I'm trying to get insight into the eigenvalue spectrum of a square matrix (large N, symmetric,positive semi definite matrix) using Fourier transforms (I've tried transforming a bunch of thigngs: the diagonasl, the skew diagonal, the entrie matrix, each eigenvector, etc... Is there anything I should be looking at specifically?

  • $\begingroup$ What kind of matrix do you have? After discrete Fourier transform the resulting matrix is unitary equivalent to the original matrix, so all the eigenvalues are the same. However the Fourier transform diagonalizes the matrix only for some very special matrices (Circular matrices). $\endgroup$ – Fabian Mar 5 '11 at 9:04

I know of a couple of results for certain special types of matrices.

1) If you have a circulant matrix, the eigenvalues of that matrix will be the discrete Fourier transform of the first row of the circulant matrix.

2) For the second one I don't remember all the technical details. Suppose you have a symmetric Toeplitz matrix whose size goes to infinity, certain functions of the eigenvalues can be written in terms of the Fourier transform of the first row of the matrix. In fact, the expression is written as a Riemann sum, which becomes a Riemann integral in the Fourier domain. In other words, you can write an unwieldy expression involving the eigenvalues of this growing matrix in terms of an integral which is usually much easier to evaluate.

If you want to learn more, check out Robert Gray's notes at http://ee.stanford.edu/~gray/toeplitz.pdf for an introduction to these kinds of results.

There is also the book by Grenander and Szego which gives a lot more detail and proves some more general results here


  • $\begingroup$ thank you, that was very helpful. in fact my matrix is a symm toeplitz, so your second point is useful. $\endgroup$ – user7815 Mar 5 '11 at 14:45
  • 1
    $\begingroup$ In fact, such problems often come up when you have a discrete time stationary random process and the symmetric toeplitz matrix represents the autocorrelation of the process. Then, the final expression is usually an integral that is written in terms of the power spectral density of the process. $\endgroup$ – svenkatr Mar 5 '11 at 17:23

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy