Crossings of smooth shot noise processes (Q1931317)

A shot noise process, \[ X_t=\sum_i \beta_i g(t-\tau_i), \quad t\in\mathbb{R}, \] is considered, where \({g}\) is a given deterministic measurable function (the kernel function of the process), \({T_i}\) are the points of a Poisson point process on the line of intensity \(\lambda \nu (ds)\) (\(\lambda>0\) and \(\nu\) is a positive \(\sigma\)-finite measure on \(\mathbb{R}\)) and the \({\beta_i}\) are independent copies of a random variable \(\beta\) (called the impulse) independent of \({\tau_i}\). The authors consider the expected number of level crossings of the shot noise process \({X}\) when the kernel function is smooth. Integral formulas for the mean number of level crossings function \[ C_X (\alpha, [a,b]) = \operatorname{E} (N_X (\alpha, [a,b])) = \operatorname{E} (\sharp \{t \in [a,b]\mid X(t) = \alpha\}) \quad \forall \alpha \in \mathbb{R} \] are obtained. The introduction contains an overview of what is known on the problem. In Section 2, crossings for second-order processes \({X}\) which are both almost surely and mean-square continuously differentiable are considered. Contrary to most known statements on crossings based on assumptions on density probabilities, the authors revisit these results with point of view based on characteristic functions. Theorem 1 contains the Fourier transform of \(C_X(\alpha, [a,b]) \in L^1 (\mathbb{R})\) and its expression is computed by the help of the joint characteristic function of \((X(t), X'(t))\). The mean number of crossings for a given level is considered in Subsection 2.2. An upper bound on \(C_X (\alpha, [a,b])\) is obtained. For the sum \(X=\sum^n_{j=1} X_j\), \(n \geq 2\), of independent real-valued processes almost surely and mean-square two times continuously differentiable on \([a,b]\), the upper estimate for \(C_X (\alpha, [a,b])\) is obtained. At last, the continuity in \(\alpha\) of \(C_X (\alpha, [a,b])\) is discussed. In Section 3, shot noise processes, given by (1), are considered. Under certain conditions on the kernel \({g}\) the process \({X}\) is almost surely and mean square continuously differentiable on \(\mathbb{R}\) with \[ X'(t)=\sum_i \beta_i g'(t-\tau_i), \quad t\in\mathbb{R}. \] By iterating, it is possible to obtain higher-order smoothness. As an example, the random process with Gaussian kernel is presented. Then, an \(\mathbb{R}^d\)-valued shot noise process given on \(\mathbb{R}\), \[ Y(t)=\sum_2 \beta_i h(t-\tau_i), \] where \(h: \mathbb{R} \rightarrow \mathbb{R}^d\) is a deterministic measurable vectorial function in \(L^1(\mathbb{R})\), is considered. Let \(Y_T\) is the process given by (2) with \(\nu_T(ds) = \mathbf{1}_{[-T,T]} (s) ds\). Moreover, for any \(M>T\) denote \(Y_M = Y_T+(Y_M-Y_T)\). The existence of a density and the continuity of the process \(Y_T\) are considered. When \({h}\) has compact support, then exists \(A>0\) such that \(h(s)=0\) for \(|s| > A\) and for \(T \geq A\), \(X_T(0)\) do not have a density. Proposition 4 gives a criterion for \(Y_T(0)\) to have a density. Proposition 5 contains conditions on the kernel \({g}\) that imply \(\frac{\operatorname{P}\{|X(t)| \leq \varepsilon\}}{\varepsilon} \rightarrow \infty\) as \(\varepsilon \rightarrow 0\). Thus, the density of \({X(t)}\) (if it exists) is not bounded in a neighborhood of 0. Let \(X_T= \sharp \{i:\tau_i \in [-T,T]\}\). The conditions of Theorem 2 implies that, for all \(M \geq T\), the mean \(\operatorname{E}(N_{X_M} (\alpha, [a,b])|\gamma_T \geq k_0)\) is continuous in \(\alpha\) on \(\mathbb{R}\) if \(k_0 \geq 8\), and, as \(M \rightarrow \infty\), it converges to \(\operatorname{E}(N_X (\alpha, [a,b])|\gamma_T \geq k_0)\) uniformly on \(\alpha \in \mathbb{R}\). In Section 4, a strictly stationary shot noise process \(X_\lambda\) given by (1) with intensity \(\lambda\) is considered. Let \[ Z_\lambda(t) = \frac{1}{\sqrt{\lambda}} (X_\lambda(t) - \operatorname{E}(X_\lambda(t))), \quad t \in \mathbb{R}. \] Let \(\beta \in L^2(\Omega)\). For certain \({g}\), the process \[ Y_\lambda = \left(\begin{matrix} Z_\lambda \\ Z'_\lambda \end{matrix} \right) \overset{f.d.d.}{\longrightarrow}_{\lambda \rightarrow +\infty} \sqrt{\operatorname{E}(\beta^2)} \left(\begin{matrix} B \\ B' \end{matrix} \right), \] where \({B}\) is a stationary centered Gaussian process almost surely and mean square continuously differentiable with \[ \operatorname{Cov} (B(t), B(t')) = \int_{\mathbb{R}} g(t-s) g(t'-s)\, ds. \] In this section, it is shown how and in which sense the mean number of crossings function converges to the one of a Gaussian process as the intensity \(\lambda\) goes to infinity. The convergence rates are given. In Section 5, the application of shot noise processes in physics is of interest. The physical model is given by sources that produce each a potential in such a way that the global potential at a point is the sum of all individual potentials. This can be modeled as a shot noise process. \[ X_{\lambda, \sigma} = \sum_i \frac{1}{\sigma \sqrt{2\pi}} e^{-(t-\tau_i)^2 / 2 \sigma^2}, \quad t \in \mathbb{R}, \] where \(\{T_i\}\) are the points of a Poisson process of intensity \(\lambda>0\) on \(\mathbb{R}\). Let \[ g_\sigma(t) = \frac{1}{\sigma \sqrt{2\pi}} e^{-t^2 / 2 \sigma^2}. \] The mean number of local extrema of this process as a function of \(\sigma\) is of interest. Thanks to the heat equation and also to scaling properties between \({T}\) and \(\lambda\), it is proved that the mean number of local extrema is decreasing function of \(\sigma\) and its asymptotics for \(\sigma\) small and large are given. The Appendix contains the stationary phase estimate for oscillatory integrals and the proofs of Proposition 8 and Lemma 3 on the uniform convergence for the mean number of crossings function.