needed only in `dsge_likelihood.m`.
* Removed `kronflag` as input. `options_.analytic_derivation_mode` is now used instead of `kronflag`.
* Removed `indvar`, an index that was used to selected specific variables in the derivatives. This is not needed, as we always compute the parameter derivatives for all variables first and then select a subset of variables. The selection now takes place in other functions, like `dsge_likelihood.m`.
* Introduced some checks: (i) deterministic exogenous variables are not supported, (ii) Kronecker method only compatible with first-order approximation so reset to sylvester method, (iii) for purely backward or forward models we need to be careful with the rows in `M_.lead_la g_incidence`, (iv) if `_params_derivs.m` files are missing an error is thrown.
* For numerical derivatives, if mod file does not contain an `estimated_params_block`, a temporary one with the most important parameter information is created.
## `unfold_g4.m`
* When evaluating g3 and g4 one needs to take into account that these do not contain symmetric elements, so one needs to use `unfold_g3.m` and the new function `unfold_g4.m`. This returns an unfolded version of the same matrix (i.e. with symmetric elements).
***
# New test models
`.gitignore` and `Makefile.am` are changed accordingly. Also now it is possible to run test suite on analytic_derivatives, i.e. run `make check m/analytic_derivatives`
## `analytic_derivatives/BrockMirman_PertParamsDerivs.mod`
* This is the Brock Mirman model, where we know the exact policy function $g$ for capital and consumption. As this does not imply a nonzero $g_{\sigma\sigma}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ I added some artificial equations to get nonzero solution matrices with respect to $\sigma$. The true perturbation solution matrices $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ are then computed analytically with Matlab's symbolic toolbox and saved in `nBrockMirmanSYM.mat`. There is a preprocessor flag that recreates these analytical computations if changes are needed (and to check whether I made some errors here ;-) )
* Then solution matrices up to third order and their parameter Jacobians are then compared to the ones computed by Dynare's `k_order_solver` and by `get_perturbation_params_derivs` for all `analytic_derivation_mode`'s. There will be an error if the maximum absolute deviation is too large, i.e. for numerical derivatives (`analytic_derivation_mode=-1|-2`) the tolerance is choosen lower (around 1e-5); for analytical methods we are stricter: around 1e-13 for first-order, 1e-12 for second order, and 1e-11 for third-order.
* As a side note, this mod file also checks Dynare's `k_order_solver` algorithm and throws an error if something is wrong.
* This test model shows that the new functionality works well. And analytical derivatives perform way better and accurate than numerical ones, even for this small model.
## `analytic_derivatives/burnside_3_order_PertParamsDerivs.mod`
* This builds upon `tests/k_order_perturbation/burnside_k_order.mod` and computes the true parameter derivatives analytically by hand.
* This test model also shows that the new functionality works well.
## `analytic_derivatives/LindeTrabandt2019.mod`
* Shows that the new functionality also works for medium-sized models, i.e. a SW type model solved at third order with 35 variables (11 states). 2 shocks and 20 parameters.
* This mod file can be used to tweak the speed of the computations in the future.
* Compares numerical versus analytical parameter derivatives (for first, second and third order). Note that this model clearly shows that numerical ones are quite different than analytical ones even at first order!
## `identification/LindeTrabandt2019_xfail.mod`
* This model is a check for issue Dynare/dynare#1595, see fjaco.m below, and will fail.
* Removed `analytic_derivatives/ls2003.mod` as this mod file is neither in the testsuite nor does it work.
***
# Detailed changes in other functions
## `get_first_order_solution_params_derivs.m`
* Deleted, or actually, renamed to `get_perturbation_params_derivs.m`, as this function now is able to compute the derivatives up to third order
## `identification_numerical_objective.m`
* `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
## `dsge_likelihood.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `T`matrix anymore, but instead of the dynare solution matrix `ghx`. So we recreate `T` here as this amounts to adding some zeros and focusing on selected variables only.
* Added some checks to make sure the first-order approximation is selected.
* Removed `kron_flag` as input, as `get_perturbation_params_derivs` looks into `options_.analytic_derivation_mode` for `kron_flag`.
## `dynare_identification.m`
* make sure that setting `analytic_derivation_mode` is set both in `options_ident` and `options_`. Note that at the end of the function we restore the `options_` structure, so all changes are local. In a next merge request, I will remove the global variables to make all variables local.
## `get_identification_jacobians.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `A` and `B` matrix anymore, but instead of the dynare solution matrix `ghx` and `ghu`. So we recreate these matrices here instead of in `get_perturbation_params_derivs.m`.
* Added `str2func` for better function handles in `fjaco.m`.
## `fjaco.m`
* make `tol`an option, which can be adjusted by changing `options_.dynatol.x`for identification and parameter derivatives purposes.
* include a check and an informative error message, if numerical derivatives (two-sided finite difference method) yield errors in `resol.m` for identification and parameter derivatives purposes. This closes issue Dynare/dynare#1595.
* Changed year of copyright to 2010-2017,2019
***
# Further suggestions and questions
* Ones this is merged, I will merge request an improvement of the identification toolbox, which will work up to third order using the pruned state space. This will also remove some issues and bugs, and also I will remove global variables in this request.
* The third-order derivatives can be further improved by taking sparsity into account and use mex versions for kronecker products etc. I leave this for further testing (and if anybody actually uses this ;-) )
2019-12-17 19:17:09 +01:00
dRHSxss ( : , : , stderrparam_nbr + corrparam_nbr + jp ) = dA ( : , : , jp ) * oo . dr . ghxss + dB ( : , : , jp ) * oo . dr . ghxss * oo . dr . ghx ( idx_states , : ) + B * oo . dr . ghxss * dghx ( idx_states , : , jp ) ;
dRHSxss ( : , : , stderrparam_nbr + corrparam_nbr + jp ) = dRHSxss ( : , : , stderrparam_nbr + corrparam_nbr + jp ) + dfyp_ghxxp ( : , : , jp ) * G_x_ss + fyp_ghxxp * dG_x_ss ; %m
dRHSxss ( : , : , stderrparam_nbr + corrparam_nbr + jp ) = dRHSxss ( : , : , stderrparam_nbr + corrparam_nbr + jp ) + dg2 ( : , k2yy0ex0 ( kyy0ex0 , kyy0ex0 ) + ( jp - 1 ) * yy0ex0_nbr ^2 ) * Z_x_ss + g2 ( : , k2yy0ex0 ( kyy0ex0 , kyy0ex0 ) ) * dZ_x_ss ;
dRHSxss ( : , : , stderrparam_nbr + corrparam_nbr + jp ) = dRHSxss ( : , : , stderrparam_nbr + corrparam_nbr + jp ) + dFxupup * Ix_vecSig_e ; %missing contribution by dSigma_e
2019-12-20 16:28:06 +01:00
%Compute dRHSuss
Added parameter derivatives of perturbation solution up to 3 order
# Preliminary comments
I finished the identification toolbox at orders two and three using the pruned state space system, but before I merge request this, I decided to first merge the new functionality to compute parameter derivatives of perturbation solution matrices at higher orders. So after this is approved, I merge the identification toolbox.
I guess @rattoma, @sebastien, and @michel are best choices to review this.
I outline the main idea first and then provide some more detailed changes I made to the functions.
***
# Main idea
This merge request is concerned with the *analytical*computation of the parameter derivatives of first, second and third order perturbation solution matrices, i.e. using _closed-form_ expressions to efficiently compute the derivative of $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ *with respect to model parameters* $\theta$. Note that $\theta$ contains model parameters, stderr and corr parameters of shocks. stderr and corr parameters of measurement errors are not yet supported, (they can easily be included as exogenous shocks). The availability of such derivatives is beneficial in terms of more reliable analysis of model sensitivity and parameter identifiability as well as more efficient estimation methods, in particular for models solved up to third order, as it is well-known that numerical derivatives are a tricky business, especially for large models.
References for my approach are:
* Iskrev (2008, 2010) and Schmitt-Grohé and Uribe (2012, Appendix) who were the first to compute the parameter derivatives analytically at first order, however, using inefficient (sparse) Kronecker products.
* Mutschler (2015) who provides the expressions for a second-order, but again using inefficient (sparse) Kronecker products.
* Ratto and Iskrev (2012) who show how the first-order system can be solved accurately, fast and efficiently using existing numerical algorithms for generalized Sylvester equations by taking the parameter derivative with respect to each parameter separately.
* Julliard and Kamenik (2004) who provide the perturbation solution equation system in tensor notation at any order k.
* Levintal (2017) who introduces permutation matrices to express the perturbation solution equation system in matrix notation up to fifth order.
Note that @rattoma already implemented the parameter derivatives of $g_x$ and $g_u$ analytically (and numerically), and I rely heavily on his work in `get_first_order_solution_params_derivs.m` (previously `getH.m`). My additions are mainly to this function and thus it is renamed to `get_perturbation_params_derivs.m`.
The basic idea of this merge request is to take the second- and third-order perturbation solution systems in Julliard and Kamenik (2004), unfold these into an equivalent matrix representation using permutation matrices as in Levintal (2017). Then extending Ratto and Iskrev (2012) one takes the derivative with respect to each parameter separately and gets a computational problem that is linear, albeit large, as it involves either solving generalized Sylvester equations or taking inverses of highly sparse matrices. I will now briefly summarize the perturbation solution system at third order and the system that results when taking the derivative with respect to parameters.
## Perturbation Solution
The following systems arise at first, second, and third order:
$(ghx): f_{x} z_{x} = f_{y_{-}^*} + f_{y_0} g_{x} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{x}= A g_{x} + f_{y_{-}^*}=0$
$(ghu): f_{z} z_{u} = f_{y_0} g_{u} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{u} + f_{u}= A g_u + f_u = 0$
$(ghxx) : A g_{xx} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{x}\right) + f_{zz} \left( z_{x} \otimes z_{x} \right) = 0$
$(ghxu) : A g_{xu} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{x} \otimes z_{u} \right) = 0$
$(ghuu) : A g_{uu} + B g_{xx} \left(g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{u} \otimes z_{u} \right) = 0$
$(ghs2) : (A+B) g_{\sigma\sigma} + \left( f_{y^{**}_{+}y^{**}_{+}} \left(g^{**}_{u} \otimes g^{**}_{u}\right) + f_{y^{**}_{+}} g^{**}_{uu}\right)vec(\Sigma) = 0$
$(ghxxx) : A g_{xxx} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{x}\right) + f_{y_{+}}g^{**}_{xx} \left(g^{*}_x \otimes g^{*}_{xx}\right)P_{x\_xx} + f_{zz} \left( z_{x} \otimes z_{xx} \right)P_{x\_xx} + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{x} \right) = 0$
$(ghxxu) : A g_{xxu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{u} \right) + f_{zz} \left( \left( z_{x} \otimes z_{xu} \right)P_{x\_xu} + \left(z_{xx} \otimes z_{u}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{x} \otimes g^{*}_{xu}\right)P_{x\_xu} + \left(g^{*}_{xx} \otimes g^{*}_{u}\right) \right) = 0$
$(ghxuu) : A g_{xuu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( \left( z_{xu} \otimes z_{u} \right)P_{xu\_u} + \left(z_{x} \otimes z_{uu}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{xu} \otimes g^{*}_{u}\right)P_{xu\_u} + \left(g^{*}_{x} \otimes g^{*}_{uu}\right) \right) = 0$
$(ghuuu) : A g_{uuu} + B g_{xxx} \left(g^{*}_{u} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{u} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( z_{u} \otimes z_{uu} \right)P_{u\_uu} + f_{y_{+}}g^{**}_{xx} \left(g^{*}_{u} \otimes g^{*}_{uu}\right)P_{u\_uu} = 0$
$(ghx\sigma\sigma) : A g_{x\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_x + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{x} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{x} \otimes z_{\sigma\sigma}\right) + F_{xu_{+}u_{+}}\left(I_{n_x} \otimes vec(\Sigma)\right) = 0$
$F_{xu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_x^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{xu_{+}} \otimes z_{u_{+}} \right)P_{xu\_u} + \left(z_{x} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{x} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
$(ghu\sigma\sigma) : A g_{u\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_{u} + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{u} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{u} \otimes z_{\sigma\sigma}\right) + F_{uu_{+}u_{+}}\left(I_{n_u} \otimes vec(\Sigma_u)\right) = 0$
$F_{uu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_u^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{uu_{+}} \otimes z_{u_{+}} \right)P_{uu\_u} + \left(z_{u} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{u} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
A and B are the common perturbation matrices:
$A = f_{y_0} + \begin{pmatrix} \underbrace{0}_{n\times n_{static}} &\vdots& \underbrace{f_{y^{**}_{+}} \cdot g^{**}_{x}}_{n \times n_{spred}} &\vdots& \underbrace{0}_{n\times n_{frwd}} \end{pmatrix}$and $B = \begin{pmatrix} \underbrace{0}_{n \times n_{static}}&\vdots & \underbrace{0}_{n \times n_{pred}} & \vdots & \underbrace{f_{y^{**}_{+}}}_{n \times n_{sfwrd}} \end{pmatrix}$
and $z=(y_{-}^{\ast}; y; y_{+}^{\ast\ast}; u)$ denotes the dynamic model variables as in `M_.lead_lag_incidence`, $y^\ast$ denote state variables, $y^{\ast\ast}$ denote forward looking variables, $y_+$ denote the variables with a lead, $y_{-}$ denote variables with a lag, $y_0$ denote variables at period t, $f$ the model equations, and $f_z$ the first-order dynamic model derivatives, $f_{zz}$ the second-order dynamic derivatives, and $f_{zzz}$ the third-order dynamic model derivatives. Then:
$z_{x} = \begin{pmatrix}I\\g_{x}\\g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{u} =\begin{pmatrix}0\\g_{u}\\g^{**}_{x} \cdot g^{*}_{u}\\I\end{pmatrix}$, $z_{u_{+}} =\begin{pmatrix}0\\0\\g^{**}_{u}\\0\end{pmatrix}$
$z_{xx} = \begin{pmatrix} 0\\g_{xx}\\g^{**}_{x} \left( g^{*}_x \otimes g^{*}_{x} \right) + g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{xu} =\begin{pmatrix}0\\g_{xu}\\g^{**}_{xx} \left( g^{*}_x \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{xu}\\0\end{pmatrix}$, $z_{uu} =\begin{pmatrix}0\\g_{uu}\\g^{**}_{xx} \left( g^{*}_u \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{uu}\\0\end{pmatrix}$,
$z_{xu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_x \otimes I \right)\\0\end{pmatrix}$, $z_{uu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_{u} \otimes I \right)\\0\end{pmatrix}$, $z_{u_{+}u_{+}} =\begin{pmatrix}0\\0\\g^{\ast\ast}_{uu}\\0\end{pmatrix}$, $z_{\sigma\sigma} = \begin{pmatrix}0\\ g_{\sigma\sigma}\\ g^{\ast\ast}_{x}g^{\ast}_{\sigma\sigma} + g^{\ast\ast}_{\sigma\sigma}\\0 \end{pmatrix}$
$P$ are permutation matrices that can be computed using Matlab's `ipermute` function.
## Parameter derivatives of perturbation solutions
First, we need the parameter derivatives of first, second, third, and fourth derivatives of the dynamic model (i.e. g1,g2,g3,g4 in dynamic files), I make use of the implicit function theorem: Let $f_{z^k}$ denote the kth derivative (wrt all dynamic variables) of the dynamic model, then let $df_{z^k}$ denote the first-derivative (wrt all model parameters) of $f_{z^k}$ evaluated at the steady state. Note that $f_{z^k}$ is a function of both the model parameters $\theta$ and of the steady state of all dynamic variables $\bar{z}$, which also depend on the parameters. Hence, implicitly $f_{z^k}=f_{z^k}(\theta,\bar{z}(\theta))$ and $df_{z^k}$ consists of two parts:
1. direct derivative wrt to all model parameters given by the preprocessor in the `_params_derivs.m` files
2. contribution of derivative of steady state of dynamic variables (wrt all model parameters): $f_{z^{k+1}} \cdot d\bar{z}$
Note that we already have functionality to compute $d\bar{z}$ analytically.
Having this, the above perturbation systems are basically equations of the following types
$AX +BXC = RHS$ or $AX = RHS$
Now when taking the derivative (wrt to one single parameter $\theta_j$), we get
$A\mathrm{d}\{X\} + B\mathrm{d}\{X\}C = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X - \mathrm{d}\{B\}XC - BX\mathrm{d}\{C\}$
or
$A\mathrm{d}\{X\} = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X$
The first one is a Sylvester type equation, the second one can be solved by taking the inverse of $A$. The only diffculty and tedious work arrises in computing (the highly sparse) derivatives of $RHS$.
***
# New functions: `
## get_perturbation_params_derivs.m`and `get_perturbation_params_derivs_numerical_objective.m`
* The parameter derivatives up to third order are computed in the new function`get_perturbation_params_derivs.m` both analytically and numerically. For numerical derivatives `get_perturbation_params_derivs_numerical_objective.m` is the objective for `fjaco.m` or `hessian_sparse.m` or `hessian.m`.
* `get_perturbation_params_derivs.m` is basically an extended version of the previous `get_first_order_solution_params_derivs.m` function.
* * `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It is used for numerical derivatives, whenever `analytic_derivation_mode=-1|-2`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
* Detailed changes:
* Most important: notation of this function is now in accordance to the k_order_solver, i.e. we do not compute derivatives of Kalman transition matrices A and B, but rather the solution matrices ghx,ghu,ghxx,ghxu,ghuu,ghs2,ghxxx,ghxxu,ghxuu,ghuuu,ghxss,ghuss in accordance with notation used in `oo_.dr`. As a byproduct at first-order, focusing on ghx and ghu instead of Kalman transition matrices A and B makes the computations slightly faster for large models (e.g. for Quest the computations were faster by a couple of seconds, not much, but okay).
* Removed use of `kstate`, see also Dynare/dynare#1653 and Dynare/dynare!1656
* Output arguments are stored in a structure `DERIVS`, there is also a flag `d2flag` that computes parameter hessians needed only in `dsge_likelihood.m`.
* Removed `kronflag` as input. `options_.analytic_derivation_mode` is now used instead of `kronflag`.
* Removed `indvar`, an index that was used to selected specific variables in the derivatives. This is not needed, as we always compute the parameter derivatives for all variables first and then select a subset of variables. The selection now takes place in other functions, like `dsge_likelihood.m`.
* Introduced some checks: (i) deterministic exogenous variables are not supported, (ii) Kronecker method only compatible with first-order approximation so reset to sylvester method, (iii) for purely backward or forward models we need to be careful with the rows in `M_.lead_la g_incidence`, (iv) if `_params_derivs.m` files are missing an error is thrown.
* For numerical derivatives, if mod file does not contain an `estimated_params_block`, a temporary one with the most important parameter information is created.
## `unfold_g4.m`
* When evaluating g3 and g4 one needs to take into account that these do not contain symmetric elements, so one needs to use `unfold_g3.m` and the new function `unfold_g4.m`. This returns an unfolded version of the same matrix (i.e. with symmetric elements).
***
# New test models
`.gitignore` and `Makefile.am` are changed accordingly. Also now it is possible to run test suite on analytic_derivatives, i.e. run `make check m/analytic_derivatives`
## `analytic_derivatives/BrockMirman_PertParamsDerivs.mod`
* This is the Brock Mirman model, where we know the exact policy function $g$ for capital and consumption. As this does not imply a nonzero $g_{\sigma\sigma}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ I added some artificial equations to get nonzero solution matrices with respect to $\sigma$. The true perturbation solution matrices $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ are then computed analytically with Matlab's symbolic toolbox and saved in `nBrockMirmanSYM.mat`. There is a preprocessor flag that recreates these analytical computations if changes are needed (and to check whether I made some errors here ;-) )
* Then solution matrices up to third order and their parameter Jacobians are then compared to the ones computed by Dynare's `k_order_solver` and by `get_perturbation_params_derivs` for all `analytic_derivation_mode`'s. There will be an error if the maximum absolute deviation is too large, i.e. for numerical derivatives (`analytic_derivation_mode=-1|-2`) the tolerance is choosen lower (around 1e-5); for analytical methods we are stricter: around 1e-13 for first-order, 1e-12 for second order, and 1e-11 for third-order.
* As a side note, this mod file also checks Dynare's `k_order_solver` algorithm and throws an error if something is wrong.
* This test model shows that the new functionality works well. And analytical derivatives perform way better and accurate than numerical ones, even for this small model.
## `analytic_derivatives/burnside_3_order_PertParamsDerivs.mod`
* This builds upon `tests/k_order_perturbation/burnside_k_order.mod` and computes the true parameter derivatives analytically by hand.
* This test model also shows that the new functionality works well.
## `analytic_derivatives/LindeTrabandt2019.mod`
* Shows that the new functionality also works for medium-sized models, i.e. a SW type model solved at third order with 35 variables (11 states). 2 shocks and 20 parameters.
* This mod file can be used to tweak the speed of the computations in the future.
* Compares numerical versus analytical parameter derivatives (for first, second and third order). Note that this model clearly shows that numerical ones are quite different than analytical ones even at first order!
## `identification/LindeTrabandt2019_xfail.mod`
* This model is a check for issue Dynare/dynare#1595, see fjaco.m below, and will fail.
* Removed `analytic_derivatives/ls2003.mod` as this mod file is neither in the testsuite nor does it work.
***
# Detailed changes in other functions
## `get_first_order_solution_params_derivs.m`
* Deleted, or actually, renamed to `get_perturbation_params_derivs.m`, as this function now is able to compute the derivatives up to third order
## `identification_numerical_objective.m`
* `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
## `dsge_likelihood.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `T`matrix anymore, but instead of the dynare solution matrix `ghx`. So we recreate `T` here as this amounts to adding some zeros and focusing on selected variables only.
* Added some checks to make sure the first-order approximation is selected.
* Removed `kron_flag` as input, as `get_perturbation_params_derivs` looks into `options_.analytic_derivation_mode` for `kron_flag`.
## `dynare_identification.m`
* make sure that setting `analytic_derivation_mode` is set both in `options_ident` and `options_`. Note that at the end of the function we restore the `options_` structure, so all changes are local. In a next merge request, I will remove the global variables to make all variables local.
## `get_identification_jacobians.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `A` and `B` matrix anymore, but instead of the dynare solution matrix `ghx` and `ghu`. So we recreate these matrices here instead of in `get_perturbation_params_derivs.m`.
* Added `str2func` for better function handles in `fjaco.m`.
## `fjaco.m`
* make `tol`an option, which can be adjusted by changing `options_.dynatol.x`for identification and parameter derivatives purposes.
* include a check and an informative error message, if numerical derivatives (two-sided finite difference method) yield errors in `resol.m` for identification and parameter derivatives purposes. This closes issue Dynare/dynare#1595.
* Changed year of copyright to 2010-2017,2019
***
# Further suggestions and questions
* Ones this is merged, I will merge request an improvement of the identification toolbox, which will work up to third order using the pruned state space. This will also remove some issues and bugs, and also I will remove global variables in this request.
* The third-order derivatives can be further improved by taking sparsity into account and use mex versions for kronecker products etc. I leave this for further testing (and if anybody actually uses this ;-) )
2019-12-17 19:17:09 +01:00
dzuup = [ zeros ( M . nspred , M . exo_nbr ^2 ) ;
zeros ( length ( nonzeros ( kcurr ) ) , M . exo_nbr ^2 ) ;
dghxu ( klead ~= 0 , : , jp ) * kron ( oo . dr . ghu ( idx_states , : ) , eye ( M . exo_nbr ) ) + oo . dr . ghxu ( klead ~= 0 , : ) * kron ( dghu ( idx_states , : , jp ) , eye ( M . exo_nbr ) ) ;
zeros ( M . exo_nbr , M . exo_nbr ^2 ) ] ;
dG_u_ss = kron ( dghu ( idx_states , : , jp ) , oo . dr . ghs2 ( idx_states , : ) ) + kron ( oo . dr . ghu ( idx_states , : ) , dghs2 ( idx_states , stderrparam_nbr + corrparam_nbr + jp ) ) ;
dZ_u_ss = kron ( dzu ( : , : , jp ) , zss ) + kron ( zu , dzss ) ;
dZ_u_upup = kron ( dzu ( : , : , jp ) , zupup ) + kron ( zu , dzupup ) ;
dZ_uup_up = kron ( dzuup , zup ) + kron ( zuup , dzup ) ;
2019-12-20 16:28:06 +01:00
dZ_u_up_up = kron ( dzu ( : , : , jp ) , Z_up_up ) + kron ( zu , dZ_up_up ) ;
Added parameter derivatives of perturbation solution up to 3 order
# Preliminary comments
I finished the identification toolbox at orders two and three using the pruned state space system, but before I merge request this, I decided to first merge the new functionality to compute parameter derivatives of perturbation solution matrices at higher orders. So after this is approved, I merge the identification toolbox.
I guess @rattoma, @sebastien, and @michel are best choices to review this.
I outline the main idea first and then provide some more detailed changes I made to the functions.
***
# Main idea
This merge request is concerned with the *analytical*computation of the parameter derivatives of first, second and third order perturbation solution matrices, i.e. using _closed-form_ expressions to efficiently compute the derivative of $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ *with respect to model parameters* $\theta$. Note that $\theta$ contains model parameters, stderr and corr parameters of shocks. stderr and corr parameters of measurement errors are not yet supported, (they can easily be included as exogenous shocks). The availability of such derivatives is beneficial in terms of more reliable analysis of model sensitivity and parameter identifiability as well as more efficient estimation methods, in particular for models solved up to third order, as it is well-known that numerical derivatives are a tricky business, especially for large models.
References for my approach are:
* Iskrev (2008, 2010) and Schmitt-Grohé and Uribe (2012, Appendix) who were the first to compute the parameter derivatives analytically at first order, however, using inefficient (sparse) Kronecker products.
* Mutschler (2015) who provides the expressions for a second-order, but again using inefficient (sparse) Kronecker products.
* Ratto and Iskrev (2012) who show how the first-order system can be solved accurately, fast and efficiently using existing numerical algorithms for generalized Sylvester equations by taking the parameter derivative with respect to each parameter separately.
* Julliard and Kamenik (2004) who provide the perturbation solution equation system in tensor notation at any order k.
* Levintal (2017) who introduces permutation matrices to express the perturbation solution equation system in matrix notation up to fifth order.
Note that @rattoma already implemented the parameter derivatives of $g_x$ and $g_u$ analytically (and numerically), and I rely heavily on his work in `get_first_order_solution_params_derivs.m` (previously `getH.m`). My additions are mainly to this function and thus it is renamed to `get_perturbation_params_derivs.m`.
The basic idea of this merge request is to take the second- and third-order perturbation solution systems in Julliard and Kamenik (2004), unfold these into an equivalent matrix representation using permutation matrices as in Levintal (2017). Then extending Ratto and Iskrev (2012) one takes the derivative with respect to each parameter separately and gets a computational problem that is linear, albeit large, as it involves either solving generalized Sylvester equations or taking inverses of highly sparse matrices. I will now briefly summarize the perturbation solution system at third order and the system that results when taking the derivative with respect to parameters.
