1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 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 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309
|
function [R,indef, E, P]=chol_SE(A,pivoting)
% [R,indef, E, P]=chol_SE(A,pivoting)
% Performs a (modified) Cholesky factorization of the form
%
% P'*A*P + E = R'*R
%
% As detailed in Schnabel/Eskow (1990), the factorization has 2 phases:
% Phase 1: While A can still be positive definite, pivot on the maximum diagonal element.
% Check that the standard Cholesky update would result in a positive diagonal
% at the current iteration. If so, continue with the normal Cholesky update.
% Otherwise switch to phase 2.
% If A is safely positive definite, stage 1 is never left, resulting in
% the standard Cholesky decomposition.
%
% Phase 2: Pivot on the minimum of the negatives of the lower Gershgorin bound
% estimates. To prevent negative diagonals, compute the amount to add to the
% pivot element and add this. Then, do the Cholesky update and update the estimates of the
% Gershgorin bounds.
%
% Notes:
% - During factorization, L=R' is stored in the lower triangle of the original matrix A,
% miminizing the memory requirements
% - Conforming with the original Schnabel/Eskow (1990) algorithm:
% - at each iteration the updated Gershgorin bounds are estimated instead of recomputed,
% reducing the computational requirements from o(n^3) to o (n^2)
% - For the last 2 by 2 submatrix, an eigenvalue-based decomposition is used
% - While pivoting is not necessary, it improves the size of E, the add-on on to the diagonal. But this comes at
% the cost of introduding a permutation.
%
%
% INPUTS
% - A [n*n] Matrix to be factorized
% - pivoting [scalar] dummy whether pivoting is used
%
% OUTPUTS
% - R [n*n] originally stored in lower triangular portion of A, including the main diagonal
%
% - E [n*1] Elements added to the diagonal of A
% - P [n*1] record of how the rows and columns of the matrix were permuted while
% performing the decomposition
%
% REFERENCES:
% This implementation is based on
%
% Robert B. Schnabel and Elizabeth Eskow. 1990. "A New Modified Cholesky
% Factorization," SIAM Journal of Scientific Statistical Computating,
% 11, 6: 1136-58.
%
% Elizabeth Eskow and Robert B. Schnabel 1991. "Algorithm 695 - Software for a New Modified Cholesky
% Factorization," ACM Transactions on Mathematical Software, Vol 17, No 3: 306-312
%
%
% Author: Johannes Pfeifer based on Eskow/Schnabel (1991)
% Copyright (C) 2015 Johannes Pfeifer
% Copyright (C) 2015-2017 Dynare Team
%
% 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/>.
if sum(sum(abs(A-A'))) > 0
error('A is not symmetric')
end
if nargin==1
pivoting=0;
end
n=size(A,1);
tau1=eps^(1/3); %tolerance parameter for determining when to switch phase 2
tau2=eps^(1/3); %tolerance used for determining the maximum condition number of the final 2X2 submatrix.
phase1 = 1;
delta = 0;
P=1:n;
g=zeros(n,1);
E=zeros(n,1);
% Find the maximum magnitude of the diagonal elements. If any diagonal element is negative, then phase1 is false.
gammma=max(diag(A));
if any(diag(A)) < 0
phase1 = 0;
end
taugam = tau1*gammma;
% If not in phase1, then calculate the initial Gershgorin bounds needed for the start of phase2.
if ~phase1
g=gersh_nested(A,1,n);
end
% check for n=1
if n==1
delta = tau2*abs(A(1,1)) - A(1,1);
if delta > 0
E(1) = delta;
end
if A(1,1) == 0
E(1) = tau2;
end
A(1,1)=sqrt(A(1,1)+E(1));
end
for j = 1:n-1
% PHASE 1
if phase1
if pivoting==1
% Find index of maximum diagonal element A(i,i) where i>=j
[tmp,imaxd] = max(diag(A(j:n,j:n)));
imaxd=imaxd+j-1;
% Pivot to the top the row and column with the max diag
if (imaxd ~= j)
% Swap row j with row of max diag
for ii = 1:j-1
temp = A(j,ii);
A(j,ii) = A(imaxd,ii);
A(imaxd,ii) = temp;
end
% Swap colj and row maxdiag between j and maxdiag
for ii = j+1:imaxd-1
temp = A(ii,j);
A(ii,j) = A(imaxd,ii);
A(imaxd,ii) = temp;
end
% Swap column j with column of max diag
for ii = imaxd+1:n
temp = A(ii,j);
A(ii,j) = A(ii,imaxd);
A(ii,imaxd) = temp;
end
% Swap diag elements
temp = A(j,j);
A(j,j) = A(imaxd,imaxd);
A(imaxd,imaxd) = temp;
% Swap elements of the permutation vector
itemp = P(j);
P(j) = P(imaxd);
P(imaxd) = itemp;
end
end
% check to see whether the normal Cholesky update for this
% iteration would result in a positive diagonal,
% and if not then switch to phase 2.
