According to Paley-Wiener theorem, a practically realizable frequency selective filter shouldn't have a constant magnitude for a range of frequencies. Why is this so? How do we prove and understand it?!

To summarize, causality has very important implications in the design of frequency-selective filters. These are: (a) the frequency response $H(\omega)$ cannot be zero, except at a finite set of points in frequency; (b) the magnitude $|H(\omega)|$ cannot be constant in any finite range of frequencies and the transition from passband to stopband cannot be infinitely sharp [this is a consequence of the Gibbs phenomenon, which results from the truncation of $h(n)$ to achieve causality]; and (c) the real and imaginary parts of $H(\omega)$ are interdependent and are related by the discrete Hilbert transform. As a consequence, the magnitude $|H(\omega)|$ and phase $\Theta(\omega)$ of $H(\omega)$ cannot be chosen arbitrarily.

Point b in the quote. Source: Proakis ch.10 4th edition

  • 1
    $\begingroup$ Hi. Rather than constant gain, a constant frequency response (magnitude and phase) should be what you are asking for. $\endgroup$ – Juancho Jan 8 '19 at 12:06
  • $\begingroup$ Sorry I meant magnitude of H(w) $\endgroup$ – Trilok Girish Kamagond Jan 9 '19 at 8:55
  • $\begingroup$ I didn't understand the difference between gain and magnitude? $\endgroup$ – Trilok Girish Kamagond Jan 9 '19 at 9:02
  • 1
    $\begingroup$ You may want to limit the question to frequency selective filters (as in the quote) which attenuate some frequencies while pass others largely unattenuated. That would exclude all-pass filters. $\endgroup$ – Olli Niemitalo Jan 9 '19 at 9:23
  • 1
    $\begingroup$ Thank you so much! I may limit the question to frequency selective filters $\endgroup$ – Trilok Girish Kamagond Jan 10 '19 at 2:27

Note that all-pass filters with a perfectly constant magnitude response over the whole frequency range are realizable. A first-order discrete-time all-pass filter has the following transfer function:


If $|a|>1$, the system $(1)$ is causal and stable. It is straightforward to show that


i.e., the system's magnitude response is constant.

What is not possible is to have a constant magnitude response over only a part of the possible frequency range $[0,2\pi]$ (or, in continuous time, $(-\infty,\infty)$).

This can be seen as follows. First, design an ideal frequency-selective filter by computing the IDFT of some ideal frequency response (with a constant gain in the pass band(s), and a gain of zero in the stop band(s)). Let's consider discrete-time filters, but the discussion is basically the same for continuous-time filters. Let the resulting ideal impulse response be $h_{id}[n]$. First, we shift that impulse response to the right to have most of its energy at positive indices $n>0$:

$$\tilde{h}_{id}[n]=h_{id}[n-n_0],\qquad n_0>0\tag{3}$$

In order to obtain a causal filter, we need to multiply that shifted impulse response by a sequence $f[n]$ satisfying

$$f[n]=\begin{cases}0,&n<0\\1,&n>n_1\ge 0\end{cases}\tag{4}$$

with some non-negative $n_1$. The values of $f[n]$ between $n=0$ and $n=n_1$ can be chosen arbitrarily, but in practice we would choose them such that $f[n]$ increases monotonically in the interval $[0,n_1]$. The final impulse response of our causal filter is


Note that $h[n]$ is still infinitely long.

In the frequency domain this corresponds to

$$H(e^{j\omega})=\left(H_{id}(e^{j\omega})e^{-jn_0\omega}\right)\star F(e^{j\omega})\tag{6}$$

where $\star$ denotes convolution. This convolution destroys the piecewise constant property of the original frequency response $H_{id}(e^{j\omega})$, and the magnitude of the resulting causal frequency response $H(e^{j\omega})$ is neither piecewise constant, nor does it have infinitely sharp transitions.

  • $\begingroup$ Thank you so much @Matt L. Very satisfying answer. Can u please tell why in practice f(n) should have monotonically increasing values between 0 to n1? $\endgroup$ – Trilok Girish Kamagond Jan 10 '19 at 2:46
  • $\begingroup$ @TrilokGirishKamagond: That's just a reasonable choice because you want a function that smoothly moves from zero to one, just like a one-sided window function. $\endgroup$ – Matt L. Jan 10 '19 at 8:02
  • $\begingroup$ You show that truncating the impulse response of a filter with a magnitude frequency response with a flat segment no longer gives such a filter. But this leaves out what can or cannot be achieved by truncating the impulse response of a filter that does not originally have that flat segment in its magnitude frequency response. $\endgroup$ – Olli Niemitalo Jan 15 '19 at 9:17
  • 1
    $\begingroup$ @OlliNiemitalo: Yes, that's right, the whole answer clearly doesn't prove anything, it's just supposed to make things plausible. $\endgroup$ – Matt L. Jan 15 '19 at 10:13
  • $\begingroup$ @Matt L I have a small doubt regarding multiplication of f(n) to hid[n] to get a causal system. I feel hid[n] is already causal as it depends on past values only as long as n0 is large enough. So what can I understand from the multiplication by f(n)? And when n > n0 won't the system become non causal? $\endgroup$ – Trilok Girish Kamagond Jan 17 '19 at 12:03

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.