## Perturbation Solution
The following systems arise at first, second, and third order:
$(ghx): f_{x} z_{x} = f_{y_{-}^*} + f_{y_0} g_{x} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{x}= A g_{x} + f_{y_{-}^*}=0$
$(ghu): f_{z} z_{u} = f_{y_0} g_{u} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{u} + f_{u}= A g_u + f_u = 0$
$(ghxx) : A g_{xx} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{x}\right) + f_{zz} \left( z_{x} \otimes z_{x} \right) = 0$
$(ghxu) : A g_{xu} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{x} \otimes z_{u} \right) = 0$
$(ghuu) : A g_{uu} + B g_{xx} \left(g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{u} \otimes z_{u} \right) = 0$
$(ghs2) : (A+B) g_{\sigma\sigma} + \left( f_{y^{**}_{+}y^{**}_{+}} \left(g^{**}_{u} \otimes g^{**}_{u}\right) + f_{y^{**}_{+}} g^{**}_{uu}\right)vec(\Sigma) = 0$
$(ghxxx) : A g_{xxx} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{x}\right) + f_{y_{+}}g^{**}_{xx} \left(g^{*}_x \otimes g^{*}_{xx}\right)P_{x\_xx} + f_{zz} \left( z_{x} \otimes z_{xx} \right)P_{x\_xx} + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{x} \right) = 0$
$(ghxxu) : A g_{xxu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{u} \right) + f_{zz} \left( \left( z_{x} \otimes z_{xu} \right)P_{x\_xu} + \left(z_{xx} \otimes z_{u}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{x} \otimes g^{*}_{xu}\right)P_{x\_xu} + \left(g^{*}_{xx} \otimes g^{*}_{u}\right) \right) = 0$
$(ghxuu) : A g_{xuu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( \left( z_{xu} \otimes z_{u} \right)P_{xu\_u} + \left(z_{x} \otimes z_{uu}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{xu} \otimes g^{*}_{u}\right)P_{xu\_u} + \left(g^{*}_{x} \otimes g^{*}_{uu}\right) \right) = 0$
$(ghuuu) : A g_{uuu} + B g_{xxx} \left(g^{*}_{u} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{u} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( z_{u} \otimes z_{uu} \right)P_{u\_uu} + f_{y_{+}}g^{**}_{xx} \left(g^{*}_{u} \otimes g^{*}_{uu}\right)P_{u\_uu} = 0$
$(ghx\sigma\sigma) : A g_{x\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_x + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{x} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{x} \otimes z_{\sigma\sigma}\right) + F_{xu_{+}u_{+}}\left(I_{n_x} \otimes vec(\Sigma)\right) = 0$
$F_{xu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_x^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{xu_{+}} \otimes z_{u_{+}} \right)P_{xu\_u} + \left(z_{x} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{x} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
$(ghu\sigma\sigma) : A g_{u\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_{u} + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{u} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{u} \otimes z_{\sigma\sigma}\right) + F_{uu_{+}u_{+}}\left(I_{n_u} \otimes vec(\Sigma_u)\right) = 0$
$F_{uu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_u^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{uu_{+}} \otimes z_{u_{+}} \right)P_{uu\_u} + \left(z_{u} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{u} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
A and B are the common perturbation matrices:
$A = f_{y_0} + \begin{pmatrix} \underbrace{0}_{n\times n_{static}} &\vdots& \underbrace{f_{y^{**}_{+}} \cdot g^{**}_{x}}_{n \times n_{spred}} &\vdots& \underbrace{0}_{n\times n_{frwd}} \end{pmatrix}$and $B = \begin{pmatrix} \underbrace{0}_{n \times n_{static}}&\vdots & \underbrace{0}_{n \times n_{pred}} & \vdots & \underbrace{f_{y^{**}_{+}}}_{n \times n_{sfwrd}} \end{pmatrix}$
and $z=(y_{-}^{\ast}; y; y_{+}^{\ast\ast}; u)$ denotes the dynamic model variables as in `M_.lead_lag_incidence`, $y^\ast$ denote state variables, $y^{\ast\ast}$ denote forward looking variables, $y_+$ denote the variables with a lead, $y_{-}$ denote variables with a lag, $y_0$ denote variables at period t, $f$ the model equations, and $f_z$ the first-order dynamic model derivatives, $f_{zz}$ the second-order dynamic derivatives, and $f_{zzz}$ the third-order dynamic model derivatives. Then:
$z_{x} = \begin{pmatrix}I\\g_{x}\\g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{u} =\begin{pmatrix}0\\g_{u}\\g^{**}_{x} \cdot g^{*}_{u}\\I\end{pmatrix}$, $z_{u_{+}} =\begin{pmatrix}0\\0\\g^{**}_{u}\\0\end{pmatrix}$
$z_{xx} = \begin{pmatrix} 0\\g_{xx}\\g^{**}_{x} \left( g^{*}_x \otimes g^{*}_{x} \right) + g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{xu} =\begin{pmatrix}0\\g_{xu}\\g^{**}_{xx} \left( g^{*}_x \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{xu}\\0\end{pmatrix}$, $z_{uu} =\begin{pmatrix}0\\g_{uu}\\g^{**}_{xx} \left( g^{*}_u \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{uu}\\0\end{pmatrix}$,
$z_{xu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_x \otimes I \right)\\0\end{pmatrix}$, $z_{uu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_{u} \otimes I \right)\\0\end{pmatrix}$, $z_{u_{+}u_{+}} =\begin{pmatrix}0\\0\\g^{\ast\ast}_{uu}\\0\end{pmatrix}$, $z_{\sigma\sigma} = \begin{pmatrix}0\\ g_{\sigma\sigma}\\ g^{\ast\ast}_{x}g^{\ast}_{\sigma\sigma} + g^{\ast\ast}_{\sigma\sigma}\\0 \end{pmatrix}$
$P$ are permutation matrices that can be computed using Matlab's `ipermute` function.
## Parameter derivatives of perturbation solutions
First, we need the parameter derivatives of first, second, third, and fourth derivatives of the dynamic model (i.e. g1,g2,g3,g4 in dynamic files), I make use of the implicit function theorem: Let $f_{z^k}$ denote the kth derivative (wrt all dynamic variables) of the dynamic model, then let $df_{z^k}$ denote the first-derivative (wrt all model parameters) of $f_{z^k}$ evaluated at the steady state. Note that $f_{z^k}$ is a function of both the model parameters $\theta$ and of the steady state of all dynamic variables $\bar{z}$, which also depend on the parameters. Hence, implicitly $f_{z^k}=f_{z^k}(\theta,\bar{z}(\theta))$ and $df_{z^k}$ consists of two parts:
1. direct derivative wrt to all model parameters given by the preprocessor in the `_params_derivs.m` files
2. contribution of derivative of steady state of dynamic variables (wrt all model parameters): $f_{z^{k+1}} \cdot d\bar{z}$
Note that we already have functionality to compute $d\bar{z}$ analytically.
Having this, the above perturbation systems are basically equations of the following types
$AX +BXC = RHS$ or $AX = RHS$
Now when taking the derivative (wrt to one single parameter $\theta_j$), we get
$A\mathrm{d}\{X\} + B\mathrm{d}\{X\}C = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X - \mathrm{d}\{B\}XC - BX\mathrm{d}\{C\}$
or
$A\mathrm{d}\{X\} = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X$
The first one is a Sylvester type equation, the second one can be solved by taking the inverse of $A$. The only diffculty and tedious work arrises in computing (the highly sparse) derivatives of $RHS$.
***
# New functions: `
## get_perturbation_params_derivs.m`and `get_perturbation_params_derivs_numerical_objective.m`
* The parameter derivatives up to third order are computed in the new function`get_perturbation_params_derivs.m` both analytically and numerically. For numerical derivatives `get_perturbation_params_derivs_numerical_objective.m` is the objective for `fjaco.m` or `hessian_sparse.m` or `hessian.m`.
* `get_perturbation_params_derivs.m` is basically an extended version of the previous `get_first_order_solution_params_derivs.m` function.
* * `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It is used for numerical derivatives, whenever `analytic_derivation_mode=-1|-2`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
* Detailed changes:
* Most important: notation of this function is now in accordance to the k_order_solver, i.e. we do not compute derivatives of Kalman transition matrices A and B, but rather the solution matrices ghx,ghu,ghxx,ghxu,ghuu,ghs2,ghxxx,ghxxu,ghxuu,ghuuu,ghxss,ghuss in accordance with notation used in `oo_.dr`. As a byproduct at first-order, focusing on ghx and ghu instead of Kalman transition matrices A and B makes the computations slightly faster for large models (e.g. for Quest the computations were faster by a couple of seconds, not much, but okay).
* Removed use of `kstate`, see also Dynare/dynare#1653 and Dynare/dynare!1656
* Output arguments are stored in a structure `DERIVS`, there is also a flag `d2flag` that computes parameter hessians needed only in `dsge_likelihood.m`.
* Removed `kronflag` as input. `options_.analytic_derivation_mode` is now used instead of `kronflag`.
* Removed `indvar`, an index that was used to selected specific variables in the derivatives. This is not needed, as we always compute the parameter derivatives for all variables first and then select a subset of variables. The selection now takes place in other functions, like `dsge_likelihood.m`.
* Introduced some checks: (i) deterministic exogenous variables are not supported, (ii) Kronecker method only compatible with first-order approximation so reset to sylvester method, (iii) for purely backward or forward models we need to be careful with the rows in `M_.lead_la g_incidence`, (iv) if `_params_derivs.m` files are missing an error is thrown.
* For numerical derivatives, if mod file does not contain an `estimated_params_block`, a temporary one with the most important parameter information is created.
## `unfold_g4.m`
* When evaluating g3 and g4 one needs to take into account that these do not contain symmetric elements, so one needs to use `unfold_g3.m` and the new function `unfold_g4.m`. This returns an unfolded version of the same matrix (i.e. with symmetric elements).
***
# New test models
`.gitignore` and `Makefile.am` are changed accordingly. Also now it is possible to run test suite on analytic_derivatives, i.e. run `make check m/analytic_derivatives`
## `analytic_derivatives/BrockMirman_PertParamsDerivs.mod`
* This is the Brock Mirman model, where we know the exact policy function $g$ for capital and consumption. As this does not imply a nonzero $g_{\sigma\sigma}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ I added some artificial equations to get nonzero solution matrices with respect to $\sigma$. The true perturbation solution matrices $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ are then computed analytically with Matlab's symbolic toolbox and saved in `nBrockMirmanSYM.mat`. There is a preprocessor flag that recreates these analytical computations if changes are needed (and to check whether I made some errors here ;-) )
* Then solution matrices up to third order and their parameter Jacobians are then compared to the ones computed by Dynare's `k_order_solver` and by `get_perturbation_params_derivs` for all `analytic_derivation_mode`'s. There will be an error if the maximum absolute deviation is too large, i.e. for numerical derivatives (`analytic_derivation_mode=-1|-2`) the tolerance is choosen lower (around 1e-5); for analytical methods we are stricter: around 1e-13 for first-order, 1e-12 for second order, and 1e-11 for third-order.
* As a side note, this mod file also checks Dynare's `k_order_solver` algorithm and throws an error if something is wrong.
* This test model shows that the new functionality works well. And analytical derivatives perform way better and accurate than numerical ones, even for this small model.
## `analytic_derivatives/burnside_3_order_PertParamsDerivs.mod`
* This builds upon `tests/k_order_perturbation/burnside_k_order.mod` and computes the true parameter derivatives analytically by hand.
* This test model also shows that the new functionality works well.
## `analytic_derivatives/LindeTrabandt2019.mod`
* Shows that the new functionality also works for medium-sized models, i.e. a SW type model solved at third order with 35 variables (11 states). 2 shocks and 20 parameters.
* This mod file can be used to tweak the speed of the computations in the future.
* Compares numerical versus analytical parameter derivatives (for first, second and third order). Note that this model clearly shows that numerical ones are quite different than analytical ones even at first order!
## `identification/LindeTrabandt2019_xfail.mod`
* This model is a check for issue Dynare/dynare#1595, see fjaco.m below, and will fail.
* Removed `analytic_derivatives/ls2003.mod` as this mod file is neither in the testsuite nor does it work.
***
# Detailed changes in other functions
## `get_first_order_solution_params_derivs.m`
* Deleted, or actually, renamed to `get_perturbation_params_derivs.m`, as this function now is able to compute the derivatives up to third order
## `identification_numerical_objective.m`
* `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
## `dsge_likelihood.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `T`matrix anymore, but instead of the dynare solution matrix `ghx`. So we recreate `T` here as this amounts to adding some zeros and focusing on selected variables only.
* Added some checks to make sure the first-order approximation is selected.
* Removed `kron_flag` as input, as `get_perturbation_params_derivs` looks into `options_.analytic_derivation_mode` for `kron_flag`.
## `dynare_identification.m`
* make sure that setting `analytic_derivation_mode` is set both in `options_ident` and `options_`. Note that at the end of the function we restore the `options_` structure, so all changes are local. In a next merge request, I will remove the global variables to make all variables local.
## `get_identification_jacobians.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `A` and `B` matrix anymore, but instead of the dynare solution matrix `ghx` and `ghu`. So we recreate these matrices here instead of in `get_perturbation_params_derivs.m`.
* Added `str2func` for better function handles in `fjaco.m`.
## `fjaco.m`
* make `tol`an option, which can be adjusted by changing `options_.dynatol.x`for identification and parameter derivatives purposes.
* include a check and an informative error message, if numerical derivatives (two-sided finite difference method) yield errors in `resol.m` for identification and parameter derivatives purposes. This closes issue Dynare/dynare#1595.
* Changed year of copyright to 2010-2017,2019
***
# Further suggestions and questions
* Ones this is merged, I will merge request an improvement of the identification toolbox, which will work up to third order using the pruned state space. This will also remove some issues and bugs, and also I will remove global variables in this request.
* The third-order derivatives can be further improved by taking sparsity into account and use mex versions for kronecker products etc. I leave this for further testing (and if anybody actually uses this ;-) )
2019-12-17 19:17:09 +01:00
daux_ZP_uup_up_Z_u_upup = dZ_uup_up * P_uu_u + dZ_u_upup ;
dFuupup = dg3 ( : , k3yy0ex0 ( kyy0ex0 , kyy0ex0 , kyy0ex0 ) + ( jp - 1 ) * yy0ex0_nbr ^3 ) * Z_u_up_up + g3 ( : , k3yy0ex0 ( kyy0ex0 , kyy0ex0 , kyy0ex0 ) ) * dZ_u_up_up ...
2019-12-20 16:28:06 +01:00
+ dg2 ( : , k2yy0ex0 ( kyy0ex0 , kyy0ex0 ) + ( jp - 1 ) * yy0ex0_nbr ^2 ) * aux_ZP_uup_up_Z_u_upup + g2 ( : , k2yy0ex0 ( kyy0ex0 , kyy0ex0 ) ) * daux_ZP_uup_up_Z_u_upup ...
+ dg1 ( : , nonzeros ( klead ) , jp ) * oo . dr . ghxuu ( klead ~= 0 , : ) * kron ( oo . dr . ghu ( idx_states , : ) , eye ( M . exo_nbr ^2 ) ) + g1 ( : , nonzeros ( klead ) ) * dghxuu ( klead ~= 0 , : , jp ) * kron ( oo . dr . ghu ( idx_states , : ) , eye ( M . exo_nbr ^2 ) ) + g1 ( : , nonzeros ( klead ) ) * oo . dr . ghxuu ( klead ~= 0 , : ) * kron ( dghu ( idx_states , : , jp ) , eye ( M . exo_nbr ^2 ) ) ;
Added parameter derivatives of perturbation solution up to 3 order
# Preliminary comments
I finished the identification toolbox at orders two and three using the pruned state space system, but before I merge request this, I decided to first merge the new functionality to compute parameter derivatives of perturbation solution matrices at higher orders. So after this is approved, I merge the identification toolbox.
I guess @rattoma, @sebastien, and @michel are best choices to review this.
I outline the main idea first and then provide some more detailed changes I made to the functions.
***
# Main idea
This merge request is concerned with the *analytical*computation of the parameter derivatives of first, second and third order perturbation solution matrices, i.e. using _closed-form_ expressions to efficiently compute the derivative of $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ *with respect to model parameters* $\theta$. Note that $\theta$ contains model parameters, stderr and corr parameters of shocks. stderr and corr parameters of measurement errors are not yet supported, (they can easily be included as exogenous shocks). The availability of such derivatives is beneficial in terms of more reliable analysis of model sensitivity and parameter identifiability as well as more efficient estimation methods, in particular for models solved up to third order, as it is well-known that numerical derivatives are a tricky business, especially for large models.
References for my approach are:
* Iskrev (2008, 2010) and Schmitt-Grohé and Uribe (2012, Appendix) who were the first to compute the parameter derivatives analytically at first order, however, using inefficient (sparse) Kronecker products.
* Mutschler (2015) who provides the expressions for a second-order, but again using inefficient (sparse) Kronecker products.
* Ratto and Iskrev (2012) who show how the first-order system can be solved accurately, fast and efficiently using existing numerical algorithms for generalized Sylvester equations by taking the parameter derivative with respect to each parameter separately.
* Julliard and Kamenik (2004) who provide the perturbation solution equation system in tensor notation at any order k.
* Levintal (2017) who introduces permutation matrices to express the perturbation solution equation system in matrix notation up to fifth order.
Note that @rattoma already implemented the parameter derivatives of $g_x$ and $g_u$ analytically (and numerically), and I rely heavily on his work in `get_first_order_solution_params_derivs.m` (previously `getH.m`). My additions are mainly to this function and thus it is renamed to `get_perturbation_params_derivs.m`.
The basic idea of this merge request is to take the second- and third-order perturbation solution systems in Julliard and Kamenik (2004), unfold these into an equivalent matrix representation using permutation matrices as in Levintal (2017). Then extending Ratto and Iskrev (2012) one takes the derivative with respect to each parameter separately and gets a computational problem that is linear, albeit large, as it involves either solving generalized Sylvester equations or taking inverses of highly sparse matrices. I will now briefly summarize the perturbation solution system at third order and the system that results when taking the derivative with respect to parameters.
## Perturbation Solution
The following systems arise at first, second, and third order:
$(ghx): f_{x} z_{x} = f_{y_{-}^*} + f_{y_0} g_{x} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{x}= A g_{x} + f_{y_{-}^*}=0$
$(ghu): f_{z} z_{u} = f_{y_0} g_{u} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{u} + f_{u}= A g_u + f_u = 0$
$(ghxx) : A g_{xx} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{x}\right) + f_{zz} \left( z_{x} \otimes z_{x} \right) = 0$
$(ghxu) : A g_{xu} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{x} \otimes z_{u} \right) = 0$
$(ghuu) : A g_{uu} + B g_{xx} \left(g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{u} \otimes z_{u} \right) = 0$
$(ghs2) : (A+B) g_{\sigma\sigma} + \left( f_{y^{**}_{+}y^{**}_{+}} \left(g^{**}_{u} \otimes g^{**}_{u}\right) + f_{y^{**}_{+}} g^{**}_{uu}\right)vec(\Sigma) = 0$
$(ghxxx) : A g_{xxx} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{x}\right) + f_{y_{+}}g^{**}_{xx} \left(g^{*}_x \otimes g^{*}_{xx}\right)P_{x\_xx} + f_{zz} \left( z_{x} \otimes z_{xx} \right)P_{x\_xx} + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{x} \right) = 0$
$(ghxxu) : A g_{xxu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{u} \right) + f_{zz} \left( \left( z_{x} \otimes z_{xu} \right)P_{x\_xu} + \left(z_{xx} \otimes z_{u}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{x} \otimes g^{*}_{xu}\right)P_{x\_xu} + \left(g^{*}_{xx} \otimes g^{*}_{u}\right) \right) = 0$
$(ghxuu) : A g_{xuu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( \left( z_{xu} \otimes z_{u} \right)P_{xu\_u} + \left(z_{x} \otimes z_{uu}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{xu} \otimes g^{*}_{u}\right)P_{xu\_u} + \left(g^{*}_{x} \otimes g^{*}_{uu}\right) \right) = 0$
$(ghuuu) : A g_{uuu} + B g_{xxx} \left(g^{*}_{u} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{u} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( z_{u} \otimes z_{uu} \right)P_{u\_uu} + f_{y_{+}}g^{**}_{xx} \left(g^{*}_{u} \otimes g^{*}_{uu}\right)P_{u\_uu} = 0$
$(ghx\sigma\sigma) : A g_{x\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_x + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{x} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{x} \otimes z_{\sigma\sigma}\right) + F_{xu_{+}u_{+}}\left(I_{n_x} \otimes vec(\Sigma)\right) = 0$
$F_{xu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_x^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{xu_{+}} \otimes z_{u_{+}} \right)P_{xu\_u} + \left(z_{x} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{x} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
$(ghu\sigma\sigma) : A g_{u\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_{u} + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{u} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{u} \otimes z_{\sigma\sigma}\right) + F_{uu_{+}u_{+}}\left(I_{n_u} \otimes vec(\Sigma_u)\right) = 0$
$F_{uu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_u^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{uu_{+}} \otimes z_{u_{+}} \right)P_{uu\_u} + \left(z_{u} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{u} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
A and B are the common perturbation matrices:
$A = f_{y_0} + \begin{pmatrix} \underbrace{0}_{n\times n_{static}} &\vdots& \underbrace{f_{y^{**}_{+}} \cdot g^{**}_{x}}_{n \times n_{spred}} &\vdots& \underbrace{0}_{n\times n_{frwd}} \end{pmatrix}$and $B = \begin{pmatrix} \underbrace{0}_{n \times n_{static}}&\vdots & \underbrace{0}_{n \times n_{pred}} & \vdots & \underbrace{f_{y^{**}_{+}}}_{n \times n_{sfwrd}} \end{pmatrix}$
and $z=(y_{-}^{\ast}; y; y_{+}^{\ast\ast}; u)$ denotes the dynamic model variables as in `M_.lead_lag_incidence`, $y^\ast$ denote state variables, $y^{\ast\ast}$ denote forward looking variables, $y_+$ denote the variables with a lead, $y_{-}$ denote variables with a lag, $y_0$ denote variables at period t, $f$ the model equations, and $f_z$ the first-order dynamic model derivatives, $f_{zz}$ the second-order dynamic derivatives, and $f_{zzz}$ the third-order dynamic model derivatives. Then:
$z_{x} = \begin{pmatrix}I\\g_{x}\\g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{u} =\begin{pmatrix}0\\g_{u}\\g^{**}_{x} \cdot g^{*}_{u}\\I\end{pmatrix}$, $z_{u_{+}} =\begin{pmatrix}0\\0\\g^{**}_{u}\\0\end{pmatrix}$
$z_{xx} = \begin{pmatrix} 0\\g_{xx}\\g^{**}_{x} \left( g^{*}_x \otimes g^{*}_{x} \right) + g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{xu} =\begin{pmatrix}0\\g_{xu}\\g^{**}_{xx} \left( g^{*}_x \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{xu}\\0\end{pmatrix}$, $z_{uu} =\begin{pmatrix}0\\g_{uu}\\g^{**}_{xx} \left( g^{*}_u \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{uu}\\0\end{pmatrix}$,
$z_{xu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_x \otimes I \right)\\0\end{pmatrix}$, $z_{uu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_{u} \otimes I \right)\\0\end{pmatrix}$, $z_{u_{+}u_{+}} =\begin{pmatrix}0\\0\\g^{\ast\ast}_{uu}\\0\end{pmatrix}$, $z_{\sigma\sigma} = \begin{pmatrix}0\\ g_{\sigma\sigma}\\ g^{\ast\ast}_{x}g^{\ast}_{\sigma\sigma} + g^{\ast\ast}_{\sigma\sigma}\\0 \end{pmatrix}$
$P$ are permutation matrices that can be computed using Matlab's `ipermute` function.
## Parameter derivatives of perturbation solutions
First, we need the parameter derivatives of first, second, third, and fourth derivatives of the dynamic model (i.e. g1,g2,g3,g4 in dynamic files), I make use of the implicit function theorem: Let $f_{z^k}$ denote the kth derivative (wrt all dynamic variables) of the dynamic model, then let $df_{z^k}$ denote the first-derivative (wrt all model parameters) of $f_{z^k}$ evaluated at the steady state. Note that $f_{z^k}$ is a function of both the model parameters $\theta$ and of the steady state of all dynamic variables $\bar{z}$, which also depend on the parameters. Hence, implicitly $f_{z^k}=f_{z^k}(\theta,\bar{z}(\theta))$ and $df_{z^k}$ consists of two parts:
1. direct derivative wrt to all model parameters given by the preprocessor in the `_params_derivs.m` files
2. contribution of derivative of steady state of dynamic variables (wrt all model parameters): $f_{z^{k+1}} \cdot d\bar{z}$
Note that we already have functionality to compute $d\bar{z}$ analytically.
Having this, the above perturbation systems are basically equations of the following types
$AX +BXC = RHS$ or $AX = RHS$
Now when taking the derivative (wrt to one single parameter $\theta_j$), we get
$A\mathrm{d}\{X\} + B\mathrm{d}\{X\}C = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X - \mathrm{d}\{B\}XC - BX\mathrm{d}\{C\}$
or
$A\mathrm{d}\{X\} = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X$
The first one is a Sylvester type equation, the second one can be solved by taking the inverse of $A$. The only diffculty and tedious work arrises in computing (the highly sparse) derivatives of $RHS$.
***
# New functions: `
## get_perturbation_params_derivs.m`and `get_perturbation_params_derivs_numerical_objective.m`
* The parameter derivatives up to third order are computed in the new function`get_perturbation_params_derivs.m` both analytically and numerically. For numerical derivatives `get_perturbation_params_derivs_numerical_objective.m` is the objective for `fjaco.m` or `hessian_sparse.m` or `hessian.m`.
* `get_perturbation_params_derivs.m` is basically an extended version of the previous `get_first_order_solution_params_derivs.m` function.
* * `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It is used for numerical derivatives, whenever `analytic_derivation_mode=-1|-2`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
* Detailed changes:
* Most important: notation of this function is now in accordance to the k_order_solver, i.e. we do not compute derivatives of Kalman transition matrices A and B, but rather the solution matrices ghx,ghu,ghxx,ghxu,ghuu,ghs2,ghxxx,ghxxu,ghxuu,ghuuu,ghxss,ghuss in accordance with notation used in `oo_.dr`. As a byproduct at first-order, focusing on ghx and ghu instead of Kalman transition matrices A and B makes the computations slightly faster for large models (e.g. for Quest the computations were faster by a couple of seconds, not much, but okay).
* Removed use of `kstate`, see also Dynare/dynare#1653 and Dynare/dynare!1656
* Output arguments are stored in a structure `DERIVS`, there is also a flag `d2flag` that computes parameter hessians needed only in `dsge_likelihood.m`.
* Removed `kronflag` as input. `options_.analytic_derivation_mode` is now used instead of `kronflag`.
* Removed `indvar`, an index that was used to selected specific variables in the derivatives. This is not needed, as we always compute the parameter derivatives for all variables first and then select a subset of variables. The selection now takes place in other functions, like `dsge_likelihood.m`.
* Introduced some checks: (i) deterministic exogenous variables are not supported, (ii) Kronecker method only compatible with first-order approximation so reset to sylvester method, (iii) for purely backward or forward models we need to be careful with the rows in `M_.lead_la g_incidence`, (iv) if `_params_derivs.m` files are missing an error is thrown.
* For numerical derivatives, if mod file does not contain an `estimated_params_block`, a temporary one with the most important parameter information is created.
## `unfold_g4.m`
* When evaluating g3 and g4 one needs to take into account that these do not contain symmetric elements, so one needs to use `unfold_g3.m` and the new function `unfold_g4.m`. This returns an unfolded version of the same matrix (i.e. with symmetric elements).
***
# New test models
`.gitignore` and `Makefile.am` are changed accordingly. Also now it is possible to run test suite on analytic_derivatives, i.e. run `make check m/analytic_derivatives`
## `analytic_derivatives/BrockMirman_PertParamsDerivs.mod`
* This is the Brock Mirman model, where we know the exact policy function $g$ for capital and consumption. As this does not imply a nonzero $g_{\sigma\sigma}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ I added some artificial equations to get nonzero solution matrices with respect to $\sigma$. The true perturbation solution matrices $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ are then computed analytically with Matlab's symbolic toolbox and saved in `nBrockMirmanSYM.mat`. There is a preprocessor flag that recreates these analytical computations if changes are needed (and to check whether I made some errors here ;-) )
* Then solution matrices up to third order and their parameter Jacobians are then compared to the ones computed by Dynare's `k_order_solver` and by `get_perturbation_params_derivs` for all `analytic_derivation_mode`'s. There will be an error if the maximum absolute deviation is too large, i.e. for numerical derivatives (`analytic_derivation_mode=-1|-2`) the tolerance is choosen lower (around 1e-5); for analytical methods we are stricter: around 1e-13 for first-order, 1e-12 for second order, and 1e-11 for third-order.
