Spectral Leakage

The primary cause of spectral leakage. When the sampling window does not contain an exact integer number of cycles of a given frequency component, the DFT assumes a discontinuity at the boundary. This discontinuity introduces broadband energy that spreads into adjacent bins.

• A 1 kHz sine wave sampled at 8 kHz with a 1024-point FFT (exactly 128 cycles) produces a single clean bin
• The same signal with a 1000-point FFT (125 cycles) smears energy across dozens of bins
• The sidelobe structure follows a sinc function pattern, decaying at only 6 dB per octave for a rectangular window

The standard mitigation technique for spectral leakage. A window function smoothly tapers the signal to zero at the boundaries of the sampling interval, eliminating the artificial discontinuity that causes leakage.

• Hann window: 18 dB/octave roll-off, -31.5 dB first sidelobe — good general-purpose choice
• Hamming window: 6 dB/octave roll-off, -43 dB first sidelobe — minimizes nearest-sidelobe interference
• Blackman-Harris: 30 dB/octave roll-off, -92 dB first sidelobe — extreme dynamic range applications
• Kaiser window: Adjustable β parameter trades mainlobe width for sidelobe suppression

The trade-off is always mainlobe broadening, which reduces frequency resolution.

A direct consequence of spectral leakage where a signal's apparent amplitude varies depending on its position relative to the center of a frequency bin. When a tone falls exactly halfway between two bins, the measured amplitude can be reduced by up to 3.92 dB with a rectangular window.

• Maximum scalloping loss occurs at bin center offsets of ±0.5 bins
• Flat-top windows are specifically designed to minimize this effect, reducing scalloping loss to < 0.01 dB
• Critical for applications requiring accurate amplitude measurements rather than frequency resolution
• Spectrum analyzers often apply flat-top windows for power measurements and Hann windows for spectral surveys

Coherent sampling eliminates spectral leakage entirely by ensuring the sampling clock and signal frequency are phase-locked, guaranteeing an integer number of cycles in the capture window.

• Requires frequency synthesis where fs/fin = N/M (integer ratio)
• Common in ADC testing and precision measurement systems
• Incoherent sampling is the default in real-world spectrum monitoring where signals are unknown and asynchronous
• In wideband spectrum awareness, incoherent sampling is unavoidable — windowing is mandatory
• Coherent gain is the sum of window coefficients; incoherent gain accounts for the window's noise bandwidth

Zero-padding does not reduce spectral leakage — it only interpolates the existing DFT spectrum to reveal the underlying continuous Fourier transform more finely.

• Appending zeros increases the DFT length N, reducing bin spacing from fs/N to fs/N_padded
• The sidelobe structure remains unchanged because the windowed signal's energy distribution is already fixed
• Useful for resolving closely spaced tones that would otherwise fall in the same bin
• Often combined with peak interpolation algorithms (quadratic, Gaussian) to estimate true frequency between bins
• Computational cost increases with N log N, but modern FFT libraries handle large zero-padded transforms efficiently

In polyphase filter bank channelizers, spectral leakage manifests as aliasing between adjacent sub-bands if the prototype filter's stopband attenuation is insufficient.

• Each channel's filter must suppress energy from neighboring channels by 80–100 dB for high-dynamic-range applications
• Oversampled channelizers (output rate > 2× channel bandwidth) relax filter requirements at the cost of higher throughput
• Leakage between channels creates false detections in spectrum sensing — a weak signal in one channel appears as energy in adjacent channels
• Modern wideband SIGINT systems use weighted overlap-add (WOLA) structures to dynamically adjust windowing per channel
• The filter bank's reconstruction error quantifies total leakage when channels are recombined

The primary mitigation technique for spectral leakage. A window function is a mathematical weighting applied to a finite time-domain sample block before the FFT to taper the signal to zero at the boundaries, reducing the discontinuities that cause leakage.

• Hann Window: Excellent frequency resolution and side-lobe roll-off, ideal for general spectral analysis.
• Hamming Window: Minimizes the nearest side-lobe level, useful for separating closely spaced signals of different amplitudes.
• Blackman-Harris Window: Maximizes dynamic range by drastically suppressing side-lobes at the cost of a wider main lobe.
• Kaiser Window: Offers a tunable parameter (beta) to trade off between main-lobe width and side-lobe level.

The DFT assumes the input signal is exactly one period of an infinitely repeating sequence. Spectral leakage occurs when this assumption is violated.

• Implicit Periodicity: The DFT treats the finite input as a single period of a periodic signal. A non-integer number of cycles creates a sharp discontinuity at the block boundary.
• Frequency Bins: The DFT output is a set of discrete frequency bins spaced by fs / N. A signal frequency falling between these bins (a non-coherent signal) will have its energy smeared across multiple adjacent bins.
• Coherent Sampling: Sampling exactly an integer number of signal cycles eliminates leakage entirely, but is rarely achievable in real-world wideband monitoring.

A direct consequence of spectral leakage that affects amplitude accuracy. Scalloping loss is the maximum reduction in a signal's measured magnitude when its frequency falls exactly halfway between two DFT bins.

• Worst-Case Amplitude Error: For a rectangular window (no windowing), scalloping loss can reach -3.92 dB.
• Flat-Top Windows: Specialized windows like the ISO 7165-1 flat-top are designed to minimize scalloping loss to < 0.01 dB, making them essential for precision amplitude measurements rather than frequency resolution.
• Impact on Wideband Monitoring: In dynamic spectrum awareness, scalloping loss can cause a weak signal to be missed if its frequency falls between bins and the amplitude drops below the detection threshold.

A post-processing technique that appends zeros to the end of the time-domain data block before the FFT. While it does not reduce spectral leakage, it interpolates the spectrum to provide a smoother visual representation.

• Mechanism: Zero-padding increases the DFT length N, which decreases the spacing between frequency bins, effectively computing the Discrete-Time Fourier Transform (DTFT) at more points.
• Spectral Interpolation: Reveals the true shape of the leakage pattern (the window's side-lobes) and helps resolve closely spaced peaks that might merge in a coarse bin grid.
• No New Information: It does not improve the fundamental ability to separate two closely spaced signals; that is determined by the original window's main-lobe width.

A key metric for selecting a window function, describing how quickly the amplitude of leakage side-lobes decreases as frequency offset from the main lobe increases.

• dB per Octave: Measured as the rate of side-lobe amplitude decay. The Hann window rolls off at -18 dB/octave, while a rectangular window rolls off at only -6 dB/octave.
• Dynamic Range Impact: A slow roll-off means a strong signal can mask a much weaker adjacent signal far from its center frequency, limiting the spurious-free dynamic range (SFDR) of the spectrum analyzer.
• Wideband Context: In a wideband spectrogram, poor side-lobe roll-off from a strong narrowband emitter can obscure the presence of low-power wideband signals like spread-spectrum transmissions.

A technique used in real-time spectrogram generation where consecutive FFT blocks overlap in time to reduce the temporal smearing caused by windowing.

• Overlap Percentage: Common values are 50% or 75%. A 50% overlap with a Hann window provides a constant-amplitude summation, meaning no signal events are lost in the gaps between windows.
• Mitigating Window Attenuation: Windowing attenuates data at block edges. Overlap ensures that data attenuated in one block is near the center of the next, preserving all signal energy.
• Computational Cost: Increases the FFT processing rate proportionally to the overlap factor, demanding higher throughput from FPGA-based polyphase filter banks or FFT cores.