Skip to content

Reading a spectrum: coherent sampling and FFT leakage

Demonstration content

This is an educational synthetic experiment, not a measured RF result or a claim of professional project experience. Author identity and repository links are pending.

The engineering question

Why does an FFT show energy in many bins when the input contains only one tone? A finite observation does not necessarily contain an integer number of signal periods. Its periodic extension can have a discontinuity, spreading energy across the discrete frequency grid.

Sampling model

For a complex sinusoid sampled at \(f_s\), let

\[ x[n] = e^{j2\pi f_0 n/f_s},\qquad n=0,\ldots,N-1. \]

The DFT evaluates the spectrum at \(f_k=kf_s/N\):

\[ X[k]=\sum_{n=0}^{N-1}x[n]e^{-j2\pi kn/N}. \]

If \(f_0N/f_s\) is an integer, the rectangular-window DFT places the tone in one bin (up to floating-point error). Otherwise the sampled Dirichlet kernel distributes it across bins. Zero padding interpolates that spectrum; it does not improve the physical resolution set by observation time.

Experiment architecture

flowchart LR
    A[Synthetic complex tone] --> B[Finite observation]
    B --> C[Rectangular or Hann window]
    C --> D[DFT]
    D --> E[Normalize by window sum]
    E --> F[Compare spectral leakage]

Reproduce the figure

Download the Python experiment. It uses the standard library only. Run from the repository root:

python docs/experiments/python/spectrum.py --check --output docs/assets/figures/leakage.svg

The experiment uses \(N=128\), \(f_s=128\) Hz, and tones at 12 Hz and 12.5 Hz. It computes a direct DFT so that the implementation is easy to inspect; this is not an efficient replacement for an FFT library.

spectrum = [abs(sum(x[n] * w[n] * cmath.exp(-2j * math.pi * k * n / N)
                    for n in range(N))) / sum(w)
            for k in range(N)]

DFT magnitude comparing bin-aligned and half-bin tones with rectangular and Hann windows

Interpreting the result

The bin-aligned tone has unit amplitude at bin 12. A half-bin offset produces two dominant bins and visible sidelobes. The periodic Hann window suppresses distant sidelobes while widening the main lobe. The normalization by the window sum corrects coherent gain for a bin-centred tone; it does not remove scalloping loss for a tone between bins.

Decision Benefit Cost
Rectangular window Narrow main lobe Higher sidelobes for off-bin tones
Hann window Lower distant sidelobes Wider main lobe
Longer observation Finer bin spacing More latency and memory

Validation and limits

The script asserts the bin-centred peak and numerical off-bin residual, and checks that the Hann window reduces distant-bin energy for the half-bin case. These are synthetic correctness checks. No RF hardware, noise, quantization, Doppler dynamics or execution-time benchmarks are included.

A C++17 companion verifies coherent-bin accumulation with the same sample count. It is a correctness illustration, not an optimized embedded implementation.

References

Reproducible spectral analysis bench

Source repository: [OWNER/REPOSITORY — configure before publication] (placeholder).

Share on LinkedIn ↗