Hyper-positive definite functions I: scalar case, branching-type stationary stochastic processes (Q2243159)

Going through the paper is a fascinating journey from probability theory to functional analysis and then back. The premise of the paper lies in an entirely probabilistic question: How do you define a branching-type stationary stochastic process (SSP, in short) on rooted infinite trees? In order to answer this question, the authors begin by first obtaining necessary and sufficient conditions for the existence of branching-type SSPs on rooted trees. Then, the authors prove a variant of the Herglotz-Bochner theorem for rooted homogeneous trees. Finally, using the classical Szegö First Theorem, the authors give an explicit formula for the \(L^2\)-distance between the random variable at the root of the infinite tree and the closed linear span of all other random variables in the SSP in the \(L^2\) space. An SSP is a zero-mean stochastic process \(X = (X_n)_{n \in \mathbb{N}}\) indexed by \(\mathbb{N}\), the set of nonnegative whole numbers, such that each random variable \(X_n\) admits a finite second order moment and \(\mathsf{Cov}(X_n, X_{n+k})\) is independent of \(n\). The primary mathematical object of study in this paper is a branching-type SSP, which is formally defined as a zero mean square-integrable stochastic process \((X_{\sigma})_{\sigma \in T}\) indexed by a rooted tree \(T\) with root \(e\) satisfying the following conditions: \begin{itemize} \item[1.] Restriction of \((X_{\sigma})_{\sigma \in T}\) on rooted geodesics (\(T\) is assumed to be equipped with the usual graph distance) is a classical SSP when the geodesic is identified with \(\mathbb{N}\). \item[2.] The family of these SSPs restricted on the geodesics share a common spectral measure, which is called the spectral measure of the branching-type SSP. \item[3.] Each pair of random variables on two vertices that are not on the same rooted geodesic of the tree is uncorrelated. \end{itemize} Such a branching-type SSP on an infinite rooted tree \(T\) exists if and only if \(T\) is of uniformly bounded degree (Theorem 1.1). The crux of this result lies in an embedding of the tree with uniformly bounded degree into a rooted homogeneous tree, for which the spectral measure of a branching-type SSP can be completely characterized in terms of the Poisson convolution of a positive Radon measure (Theorem 1.2). More precisely, Theorem 1.2 states that a measure \(\nu\) on \(\mathbb{T} = \mathbb{R}/2\pi \mathbb{Z}\) is the spectral measure of a branching-type SSP on \(T_q\), the set of vertices of the infinite rooted \(q\)-homogeneous tree, if and only if there exists a positive Radon measure \(\mu\) on \(\mathbb{T}\) such that \(\nu\) is the \((1/\sqrt{q})\)-Poisson convolution of \(\mu\), i.e., the measure \(\nu\) satisfies \[ \nu (e^{i\theta}) = \int_{\mathbb{T}} \frac{1}{2\pi} \frac{ (q-1) }{ |\sqrt{q} - e^{i (\theta - \theta')} |^2 } \mathrm{d} \mu(\theta'), \] for some positive Radon measure \(\mu\). An alternative way to view branching-type SSPs is to embed (one dimensional) SSPs into a stochastic process with a given structure on a tree. More precisely, one could ask under what condition(s) on the spectral measure of a given SSP \((X_n)_{n \in \mathbb{N}}\) and a rooted tree \(T\), the stochastic processes indexed by the vertices along geodesics of the tree \(T\) can be embedded into a branching-type SSP on \(T\). While Theorem 1.2 gives a complete answer to this question for rooted homogeneous trees in terms of Poisson convolutions, the picture is far from clear for other rooted trees. For the special case of trees where only the root has \(q\) descendants and all other vertices have only one descendant, denoted by \(T(q;1)\), the authors are able to answer the question of embedding (Theorem 1.5): A positive Radon measure \(\mu\) is the spectral measure of a branching-type SSP on \(T(q;1)\) if and only if the measure \(\mu\) satisfies the following inequality: \[ \exp\left( \int_{\mathbb{T}} \log( \frac{\mathrm{d}\mu_{a}}{\mathrm{d}h} ) \mathrm{d}h \right) \ge (1- q^{-1})\mu(\mathbb{T} ), \] where \(\mu_a\) is the absolutely continuous part of \(\mu\) with respect to the normalized Haar measure \(h\) on \(\mathbb{T}\). In my view, the main contribution of the paper is the characterization of hyper-positive definite functions. A function \(\alpha\) on \(\mathbb{N}\) is defined to be \(q\)-hyper-positive definite (\(q\)-HPD, in short) if there exists a branching-type SSP \((X_\sigma)_{\sigma \in T_q}\) on \(T_q\) whose spectral measure \(\mu\) satisfies \[\alpha(n) = \int_{\mathbb{T}} e^{-i n \theta} \mathrm{d}\mu(\theta).\] The authors prove the following version of Herglotz-Bochner Theorem (Theorem 1.8 in the paper) for \(q\)-HPD functions: A function \(\alpha\) on \(\mathbb{N}\) is \(q\)-HPD if the map \(n \mapsto q^{n/2}\alpha(n)\) is positive definite on \(\mathbb{N}\), which consequently means that the function \(\alpha\) is \(q\)-HPD if and only if there exists a positive Radon measure \(\mu\) such that \[ \alpha(n) = q^{-n/2}\int_{\mathbb{T}} e^{-in\theta} \mathrm{d}\mu(\theta). \] This is a very powerful result in that it characterizes \emph{all} \(q\)-HPD functions. The authors go on to compute an explicit expression for the \(L^2\)-distance between the random variable at the root of the infinite rooted \(q\)-homogeneous tree \(T_q\) and the closed linear span of all other random variables in the SSP in the \(L^2\) space. As a byproduct of their investigation, the authors discover new hyper-contractive inequalities for Hankel operators, which are defined on the dense subset of analytic trigonometric polynomials as the orthogonal (Riesz) projection onto the space \[ H_{-}^{2}(\mathbb{T}) = \{ f(e^{i\theta}) = \sum_{n=0}^{\infty}a_n e^{-in\theta} \mid \sum_{n=0}^{\infty} |a_n|^2 < \infty \}. \] The inequalities seem to be of independent interest and may find applications outside probability theory and functional analysis. The paper, the first in a series of papers, tackles a problem, which, I think, is more interesting from a probabilistic perspective than a functional analytic one. Regardless of whether this personal opinion of mine is shared by my colleagues in the two communities, I hope this beautiful interplay between probability theory and functional analysis will inspire many to investigate similar problems in the future. Many congratulations to the authors for the explicit calculations.