Topology-guided sampling of nonhomogeneous random processes (Q988765): Difference between revisions

For a function \(u:[a,b]\to \mathbb{R}\) and a threshold function \(\mu:[a,b]\to \mathbb{R}\) the \textit{generalized nodal domain} of \(u\) is a set \(N_{\mu}^{\pm}=\{x\in [a,b]: \pm(u(x)-\mu(x))\geq 0\}\). Taking the grid points \(a=x_0 < x_1 < \dots < x_M=b\) consider the \textit{cubical approximations} \(Q_M^{\pm}\) of such \(N_{\mu}^{\pm}\) defined as \(Q_{\mu,M}^{\pm}:=\bigcup\{[x_k,x_{k+1}]: \pm((u-\mu)(x_k))\geq 0, k=0,\dots,M\}\) where \(x_{M+1}=x_M=b\). Let \(\beta_0(X)\) be a number of components of a set \(X\subset \mathbb{R}\). If \(u\) is a random process, then under specified conditions the authors prove the lower bound for \(\mathbb{P}(\beta_0(N_{\mu}^{\pm})=\beta_0(Q_{\mu}^{\pm}))\) generalizing Theorem 1.3 in their previous paper [Ann. Appl. Probab. 17, 980--1018 (2007; Zbl 1131.60047)]. Special attention is paid to the Gaussian processes \(u\). Thus explicit probabilistic a priori bounds are provided for the suitability of certain discretization sizes and the choice of the sampling points \(x_0,\dots,x_M\) to minimize the error probability is tackled. Applications are given for various stochastic processes, e.g., \(u(x)= \sum_{k=0}^{N}g_k \varphi_k(x)\) where \(x\in [a,b]\), \(N\leq \infty\), \(g_1,g_2,\dots\) are mean zero Gaussian random variables (defined on the same probability space) and \(\varphi_1,\varphi_2,\dots\) are smooth real-valued functions on \([a,b]\). The optimal size of the grid and the relationship with the density of zeros of \(u\) are discussed as well.