Default rank revealing tolerance

11 vues (au cours des 30 derniers jours)

Afficher commentaires plus anciens

Bruno Luong le 16 Fév 2024

1
Lien

Utiliser le lien direct vers cette question

https://fr.mathworks.com/matlabcentral/answers/2082808-default-rank-revealing-tolerance

Commenté : Paul le 17 Fév 2024

Ouvrir dans MATLAB Online

In few MATLAB functions such as RANK, PINV, ORTH there is a parameter TOL.

MATLAB uses the default value as

% tol = max(size(A)) * eps(norm(A))

where A the matrix under study.

I have two quetions:

First question: what is the theoretical basis of such formula? Especially why the max(size(A)) factor?

In my simple test here, it seems the error or the singular value grows like a square root of the dimension m. Here is the script to create a matrix A of size m x 4 of rank 3. I use SVD to compute the singulars values, s(4) ideally should be 0, s1(1:3) about unity, but I see experimentally s(4) grow like sqrt(m) as showed in this script:

m0 = 1e5;
s4 = [];
m = [];
while m0 < 1e8
    [Q, ~] =qr(rand(m0,3),0);
    X = rand(3,4);
    A = Q * X;
    [~,s,~] = svd(A,0,'vector');
    s4(end+1) = s(4);
    m(end+1) = m0;
    m0 = ceil(m0*1.1);
end
P = polyfit(log(m),log(s4),1);
p = P(1);
c=exp(P(2));
figure
loglog(m, s4);
hold on
h = loglog(m, c*m.^p);
xlabel('m')
ylabel('s(4)')
legend(h, sprintf('%g*m^{%g}', c, p))

PS: There is a floor noise about 1e-16 for m < 1e5, that could affect the fit, so I take the m value above 1e5 up to 1e8 .

Intuitively the square-root relationship seems reasonable to me since we could consider truncation errors in each component are somewhat random and independent.

Second question. The function lsqminnorm also have a default tol but is it NOT specified beside telling it computed from QR decomposition. It describes how the rank k is used to approximate the solution, but I don't see how the tol value is defined to estimate in turn the rank k. I would expect something like

% tol = eps(abs(R(1,1))) * max(size(A));

where R is the triangular matrix returned by qr, meaning abs(R(1,1)) = max(vecnorm(A)).

Can someone (TMW staff) shed a light and give the method to determine the default tol value implemented within lsqminnorm?

8 commentaires
Afficher 6 commentaires plus anciensMasquer 6 commentaires plus anciens

Bruno Luong le 17 Fév 2024

Modifié(e) : Bruno Luong le 17 Fév 2024

https://www.netlib.org/lapack/lug/node97.html eqt UERRBD(i)

https://core.ac.uk/download/pdf/216041984.pdf Lemme 4.2, for definition of modestly growing

Paul le 17 Fév 2024

Thanks for the links. I checked the LINPACK User's Guide from siam.org and I didn't see the TMW formula in Chapter 11, Section 1 (at least not explicitly).

Connectez-vous pour commenter.

Connectez-vous pour répondre à cette question.