Nonstationary Gabor Frames

• Nonstationary Gabor Frames are a generalization of classical Gabor systems that adapt window functions and lattice parameters to locally optimize time-frequency resolution.
• They employ variable windows and nonuniform parameter grids along with Walnut-type representations to ensure bounded, invertible analysis and efficient reconstruction.
• Practical applications include audio signal processing, sparse representation theory, and nonlinear approximation, where adaptive time-frequency tiling improves performance.

Nonstationary Gabor frames (NSGFs) are a broad generalization of classical Gabor frames, designed for analyzing signals where time-frequency resolution needs adaptive and local control. Unlike the stationary Gabor framework—where window functions and lattice parameters remain uniform across the signal—NSGFs admit sequences of windows, time/shifts, and modulation parameters that can vary over time or frequency. This flexibility enables locally optimal time-frequency tiling foundational to numerous applications, especially in audio signal processing, sparse representation theory, and nonlinear approximation. The mathematical foundation of NSGFs involves atomic systems parameterized by variable windows and nonuniform lattice, with frame-theoretic conditions ensuring bounded, invertible analysis and synthesis operators. The "painless" NSGF regime admits explicit inversion in Fourier or time domain, while general NSGF constructions employ Walnut-type representations and operator-theoretic perturbation and duality techniques.

1. Mathematical Formulation of Nonstationary Gabor Frames

Let $\{g_k\}_{k\in K}\subset L^2(\mathbb{R})$ be a family of window functions, and $\{a_k\}, \{b_k\}_{k\in K}$ strictly positive time and frequency shift parameters. The nonstationary Gabor system is generated as

$$\mathcal{G} = \{ M_{m\,b_k}\,T_{n\,a_k}\,g_k : k\in K,\, n,m\in \mathbb{Z} \}.$$

The frame condition requires the existence of $0 < A \leq B < \infty$ such that for all $f\in L^2(\mathbb{R})$,

$$A\,\|f\|_2^2 \leq \sum_{k\in K} \sum_{n,m\in\mathbb{Z}} |\langle f, M_{m\,b_k}T_{n\,a_k}g_k \rangle|^2 \leq B\,\|f\|_2^2.$$

Frame operator $S$ acts as

$$Sf = \sum_{k,n,m} \langle f, M_{m\,b_k}\,T_{n\,a_k}g_k \rangle M_{m\,b_k}\,T_{n\,a_k}g_k,$$

with invertibility ensured under the frame bounds. In many important applications, one seeks to tune the windows and lattice locally (for instance, shorter windows at signal transients, longer windows for stationary regions).

The associated canonical dual frame system is

$$\widetilde{\mathcal{G}} = \{ M_{m\,b_k}\,T_{n\,a_k}\,\widetilde{g}_k : \widetilde{g}_k = S^{-1}g_k \}.$$

Among various representations, the Fourier domain (under sufficient decay and compact support hypotheses) simplifies the frame operator to pointwise multiplication: $$\widehat{Sf}(\omega) = m(\omega)\,\widehat{f}(\omega), \qquad m(\omega) = \sum_{k,m} \frac{1}{|a_k|} |\widehat{g_k}(\omega - m\,b_k)|^2.$$ The canonical dual windows then satisfy $$\widehat{\widetilde{g}_k}(\omega) = \widehat{g_k}(\omega) / m(\omega)$$ (Jindal et al., 2022).

2. Existence, Painless Case, and Perturbation Theory

NSGF existence results generalize the "painless" expansion paradigm from stationary systems. If the family of windows $\{g_k\}$ satisfies uniform energy distribution and decay conditions, NSGFs can be constructed for sufficiently dense ($b_k$) sampling lattice (Dörfler et al., 2011). In the "painless" case, each window $g_k$ is compactly supported on a region of length $\leq 1 / b_k$; all off-diagonal Walnut terms vanish: $$Sf(t) = G_0(t) f(t), \qquad G_0(t) = \sum_k b_k^{-1} |g_k(t)|^2.$$ Invertibility is guaranteed when $A \leq G_0(t) \leq B$. The canonical dual atoms can then be expressed pointwise by

$$\widetilde{g}_{k,m}(t) = e^{2\pi i m b_k t} G_0(t)^{-1} g_k(t).$$

Perturbation results extend stability to non-compactly supported windows provided the deviation remains sufficiently controlled, e.g., $|g_k(t)-h_k(t)| \leq C_k(1 + |t-a_k|)^{-p_k}$, with convergence of frame bounds ensured by diagonal dominance and tail bounds (Dörfler et al., 2011).

3. Walnut Representations, Dual Frame Structures, and Operator Theory

General NSGF constructions may lack full diagonalization. Instead, the frame operator $S$ and its inverse admit Walnut-type representations: $$Sf(t) = \sum_{n,k} \omega_{n,k}(t) f(t - k b_n^{-1}),$$ where $$\omega_{n,k}(t) = b_n^{-1} g_n(t) \overline{g_n(t - k b_n^{-1})}$$ (Holighaus, 2013). The inverse $S^{-1}$ is written as an infinite sum over translated and weighted forms, with support shrinking in higher order terms. Duality criteria are given by conditions on cross-systems: $$\Omega_0(t) = \sum_n b_n^{-1} h_n(t) \overline{g_n(t)},\quad \Omega_x(t) = \sum_{(n,m):m b_n^{-1}=x} b_n^{-1} h_n(t) \overline{g_n(t-x)},$$ with the duality holding iff $\Omega_0(t) \equiv 1$ and $\Omega_x(t) \equiv 0$ for $x \ne 0$. These conditions can be both sufficient and necessary under regularity assumptions (Holighaus, 2013).

