# How to rebin FFT results

I converted time domain waveform signal to frequency domain using FFT, The bin size was 1 HZ. We saved the data but now for different analysis we need to change the binning to 100Hz. Can I combine all 100 frequencies in each bin by adding them up as sum of magnitudes divided by frequencies ? Do I do this for both the real and imaginary parts ?

• explicitly, why are you "rebinning"? Is the size of your FFT being changed? Sep 19, 2023 at 23:57
• It is many MGz of data for each waveform and there are many thousands of them. Some algorithms that use it do not need that resolution so we are trying to shrink the data traveling over the network. Sep 20, 2023 at 0:00
• so you have the same sample rate, but you are changing the window size? Sep 20, 2023 at 0:01
• Correct, let's say we have 100 millions of datapoints over 1 second. The data size from FFT with window size 1 Hz is 50mln and it is saved on the server. If the clients request this data with 50 Hz windows size was want to combine the data on the server and and return only 5 mln data points instead of 50 mln. Sep 20, 2023 at 0:07
• Or, like you said, sum bins 1to50, 51to100, 101to150 etc but whoever’s asking should know that you’re providing them with the sum of magnitudes in these frequency bands.
– Jdip
Sep 20, 2023 at 1:23

If, however, a power spectral density is desired, in which the result is only a magnitude spectrum with no phase, then all samples could be used such that groups of 100 bins associates with each 100 Hz sub-band should be sum-squared (for a power qty and using $$10Log_{10}()$$ to express in dB) or root-sum-squared (for a magnitude qty and using $$20Log_{10}()$$ to express in dB). The squaring for the complex FFT samples should be complex conjugate products, which results in being the square of the absolute magnitudes for each frequency bin. But even for this purpose (unless the ability to process the inverse FFT of the original sequence was not feasible for some reason), for the case of a power spectral density my approach would be to recover the original time domain sequence using the inverse FFT and then compute the power spectral density using the Welch Method, which comes down to taking smaller FFT’s of each shorter block in time over the entire time domain sequence and averaging the individual magnitude spectrums. This results in a power spectral density with much less variation or noise in the result as I detail in this post: https://dsp.stackexchange.com/a/87731/21048