### Does zero-padding a FFT matter?

Zero-padding a Fast Fourier Transform (FFT) can increase the resolution of the frequency domain results (see FFT Zero Padding). This is useful when you are looking to determine something like a dominant frequency over a narrow band with limited data.

Here, I examine the case of a slow-wave (<4 Hz) signal. I am detecting the dominant frequency as determined by the power spectrum, and then examining how the phase changes as the FFT is zero-padded by a factor of *n*, where *n*=1 is no zero-padding.

## Methods & Results

I used 10,000 trials of slow-wave data that were recorded at 250 Hz (each trial was 1,024 samples). These data were subject to a slow-wave IIR elliptical filter. For each trial, I padded the data by a factor *n*; if *n*=2, the 1,024 zeros are added to the original signal. Below shows a single trial and how changing *n* ‘smooths’ the power spectrum (middle) leading to a more accurate assessment of phase (bottom).

The plot below shows how different each pad factor was to the last pad iteration, wrapped to pi. This assumes the last pad iteration (e.g., *n*=10) is the most accurate estimation of the actual phase of the signal at the dominant frequency.

## Interpretation & Use

The error in the phase of the FFT decreases as padding increases (i.e. converges), as expected, where *n*=10 is the trivial case and equals 0. The degree to which this matters in application-dependent—I don’t think a p-value really helps the interpretation, especially because these data are particular to the slow-wave use case. However, it is helpful to know *how much* of the data lies *how far* away.

This plot says that ~70% data have an ‘error’ greater than ±π/8 without any zero-padding (*n*=1) and decreases to ~50% for *n*=2. The practical question is really how much error is acceptable to *your* system. As put in MATLAB – FFT and Zero Padding, if you have the choice between zero-padding and including more data, always chose more data.

## Recent Comments

## Archives

- April 2023
- January 2023
- November 2022
- May 2022
- March 2022
- January 2022
- December 2021
- April 2021
- December 2020
- October 2020
- August 2020
- July 2020
- March 2020
- February 2020
- January 2020
- December 2019
- November 2019
- October 2019
- January 2019
- December 2018
- November 2018
- August 2018
- July 2018
- April 2018
- March 2018
- November 2017
- October 2017
- February 2017
- October 2016
- August 2016
- July 2016
- November 2015
- October 2013
- February 2013
- January 2013
- August 2012
- July 2012
- June 2012
- May 2012
- April 2012
- February 2012
- December 2011