The Neumann series allows for operator-norm approximation of inverses,

$$S^{-1} = \frac{2}{A+B} \sum_{\ell=0}^\infty \left( I - \frac{2}{A+B} S \right)^\ell,$$

with exponential decay in error rate proportional to $(B-A)/(B+A)$ (Jindal et al., 2022).

4. Decomposition Spaces and Sparse NSGF Expansions

NSGF coefficients characterize smoothness and sparsity via decomposition spaces. Let $\mathcal{Q}$ be a structured covering of the frequency domain, and $D(\mathcal{Q}, L^p, \ell^q_{w^s})$ the corresponding decomposition space, with norm

$$\|f\|_{D(\mathcal{Q}, L^p, \ell^q_{w^s})} = \Big\| \{ |Q_T|^{\frac{1}{p} - \frac{1}{2}} \|\psi_T(D) f\|_{L^p} \}_{T \in \mathcal{T}} \Big\|_{\ell^q_{w^s}},$$

where $\{\psi_T\}$ are a bounded admissible partition of unity. The NSGF-induced coefficient map is bounded and norm-equivalent,

$$\|f\|_{D(\mathcal{Q}, L^p, \ell^q_{w^s})} \asymp \|\{\langle f, h^p_{T,n} \rangle\}\|_{\ell^q_{w^s}}.$$

Thus, membership in $D(\mathcal{Q}, L^p, \ell^q_{w^s})$ is equivalent to sparseness of NSGF expansions (Ottosen et al., 2016, Ottosen et al., 2017).

Nonlinear approximation error is tightly controlled: for $0 < \tau < 1$, one has a Jackson-type estimate

$$\| f - f_N \|_{D(\mathcal{Q}, L^2, \ell^2)} \leq C N^{-\alpha} \|f\|_{D(\mathcal{Q}, L^\tau, \ell^\tau_{w^s})},$$

with $\alpha = 1/\tau - 1/p > 0$. Thresholding NSGF coefficients thus achieves near-optimal compression rates for sparse signals (Ottosen et al., 2016).

5. Continuous NSGFs, Reproducing Pairs, and Affine Group Representations

The continuous nonstationary Gabor transform generalizes discrete NSGFs to LCA groups, with time-frequency atoms parameterized over $G \times Y$ for locally compact abelian group $G$ and measure space $Y$,

$$\Psi(x, y) = T_x \psi_y, \quad \Phi(x, y) = T_x \varphi_y.$$

Analysis and synthesis operators are constructed via integration, and the frame operator becomes a Fourier multiplier,

$$m_{\Psi, \Phi}(\xi) = \int_Y \overline{\hat{\psi}_y(\xi)} \hat{\varphi}_y(\xi) \, d\mu(y),$$

with invertibility and frame bounds following from $A \leq |m_{\Psi, \Phi}(\xi)| \leq B$ almost everywhere (Speckbacher et al., 2014, Speckbacher et al., 2015). It is proven that, although canonical dual frames always exist for continuous frames, dual systems of identical structural form cannot always realize the identity resolution operator. Explicit counterexamples constructed from Weyl–Heisenberg subgroup actions demonstrate impossibility for certain windows in $L^2(\mathbb{R})$ (Speckbacher et al., 2014, Speckbacher et al., 2015).

6. Explicit Constructions, Parametric NSGF Families, and Generalization

Explicit nonstationary frame constructions are available, including $\alpha$-DOST systems, which interpolate between classical Gabor frames ($\alpha=0$) and Stockwell-type DOST bases ($\alpha=1$). For each band index $p$, time index $k$,

$$\psi_{\alpha;p,k}(t) = \frac{1}{\sqrt{\beta_\alpha(p)}} \sum_{\eta \in \mu I_{\alpha;p}} e^{i\eta (t - \nu k / \beta_\alpha(p))} \varphi(t - \nu k / \beta_\alpha(p)),$$

with frame conditions and Walnut-type representations ensuring invertibility if frequency partition and window decay conditions are met (Battisti et al., 2015). The parameter $\alpha$ governs the growth of frequency bands, interpolating between uniform tilings ($\alpha=0$) and dyadic, increasing-width tilings ($\alpha=1$).

This construction allows direct adaptation to multidimensional settings, computation via FFT on partitioned bands, and perfect reconstruction using conjugate filters. Applications include seismic imaging, adaptive analysis, and connection to modulation spaces.

7. Applications and Computational Implications

NSGFs form the foundation for adaptive, real-time signal-processing systems such as phase vocoders, constant-Q filter banks, and reduction-based denoising and source separation. By varying windows and lattice parameters according to transient detection or frequency characteristics, NSGFs yield time-frequency representations with improved sparsity and resolution trade-offs (Ottosen et al., 2016). Fast reconstruction is feasible in the painless regime, while iterative inversion and preconditioning remain viable for general NSGF systems.

Computationally, NSGFs admit analysis-synthesis pipelines compatible with FFT-based implementations, leveraging local windows and partitioned frequency domains for efficiency. Practical examples in music and audio clearly demonstrate superiority in sparsity and error-rate metrics compared to stationary Gabor expansions, with adaptive NSGFs achieving lower absolute approximation errors for thresholded expansions (Ottosen et al., 2017, Ottosen et al., 2016).

---

Primary references: Ottosen & Nielsen (Ottosen et al., 2016), Dörfler & Matusiak (Dörfler et al., 2011, Dörfler et al., 2013, Holighaus, 2013), Speckbacher & Balazs (Speckbacher et al., 2014, Speckbacher et al., 2015), Jindal & Vashisht (Jindal et al., 2022), Battisti & Berra (Battisti et al., 2015), and relevant applied work (Ottosen et al., 2016, Ottosen et al., 2017).