* As a side note, this mod file also checks Dynare's `k_order_solver` algorithm and throws an error if something is wrong.
* This test model shows that the new functionality works well. And analytical derivatives perform way better and accurate than numerical ones, even for this small model.
## `analytic_derivatives/burnside_3_order_PertParamsDerivs.mod`
* This builds upon `tests/k_order_perturbation/burnside_k_order.mod` and computes the true parameter derivatives analytically by hand.
* This test model also shows that the new functionality works well.
## `analytic_derivatives/LindeTrabandt2019.mod`
* Shows that the new functionality also works for medium-sized models, i.e. a SW type model solved at third order with 35 variables (11 states). 2 shocks and 20 parameters.
* This mod file can be used to tweak the speed of the computations in the future.
* Compares numerical versus analytical parameter derivatives (for first, second and third order). Note that this model clearly shows that numerical ones are quite different than analytical ones even at first order!
## `identification/LindeTrabandt2019_xfail.mod`
* This model is a check for issue Dynare/dynare#1595, see fjaco.m below, and will fail.
* Removed `analytic_derivatives/ls2003.mod` as this mod file is neither in the testsuite nor does it work.
***
# Detailed changes in other functions
## `get_first_order_solution_params_derivs.m`
* Deleted, or actually, renamed to `get_perturbation_params_derivs.m`, as this function now is able to compute the derivatives up to third order
## `identification_numerical_objective.m`
* `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
## `dsge_likelihood.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `T`matrix anymore, but instead of the dynare solution matrix `ghx`. So we recreate `T` here as this amounts to adding some zeros and focusing on selected variables only.
* Added some checks to make sure the first-order approximation is selected.
* Removed `kron_flag` as input, as `get_perturbation_params_derivs` looks into `options_.analytic_derivation_mode` for `kron_flag`.
## `dynare_identification.m`
* make sure that setting `analytic_derivation_mode` is set both in `options_ident` and `options_`. Note that at the end of the function we restore the `options_` structure, so all changes are local. In a next merge request, I will remove the global variables to make all variables local.
## `get_identification_jacobians.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `A` and `B` matrix anymore, but instead of the dynare solution matrix `ghx` and `ghu`. So we recreate these matrices here instead of in `get_perturbation_params_derivs.m`.
* Added `str2func` for better function handles in `fjaco.m`.
## `fjaco.m`
* make `tol`an option, which can be adjusted by changing `options_.dynatol.x`for identification and parameter derivatives purposes.
* include a check and an informative error message, if numerical derivatives (two-sided finite difference method) yield errors in `resol.m` for identification and parameter derivatives purposes. This closes issue Dynare/dynare#1595.
* Changed year of copyright to 2010-2017,2019
***
# Further suggestions and questions
* Ones this is merged, I will merge request an improvement of the identification toolbox, which will work up to third order using the pruned state space. This will also remove some issues and bugs, and also I will remove global variables in this request.
* The third-order derivatives can be further improved by taking sparsity into account and use mex versions for kronecker products etc. I leave this for further testing (and if anybody actually uses this ;-) )
2019-12-17 19:17:09 +01:00
dRHSuss ( : , : , stderrparam_nbr + corrparam_nbr + jp ) = dA ( : , : , jp ) * oo . dr . ghuss + dB ( : , : , jp ) * oo . dr . ghxss * oo . dr . ghu ( idx_states , : ) + B * oo . dr . ghxss * dghu ( idx_states , : , jp ) ; %missing dghxss
dRHSuss ( : , : , stderrparam_nbr + corrparam_nbr + jp ) = dRHSuss ( : , : , stderrparam_nbr + corrparam_nbr + jp ) + dfyp_ghxxp ( : , : , jp ) * G_u_ss + fyp_ghxxp * dG_u_ss ;
dRHSuss ( : , : , stderrparam_nbr + corrparam_nbr + jp ) = dRHSuss ( : , : , stderrparam_nbr + corrparam_nbr + jp ) + dg2 ( : , k2yy0ex0 ( kyy0ex0 , kyy0ex0 ) + ( jp - 1 ) * yy0ex0_nbr ^2 ) * Z_u_ss + g2 ( : , k2yy0ex0 ( kyy0ex0 , kyy0ex0 ) ) * dZ_u_ss ;
2019-12-20 16:28:06 +01:00
dRHSuss ( : , : , stderrparam_nbr + corrparam_nbr + jp ) = dRHSuss ( : , : , stderrparam_nbr + corrparam_nbr + jp ) + dFuupup * Iu_vecSig_e ; %contribution by dSigma_e only for stderr and corr params
Added parameter derivatives of perturbation solution up to 3 order
# Preliminary comments
I finished the identification toolbox at orders two and three using the pruned state space system, but before I merge request this, I decided to first merge the new functionality to compute parameter derivatives of perturbation solution matrices at higher orders. So after this is approved, I merge the identification toolbox.
I guess @rattoma, @sebastien, and @michel are best choices to review this.
I outline the main idea first and then provide some more detailed changes I made to the functions.
***
# Main idea
This merge request is concerned with the *analytical*computation of the parameter derivatives of first, second and third order perturbation solution matrices, i.e. using _closed-form_ expressions to efficiently compute the derivative of $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ *with respect to model parameters* $\theta$. Note that $\theta$ contains model parameters, stderr and corr parameters of shocks. stderr and corr parameters of measurement errors are not yet supported, (they can easily be included as exogenous shocks). The availability of such derivatives is beneficial in terms of more reliable analysis of model sensitivity and parameter identifiability as well as more efficient estimation methods, in particular for models solved up to third order, as it is well-known that numerical derivatives are a tricky business, especially for large models.
References for my approach are:
* Iskrev (2008, 2010) and Schmitt-Grohé and Uribe (2012, Appendix) who were the first to compute the parameter derivatives analytically at first order, however, using inefficient (sparse) Kronecker products.
* Mutschler (2015) who provides the expressions for a second-order, but again using inefficient (sparse) Kronecker products.
* Ratto and Iskrev (2012) who show how the first-order system can be solved accurately, fast and efficiently using existing numerical algorithms for generalized Sylvester equations by taking the parameter derivative with respect to each parameter separately.
* Julliard and Kamenik (2004) who provide the perturbation solution equation system in tensor notation at any order k.
* Levintal (2017) who introduces permutation matrices to express the perturbation solution equation system in matrix notation up to fifth order.
Note that @rattoma already implemented the parameter derivatives of $g_x$ and $g_u$ analytically (and numerically), and I rely heavily on his work in `get_first_order_solution_params_derivs.m` (previously `getH.m`). My additions are mainly to this function and thus it is renamed to `get_perturbation_params_derivs.m`.
The basic idea of this merge request is to take the second- and third-order perturbation solution systems in Julliard and Kamenik (2004), unfold these into an equivalent matrix representation using permutation matrices as in Levintal (2017). Then extending Ratto and Iskrev (2012) one takes the derivative with respect to each parameter separately and gets a computational problem that is linear, albeit large, as it involves either solving generalized Sylvester equations or taking inverses of highly sparse matrices. I will now briefly summarize the perturbation solution system at third order and the system that results when taking the derivative with respect to parameters.
## Perturbation Solution
The following systems arise at first, second, and third order:
$(ghx): f_{x} z_{x} = f_{y_{-}^*} + f_{y_0} g_{x} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{x}= A g_{x} + f_{y_{-}^*}=0$
$(ghu): f_{z} z_{u} = f_{y_0} g_{u} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{u} + f_{u}= A g_u + f_u = 0$
$(ghxx) : A g_{xx} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{x}\right) + f_{zz} \left( z_{x} \otimes z_{x} \right) = 0$
$(ghxu) : A g_{xu} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{x} \otimes z_{u} \right) = 0$
$(ghuu) : A g_{uu} + B g_{xx} \left(g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{u} \otimes z_{u} \right) = 0$
$(ghs2) : (A+B) g_{\sigma\sigma} + \left( f_{y^{**}_{+}y^{**}_{+}} \left(g^{**}_{u} \otimes g^{**}_{u}\right) + f_{y^{**}_{+}} g^{**}_{uu}\right)vec(\Sigma) = 0$
$(ghxxx) : A g_{xxx} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{x}\right) + f_{y_{+}}g^{**}_{xx} \left(g^{*}_x \otimes g^{*}_{xx}\right)P_{x\_xx} + f_{zz} \left( z_{x} \otimes z_{xx} \right)P_{x\_xx} + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{x} \right) = 0$
$(ghxxu) : A g_{xxu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{u} \right) + f_{zz} \left( \left( z_{x} \otimes z_{xu} \right)P_{x\_xu} + \left(z_{xx} \otimes z_{u}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{x} \otimes g^{*}_{xu}\right)P_{x\_xu} + \left(g^{*}_{xx} \otimes g^{*}_{u}\right) \right) = 0$
$(ghxuu) : A g_{xuu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( \left( z_{xu} \otimes z_{u} \right)P_{xu\_u} + \left(z_{x} \otimes z_{uu}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{xu} \otimes g^{*}_{u}\right)P_{xu\_u} + \left(g^{*}_{x} \otimes g^{*}_{uu}\right) \right) = 0$
$(ghuuu) : A g_{uuu} + B g_{xxx} \left(g^{*}_{u} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{u} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( z_{u} \otimes z_{uu} \right)P_{u\_uu} + f_{y_{+}}g^{**}_{xx} \left(g^{*}_{u} \otimes g^{*}_{uu}\right)P_{u\_uu} = 0$
$(ghx\sigma\sigma) : A g_{x\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_x + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{x} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{x} \otimes z_{\sigma\sigma}\right) + F_{xu_{+}u_{+}}\left(I_{n_x} \otimes vec(\Sigma)\right) = 0$
$F_{xu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_x^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{xu_{+}} \otimes z_{u_{+}} \right)P_{xu\_u} + \left(z_{x} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{x} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
$(ghu\sigma\sigma) : A g_{u\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_{u} + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{u} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{u} \otimes z_{\sigma\sigma}\right) + F_{uu_{+}u_{+}}\left(I_{n_u} \otimes vec(\Sigma_u)\right) = 0$
$F_{uu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_u^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{uu_{+}} \otimes z_{u_{+}} \right)P_{uu\_u} + \left(z_{u} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{u} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
A and B are the common perturbation matrices:
$A = f_{y_0} + \begin{pmatrix} \underbrace{0}_{n\times n_{static}} &\vdots& \underbrace{f_{y^{**}_{+}} \cdot g^{**}_{x}}_{n \times n_{spred}} &\vdots& \underbrace{0}_{n\times n_{frwd}} \end{pmatrix}$and $B = \begin{pmatrix} \underbrace{0}_{n \times n_{static}}&\vdots & \underbrace{0}_{n \times n_{pred}} & \vdots & \underbrace{f_{y^{**}_{+}}}_{n \times n_{sfwrd}} \end{pmatrix}$
and $z=(y_{-}^{\ast}; y; y_{+}^{\ast\ast}; u)$ denotes the dynamic model variables as in `M_.lead_lag_incidence`, $y^\ast$ denote state variables, $y^{\ast\ast}$ denote forward looking variables, $y_+$ denote the variables with a lead, $y_{-}$ denote variables with a lag, $y_0$ denote variables at period t, $f$ the model equations, and $f_z$ the first-order dynamic model derivatives, $f_{zz}$ the second-order dynamic derivatives, and $f_{zzz}$ the third-order dynamic model derivatives. Then:
$z_{x} = \begin{pmatrix}I\\g_{x}\\g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{u} =\begin{pmatrix}0\\g_{u}\\g^{**}_{x} \cdot g^{*}_{u}\\I\end{pmatrix}$, $z_{u_{+}} =\begin{pmatrix}0\\0\\g^{**}_{u}\\0\end{pmatrix}$
$z_{xx} = \begin{pmatrix} 0\\g_{xx}\\g^{**}_{x} \left( g^{*}_x \otimes g^{*}_{x} \right) + g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{xu} =\begin{pmatrix}0\\g_{xu}\\g^{**}_{xx} \left( g^{*}_x \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{xu}\\0\end{pmatrix}$, $z_{uu} =\begin{pmatrix}0\\g_{uu}\\g^{**}_{xx} \left( g^{*}_u \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{uu}\\0\end{pmatrix}$,
$z_{xu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_x \otimes I \right)\\0\end{pmatrix}$, $z_{uu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_{u} \otimes I \right)\\0\end{pmatrix}$, $z_{u_{+}u_{+}} =\begin{pmatrix}0\\0\\g^{\ast\ast}_{uu}\\0\end{pmatrix}$, $z_{\sigma\sigma} = \begin{pmatrix}0\\ g_{\sigma\sigma}\\ g^{\ast\ast}_{x}g^{\ast}_{\sigma\sigma} + g^{\ast\ast}_{\sigma\sigma}\\0 \end{pmatrix}$
$P$ are permutation matrices that can be computed using Matlab's `ipermute` function.
## Parameter derivatives of perturbation solutions
First, we need the parameter derivatives of first, second, third, and fourth derivatives of the dynamic model (i.e. g1,g2,g3,g4 in dynamic files), I make use of the implicit function theorem: Let $f_{z^k}$ denote the kth derivative (wrt all dynamic variables) of the dynamic model, then let $df_{z^k}$ denote the first-derivative (wrt all model parameters) of $f_{z^k}$ evaluated at the steady state. Note that $f_{z^k}$ is a function of both the model parameters $\theta$ and of the steady state of all dynamic variables $\bar{z}$, which also depend on the parameters. Hence, implicitly $f_{z^k}=f_{z^k}(\theta,\bar{z}(\theta))$ and $df_{z^k}$ consists of two parts:
1. direct derivative wrt to all model parameters given by the preprocessor in the `_params_derivs.m` files
2. contribution of derivative of steady state of dynamic variables (wrt all model parameters): $f_{z^{k+1}} \cdot d\bar{z}$
Note that we already have functionality to compute $d\bar{z}$ analytically.
Having this, the above perturbation systems are basically equations of the following types
$AX +BXC = RHS$ or $AX = RHS$
Now when taking the derivative (wrt to one single parameter $\theta_j$), we get
$A\mathrm{d}\{X\} + B\mathrm{d}\{X\}C = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X - \mathrm{d}\{B\}XC - BX\mathrm{d}\{C\}$
or
$A\mathrm{d}\{X\} = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X$
The first one is a Sylvester type equation, the second one can be solved by taking the inverse of $A$. The only diffculty and tedious work arrises in computing (the highly sparse) derivatives of $RHS$.
***
# New functions: `
## get_perturbation_params_derivs.m`and `get_perturbation_params_derivs_numerical_objective.m`
* The parameter derivatives up to third order are computed in the new function`get_perturbation_params_derivs.m` both analytically and numerically. For numerical derivatives `get_perturbation_params_derivs_numerical_objective.m` is the objective for `fjaco.m` or `hessian_sparse.m` or `hessian.m`.
* `get_perturbation_params_derivs.m` is basically an extended version of the previous `get_first_order_solution_params_derivs.m` function.
* * `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It is used for numerical derivatives, whenever `analytic_derivation_mode=-1|-2`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
* Detailed changes:
* Most important: notation of this function is now in accordance to the k_order_solver, i.e. we do not compute derivatives of Kalman transition matrices A and B, but rather the solution matrices ghx,ghu,ghxx,ghxu,ghuu,ghs2,ghxxx,ghxxu,ghxuu,ghuuu,ghxss,ghuss in accordance with notation used in `oo_.dr`. As a byproduct at first-order, focusing on ghx and ghu instead of Kalman transition matrices A and B makes the computations slightly faster for large models (e.g. for Quest the computations were faster by a couple of seconds, not much, but okay).
* Removed use of `kstate`, see also Dynare/dynare#1653 and Dynare/dynare!1656
* Output arguments are stored in a structure `DERIVS`, there is also a flag `d2flag` that computes parameter hessians needed only in `dsge_likelihood.m`.
* Removed `kronflag` as input. `options_.analytic_derivation_mode` is now used instead of `kronflag`.
* Removed `indvar`, an index that was used to selected specific variables in the derivatives. This is not needed, as we always compute the parameter derivatives for all variables first and then select a subset of variables. The selection now takes place in other functions, like `dsge_likelihood.m`.
* Introduced some checks: (i) deterministic exogenous variables are not supported, (ii) Kronecker method only compatible with first-order approximation so reset to sylvester method, (iii) for purely backward or forward models we need to be careful with the rows in `M_.lead_la g_incidence`, (iv) if `_params_derivs.m` files are missing an error is thrown.
* For numerical derivatives, if mod file does not contain an `estimated_params_block`, a temporary one with the most important parameter information is created.
## `unfold_g4.m`
* When evaluating g3 and g4 one needs to take into account that these do not contain symmetric elements, so one needs to use `unfold_g3.m` and the new function `unfold_g4.m`. This returns an unfolded version of the same matrix (i.e. with symmetric elements).
***
# New test models
`.gitignore` and `Makefile.am` are changed accordingly. Also now it is possible to run test suite on analytic_derivatives, i.e. run `make check m/analytic_derivatives`
## `analytic_derivatives/BrockMirman_PertParamsDerivs.mod`
* This is the Brock Mirman model, where we know the exact policy function $g$ for capital and consumption. As this does not imply a nonzero $g_{\sigma\sigma}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ I added some artificial equations to get nonzero solution matrices with respect to $\sigma$. The true perturbation solution matrices $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ are then computed analytically with Matlab's symbolic toolbox and saved in `nBrockMirmanSYM.mat`. There is a preprocessor flag that recreates these analytical computations if changes are needed (and to check whether I made some errors here ;-) )
* Then solution matrices up to third order and their parameter Jacobians are then compared to the ones computed by Dynare's `k_order_solver` and by `get_perturbation_params_derivs` for all `analytic_derivation_mode`'s. There will be an error if the maximum absolute deviation is too large, i.e. for numerical derivatives (`analytic_derivation_mode=-1|-2`) the tolerance is choosen lower (around 1e-5); for analytical methods we are stricter: around 1e-13 for first-order, 1e-12 for second order, and 1e-11 for third-order.
* As a side note, this mod file also checks Dynare's `k_order_solver` algorithm and throws an error if something is wrong.
* This test model shows that the new functionality works well. And analytical derivatives perform way better and accurate than numerical ones, even for this small model.
## `analytic_derivatives/burnside_3_order_PertParamsDerivs.mod`
* This builds upon `tests/k_order_perturbation/burnside_k_order.mod` and computes the true parameter derivatives analytically by hand.
* This test model also shows that the new functionality works well.
## `analytic_derivatives/LindeTrabandt2019.mod`
* Shows that the new functionality also works for medium-sized models, i.e. a SW type model solved at third order with 35 variables (11 states). 2 shocks and 20 parameters.
* This mod file can be used to tweak the speed of the computations in the future.
* Compares numerical versus analytical parameter derivatives (for first, second and third order). Note that this model clearly shows that numerical ones are quite different than analytical ones even at first order!
## `identification/LindeTrabandt2019_xfail.mod`
* This model is a check for issue Dynare/dynare#1595, see fjaco.m below, and will fail.
* Removed `analytic_derivatives/ls2003.mod` as this mod file is neither in the testsuite nor does it work.
***
# Detailed changes in other functions
## `get_first_order_solution_params_derivs.m`
* Deleted, or actually, renamed to `get_perturbation_params_derivs.m`, as this function now is able to compute the derivatives up to third order
## `identification_numerical_objective.m`
* `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
## `dsge_likelihood.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `T`matrix anymore, but instead of the dynare solution matrix `ghx`. So we recreate `T` here as this amounts to adding some zeros and focusing on selected variables only.
* Added some checks to make sure the first-order approximation is selected.
* Removed `kron_flag` as input, as `get_perturbation_params_derivs` looks into `options_.analytic_derivation_mode` for `kron_flag`.
## `dynare_identification.m`
* make sure that setting `analytic_derivation_mode` is set both in `options_ident` and `options_`. Note that at the end of the function we restore the `options_` structure, so all changes are local. In a next merge request, I will remove the global variables to make all variables local.
## `get_identification_jacobians.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `A` and `B` matrix anymore, but instead of the dynare solution matrix `ghx` and `ghu`. So we recreate these matrices here instead of in `get_perturbation_params_derivs.m`.
* Added `str2func` for better function handles in `fjaco.m`.
## `fjaco.m`
* make `tol`an option, which can be adjusted by changing `options_.dynatol.x`for identification and parameter derivatives purposes.
* include a check and an informative error message, if numerical derivatives (two-sided finite difference method) yield errors in `resol.m` for identification and parameter derivatives purposes. This closes issue Dynare/dynare#1595.
* Changed year of copyright to 2010-2017,2019
***
# Further suggestions and questions
* Ones this is merged, I will merge request an improvement of the identification toolbox, which will work up to third order using the pruned state space. This will also remove some issues and bugs, and also I will remove global variables in this request.
* The third-order derivatives can be further improved by taking sparsity into account and use mex versions for kronecker products etc. I leave this for further testing (and if anybody actually uses this ;-) )
2019-12-17 19:17:09 +01:00
end
%Add contribution for stderr and corr params to dRHSxss and dRHSuss
if ~ isempty ( indpstderr )
for jp = 1 : stderrparam_nbr
dzss = [ zeros ( M . nspred , 1 ) ;
dghs2 ( kcurr ~= 0 , jp ) ;
dghs2 ( klead ~= 0 , jp ) + oo . dr . ghx ( klead ~= 0 , : ) * dghs2 ( idx_states , jp ) ;
zeros ( M . exo_nbr , 1 ) ] ;
dG_x_ss = kron ( oo . dr . ghx ( idx_states , : ) , dghs2 ( idx_states , jp ) ) ;
dZ_x_ss = kron ( zx , dzss ) ;
dRHSxss ( : , : , jp ) = Fxupup * kron ( speye ( M . nspred ) , vec ( dSigma_e ( : , : , jp ) ) ) ;
dRHSxss ( : , : , jp ) = dRHSxss ( : , : , jp ) + fyp_ghxxp * dG_x_ss ;
dRHSxss ( : , : , jp ) = dRHSxss ( : , : , jp ) + g2 ( : , k2yy0ex0 ( kyy0ex0 , kyy0ex0 ) ) * dZ_x_ss ;
2019-12-20 16:28:06 +01:00
Added parameter derivatives of perturbation solution up to 3 order
# Preliminary comments
I finished the identification toolbox at orders two and three using the pruned state space system, but before I merge request this, I decided to first merge the new functionality to compute parameter derivatives of perturbation solution matrices at higher orders. So after this is approved, I merge the identification toolbox.
I guess @rattoma, @sebastien, and @michel are best choices to review this.
I outline the main idea first and then provide some more detailed changes I made to the functions.
***
# Main idea
This merge request is concerned with the *analytical*computation of the parameter derivatives of first, second and third order perturbation solution matrices, i.e. using _closed-form_ expressions to efficiently compute the derivative of $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ *with respect to model parameters* $\theta$. Note that $\theta$ contains model parameters, stderr and corr parameters of shocks. stderr and corr parameters of measurement errors are not yet supported, (they can easily be included as exogenous shocks). The availability of such derivatives is beneficial in terms of more reliable analysis of model sensitivity and parameter identifiability as well as more efficient estimation methods, in particular for models solved up to third order, as it is well-known that numerical derivatives are a tricky business, especially for large models.
References for my approach are:
* Iskrev (2008, 2010) and Schmitt-Grohé and Uribe (2012, Appendix) who were the first to compute the parameter derivatives analytically at first order, however, using inefficient (sparse) Kronecker products.
* Mutschler (2015) who provides the expressions for a second-order, but again using inefficient (sparse) Kronecker products.
* Ratto and Iskrev (2012) who show how the first-order system can be solved accurately, fast and efficiently using existing numerical algorithms for generalized Sylvester equations by taking the parameter derivative with respect to each parameter separately.
* Julliard and Kamenik (2004) who provide the perturbation solution equation system in tensor notation at any order k.
* Levintal (2017) who introduces permutation matrices to express the perturbation solution equation system in matrix notation up to fifth order.
Note that @rattoma already implemented the parameter derivatives of $g_x$ and $g_u$ analytically (and numerically), and I rely heavily on his work in `get_first_order_solution_params_derivs.m` (previously `getH.m`). My additions are mainly to this function and thus it is renamed to `get_perturbation_params_derivs.m`.
The basic idea of this merge request is to take the second- and third-order perturbation solution systems in Julliard and Kamenik (2004), unfold these into an equivalent matrix representation using permutation matrices as in Levintal (2017). Then extending Ratto and Iskrev (2012) one takes the derivative with respect to each parameter separately and gets a computational problem that is linear, albeit large, as it involves either solving generalized Sylvester equations or taking inverses of highly sparse matrices. I will now briefly summarize the perturbation solution system at third order and the system that results when taking the derivative with respect to parameters.
