A cascade of recursive (i.e. feedback) comb filters at subharmonics of a desired frequency should be able to provide a high-Q bandpass filter at very low computational cost. For example, to isolate a 256-Hz signal with a sampling rate of 1024 Hz, you can subtract the sample 2, 6, 10, 14, 18, or 22 samples ago. Suppose you start by subtracting the sample X[i-14] 14 samples ago, which is 3½ cycles; this produces a new signal A. Now add the sample 16 samples ago from that new signal A[i] + A[i-16]; this produces a new signal B. Now subtract the sample 18 samples ago from that new signal B[i] - B[i-18]. This produces a new signal C. Now add the sample 20 samples ago C[i] + C[i-20], producing a new signal D.
This signal C is the input signal filtered with the product of the frequency responses of the three component filters; that filter has exact nulls at every place any one of the filters had an exact null, including DC, and it has peaks at multiples of the least common multiple of their periods, which I think is well above the Nyquist frequency in this case.
Wait, I’m confusing feedforward and feedback implementations.
Hmm, with unit impulse amplitude this ends up being a linearly growing signal if you make it feedback. You probably need to apply some kind of Hogenauer-style limit thing to keep its amplitude below a limit.