Let's now apply the above technique to the design of an octave filter
first sight, this appears to be a natural fit, because it is
immediately easy to partition a power of 2 (typical for the FFT size
N) into octaves, each having a width in bins that is a power of 2. For
example, when N == 8, we have the following stack of
frequency-response vectors for the rectangular-window, no-zero-padding,
Fast Octave Filter Banks
H = [ ... 0 0 0 0 1 1 1 1 ; ... 0 0 1 1 0 0 0 0 ; ... 0 1 0 0 0 0 0 0 ; ... 1 0 0 0 0 0 0 0 ];Figure 10.31 depicts the resulting filter bank schematically in the frequency domain.11.20 Thus, H(1,:) is the frequency-response for the top (first) octave, H(2,:) is the frequency-response for the next-to-top (second) octave, H(3,:) is the next octave down, and H(4,:) is the ``remainder'' of the spectrum. (In every octave filter bank, there is a final low-frequency band.
10.30. The division by N called for by ifft(N) is often spread out among the ifft butterfly stages as divisions by 2 (one-bit right-shifts in fixed-point arithmetic) after each of them. For conservation of dynamic range, half of the right-shifts can be spread out among every other forward-FFT butterfly stage  as well. The simple spectral partitioning of Fig.10.31 is appropriate for complex signals--those for which the spectrum is regarded as spanning 0 Hz to the sampling rate. For real signals, we need a spectral partition more like that in Fig.10.32. 10.32 is 14--not a power of 2. (The previous length 8 complex case maps to length 14 real case because the dc and Nyquist-limit samples do not have complex-conjugate counterparts.) Discarding the sample at the Nyquist limit--number 8 in Fig.10.32--does not help, since that gives 13 samples--still not a power of 2. In summary, there is no obvious way to octave-partition the spectral samples of a real signal while maintaining the power-of-2 condition for each band and symmetrically partitioning positive and negative frequencies. A real signal can of course be converted to its corresponding ``analytic signal'' by filtering out the negative-frequency components. This is normally done by designing a Hilbert transform filter [133,224]. However, such filters are large-order FIR filters, exactly like we are trying to design! If we design our filter bank properly, we can use it to zero the negative-frequency components.
Spectral Rotation of Real Signals
Improving the Channel Filters