STFT Formalism in Time–Frequency Analysis

• STFT is a time–frequency analysis tool that uses windowed spectral decompositions to characterize non-stationary signals with an exact inversion framework.
• Adaptive STFT variants optimize window parameters to balance temporal and spectral resolution for complex, dynamic signal structures.
• Extensions such as synchrosqueezing, finite-dimensional formulations, and operator-theoretic approaches enhance reconstruction accuracy and provide advanced analytic flexibility.

The short-time Fourier transform (STFT) is a foundational tool for localized time–frequency analysis of signals, enabling precise characterization of non-stationary structures via windowed spectral decompositions. Through flexible parameterization of the analysis window and its temporal shifts, the STFT provides a continuous or discrete mapping of signals onto joint time–frequency domains, with exact inversion and a mathematically well-characterized trade-off between temporal and spectral resolution. In advanced variants, learnable or adaptive STFT parameterizations, operator-theoretical generalizations, phase-distribution analysis, finite-dimensional toric settings, and connections to reassignment techniques such as synchrosqueezing extend both the theoretical power and practical impact of the formalism.

1. Mathematical Definition and Basic Properties

Let $x:\mathbb{R}\to\mathbb{C}$ be a signal, and $g:\mathbb{R}\to\mathbb{C}$ a window function, smooth and rapidly decaying, with $\|g\|_2^2 > 0$. The continuous STFT is

$$\mathrm{STFT}_x(t, \omega) = \int_{-\infty}^{\infty} x(\tau)\,g(\tau - t) \,e^{-i\omega \tau} \,d\tau$$

where $t$ is the analysis time and $\omega$ is the angular frequency. In discrete time for a signal $x[n]$, window $w[n]$, DFT length $N$, hop size $H$, and frame index $g:\mathbb{R}\to\mathbb{C}$0, one has

$g:\mathbb{R}\to\mathbb{C}$1

with frequency bin $g:\mathbb{R}\to\mathbb{C}$2 (Abdalla, 2023, Zhao et al., 2020, Leiber et al., 26 Jun 2025).

Key linearity and covariance properties include:

• Linearity: $g:\mathbb{R}\to\mathbb{C}$3
• Time shift: $g:\mathbb{R}\to\mathbb{C}$4
• Frequency shift: $g:\mathbb{R}\to\mathbb{C}$5
• Energy preservation (Moyal's formula): $g:\mathbb{R}\to\mathbb{C}$6 (Abdalla, 2023).

Exact signal reconstruction is given by the overlap–add formula: $g:\mathbb{R}\to\mathbb{C}$7

2. Time–Frequency Resolution and Window Trade-Offs

The core limitation and flexibility of the STFT lies in the time–frequency uncertainty trade-off. Given window $g:\mathbb{R}\to\mathbb{C}$8,

$g:\mathbb{R}\to\mathbb{C}$9

with $\|g\|_2^2 > 0$0 (Heisenberg–Gabor uncertainty) (Abdalla, 2023).

A narrow window provides high temporal, but poor frequency, resolution; a wider window improves frequency localization at the expense of smearing time details. In practical scenarios, window choice (e.g., Gaussian, Hamming, Slepian) and support may be static or adapted to the local signal structure, or even learned as real-valued differentiable parameters (Zhao et al., 2020, Leiber et al., 26 Jun 2025, Li et al., 2018).

Adaptive and quilted STFT approaches allow time and/or frequency-dependent window parameters to locally optimize concentration or separation in the time–frequency domain, providing superiority for signals with non-stationary or highly dynamic features (Berrian et al., 2017, Li et al., 2018).

3. Inversion, Discretization, and Learnable Parameterization

In both theory and practice, effective use of the STFT depends on well-designed analysis–synthesis pairs and window/hop settings. For discrete-time STFT

$\|g\|_2^2 > 0$1

reconstruction is obtained by

$\|g\|_2^2 > 0$2

where $\|g\|_2^2 > 0$3 is a normalization constant determined by the window overlap–add condition

$\|g\|_2^2 > 0$4

(Zhao et al., 2020).

Modern approaches treat window length, shape, and hop size as real-valued and differentiable parameters, enabling direct optimization via gradient-based methods (Leiber et al., 26 Jun 2025). This formalism is compatible with arbitrary differentiable cost functions, seamless backpropagation, and integration of the DSTFT as the initial layer in neural networks, facilitating joint parameter and network weight learning. Resulting time–frequency representations can thus be tailored for concentration, sparsity, classification, or accuracy within downstream tasks, eliminating the need for computationally expensive discrete hyperparameter searches (Zhao et al., 2020, Leiber et al., 26 Jun 2025).

4. Phase Structure and Statistical Properties

A classical modeling assumption posits that the phase $\|g\|_2^2 > 0$5 of STFT coefficients $\|g\|_2^2 > 0$6 is uniformly distributed. Recent analysis demonstrates that this global uniformity assumption is misleading: the phase distribution per frequency bin or magnitude stratum is generally nonuniform and exhibits systematic lobe structures, fundamentally tied to the window type and spectral localization properties (Voran, 2024).

