An algorithm to separate nonstationary part of a signal using mid-prediction filter

In the autoregressive model, both end-prediction (prediction using the past N values) and mid-prediction (prediction using N/2 past and N/2 future values) filters may be used. If the nonstationary part consists of random impulsive waves of low occurrence rate, it may be separated as an error signal corrupted with “carried over” errors. The latter is removed using the signal-inversion technique. Impulses with slowly rising and falling edges are sometimes recovered better by the mid-prediction filter because it provides higher gain at low frequencies