In the next two sections we look at two alternative forms of the phase response: phase delay and group delay. After considering some examples and special cases, poles and zeros of the transfer function are discussed in the next chapter.
The phase response of an LTI filter gives the radian phase shift added to the phase of each sinusoidal component of the input signal. It is often more intuitive to consider instead the phase delay, defined as
and it can be clearly seen in this form that the phase delay expresses the phase response as a time delay in seconds.
In working with phase delay, it is often necessary to ``unwrap'' the phase response . Phase unwrapping ensures that all appropriate multiples of have been included in . We defined simply as the complex angle of the frequency response , and this is not sufficient for obtaining a phase response which can be converted to true time delay. If multiples of are discarded, as is done in the definition of complex angle, the phase delay is modified by multiples of the sinusoidal period. Since LTI filter analysis is based on sinusoids without beginning or end, one cannot in principle distinguish between ``true'' phase delay and a phase delay with discarded sinusoidal periods when looking at a sinusoidal output at any given frequency. Nevertheless, it is often useful to define the filter phase response as a continuous function of frequency with the property that or (for real filters). This specifies how to unwrap the phase response at all frequencies where the amplitude response is finite and nonzero. When the amplitude response goes to zero or infinity at some frequency, we can try to take a limit from below and above that frequency.
Matlab and Octave have a function called
implements a numerical algorithm for phase unwrapping.
Figures 7.6.2 and 7.6.2 show the effect of the
unwrap function on the phase response of the example elliptic
lowpass filter of §7.5.2, modified to contract the zeros from
the unit circle to a circle of radius in the plane:
[B,A] = ellip(4,1,20,0.5); % design lowpass filter B = B .* (0.95).^[1:length(B)]; % contract zeros by 0.95 [H,w] = freqz(B,A); % frequency response theta = angle(H); % phase response thetauw = unwrap(theta); % unwrapped phase responseIn Fig.7.6.2, the phase-response minimum has ``wrapped around'' to the top of the plot. In Fig.7.6.2, the phase response is continuous. We have contracted the zeros away from the unit circle in this example, because the phase response really does switch discontinuously by radians when frequency passes through a point where the phases crosses zero along the unit circle (see Fig.7.3(b)). The unwrap function need not modify these discontinuities, but it is free to add or subtract any integer multiple of in order to obtain the ``best looking'' discontinuity. Typically, for best results, such discontinuities should alternate between and , making the phase response resemble a distorted ``square wave'', as in Fig.7.3(b). A more precise example appears in Fig.10.2.
An example of a linear phase response is that of the simplest lowpass filter, . Thus, both the phase delay and the group delay of the simplest lowpass filter are equal to half a sample at every frequency.
For any reasonably smooth phase function, the group delay may be interpreted as the time delay of the amplitude envelope of a sinusoid at frequency . The bandwidth of the amplitude envelope in this interpretation must be restricted to a frequency interval over which the phase response is approximately linear. We derive this result in the next subsection.
Thus, the name ``group delay'' for refers to the fact that it specifies the delay experienced by a narrow-band ``group'' of sinusoidal components which have frequencies within a narrow frequency interval about . The width of this interval is limited to that over which is approximately constant.
Derivation of Group Delay as Modulation Delay
Suppose we write a narrowband signal centered at frequency as
where is defined as the carrier frequency (in radians per sample), and is some ``lowpass'' amplitude modulation signal. The modulation can be complex-valued to represent either phase or amplitude modulation or both. By ``lowpass,'' we mean that the spectrum of is concentrated near dc, i.e.,
Using the above frequency-domain expansion of , can be written as
Assuming the phase response is approximately linear over the narrow frequency interval , we can write
where we also used the definition of phase delay, , in the last step. In this expression we can already see that the carrier sinusoid is delayed by the phase delay, while the amplitude-envelope frequency-component is delayed by the group delay. Integrating over to recombine the sinusoidal components (i.e., using a Fourier superposition integral for ) gives
We have shown that, for narrowband signals expressed as in Eq.(7.6) as a modulation envelope times a sinusoidal carrier, the carrier wave is delayed by the filter phase delay, while the modulation is delayed by the filter group delay, provided that the filter phase response is approximately linear over the narrowband frequency interval.
Group Delay Examples in Matlab
Figure 7.6 compares the group delay responses for a number of classic lowpass filters, including the example of Fig.7.2. The matlab code is listed in Fig.7.5. See, e.g., Parks and Burrus  for a discussion of Butterworth, Chebyshev, and Elliptic Function digital filter design. See also §I.2 for details on the Butterworth case. The various types may be summarized as follows:
- Butterworth filters are maximally flat in middle of the passband.
- Chebyshev Type I filters are ``equiripple'' in the passband and ``Butterworth'' in the stopband.
