14
$\begingroup$

Consider $0\leq \alpha\leq 1$, and let $A_{\alpha}$ be the Toeplitz $n\times n$ matrix given by $$ A_\alpha := \begin{bmatrix} 1 & \alpha & \alpha^2 & \ldots &\alpha^{n-1} \\\ \alpha & 1 & \alpha & \ddots & \vdots \\\ \alpha^2 & \alpha & \ddots & \ddots & \alpha^2 \\\ \vdots & \ddots & \ddots & 1 & \alpha \\\ \alpha^{n-1} & \ldots & \alpha^2 & \alpha & 1 \end{bmatrix}. $$

We can decompose $A_{\alpha}=U_{\alpha}D_{\alpha}U_{\alpha}^{*}$ where $U_{\alpha}$ is a unitary matrix and $D_{\alpha}$ is a diagonal matrix.

  • Are the matrices $U_{\alpha}$ and $D_{\alpha}$ explicitly known as a function of $\alpha$ and $n$? Can anyone point me to the right reference.

  • What is known about the spectral limit distribution of these matrices as $n\to\infty$? More specifically, can we compute the limit moments $$ \gamma_{k}:=\lim_{n\to\infty}{\frac{1}{n}\mathrm{Tr}\Big(A_{\alpha}^{k}\Big)} $$ for $k\geq 1$?

Thanks!

$\endgroup$
2
  • $\begingroup$ I took the liberty of changing your title to something more detailed - hope that's OK $\endgroup$
    – Yemon Choi
    Commented Jun 17, 2011 at 22:48
  • 3
    $\begingroup$ An answer to the second point for a generic Toeplitz matrix is given by the Szego-Tyrtyshnikov-Zamarashkin-Tilli theorem. $\endgroup$ Commented Jun 18, 2011 at 19:21

3 Answers 3

16
$\begingroup$

There's likely no explicit diagonalization of $A_\alpha$ except when $n$ is very small or in special cases like $\alpha = 0$ and $\alpha = 1$. Nevertheless each "limit moment" $\gamma_k$ can be computed as a rational function of $\alpha$, and this can be used to describe for each $\alpha$ the distribution of eigenvalues of $A_\alpha$ as $n \rightarrow \infty$.

To diagonalize $A_\alpha$ explicitly we'd need to know the eigenvalues; these are roots of the degree-$n$ characteristic polynomial $\chi_{A_\alpha}$, and it's often too much to expect that a family of such polynomials can be factored for each $n$. Here $\chi_{A_\alpha}$ does split into two factors $\chi^\pm_{A_\alpha}$ of equal or nearly equal degree, but usually that's as far as we can go. The factorization arises because $A_\alpha$ commutes with the involution, call it $\iota$, that takes each coordinate $a_k$ to $a_{n+1-k}$, so the $\pm1$ eigenspaces of $\iota$ are invariant subspaces of $A_\alpha$. The factor $\chi^\pm_{A_\alpha}$ is the characteristic polynomial of the restriction of $A_\alpha$ to the $\pm1$ subspace. But once $n$ is at all large it seems there's nothing to be done with these factors $\chi^\pm_{A_\alpha}$. For example, trying "random" rational values for $\alpha$ yields polynomials whose Galois group is the full symmetric group. Thus if you ask gp

f(a,n) = factor(charpoly(matrix(n,n,i,j,a^abs(i-j))))
F = f(1/2,21)
vector(#F[,1], n, polgalois(F[n,1]))

you'll see that for $n=21$ the factors of $A_{1/2}$ have Galois groups $S_{10}$ and $S_{11}$.

There are some special values of $\alpha$ for which one can find the roots of $\chi_{A_\alpha}$ explicitly. Most obviously, $A_0$ is the identity matrix. Also $A_1$ is the all-ones matrix, with one eigenvalue of $N$ and all other eigenvalues zero. The OP required $\alpha \in [0,1]$, but $A_{-1}$ has rank 2 so its eigenvalues are easy too. In each of these cases there's no unique diagonalization because there's an eigenvalue with high multiplicity.