## Perturbation Solution
The following systems arise at first, second, and third order:
$(ghx): f_{x} z_{x} = f_{y_{-}^*} + f_{y_0} g_{x} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{x}= A g_{x} + f_{y_{-}^*}=0$
$(ghu): f_{z} z_{u} = f_{y_0} g_{u} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{u} + f_{u}= A g_u + f_u = 0$
$(ghxx) : A g_{xx} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{x}\right) + f_{zz} \left( z_{x} \otimes z_{x} \right) = 0$
$(ghxu) : A g_{xu} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{x} \otimes z_{u} \right) = 0$
$(ghuu) : A g_{uu} + B g_{xx} \left(g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{u} \otimes z_{u} \right) = 0$
$(ghs2) : (A+B) g_{\sigma\sigma} + \left( f_{y^{**}_{+}y^{**}_{+}} \left(g^{**}_{u} \otimes g^{**}_{u}\right) + f_{y^{**}_{+}} g^{**}_{uu}\right)vec(\Sigma) = 0$
$(ghxxx) : A g_{xxx} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{x}\right) + f_{y_{+}}g^{**}_{xx} \left(g^{*}_x \otimes g^{*}_{xx}\right)P_{x\_xx} + f_{zz} \left( z_{x} \otimes z_{xx} \right)P_{x\_xx} + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{x} \right) = 0$
$(ghxxu) : A g_{xxu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{u} \right) + f_{zz} \left( \left( z_{x} \otimes z_{xu} \right)P_{x\_xu} + \left(z_{xx} \otimes z_{u}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{x} \otimes g^{*}_{xu}\right)P_{x\_xu} + \left(g^{*}_{xx} \otimes g^{*}_{u}\right) \right) = 0$
$(ghxuu) : A g_{xuu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( \left( z_{xu} \otimes z_{u} \right)P_{xu\_u} + \left(z_{x} \otimes z_{uu}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{xu} \otimes g^{*}_{u}\right)P_{xu\_u} + \left(g^{*}_{x} \otimes g^{*}_{uu}\right) \right) = 0$
$(ghuuu) : A g_{uuu} + B g_{xxx} \left(g^{*}_{u} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{u} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( z_{u} \otimes z_{uu} \right)P_{u\_uu} + f_{y_{+}}g^{**}_{xx} \left(g^{*}_{u} \otimes g^{*}_{uu}\right)P_{u\_uu} = 0$
$(ghx\sigma\sigma) : A g_{x\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_x + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{x} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{x} \otimes z_{\sigma\sigma}\right) + F_{xu_{+}u_{+}}\left(I_{n_x} \otimes vec(\Sigma)\right) = 0$
$F_{xu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_x^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{xu_{+}} \otimes z_{u_{+}} \right)P_{xu\_u} + \left(z_{x} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{x} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
$(ghu\sigma\sigma) : A g_{u\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_{u} + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{u} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{u} \otimes z_{\sigma\sigma}\right) + F_{uu_{+}u_{+}}\left(I_{n_u} \otimes vec(\Sigma_u)\right) = 0$
$F_{uu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_u^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{uu_{+}} \otimes z_{u_{+}} \right)P_{uu\_u} + \left(z_{u} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{u} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
A and B are the common perturbation matrices:
$A = f_{y_0} + \begin{pmatrix} \underbrace{0}_{n\times n_{static}} &\vdots& \underbrace{f_{y^{**}_{+}} \cdot g^{**}_{x}}_{n \times n_{spred}} &\vdots& \underbrace{0}_{n\times n_{frwd}} \end{pmatrix}$and $B = \begin{pmatrix} \underbrace{0}_{n \times n_{static}}&\vdots & \underbrace{0}_{n \times n_{pred}} & \vdots & \underbrace{f_{y^{**}_{+}}}_{n \times n_{sfwrd}} \end{pmatrix}$
and $z=(y_{-}^{\ast}; y; y_{+}^{\ast\ast}; u)$ denotes the dynamic model variables as in `M_.lead_lag_incidence`, $y^\ast$ denote state variables, $y^{\ast\ast}$ denote forward looking variables, $y_+$ denote the variables with a lead, $y_{-}$ denote variables with a lag, $y_0$ denote variables at period t, $f$ the model equations, and $f_z$ the first-order dynamic model derivatives, $f_{zz}$ the second-order dynamic derivatives, and $f_{zzz}$ the third-order dynamic model derivatives. Then:
$z_{x} = \begin{pmatrix}I\\g_{x}\\g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{u} =\begin{pmatrix}0\\g_{u}\\g^{**}_{x} \cdot g^{*}_{u}\\I\end{pmatrix}$, $z_{u_{+}} =\begin{pmatrix}0\\0\\g^{**}_{u}\\0\end{pmatrix}$
$z_{xx} = \begin{pmatrix} 0\\g_{xx}\\g^{**}_{x} \left( g^{*}_x \otimes g^{*}_{x} \right) + g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{xu} =\begin{pmatrix}0\\g_{xu}\\g^{**}_{xx} \left( g^{*}_x \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{xu}\\0\end{pmatrix}$, $z_{uu} =\begin{pmatrix}0\\g_{uu}\\g^{**}_{xx} \left( g^{*}_u \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{uu}\\0\end{pmatrix}$,
$z_{xu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_x \otimes I \right)\\0\end{pmatrix}$, $z_{uu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_{u} \otimes I \right)\\0\end{pmatrix}$, $z_{u_{+}u_{+}} =\begin{pmatrix}0\\0\\g^{\ast\ast}_{uu}\\0\end{pmatrix}$, $z_{\sigma\sigma} = \begin{pmatrix}0\\ g_{\sigma\sigma}\\ g^{\ast\ast}_{x}g^{\ast}_{\sigma\sigma} + g^{\ast\ast}_{\sigma\sigma}\\0 \end{pmatrix}$
$P$ are permutation matrices that can be computed using Matlab's `ipermute` function.
## Parameter derivatives of perturbation solutions
First, we need the parameter derivatives of first, second, third, and fourth derivatives of the dynamic model (i.e. g1,g2,g3,g4 in dynamic files), I make use of the implicit function theorem: Let $f_{z^k}$ denote the kth derivative (wrt all dynamic variables) of the dynamic model, then let $df_{z^k}$ denote the first-derivative (wrt all model parameters) of $f_{z^k}$ evaluated at the steady state. Note that $f_{z^k}$ is a function of both the model parameters $\theta$ and of the steady state of all dynamic variables $\bar{z}$, which also depend on the parameters. Hence, implicitly $f_{z^k}=f_{z^k}(\theta,\bar{z}(\theta))$ and $df_{z^k}$ consists of two parts:
1. direct derivative wrt to all model parameters given by the preprocessor in the `_params_derivs.m` files
2. contribution of derivative of steady state of dynamic variables (wrt all model parameters): $f_{z^{k+1}} \cdot d\bar{z}$
Note that we already have functionality to compute $d\bar{z}$ analytically.
Having this, the above perturbation systems are basically equations of the following types
$AX +BXC = RHS$ or $AX = RHS$
Now when taking the derivative (wrt to one single parameter $\theta_j$), we get
$A\mathrm{d}\{X\} + B\mathrm{d}\{X\}C = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X - \mathrm{d}\{B\}XC - BX\mathrm{d}\{C\}$
or
$A\mathrm{d}\{X\} = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X$
The first one is a Sylvester type equation, the second one can be solved by taking the inverse of $A$. The only diffculty and tedious work arrises in computing (the highly sparse) derivatives of $RHS$.
***
# New functions: `
## get_perturbation_params_derivs.m`and `get_perturbation_params_derivs_numerical_objective.m`
* The parameter derivatives up to third order are computed in the new function`get_perturbation_params_derivs.m` both analytically and numerically. For numerical derivatives `get_perturbation_params_derivs_numerical_objective.m` is the objective for `fjaco.m` or `hessian_sparse.m` or `hessian.m`.
* `get_perturbation_params_derivs.m` is basically an extended version of the previous `get_first_order_solution_params_derivs.m` function.
* * `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It is used for numerical derivatives, whenever `analytic_derivation_mode=-1|-2`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
* Detailed changes:
* Most important: notation of this function is now in accordance to the k_order_solver, i.e. we do not compute derivatives of Kalman transition matrices A and B, but rather the solution matrices ghx,ghu,ghxx,ghxu,ghuu,ghs2,ghxxx,ghxxu,ghxuu,ghuuu,ghxss,ghuss in accordance with notation used in `oo_.dr`. As a byproduct at first-order, focusing on ghx and ghu instead of Kalman transition matrices A and B makes the computations slightly faster for large models (e.g. for Quest the computations were faster by a couple of seconds, not much, but okay).
* Removed use of `kstate`, see also Dynare/dynare#1653 and Dynare/dynare!1656
* Output arguments are stored in a structure `DERIVS`, there is also a flag `d2flag` that computes parameter hessians needed only in `dsge_likelihood.m`.
* Removed `kronflag` as input. `options_.analytic_derivation_mode` is now used instead of `kronflag`.
* Removed `indvar`, an index that was used to selected specific variables in the derivatives. This is not needed, as we always compute the parameter derivatives for all variables first and then select a subset of variables. The selection now takes place in other functions, like `dsge_likelihood.m`.
* Introduced some checks: (i) deterministic exogenous variables are not supported, (ii) Kronecker method only compatible with first-order approximation so reset to sylvester method, (iii) for purely backward or forward models we need to be careful with the rows in `M_.lead_la g_incidence`, (iv) if `_params_derivs.m` files are missing an error is thrown.
* For numerical derivatives, if mod file does not contain an `estimated_params_block`, a temporary one with the most important parameter information is created.
## `unfold_g4.m`
* When evaluating g3 and g4 one needs to take into account that these do not contain symmetric elements, so one needs to use `unfold_g3.m` and the new function `unfold_g4.m`. This returns an unfolded version of the same matrix (i.e. with symmetric elements).
***
# New test models
`.gitignore` and `Makefile.am` are changed accordingly. Also now it is possible to run test suite on analytic_derivatives, i.e. run `make check m/analytic_derivatives`
## `analytic_derivatives/BrockMirman_PertParamsDerivs.mod`
* This is the Brock Mirman model, where we know the exact policy function $g$ for capital and consumption. As this does not imply a nonzero $g_{\sigma\sigma}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ I added some artificial equations to get nonzero solution matrices with respect to $\sigma$. The true perturbation solution matrices $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ are then computed analytically with Matlab's symbolic toolbox and saved in `nBrockMirmanSYM.mat`. There is a preprocessor flag that recreates these analytical computations if changes are needed (and to check whether I made some errors here ;-) )
* Then solution matrices up to third order and their parameter Jacobians are then compared to the ones computed by Dynare's `k_order_solver` and by `get_perturbation_params_derivs` for all `analytic_derivation_mode`'s. There will be an error if the maximum absolute deviation is too large, i.e. for numerical derivatives (`analytic_derivation_mode=-1|-2`) the tolerance is choosen lower (around 1e-5); for analytical methods we are stricter: around 1e-13 for first-order, 1e-12 for second order, and 1e-11 for third-order.
* As a side note, this mod file also checks Dynare's `k_order_solver` algorithm and throws an error if something is wrong.
* This test model shows that the new functionality works well. And analytical derivatives perform way better and accurate than numerical ones, even for this small model.
## `analytic_derivatives/burnside_3_order_PertParamsDerivs.mod`
* This builds upon `tests/k_order_perturbation/burnside_k_order.mod` and computes the true parameter derivatives analytically by hand.
* This test model also shows that the new functionality works well.
## `analytic_derivatives/LindeTrabandt2019.mod`
* Shows that the new functionality also works for medium-sized models, i.e. a SW type model solved at third order with 35 variables (11 states). 2 shocks and 20 parameters.
* This mod file can be used to tweak the speed of the computations in the future.
* Compares numerical versus analytical parameter derivatives (for first, second and third order). Note that this model clearly shows that numerical ones are quite different than analytical ones even at first order!
## `identification/LindeTrabandt2019_xfail.mod`
* This model is a check for issue Dynare/dynare#1595, see fjaco.m below, and will fail.
* Removed `analytic_derivatives/ls2003.mod` as this mod file is neither in the testsuite nor does it work.
***
# Detailed changes in other functions
## `get_first_order_solution_params_derivs.m`
* Deleted, or actually, renamed to `get_perturbation_params_derivs.m`, as this function now is able to compute the derivatives up to third order
## `identification_numerical_objective.m`
* `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
## `dsge_likelihood.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `T`matrix anymore, but instead of the dynare solution matrix `ghx`. So we recreate `T` here as this amounts to adding some zeros and focusing on selected variables only.
* Added some checks to make sure the first-order approximation is selected.
* Removed `kron_flag` as input, as `get_perturbation_params_derivs` looks into `options_.analytic_derivation_mode` for `kron_flag`.
## `dynare_identification.m`
* make sure that setting `analytic_derivation_mode` is set both in `options_ident` and `options_`. Note that at the end of the function we restore the `options_` structure, so all changes are local. In a next merge request, I will remove the global variables to make all variables local.
## `get_identification_jacobians.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `A` and `B` matrix anymore, but instead of the dynare solution matrix `ghx` and `ghu`. So we recreate these matrices here instead of in `get_perturbation_params_derivs.m`.
* Added `str2func` for better function handles in `fjaco.m`.
## `fjaco.m`
* make `tol`an option, which can be adjusted by changing `options_.dynatol.x`for identification and parameter derivatives purposes.
* include a check and an informative error message, if numerical derivatives (two-sided finite difference method) yield errors in `resol.m` for identification and parameter derivatives purposes. This closes issue Dynare/dynare#1595.
* Changed year of copyright to 2010-2017,2019
***
# Further suggestions and questions
* Ones this is merged, I will merge request an improvement of the identification toolbox, which will work up to third order using the pruned state space. This will also remove some issues and bugs, and also I will remove global variables in this request.
* The third-order derivatives can be further improved by taking sparsity into account and use mex versions for kronecker products etc. I leave this for further testing (and if anybody actually uses this ;-) )
2019-12-17 19:17:09 +01:00
dG_u_ss = kron ( oo . dr . ghu ( idx_states , : ) , dghs2 ( idx_states , jp ) ) ;
dZ_u_ss = kron ( zu , dzss ) ;
dRHSuss ( : , : , jp ) = Fuupup * kron ( speye ( M . exo_nbr ) , vec ( dSigma_e ( : , : , jp ) ) ) ;
dRHSuss ( : , : , jp ) = dRHSuss ( : , : , jp ) + fyp_ghxxp * dG_u_ss ;
2019-12-20 16:28:06 +01:00
dRHSuss ( : , : , jp ) = dRHSuss ( : , : , jp ) + g2 ( : , k2yy0ex0 ( kyy0ex0 , kyy0ex0 ) ) * dZ_u_ss ;
Added parameter derivatives of perturbation solution up to 3 order
# Preliminary comments
I finished the identification toolbox at orders two and three using the pruned state space system, but before I merge request this, I decided to first merge the new functionality to compute parameter derivatives of perturbation solution matrices at higher orders. So after this is approved, I merge the identification toolbox.
I guess @rattoma, @sebastien, and @michel are best choices to review this.
I outline the main idea first and then provide some more detailed changes I made to the functions.
***
# Main idea
This merge request is concerned with the *analytical*computation of the parameter derivatives of first, second and third order perturbation solution matrices, i.e. using _closed-form_ expressions to efficiently compute the derivative of $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ *with respect to model parameters* $\theta$. Note that $\theta$ contains model parameters, stderr and corr parameters of shocks. stderr and corr parameters of measurement errors are not yet supported, (they can easily be included as exogenous shocks). The availability of such derivatives is beneficial in terms of more reliable analysis of model sensitivity and parameter identifiability as well as more efficient estimation methods, in particular for models solved up to third order, as it is well-known that numerical derivatives are a tricky business, especially for large models.
References for my approach are:
* Iskrev (2008, 2010) and Schmitt-Grohé and Uribe (2012, Appendix) who were the first to compute the parameter derivatives analytically at first order, however, using inefficient (sparse) Kronecker products.
* Mutschler (2015) who provides the expressions for a second-order, but again using inefficient (sparse) Kronecker products.
* Ratto and Iskrev (2012) who show how the first-order system can be solved accurately, fast and efficiently using existing numerical algorithms for generalized Sylvester equations by taking the parameter derivative with respect to each parameter separately.
* Julliard and Kamenik (2004) who provide the perturbation solution equation system in tensor notation at any order k.
* Levintal (2017) who introduces permutation matrices to express the perturbation solution equation system in matrix notation up to fifth order.
Note that @rattoma already implemented the parameter derivatives of $g_x$ and $g_u$ analytically (and numerically), and I rely heavily on his work in `get_first_order_solution_params_derivs.m` (previously `getH.m`). My additions are mainly to this function and thus it is renamed to `get_perturbation_params_derivs.m`.
The basic idea of this merge request is to take the second- and third-order perturbation solution systems in Julliard and Kamenik (2004), unfold these into an equivalent matrix representation using permutation matrices as in Levintal (2017). Then extending Ratto and Iskrev (2012) one takes the derivative with respect to each parameter separately and gets a computational problem that is linear, albeit large, as it involves either solving generalized Sylvester equations or taking inverses of highly sparse matrices. I will now briefly summarize the perturbation solution system at third order and the system that results when taking the derivative with respect to parameters.
## Perturbation Solution
The following systems arise at first, second, and third order:
$(ghx): f_{x} z_{x} = f_{y_{-}^*} + f_{y_0} g_{x} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{x}= A g_{x} + f_{y_{-}^*}=0$
$(ghu): f_{z} z_{u} = f_{y_0} g_{u} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{u} + f_{u}= A g_u + f_u = 0$
$(ghxx) : A g_{xx} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{x}\right) + f_{zz} \left( z_{x} \otimes z_{x} \right) = 0$
$(ghxu) : A g_{xu} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{x} \otimes z_{u} \right) = 0$
$(ghuu) : A g_{uu} + B g_{xx} \left(g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{u} \otimes z_{u} \right) = 0$
$(ghs2) : (A+B) g_{\sigma\sigma} + \left( f_{y^{**}_{+}y^{**}_{+}} \left(g^{**}_{u} \otimes g^{**}_{u}\right) + f_{y^{**}_{+}} g^{**}_{uu}\right)vec(\Sigma) = 0$
$(ghxxx) : A g_{xxx} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{x}\right) + f_{y_{+}}g^{**}_{xx} \left(g^{*}_x \otimes g^{*}_{xx}\right)P_{x\_xx} + f_{zz} \left( z_{x} \otimes z_{xx} \right)P_{x\_xx} + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{x} \right) = 0$
$(ghxxu) : A g_{xxu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{u} \right) + f_{zz} \left( \left( z_{x} \otimes z_{xu} \right)P_{x\_xu} + \left(z_{xx} \otimes z_{u}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{x} \otimes g^{*}_{xu}\right)P_{x\_xu} + \left(g^{*}_{xx} \otimes g^{*}_{u}\right) \right) = 0$
$(ghxuu) : A g_{xuu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( \left( z_{xu} \otimes z_{u} \right)P_{xu\_u} + \left(z_{x} \otimes z_{uu}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{xu} \otimes g^{*}_{u}\right)P_{xu\_u} + \left(g^{*}_{x} \otimes g^{*}_{uu}\right) \right) = 0$
$(ghuuu) : A g_{uuu} + B g_{xxx} \left(g^{*}_{u} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{u} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( z_{u} \otimes z_{uu} \right)P_{u\_uu} + f_{y_{+}}g^{**}_{xx} \left(g^{*}_{u} \otimes g^{*}_{uu}\right)P_{u\_uu} = 0$
$(ghx\sigma\sigma) : A g_{x\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_x + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{x} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{x} \otimes z_{\sigma\sigma}\right) + F_{xu_{+}u_{+}}\left(I_{n_x} \otimes vec(\Sigma)\right) = 0$
$F_{xu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_x^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{xu_{+}} \otimes z_{u_{+}} \right)P_{xu\_u} + \left(z_{x} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{x} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
$(ghu\sigma\sigma) : A g_{u\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_{u} + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{u} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{u} \otimes z_{\sigma\sigma}\right) + F_{uu_{+}u_{+}}\left(I_{n_u} \otimes vec(\Sigma_u)\right) = 0$
$F_{uu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_u^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{uu_{+}} \otimes z_{u_{+}} \right)P_{uu\_u} + \left(z_{u} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{u} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
A and B are the common perturbation matrices:
$A = f_{y_0} + \begin{pmatrix} \underbrace{0}_{n\times n_{static}} &\vdots& \underbrace{f_{y^{**}_{+}} \cdot g^{**}_{x}}_{n \times n_{spred}} &\vdots& \underbrace{0}_{n\times n_{frwd}} \end{pmatrix}$and $B = \begin{pmatrix} \underbrace{0}_{n \times n_{static}}&\vdots & \underbrace{0}_{n \times n_{pred}} & \vdots & \underbrace{f_{y^{**}_{+}}}_{n \times n_{sfwrd}} \end{pmatrix}$
and $z=(y_{-}^{\ast}; y; y_{+}^{\ast\ast}; u)$ denotes the dynamic model variables as in `M_.lead_lag_incidence`, $y^\ast$ denote state variables, $y^{\ast\ast}$ denote forward looking variables, $y_+$ denote the variables with a lead, $y_{-}$ denote variables with a lag, $y_0$ denote variables at period t, $f$ the model equations, and $f_z$ the first-order dynamic model derivatives, $f_{zz}$ the second-order dynamic derivatives, and $f_{zzz}$ the third-order dynamic model derivatives. Then:
$z_{x} = \begin{pmatrix}I\\g_{x}\\g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{u} =\begin{pmatrix}0\\g_{u}\\g^{**}_{x} \cdot g^{*}_{u}\\I\end{pmatrix}$, $z_{u_{+}} =\begin{pmatrix}0\\0\\g^{**}_{u}\\0\end{pmatrix}$
$z_{xx} = \begin{pmatrix} 0\\g_{xx}\\g^{**}_{x} \left( g^{*}_x \otimes g^{*}_{x} \right) + g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{xu} =\begin{pmatrix}0\\g_{xu}\\g^{**}_{xx} \left( g^{*}_x \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{xu}\\0\end{pmatrix}$, $z_{uu} =\begin{pmatrix}0\\g_{uu}\\g^{**}_{xx} \left( g^{*}_u \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{uu}\\0\end{pmatrix}$,
$z_{xu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_x \otimes I \right)\\0\end{pmatrix}$, $z_{uu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_{u} \otimes I \right)\\0\end{pmatrix}$, $z_{u_{+}u_{+}} =\begin{pmatrix}0\\0\\g^{\ast\ast}_{uu}\\0\end{pmatrix}$, $z_{\sigma\sigma} = \begin{pmatrix}0\\ g_{\sigma\sigma}\\ g^{\ast\ast}_{x}g^{\ast}_{\sigma\sigma} + g^{\ast\ast}_{\sigma\sigma}\\0 \end{pmatrix}$
$P$ are permutation matrices that can be computed using Matlab's `ipermute` function.
## Parameter derivatives of perturbation solutions
First, we need the parameter derivatives of first, second, third, and fourth derivatives of the dynamic model (i.e. g1,g2,g3,g4 in dynamic files), I make use of the implicit function theorem: Let $f_{z^k}$ denote the kth derivative (wrt all dynamic variables) of the dynamic model, then let $df_{z^k}$ denote the first-derivative (wrt all model parameters) of $f_{z^k}$ evaluated at the steady state. Note that $f_{z^k}$ is a function of both the model parameters $\theta$ and of the steady state of all dynamic variables $\bar{z}$, which also depend on the parameters. Hence, implicitly $f_{z^k}=f_{z^k}(\theta,\bar{z}(\theta))$ and $df_{z^k}$ consists of two parts:
1. direct derivative wrt to all model parameters given by the preprocessor in the `_params_derivs.m` files
2. contribution of derivative of steady state of dynamic variables (wrt all model parameters): $f_{z^{k+1}} \cdot d\bar{z}$
Note that we already have functionality to compute $d\bar{z}$ analytically.
Having this, the above perturbation systems are basically equations of the following types
$AX +BXC = RHS$ or $AX = RHS$
Now when taking the derivative (wrt to one single parameter $\theta_j$), we get
$A\mathrm{d}\{X\} + B\mathrm{d}\{X\}C = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X - \mathrm{d}\{B\}XC - BX\mathrm{d}\{C\}$
or
$A\mathrm{d}\{X\} = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X$
The first one is a Sylvester type equation, the second one can be solved by taking the inverse of $A$. The only diffculty and tedious work arrises in computing (the highly sparse) derivatives of $RHS$.
***
# New functions: `
## get_perturbation_params_derivs.m`and `get_perturbation_params_derivs_numerical_objective.m`
* The parameter derivatives up to third order are computed in the new function`get_perturbation_params_derivs.m` both analytically and numerically. For numerical derivatives `get_perturbation_params_derivs_numerical_objective.m` is the objective for `fjaco.m` or `hessian_sparse.m` or `hessian.m`.
* `get_perturbation_params_derivs.m` is basically an extended version of the previous `get_first_order_solution_params_derivs.m` function.
* * `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It is used for numerical derivatives, whenever `analytic_derivation_mode=-1|-2`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
* Detailed changes:
* Most important: notation of this function is now in accordance to the k_order_solver, i.e. we do not compute derivatives of Kalman transition matrices A and B, but rather the solution matrices ghx,ghu,ghxx,ghxu,ghuu,ghs2,ghxxx,ghxxu,ghxuu,ghuuu,ghxss,ghuss in accordance with notation used in `oo_.dr`. As a byproduct at first-order, focusing on ghx and ghu instead of Kalman transition matrices A and B makes the computations slightly faster for large models (e.g. for Quest the computations were faster by a couple of seconds, not much, but okay).
* Removed use of `kstate`, see also Dynare/dynare#1653 and Dynare/dynare!1656
* Output arguments are stored in a structure `DERIVS`, there is also a flag `d2flag` that computes parameter hessians needed only in `dsge_likelihood.m`.
* Removed `kronflag` as input. `options_.analytic_derivation_mode` is now used instead of `kronflag`.
* Removed `indvar`, an index that was used to selected specific variables in the derivatives. This is not needed, as we always compute the parameter derivatives for all variables first and then select a subset of variables. The selection now takes place in other functions, like `dsge_likelihood.m`.
* Introduced some checks: (i) deterministic exogenous variables are not supported, (ii) Kronecker method only compatible with first-order approximation so reset to sylvester method, (iii) for purely backward or forward models we need to be careful with the rows in `M_.lead_la g_incidence`, (iv) if `_params_derivs.m` files are missing an error is thrown.
* For numerical derivatives, if mod file does not contain an `estimated_params_block`, a temporary one with the most important parameter information is created.
## `unfold_g4.m`
* When evaluating g3 and g4 one needs to take into account that these do not contain symmetric elements, so one needs to use `unfold_g3.m` and the new function `unfold_g4.m`. This returns an unfolded version of the same matrix (i.e. with symmetric elements).
***
# New test models
`.gitignore` and `Makefile.am` are changed accordingly. Also now it is possible to run test suite on analytic_derivatives, i.e. run `make check m/analytic_derivatives`
## `analytic_derivatives/BrockMirman_PertParamsDerivs.mod`
* This is the Brock Mirman model, where we know the exact policy function $g$ for capital and consumption. As this does not imply a nonzero $g_{\sigma\sigma}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ I added some artificial equations to get nonzero solution matrices with respect to $\sigma$. The true perturbation solution matrices $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ are then computed analytically with Matlab's symbolic toolbox and saved in `nBrockMirmanSYM.mat`. There is a preprocessor flag that recreates these analytical computations if changes are needed (and to check whether I made some errors here ;-) )
* Then solution matrices up to third order and their parameter Jacobians are then compared to the ones computed by Dynare's `k_order_solver` and by `get_perturbation_params_derivs` for all `analytic_derivation_mode`'s. There will be an error if the maximum absolute deviation is too large, i.e. for numerical derivatives (`analytic_derivation_mode=-1|-2`) the tolerance is choosen lower (around 1e-5); for analytical methods we are stricter: around 1e-13 for first-order, 1e-12 for second order, and 1e-11 for third-order.
* As a side note, this mod file also checks Dynare's `k_order_solver` algorithm and throws an error if something is wrong.
* This test model shows that the new functionality works well. And analytical derivatives perform way better and accurate than numerical ones, even for this small model.
## `analytic_derivatives/burnside_3_order_PertParamsDerivs.mod`
* This builds upon `tests/k_order_perturbation/burnside_k_order.mod` and computes the true parameter derivatives analytically by hand.
* This test model also shows that the new functionality works well.
## `analytic_derivatives/LindeTrabandt2019.mod`
* Shows that the new functionality also works for medium-sized models, i.e. a SW type model solved at third order with 35 variables (11 states). 2 shocks and 20 parameters.
* This mod file can be used to tweak the speed of the computations in the future.
* Compares numerical versus analytical parameter derivatives (for first, second and third order). Note that this model clearly shows that numerical ones are quite different than analytical ones even at first order!
## `identification/LindeTrabandt2019_xfail.mod`
* This model is a check for issue Dynare/dynare#1595, see fjaco.m below, and will fail.
* Removed `analytic_derivatives/ls2003.mod` as this mod file is neither in the testsuite nor does it work.
***
# Detailed changes in other functions
## `get_first_order_solution_params_derivs.m`
* Deleted, or actually, renamed to `get_perturbation_params_derivs.m`, as this function now is able to compute the derivatives up to third order
## `identification_numerical_objective.m`
* `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
## `dsge_likelihood.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `T`matrix anymore, but instead of the dynare solution matrix `ghx`. So we recreate `T` here as this amounts to adding some zeros and focusing on selected variables only.
* Added some checks to make sure the first-order approximation is selected.
* Removed `kron_flag` as input, as `get_perturbation_params_derivs` looks into `options_.analytic_derivation_mode` for `kron_flag`.
## `dynare_identification.m`
* make sure that setting `analytic_derivation_mode` is set both in `options_ident` and `options_`. Note that at the end of the function we restore the `options_` structure, so all changes are local. In a next merge request, I will remove the global variables to make all variables local.
## `get_identification_jacobians.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `A` and `B` matrix anymore, but instead of the dynare solution matrix `ghx` and `ghu`. So we recreate these matrices here instead of in `get_perturbation_params_derivs.m`.
* Added `str2func` for better function handles in `fjaco.m`.
## `fjaco.m`
* make `tol`an option, which can be adjusted by changing `options_.dynatol.x`for identification and parameter derivatives purposes.
