My question is related to a previous one linked here. I am interested in non-linear phase FIR filters with a specific desired phase response.

After I tried the options in the linked question I also stumbled upon the idea to separately define weights for phase and magnitude responses developed in Matt's PhD thesis linked here (see section 2.1.3).

The thesis states that a discrete approximation variant can be similarly developed which is what I'm after. While Matt provides a MATLAB implementation, it follows the continuous approximation, and therefore the desired phase response is limited, and only allows for resampling data (fractional delay filter). On the other hand, I have measurements of the required phase response on a grid.


When I try to formulate the problem on a discrete frequency grid I get to a problem. After applying the approximation, the contribution of passband to approximation error reads:

$$\tilde\epsilon_p = \sum_{i=0}^{L-1}\left(W_m(\omega_i)\{Re[\mathbf{e}^H(\omega_i)\mathbf{h}e^{-j\phi_d(\omega_i)}]-|D(e^{j\omega_i})|\}^2 + W_\phi(\omega_i)\{Im[\mathbf{e}^H(\omega_i)\mathbf{h}e^{-j\phi_d(\omega_i)}]\}^2\right) $$

However, I'm not sure how to proceed due to the separate treatment of the real and imaginary parts. How would I get the normal equation form or solve it another way?



Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.