As for the limit moments $\gamma_k$: if $\alpha=1$ then clearly ${\rm Tr}(A_\alpha^k) = n^k$ so $\gamma_k=\infty$ once $k>1$. So we assume $\alpha < 1$, and then we may as well take $\alpha \in {\bf C}$ with $|\alpha| < 1$. Then $\gamma_1$, $\gamma_2$, $\gamma_3$, $\gamma_4$, $\gamma_5$, etc. are $$ 1, \ \frac{1+\alpha}{1-\alpha},\ \frac{1+4\alpha+\alpha^2}{(1-\alpha)^2},\ \frac{1+9\alpha+9\alpha^2+\alpha^3}{(1-\alpha)^3},\ \frac{1+16\alpha+36\alpha^2+16\alpha^3+\alpha^4}{(1-\alpha)^4}, \ldots $$ and in general $\gamma_k = P_{k-1}(\alpha) / (1-\alpha)^{k-1}$ where $$ P_m(X) := \sum_{j=0}^m \left({m \atop j}\right)^2 X^j $$ is the polynomial obtained from the binomial expansion of $(1+X)^m$ by squaring each coefficient. These $P_m$ don't have an entirely elementary formula, but they can be written as hypergeometric polynomials, or (if memory serves) expressed in terms of Legendre polynomials, or manipulated using the generating function $$ \sum_{m=0}^\infty P_m(X) t^m = \left((\alpha-1)^2 t^2 - 2(\alpha+1)t + 1\right)^{-1/2} $$ if I did this right (I guessed the formula using the technique I described here a few weeks ago: Determining a generating function (of a restricted form)).

To get that formula for $\gamma_k$, we first find an integral representation, which I gather is a special case of the "Szegő-Tyrtyshnikov-Zamarashkin-Tilli theorem" that F. Poloni mentioned in his comment. While general Toeplitz matrices cannot be diagonalized explicitly, circulant ones can. So we compare $A_\alpha$ with the circulant matrix $A'_\alpha$ whose $(i,j)$ entry is $\alpha^{\min(|i-j|,n-|i-j|)}$. For each $\alpha$ and $k$, the $k$-th powers of $A_\alpha$ and $A'_\alpha$ differ by $O(1)$ as $n \rightarrow \infty$. [This was somewhat annoying to check; maybe there's a nice way to do it.] Thus $A_\alpha$ and $A'_\alpha$ have the same limit moments -- and the moments of $A'_\alpha$ can be computed by finding its eigenvalues. Being circulant, $A'_\alpha$ is explicitly diagonalized by the discrete Fourier transform on ${\bf Z} / n {\bf Z}$, with an eigenvalue $\lambda_z = \sum_{j=0}^{n-1} \alpha^{\min(j,n-j)} z^j$ for each $n$th root of unity $z = \exp(2\pi i r/n)$. For large $n$ we can approximate $\lambda_z$ by $$ f_\alpha(z) = \sum_{j=-\infty}^\infty \alpha^{|j|} z^j = \frac1{1-\alpha z} + \frac{\alpha z^{-1}} {1 - \alpha z^{-1}} = \frac{1-\alpha^2}{ (1-\alpha z)(1 - \alpha z^{-1}) } $$ and deduce that $$ \gamma_k = \frac{1}{2\pi} \int_{-\pi}^{\pi} f_\alpha(e^{i\theta})^k d\theta. $$ This also means that as $n \rightarrow \infty$ the eigenvalues of $A_\alpha$ tend to the same distribution as the image of the uniform distribution on the unit circle $|z|=1$ under $f_\alpha$, in the sense that for any continuous function $\phi$ on ${\bf C}$ the average of $\phi(\lambda)$ over the eigenvalues approaches $(2\pi)^{-1} \int_{-\pi}^{\pi} \phi(f_\alpha(e^{i\theta}))d\theta$. For real $\alpha$ in $(0,1)$, this distribution is supported on the interval $((1+\alpha)/(1-\alpha), (1-\alpha)/(1+\alpha))$.

In our case we can evalute the integral for $\gamma_k$ by writing it as the contour integral

$$ \frac1{2\pi i}\oint_{|z|=1} f_\alpha(z)^k \frac{dz}{z}. $$

For each $k \geq 1$ the integrand has a pole of order $k$ at $z = \alpha$ and no other poles in $|z| \leq 1$; evaluating the residue at this pole yields the formula $\gamma_k = P_{k-1}(\alpha) / (1-\alpha)^{k-1}$ given above.

