Adaptive time-frequency analysis of nonstationary signals

We consider the problem of time-frequency analysis of nonstationary signals. Instead of the usual one-sided line prediction (OSP), we propose to use the so called two-sided linear prediction (TSP) approach. In the TSP, the signal is modeled as a linear combination of its past and future values. The vector of prediction coefficients is estimated adaptively using an algorithm based on the least mean square, noisy gradient technique. Numerical examples show that TSP requires a smaller order than OSP to achieve better resolution, and consequently is computationally efficient.