I have to process a RF signal band with bandwith between 75MHz and 100MHz to divide it in up to 8 contiguous subbands, each one with a bandwith programmable between 50kHz and 25MHz, in 25kHz steps.

The objective is to have the capability to attenuate each sub-band differently.

Transition skirts should be aprox. 6dB at 25kHz, on-band ripple less than 0.5dB and out-of-band attenuation better than -85dB at > 500kHz of the transition frequencies.
I would prefer a linear phase filter to ensure a constant group delay (constant, but as low as possible).

My first thoughts were to downconvert the RF signal band to a lower band (or to 0Hz to create a complex I+Q signal), then use 8 separate FIR filters, and finally upconvert the summing result. However, I guess that the contiguity between adjacent filters may lead to some ways to reduce the computation cost. May you give me some suggestions?


This sounds like the classical job of a polyphase filterbank, and a cascade of polyphase synthesizers. Yes, do this in complex baseband.

First, you use your filterbank to divide the 25 MHz in 25 kHz wide channels. Then, you use polyphase synthesizers to combine subsets these channels back to the channel widths you need, and then you use another one to combine these back to 25 MHz of bandwidth.

Another option would be to simply have a prepared table of filter coefficients for your potential subband widths, and just shift these in frequency (which is rather trivial) accordingly to get out your subbands of choice, then combine these back together with a bog-normal adder.

I'm assuming none of this requires that group delay in all these subbands is identical, for example. Otherwise, the design approach would have to amount for different filter delays.

  • $\begingroup$ Thanks for your answer, Marcus. I agree with your solution, but my interest is about reducing hardware requirements. With an input band of 100Mhz I should need to generate 4000 channels of 25kHz, and then combine many of them to recreate bigger subbands. I was looking for a way to avoid divide-then-combine operations. Maybe some hardware may be simplified, don't you think? By the way, what means "bog-normal adder"? $\endgroup$ Mar 14 '18 at 13:51
  • $\begingroup$ You don't have an input bandwidth of 100 MHz, you've got 25 MHz: as said, use equivalent complex baseband, so that you don't have to sample a 100 MHz signal. Using a PFB channelizer is already a highly optimized method, so no, I don't think you'll find something that will be smaller or faster. With "bog-normal adder" I meant an adder. $\endgroup$ Mar 14 '18 at 14:11

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.