Must the number of samples be a power of two?

No, although some lengths are faster for particular implementations.

Are FFT outputs always real?

No. Real inputs generally yield complex coefficients with conjugate symmetry.

Does the largest frequency peak prove causation?

No. It describes a feature of the analyzed record, whose explanation requires contextual evidence.

AI Glossary / Time series and signal processing / Intermediate

Fast Fourier transform

Definition

A fast Fourier transform, or FFT, is an algorithm for computing a discrete Fourier transform efficiently. It converts a finite sequence of samples into coefficients representing discrete frequency components. The FFT is a computational method, while the discrete Fourier transform is the mathematical transformation being calculated. Different FFT algorithms can calculate the same transform.

What it means

A sound recording or vibration trace is usually stored as a sequence of values over time. A frequency representation can reveal recurring oscillations that are harder to see in that sequence. Each Fourier coefficient carries magnitude and phase information, subject to the chosen convention.
An FFT makes this calculation practical for large records. It does not automatically decide which peaks are meaningful, identify their physical cause, or establish whether a component will persist in future data.

How it works

A direct discrete Fourier transform combines every input sample with every frequency index. Common FFT algorithms reuse intermediate calculations to reduce the typical operation count from order N squared to order N log N.
For an even-length input, one familiar construction separates even-indexed and odd-indexed samples, transforms those shorter sequences, and combines them using complex rotation factors. Repeating this decomposition produces efficient computation. Power-of-two lengths suit a basic radix-two algorithm, but modern libraries support many other lengths. Sampling frequency determines the mapping from coefficient index to physical frequency; normalization determines how coefficient magnitudes should be interpreted.

Three practical examples

A maintenance engineer transforms machine-vibration samples to examine frequency components associated with rotating parts. A peak supports investigation but does not by itself diagnose a fault.
A speech application uses short-time Fourier calculations as part of an audio representation. The window length trades temporal detail against frequency discrimination.
An image-processing program uses multidimensional FFTs to accelerate suitable convolution operations. It pads arrays appropriately when linear, rather than circular, convolution is intended.

Think of it this way

Think of recognizing musical notes in a chord. The time waveform is the combined sound, while a frequency representation separates contributions by oscillation rate. The FFT is the efficient calculation used to obtain that representation. Unlike a listener, it has no built-in understanding of instruments, musical intent, or whether a recording contains unwanted noise.

Worked example

Take an original four-sample sequence x = [3, 1, −1, 1], sampled at 8 samples per second. Use the forward DFT convention with a negative exponential and no forward normalization. Its frequency spacing is 8/4 = 2 Hz.
For a small radix-two calculation, the even-indexed samples [3, −1] have a two-point transform [2, 4]. The odd-indexed samples [1, 1] have transform [2, 0]. Combining them gives X0 = 2 + 2 = 4, X2 = 2 − 2 = 0, X1 = 4 + (−i) × 0 = 4, and X3 = 4 − (−i) × 0 = 4. Thus the full transform is [4, 4, 0, 4].
The mean is X0/4 = 1. The positive-frequency bin at 2 Hz has sinusoidal amplitude 2 × 4/4 = 2 under this real, bin-centered example; its conjugate partner represents −2 Hz, not a separate physical oscillation. Reconstructing 1 + 2 cos(2πn/4) returns the original samples. An independent energy check gives 3 squared + 1 squared + (−1) squared + 1 squared = 12, matching (16 + 16 + 0 + 16)/4. This exact toy record avoids leakage and is not a general amplitude-calibration recipe.

Illustrative case study

Illustrative fictional case: Beacon Pumps compares vibration spectra before and after maintenance. A technician initially interprets a taller displayed peak as a larger physical vibration, but discovers that the second recording used a different window and scaling.
The team standardizes acquisition settings, documents the window correction, and checks a known test signal before comparing equipment. It retains the original time traces because a short transient can matter even when an averaged spectrum looks ordinary. This fictional case illustrates measurement discipline, not a claim of predictive-maintenance accuracy.

Common misconceptions and mistakes

An FFT does not create additional observed information. Zero-padding makes the displayed frequency grid denser but does not replace a longer measurement when separating nearby components. A transform also cannot reverse aliasing that occurred during sampling.
Do not assume that every record fits an integer number of cycles. Finite-window boundaries can spread energy across bins, producing spectral leakage. Distinguish magnitude, power, and power spectral density; their units and scaling differ. Finally, a transform is reversible when the necessary coefficients and convention are retained, but discarding phase or selected frequencies changes that situation.

Business use

Businesses use FFTs in diagnostics, communications, audio processing, imaging, and efficient numerical computation. Their speed can reduce processing costs, especially when many transforms or large convolutions are required.
A useful specification includes sampling rate, record length, window choice, detrending, frequency scaling, and acceptable latency. Validate outputs with signals of known frequency and amplitude. Monitor whether preprocessing suppresses the very feature the downstream decision needs, and retain enough acquisition metadata to reproduce the analysis.

Check your understanding

What is the frequency-bin spacing in the example?

Answer: Two hertz.

Why is the mean 1 rather than 4?

Answer: The unnormalized zero-frequency coefficient is the sum of four samples; dividing by four gives the mean.

Would doubling the record by appending zeros double the observation duration?

Answer: No. The additional values are computational padding, not new measurements.

Related terms

Learn next

calculate a small DFT directly and compare it with an FFT result. Then vary record length and window choice using a known sinusoid, keeping amplitude normalization explicit. Study sampling and aliasing before interpreting field recordings.

Sources

Sources: SciPy, official Discrete Fourier Transforms tutorial: https://docs.scipy.org/doc/scipy/tutorial/fft.html . FFTW, What FFTW Really Computes, official transform conventions: https://www.fftw.org/fftw3_doc/What-FFTW-Really-Computes.html . Primary documentation consulted. The four-sample calculation, independent energy check, and fictional case were created for this entry.

Previous: Fast decay Next: Fast marching method