Table: Manifestations and origins of nonuniform phase in STFT

Factor	Manifestation	Governing Mechanism
Frequency bin	2- or 4-lobed phase histograms	Nonlinear mapping θ→φ_k
Magnitude range	Pronounced lobed patterns at low	Tone-induced concentration
Window shape	Strength of lobes varies	Sidelobe suppression

For rectangular windows, the phase distribution is analytically determined (see Eq. 26 in (Voran, 2024)), with explicit peak locations

$\|g\|_2^2 > 0$7

where $\|g\|_2^2 > 0$8 reduces to $\|g\|_2^2 > 0$9. Move toward Hann or Hamming windows increases uniformity, but measurable nonuniformity persists at many bins. These structures impact quantization, statistical modeling, and audio perception, and should be leveraged as per-frequency or per-magnitude priors in STFT-based algorithms for optimal performance (Voran, 2024).

5. Operator-Theoretic and Clifford Generalizations

The STFT supports a rich operator-theoretic formalism: given analysis and synthesis windows $$\mathrm{STFT}_x(t, \omega) = \int_{-\infty}^{\infty} x(\tau)\,g(\tau - t) \,e^{-i\omega \tau} \,d\tau$$0 and a symbol $$\mathrm{STFT}_x(t, \omega) = \int_{-\infty}^{\infty} x(\tau)\,g(\tau - t) \,e^{-i\omega \tau} \,d\tau$$1, the localization (anti-Wick) operator

$$\mathrm{STFT}_x(t, \omega) = \int_{-\infty}^{\infty} x(\tau)\,g(\tau - t) \,e^{-i\omega \tau} \,d\tau$$2

smooths the symbol via convolution with the window correlation function and never equals a true Fourier multiplier except in degenerate window cases. This smoothing effect has consequences for continuity and $$\mathrm{STFT}_x(t, \omega) = \int_{-\infty}^{\infty} x(\tau)\,g(\tau - t) \,e^{-i\omega \tau} \,d\tau$$3 bounds of the associated operators. In the discrete setting, the theory finds analogues in Gabor multipliers versus LTI filters, with exact equivalence only under restrictive conditions (Balazs et al., 2022).

Clifford extensions of the STFT (CSTFT) on $$\mathrm{STFT}_x(t, \omega) = \int_{-\infty}^{\infty} x(\tau)\,g(\tau - t) \,e^{-i\omega \tau} \,d\tau$$4 with even $$\mathrm{STFT}_x(t, \omega) = \int_{-\infty}^{\infty} x(\tau)\,g(\tau - t) \,e^{-i\omega \tau} \,d\tau$$5, via the Clifford–Fourier transform kernel $$\mathrm{STFT}_x(t, \omega) = \int_{-\infty}^{\infty} x(\tau)\,g(\tau - t) \,e^{-i\omega \tau} \,d\tau$$6, preserve orthogonality, inversion, and reproducing kernel properties, and satisfy analysis-specific uncertainty principles, with explicit polynomial-growth bounds in the phase space (Martino, 2021).

6. Finite-Dimensional, Toric, and Frame-Theoretic Settings

On finite-dimensional settings, the STFT can be consistently extended onto the flat torus $$\mathrm{STFT}_x(t, \omega) = \int_{-\infty}^{\infty} x(\tau)\,g(\tau - t) \,e^{-i\omega \tau} \,d\tau$$7, using subspaces $$\mathrm{STFT}_x(t, \omega) = \int_{-\infty}^{\infty} x(\tau)\,g(\tau - t) \,e^{-i\omega \tau} \,d\tau$$8 of periodized delta trains. The STFT on $$\mathrm{STFT}_x(t, \omega) = \int_{-\infty}^{\infty} x(\tau)\,g(\tau - t) \,e^{-i\omega \tau} \,d\tau$$9, with window in the Feichtinger algebra, is a continuous extension of the finite discrete Gabor transform and admits a version of Moyal's formula and a toric sampling theorem for periodic Gaussian windows. For odd $t$0, every $t$1 distinct lattice points yield a full-spark Gabor frame. The formalism facilitates exact theoretical analyses of phase-space sampling, zero-detection in noisy spectrograms, and the explicit study of random analytic functions arising from noise (Abreu et al., 2022).

7. Extensions: Adaptive STFT and Synchrosqueezing

To overcome the limitations of fixed windowing, the adaptive STFT permits time- and/or frequency-varying window parameters $t$2, tailored by entropy-minimization or ridge-support separation to local instantaneous frequencies, and enabling optimized multicomponent separation (Li et al., 2018). "Quilted" (regionwise adaptive) STFTs allow window shape and length adaptation per time–frequency tile, achieving high local resolution (Berrian et al., 2017).

The synchrosqueezing transform (SST) sharpens the classical STFT by reassigning coefficients along instantaneous frequency estimates

$t$3

with the SST itself given by

$t$4

yielding highly concentrated time–frequency representations and permitting accurate inversion and mode extraction. Adaptive and 2nd-order SST variants further enhance concentration for signals with fast-varying frequencies (Abdalla, 2023, Li et al., 2018, Berrian et al., 2017).

---

In summary, the STFT formalism comprises a robust mathematical and computational infrastructure for time–frequency representation, adapts to complex signal behaviors via advanced parameterizations, supports operator-theoretic and frame-theoretic generalizations, and serves as the analytic backbone for refinement and reassignment methods such as SST, with ongoing extensions to adaptive, learnable, and finite phase-space settings.