Short-Time Fourier Transform (STFT)

The STFT operates by multiplying the signal by a window function of finite duration before applying the Fourier transform. This isolates a local section of the signal, assuming quasi-stationarity within that short interval. The window slides along the time axis, producing a sequence of spectra that reveals how frequency content evolves. Common windows include Hann, Hamming, and Gaussian; the choice of window dictates side-lobe suppression and spectral leakage characteristics.

A fundamental physical limit governs STFT resolution: time and frequency localization cannot be simultaneously arbitrarily precise. The product of the time resolution (Δt) and frequency resolution (Δf) has a strict lower bound:

• Narrow window: Good time resolution, poor frequency resolution.
• Wide window: Good frequency resolution, poor time resolution. This trade-off is intrinsic and cannot be circumvented by changing the window shape; it is a property of the Fourier transform itself.

Once a window length is selected, the STFT analyzes the entire signal with a constant resolution across all frequencies. This is a critical distinction from multi-resolution techniques like the Continuous Wavelet Transform (CWT). In the STFT, the time-frequency plane is tiled with rectangles of identical dimensions. This fixed tiling is suboptimal for signals containing both short, high-frequency bursts and long, low-frequency drifts simultaneously.

The squared magnitude of the STFT, denoted |X(t, f)|², is called the spectrogram. It is a real-valued, non-negative energy distribution that visualizes how signal power is distributed over time and frequency. The spectrogram discards phase information, providing an intuitive, though resolution-limited, view of the signal's spectral dynamics. It is widely used in audio processing, biomedical signal analysis, and RF fingerprinting.

The STFT is an invertible transform, provided the window function satisfies the constant overlap-add (COLA) constraint. By overlapping windowed segments (typically 50% or 75%) and summing the inverse transforms, the original time-domain signal can be perfectly reconstructed. This property is essential for applications like noise reduction and source separation, where processing is performed in the time-frequency domain before resynthesis.

In practice, the STFT is computed using the Fast Fourier Transform (FFT) on discrete, zero-padded windowed frames. The hop size—the number of samples the window advances between successive frames—controls the temporal density of the output. A smaller hop size yields a smoother, over-sampled time-frequency representation at increased computational cost, while a hop size equal to the window length produces a critically sampled, block-based analysis.

The visual output of the STFT, representing the squared magnitude of the transform. It displays signal energy as a function of time and frequency on a two-dimensional heatmap. In RF fingerprinting, the spectrogram serves as the primary visual input for convolutional neural networks, allowing models to learn transient and steady-state patterns. The resolution of the spectrogram is fundamentally governed by the Heisenberg-Gabor uncertainty principle.

An alternative to the STFT that provides a multi-resolution analysis. Unlike the fixed window of the STFT, the CWT uses a scaling wavelet that dilates for low frequencies and contracts for high frequencies. This results in superior time resolution for high-frequency transients and better frequency resolution for low-frequency components, making it ideal for analyzing the turn-on transients of RF emitters.

A quadratic time-frequency distribution offering the highest possible joint resolution. It computes the Fourier transform of the signal's instantaneous autocorrelation. While it perfectly localizes linear chirps, it suffers from severe cross-term interference when analyzing multi-component signals, creating spurious artifacts that obscure the true signal structure. Cohen's class distributions were developed to suppress these artifacts.

A specific case of the STFT that uses a Gaussian window function. The Gaussian window is unique because it achieves the theoretical lower bound of the uncertainty principle, providing the optimal trade-off between time and frequency localization. This transform is named after Dennis Gabor, who first proposed the analysis of signals in time-frequency atoms, laying the groundwork for the STFT.

A post-processing reassignment technique applied to the STFT or CWT. The SST sharpens a blurry time-frequency representation by reallocating the transform coefficients along the frequency axis based on an instantaneous frequency estimate. This concentrates the smeared energy onto the true time-frequency ridges, significantly improving the readability of the representation without violating the uncertainty principle.

A variation where frequency bins are geometrically spaced and the window length varies inversely with frequency. This maintains a constant Q-factor (ratio of center frequency to bandwidth). The CQT mirrors the human auditory system and is highly efficient for analyzing harmonic signals where spectral content follows a logarithmic scale, often used in audio processing and music analysis.