* include a check and an informative error message, if numerical derivatives (two-sided finite difference method) yield errors in `resol.m` for identification and parameter derivatives purposes. This closes issue Dynare/dynare#1595.
* Changed year of copyright to 2010-2017,2019
***
# Further suggestions and questions
* Ones this is merged, I will merge request an improvement of the identification toolbox, which will work up to third order using the pruned state space. This will also remove some issues and bugs, and also I will remove global variables in this request.
* The third-order derivatives can be further improved by taking sparsity into account and use mex versions for kronecker products etc. I leave this for further testing (and if anybody actually uses this ;-) )
2019-12-17 19:17:09 +01:00
end
end
if ~ isempty ( indpcorr )
for jp = ( stderrparam_nbr + 1 ) : ( stderrparam_nbr + corrparam_nbr )
dzss = [ zeros ( M . nspred , 1 ) ;
dghs2 ( kcurr ~= 0 , jp ) ;
dghs2 ( klead ~= 0 , jp ) + oo . dr . ghx ( klead ~= 0 , : ) * dghs2 ( idx_states , jp ) ;
zeros ( M . exo_nbr , 1 ) ] ;
dG_x_ss = kron ( oo . dr . ghx ( idx_states , : ) , dghs2 ( idx_states , jp ) ) ;
dZ_x_ss = kron ( zx , dzss ) ;
dRHSxss ( : , : , jp ) = Fxupup * kron ( speye ( M . nspred ) , vec ( dSigma_e ( : , : , jp ) ) ) ;
dRHSxss ( : , : , jp ) = dRHSxss ( : , : , jp ) + fyp_ghxxp * dG_x_ss ;
dRHSxss ( : , : , jp ) = dRHSxss ( : , : , jp ) + g2 ( : , k2yy0ex0 ( kyy0ex0 , kyy0ex0 ) ) * dZ_x_ss ;
2019-12-20 16:28:06 +01:00
Added parameter derivatives of perturbation solution up to 3 order
# Preliminary comments
I finished the identification toolbox at orders two and three using the pruned state space system, but before I merge request this, I decided to first merge the new functionality to compute parameter derivatives of perturbation solution matrices at higher orders. So after this is approved, I merge the identification toolbox.
I guess @rattoma, @sebastien, and @michel are best choices to review this.
I outline the main idea first and then provide some more detailed changes I made to the functions.
***
# Main idea
This merge request is concerned with the *analytical*computation of the parameter derivatives of first, second and third order perturbation solution matrices, i.e. using _closed-form_ expressions to efficiently compute the derivative of $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ *with respect to model parameters* $\theta$. Note that $\theta$ contains model parameters, stderr and corr parameters of shocks. stderr and corr parameters of measurement errors are not yet supported, (they can easily be included as exogenous shocks). The availability of such derivatives is beneficial in terms of more reliable analysis of model sensitivity and parameter identifiability as well as more efficient estimation methods, in particular for models solved up to third order, as it is well-known that numerical derivatives are a tricky business, especially for large models.
References for my approach are:
* Iskrev (2008, 2010) and Schmitt-Grohé and Uribe (2012, Appendix) who were the first to compute the parameter derivatives analytically at first order, however, using inefficient (sparse) Kronecker products.
* Mutschler (2015) who provides the expressions for a second-order, but again using inefficient (sparse) Kronecker products.
* Ratto and Iskrev (2012) who show how the first-order system can be solved accurately, fast and efficiently using existing numerical algorithms for generalized Sylvester equations by taking the parameter derivative with respect to each parameter separately.
* Julliard and Kamenik (2004) who provide the perturbation solution equation system in tensor notation at any order k.
* Levintal (2017) who introduces permutation matrices to express the perturbation solution equation system in matrix notation up to fifth order.
Note that @rattoma already implemented the parameter derivatives of $g_x$ and $g_u$ analytically (and numerically), and I rely heavily on his work in `get_first_order_solution_params_derivs.m` (previously `getH.m`). My additions are mainly to this function and thus it is renamed to `get_perturbation_params_derivs.m`.
The basic idea of this merge request is to take the second- and third-order perturbation solution systems in Julliard and Kamenik (2004), unfold these into an equivalent matrix representation using permutation matrices as in Levintal (2017). Then extending Ratto and Iskrev (2012) one takes the derivative with respect to each parameter separately and gets a computational problem that is linear, albeit large, as it involves either solving generalized Sylvester equations or taking inverses of highly sparse matrices. I will now briefly summarize the perturbation solution system at third order and the system that results when taking the derivative with respect to parameters.
## Perturbation Solution
The following systems arise at first, second, and third order:
$(ghx): f_{x} z_{x} = f_{y_{-}^*} + f_{y_0} g_{x} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{x}= A g_{x} + f_{y_{-}^*}=0$
$(ghu): f_{z} z_{u} = f_{y_0} g_{u} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{u} + f_{u}= A g_u + f_u = 0$
$(ghxx) : A g_{xx} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{x}\right) + f_{zz} \left( z_{x} \otimes z_{x} \right) = 0$
$(ghxu) : A g_{xu} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{x} \otimes z_{u} \right) = 0$
$(ghuu) : A g_{uu} + B g_{xx} \left(g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{u} \otimes z_{u} \right) = 0$
$(ghs2) : (A+B) g_{\sigma\sigma} + \left( f_{y^{**}_{+}y^{**}_{+}} \left(g^{**}_{u} \otimes g^{**}_{u}\right) + f_{y^{**}_{+}} g^{**}_{uu}\right)vec(\Sigma) = 0$
$(ghxxx) : A g_{xxx} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{x}\right) + f_{y_{+}}g^{**}_{xx} \left(g^{*}_x \otimes g^{*}_{xx}\right)P_{x\_xx} + f_{zz} \left( z_{x} \otimes z_{xx} \right)P_{x\_xx} + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{x} \right) = 0$
$(ghxxu) : A g_{xxu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{u} \right) + f_{zz} \left( \left( z_{x} \otimes z_{xu} \right)P_{x\_xu} + \left(z_{xx} \otimes z_{u}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{x} \otimes g^{*}_{xu}\right)P_{x\_xu} + \left(g^{*}_{xx} \otimes g^{*}_{u}\right) \right) = 0$
$(ghxuu) : A g_{xuu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( \left( z_{xu} \otimes z_{u} \right)P_{xu\_u} + \left(z_{x} \otimes z_{uu}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{xu} \otimes g^{*}_{u}\right)P_{xu\_u} + \left(g^{*}_{x} \otimes g^{*}_{uu}\right) \right) = 0$
$(ghuuu) : A g_{uuu} + B g_{xxx} \left(g^{*}_{u} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{u} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( z_{u} \otimes z_{uu} \right)P_{u\_uu} + f_{y_{+}}g^{**}_{xx} \left(g^{*}_{u} \otimes g^{*}_{uu}\right)P_{u\_uu} = 0$
$(ghx\sigma\sigma) : A g_{x\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_x + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{x} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{x} \otimes z_{\sigma\sigma}\right) + F_{xu_{+}u_{+}}\left(I_{n_x} \otimes vec(\Sigma)\right) = 0$
$F_{xu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_x^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{xu_{+}} \otimes z_{u_{+}} \right)P_{xu\_u} + \left(z_{x} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{x} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
$(ghu\sigma\sigma) : A g_{u\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_{u} + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{u} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{u} \otimes z_{\sigma\sigma}\right) + F_{uu_{+}u_{+}}\left(I_{n_u} \otimes vec(\Sigma_u)\right) = 0$
$F_{uu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_u^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{uu_{+}} \otimes z_{u_{+}} \right)P_{uu\_u} + \left(z_{u} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{u} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
A and B are the common perturbation matrices:
$A = f_{y_0} + \begin{pmatrix} \underbrace{0}_{n\times n_{static}} &\vdots& \underbrace{f_{y^{**}_{+}} \cdot g^{**}_{x}}_{n \times n_{spred}} &\vdots& \underbrace{0}_{n\times n_{frwd}} \end{pmatrix}$and $B = \begin{pmatrix} \underbrace{0}_{n \times n_{static}}&\vdots & \underbrace{0}_{n \times n_{pred}} & \vdots & \underbrace{f_{y^{**}_{+}}}_{n \times n_{sfwrd}} \end{pmatrix}$
and $z=(y_{-}^{\ast}; y; y_{+}^{\ast\ast}; u)$ denotes the dynamic model variables as in `M_.lead_lag_incidence`, $y^\ast$ denote state variables, $y^{\ast\ast}$ denote forward looking variables, $y_+$ denote the variables with a lead, $y_{-}$ denote variables with a lag, $y_0$ denote variables at period t, $f$ the model equations, and $f_z$ the first-order dynamic model derivatives, $f_{zz}$ the second-order dynamic derivatives, and $f_{zzz}$ the third-order dynamic model derivatives. Then:
$z_{x} = \begin{pmatrix}I\\g_{x}\\g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{u} =\begin{pmatrix}0\\g_{u}\\g^{**}_{x} \cdot g^{*}_{u}\\I\end{pmatrix}$, $z_{u_{+}} =\begin{pmatrix}0\\0\\g^{**}_{u}\\0\end{pmatrix}$
$z_{xx} = \begin{pmatrix} 0\\g_{xx}\\g^{**}_{x} \left( g^{*}_x \otimes g^{*}_{x} \right) + g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{xu} =\begin{pmatrix}0\\g_{xu}\\g^{**}_{xx} \left( g^{*}_x \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{xu}\\0\end{pmatrix}$, $z_{uu} =\begin{pmatrix}0\\g_{uu}\\g^{**}_{xx} \left( g^{*}_u \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{uu}\\0\end{pmatrix}$,
$z_{xu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_x \otimes I \right)\\0\end{pmatrix}$, $z_{uu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_{u} \otimes I \right)\\0\end{pmatrix}$, $z_{u_{+}u_{+}} =\begin{pmatrix}0\\0\\g^{\ast\ast}_{uu}\\0\end{pmatrix}$, $z_{\sigma\sigma} = \begin{pmatrix}0\\ g_{\sigma\sigma}\\ g^{\ast\ast}_{x}g^{\ast}_{\sigma\sigma} + g^{\ast\ast}_{\sigma\sigma}\\0 \end{pmatrix}$
$P$ are permutation matrices that can be computed using Matlab's `ipermute` function.
## Parameter derivatives of perturbation solutions
First, we need the parameter derivatives of first, second, third, and fourth derivatives of the dynamic model (i.e. g1,g2,g3,g4 in dynamic files), I make use of the implicit function theorem: Let $f_{z^k}$ denote the kth derivative (wrt all dynamic variables) of the dynamic model, then let $df_{z^k}$ denote the first-derivative (wrt all model parameters) of $f_{z^k}$ evaluated at the steady state. Note that $f_{z^k}$ is a function of both the model parameters $\theta$ and of the steady state of all dynamic variables $\bar{z}$, which also depend on the parameters. Hence, implicitly $f_{z^k}=f_{z^k}(\theta,\bar{z}(\theta))$ and $df_{z^k}$ consists of two parts:
1. direct derivative wrt to all model parameters given by the preprocessor in the `_params_derivs.m` files
2. contribution of derivative of steady state of dynamic variables (wrt all model parameters): $f_{z^{k+1}} \cdot d\bar{z}$
Note that we already have functionality to compute $d\bar{z}$ analytically.
Having this, the above perturbation systems are basically equations of the following types
$AX +BXC = RHS$ or $AX = RHS$
Now when taking the derivative (wrt to one single parameter $\theta_j$), we get
$A\mathrm{d}\{X\} + B\mathrm{d}\{X\}C = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X - \mathrm{d}\{B\}XC - BX\mathrm{d}\{C\}$
or
$A\mathrm{d}\{X\} = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X$
The first one is a Sylvester type equation, the second one can be solved by taking the inverse of $A$. The only diffculty and tedious work arrises in computing (the highly sparse) derivatives of $RHS$.
***
# New functions: `
## get_perturbation_params_derivs.m`and `get_perturbation_params_derivs_numerical_objective.m`
* The parameter derivatives up to third order are computed in the new function`get_perturbation_params_derivs.m` both analytically and numerically. For numerical derivatives `get_perturbation_params_derivs_numerical_objective.m` is the objective for `fjaco.m` or `hessian_sparse.m` or `hessian.m`.
* `get_perturbation_params_derivs.m` is basically an extended version of the previous `get_first_order_solution_params_derivs.m` function.
* * `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It is used for numerical derivatives, whenever `analytic_derivation_mode=-1|-2`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
* Detailed changes:
* Most important: notation of this function is now in accordance to the k_order_solver, i.e. we do not compute derivatives of Kalman transition matrices A and B, but rather the solution matrices ghx,ghu,ghxx,ghxu,ghuu,ghs2,ghxxx,ghxxu,ghxuu,ghuuu,ghxss,ghuss in accordance with notation used in `oo_.dr`. As a byproduct at first-order, focusing on ghx and ghu instead of Kalman transition matrices A and B makes the computations slightly faster for large models (e.g. for Quest the computations were faster by a couple of seconds, not much, but okay).
* Removed use of `kstate`, see also Dynare/dynare#1653 and Dynare/dynare!1656
* Output arguments are stored in a structure `DERIVS`, there is also a flag `d2flag` that computes parameter hessians needed only in `dsge_likelihood.m`.
* Removed `kronflag` as input. `options_.analytic_derivation_mode` is now used instead of `kronflag`.
* Removed `indvar`, an index that was used to selected specific variables in the derivatives. This is not needed, as we always compute the parameter derivatives for all variables first and then select a subset of variables. The selection now takes place in other functions, like `dsge_likelihood.m`.
* Introduced some checks: (i) deterministic exogenous variables are not supported, (ii) Kronecker method only compatible with first-order approximation so reset to sylvester method, (iii) for purely backward or forward models we need to be careful with the rows in `M_.lead_la g_incidence`, (iv) if `_params_derivs.m` files are missing an error is thrown.
* For numerical derivatives, if mod file does not contain an `estimated_params_block`, a temporary one with the most important parameter information is created.
## `unfold_g4.m`
* When evaluating g3 and g4 one needs to take into account that these do not contain symmetric elements, so one needs to use `unfold_g3.m` and the new function `unfold_g4.m`. This returns an unfolded version of the same matrix (i.e. with symmetric elements).
***
# New test models
`.gitignore` and `Makefile.am` are changed accordingly. Also now it is possible to run test suite on analytic_derivatives, i.e. run `make check m/analytic_derivatives`
## `analytic_derivatives/BrockMirman_PertParamsDerivs.mod`
* This is the Brock Mirman model, where we know the exact policy function $g$ for capital and consumption. As this does not imply a nonzero $g_{\sigma\sigma}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ I added some artificial equations to get nonzero solution matrices with respect to $\sigma$. The true perturbation solution matrices $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ are then computed analytically with Matlab's symbolic toolbox and saved in `nBrockMirmanSYM.mat`. There is a preprocessor flag that recreates these analytical computations if changes are needed (and to check whether I made some errors here ;-) )
* Then solution matrices up to third order and their parameter Jacobians are then compared to the ones computed by Dynare's `k_order_solver` and by `get_perturbation_params_derivs` for all `analytic_derivation_mode`'s. There will be an error if the maximum absolute deviation is too large, i.e. for numerical derivatives (`analytic_derivation_mode=-1|-2`) the tolerance is choosen lower (around 1e-5); for analytical methods we are stricter: around 1e-13 for first-order, 1e-12 for second order, and 1e-11 for third-order.
* As a side note, this mod file also checks Dynare's `k_order_solver` algorithm and throws an error if something is wrong.
* This test model shows that the new functionality works well. And analytical derivatives perform way better and accurate than numerical ones, even for this small model.
## `analytic_derivatives/burnside_3_order_PertParamsDerivs.mod`
* This builds upon `tests/k_order_perturbation/burnside_k_order.mod` and computes the true parameter derivatives analytically by hand.
* This test model also shows that the new functionality works well.
## `analytic_derivatives/LindeTrabandt2019.mod`
* Shows that the new functionality also works for medium-sized models, i.e. a SW type model solved at third order with 35 variables (11 states). 2 shocks and 20 parameters.
* This mod file can be used to tweak the speed of the computations in the future.
* Compares numerical versus analytical parameter derivatives (for first, second and third order). Note that this model clearly shows that numerical ones are quite different than analytical ones even at first order!
## `identification/LindeTrabandt2019_xfail.mod`
* This model is a check for issue Dynare/dynare#1595, see fjaco.m below, and will fail.
* Removed `analytic_derivatives/ls2003.mod` as this mod file is neither in the testsuite nor does it work.
***
# Detailed changes in other functions
## `get_first_order_solution_params_derivs.m`
* Deleted, or actually, renamed to `get_perturbation_params_derivs.m`, as this function now is able to compute the derivatives up to third order
## `identification_numerical_objective.m`
* `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
## `dsge_likelihood.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `T`matrix anymore, but instead of the dynare solution matrix `ghx`. So we recreate `T` here as this amounts to adding some zeros and focusing on selected variables only.
* Added some checks to make sure the first-order approximation is selected.
* Removed `kron_flag` as input, as `get_perturbation_params_derivs` looks into `options_.analytic_derivation_mode` for `kron_flag`.
## `dynare_identification.m`
* make sure that setting `analytic_derivation_mode` is set both in `options_ident` and `options_`. Note that at the end of the function we restore the `options_` structure, so all changes are local. In a next merge request, I will remove the global variables to make all variables local.
## `get_identification_jacobians.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `A` and `B` matrix anymore, but instead of the dynare solution matrix `ghx` and `ghu`. So we recreate these matrices here instead of in `get_perturbation_params_derivs.m`.
* Added `str2func` for better function handles in `fjaco.m`.
## `fjaco.m`
* make `tol`an option, which can be adjusted by changing `options_.dynatol.x`for identification and parameter derivatives purposes.
* include a check and an informative error message, if numerical derivatives (two-sided finite difference method) yield errors in `resol.m` for identification and parameter derivatives purposes. This closes issue Dynare/dynare#1595.
* Changed year of copyright to 2010-2017,2019
***
# Further suggestions and questions
* Ones this is merged, I will merge request an improvement of the identification toolbox, which will work up to third order using the pruned state space. This will also remove some issues and bugs, and also I will remove global variables in this request.
* The third-order derivatives can be further improved by taking sparsity into account and use mex versions for kronecker products etc. I leave this for further testing (and if anybody actually uses this ;-) )
2019-12-17 19:17:09 +01:00
dG_u_ss = kron ( oo . dr . ghu ( idx_states , : ) , dghs2 ( idx_states , jp ) ) ;
dZ_u_ss = kron ( zu , dzss ) ;
dRHSuss ( : , : , jp ) = Fuupup * kron ( speye ( M . exo_nbr ) , vec ( dSigma_e ( : , : , jp ) ) ) ;
dRHSuss ( : , : , jp ) = dRHSuss ( : , : , jp ) + fyp_ghxxp * dG_u_ss ;
2019-12-20 16:28:06 +01:00
dRHSuss ( : , : , jp ) = dRHSuss ( : , : , jp ) + g2 ( : , k2yy0ex0 ( kyy0ex0 , kyy0ex0 ) ) * dZ_u_ss ;
Added parameter derivatives of perturbation solution up to 3 order
# Preliminary comments
I finished the identification toolbox at orders two and three using the pruned state space system, but before I merge request this, I decided to first merge the new functionality to compute parameter derivatives of perturbation solution matrices at higher orders. So after this is approved, I merge the identification toolbox.
I guess @rattoma, @sebastien, and @michel are best choices to review this.
I outline the main idea first and then provide some more detailed changes I made to the functions.
***
# Main idea
This merge request is concerned with the *analytical*computation of the parameter derivatives of first, second and third order perturbation solution matrices, i.e. using _closed-form_ expressions to efficiently compute the derivative of $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ *with respect to model parameters* $\theta$. Note that $\theta$ contains model parameters, stderr and corr parameters of shocks. stderr and corr parameters of measurement errors are not yet supported, (they can easily be included as exogenous shocks). The availability of such derivatives is beneficial in terms of more reliable analysis of model sensitivity and parameter identifiability as well as more efficient estimation methods, in particular for models solved up to third order, as it is well-known that numerical derivatives are a tricky business, especially for large models.
References for my approach are:
* Iskrev (2008, 2010) and Schmitt-Grohé and Uribe (2012, Appendix) who were the first to compute the parameter derivatives analytically at first order, however, using inefficient (sparse) Kronecker products.
* Mutschler (2015) who provides the expressions for a second-order, but again using inefficient (sparse) Kronecker products.
* Ratto and Iskrev (2012) who show how the first-order system can be solved accurately, fast and efficiently using existing numerical algorithms for generalized Sylvester equations by taking the parameter derivative with respect to each parameter separately.
* Julliard and Kamenik (2004) who provide the perturbation solution equation system in tensor notation at any order k.
* Levintal (2017) who introduces permutation matrices to express the perturbation solution equation system in matrix notation up to fifth order.
Note that @rattoma already implemented the parameter derivatives of $g_x$ and $g_u$ analytically (and numerically), and I rely heavily on his work in `get_first_order_solution_params_derivs.m` (previously `getH.m`). My additions are mainly to this function and thus it is renamed to `get_perturbation_params_derivs.m`.
The basic idea of this merge request is to take the second- and third-order perturbation solution systems in Julliard and Kamenik (2004), unfold these into an equivalent matrix representation using permutation matrices as in Levintal (2017). Then extending Ratto and Iskrev (2012) one takes the derivative with respect to each parameter separately and gets a computational problem that is linear, albeit large, as it involves either solving generalized Sylvester equations or taking inverses of highly sparse matrices. I will now briefly summarize the perturbation solution system at third order and the system that results when taking the derivative with respect to parameters.
## Perturbation Solution
The following systems arise at first, second, and third order:
$(ghx): f_{x} z_{x} = f_{y_{-}^*} + f_{y_0} g_{x} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{x}= A g_{x} + f_{y_{-}^*}=0$
$(ghu): f_{z} z_{u} = f_{y_0} g_{u} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{u} + f_{u}= A g_u + f_u = 0$
$(ghxx) : A g_{xx} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{x}\right) + f_{zz} \left( z_{x} \otimes z_{x} \right) = 0$
$(ghxu) : A g_{xu} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{x} \otimes z_{u} \right) = 0$
$(ghuu) : A g_{uu} + B g_{xx} \left(g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{u} \otimes z_{u} \right) = 0$
$(ghs2) : (A+B) g_{\sigma\sigma} + \left( f_{y^{**}_{+}y^{**}_{+}} \left(g^{**}_{u} \otimes g^{**}_{u}\right) + f_{y^{**}_{+}} g^{**}_{uu}\right)vec(\Sigma) = 0$
$(ghxxx) : A g_{xxx} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{x}\right) + f_{y_{+}}g^{**}_{xx} \left(g^{*}_x \otimes g^{*}_{xx}\right)P_{x\_xx} + f_{zz} \left( z_{x} \otimes z_{xx} \right)P_{x\_xx} + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{x} \right) = 0$
$(ghxxu) : A g_{xxu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{u} \right) + f_{zz} \left( \left( z_{x} \otimes z_{xu} \right)P_{x\_xu} + \left(z_{xx} \otimes z_{u}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{x} \otimes g^{*}_{xu}\right)P_{x\_xu} + \left(g^{*}_{xx} \otimes g^{*}_{u}\right) \right) = 0$
$(ghxuu) : A g_{xuu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( \left( z_{xu} \otimes z_{u} \right)P_{xu\_u} + \left(z_{x} \otimes z_{uu}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{xu} \otimes g^{*}_{u}\right)P_{xu\_u} + \left(g^{*}_{x} \otimes g^{*}_{uu}\right) \right) = 0$
$(ghuuu) : A g_{uuu} + B g_{xxx} \left(g^{*}_{u} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{u} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( z_{u} \otimes z_{uu} \right)P_{u\_uu} + f_{y_{+}}g^{**}_{xx} \left(g^{*}_{u} \otimes g^{*}_{uu}\right)P_{u\_uu} = 0$
$(ghx\sigma\sigma) : A g_{x\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_x + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{x} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{x} \otimes z_{\sigma\sigma}\right) + F_{xu_{+}u_{+}}\left(I_{n_x} \otimes vec(\Sigma)\right) = 0$
$F_{xu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_x^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{xu_{+}} \otimes z_{u_{+}} \right)P_{xu\_u} + \left(z_{x} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{x} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
$(ghu\sigma\sigma) : A g_{u\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_{u} + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{u} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{u} \otimes z_{\sigma\sigma}\right) + F_{uu_{+}u_{+}}\left(I_{n_u} \otimes vec(\Sigma_u)\right) = 0$
$F_{uu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_u^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{uu_{+}} \otimes z_{u_{+}} \right)P_{uu\_u} + \left(z_{u} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{u} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
A and B are the common perturbation matrices:
$A = f_{y_0} + \begin{pmatrix} \underbrace{0}_{n\times n_{static}} &\vdots& \underbrace{f_{y^{**}_{+}} \cdot g^{**}_{x}}_{n \times n_{spred}} &\vdots& \underbrace{0}_{n\times n_{frwd}} \end{pmatrix}$and $B = \begin{pmatrix} \underbrace{0}_{n \times n_{static}}&\vdots & \underbrace{0}_{n \times n_{pred}} & \vdots & \underbrace{f_{y^{**}_{+}}}_{n \times n_{sfwrd}} \end{pmatrix}$
and $z=(y_{-}^{\ast}; y; y_{+}^{\ast\ast}; u)$ denotes the dynamic model variables as in `M_.lead_lag_incidence`, $y^\ast$ denote state variables, $y^{\ast\ast}$ denote forward looking variables, $y_+$ denote the variables with a lead, $y_{-}$ denote variables with a lag, $y_0$ denote variables at period t, $f$ the model equations, and $f_z$ the first-order dynamic model derivatives, $f_{zz}$ the second-order dynamic derivatives, and $f_{zzz}$ the third-order dynamic model derivatives. Then:
$z_{x} = \begin{pmatrix}I\\g_{x}\\g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{u} =\begin{pmatrix}0\\g_{u}\\g^{**}_{x} \cdot g^{*}_{u}\\I\end{pmatrix}$, $z_{u_{+}} =\begin{pmatrix}0\\0\\g^{**}_{u}\\0\end{pmatrix}$
$z_{xx} = \begin{pmatrix} 0\\g_{xx}\\g^{**}_{x} \left( g^{*}_x \otimes g^{*}_{x} \right) + g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{xu} =\begin{pmatrix}0\\g_{xu}\\g^{**}_{xx} \left( g^{*}_x \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{xu}\\0\end{pmatrix}$, $z_{uu} =\begin{pmatrix}0\\g_{uu}\\g^{**}_{xx} \left( g^{*}_u \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{uu}\\0\end{pmatrix}$,
$z_{xu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_x \otimes I \right)\\0\end{pmatrix}$, $z_{uu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_{u} \otimes I \right)\\0\end{pmatrix}$, $z_{u_{+}u_{+}} =\begin{pmatrix}0\\0\\g^{\ast\ast}_{uu}\\0\end{pmatrix}$, $z_{\sigma\sigma} = \begin{pmatrix}0\\ g_{\sigma\sigma}\\ g^{\ast\ast}_{x}g^{\ast}_{\sigma\sigma} + g^{\ast\ast}_{\sigma\sigma}\\0 \end{pmatrix}$
$P$ are permutation matrices that can be computed using Matlab's `ipermute` function.
