Filtering with degenerate observation noise: a stochastic approximation approach (Q2151896)

Filtering problems for partially observed continuous-time control systems are extensively studied. A usual assumption for such problem is the non-degeneration of the noise. With such assumption, various filtering schemes were developed. Nevertheless, in many applications, the non-degeneration of the noise may not hold and the traditional filtering methods do not work. In this interesting paper, the authors develop a viable scheme to address possible degenerate observation noise by introducing a recursive filtering equation in which the gain matrix is a matrix-valued parameter to be determined. They adopt the Monte Carlo training procedure used for deep filtering to determine the best gain matrix. They focus on computational experiments and demonstrate the performance of the SA filter and its robustness. They also compare the SA filter with the (extended) Kalman-Bucy filter and the deep filter in both linear and nonlinear models.