Can you please help me find the distribution of eigenvalues of a Toeplitz matrix $\mathbf{K}$ that is constructed as follows: $$\mathbf{K}=\left[ \begin{array}{cccc} 1 & \rho & \ldots & \, \, \rho^{N-1} \\ \rho & 1 & \ldots & \, \,\rho^{N-2}\\ \vdots & \vdots & \ddots & \vdots \\ \rho^{N-1} & \rho^{N-2} & \ldots & 1 \\ \end{array} \right].$$ where $0 \leq \rho < 1$.

Thanks a lot in advance,



migrated from stats.stackexchange.com Aug 15 '11 at 13:40

This question came from our site for people interested in statistics, machine learning, data analysis, data mining, and data visualization.

  • $\begingroup$ what do you mean by 'distribution' ? Are all the elements numbers (as opposed to random variables) ? Do you need this in analytical form or is a numerical solution sufficient ? $\endgroup$ – Andre Holzner Aug 15 '11 at 11:33
  • $\begingroup$ @ Andre, Thanks a lot for your comment, you are right. Then, there is any analytical equation to evaluate eigenvalues? $\endgroup$ – Farzad Aug 15 '11 at 12:23
  • $\begingroup$ Because this question is purely about the eigenvalues of these matrices, it is more suited for posting on math.SE. $\endgroup$ – whuber Aug 15 '11 at 13:40

For large $N$, the eigenvalues of $K$ are approximately distributed as $2\pi f(\lambda)$ evaluated at frequencies $-\pi + 2\pi j/N$; $f(\lambda)$ is the spectral density associated to the covariance sequence in your Toeplitz matrix. See for instance Hannan, E.J. Time Series Analysis, Chap. 1 towards the end.

You may find more details in Grenander & Szego, Toeplitz forms and their applications, but I do not have that book at hand and cannot say from memory if it will answer your question more precisely.

  • $\begingroup$ @Tusell, thanks a lot for your answer. I have some questions: 1) I think in frequency range $-\pi+2\pi j/N$ there should be $\lambda$ parameter aomewhere, 2) is $f(\lambda) = |\frac{-2\ln \rho}{(\ln \rho)^2+ (2\pi/N\lambda)}|^2$? $\endgroup$ – Farzad Aug 17 '11 at 10:42
  • $\begingroup$ @Farzad: sorry I couldn't answer earlier. Was on a holiday and not connected. $\lambda$ takes values $\lambda_j = -\pi+2\pi j/N$ for integer values of $0 \le j \le N$. The spectral density is the Fourier transform of the covariance sequence, hence for your case would be something like $\frac{1}{2\pi}(1 + \sum_{j}\rho^{|j|}\cos(\lambda_j))$. $\endgroup$ – F. Tusell Sep 1 '11 at 10:44
  • $\begingroup$ @Farzad The range $\lambda_j = -\pi + 2\pi j /N$ for $0\leq j \leq N$ doesn't seem right. With that range, I end up with N + 1 values, but the matrix $\mathbf{K}$ will have only ${N}$ eigenvalues. I think it should be $0 \leq j < N$. $\endgroup$ – sauravrt Mar 22 '12 at 17:39

Your Answer

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

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