我正在尝试编写一个简单的奇异值分解(SVD)实现。我使用单边Jacobi算法,因为它看起来很简单。该算法在here中进行了描述,该here有一个简单的Matlab代码(练习4)。我已经实现了相同的代码,它工作得很好,我的意思是SVD(A) =U*S* V‘(或使用的任何其他表示法),对于某些矩阵,结果与Matlab的SVD产生的结果相同(除了这个函数不对奇异值进行排序)。但是我的问题是,当矩阵A有一个0的奇异值时,U和V不再是它们应该是的么正矩阵!
有没有办法更新这个算法,让它也适用于奇异值为0的情况?如果没有,有没有另一种同样容易实现的SVD算法?任何帮助都是非常感谢的。
这是我的Matlab代码,它与上面链接中的代码基本相同,只是完成了一些微小的更改。
function [U,S,V] = jacobi_SVD(A)
TOL=1.e-4;
n=size(A,1);
U=A';
V=eye(n);
converge=TOL+1;
while converge>TOL
converge=0;
for i=1:n-1
for j=i+1:n
% compute [alpha gamma;gamma beta]=(i,j) submatrix of U*U'
alpha=sumsqr(U(i, :));
beta=sumsqr(U(j, :));
gamma=sum(U(i, :).* U(j, :));
converge=max(converge,abs(gamma)/sqrt(alpha*beta));
% compute Jacobi rotation that diagonalizes
% [alpha gamma;gamma beta]
zeta=(beta-alpha)/(2*gamma);
t=sign(zeta)/(abs(zeta)+sqrt(1+zeta^2));
c= 1.0 / (sqrt(1 + t * t));
s= c * t;
% update columns i and j of U
t=U(i, :);
U(i, :)=c*t-s*U(j, :);
U(j, :)=s*t+c*U(j, :);
% update matrix V of right singular vectors
t=V(i, :);
V(i, :)=c*t-s*V(j, :);
V(j, :)=s*t+c*V(j, :);
end
end
end
% the singular values are the norms of the columns of U
% the left singular vectors are the normalized columns of U
for j=1:n
singvals(j)=norm(U(j, :));
U(j, :)=U(j, :)/singvals(j);
end
S=diag(singvals);
U = U';
V = V'; %return V, not V'
end发布于 2018-06-08 00:18:25
我不能让你的代码运行,因为当你评估alpha和beta时,你得到了你还没有定义的sumsq。我在Matlab website.上找到了一些使用QR分解(格拉姆-施密特)的简单代码。
function [u,s,v] = svdsim(a,tol)
%SVDSIM simple SVD program
%
% A simple program that demonstrates how to use the
% QR decomposition to perform the SVD of a matrix.
% A may be rectangular and complex.
%
% usage: [U,S,V]= SVDSIM(A)
% or S = SVDSIM(A)
%
% with A = U*S*V' , S>=0 , U'*U = Iu , and V'*V = Iv
%
% The idea is to use the QR decomposition on A to gradually "pull" U out from
% the left and then use QR on A transposed to "pull" V out from the right.
% This process makes A lower triangular and then upper triangular alternately.
% Eventually, A becomes both upper and lower triangular at the same time,
% (i.e. Diagonal) with the singular values on the diagonal.
%
% Matlab's own SVD routine should always be the first choice to use,
% but this routine provides a simple "algorithmic alternative"
% depending on the users' needs.
%
%see also: SVD, EIG, QR, BIDIAG, HESS
%
% Paul Godfrey
% October 23, 2006
if ~exist('tol','var')
tol=eps*1024;
end
%reserve space in advance
sizea=size(a);
loopmax=100*max(sizea);
loopcount=0;
% or use Bidiag(A) to initialize U, S, and V
u=eye(sizea(1));
s=a';
v=eye(sizea(2));
Err=realmax;
while Err>tol & loopcount<loopmax ;
% log10([Err tol loopcount loopmax]); pause
[q,s]=qr(s'); u=u*q;
[q,s]=qr(s'); v=v*q;
% exit when we get "close"
e=triu(s,1);
E=norm(e(:));
F=norm(diag(s));
if F==0, F=1;end
Err=E/F;
loopcount=loopcount+1;
end
% [Err/tol loopcount/loopmax]
%fix the signs in S
ss=diag(s);
s=zeros(sizea);
for n=1:length(ss)
ssn=ss(n);
s(n,n)=abs(ssn);
if ssn<0
u(:,n)=-u(:,n);
end
end
if nargout<=1
u=diag(s);
end
return在我的经验中,通常情况下你实际上没有零。数值精度给你留下了大致接近的东西,例如,下面的内容。如果我想创建一个5 x 5的矩阵,但它的秩是3,我可以这样做。
A = randn(5,3)*randn(5,3);
[U,S,Vt] = svdsim(A,1e-8);
S =
6.3812 0 0 0 0
0 2.0027 0 0 0
0 0 1.0240 0 0
0 0 0 0.0000 0
0 0 0 0 0.0000现在它看起来就像是你有零。但如果你仔细观察。
format long
>> S(4,4)
ans =
3.418057860623250e-16
S(5,5)
ans =
9.725444388260210e-17我要注意的是,这是机器epsilon,对于所有密集的目的,它是0。
https://stackoverflow.com/questions/50745447
复制相似问题