A \(p \times p\) real Toeplitz matrix has elements \(A_{u v} = a_{u - v}\) and is fully defined by the elements of \(a_{n}\) for \(-p+1 \le n \le p-1\).

\[A = \begin{bmatrix} a_{0} & a_{-1} & a_{-2} & \cdots & a_{-p+1} \\ a_{1} & a_{0} & a_{-1} & \cdots & a_{-p+2} \\ a_{2} & a_{1} & a_{0} & \cdots & a_{-p+3} \\ \vdots & \vdots & \vdots & \ddots & \vdots \\ a_{p-1} & a_{p-2} & a_{p-3} & \cdots & a_{0} \end{bmatrix}\]

If this is a symmetric matrix then the signal has even symmetry \(a_{n} = a_{-n}\). This post is going to look at two ways of generating symmetric Toeplitz matrices which are positive semi-definite.

Auto-correlation of an infinite signal

One way to generate symmetric Toeplitz matrices which are positive semi-definite is to compute the auto-covariance of an infinite signal. An easy way to obtain an infinite signal is to use the periodic extension of a finite signal. The period \(m\) should be at least \(2p - 1\) to avoid periodic effects. If \(m = p\) then the matrix will be circulant Toeplitz. The Fourier transform can be used to compute the auto-correlation.

p = 100;
m = 2*p-1;
x = randn(m, 1);
a = ifft(conj(fft(x)) .* fft(x)) / m;
a = a(1:p);
A = toeplitz(a, a);

Sampled continuous function

It is possible to consider an infinite Toeplitz matrix with elements \(a_{n}\) for \(n \in \mathcal{Z}\). This matrix maps infinite sequences to infinite sequences such that \(y = A x\) implies

\[y_{u} = \sum_{v \in \mathcal{Z}} a_{u-v} z_{v} .\]

The infinite sequence \(a_{n}\) defines a \(2 \pi\)-periodic function in its discrete-time Fourier transform (DTFT)

\[\hat{a}(\theta) = \sum_{n \in \mathcal{Z}} a_{n} e^{-i n \theta} .\]

The values that this function takes in a period are the eigenvalues of the infinite Toeplitz matrix. Therefore, an infinite Toeplitz matrix will be positive semi-definite if and only if the DTFT of its elements is a non-negative function.

It can be shown that if a continuous function has a positive Fourier transform, then sampling the function will result in a sequence with a positive DTFT. The DTFT of a sampled continuous function \(a_{n} = x(n)\) can be written as the continuous transform of the function multiplied by a “Dirac comb” or “impulse train”

\[\hat{a}(\theta) = \sum_{n \in \mathcal{Z}} x(n) e^{-i n \theta} = \int_{t \in \mathcal{R}} x(t) \left[\sum_{n \in \mathcal{Z}} \delta(t-n)\right] e^{-i t \theta} dt .\]

Using the transform of a Dirac comb and the fact that multiplication in time is convolution in frequency,

\[\hat{a}(\theta) = \mathcal{F} \left\{ x(t) \sum_{n \in \mathcal{Z}} \delta(t-n) \right\}(\theta) = \sum_{n \in \mathcal{Z}} \hat{x}(\theta - 2 \pi n) .\]

Therefore if the continuous transform of \(x(t)\) is non-negative \(\hat{x}(\theta) \ge 0\), then the sequence obtained by sampling that function \(a_{n} = x(n)\) will have a non-negative DTFT \(\hat{a}(\theta) \ge 0\). This gives us many ways to construct infinite symmetric Toeplitz matrices which are positive semi-definite. For example, the elements can be sampled from a Gaussian (transform is a Gaussian), triangle (transform is a squared sinc), sinc (transform is a rectangle), squared sinc (transform is a triangle), cosine, sech, Laplacian density function, etc.

Finally, we need a way to go from infinite matrices to finite matrices. This is easily achieved using the fact that any subset of the rows and columns of a positive semi-definite matrix is itself semi-definite. If \(x^{T} A x\) for all \(x\), then \(y^{T} P^{T} A P y = (P y)^{T} A (P y) \ge 0\) for all \(y\). Therefore we can simply take a finite section of the infinite matrix and be guaranteed that it is positive semi-definite.

p = 100;
t = 0:p-1;
a = exp(-abs(t)/10);
A = toeplitz(a, a);

This gives us two alternative methods by which to obtain symmetric positive semi-definite Toeplitz matrices.