1) allow to compute derivatives starting from NUMERICAL derivatives of jacobian and steady state: this has a minor cost in accuracy and allow apply without errors identification and estimation with numerical derivatives;
2) added trap in dynare_estimation_init: if steadystate changes param values, automaticly shifts to numerical derivs of jacoban and steady state + analytic derivatives of all the rest;
3) bug fixes for 2nd order derivatives w.r.t. model parameters;
If options_.prior_trunc is set to zero (the default is strictly positive) then prior_correction is infinite because the prior density is zero (this is not true for the uniform prior)... This does not help the optimizer. Even if we do not fall in this case (because options_.prior_trunc>0 or becuase only uniform priors are used for the bounded parameters) the meaning of this correction is unclear.
This can improve a bit optimization routines when parameter go beyond prior bounds during line search algorithms or when numerical gradient is computed.
- Adds the tolerance criteria for the iterative solvers (sylvester_fixed_point_tol, lyapunov_fixed_point_tol and lyapunov_doubling_tol)
- Updates the reference manual
penalty in dsge_likelihood.m and dsge_likelihood_hh.m Removed misleading
initialization code. Added call to dsge_likelihood_hh in
initial_estimation_checks to initialize persistent variable in that
function as well.