## Parameter derivatives of perturbation solutions
First, we need the parameter derivatives of first, second, third, and fourth derivatives of the dynamic model (i.e. g1,g2,g3,g4 in dynamic files), I make use of the implicit function theorem: Let $f_{z^k}$ denote the kth derivative (wrt all dynamic variables) of the dynamic model, then let $df_{z^k}$ denote the first-derivative (wrt all model parameters) of $f_{z^k}$ evaluated at the steady state. Note that $f_{z^k}$ is a function of both the model parameters $\theta$ and of the steady state of all dynamic variables $\bar{z}$, which also depend on the parameters. Hence, implicitly $f_{z^k}=f_{z^k}(\theta,\bar{z}(\theta))$ and $df_{z^k}$ consists of two parts:
1. direct derivative wrt to all model parameters given by the preprocessor in the `_params_derivs.m` files
2. contribution of derivative of steady state of dynamic variables (wrt all model parameters): $f_{z^{k+1}} \cdot d\bar{z}$
Note that we already have functionality to compute $d\bar{z}$ analytically.
Having this, the above perturbation systems are basically equations of the following types
$AX +BXC = RHS$ or $AX = RHS$
Now when taking the derivative (wrt to one single parameter $\theta_j$), we get
$A\mathrm{d}\{X\} + B\mathrm{d}\{X\}C = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X - \mathrm{d}\{B\}XC - BX\mathrm{d}\{C\}$
or
$A\mathrm{d}\{X\} = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X$
The first one is a Sylvester type equation, the second one can be solved by taking the inverse of $A$. The only diffculty and tedious work arrises in computing (the highly sparse) derivatives of $RHS$.
***
# New functions: `
## get_perturbation_params_derivs.m`and `get_perturbation_params_derivs_numerical_objective.m`
* The parameter derivatives up to third order are computed in the new function`get_perturbation_params_derivs.m` both analytically and numerically. For numerical derivatives `get_perturbation_params_derivs_numerical_objective.m` is the objective for `fjaco.m` or `hessian_sparse.m` or `hessian.m`.
* `get_perturbation_params_derivs.m` is basically an extended version of the previous `get_first_order_solution_params_derivs.m` function.
* * `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It is used for numerical derivatives, whenever `analytic_derivation_mode=-1|-2`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
* Detailed changes:
* Most important: notation of this function is now in accordance to the k_order_solver, i.e. we do not compute derivatives of Kalman transition matrices A and B, but rather the solution matrices ghx,ghu,ghxx,ghxu,ghuu,ghs2,ghxxx,ghxxu,ghxuu,ghuuu,ghxss,ghuss in accordance with notation used in `oo_.dr`. As a byproduct at first-order, focusing on ghx and ghu instead of Kalman transition matrices A and B makes the computations slightly faster for large models (e.g. for Quest the computations were faster by a couple of seconds, not much, but okay).
* Removed use of `kstate`, see also Dynare/dynare#1653 and Dynare/dynare!1656
* Output arguments are stored in a structure `DERIVS`, there is also a flag `d2flag` that computes parameter hessians needed only in `dsge_likelihood.m`.
* Removed `kronflag` as input. `options_.analytic_derivation_mode` is now used instead of `kronflag`.
* Removed `indvar`, an index that was used to selected specific variables in the derivatives. This is not needed, as we always compute the parameter derivatives for all variables first and then select a subset of variables. The selection now takes place in other functions, like `dsge_likelihood.m`.
* Introduced some checks: (i) deterministic exogenous variables are not supported, (ii) Kronecker method only compatible with first-order approximation so reset to sylvester method, (iii) for purely backward or forward models we need to be careful with the rows in `M_.lead_la g_incidence`, (iv) if `_params_derivs.m` files are missing an error is thrown.
* For numerical derivatives, if mod file does not contain an `estimated_params_block`, a temporary one with the most important parameter information is created.
## `unfold_g4.m`
* When evaluating g3 and g4 one needs to take into account that these do not contain symmetric elements, so one needs to use `unfold_g3.m` and the new function `unfold_g4.m`. This returns an unfolded version of the same matrix (i.e. with symmetric elements).
***
# New test models
`.gitignore` and `Makefile.am` are changed accordingly. Also now it is possible to run test suite on analytic_derivatives, i.e. run `make check m/analytic_derivatives`
## `analytic_derivatives/BrockMirman_PertParamsDerivs.mod`
* This is the Brock Mirman model, where we know the exact policy function $g$ for capital and consumption. As this does not imply a nonzero $g_{\sigma\sigma}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ I added some artificial equations to get nonzero solution matrices with respect to $\sigma$. The true perturbation solution matrices $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ are then computed analytically with Matlab's symbolic toolbox and saved in `nBrockMirmanSYM.mat`. There is a preprocessor flag that recreates these analytical computations if changes are needed (and to check whether I made some errors here ;-) )
* Then solution matrices up to third order and their parameter Jacobians are then compared to the ones computed by Dynare's `k_order_solver` and by `get_perturbation_params_derivs` for all `analytic_derivation_mode`'s. There will be an error if the maximum absolute deviation is too large, i.e. for numerical derivatives (`analytic_derivation_mode=-1|-2`) the tolerance is choosen lower (around 1e-5); for analytical methods we are stricter: around 1e-13 for first-order, 1e-12 for second order, and 1e-11 for third-order.
* As a side note, this mod file also checks Dynare's `k_order_solver` algorithm and throws an error if something is wrong.
* This test model shows that the new functionality works well. And analytical derivatives perform way better and accurate than numerical ones, even for this small model.
## `analytic_derivatives/burnside_3_order_PertParamsDerivs.mod`
* This builds upon `tests/k_order_perturbation/burnside_k_order.mod` and computes the true parameter derivatives analytically by hand.
* This test model also shows that the new functionality works well.
## `analytic_derivatives/LindeTrabandt2019.mod`
* Shows that the new functionality also works for medium-sized models, i.e. a SW type model solved at third order with 35 variables (11 states). 2 shocks and 20 parameters.
* This mod file can be used to tweak the speed of the computations in the future.
* Compares numerical versus analytical parameter derivatives (for first, second and third order). Note that this model clearly shows that numerical ones are quite different than analytical ones even at first order!
## `identification/LindeTrabandt2019_xfail.mod`
* This model is a check for issue Dynare/dynare#1595, see fjaco.m below, and will fail.
* Removed `analytic_derivatives/ls2003.mod` as this mod file is neither in the testsuite nor does it work.
***
# Detailed changes in other functions
## `get_first_order_solution_params_derivs.m`
* Deleted, or actually, renamed to `get_perturbation_params_derivs.m`, as this function now is able to compute the derivatives up to third order
## `identification_numerical_objective.m`
* `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
## `dsge_likelihood.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `T`matrix anymore, but instead of the dynare solution matrix `ghx`. So we recreate `T` here as this amounts to adding some zeros and focusing on selected variables only.
* Added some checks to make sure the first-order approximation is selected.
* Removed `kron_flag` as input, as `get_perturbation_params_derivs` looks into `options_.analytic_derivation_mode` for `kron_flag`.
## `dynare_identification.m`
* make sure that setting `analytic_derivation_mode` is set both in `options_ident` and `options_`. Note that at the end of the function we restore the `options_` structure, so all changes are local. In a next merge request, I will remove the global variables to make all variables local.
## `get_identification_jacobians.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `A` and `B` matrix anymore, but instead of the dynare solution matrix `ghx` and `ghu`. So we recreate these matrices here instead of in `get_perturbation_params_derivs.m`.
* Added `str2func` for better function handles in `fjaco.m`.
## `fjaco.m`
* make `tol`an option, which can be adjusted by changing `options_.dynatol.x`for identification and parameter derivatives purposes.
* include a check and an informative error message, if numerical derivatives (two-sided finite difference method) yield errors in `resol.m` for identification and parameter derivatives purposes. This closes issue Dynare/dynare#1595.
* Changed year of copyright to 2010-2017,2019
***
# Further suggestions and questions
* Ones this is merged, I will merge request an improvement of the identification toolbox, which will work up to third order using the pruned state space. This will also remove some issues and bugs, and also I will remove global variables in this request.
* The third-order derivatives can be further improved by taking sparsity into account and use mex versions for kronecker products etc. I leave this for further testing (and if anybody actually uses this ;-) )
2019-12-17 19:17:09 +01:00
end
end
dRHSxss = - dRHSxss ;
%use iterated generalized sylvester equation to compute dghxss
dghxss = sylvester3 ( A , B , oo . dr . ghx ( idx_states , : ) , dRHSxss ) ;
flag = 1 ; icount = 0 ;
while flag && icount < 4
[ dghxss , flag ] = sylvester3a ( dghxss , A , B , oo . dr . ghx ( idx_states , : ) , dRHSxss ) ;
icount = icount + 1 ;
end
%Add contribution by dghxss to dRHSuss and compute it
dghuss = zeros ( M . endo_nbr , M . exo_nbr , totparam_nbr ) ;
for jp = 1 : totparam_nbr
dRHS = dRHSuss ( : , : , jp ) + B * dghxss ( : , : , jp ) * oo . dr . ghu ( idx_states , : ) ;
dghuss ( : , : , jp ) = invA * ( - dRHS ) ;
end
end
%% Store into structure
DERIVS . dg1 = dg1 ;
DERIVS . dSigma_e = dSigma_e ;
2019-12-20 00:40:00 +01:00
DERIVS . dCorrelation_matrix = dCorrelation_matrix ;
Added parameter derivatives of perturbation solution up to 3 order
# Preliminary comments
I finished the identification toolbox at orders two and three using the pruned state space system, but before I merge request this, I decided to first merge the new functionality to compute parameter derivatives of perturbation solution matrices at higher orders. So after this is approved, I merge the identification toolbox.
I guess @rattoma, @sebastien, and @michel are best choices to review this.
I outline the main idea first and then provide some more detailed changes I made to the functions.
***
# Main idea
This merge request is concerned with the *analytical*computation of the parameter derivatives of first, second and third order perturbation solution matrices, i.e. using _closed-form_ expressions to efficiently compute the derivative of $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ *with respect to model parameters* $\theta$. Note that $\theta$ contains model parameters, stderr and corr parameters of shocks. stderr and corr parameters of measurement errors are not yet supported, (they can easily be included as exogenous shocks). The availability of such derivatives is beneficial in terms of more reliable analysis of model sensitivity and parameter identifiability as well as more efficient estimation methods, in particular for models solved up to third order, as it is well-known that numerical derivatives are a tricky business, especially for large models.
References for my approach are:
* Iskrev (2008, 2010) and Schmitt-Grohé and Uribe (2012, Appendix) who were the first to compute the parameter derivatives analytically at first order, however, using inefficient (sparse) Kronecker products.
* Mutschler (2015) who provides the expressions for a second-order, but again using inefficient (sparse) Kronecker products.
* Ratto and Iskrev (2012) who show how the first-order system can be solved accurately, fast and efficiently using existing numerical algorithms for generalized Sylvester equations by taking the parameter derivative with respect to each parameter separately.
* Julliard and Kamenik (2004) who provide the perturbation solution equation system in tensor notation at any order k.
* Levintal (2017) who introduces permutation matrices to express the perturbation solution equation system in matrix notation up to fifth order.
Note that @rattoma already implemented the parameter derivatives of $g_x$ and $g_u$ analytically (and numerically), and I rely heavily on his work in `get_first_order_solution_params_derivs.m` (previously `getH.m`). My additions are mainly to this function and thus it is renamed to `get_perturbation_params_derivs.m`.
The basic idea of this merge request is to take the second- and third-order perturbation solution systems in Julliard and Kamenik (2004), unfold these into an equivalent matrix representation using permutation matrices as in Levintal (2017). Then extending Ratto and Iskrev (2012) one takes the derivative with respect to each parameter separately and gets a computational problem that is linear, albeit large, as it involves either solving generalized Sylvester equations or taking inverses of highly sparse matrices. I will now briefly summarize the perturbation solution system at third order and the system that results when taking the derivative with respect to parameters.
## Perturbation Solution
The following systems arise at first, second, and third order:
$(ghx): f_{x} z_{x} = f_{y_{-}^*} + f_{y_0} g_{x} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{x}= A g_{x} + f_{y_{-}^*}=0$
$(ghu): f_{z} z_{u} = f_{y_0} g_{u} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{u} + f_{u}= A g_u + f_u = 0$
$(ghxx) : A g_{xx} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{x}\right) + f_{zz} \left( z_{x} \otimes z_{x} \right) = 0$
$(ghxu) : A g_{xu} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{x} \otimes z_{u} \right) = 0$
$(ghuu) : A g_{uu} + B g_{xx} \left(g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{u} \otimes z_{u} \right) = 0$
$(ghs2) : (A+B) g_{\sigma\sigma} + \left( f_{y^{**}_{+}y^{**}_{+}} \left(g^{**}_{u} \otimes g^{**}_{u}\right) + f_{y^{**}_{+}} g^{**}_{uu}\right)vec(\Sigma) = 0$
$(ghxxx) : A g_{xxx} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{x}\right) + f_{y_{+}}g^{**}_{xx} \left(g^{*}_x \otimes g^{*}_{xx}\right)P_{x\_xx} + f_{zz} \left( z_{x} \otimes z_{xx} \right)P_{x\_xx} + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{x} \right) = 0$
$(ghxxu) : A g_{xxu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{u} \right) + f_{zz} \left( \left( z_{x} \otimes z_{xu} \right)P_{x\_xu} + \left(z_{xx} \otimes z_{u}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{x} \otimes g^{*}_{xu}\right)P_{x\_xu} + \left(g^{*}_{xx} \otimes g^{*}_{u}\right) \right) = 0$
$(ghxuu) : A g_{xuu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( \left( z_{xu} \otimes z_{u} \right)P_{xu\_u} + \left(z_{x} \otimes z_{uu}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{xu} \otimes g^{*}_{u}\right)P_{xu\_u} + \left(g^{*}_{x} \otimes g^{*}_{uu}\right) \right) = 0$
$(ghuuu) : A g_{uuu} + B g_{xxx} \left(g^{*}_{u} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{u} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( z_{u} \otimes z_{uu} \right)P_{u\_uu} + f_{y_{+}}g^{**}_{xx} \left(g^{*}_{u} \otimes g^{*}_{uu}\right)P_{u\_uu} = 0$
$(ghx\sigma\sigma) : A g_{x\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_x + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{x} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{x} \otimes z_{\sigma\sigma}\right) + F_{xu_{+}u_{+}}\left(I_{n_x} \otimes vec(\Sigma)\right) = 0$
$F_{xu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_x^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{xu_{+}} \otimes z_{u_{+}} \right)P_{xu\_u} + \left(z_{x} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{x} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
$(ghu\sigma\sigma) : A g_{u\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_{u} + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{u} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{u} \otimes z_{\sigma\sigma}\right) + F_{uu_{+}u_{+}}\left(I_{n_u} \otimes vec(\Sigma_u)\right) = 0$
$F_{uu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_u^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{uu_{+}} \otimes z_{u_{+}} \right)P_{uu\_u} + \left(z_{u} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{u} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
A and B are the common perturbation matrices:
$A = f_{y_0} + \begin{pmatrix} \underbrace{0}_{n\times n_{static}} &\vdots& \underbrace{f_{y^{**}_{+}} \cdot g^{**}_{x}}_{n \times n_{spred}} &\vdots& \underbrace{0}_{n\times n_{frwd}} \end{pmatrix}$and $B = \begin{pmatrix} \underbrace{0}_{n \times n_{static}}&\vdots & \underbrace{0}_{n \times n_{pred}} & \vdots & \underbrace{f_{y^{**}_{+}}}_{n \times n_{sfwrd}} \end{pmatrix}$
and $z=(y_{-}^{\ast}; y; y_{+}^{\ast\ast}; u)$ denotes the dynamic model variables as in `M_.lead_lag_incidence`, $y^\ast$ denote state variables, $y^{\ast\ast}$ denote forward looking variables, $y_+$ denote the variables with a lead, $y_{-}$ denote variables with a lag, $y_0$ denote variables at period t, $f$ the model equations, and $f_z$ the first-order dynamic model derivatives, $f_{zz}$ the second-order dynamic derivatives, and $f_{zzz}$ the third-order dynamic model derivatives. Then:
$z_{x} = \begin{pmatrix}I\\g_{x}\\g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{u} =\begin{pmatrix}0\\g_{u}\\g^{**}_{x} \cdot g^{*}_{u}\\I\end{pmatrix}$, $z_{u_{+}} =\begin{pmatrix}0\\0\\g^{**}_{u}\\0\end{pmatrix}$
$z_{xx} = \begin{pmatrix} 0\\g_{xx}\\g^{**}_{x} \left( g^{*}_x \otimes g^{*}_{x} \right) + g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{xu} =\begin{pmatrix}0\\g_{xu}\\g^{**}_{xx} \left( g^{*}_x \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{xu}\\0\end{pmatrix}$, $z_{uu} =\begin{pmatrix}0\\g_{uu}\\g^{**}_{xx} \left( g^{*}_u \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{uu}\\0\end{pmatrix}$,
$z_{xu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_x \otimes I \right)\\0\end{pmatrix}$, $z_{uu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_{u} \otimes I \right)\\0\end{pmatrix}$, $z_{u_{+}u_{+}} =\begin{pmatrix}0\\0\\g^{\ast\ast}_{uu}\\0\end{pmatrix}$, $z_{\sigma\sigma} = \begin{pmatrix}0\\ g_{\sigma\sigma}\\ g^{\ast\ast}_{x}g^{\ast}_{\sigma\sigma} + g^{\ast\ast}_{\sigma\sigma}\\0 \end{pmatrix}$
$P$ are permutation matrices that can be computed using Matlab's `ipermute` function.
## Parameter derivatives of perturbation solutions
First, we need the parameter derivatives of first, second, third, and fourth derivatives of the dynamic model (i.e. g1,g2,g3,g4 in dynamic files), I make use of the implicit function theorem: Let $f_{z^k}$ denote the kth derivative (wrt all dynamic variables) of the dynamic model, then let $df_{z^k}$ denote the first-derivative (wrt all model parameters) of $f_{z^k}$ evaluated at the steady state. Note that $f_{z^k}$ is a function of both the model parameters $\theta$ and of the steady state of all dynamic variables $\bar{z}$, which also depend on the parameters. Hence, implicitly $f_{z^k}=f_{z^k}(\theta,\bar{z}(\theta))$ and $df_{z^k}$ consists of two parts:
1. direct derivative wrt to all model parameters given by the preprocessor in the `_params_derivs.m` files
2. contribution of derivative of steady state of dynamic variables (wrt all model parameters): $f_{z^{k+1}} \cdot d\bar{z}$
Note that we already have functionality to compute $d\bar{z}$ analytically.
Having this, the above perturbation systems are basically equations of the following types
$AX +BXC = RHS$ or $AX = RHS$
Now when taking the derivative (wrt to one single parameter $\theta_j$), we get
$A\mathrm{d}\{X\} + B\mathrm{d}\{X\}C = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X - \mathrm{d}\{B\}XC - BX\mathrm{d}\{C\}$
or
$A\mathrm{d}\{X\} = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X$
The first one is a Sylvester type equation, the second one can be solved by taking the inverse of $A$. The only diffculty and tedious work arrises in computing (the highly sparse) derivatives of $RHS$.
***
# New functions: `
## get_perturbation_params_derivs.m`and `get_perturbation_params_derivs_numerical_objective.m`
* The parameter derivatives up to third order are computed in the new function`get_perturbation_params_derivs.m` both analytically and numerically. For numerical derivatives `get_perturbation_params_derivs_numerical_objective.m` is the objective for `fjaco.m` or `hessian_sparse.m` or `hessian.m`.
* `get_perturbation_params_derivs.m` is basically an extended version of the previous `get_first_order_solution_params_derivs.m` function.
* * `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It is used for numerical derivatives, whenever `analytic_derivation_mode=-1|-2`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
* Detailed changes:
* Most important: notation of this function is now in accordance to the k_order_solver, i.e. we do not compute derivatives of Kalman transition matrices A and B, but rather the solution matrices ghx,ghu,ghxx,ghxu,ghuu,ghs2,ghxxx,ghxxu,ghxuu,ghuuu,ghxss,ghuss in accordance with notation used in `oo_.dr`. As a byproduct at first-order, focusing on ghx and ghu instead of Kalman transition matrices A and B makes the computations slightly faster for large models (e.g. for Quest the computations were faster by a couple of seconds, not much, but okay).
* Removed use of `kstate`, see also Dynare/dynare#1653 and Dynare/dynare!1656
* Output arguments are stored in a structure `DERIVS`, there is also a flag `d2flag` that computes parameter hessians needed only in `dsge_likelihood.m`.
* Removed `kronflag` as input. `options_.analytic_derivation_mode` is now used instead of `kronflag`.
* Removed `indvar`, an index that was used to selected specific variables in the derivatives. This is not needed, as we always compute the parameter derivatives for all variables first and then select a subset of variables. The selection now takes place in other functions, like `dsge_likelihood.m`.
* Introduced some checks: (i) deterministic exogenous variables are not supported, (ii) Kronecker method only compatible with first-order approximation so reset to sylvester method, (iii) for purely backward or forward models we need to be careful with the rows in `M_.lead_la g_incidence`, (iv) if `_params_derivs.m` files are missing an error is thrown.
* For numerical derivatives, if mod file does not contain an `estimated_params_block`, a temporary one with the most important parameter information is created.
## `unfold_g4.m`
* When evaluating g3 and g4 one needs to take into account that these do not contain symmetric elements, so one needs to use `unfold_g3.m` and the new function `unfold_g4.m`. This returns an unfolded version of the same matrix (i.e. with symmetric elements).
***
# New test models
`.gitignore` and `Makefile.am` are changed accordingly. Also now it is possible to run test suite on analytic_derivatives, i.e. run `make check m/analytic_derivatives`
## `analytic_derivatives/BrockMirman_PertParamsDerivs.mod`
* This is the Brock Mirman model, where we know the exact policy function $g$ for capital and consumption. As this does not imply a nonzero $g_{\sigma\sigma}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ I added some artificial equations to get nonzero solution matrices with respect to $\sigma$. The true perturbation solution matrices $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ are then computed analytically with Matlab's symbolic toolbox and saved in `nBrockMirmanSYM.mat`. There is a preprocessor flag that recreates these analytical computations if changes are needed (and to check whether I made some errors here ;-) )
* Then solution matrices up to third order and their parameter Jacobians are then compared to the ones computed by Dynare's `k_order_solver` and by `get_perturbation_params_derivs` for all `analytic_derivation_mode`'s. There will be an error if the maximum absolute deviation is too large, i.e. for numerical derivatives (`analytic_derivation_mode=-1|-2`) the tolerance is choosen lower (around 1e-5); for analytical methods we are stricter: around 1e-13 for first-order, 1e-12 for second order, and 1e-11 for third-order.
* As a side note, this mod file also checks Dynare's `k_order_solver` algorithm and throws an error if something is wrong.
* This test model shows that the new functionality works well. And analytical derivatives perform way better and accurate than numerical ones, even for this small model.
## `analytic_derivatives/burnside_3_order_PertParamsDerivs.mod`
* This builds upon `tests/k_order_perturbation/burnside_k_order.mod` and computes the true parameter derivatives analytically by hand.
* This test model also shows that the new functionality works well.
## `analytic_derivatives/LindeTrabandt2019.mod`
* Shows that the new functionality also works for medium-sized models, i.e. a SW type model solved at third order with 35 variables (11 states). 2 shocks and 20 parameters.
* This mod file can be used to tweak the speed of the computations in the future.
* Compares numerical versus analytical parameter derivatives (for first, second and third order). Note that this model clearly shows that numerical ones are quite different than analytical ones even at first order!
## `identification/LindeTrabandt2019_xfail.mod`
* This model is a check for issue Dynare/dynare#1595, see fjaco.m below, and will fail.
* Removed `analytic_derivatives/ls2003.mod` as this mod file is neither in the testsuite nor does it work.
***
# Detailed changes in other functions
## `get_first_order_solution_params_derivs.m`
* Deleted, or actually, renamed to `get_perturbation_params_derivs.m`, as this function now is able to compute the derivatives up to third order
## `identification_numerical_objective.m`
* `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
## `dsge_likelihood.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `T`matrix anymore, but instead of the dynare solution matrix `ghx`. So we recreate `T` here as this amounts to adding some zeros and focusing on selected variables only.
* Added some checks to make sure the first-order approximation is selected.
* Removed `kron_flag` as input, as `get_perturbation_params_derivs` looks into `options_.analytic_derivation_mode` for `kron_flag`.
## `dynare_identification.m`
* make sure that setting `analytic_derivation_mode` is set both in `options_ident` and `options_`. Note that at the end of the function we restore the `options_` structure, so all changes are local. In a next merge request, I will remove the global variables to make all variables local.
## `get_identification_jacobians.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `A` and `B` matrix anymore, but instead of the dynare solution matrix `ghx` and `ghu`. So we recreate these matrices here instead of in `get_perturbation_params_derivs.m`.
* Added `str2func` for better function handles in `fjaco.m`.
## `fjaco.m`
* make `tol`an option, which can be adjusted by changing `options_.dynatol.x`for identification and parameter derivatives purposes.
* include a check and an informative error message, if numerical derivatives (two-sided finite difference method) yield errors in `resol.m` for identification and parameter derivatives purposes. This closes issue Dynare/dynare#1595.
* Changed year of copyright to 2010-2017,2019
***
# Further suggestions and questions
* Ones this is merged, I will merge request an improvement of the identification toolbox, which will work up to third order using the pruned state space. This will also remove some issues and bugs, and also I will remove global variables in this request.
