Remez Exchange Algorithm

The Remez multiple exchange algorithm works by moving the frequency samples each iteration to points of maximum error (on a denser grid). Remez iterations could be added to our formulation as well. The Remez multiple exchange algorithm (function firpm [formerly remez] in the Matlab Signal Processing Toolbox, and still remez in Octave) is normally faster than a linear programming formulation, which can be regarded as a single exchange method [224, p. 140]. Another reason for the speed of firpm is that it solves the following equations non-iteratively for the filter exhibiting the desired error alternation over the current set of extremal frequencies:

$$\left[ \begin{array}{c} H(\omega_1) \\ H(\omega_2) \\ \vdots \\ H(\omega_{K}) \end{array} \right] = \left[ \begin{array}{cccccc} 1 & 2\cos(\omega_1) & \dots & 2\cos(\omega_1L) & \frac{1}{W(\omega_1)} \\ 1 & 2\cos(\omega_2) & \dots & 2\cos(\omega_2L) & \frac{-1}{W(\omega_2)} \\ \vdots & & & \\ 1 & 2\cos(\omega_{K}) & \dots & 2\cos(\omega_{K}L) & \frac{(-1)^{K}}{W(\omega_{K})} \end{array} \right] \left[ \begin{array}{c} h_0 \\ h_1 \\ \vdots \\ h_{L} \\ \delta \end{array} \right]$$ (4.76)

where $W(\omega_k)\delta$ is the weighted ripple amplitude at frequency $\omega_k$ . ( $W(\omega_k)$ is an arbitrary ripple weighting function.) Note that the desired frequency-response amplitude $H(\omega_k)$ is also arbitrary at each frequency sample.