Take the 2-minute tour ×
Mathematics Stack Exchange is a question and answer site for people studying math at any level and professionals in related fields. It's 100% free, no registration required.

Let $w(x,y)$ be a real-valued symmetric function over the $[0,1]$x$[0,1]$ interval. We also know that if $n$ is an integer, and you pick $n$ values $x_i$ in the $[0,1]$ interval, then the matrix $\Sigma_N$ whose entry $(i,j)$ is given by $w(x_i,x_j)$ is symmetric positive definite. To me, $\Sigma_N$ is a covariance matrix so I call $w$ a covariance function.

What is the determinant of $w$? How can it be defined, and more importantly calculated for a given function $w$?

This is as far as I've gotten: If the matrix is diagonal, that is if $w(x,y)=0$ if $x\ne y$, then

$$\det \Sigma_N = \prod_{i=1}^N w(x_i,x_i) $$ so that when $N$ tends to infinity and taking $x_i$ points with even spacing and $x_1=0$ and $x_N=1$ $$\lim_{N\to +\infty} \frac{1}{N-1} \log\det\Sigma_N = \int_0^1 \log w(x,x)\text{d} x$$ How can you generalize that to arbitrary covariance functions?

I'm no mathematician, feel free to make this more rigorous.

share|improve this question
add comment

1 Answer 1

I cannot provide you with a solution of the general case (as I guess is difficult). But there is an important class of matrices for which the limit can be taken rather explicitly (for the other matrices it just remains a limit :-)).

Imagine that the matrix $w(x_i,y_j)$ just depends on the difference $i-j$. These matrices are called Toeplitz matrices and they appear in real life more often than you think. Without the discretization, we introduce $f(x-y) = w(x,y)$.

Take a look at the Fourier transform of $w(x,y)$. It is also a symmetric matrix $$\hat{w}_{m,n}= \int_{[0,1]^2}\! dx dy\, e^{2\pi i (xm + yn)} w(x,y) = \int\! dx dy_{[0,1]^2}\, e^{2\pi i (xm + yn)} f(x-y) = \hat f_{m-n},$$ where $\hat f(m) = \int \!dx\,f(x) e^{2\pi i xm}$. Note that the Fourier transform is unitary (up to a factor $N^{-1}$ see below), so we can calculate the determinant of $\hat{w}$ (not very rigorous but in principle you can write everything out as limits and convince yourself).

Now, we can use a result due to Szegő:

let $w_N$ be the $N\times N$ Toeplitz matrix with the elements $$ (\hat w_N)_{m,n} = \hat f_{m-n}. $$ Then (for $N \to \infty$) $$\log \det \hat{w}_N \sim N [\log f]_0 + \sum_{n=1}^\infty n [\log f]_n [\log f]_{-n}$$ where $$[\log f]_n = \int_0^1\!dx\,\log f(x) e^{2\pi i m x}.$$

We obtain that the logarithm of the functional determinant of $w(x,y)$ is given by (the $1/N$ comes from the Fourier transform) $$ \lim_{N\to\infty} \frac{1}{N} \log \det \hat w_{N} = \int_0^1\!dx\,\log f(x),$$ which gives your result when all the elements on the diagonal are equal.

Note that all of this is not yet rigorous (I just wanted to give you a short summary). In principle you should in the beginning introduce $N$ by only allowing $w$ to have up to $N$ Fourier components. Then use Szegő and in the end let $N\to\infty$.

share|improve this answer
    
wonderful! Is an extension possible to small deviations from the Toeplitz case? Say to a matrix M=T+E where T is toeplitz and E is any matrix but M and T have almost the same norm? –  yannick Jan 13 '12 at 19:18
add comment

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.