* The third-order derivatives can be further improved by taking sparsity into account and use mex versions for kronecker products etc. I leave this for further testing (and if anybody actually uses this ;-) )
2019-12-17 19:17:09 +01:00
DERIVS . dYss = dYss ;
DERIVS . dghx = cat ( 3 , zeros ( M . endo_nbr , M . nspred , stderrparam_nbr + corrparam_nbr ) , dghx ) ;
DERIVS . dghu = cat ( 3 , zeros ( M . endo_nbr , M . exo_nbr , stderrparam_nbr + corrparam_nbr ) , dghu ) ;
DERIVS . dOm = dOm ;
if order > 1
DERIVS . dg2 = dg2 ;
DERIVS . dghxx = cat ( 3 , zeros ( M . endo_nbr , M . nspred ^2 , stderrparam_nbr + corrparam_nbr ) , dghxx ) ;
DERIVS . dghxu = cat ( 3 , zeros ( M . endo_nbr , M . nspred * M . exo_nbr , stderrparam_nbr + corrparam_nbr ) , dghxu ) ;
DERIVS . dghuu = cat ( 3 , zeros ( M . endo_nbr , M . exo_nbr ^2 , stderrparam_nbr + corrparam_nbr ) , dghuu ) ;
DERIVS . dghs2 = dghs2 ;
end
if order > 2
DERIVS . dg3 = dg3 ;
DERIVS . dghxxx = cat ( 3 , zeros ( M . endo_nbr , M . nspred ^3 , stderrparam_nbr + corrparam_nbr ) , dghxxx ) ;
DERIVS . dghxxu = cat ( 3 , zeros ( M . endo_nbr , M . nspred ^2 * M . exo_nbr , stderrparam_nbr + corrparam_nbr ) , dghxxu ) ;
DERIVS . dghxuu = cat ( 3 , zeros ( M . endo_nbr , M . nspred * M . exo_nbr ^2 , stderrparam_nbr + corrparam_nbr ) , dghxuu ) ;
DERIVS . dghuuu = cat ( 3 , zeros ( M . endo_nbr , M . exo_nbr ^3 , stderrparam_nbr + corrparam_nbr ) , dghuuu ) ;
DERIVS . dghxss = dghxss ;
DERIVS . dghuss = dghuss ;
end
%% Construct Hessian (wrt all selected parameters) of ghx, and Om=ghu*Sigma_e*ghu'
if d2flag
% Construct Hessian (wrt all selected parameters) of Sigma_e
% note that we only need to focus on (stderr x stderr), (stderr x corr), (corr x stderr) parameters, because derivatives wrt all other second-cross parameters are zero by construction
d2Sigma_e = zeros ( M . exo_nbr , M . exo_nbr , totparam_nbr ^2 ) ; %initialize full matrix, even though we'll reduce it later to unique upper triangular values
% Compute Hessian of Sigma_e wrt (stderr x stderr) parameters
if ~ isempty ( indp2stderrstderr )
for jp = 1 : stderrparam_nbr
for ip = 1 : jp
if jp == ip %same stderr parameters
d2Sigma_e ( indpstderr ( jp ) , indpstderr ( jp ) , indp2stderrstderr ( ip , jp ) ) = 2 ;
else %different stderr parameters
if isdiag ( M . Sigma_e ) == 0 % if there are correlated errors
d2Sigma_e ( indpstderr ( jp ) , indpstderr ( ip ) , indp2stderrstderr ( ip , jp ) ) = M . Correlation_matrix ( indpstderr ( jp ) , indpstderr ( ip ) ) ;
d2Sigma_e ( indpstderr ( ip ) , indpstderr ( jp ) , indp2stderrstderr ( ip , jp ) ) = M . Correlation_matrix ( indpstderr ( jp ) , indpstderr ( ip ) ) ; %symmetry
end
end
end
end
end
% Compute Hessian of Sigma_e wrt (stderr x corr) parameters
if ~ isempty ( indp2stderrcorr )
for jp = 1 : stderrparam_nbr
for ip = 1 : corrparam_nbr
if indpstderr ( jp ) == indpcorr ( ip , 1 ) %if stderr is equal to first index of corr parameter, then derivative is equal to stderr of second index
d2Sigma_e ( indpstderr ( jp ) , indpcorr ( ip , 2 ) , indp2stderrcorr ( jp , ip ) ) = sqrt ( M . Sigma_e ( indpcorr ( ip , 2 ) , indpcorr ( ip , 2 ) ) ) ;
d2Sigma_e ( indpcorr ( ip , 2 ) , indpstderr ( jp ) , indp2stderrcorr ( jp , ip ) ) = sqrt ( M . Sigma_e ( indpcorr ( ip , 2 ) , indpcorr ( ip , 2 ) ) ) ; % symmetry
end
if indpstderr ( jp ) == indpcorr ( ip , 2 ) %if stderr is equal to second index of corr parameter, then derivative is equal to stderr of first index
d2Sigma_e ( indpstderr ( jp ) , indpcorr ( ip , 1 ) , indp2stderrcorr ( jp , ip ) ) = sqrt ( M . Sigma_e ( indpcorr ( ip , 1 ) , indpcorr ( ip , 1 ) ) ) ;
d2Sigma_e ( indpcorr ( ip , 1 ) , indpstderr ( jp ) , indp2stderrcorr ( jp , ip ) ) = sqrt ( M . Sigma_e ( indpcorr ( ip , 1 ) , indpcorr ( ip , 1 ) ) ) ; % symmetry
end
end
end
end
d2Sigma_e = d2Sigma_e ( : , : , indp2tottot2 ) ; %focus on upper triangular hessian values only
% Construct nonzero derivatives wrt to t+1, i.e. GAM1=-f_{y^+} in Villemot (2011)
GAM1 = zeros ( M . endo_nbr , M . endo_nbr ) ;
GAM1 ( : , klead ~= 0 , : ) = - g1 ( : , nonzeros ( klead ) ) ;
dGAM1 = zeros ( M . endo_nbr , M . endo_nbr , modparam_nbr ) ;
dGAM1 ( : , klead ~= 0 , : ) = - dg1 ( : , nonzeros ( klead ) , : ) ;
indind = ismember ( d2g1 ( : , 2 ) , nonzeros ( klead ) ) ;
tmp = d2g1 ( indind , : ) ;
tmp ( : , end ) = - tmp ( : , end ) ;
d2GAM1 = tmp ;
indklead = find ( klead ~= 0 ) ;
for j = 1 : size ( tmp , 1 )
inxinx = ( nonzeros ( klead ) == tmp ( j , 2 ) ) ;
d2GAM1 ( j , 2 ) = indklead ( inxinx ) ;
end
% Construct nonzero derivatives wrt to t, i.e. GAM0=f_{y^0} in Villemot (2011)
GAM0 = zeros ( M . endo_nbr , M . endo_nbr ) ;
GAM0 ( : , kcurr ~= 0 , : ) = g1 ( : , nonzeros ( kcurr ) ) ;
dGAM0 = zeros ( M . endo_nbr , M . endo_nbr , modparam_nbr ) ;
dGAM0 ( : , kcurr ~= 0 , : ) = dg1 ( : , nonzeros ( kcurr ) , : ) ;
indind = ismember ( d2g1 ( : , 2 ) , nonzeros ( kcurr ) ) ;
tmp = d2g1 ( indind , : ) ;
d2GAM0 = tmp ;
indkcurr = find ( kcurr ~= 0 ) ;
for j = 1 : size ( tmp , 1 )
inxinx = ( nonzeros ( kcurr ) == tmp ( j , 2 ) ) ;
d2GAM0 ( j , 2 ) = indkcurr ( inxinx ) ;
end
% Construct nonzero derivatives wrt to t-1, i.e. GAM2=-f_{y^-} in Villemot (2011)
% GAM2 = zeros(M.endo_nbr,M.endo_nbr);
% GAM2(:,klag~=0) = -g1(:,nonzeros(klag));
% dGAM2 = zeros(M.endo_nbr,M.endo_nbr,modparam_nbr);
% dGAM2(:,klag~=0) = -dg1(:,nonzeros(klag),:);
indind = ismember ( d2g1 ( : , 2 ) , nonzeros ( klag ) ) ;
tmp = d2g1 ( indind , : ) ;
tmp ( : , end ) = - tmp ( : , end ) ;
d2GAM2 = tmp ;
indklag = find ( klag ~= 0 ) ;
for j = 1 : size ( tmp , 1 )
inxinx = ( nonzeros ( klag ) == tmp ( j , 2 ) ) ;
d2GAM2 ( j , 2 ) = indklag ( inxinx ) ;
end
% Construct nonzero derivatives wrt to u_t, i.e. GAM3=-f_{u} in Villemot (2011)
% GAM3 = -g1(:,yy0_nbr+1:end);
% dGAM3 = -dg1(:,yy0_nbr+1:end,:);
cols_ex = yy0_nbr + ( 1 : yy0ex0_nbr ) ;
indind = ismember ( d2g1 ( : , 2 ) , cols_ex ) ;
tmp = d2g1 ( indind , : ) ;
tmp ( : , end ) = - tmp ( : , end ) ;
d2GAM3 = tmp ;
for j = 1 : size ( tmp , 1 )
inxinx = find ( cols_ex == tmp ( j , 2 ) ) ;
d2GAM3 ( j , 2 ) = inxinx ;
end
clear d2g1 tmp
% Compute Hessian (wrt selected params) of ghx using generalized sylvester equations, see equations 17 and 18 in Ratto and Iskrev (2012)
2019-12-20 16:28:06 +01:00
% solves MM*d2KalmanA+N*d2KalmanA*P = QQ where d2KalmanA are second order derivatives (wrt model parameters) of KalmanA
Added parameter derivatives of perturbation solution up to 3 order
# Preliminary comments
I finished the identification toolbox at orders two and three using the pruned state space system, but before I merge request this, I decided to first merge the new functionality to compute parameter derivatives of perturbation solution matrices at higher orders. So after this is approved, I merge the identification toolbox.
I guess @rattoma, @sebastien, and @michel are best choices to review this.
I outline the main idea first and then provide some more detailed changes I made to the functions.
***
# Main idea
This merge request is concerned with the *analytical*computation of the parameter derivatives of first, second and third order perturbation solution matrices, i.e. using _closed-form_ expressions to efficiently compute the derivative of $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ *with respect to model parameters* $\theta$. Note that $\theta$ contains model parameters, stderr and corr parameters of shocks. stderr and corr parameters of measurement errors are not yet supported, (they can easily be included as exogenous shocks). The availability of such derivatives is beneficial in terms of more reliable analysis of model sensitivity and parameter identifiability as well as more efficient estimation methods, in particular for models solved up to third order, as it is well-known that numerical derivatives are a tricky business, especially for large models.
References for my approach are:
* Iskrev (2008, 2010) and Schmitt-Grohé and Uribe (2012, Appendix) who were the first to compute the parameter derivatives analytically at first order, however, using inefficient (sparse) Kronecker products.
* Mutschler (2015) who provides the expressions for a second-order, but again using inefficient (sparse) Kronecker products.
* Ratto and Iskrev (2012) who show how the first-order system can be solved accurately, fast and efficiently using existing numerical algorithms for generalized Sylvester equations by taking the parameter derivative with respect to each parameter separately.
* Julliard and Kamenik (2004) who provide the perturbation solution equation system in tensor notation at any order k.
* Levintal (2017) who introduces permutation matrices to express the perturbation solution equation system in matrix notation up to fifth order.
Note that @rattoma already implemented the parameter derivatives of $g_x$ and $g_u$ analytically (and numerically), and I rely heavily on his work in `get_first_order_solution_params_derivs.m` (previously `getH.m`). My additions are mainly to this function and thus it is renamed to `get_perturbation_params_derivs.m`.
The basic idea of this merge request is to take the second- and third-order perturbation solution systems in Julliard and Kamenik (2004), unfold these into an equivalent matrix representation using permutation matrices as in Levintal (2017). Then extending Ratto and Iskrev (2012) one takes the derivative with respect to each parameter separately and gets a computational problem that is linear, albeit large, as it involves either solving generalized Sylvester equations or taking inverses of highly sparse matrices. I will now briefly summarize the perturbation solution system at third order and the system that results when taking the derivative with respect to parameters.
## Perturbation Solution
The following systems arise at first, second, and third order:
$(ghx): f_{x} z_{x} = f_{y_{-}^*} + f_{y_0} g_{x} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{x}= A g_{x} + f_{y_{-}^*}=0$
$(ghu): f_{z} z_{u} = f_{y_0} g_{u} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{u} + f_{u}= A g_u + f_u = 0$
$(ghxx) : A g_{xx} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{x}\right) + f_{zz} \left( z_{x} \otimes z_{x} \right) = 0$
$(ghxu) : A g_{xu} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{x} \otimes z_{u} \right) = 0$
$(ghuu) : A g_{uu} + B g_{xx} \left(g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{u} \otimes z_{u} \right) = 0$
$(ghs2) : (A+B) g_{\sigma\sigma} + \left( f_{y^{**}_{+}y^{**}_{+}} \left(g^{**}_{u} \otimes g^{**}_{u}\right) + f_{y^{**}_{+}} g^{**}_{uu}\right)vec(\Sigma) = 0$
$(ghxxx) : A g_{xxx} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{x}\right) + f_{y_{+}}g^{**}_{xx} \left(g^{*}_x \otimes g^{*}_{xx}\right)P_{x\_xx} + f_{zz} \left( z_{x} \otimes z_{xx} \right)P_{x\_xx} + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{x} \right) = 0$
$(ghxxu) : A g_{xxu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{u} \right) + f_{zz} \left( \left( z_{x} \otimes z_{xu} \right)P_{x\_xu} + \left(z_{xx} \otimes z_{u}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{x} \otimes g^{*}_{xu}\right)P_{x\_xu} + \left(g^{*}_{xx} \otimes g^{*}_{u}\right) \right) = 0$
$(ghxuu) : A g_{xuu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( \left( z_{xu} \otimes z_{u} \right)P_{xu\_u} + \left(z_{x} \otimes z_{uu}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{xu} \otimes g^{*}_{u}\right)P_{xu\_u} + \left(g^{*}_{x} \otimes g^{*}_{uu}\right) \right) = 0$
$(ghuuu) : A g_{uuu} + B g_{xxx} \left(g^{*}_{u} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{u} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( z_{u} \otimes z_{uu} \right)P_{u\_uu} + f_{y_{+}}g^{**}_{xx} \left(g^{*}_{u} \otimes g^{*}_{uu}\right)P_{u\_uu} = 0$
$(ghx\sigma\sigma) : A g_{x\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_x + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{x} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{x} \otimes z_{\sigma\sigma}\right) + F_{xu_{+}u_{+}}\left(I_{n_x} \otimes vec(\Sigma)\right) = 0$
$F_{xu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_x^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{xu_{+}} \otimes z_{u_{+}} \right)P_{xu\_u} + \left(z_{x} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{x} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
$(ghu\sigma\sigma) : A g_{u\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_{u} + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{u} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{u} \otimes z_{\sigma\sigma}\right) + F_{uu_{+}u_{+}}\left(I_{n_u} \otimes vec(\Sigma_u)\right) = 0$
$F_{uu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_u^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{uu_{+}} \otimes z_{u_{+}} \right)P_{uu\_u} + \left(z_{u} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{u} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
A and B are the common perturbation matrices:
$A = f_{y_0} + \begin{pmatrix} \underbrace{0}_{n\times n_{static}} &\vdots& \underbrace{f_{y^{**}_{+}} \cdot g^{**}_{x}}_{n \times n_{spred}} &\vdots& \underbrace{0}_{n\times n_{frwd}} \end{pmatrix}$and $B = \begin{pmatrix} \underbrace{0}_{n \times n_{static}}&\vdots & \underbrace{0}_{n \times n_{pred}} & \vdots & \underbrace{f_{y^{**}_{+}}}_{n \times n_{sfwrd}} \end{pmatrix}$
and $z=(y_{-}^{\ast}; y; y_{+}^{\ast\ast}; u)$ denotes the dynamic model variables as in `M_.lead_lag_incidence`, $y^\ast$ denote state variables, $y^{\ast\ast}$ denote forward looking variables, $y_+$ denote the variables with a lead, $y_{-}$ denote variables with a lag, $y_0$ denote variables at period t, $f$ the model equations, and $f_z$ the first-order dynamic model derivatives, $f_{zz}$ the second-order dynamic derivatives, and $f_{zzz}$ the third-order dynamic model derivatives. Then:
$z_{x} = \begin{pmatrix}I\\g_{x}\\g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{u} =\begin{pmatrix}0\\g_{u}\\g^{**}_{x} \cdot g^{*}_{u}\\I\end{pmatrix}$, $z_{u_{+}} =\begin{pmatrix}0\\0\\g^{**}_{u}\\0\end{pmatrix}$
$z_{xx} = \begin{pmatrix} 0\\g_{xx}\\g^{**}_{x} \left( g^{*}_x \otimes g^{*}_{x} \right) + g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{xu} =\begin{pmatrix}0\\g_{xu}\\g^{**}_{xx} \left( g^{*}_x \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{xu}\\0\end{pmatrix}$, $z_{uu} =\begin{pmatrix}0\\g_{uu}\\g^{**}_{xx} \left( g^{*}_u \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{uu}\\0\end{pmatrix}$,
$z_{xu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_x \otimes I \right)\\0\end{pmatrix}$, $z_{uu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_{u} \otimes I \right)\\0\end{pmatrix}$, $z_{u_{+}u_{+}} =\begin{pmatrix}0\\0\\g^{\ast\ast}_{uu}\\0\end{pmatrix}$, $z_{\sigma\sigma} = \begin{pmatrix}0\\ g_{\sigma\sigma}\\ g^{\ast\ast}_{x}g^{\ast}_{\sigma\sigma} + g^{\ast\ast}_{\sigma\sigma}\\0 \end{pmatrix}$
$P$ are permutation matrices that can be computed using Matlab's `ipermute` function.
## Parameter derivatives of perturbation solutions
First, we need the parameter derivatives of first, second, third, and fourth derivatives of the dynamic model (i.e. g1,g2,g3,g4 in dynamic files), I make use of the implicit function theorem: Let $f_{z^k}$ denote the kth derivative (wrt all dynamic variables) of the dynamic model, then let $df_{z^k}$ denote the first-derivative (wrt all model parameters) of $f_{z^k}$ evaluated at the steady state. Note that $f_{z^k}$ is a function of both the model parameters $\theta$ and of the steady state of all dynamic variables $\bar{z}$, which also depend on the parameters. Hence, implicitly $f_{z^k}=f_{z^k}(\theta,\bar{z}(\theta))$ and $df_{z^k}$ consists of two parts:
1. direct derivative wrt to all model parameters given by the preprocessor in the `_params_derivs.m` files
2. contribution of derivative of steady state of dynamic variables (wrt all model parameters): $f_{z^{k+1}} \cdot d\bar{z}$
Note that we already have functionality to compute $d\bar{z}$ analytically.
Having this, the above perturbation systems are basically equations of the following types
$AX +BXC = RHS$ or $AX = RHS$
Now when taking the derivative (wrt to one single parameter $\theta_j$), we get
$A\mathrm{d}\{X\} + B\mathrm{d}\{X\}C = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X - \mathrm{d}\{B\}XC - BX\mathrm{d}\{C\}$
or
$A\mathrm{d}\{X\} = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X$
The first one is a Sylvester type equation, the second one can be solved by taking the inverse of $A$. The only diffculty and tedious work arrises in computing (the highly sparse) derivatives of $RHS$.
***
# New functions: `
## get_perturbation_params_derivs.m`and `get_perturbation_params_derivs_numerical_objective.m`
* The parameter derivatives up to third order are computed in the new function`get_perturbation_params_derivs.m` both analytically and numerically. For numerical derivatives `get_perturbation_params_derivs_numerical_objective.m` is the objective for `fjaco.m` or `hessian_sparse.m` or `hessian.m`.
* `get_perturbation_params_derivs.m` is basically an extended version of the previous `get_first_order_solution_params_derivs.m` function.
* * `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It is used for numerical derivatives, whenever `analytic_derivation_mode=-1|-2`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
* Detailed changes:
* Most important: notation of this function is now in accordance to the k_order_solver, i.e. we do not compute derivatives of Kalman transition matrices A and B, but rather the solution matrices ghx,ghu,ghxx,ghxu,ghuu,ghs2,ghxxx,ghxxu,ghxuu,ghuuu,ghxss,ghuss in accordance with notation used in `oo_.dr`. As a byproduct at first-order, focusing on ghx and ghu instead of Kalman transition matrices A and B makes the computations slightly faster for large models (e.g. for Quest the computations were faster by a couple of seconds, not much, but okay).
* Removed use of `kstate`, see also Dynare/dynare#1653 and Dynare/dynare!1656
* Output arguments are stored in a structure `DERIVS`, there is also a flag `d2flag` that computes parameter hessians needed only in `dsge_likelihood.m`.
* Removed `kronflag` as input. `options_.analytic_derivation_mode` is now used instead of `kronflag`.
* Removed `indvar`, an index that was used to selected specific variables in the derivatives. This is not needed, as we always compute the parameter derivatives for all variables first and then select a subset of variables. The selection now takes place in other functions, like `dsge_likelihood.m`.
* Introduced some checks: (i) deterministic exogenous variables are not supported, (ii) Kronecker method only compatible with first-order approximation so reset to sylvester method, (iii) for purely backward or forward models we need to be careful with the rows in `M_.lead_la g_incidence`, (iv) if `_params_derivs.m` files are missing an error is thrown.
* For numerical derivatives, if mod file does not contain an `estimated_params_block`, a temporary one with the most important parameter information is created.
## `unfold_g4.m`
* When evaluating g3 and g4 one needs to take into account that these do not contain symmetric elements, so one needs to use `unfold_g3.m` and the new function `unfold_g4.m`. This returns an unfolded version of the same matrix (i.e. with symmetric elements).
***
# New test models
`.gitignore` and `Makefile.am` are changed accordingly. Also now it is possible to run test suite on analytic_derivatives, i.e. run `make check m/analytic_derivatives`
## `analytic_derivatives/BrockMirman_PertParamsDerivs.mod`
* This is the Brock Mirman model, where we know the exact policy function $g$ for capital and consumption. As this does not imply a nonzero $g_{\sigma\sigma}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ I added some artificial equations to get nonzero solution matrices with respect to $\sigma$. The true perturbation solution matrices $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ are then computed analytically with Matlab's symbolic toolbox and saved in `nBrockMirmanSYM.mat`. There is a preprocessor flag that recreates these analytical computations if changes are needed (and to check whether I made some errors here ;-) )
* Then solution matrices up to third order and their parameter Jacobians are then compared to the ones computed by Dynare's `k_order_solver` and by `get_perturbation_params_derivs` for all `analytic_derivation_mode`'s. There will be an error if the maximum absolute deviation is too large, i.e. for numerical derivatives (`analytic_derivation_mode=-1|-2`) the tolerance is choosen lower (around 1e-5); for analytical methods we are stricter: around 1e-13 for first-order, 1e-12 for second order, and 1e-11 for third-order.
* As a side note, this mod file also checks Dynare's `k_order_solver` algorithm and throws an error if something is wrong.
* This test model shows that the new functionality works well. And analytical derivatives perform way better and accurate than numerical ones, even for this small model.
## `analytic_derivatives/burnside_3_order_PertParamsDerivs.mod`
* This builds upon `tests/k_order_perturbation/burnside_k_order.mod` and computes the true parameter derivatives analytically by hand.
* This test model also shows that the new functionality works well.
## `analytic_derivatives/LindeTrabandt2019.mod`
* Shows that the new functionality also works for medium-sized models, i.e. a SW type model solved at third order with 35 variables (11 states). 2 shocks and 20 parameters.
* This mod file can be used to tweak the speed of the computations in the future.
* Compares numerical versus analytical parameter derivatives (for first, second and third order). Note that this model clearly shows that numerical ones are quite different than analytical ones even at first order!
## `identification/LindeTrabandt2019_xfail.mod`
* This model is a check for issue Dynare/dynare#1595, see fjaco.m below, and will fail.
* Removed `analytic_derivatives/ls2003.mod` as this mod file is neither in the testsuite nor does it work.
***
# Detailed changes in other functions
## `get_first_order_solution_params_derivs.m`
* Deleted, or actually, renamed to `get_perturbation_params_derivs.m`, as this function now is able to compute the derivatives up to third order
## `identification_numerical_objective.m`
* `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
## `dsge_likelihood.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `T`matrix anymore, but instead of the dynare solution matrix `ghx`. So we recreate `T` here as this amounts to adding some zeros and focusing on selected variables only.
* Added some checks to make sure the first-order approximation is selected.
* Removed `kron_flag` as input, as `get_perturbation_params_derivs` looks into `options_.analytic_derivation_mode` for `kron_flag`.
## `dynare_identification.m`
* make sure that setting `analytic_derivation_mode` is set both in `options_ident` and `options_`. Note that at the end of the function we restore the `options_` structure, so all changes are local. In a next merge request, I will remove the global variables to make all variables local.
## `get_identification_jacobians.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `A` and `B` matrix anymore, but instead of the dynare solution matrix `ghx` and `ghu`. So we recreate these matrices here instead of in `get_perturbation_params_derivs.m`.
* Added `str2func` for better function handles in `fjaco.m`.
## `fjaco.m`
* make `tol`an option, which can be adjusted by changing `options_.dynatol.x`for identification and parameter derivatives purposes.
* include a check and an informative error message, if numerical derivatives (two-sided finite difference method) yield errors in `resol.m` for identification and parameter derivatives purposes. This closes issue Dynare/dynare#1595.
* Changed year of copyright to 2010-2017,2019
***
# Further suggestions and questions
* Ones this is merged, I will merge request an improvement of the identification toolbox, which will work up to third order using the pruned state space. This will also remove some issues and bugs, and also I will remove global variables in this request.
* The third-order derivatives can be further improved by taking sparsity into account and use mex versions for kronecker products etc. I leave this for further testing (and if anybody actually uses this ;-) )
2019-12-17 19:17:09 +01:00
QQ = zeros ( M . endo_nbr , M . endo_nbr , floor ( sqrt ( modparam_nbr2 ) ) ) ;
jcount = 0 ;
cumjcount = 0 ;
jinx = [ ] ;
x2x = sparse ( M . endo_nbr * M . endo_nbr , modparam_nbr2 ) ;
dKalmanA = zeros ( M . endo_nbr , M . endo_nbr , modparam_nbr ) ;
dKalmanA ( : , idx_states , : ) = dghx ;
MM = ( GAM0 - GAM1 * KalmanA ) ;
invMM = inv ( MM ) ;
for i = 1 : modparam_nbr
for j = 1 : i
elem1 = ( get_2nd_deriv ( d2GAM0 , M . endo_nbr , M . endo_nbr , j , i ) - get_2nd_deriv ( d2GAM1 , M . endo_nbr , M . endo_nbr , j , i ) * KalmanA ) ;
elem1 = get_2nd_deriv ( d2GAM2 , M . endo_nbr , M . endo_nbr , j , i ) - elem1 * KalmanA ;
elemj0 = dGAM0 ( : , : , j ) - dGAM1 ( : , : , j ) * KalmanA ;
elemi0 = dGAM0 ( : , : , i ) - dGAM1 ( : , : , i ) * KalmanA ;
elem2 = - elemj0 * dKalmanA ( : , : , i ) - elemi0 * dKalmanA ( : , : , j ) ;
elem2 = elem2 + ( dGAM1 ( : , : , j ) * dKalmanA ( : , : , i ) + dGAM1 ( : , : , i ) * dKalmanA ( : , : , j ) ) * KalmanA ;
elem2 = elem2 + GAM1 * ( dKalmanA ( : , : , i ) * dKalmanA ( : , : , j ) + dKalmanA ( : , : , j ) * dKalmanA ( : , : , i ) ) ;
jcount = jcount + 1 ;
jinx = [ jinx ; [ j i ] ] ;
QQ ( : , : , jcount ) = elem1 + elem2 ;
if jcount == floor ( sqrt ( modparam_nbr2 ) ) || ( j * i ) == modparam_nbr ^2
if ( j * i ) == modparam_nbr ^2
QQ = QQ ( : , : , 1 : jcount ) ;
end
xx2 = sylvester3 ( MM , - GAM1 , KalmanA , QQ ) ;
flag = 1 ;
icount = 0 ;
while flag && icount < 4
[ xx2 , flag ] = sylvester3a ( xx2 , MM , - GAM1 , KalmanA , QQ ) ;
icount = icount + 1 ;
end
x2x ( : , cumjcount + 1 : cumjcount + jcount ) = reshape ( xx2 , [ M . endo_nbr * M . endo_nbr jcount ] ) ;
cumjcount = cumjcount + jcount ;
jcount = 0 ;
jinx = [ ] ;
end
end
end
clear xx2 ;
jcount = 0 ;
icount = 0 ;
cumjcount = 0 ;
MAX_DIM_MAT = 100000000 ;
ncol = max ( 1 , floor ( MAX_DIM_MAT / ( 8 * M . endo_nbr * ( M . endo_nbr + 1 ) / 2 ) ) ) ;
ncol = min ( ncol , totparam_nbr2 ) ;
d2KalmanA = sparse ( M . endo_nbr * M . endo_nbr , totparam_nbr2 ) ;
d2Om = sparse ( M . endo_nbr * ( M . endo_nbr + 1 ) / 2 , totparam_nbr2 ) ;
d2KalmanA_tmp = zeros ( M . endo_nbr * M . endo_nbr , ncol ) ;
d2Om_tmp = zeros ( M . endo_nbr * ( M . endo_nbr + 1 ) / 2 , ncol ) ;
tmpDir = CheckPath ( ' tmp_derivs' , M . dname ) ;
offset = stderrparam_nbr + corrparam_nbr ;
% d2B = zeros(m,n,tot_param_nbr,tot_param_nbr);
for j = 1 : totparam_nbr
for i = 1 : j
jcount = jcount + 1 ;
if j < = offset %stderr and corr parameters
2019-12-20 16:28:06 +01:00
y = KalmanB * d2Sigma_e ( : , : , jcount ) * KalmanB ' ;
d2Om_tmp ( : , jcount ) = dyn_vech ( y ) ;
Added parameter derivatives of perturbation solution up to 3 order
# Preliminary comments
I finished the identification toolbox at orders two and three using the pruned state space system, but before I merge request this, I decided to first merge the new functionality to compute parameter derivatives of perturbation solution matrices at higher orders. So after this is approved, I merge the identification toolbox.
