dynare/dynare++/kord/global_check.hweb

@q $Id: global_check.hweb 431 2005-08-16 15:41:01Z kamenik $ @>
@q Copyright 2005, Ondra Kamenik @>

@*2 Global check. Start of {\tt global\_check.h} file.

The purpose of this file is to provide classes for checking error of
approximation. If $y_t=g(y^*_{t-1},u)$ is an approximate solution,
then we check for the error of residuals of the system equations. Let
$F(y^*,u,u')=f(g^{**}(g^*(y^*,u'),u),g(y^*,u),y^*,u)$, then we
calculate integral
$$E[F(y^*,u,u')]$$@q'@>
which we want to be zero for all $y^*$, and $u$.

There are a few possibilities how and where the integral is
evaluated. Currently we offer the following:

\numberedlist
\li Along shocks. The $y^*$ is set to steady state, and $u$ is set to
zero but one element is going from minus through plus shocks in few
steps. The user gives the scaling factor, for instance interval
$\langle<-3\sigma,3\sigma\rangle$ (where $sigma$ is a standard error
of the shock), and a number of steps. This is repeated for each shock
(element of the $u$ vector).
\li Along simulation. Some random simulation is run, and for each
realization of $y^*$ and $u$ along the path we evaluate the residual.
\li On ellipse. Let $V=AA^T$ be a covariance matrix of the
predetermined variables $y^*$ based on linear approximation, then we
calculate integral for points on the ellipse $\{Ax\vert\, \Vert
x\Vert_2=1\}$. The points are selected by means of low discrepancy
method and polar transformation. The shock $u$ are zeros.

\li Unconditional distribution. 

\endnumberedlist


@s ResidFunction int
@s GResidFunction int
@s GlobalChecker int
@s VectorFunction int
@s ResidFunctionSig int
@s GaussHermite int
@s SmolyakQuadrature int
@s ProductQuadrature int
@s ParameterSignal int
@s Quadrature int
@s QMCarloCubeQuadrature int

@c
#ifndef GLOBAL_CHECK_H
#define GLOBAL_CHECK_H

#include <matio.h>

#include "vector_function.h"
#include "quadrature.h"

#include "dynamic_model.h"
#include "journal.h"
#include "approximation.h"

@<|ResidFunction| class declaration@>;
@<|GResidFunction| class declaration@>;
@<|GlobalChecker| class declaration@>;
@<|ResidFunctionSig| class declaration@>;

#endif

@ This is a class for implementing |VectorFunction| interface
evaluating the residual of equations, this is
$$F(y^*,u,u')=f(g^{**}(g^*(y^*,u),u'),y^*,u)$$
is written as a function of $u'$.

When the object is constructed, one has to specify $(y^*,u)$, this is
done by |setYU| method. The object has basically two states. One is
after construction and before call to |setYU|. The second is after
call |setYU|. We distinguish between the two states, an object in the
second state contains |yplus|, |ystar|, |u|, and |hss|.

The vector |yplus| is $g^*(y^*,u)$. |ystar| is $y^*$, and polynomial
|hss| is partially evaluated $g^**(yplus, u)$.

The pointer to |DynamicModel| is important, since the |DynamicModel|
evaluates the function $f$. When copying the object, we have to make
also a copy of |DynamicModel|.

@<|ResidFunction| class declaration@>=
class ResidFunction : public VectorFunction {
protected:@;
	const Approximation& approx;
	DynamicModel* model;
	Vector* yplus;
	Vector* ystar;
	Vector* u;
	FTensorPolynomial* hss;
public:@;
	ResidFunction(const Approximation& app);
	ResidFunction(const ResidFunction& rf);
	virtual ~ResidFunction();
	virtual VectorFunction* clone() const
		{@+ return new ResidFunction(*this);@+}
	virtual void eval(const Vector& point, const ParameterSignal& sig, Vector& out);
	void setYU(const Vector& ys, const Vector& xx);
};

@ This is a |ResidFunction| wrapped with |GaussConverterFunction|.

@<|GResidFunction| class declaration@>=
class GResidFunction : public GaussConverterFunction {
public:@;
	GResidFunction(const Approximation& app)
		: GaussConverterFunction(new ResidFunction(app), app.getModel().getVcov())@+ {}
	GResidFunction(const GResidFunction& rf)
		: GaussConverterFunction(rf)@+ {}
	virtual ~GResidFunction()@+ {}
	virtual VectorFunction* clone() const
		{@+ return new GResidFunction(*this);@+}
	void setYU(const Vector& ys, const Vector& xx)
		{@+ ((ResidFunction*)func)->setYU(ys, xx);}
};


@ This is a class encapsulating checking algorithms. Its core routine
is |check|, which calculates integral $E[F(y^*,u,u')\vert y^*,u]$ for
given realizations of $y^*$ and $u$. The both are given in
matrices. The methods checking along shocks, on ellipse and anlong a
simulation path, just fill the matrices and call the core |check|.

The method |checkUnconditionalAndSave| evaluates unconditional
$E[F(y,u,u')]$.

The object also maintains a set of |GResidFunction| functions |vfs| in
order to save (possibly expensive) copying of |DynamicModel|s.

@<|GlobalChecker| class declaration@>=
class GlobalChecker {
	const Approximation& approx;
	const DynamicModel& model;
	Journal& journal;
	GResidFunction rf;
	VectorFunctionSet vfs;
public:@;
	GlobalChecker(const Approximation& app, int n, Journal& jr)
		: approx(app), model(approx.getModel()), journal(jr),
		  rf(approx), vfs(rf, n)@+ {}
	void check(int max_evals, const ConstTwoDMatrix& y,
			   const ConstTwoDMatrix& x, TwoDMatrix& out);
	void checkAlongShocksAndSave(mat_t* fd, const char* prefix,
								 int m, double mult, int max_evals);
	void checkOnEllipseAndSave(mat_t* fd, const char* prefix,
							   int m, double mult, int max_evals);
	void checkAlongSimulationAndSave(mat_t* fd, const char* prefix,
									 int m, int max_evals);
	void checkUnconditionalAndSave(mat_t* fd, const char* prefix,
								   int m, int max_evals);
protected:@;
	void check(const Quadrature& quad, int level,
			   const ConstVector& y, const ConstVector& x, Vector& out);
};


@ Signalled resid function. Not implemented yet. todo:
@<|ResidFunctionSig| class declaration@>=
class ResidFunctionSig : public ResidFunction {
public:@;
	ResidFunctionSig(const Approximation& app, const Vector& ys, const Vector& xx);
};

@ End of {\tt global\_check.h} file.