gaussian_filter.m 6.53 KB
Newer Older
1
function [LIK,lik] = gaussian_filter(ReducedForm, Y, start, ParticleOptions, ThreadsOptions)
2
3
4
5
% Evaluates the likelihood of a non-linear model approximating the
% predictive (prior) and filtered (posterior) densities for state variables
% by gaussian distributions.
% Gaussian approximation is done by:
6
7
% - a spherical-radial cubature (ref: Arasaratnam & Haykin, 2009).
% - a scaled unscented transform cubature (ref: Julier & Uhlmann 1995)
8
9
10
11
% - Monte-Carlo draws from a multivariate gaussian distribution.
% First and second moments of prior and posterior state densities are computed
% from the resulting nodes/particles and allows to generate new distributions at the
% following observation.
12
13
% Pros: The use of nodes is much faster than Monte-Carlo Gaussian particle and standard particles
% filters since it treats a lesser number of particles. Furthermore, in all cases, there is no need
14
% of resampling.
15
% Cons: estimations may be biaised if the model is truly non-gaussian
16
17
18
% since predictive and filtered densities are unimodal.
%
% INPUTS
19
20
21
22
23
%    Reduced_Form     [structure] Matlab's structure describing the reduced form model.
%    Y                [double]    matrix of original observed variables.
%    start            [double]    structural parameters.
%    ParticleOptions  [structure] Matlab's structure describing options concerning particle filtering.
%    ThreadsOptions   [structure] Matlab's structure.
24
25
26
27
28
29
30
31
32
%
% OUTPUTS
%    LIK        [double]    scalar, likelihood
%    lik        [double]    vector, density of observations in each period.
%
% REFERENCES
%
% NOTES
%   The vector "lik" is used to evaluate the jacobian of the likelihood.
Stéphane Adjemian's avatar
Stéphane Adjemian committed
33
% Copyright (C) 2009-2017 Dynare Team
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
%
% 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/>.

persistent init_flag mf0 mf1
persistent nodes2 weights2 weights_c2 number_of_particles
persistent sample_size number_of_state_variables number_of_observed_variables

% Set default
if isempty(start)
    start = 1;
end

% Set persistent variables.
if isempty(init_flag)
    mf0 = ReducedForm.mf0;
    mf1 = ReducedForm.mf1;
    sample_size = size(Y,2);
    number_of_state_variables = length(mf0);
    number_of_observed_variables = length(mf1);
66
    number_of_particles = ParticleOptions.number_of_particles;
67
68
69
70
71
    init_flag = 1;
end

% compute gaussian quadrature nodes and weights on states and shocks
if isempty(nodes2)
72
    if ParticleOptions.distribution_approximation.cubature
73
74
        [nodes2,weights2] = spherical_radial_sigma_points(number_of_state_variables);
        weights_c2 = weights2;
75
76
    elseif ParticleOptions.distribution_approximation.unscented
        [nodes2,weights2,weights_c2] = unscented_sigma_points(number_of_state_variables,ParticleOptions);
77
    else
78
        if ~ParticleOptions.distribution_approximation.montecarlo
79
80
81
82
83
            error('Estimation: This approximation for the proposal is not implemented or unknown!')
        end
    end
end

84
if ParticleOptions.distribution_approximation.montecarlo
85
86
87
88
89
90
91
92
93
94
    set_dynare_seed('default');
end

% Get covariance matrices
Q = ReducedForm.Q;
H = ReducedForm.H;
if isempty(H)
    H = 0;
    H_lower_triangular_cholesky = 0;
else
95
    H_lower_triangular_cholesky = reduced_rank_cholesky(H)';
96
97
98
99
end

% Get initial condition for the state vector.
StateVectorMean = ReducedForm.StateVectorMean;
100
StateVectorVarianceSquareRoot = reduced_rank_cholesky(ReducedForm.StateVectorVariance)';
101
state_variance_rank = size(StateVectorVarianceSquareRoot,2);
102
Q_lower_triangular_cholesky = reduced_rank_cholesky(Q)';
103
104
105
106
107
108
109
110

% Initialization of the likelihood.
const_lik = (2*pi)^(number_of_observed_variables/2) ;
lik  = NaN(sample_size,1);
LIK  = NaN;

for t=1:sample_size
    [PredictedStateMean,PredictedStateVarianceSquareRoot,StateVectorMean,StateVectorVarianceSquareRoot] = ...
111
112
        gaussian_filter_bank(ReducedForm,Y(:,t),StateVectorMean,StateVectorVarianceSquareRoot,Q_lower_triangular_cholesky,H_lower_triangular_cholesky,H,ParticleOptions,ThreadsOptions) ;
    if ParticleOptions.distribution_approximation.cubature || ParticleOptions.distribution_approximation.unscented
113
114
        StateParticles = bsxfun(@plus,StateVectorMean,StateVectorVarianceSquareRoot*nodes2') ;
        IncrementalWeights = ...
Stéphane Adjemian's avatar
Stéphane Adjemian committed
115
116
117
118
            gaussian_densities(Y(:,t),StateVectorMean,...
                               StateVectorVarianceSquareRoot,PredictedStateMean,...
                               PredictedStateVarianceSquareRoot,StateParticles,H,const_lik,...
                               weights2,weights_c2,ReducedForm,ThreadsOptions) ;
119
        SampleWeights = weights2.*IncrementalWeights ;
Stéphane Adjemian's avatar
Stéphane Adjemian committed
120
    else
121
122
        StateParticles = bsxfun(@plus,StateVectorVarianceSquareRoot*randn(state_variance_rank,number_of_particles),StateVectorMean) ;
        IncrementalWeights = ...
Stéphane Adjemian's avatar
Stéphane Adjemian committed
123
124
125
126
            gaussian_densities(Y(:,t),StateVectorMean,...
                               StateVectorVarianceSquareRoot,PredictedStateMean,...
                               PredictedStateVarianceSquareRoot,StateParticles,H,const_lik,...
                               1/number_of_particles,1/number_of_particles,ReducedForm,ThreadsOptions) ;
127
        SampleWeights = IncrementalWeights/number_of_particles ;
128
    end
129
    SampleWeights = SampleWeights + 1e-6*ones(size(SampleWeights,1),1) ;
130
131
132
    SumSampleWeights = sum(SampleWeights) ;
    lik(t) = log(SumSampleWeights) ;
    SampleWeights = SampleWeights./SumSampleWeights ;
133
134
    if not(ParticleOptions.distribution_approximation.cubature || ParticleOptions.distribution_approximation.unscented)
        if (ParticleOptions.resampling.status.generic && neff(SampleWeights)<ParticleOptions.resampling.threshold*sample_size) || ParticleOptions.resampling.status.systematic
Stéphane Adjemian's avatar
Stéphane Adjemian committed
135
            StateParticles = resample(StateParticles',SampleWeights,ParticleOptions)' ;
136
            SampleWeights = ones(number_of_particles,1)/number_of_particles;
Stéphane Adjemian's avatar
Stéphane Adjemian committed
137
        end
138
    end
139
140
    StateVectorMean = StateParticles*SampleWeights ;
    temp = bsxfun(@minus,StateParticles,StateVectorMean) ;
141
    StateVectorVarianceSquareRoot = reduced_rank_cholesky( bsxfun(@times,SampleWeights',temp)*temp' )';
142
143
144
end

LIK = -sum(lik(start:end));