I guess @rattoma, @sebastien, and @michel are best choices to review this.
I outline the main idea first and then provide some more detailed changes I made to the functions.
***
# Main idea
This merge request is concerned with the *analytical*computation of the parameter derivatives of first, second and third order perturbation solution matrices, i.e. using _closed-form_ expressions to efficiently compute the derivative of $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ *with respect to model parameters* $\theta$. Note that $\theta$ contains model parameters, stderr and corr parameters of shocks. stderr and corr parameters of measurement errors are not yet supported, (they can easily be included as exogenous shocks). The availability of such derivatives is beneficial in terms of more reliable analysis of model sensitivity and parameter identifiability as well as more efficient estimation methods, in particular for models solved up to third order, as it is well-known that numerical derivatives are a tricky business, especially for large models.
References for my approach are:
* Iskrev (2008, 2010) and Schmitt-Grohé and Uribe (2012, Appendix) who were the first to compute the parameter derivatives analytically at first order, however, using inefficient (sparse) Kronecker products.
* Mutschler (2015) who provides the expressions for a second-order, but again using inefficient (sparse) Kronecker products.
* Ratto and Iskrev (2012) who show how the first-order system can be solved accurately, fast and efficiently using existing numerical algorithms for generalized Sylvester equations by taking the parameter derivative with respect to each parameter separately.
* Julliard and Kamenik (2004) who provide the perturbation solution equation system in tensor notation at any order k.
* Levintal (2017) who introduces permutation matrices to express the perturbation solution equation system in matrix notation up to fifth order.
Note that @rattoma already implemented the parameter derivatives of $g_x$ and $g_u$ analytically (and numerically), and I rely heavily on his work in `get_first_order_solution_params_derivs.m` (previously `getH.m`). My additions are mainly to this function and thus it is renamed to `get_perturbation_params_derivs.m`.
The basic idea of this merge request is to take the second- and third-order perturbation solution systems in Julliard and Kamenik (2004), unfold these into an equivalent matrix representation using permutation matrices as in Levintal (2017). Then extending Ratto and Iskrev (2012) one takes the derivative with respect to each parameter separately and gets a computational problem that is linear, albeit large, as it involves either solving generalized Sylvester equations or taking inverses of highly sparse matrices. I will now briefly summarize the perturbation solution system at third order and the system that results when taking the derivative with respect to parameters.
## Perturbation Solution
The following systems arise at first, second, and third order:
$(ghx): f_{x} z_{x} = f_{y_{-}^*} + f_{y_0} g_{x} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{x}= A g_{x} + f_{y_{-}^*}=0$
$(ghu): f_{z} z_{u} = f_{y_0} g_{u} + f_{y_{+}^{**}} g^{**}_{x} g^{*}_{u} + f_{u}= A g_u + f_u = 0$
$(ghxx) : A g_{xx} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{x}\right) + f_{zz} \left( z_{x} \otimes z_{x} \right) = 0$
$(ghxu) : A g_{xu} + B g_{xx} \left(g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{x} \otimes z_{u} \right) = 0$
$(ghuu) : A g_{uu} + B g_{xx} \left(g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zz} \left( z_{u} \otimes z_{u} \right) = 0$
$(ghs2) : (A+B) g_{\sigma\sigma} + \left( f_{y^{**}_{+}y^{**}_{+}} \left(g^{**}_{u} \otimes g^{**}_{u}\right) + f_{y^{**}_{+}} g^{**}_{uu}\right)vec(\Sigma) = 0$
$(ghxxx) : A g_{xxx} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{x}\right) + f_{y_{+}}g^{**}_{xx} \left(g^{*}_x \otimes g^{*}_{xx}\right)P_{x\_xx} + f_{zz} \left( z_{x} \otimes z_{xx} \right)P_{x\_xx} + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{x} \right) = 0$
$(ghxxu) : A g_{xxu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{x} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{x} \otimes z_{u} \right) + f_{zz} \left( \left( z_{x} \otimes z_{xu} \right)P_{x\_xu} + \left(z_{xx} \otimes z_{u}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{x} \otimes g^{*}_{xu}\right)P_{x\_xu} + \left(g^{*}_{xx} \otimes g^{*}_{u}\right) \right) = 0$
$(ghxuu) : A g_{xuu} + B g_{xxx} \left(g^{*}_{x} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{x} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( \left( z_{xu} \otimes z_{u} \right)P_{xu\_u} + \left(z_{x} \otimes z_{uu}\right) \right) + f_{y_{+}}g^{**}_{xx} \left( \left(g^{*}_{xu} \otimes g^{*}_{u}\right)P_{xu\_u} + \left(g^{*}_{x} \otimes g^{*}_{uu}\right) \right) = 0$
$(ghuuu) : A g_{uuu} + B g_{xxx} \left(g^{*}_{u} \otimes g^{*}_{u} \otimes g^{*}_{u}\right) + f_{zzz} \left( z_{u} \otimes z_{u} \otimes z_{u} \right)+ f_{zz} \left( z_{u} \otimes z_{uu} \right)P_{u\_uu} + f_{y_{+}}g^{**}_{xx} \left(g^{*}_{u} \otimes g^{*}_{uu}\right)P_{u\_uu} = 0$
$(ghx\sigma\sigma) : A g_{x\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_x + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{x} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{x} \otimes z_{\sigma\sigma}\right) + F_{xu_{+}u_{+}}\left(I_{n_x} \otimes vec(\Sigma)\right) = 0$
$F_{xu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_x^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{xu_{+}} \otimes z_{u_{+}} \right)P_{xu\_u} + \left(z_{x} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{x} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
$(ghu\sigma\sigma) : A g_{u\sigma\sigma} + B g_{x\sigma\sigma} g^{*}_{u} + f_{y_{+}} g^{**}_{xx}\left(g^{*}_{u} \otimes g^{*}_{\sigma\sigma}\right) + f_{zz} \left(z_{u} \otimes z_{\sigma\sigma}\right) + F_{uu_{+}u_{+}}\left(I_{n_u} \otimes vec(\Sigma_u)\right) = 0$
$F_{uu_{+}u_{+}} = f_{y_{+}^{\ast\ast}} g_{xuu}^{\ast\ast} (g_u^{\ast} \otimes I_{n_u^2}) + f_{zz} \left( \left( z_{uu_{+}} \otimes z_{u_{+}} \right)P_{uu\_u} + \left(z_{u} \otimes z_{u_{+}u_{+}}\right) \right) + f_{zzz}\left(z_{u} \otimes z_{u_{+}} \otimes z_{u_{+}}\right)$
A and B are the common perturbation matrices:
$A = f_{y_0} + \begin{pmatrix} \underbrace{0}_{n\times n_{static}} &\vdots& \underbrace{f_{y^{**}_{+}} \cdot g^{**}_{x}}_{n \times n_{spred}} &\vdots& \underbrace{0}_{n\times n_{frwd}} \end{pmatrix}$and $B = \begin{pmatrix} \underbrace{0}_{n \times n_{static}}&\vdots & \underbrace{0}_{n \times n_{pred}} & \vdots & \underbrace{f_{y^{**}_{+}}}_{n \times n_{sfwrd}} \end{pmatrix}$
and $z=(y_{-}^{\ast}; y; y_{+}^{\ast\ast}; u)$ denotes the dynamic model variables as in `M_.lead_lag_incidence`, $y^\ast$ denote state variables, $y^{\ast\ast}$ denote forward looking variables, $y_+$ denote the variables with a lead, $y_{-}$ denote variables with a lag, $y_0$ denote variables at period t, $f$ the model equations, and $f_z$ the first-order dynamic model derivatives, $f_{zz}$ the second-order dynamic derivatives, and $f_{zzz}$ the third-order dynamic model derivatives. Then:
$z_{x} = \begin{pmatrix}I\\g_{x}\\g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{u} =\begin{pmatrix}0\\g_{u}\\g^{**}_{x} \cdot g^{*}_{u}\\I\end{pmatrix}$, $z_{u_{+}} =\begin{pmatrix}0\\0\\g^{**}_{u}\\0\end{pmatrix}$
$z_{xx} = \begin{pmatrix} 0\\g_{xx}\\g^{**}_{x} \left( g^{*}_x \otimes g^{*}_{x} \right) + g^{**}_{x} g^{*}_{x}\\0\end{pmatrix}$, $z_{xu} =\begin{pmatrix}0\\g_{xu}\\g^{**}_{xx} \left( g^{*}_x \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{xu}\\0\end{pmatrix}$, $z_{uu} =\begin{pmatrix}0\\g_{uu}\\g^{**}_{xx} \left( g^{*}_u \otimes g^{*}_{u} \right) + g^{**}_{x} g^{*}_{uu}\\0\end{pmatrix}$,
$z_{xu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_x \otimes I \right)\\0\end{pmatrix}$, $z_{uu_{+}} =\begin{pmatrix}0\\0\\g^{**}_{xu} \left( g^{*}_{u} \otimes I \right)\\0\end{pmatrix}$, $z_{u_{+}u_{+}} =\begin{pmatrix}0\\0\\g^{\ast\ast}_{uu}\\0\end{pmatrix}$, $z_{\sigma\sigma} = \begin{pmatrix}0\\ g_{\sigma\sigma}\\ g^{\ast\ast}_{x}g^{\ast}_{\sigma\sigma} + g^{\ast\ast}_{\sigma\sigma}\\0 \end{pmatrix}$
$P$ are permutation matrices that can be computed using Matlab's `ipermute` function.
## Parameter derivatives of perturbation solutions
First, we need the parameter derivatives of first, second, third, and fourth derivatives of the dynamic model (i.e. g1,g2,g3,g4 in dynamic files), I make use of the implicit function theorem: Let $f_{z^k}$ denote the kth derivative (wrt all dynamic variables) of the dynamic model, then let $df_{z^k}$ denote the first-derivative (wrt all model parameters) of $f_{z^k}$ evaluated at the steady state. Note that $f_{z^k}$ is a function of both the model parameters $\theta$ and of the steady state of all dynamic variables $\bar{z}$, which also depend on the parameters. Hence, implicitly $f_{z^k}=f_{z^k}(\theta,\bar{z}(\theta))$ and $df_{z^k}$ consists of two parts:
1. direct derivative wrt to all model parameters given by the preprocessor in the `_params_derivs.m` files
2. contribution of derivative of steady state of dynamic variables (wrt all model parameters): $f_{z^{k+1}} \cdot d\bar{z}$
Note that we already have functionality to compute $d\bar{z}$ analytically.
Having this, the above perturbation systems are basically equations of the following types
$AX +BXC = RHS$ or $AX = RHS$
Now when taking the derivative (wrt to one single parameter $\theta_j$), we get
$A\mathrm{d}\{X\} + B\mathrm{d}\{X\}C = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X - \mathrm{d}\{B\}XC - BX\mathrm{d}\{C\}$
or
$A\mathrm{d}\{X\} = \mathrm{d}\{RHS\} - \mathrm{d}\{A\}X$
The first one is a Sylvester type equation, the second one can be solved by taking the inverse of $A$. The only diffculty and tedious work arrises in computing (the highly sparse) derivatives of $RHS$.
***
# New functions: `
## get_perturbation_params_derivs.m`and `get_perturbation_params_derivs_numerical_objective.m`
* The parameter derivatives up to third order are computed in the new function`get_perturbation_params_derivs.m` both analytically and numerically. For numerical derivatives `get_perturbation_params_derivs_numerical_objective.m` is the objective for `fjaco.m` or `hessian_sparse.m` or `hessian.m`.
* `get_perturbation_params_derivs.m` is basically an extended version of the previous `get_first_order_solution_params_derivs.m` function.
* * `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It is used for numerical derivatives, whenever `analytic_derivation_mode=-1|-2`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
* Detailed changes:
* Most important: notation of this function is now in accordance to the k_order_solver, i.e. we do not compute derivatives of Kalman transition matrices A and B, but rather the solution matrices ghx,ghu,ghxx,ghxu,ghuu,ghs2,ghxxx,ghxxu,ghxuu,ghuuu,ghxss,ghuss in accordance with notation used in `oo_.dr`. As a byproduct at first-order, focusing on ghx and ghu instead of Kalman transition matrices A and B makes the computations slightly faster for large models (e.g. for Quest the computations were faster by a couple of seconds, not much, but okay).
* Removed use of `kstate`, see also Dynare/dynare#1653 and Dynare/dynare!1656
* Output arguments are stored in a structure `DERIVS`, there is also a flag `d2flag` that computes parameter hessians needed only in `dsge_likelihood.m`.
* Removed `kronflag` as input. `options_.analytic_derivation_mode` is now used instead of `kronflag`.
* Removed `indvar`, an index that was used to selected specific variables in the derivatives. This is not needed, as we always compute the parameter derivatives for all variables first and then select a subset of variables. The selection now takes place in other functions, like `dsge_likelihood.m`.
* Introduced some checks: (i) deterministic exogenous variables are not supported, (ii) Kronecker method only compatible with first-order approximation so reset to sylvester method, (iii) for purely backward or forward models we need to be careful with the rows in `M_.lead_la g_incidence`, (iv) if `_params_derivs.m` files are missing an error is thrown.
* For numerical derivatives, if mod file does not contain an `estimated_params_block`, a temporary one with the most important parameter information is created.
## `unfold_g4.m`
* When evaluating g3 and g4 one needs to take into account that these do not contain symmetric elements, so one needs to use `unfold_g3.m` and the new function `unfold_g4.m`. This returns an unfolded version of the same matrix (i.e. with symmetric elements).
***
# New test models
`.gitignore` and `Makefile.am` are changed accordingly. Also now it is possible to run test suite on analytic_derivatives, i.e. run `make check m/analytic_derivatives`
## `analytic_derivatives/BrockMirman_PertParamsDerivs.mod`
* This is the Brock Mirman model, where we know the exact policy function $g$ for capital and consumption. As this does not imply a nonzero $g_{\sigma\sigma}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ I added some artificial equations to get nonzero solution matrices with respect to $\sigma$. The true perturbation solution matrices $g_x$ , $g_u$, $g_{xx}$, $g_{xu}$, $g_{uu}$, $g_{\sigma\sigma}$, $g_{xxx}$, $g_{xxu}$, $g_{xuu}$, $g_{uuu}$, $g_{x\sigma\sigma}$, $g_{u\sigma\sigma}$ are then computed analytically with Matlab's symbolic toolbox and saved in `nBrockMirmanSYM.mat`. There is a preprocessor flag that recreates these analytical computations if changes are needed (and to check whether I made some errors here ;-) )
* Then solution matrices up to third order and their parameter Jacobians are then compared to the ones computed by Dynare's `k_order_solver` and by `get_perturbation_params_derivs` for all `analytic_derivation_mode`'s. There will be an error if the maximum absolute deviation is too large, i.e. for numerical derivatives (`analytic_derivation_mode=-1|-2`) the tolerance is choosen lower (around 1e-5); for analytical methods we are stricter: around 1e-13 for first-order, 1e-12 for second order, and 1e-11 for third-order.
* As a side note, this mod file also checks Dynare's `k_order_solver` algorithm and throws an error if something is wrong.
* This test model shows that the new functionality works well. And analytical derivatives perform way better and accurate than numerical ones, even for this small model.
## `analytic_derivatives/burnside_3_order_PertParamsDerivs.mod`
* This builds upon `tests/k_order_perturbation/burnside_k_order.mod` and computes the true parameter derivatives analytically by hand.
* This test model also shows that the new functionality works well.
## `analytic_derivatives/LindeTrabandt2019.mod`
* Shows that the new functionality also works for medium-sized models, i.e. a SW type model solved at third order with 35 variables (11 states). 2 shocks and 20 parameters.
* This mod file can be used to tweak the speed of the computations in the future.
* Compares numerical versus analytical parameter derivatives (for first, second and third order). Note that this model clearly shows that numerical ones are quite different than analytical ones even at first order!
## `identification/LindeTrabandt2019_xfail.mod`
* This model is a check for issue Dynare/dynare#1595, see fjaco.m below, and will fail.
* Removed `analytic_derivatives/ls2003.mod` as this mod file is neither in the testsuite nor does it work.
***
# Detailed changes in other functions
## `get_first_order_solution_params_derivs.m`
* Deleted, or actually, renamed to `get_perturbation_params_derivs.m`, as this function now is able to compute the derivatives up to third order
## `identification_numerical_objective.m`
* `get_perturbation_params_derivs_numerical_objective.m`builds upon `identification_numerical_objective.m`. It takes from `identification_numerical_objective.m` the parts that compute numerical parameter Jacobians of steady state, dynamic model equations, and perturbation solution matrices. Hence, these parts are removed in `identification_numerical_objective.m` and it only computes numerical parameter Jacobian of moments and spectrum which are needed for identification analysis in `get_identification_jacobians.m`, when `analytic_derivation_mode=-1` only.
## `dsge_likelihood.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `T`matrix anymore, but instead of the dynare solution matrix `ghx`. So we recreate `T` here as this amounts to adding some zeros and focusing on selected variables only.
* Added some checks to make sure the first-order approximation is selected.
* Removed `kron_flag` as input, as `get_perturbation_params_derivs` looks into `options_.analytic_derivation_mode` for `kron_flag`.
## `dynare_identification.m`
* make sure that setting `analytic_derivation_mode` is set both in `options_ident` and `options_`. Note that at the end of the function we restore the `options_` structure, so all changes are local. In a next merge request, I will remove the global variables to make all variables local.
## `get_identification_jacobians.m`
* As `get_first_order_solution_params_derivs.m`is renamed to `get_perturbation_params_derivs.m`, the call is adapted. That is,`get_perturbation_params_derivs` does not compute the derivatives of the Kalman transition `A` and `B` matrix anymore, but instead of the dynare solution matrix `ghx` and `ghu`. So we recreate these matrices here instead of in `get_perturbation_params_derivs.m`.
* Added `str2func` for better function handles in `fjaco.m`.
## `fjaco.m`
* make `tol`an option, which can be adjusted by changing `options_.dynatol.x`for identification and parameter derivatives purposes.
* include a check and an informative error message, if numerical derivatives (two-sided finite difference method) yield errors in `resol.m` for identification and parameter derivatives purposes. This closes issue Dynare/dynare#1595.
* Changed year of copyright to 2010-2017,2019
***
# Further suggestions and questions
* Ones this is merged, I will merge request an improvement of the identification toolbox, which will work up to third order using the pruned state space. This will also remove some issues and bugs, and also I will remove global variables in this request.
* The third-order derivatives can be further improved by taking sparsity into account and use mex versions for kronecker products etc. I leave this for further testing (and if anybody actually uses this ;-) )
2019-12-17 19:17:09 +01:00
else %model parameters
jind = j - offset ;
iind = i - offset ;
if i < = offset
y = dghu ( : , : , jind ) * dSigma_e ( : , : , i ) * KalmanB ' + KalmanB * dSigma_e ( : , : , i ) * dghu ( : , : , jind ) ' ;
% y(abs(y)<1.e-8)=0;
d2Om_tmp ( : , jcount ) = dyn_vech ( y ) ;
else
icount = icount + 1 ;
dKalmanAj = reshape ( x2x ( : , icount ) , [ M . endo_nbr M . endo_nbr ] ) ;
% x = get_2nd_deriv(x2x,m,m,iind,jind);%xx2(:,:,jcount);
elem1 = ( get_2nd_deriv ( d2GAM0 , M . endo_nbr , M . endo_nbr , iind , jind ) - get_2nd_deriv ( d2GAM1 , M . endo_nbr , M . endo_nbr , iind , jind ) * KalmanA ) ;
elem1 = elem1 - ( dGAM1 ( : , : , jind ) * dKalmanA ( : , : , iind ) + dGAM1 ( : , : , iind ) * dKalmanA ( : , : , jind ) ) ;
elemj0 = dGAM0 ( : , : , jind ) - dGAM1 ( : , : , jind ) * KalmanA - GAM1 * dKalmanA ( : , : , jind ) ;
elemi0 = dGAM0 ( : , : , iind ) - dGAM1 ( : , : , iind ) * KalmanA - GAM1 * dKalmanA ( : , : , iind ) ;
elem0 = elemj0 * dghu ( : , : , iind ) + elemi0 * dghu ( : , : , jind ) ;
y = invMM * ( get_2nd_deriv ( d2GAM3 , M . endo_nbr , M . exo_nbr , iind , jind ) - elem0 - ( elem1 - GAM1 * dKalmanAj ) * KalmanB ) ;
% d2B(:,:,j+length(indexo),i+length(indexo)) = y;
% d2B(:,:,i+length(indexo),j+length(indexo)) = y;
y = y * M . Sigma_e * KalmanB ' + KalmanB * M . Sigma_e * y ' + ...
dghu ( : , : , jind ) * M . Sigma_e * dghu ( : , : , iind ) ' + dghu ( : , : , iind ) * M . Sigma_e * dghu ( : , : , jind ) ' ;
% x(abs(x)<1.e-8)=0;
d2KalmanA_tmp ( : , jcount ) = vec ( dKalmanAj ) ;
% y(abs(y)<1.e-8)=0;
d2Om_tmp ( : , jcount ) = dyn_vech ( y ) ;
end
end
if jcount == ncol || i * j == totparam_nbr ^2
d2KalmanA ( : , cumjcount + 1 : cumjcount + jcount ) = d2KalmanA_tmp ( : , 1 : jcount ) ;
% d2KalmanA(:,:,j+length(indexo),i+length(indexo)) = x;
% d2KalmanA(:,:,i+length(indexo),j+length(indexo)) = x;
d2Om ( : , cumjcount + 1 : cumjcount + jcount ) = d2Om_tmp ( : , 1 : jcount ) ;
% d2Om(:,:,j+length(indexo),i+length(indexo)) = y;
% d2Om(:,:,i+length(indexo),j+length(indexo)) = y;
save ( [ tmpDir filesep ' d2KalmanA_' int2str ( cumjcount + 1 ) ' _' int2str ( cumjcount + jcount ) ' .mat' ] , ' d2KalmanA' )
save ( [ tmpDir filesep ' d2Om_' int2str ( cumjcount + 1 ) ' _' int2str ( cumjcount + jcount ) ' .mat' ] , ' d2Om' )
cumjcount = cumjcount + jcount ;
jcount = 0 ;
% d2KalmanA = sparse(m1*m1,tot_param_nbr*(tot_param_nbr+1)/2);
% d2Om = sparse(m1*(m1+1)/2,tot_param_nbr*(tot_param_nbr+1)/2);
d2KalmanA_tmp = zeros ( M . endo_nbr * M . endo_nbr , ncol ) ;
d2Om_tmp = zeros ( M . endo_nbr * ( M . endo_nbr + 1 ) / 2 , ncol ) ;
end
end
end
%Store into structure
DERIVS . d2Yss = d2Yss ;
DERIVS . d2KalmanA = d2KalmanA ;
DERIVS . d2Om = d2Om ;
end
return
%% AUXILIARY FUNCTIONS %%
%%%%%%%%%%%%%%%%%%%%%%%%%
function g22 = get_2nd_deriv ( gpp,m,n,i,j)
% inputs:
% - gpp: [#second_order_Jacobian_terms by 5] double Hessian matrix (wrt parameters) of a matrix
% rows: respective derivative term
% 1st column: equation number of the term appearing
% 2nd column: column number of variable in Jacobian
% 3rd column: number of the first parameter in derivative
% 4th column: number of the second parameter in derivative
% 5th column: value of the Hessian term
% - m: scalar number of equations
% - n: scalar number of variables
% - i: scalar number for which first parameter
% - j: scalar number for which second parameter
g22 = zeros ( m , n ) ;
is = find ( gpp ( : , 3 ) == i ) ;
is = is ( find ( gpp ( is , 4 ) == j ) ) ;
if ~ isempty ( is )
g22 ( sub2ind ( [ m , n ] , gpp ( is , 1 ) , gpp ( is , 2 ) ) ) = gpp ( is , 5 ) ' ;
end
return
function g22 = get_2nd_deriv_mat ( gpp,i,j,npar)
% inputs:
% - gpp: [#second_order_Jacobian_terms by 5] double Hessian matrix of (wrt parameters) of dynamic Jacobian
% rows: respective derivative term
% 1st column: equation number of the term appearing
% 2nd column: column number of variable in Jacobian of the dynamic model
% 3rd column: number of the first parameter in derivative
% 4th column: number of the second parameter in derivative
% 5th column: value of the Hessian term
% - i: scalar number for which model equation
% - j: scalar number for which variable in Jacobian of dynamic model
% - npar: scalar Number of model parameters, i.e. equals M.param_nbr
%
% output:
% g22: [npar by npar] Hessian matrix (wrt parameters) of Jacobian of dynamic model for equation i
% rows: first parameter in Hessian
% columns: second paramater in Hessian
g22 = zeros ( npar , npar ) ;
is = find ( gpp ( : , 1 ) == i ) ;
is = is ( find ( gpp ( is , 2 ) == j ) ) ;
if ~ isempty ( is )
g22 ( sub2ind ( [ npar , npar ] , gpp ( is , 3 ) , gpp ( is , 4 ) ) ) = gpp ( is , 5 ) ' ;
end
return
function r22 = get_all_resid_2nd_derivs ( rpp,m,npar)
% inputs:
% - rpp: [#second_order_residual_terms by 4] double Hessian matrix (wrt paramters) of model equations
% rows: respective derivative term
% 1st column: equation number of the term appearing
% 2nd column: number of the first parameter in derivative
% 3rd column: number of the second parameter in derivative
% 4th column: value of the Hessian term
% - m: scalar Number of residuals (or model equations), i.e. equals M.endo_nbr
% - npar: scalar Number of model parameters, i.e. equals M.param_nbr
%
% output:
% r22: [M.endo_nbr by M.param_nbr by M.param_nbr] Hessian matrix of model equations with respect to parameters
% rows: equations in order of declaration
% 1st columns: first parameter number in derivative
% 2nd columns: second parameter in derivative
r22 = zeros ( m , npar , npar ) ;
for is = 1 : size ( rpp , 1 )
% Keep symmetry in hessian, hence 2 and 3 as well as 3 and 2, i.e. d2f/(dp1 dp2) = d2f/(dp2 dp1)
r22 ( rpp ( is , 1 ) , rpp ( is , 2 ) , rpp ( is , 3 ) ) = rpp ( is , 4 ) ;
r22 ( rpp ( is , 1 ) , rpp ( is , 3 ) , rpp ( is , 2 ) ) = rpp ( is , 4 ) ;
end
return
function h2 = get_hess_deriv ( hp,i,j,m,npar)
% inputs:
% - hp: [#first_order_Hessian_terms by 5] double Jacobian matrix (wrt paramters) of dynamic Hessian
% rows: respective derivative term
% 1st column: equation number of the term appearing
% 2nd column: column number of first variable in Hessian of the dynamic model
% 3rd column: column number of second variable in Hessian of the dynamic model
% 4th column: number of the parameter in derivative
% 5th column: value of the Hessian term
% - i: scalar number for which model equation
% - j: scalar number for which first variable in Hessian of dynamic model variable
% - m: scalar Number of dynamic model variables + exogenous vars, i.e. yy0_nbr + M.exo_nbr
% - npar: scalar Number of model parameters, i.e. equals M.param_nbr
%
% output:
% h2: [(yy0_nbr + M.exo_nbr) by M.param_nbr] Jacobian matrix (wrt parameters) of dynamic Hessian
% rows: second dynamic or exogenous variables in Hessian of specific model equation of the dynamic model
% columns: parameters
h2 = zeros ( m , npar ) ;
is1 = find ( hp ( : , 1 ) == i ) ;
is = is1 ( find ( hp ( is1 , 2 ) == j ) ) ;
if ~ isempty ( is )
h2 ( sub2ind ( [ m , npar ] , hp ( is , 3 ) , hp ( is , 4 ) ) ) = hp ( is , 5 ) ' ;
end
return
