Window Function Spectral Comparison — Mainlobe Width and Sidelobe Suppression
Visualize the spectra of rectangular, Hann, Hamming and Blackman windows. Vary the window length, signal frequency and offset to see the trade-off between spectral leakage and frequency resolution.
Parameters
Window length N
samples
Signal frequency f₀
bin
Frequency offset δ
bin
Window type (0=Rect 1=Hann 2=Hamming 3=Blackman)
—
δ is the offset of the signal from an integer bin. δ=0 gives no leakage; δ=0.5 maximizes leakage and scallop loss.
Results
—
Mainlobe width (-6 dB)
—
Peak sidelobe suppression
—
Coherent gain CG
—
Peak amplitude error (at δ=0.5)
Window time waveform and spectrum
Top: window w[n] (blue) — Bottom: |W[k]| in dB (red) — Yellow dashed: signal bin f₀+δ
Theory & Key Formulas
A window function is a weighting sequence applied to a finite-length signal that tapers smoothly to zero at both ends, suppressing the spectral leakage caused by DFT periodicity. The classic windows are (n = 0, 1, …, N−1):
The window spectrum is $W[k] = \mathrm{DFT}(w[n])$. A wider mainlobe degrades frequency resolution, while lower sidelobes reduce leakage. The two are in direct trade-off.
What is the window function spectral comparison tool?
🙋
Everyone in DSP talks about window functions. Why can't we just take the FFT directly?
🎓
That's the classic trap. The FFT assumes the signal repeats every N samples. But real signals rarely fit perfectly, so the frame chops off abruptly at both ends. That discontinuity smears one true spectral line into a long tail across many bins — that's spectral leakage. Set the window type to 0 (rectangular = no window) and δ to 0.5 in the simulator, and watch how the bottom spectrum spreads out instead of staying a single sharp line.
🙋
Whoa, the rectangular spectrum is messy! Switching to 1 (Hann) cleans it up into one clean peak.
🎓
Right. A window forces the frame to fade smoothly to zero at both ends, eliminating the discontinuity. Hann does this with a cosine curve that hits zero at the edges. Hamming leaves a small offset to minimize the nearest sidelobe. Blackman uses three cosine terms to crush the far sidelobes. Sweep the window-type slider 0 to 3 and watch the sidelobe heights drop step by step.
🙋
So Blackman is the best, period? Why not always use it?
🎓
Because it's a trade-off. Lower sidelobes always come with a wider mainlobe. Check the "Mainlobe width" card: rectangular is 2 bins, Hann/Hamming are 4 bins, Blackman is 6. A wider mainlobe means worse frequency resolution — you can't separate two close-spaced tones as well. In practice, choose closer to rectangular when you need to separate nearby frequencies, and closer to Blackman when you need to spot a weak signal next to a very strong one.
🙋
What does the "Peak amplitude error" card mean? It says -1.42 dB at δ=0.5.
🎓
That's scallop loss. When the signal frequency falls exactly between two bins (δ=0.5), the mainlobe peak misses the bin center, so the measured peak is lower than the true amplitude. For Hann it can drop by up to 1.42 dB. If amplitude accuracy matters — sound level meters, vibration tests — you need a correction, or you switch to a FlatTop window. Move δ between 0 and 0.5 and watch the amplitude-error card update.
Frequently asked questions
The Kaiser window has a single shape parameter β (or α) that tunes the window continuously from a narrow-mainlobe / high-sidelobe regime close to rectangular all the way to sidelobes lower than Blackman, making it an "adjustable" window. The FlatTop window intentionally flattens the mainlobe top so that scallop loss falls below 0.01 dB; it is the standard choice for amplitude calibration on FFT analyzers and sound level meters, at the cost of a very wide mainlobe (about 9 bins).
δ=0 means the signal frequency falls exactly on a DFT bin, so there is no spectral leakage and no scallop loss (apart from the intrinsic sidelobes of the window). δ=0.5 places the signal exactly between two bins, which is the worst case for both leakage and peak amplitude error. In real signals δ is unknown, so the rule of thumb is to pick a window that meets the spec in the worst case, δ=0.5.
At N=1024 the direct DFT takes about 1 million operations; an FFT (Cooley-Tukey) takes about 10,000 — roughly 100x faster. At N=1,048,576 the DFT needs 1 trillion operations and the FFT about 20 million — a 50,000x difference. This tool uses the direct DFT because N stays below 1024 and the code is more readable, but real-time or large-scale analysis must use an FFT library function (numpy.fft.fft, MATLAB fft, SciPy, FFTW).
Yes. The exact same idea is applied in space: 2-D image FFTs (texture and X-ray diffraction analysis), antenna-array beamforming, radar pulse compression, and aperture optics all use windowing wherever a finite-length Fourier transform appears. The trade-off — mainlobe width versus sidelobe level — is identical.
Real-world applications
Acoustic and vibration analysis: Machine condition monitoring and noise measurement use FFT analysis of vibration signals to extract frequency components. Hann and Hamming are common defaults; FlatTop is chosen when amplitude calibration matters. On rotating machinery where speed varies, δ also varies, making leakage-resistant window selection essential.
Communication signal processing: OFDM modulation (Wi-Fi, 4G/5G, digital TV) applies windowing at the transmitter to suppress out-of-band emissions and inter-carrier interference. Raised-cosine (RC) and root-raised-cosine (RRC) windows are used to control adjacent-channel leakage.
Seismology and geophysics: Seismic-wave spectral analysis and microseismic monitoring split long records into short frames and apply the FFT (the STFT). If window sidelobes mask small late-arriving waves, the diagnosis is wrong, so Blackman and Kaiser are preferred.
Radar and sonar: In pulse-compression radar and active sonar, the trade-off between range resolution (mainlobe width) and weak-target detectability (sidelobe level) directly determines performance. Specialized windows such as Taylor and Chebyshev, which level the sidelobes at a specified suppression, are used to avoid missing weak targets near strong ones.
Common misconceptions and pitfalls
The most common misconception is that "applying a window distorts the signal, so it must be wrong". Yes, windowing modifies the time-domain waveform, but this is the best response to the unavoidable fact that the frame length is finite. Not modifying it — applying no window — is equivalent to applying a rectangular window, which has the worst leakage of all. The choice "no window" does not exist; what exists is "rectangular window".
Next, the assumption that "a window with lower sidelobes is always better" is wrong. Lower sidelobes mean a wider mainlobe and worse frequency resolution. Sweep the window-type slider from 0 to 3 in the simulator and watch the mainlobe in the bottom panel get wider and rounder as sidelobes drop. Choose based on the job — frequency resolution or low leakage.
Finally, do not forget the coherent-gain correction when reading amplitudes. With a Hann window the spectrum peak is half of the true signal amplitude (CG=0.5). To read amplitudes from a windowed spectrum, divide the peak by CG. If δ ≠ 0 you also need to correct for scallop loss (or switch to a FlatTop window where the correction drops below 1%). Once you use a window, you are reading "window-compensated values", not the spectrum itself.