Mathematica Stack Exchange is a question and answer site for users of Mathematica. Join them; it only takes a minute:

Sign up
Here's how it works:
  1. Anybody can ask a question
  2. Anybody can answer
  3. The best answers are voted up and rise to the top

I would like a fast method of creating a sample of random numbers which corresponds to the eigenvalues of a Wishart matrix:

definition of Wishart matrix

For M>N the eigenvalues \lambda_i are given by the jpd

eigenvalue distribution

Where $K_N$ is a normalisation constant and $\beta = 1,2$ depends on whether $X$ is real or complex. Does anyone have any suggestions for a fast way of implementing this? My goal is to play with multiple realisations of eigenvalues representing 100x100 matrices.

Many thanks in advance!

EDIT:

Ok, here is a very basic example:

n = 1000;
m = 1001;
μ = 0;
σ = 1/Sqrt[n];
AbsoluteTiming[
a = RandomVariate[NormalDistribution[μ, σ], {n, m}];
wish = a.a\[ConjugateTranspose];
ewish = Eigenvalues@wish;
]
ewishPlot = Histogram[ewish, {0.1}, "PDF"];

ηp = n σ^2 (1 + Sqrt[η])^2;
ηm = n σ^2 (1 - Sqrt[η])^2;
η = m/n;
ρmp[λ_] := 1/(2 π n σ^2 λ)Sqrt[(ηp - λ) (λ - ηm)];
ρmpPlot = Plot[ρmp[λ], {λ, 0, 4}];
Show[ρmpPlot, ewishPlot]

I guess speed won't be a problem but I would still very much like to know the best way of doing this.

share|improve this question
1  
Do you have a slow method now? Wishart matrices are trivial to construct and Eigenvalues on 100x100 machine number matrices should be very fast as well... – R. M. Mar 21 '14 at 14:04
    
Have a look here. I would've done that probably differently now, but that could be a start. – Leonid Shifrin Mar 21 '14 at 14:31
    
In fact, about 6 years ago I was generating a bunch of these for betas 2 and 4 (IIRC), to study eigenvalue statistics in the context of applications to QCD (Dirac operator in Random Matrix Theory), so I could probably dig up some of that code, but what is in the book can also be a starting point. – Leonid Shifrin Mar 21 '14 at 14:34
    
Thanks very much. This looks great! – user12876 Mar 21 '14 at 15:22
1  
Because it is implemented more generally than what you've done here. You can use a non-diagonal Sigma and it uses MultinormalDistribution to generate the data. In any case, the core part of it is implemented the same way as yours, so I don't see a reason to look for a different way. It's probably as efficient as it can be. – R. M. Mar 21 '14 at 18:11

Since all the entries of $X$ are i.i.d. normal, it can be converted into a band diagonal matrix with $\chi$ distributed components via elementary reflections. A paper discusses about this is "Matrix Models for Beta Ensembles" and can be found in the following link:

http://arxiv.org/abs/math-ph/0206043

The Wishart matrix constructed in this way is tridiagonal and can be represented as SparseArray in Mathematica. The eigenvalues can be obtained efficiently by using Eigenvalues with method "Banded".

share|improve this answer

Your Answer

 
discard

By posting your answer, you agree to the privacy policy and terms of service.

Not the answer you're looking for? Browse other questions tagged or ask your own question.