Decimation lowers the sample rate of a digital signal by keeping only every -th sample. Before that, a lowpass filter must remove every frequency that the new rate cannot represent, otherwise those frequencies fold back on top of the signal (aliasing). Sampling and aliasing are introduced in 01 Basics, Operations and Sampling.

Downsampling

  • the new sample rate is
  • for complex samples, the band that can be represented shrinks from to

Why Aliasing Appears

  • sampling makes the spectrum repeat every
  • after downsampling, the copies are times closer together
  • any energy above the new limit lands on top of the wanted band
  • once folded, it cannot be separated from the real signal

Correct Order

  • lowpass filter first, with a cutoff a little below
  • then keep every -th sample
  • a real filter needs a transition band, so the cutoff sits slightly inside the new band edge, for example 23.5 MHz for a new MHz band

Efficient Filtering

  • only every -th filter output is kept, so there is no need to compute the others
  • a polyphase structure, or a strided convolution (for example conv1d with stride=M), computes only the outputs that are kept, which cuts the cost by

Selecting A Sub-Band

To study a band that is not centred on 0 Hz:

  • first move it to 0 Hz by multiplying by a complex exponential
  • this shifts the whole spectrum by
  • then lowpass filter and decimate
  • example: from one 100 MHz wide capture, make three 50 MHz views centred at , and MHz

FIR Design For Decimation

  • the window method with a Kaiser window is common
  • the Kaiser parameter trades transition width against stopband attenuation, and more taps give a sharper transition
  • example: 201 taps with
  • an FIR filter with symmetric taps has linear phase, so all frequencies are delayed by the same amount and the waveform shape is kept