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
|
classdef factorization_svd < factorization
%FACTORIZATION_SVD A = U*S*V'
% Adds the following extra methods that act just like the builtin functions.
% Most take little or no time to compute, since they rely on the precomputed
% SVD. The exceptions are cond(F,p) and norm(F,p) when p is not 2.
%
% c = cond (F,p) the p-norm condition number. p=2 is the default.
% cond(F,2) takes no time to compute, since it was
% computed when the SVD factorization was found.
% a = norm (F,p) the p-norm. see the cond(F,p) discussion above.
% r = rank (F) returns the rank of A, precomputed by the SVD.
% Z = null (F) orthonormal basis for the null space of A
% Q = orth (F) orthonormal basis for the range of A
% C = pinv (F) the pseudo-inverse, V'*(S\V).
% [U,S,V] = svd (F) SVD of A or pinv(A), regular, economy, or rank-sized
% Copyright 2011-2012, Timothy A. Davis, http://www.suitesparse.com
methods
function F = factorization_svd (A)
%FACTORIZATION_SVD singular value decomponsition, A = U*S*V'
[f.U, f.S, f.V] = svd (full (A)) ;
[m, n] = size (A) ;
% convert S into a vector of singular values
if (isempty (A))
f.S = 0 ;
elseif (isvector (A))
f.S = f.S (1) ;
else
f.S = diag (f.S) ;
end
% compute rank(A), and save it in F
f.r = sum (f.S > max (m,n) * eps (f.S (1))) ;
F.A_rank = f.r ;
% compute cond(A), and save it in F
if (isempty (A))
F.A_cond = 0 ; % cond ([]) is zero, according to MATLAB
elseif (f.r < min (m,n))
F.A_cond = inf ; % matrix is singular
else
F.A_cond = f.S (1) / f.S (end) ;
end
F.A = A ;
F.Factors = f ;
F.kind = 'singular value decomposition: A = U*S*V''' ;
end
function e = error_check (F)
%ERROR_CHECK : return relative 1-norm of error in factorization
% meant for testing only
[U, S, V] = svd (F) ; % extracts the pre-computed SVD of A from F
e = norm (F.A - U*S*V', 1) / norm (F.A, 1) ;
end
function x = mldivide_subclass (F,b)
%MLDIVIDE_SUBCLASS x=A\b using a singular value decomposition
% Only svd(A,'econ') is needed.
f = F.Factors ;
r = f.r ;
x = f.V (:,1:r) * (diag (f.S (1:r)) \ (f.U (:,1:r)' * b)) ;
end
function x = mrdivide_subclass (b,F)
%MRDIVIDE_SUBCLASS x=b/A using a singular value decomposition
% Only svd(A,'econ') is needed.
f = F.Factors ;
r = f.r ;
x = ((b * f.V (:,1:r)) / diag (f.S (1:r))) * f.U (:,1:r)' ;
end
function c = cond (F, p)
%COND the 2-norm condition number
% cond(F,2) takes O(1) time to compute once the SVD is known.
% Otherwise, pinv(A) (or pinv(A') if F has been transposed) is
% explicitly computed using the pre-computed [U,S,V]=svd(A).
if (nargin == 1 || isequal (p,2) || isempty (F))
% The 2-norm condition number has been pre-computed.
c = F.A_cond ;
else
% Compute the p-norm of a non-empty matrix, where p is not 2.
[m, n] = size (F) ;
if (m ~= n)
error ('MATLAB:cond:normMismatchSizeA', ...
'A is rectangular. Use the 2 norm.') ;
end
r = F.A_rank ;
if (r < min (m,n))
% matrix is rank-deficient so cond (A,p) is always inf
c = inf ;
else
% The matrix is square, non-empty, and has full rank.
% One of these requires the explicit computation of pinv(A),
% where U,S,V are already pre-computed. The same result is
% computed whether F represents A or its (pseudo) inverse.
c = norm (double (F), p) * norm (double (inverse (F)), p) ;
end
end
end
function nrm = norm (F, p)
%NORM see the description of cond, above.
if (nargin == 1 || isequal (p,2) || isempty (F))
f = F.Factors ;
r = f.r ;
if (isempty (F))
nrm = 0 ;
elseif (r == 0)
if (F.is_inverse)
nrm = inf ;
else
nrm = 0 ;
end
else
if (F.is_inverse)
nrm = 1 / (f.S (r)) ; % norm (pinv (A))
else
nrm = f.S (1) ; % norm (A)
end
end
else
% If F represents the inverse, then double (F) is pinv (A),
% which is computed with V*(S\U') via mldivide. U,S,V are
% already computed, so double(F) is not too hard to compute.
nrm = norm (double (F), p) ;
end
end
function r = rank (F)
% The rank of A has been pre-computed. Just return it.
r = F.A_rank ;
end
function Z = null (F)
%NULL orthonormal basis for the null space of A
f = F.Factors ;
r = f.r ;
Z = f.V (:, r+1:end) ;
end
function Q = orth (F)
%ORTH orthonormal basis for the range of A
% This function makes theta=subspace(A,B) easy to compute,
%
f = F.Factors ;
r = f.r ;
Q = f.U (:, 1:r) ;
end
function C = pinv (F)
% PINV is just another name for inverse(factorize(A,'svd'))
C = inverse (F) ;
end
function [U,S,V] = svd (F, kind)
% SVD return the svd of A, A', pinv(A), or pinv(A'). [U,S,V]=svd(A)
% has already been computed. Truncate / transpose / reshape it
% as needed, also considering svd(",'econ') and svd(",0).
f = F.Factors ;
U = f.U ;
S = f.S ;
V = f.V ;
[m, n] = size (F.A) ;
if (nargin > 1)
switch kind
case 'econ'
% return svd(A,'econ')
k = min (m,n) ;
U = U (:, 1:k) ;
S = S (1:k) ;
V = V (:, 1:k) ;
case 'rank'
% return the rank-sized SVD
k = f.r ;
U = U (:, 1:k) ;
S = S (1:k) ;
V = V (:, 1:k) ;
case 0
% return svd(A,0)
if (m > n)
k = n ;
U = U (:, 1:k) ;
S = S (1:k) ;
end
otherwise
error ('unrecognized kind') ;
end
end
if (F.is_inverse ~= F.is_ctrans)
% returning svd(A') or svd(pinv(A)). swap U and V
T = U ;
U = V ;
V = T ;
end
if (F.is_inverse)
% returning svd(pinv(A)) or svd(pinv(A')). Invert and reverse
% S(1:r) and the corresponding singular vectors.
r = f.r ;
S = [(1 ./ S (r:-1:1)) ; (zeros (length (S) - r, 1))] ;
U (:, 1:r) = U (:, r:-1:1) ;
V (:, 1:r) = V (:, r:-1:1) ;
end
% The user expects S as a matrix of the proper size, so expand it.
k = length (S) ;
if (isempty (F))
k = 0 ;
end
S = full (sparse (1:k, 1:k, S, size (U,2), size (V,2))) ;
end
end
end
|