- Chebyshev Type II filters are ``Butterworth'' in the passband and equiripple in the stopband.
- Elliptic function filters are equiripple in both the passband and stopband.
As Fig.7.6.4 indicates, and as is well known, the Butterworth filter has the flattest group delay curve (and most gentle transition from passband to stopband) for the four types compared. The elliptic function filter has the largest amount of ``delay distortion'' near the cut-off frequency (passband edge frequency). Fundamentally, the more abrupt the transition from passband to stopband, the greater the delay-distortion across that transition, for any minimum-phase filter. (Minimum-phase filters are introduced in Chapter 11.) The delay-distortion can be compensated by delay equalization, i.e., adding delay at other frequencies in order approach an overall constant group delay versus frequency. Delay equalization is typically carried out using an allpass filter (defined in §B.2) in series with the filter to be delay-equalized .
[Bb,Ab] = butter(4,0.5); % order 4, cutoff at 0.5 * pi Hb=freqz(Bb,Ab); Db=grpdelay(Bb,Ab); [Bc1,Ac1] = cheby1(4,1,0.5); % 1 dB passband ripple Hc1=freqz(Bc1,Ac1); Dc1=grpdelay(Bc1,Ac1); [Bc2,Ac2] = cheby2(4,20,0.5); % 20 dB stopband attenuation Hc2=freqz(Bc2,Ac2); Dc2=grpdelay(Bc2,Ac2); [Be,Ae] = ellip(4,1,20,0.5); % like cheby1 + cheby2 He=freqz(Be,Ae); [De,w]=grpdelay(Be,Ae); figure(1); plot(w,abs([Hb,Hc1,Hc2,He])); grid('on'); xlabel('Frequency (rad/sample)'); ylabel('Gain'); legend('butter','cheby1','cheby2','ellip'); saveplot('../eps/grpdelaydemo1.eps'); figure(2); plot(w,[Db,Dc1,Dc2,De]); grid('on'); xlabel('Frequency (rad/sample)'); ylabel('Delay (samples)'); legend('butter','cheby1','cheby2','ellip'); saveplot('../eps/grpdelaydemo2.eps');
The definitions of phase delay and group delay apply quite naturally to the analysis of the vocoder (``voice coder'') [21,26,54,76]. The vocoder provides a bank of bandpass filters which decompose the input signal into narrow spectral ``slices.'' This is the analysis step. For synthesis (often called additive synthesis), a bank of sinusoidal oscillators is provided, having amplitude and frequency control inputs. The oscillator frequencies are tuned to the filter center frequencies, and the amplitude controls are driven by the amplitude envelopes measured in the filter-bank analysis. (Typically, some data reduction or envelope modification has taken place in the amplitude envelope set.) With these oscillators, the band slices are independently regenerated and summed together to resynthesize the signal.
Suppose we excite only channel of the vocoder with the input signal
If the phase of each channel filter is linear in frequency within the passband (or at least across the width of the spectrum of ), and if each channel filter has a flat amplitude response in its passband, then the filter output will be, by the analysis of the previous section,
where is the phase delay of the channel filter at frequency , and is the group delay at that frequency. Thus, in vocoder analysis for additive synthesis, the phase delay of the analysis filter bank gives the time delay experienced by the oscillator carrier waves, while the group delay of the analysis filter bank gives the time delay imposed on the estimated oscillator amplitude-envelope functions.
Numerical Computation of Group Delay
The definition of group delay,
A more useful form of the group delay arises from the logarithmic derivative of the frequency response. Expressing the frequency response in polar form as
Since differentiation is linear, the logarithmic derivative becomes
In this case, the derivative is simply
where denotes `` ramped'', i.e., the th coefficient of the polynomial is , for . In matlab, we may compute Br from B via the following statement:
Br = B .* [0:M]; % Compute ramped B polynomialThe group delay of an FIR filter can now be written as
D = real(fft(Br) ./ fft(B))where the fft, of course, approximates the Discrete Time Fourier Transform (DTFT). Such sampling of the frequency axis by this approximation is information-preserving whenever the number of samples (FFT length) exceeds the polynomial order . The ratio of sampled DTFTs, however, is undersampled, in general. In fact, we may have at some frequencies (``zeros on the unit circle''). The grpdelay matlab utility in §J.8 watches out for division by zero, and simply sets the group delay to zero at such frequencies. Note that the true group delay approaches infinite magnitude as either a zero or pole approaches the unit circle.
Finally, when there are both poles and zeros, we have
Straightforward differentiation yields
and this can be implemented analogous to the FIR case discussed above. However, a faster algorithm (usually) results from converting the IIR case to the FIR case:
C = conv(B,fliplr(conj(A)));It is straightforward to show (Problem 11) that
and the group delay computation thus reduces to the FIR case:
Frequency Response Analysis Problems
Frequency Response as a Ratio of DTFTs