jp1 = j+1;
if A(j,j)>0
if min((diag(A(j+1:n,j+1:n)) - A(j+1:n,j).^2/A(j,j))) < taugam %test whether stage 2 is required
phase1=0;
end
else
phase1 = 0;
end
if phase1
% Do the normal cholesky update if still in phase 1
A(j,j) = sqrt(A(j,j));
tempjj = A(j,j);
for ii = jp1: n
A(ii,j) = A(ii,j)/tempjj;
end
for ii=jp1:n
temp=A(ii,j);
for k = jp1:ii
A(ii,k) = A(ii,k)-(temp * A(k,j));
end
end
if j == n-1
A(n,n)=sqrt(A(n,n));
end
else
% Calculate the negatives of the lower Gershgorin bounds
g=gersh_nested(A,j,n);
end
end
% PHASE 2
if ~phase1
if j ~= n-1
if pivoting
% Find the minimum negative Gershgorin bound
[tmp,iming] = min(g(j:n));
iming=iming+j-1;
% Pivot to the top the row and column with the
% minimum negative Gershgorin bound
if iming ~= j
% Swap row j with row of min Gershgorin bound
for ii = 1:j-1
temp = A(j,ii);
A(j,ii) = A(iming,ii);
A(iming,ii) = temp;
end
% Swap col j with row iming from j to iming
for ii = j+1:iming-1
temp = A(ii,j);
A(ii,j) = A(iming,ii);
A(iming,ii) = temp;
end
% Swap column j with column of min Gershgorin bound
for ii = iming+1:n
temp = A(ii,j);
A(ii,j) = A(ii,iming);
A(ii,iming) = temp;
end
% Swap diagonal elements
temp = A(j,j);
A(j,j) = A(iming,iming);
A(iming,iming) = temp;
% Swap elements of the permutation vector
itemp = P(j);
P(j) = P(iming);
P(iming) = itemp;
% Swap elements of the negative Gershgorin bounds vector
temp = g(j);
g(j) = g(iming);
g(iming) = temp;
end
end
% Calculate delta and add to the diagonal. delta=max{0,-A(j,j) + max{normj,taugam},delta_previous}
% where normj=sum of |A(i,j)|,for i=1,n, delta_previous is the delta computed at the previous iter and taugam is tau1*gamma.
normj=sum(abs(A(j+1:n,j)));
delta = max([0;delta;-A(j,j)+normj;-A(j,j)+taugam]); % get adjustment based on formula on bottom of p. 309 of Eskow/Schnabel (1991)
E(j) = delta;
A(j,j) = A(j,j) + E(j);
% Update the Gershgorin bound estimates (note: g(i) is the negative of the Gershgorin lower bound.)
if A(j,j) ~= normj
temp = (normj/A(j,j)) - 1;
for ii = j+1:n
g(ii) = g(ii) + abs(A(ii,j)) * temp;
end
end
% Do the cholesky update
A(j,j) = sqrt(A(j,j));
tempjj = A(j,j);
for ii = j+1:n
A(ii,j) = A(ii,j) / tempjj;
end
for ii = j+1:n
temp = A(ii,j);
for k = j+1: ii
A(ii,k) = A(ii,k) - (temp * A(k,j));
end
end
else
% Find eigenvalues of final 2 by 2 submatrix
% Find delta such that:
% 1. the l2 condition number of the final 2X2 submatrix + delta*I <= tau2
% 2. delta >= previous delta,
% 3. min(eigvals) + delta >= tau2 * gamma, where min(eigvals) is the smallest eigenvalue of the final 2X2 submatrix
A(n-1,n)=A(n,n-1); %set value above diagonal for computation of eigenvalues
eigvals = eig(A(n-1:n,n-1:n));
delta = max([ 0 ; delta ; -min(eigvals)+tau2*max((max(eigvals)-min(eigvals))/(1-tau1),gammma) ]); %Formula 5.3.2 of Schnabel/Eskow (1990)
if delta > 0
A(n-1,n-1) = A(n-1,n-1) + delta;
A(n,n) = A(n,n) + delta;
E(n-1) = delta;
E(n) = delta;
end
% Final update
A(n-1,n-1) = sqrt(A(n-1,n-1));
A(n,n-1) = A(n,n-1)/A(n-1,n-1);
A(n,n) = A(n,n) - (A(n,n-1)^2);
A(n,n) = sqrt(A(n,n));
end
end
end
R=(tril(A))';
indef=~phase1;
Pprod=zeros(n,n);
Pprod(sub2ind([n,n],P,1:n))=1;
P=Pprod;
end
function g=gersh_nested(A,j,n)
g=zeros(n,1);
for ii = j:n
if ii == 1
sum_up_to_i = 0;
else
sum_up_to_i = sum(abs(A(ii,j:(ii-1))));
end
if ii == n
sum_after_i = 0;
else
sum_after_i = sum(abs(A((ii+1):n,ii)));
end
g(ii) = sum_up_to_i + sum_after_i- A(ii,ii);
end
end
|