$\endgroup$
1
  • 2
    $\begingroup$ You're welcome! Where did this arise? $\endgroup$ Commented Jun 20, 2011 at 2:07
9
$\begingroup$

While there seems to be no explicit diagonalization, here is a simple idea to obtain an approximation.

To simplify notation, define $T=A_x$ to be the original Toeplitz matrix (I'm using $x$ instead of $\alpha$ to save typing).

After some playing around, it can be seen that the inverse enjoys remarkable structure, namely, $$ T^{-1} = \frac{1}{1-x^2} \begin{bmatrix} 1 & -x & \cdots & \cdots & 0\\\\ -x & 1+x^2 & -x & \cdots & 0\\\\ & \ddots & \ddots & \ddots\\\\ 0 & \cdots & &1+x^2 & -x\\\\ 0 & \cdots & &-x & 1 \end{bmatrix} $$

Consider, therefore, the following Toeplitz matrix \begin{equation*} M := \begin{bmatrix} 1+x^2 & -x & \cdots & \cdots & 0\\\\ -x & 1+x^2 & -x & \cdots & 0\\\\ & \ddots & \ddots & \ddots\\\\ 0 & \cdots & &1+x^2 & -x\\\\ 0 & \cdots & &-x & 1+x^2 \end{bmatrix}, \end{equation*} for which one has closed form eigenvalues and eigenvectors, given by

\begin{equation*} \lambda_k = (1+x^2)-2x\cos\left(\frac{k\pi}{n+1}\right),\quad 1 \le k \le n, \end{equation*} and \begin{equation*} v_{ik} = \sin\left(\frac{ik\pi}{n+1}\right),\quad 1 \le i \le n, 1 \le k \le n. \end{equation*} These eigenvalues (after scaling by $1-x^2$) and eigenvectors may be approximately substituted for those of $T^{-1}$.

Simple experimentation reveals that the eigenvalues of the matrix \begin{equation*} M' := \begin{bmatrix} 1 & -x & \cdots & \cdots & 0\\\\ -x & 1+x^2 & -x & \cdots & 0\\\\ & \ddots & \ddots & \ddots\\\\ 0 & \cdots & &1+x^2 & -x\\\\ 0 & \cdots & &-x & 1 \end{bmatrix} \end{equation*} satisfy $$|\lambda(M) - \lambda(M')| \le 4x/n,$$ where the bound can be made tighter by closer analysis (note that $\lambda(M) \ge \lambda(M')$ also holds). Similar results can also be shown for the eigenvectors, but I haven't had the time to prove that.



EDIT

This journal article shows how to compute eigenvalues and eigenvectors for matrices that look like $T^{-1}$ above. In particular, it implies for example, that the explicit eigenvalues of $T^{-1}$ above are given by

$$\lambda_k = 1 - 2x\cos\theta_k,$$

where, for $k=1,\ldots,n$, the angle $\theta_k$ is a root of $$\sin(n+1)\theta + x^2\sin(n-1)\theta - x\sin(n\theta).$$

Formulae for eigenvectors can also be found in terms of the $x$ and $\theta_k$ as stated above.

However, it seems that in our case, we'll have to numerically solve for $\theta_k$. Modulo that, I guess, this is as close as we'll get to explicit eigenvalues and eigenvectors of $T^{-1}$ (and thereby of $A_x$).

$\endgroup$
2
  • $\begingroup$ That's a nice observation about $T^{-1}$. Maybe revert "$x$" to the OP's "$\alpha$", or at least make the change explicit by writing $T = A_x$ ? $\endgroup$ Commented Jun 22, 2011 at 4:43
  • $\begingroup$ Thanks Noam; I added a sentence incorporating your suggestion. $\endgroup$
    – Suvrit
    Commented Jun 22, 2011 at 13:57
1
$\begingroup$

This class of matrices may have additional nice properties. You might look at the minors when alpha is less than 1; I am thinking in this case, you may have all positive minors, in other words, a totally positive matrix.

$\endgroup$
1
  • $\begingroup$ It's at least positive definite for α in (-1,1), being proportional to the Gram matrix of the first n vectors of the following sequence of vectors in l2: $(1,α,α^2,α^3,α^4,...)$, $(0,1,α,α^2,α^3,...)$, $(0,0,1,α,α^2,...)$, $(0,0,0,1,α,...)$, etc. $\endgroup$ Commented Jun 22, 2011 at 18:38

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .