How to quickly do Cholesky factorization for many small matrices?

4 vues (au cours des 30 derniers jours)
Stephan Orzada
Stephan Orzada le 23 Juil 2021
Commenté : Stephan Orzada le 1 Août 2021
In my project, I have to check many small matrices for positive-semidefiniteness (PSDness). I use chol() since it is much faster than using eig() to check for negative eigenvalues. Still it is very slow and the most time consuming part of my calculations, since I have to check 10e6 32x32 matrices over and over again. I already tried to use parfor and batch functions but it doesn't work (I also read this in the forums). However, it occured to me that the processor utilization is very low, so I checked what happens when I run the code on two instances of matlab on the same computer. The processor utilization doubled and each script finished in the same time it would have taken if they had been running alone. So it is possible to let chol() run in parallel on the same system. Any ideas how to do this in the same instance of matlab without incurring the problems you get with chol() in a parfor loop?
  12 commentaires
Bruno Luong
Bruno Luong le 31 Juil 2021
Modifié(e) : Bruno Luong le 31 Juil 2021
In real life 0 eigen value does not exist exactly. CHOL, EIG, EIGS all might fail due to round off.
And Stephan Orzada end up programming his own Cholesky decomposition for his own need, not MATLAB CHOL.
Stephan Orzada
Stephan Orzada le 1 Août 2021
You are right. For my purpose, I don't care whether it is positive semidefinite or positive definite. Mathematically there is a difference, but it is negligible in many practical cases.

Connectez-vous pour commenter.

Réponses (1)

Bruno Luong
Bruno Luong le 23 Juil 2021
  2 commentaires
Stephan Orzada
Stephan Orzada le 23 Juil 2021
Modifié(e) : Stephan Orzada le 23 Juil 2021
Thank you, but unfortunately this doesn't work, since chol in MMX ignores the imaginary part of the complex matrices, and anyway does not support the flag which I need for determining PSDness. I couldn't even tell from the results, because even if the matrix is not PSD (for which the Cholesky decomposition doesn't work) it produces a "result" without any error messages.
Bruno Luong
Bruno Luong le 24 Juil 2021
Modifié(e) : Bruno Luong le 24 Juil 2021
Indeed mmx only deal with real matrix.
But if you are willing to modify the code, you might change the function dpotrf in line 284 of mmx.cpp to zpotrf
Of course you have to take care of retriving MATLAB complex internal interleaved data.
You might ask the author if he can gives you a hand for such task.

Connectez-vous pour commenter.

Catégories

En savoir plus sur Linear Algebra dans Help Center et File Exchange

Produits


Version

R2021a

Community Treasure Hunt

Find the treasures in MATLAB Central and discover how the community can help you!

Start Hunting!

Translated by