Fast moving average

A fast implementation of the moving average filter for long kernels.

In terms of behavior, this is an alternative to filter() for a moving-average kernel. The running time does not grow with filter length beyond N=500.

The code uses a variant of the cumsum-trick, although not the "garden variety" but in a way that does not run into numerical issues for long data arrays.

The function is suitable for incremental (online) processing.