On error-free filtering of singular processes under nonstationary distortions (Q5930873)

This paper is about a process \(\{\underline W_i,i\) integers\} built a follows: \[ \underline W_i= {U_i\brack V_i}, \quad 1\leq i\leq n, \] where \(U\) is a (unobservable) ``signal process'' and \(V\) is an ``observation process.'' One wants to recover without error (with probability one) the value of \(U_i\) knowing either the past to time \(i\) of \(V\) or the whole of \(V\). Several cases are considered, all of which but one have an observed process of the form ``signal-plus-noise,'' signal and noise independent. The noise process is not assumed to be stationary, but the probability laws of the signal and noise processes are assumed known. The signal process is assumed to be either a singular or completely singular, stationary process (which is a way to define a ``slowly varying process''). The method used in all cases consists in adding to the observed process an independent noise process so that the result is stationary, and then to use known results for error-free filtering of stationary processes. The choice of the added noise process depends on the kuown distributional properties of signal and noise. The cases accounted for in the paper are: 1. When \(V_i=U_i+B_i\), (a) the \(B_i\)'s are independent normal with bounded variance; (b) the \(B_i\)'s form a Gaussian process such that the largest eigenvalue of the covariance matrix of the variables \(B_i\), \(B_{i+1}, \dots, B_{i+p}\) is bounded, for all \(i\) and \(p\). 2. When \(V_i=U_i \oplus B_i\) (addition modulo 2), \(U_i\) and \(B_i\) have values in \(\{0,1\}\), the \(B_i\)'s are independent and \(|P(B_i=0) -1/2|\geq\alpha >0\), 3. when \(U_i\) and \(V_i\) take values in finite sets of respective sizes \(m\leq n\) \((\underline U= \{U_i\), all \(i\})\), given that \(P(V_i=v_j \mid\underline U)= P(V_i=v_j \mid U_i)\), \(P(V_i=v_j \mid U_i=u_k)= p^{(j)}_{j,k}\), \(P_i\) the matrix with entries \(p^{(i)}_{j,k}\), then there exists a submatrix \(P_i^{(m)}\) of \(P_i\), of size \(m\) whose determinant is strictly positive.