SVD-based Graph Fourier Transform

• SVD-based GFT is a spectral signal representation technique that uses singular value decomposition on directed and product graphs to produce real (or biorthogonal) frequency bases.
• It employs both global and Kronecker SVD constructions to achieve computational efficiency and scalable spectral analysis, enabling applications in denoising, compression, and speech enhancement.
• Extensions such as parameterized, fractional, and learnable variants enhance reconstruction accuracy, robustness, and adaptability in diverse graph signal processing tasks.

A singular value decomposition (SVD)-based graph Fourier transform (GFT) is a spectral signal representation framework for directed graphs and their product structures, where SVD replaces the traditional eigendecomposition of Hermitian matrices. This approach yields real (or biorthogonal) spectral bases and an invertible analysis–synthesis pair that generalizes the GFT to a broad class of nonsymmetric and product graphs—retaining classical cases such as the discrete Fourier transform (DFT) as special instances. SVD-based GFTs have found principal applications in denoising, compression, deep neural network representations, and graph signal processing on spatio-temporal and product graphs.

1. Mathematical Foundation: SVD of Directed Graph Operators

SVD-based GFT frameworks start by defining a representation matrix $M$ on a directed graph $G = (V, E)$, typically as the (in-degree) Laplacian $L = D - A$ (where $A$ is the adjacency matrix and $D$ the in-degree matrix), but may also use adjacency or other unified representations. For $L \in \mathbb{R}^{N \times N}$, the full SVD reads:

$$L = U \Sigma V^T = \sum_{k=0}^{N-1} \sigma_k\,u_k\,v_k^T,$$

where $U, V \in \mathbb{R}^{N \times N}$ are orthonormal, $$\Sigma = \mathrm{diag}(\sigma_0,\ldots,\sigma_{N-1})$$, and $0 = \sigma_0 \le \cdots \le \sigma_{N-1}$. The left and right singular vectors ($u_k$, $v_k$) define "frequency components" and $\sigma_k$ the associated graph "frequencies," quantifying variation over the graph topology. SVD is guaranteed for all real (not necessarily symmetric) matrices and is numerically robust (Cheng et al., 2022, Chen et al., 2022).

The SVD-based GFT for $x \in \mathbb{R}^N$ can be constructed via the orthonormal 2$N$-dimensional dilation:

$$\mathcal{F}x = \frac{1}{2} \begin{pmatrix} (U + V)^T x \ (U - V)^T x \end{pmatrix}, \qquad \mathcal{F}^{-1} \begin{pmatrix} z_1 \ z_2 \end{pmatrix} = \frac{1}{2} \left[ U(z_1 + z_2) + V(z_1 - z_2) \right].$$

This definition ensures the Parseval identity and perfect invertibility. The GFT coefficients capture "low" and "high" graph-spectral content in terms of singular values and their associated left/right modes (Chen et al., 2022).

2. SVD-Based GFTs on Directed Product Graphs

For product graphs, particularly Cartesian products $\mathcal{G}_1 \boxtimes \mathcal{G}_2$, SVD-based GFTs follow two dominant constructions:

• Direct Product-Laplacian SVD ("$\mathcal{F}_\square$"): The Laplacian of the product graph is $$L_\square = L_1 \otimes I_{N_2} + I_{N_1} \otimes L_2$$. SVD is performed globally on $L_\square$, yielding $U_\square, V_\square, \Sigma_\square$ as above. The GFT and its inverse generalize the single-graph formulas to size $N_1N_2$ (Cheng et al., 2022).
• Kronecker-SVD ("$\mathcal{F}_\otimes$"): SVDs are separately applied to $L_1$ and $L_2$, yielding $U_1, V_1$ and $U_2, V_2$. Kronecker products $U_\otimes = U_1 \otimes U_2$, $V_\otimes = V_1 \otimes V_2$ generate the spectral bases for the product graph. This allows substantial computational savings over global SVD when $N_1,N_2$ are large (Cheng et al., 2022).

The two forms coincide in the undirected (symmetric) case, reducing to the standard joint GFT. Frequencies and modes naturally combine as pairs $(\sigma_{1,i}, \sigma_{2,j})$ or their sums, enabling joint spectral ordering.

3. Extensions: Parameterized and Generalized SVD-GFTs

Parameterized SVD-GFTs introduce tunable operators such as the Unified Graph Representation Matrix (UGRM) $P^{\alpha,k} = \alpha D + (2k-1)(\alpha-1)A$, interpolating between Laplacian, adjacency, degree, and signless Laplacian cases. The SVD-based GFT is then defined on $P^{\alpha,k}$, extending the spectral framework and allowing spectral monotonicity with respect to $\alpha$ (Xie et al., 12 Oct 2025). For product graphs, both global (UGRM-GFT-I on $P^{\alpha,k}_{\boxtimes}$) and Kronecker-based (UGRM-GFT-II) variants have been introduced, providing axiswise adaptation and computational efficiency.

Fractional SVD-based GFTs generalize the operator to a fractional power (e.g., $L^\alpha = U\Sigma^\alpha V^T$ for $0<\alpha\leq1$), yielding flexible spectral analysis in the fractional Fourier domain (Li et al., 4 Jun 2025). The analytical structure, energy preservation, and efficient Kronecker extensions remain intact, enabling higher-order and multi-axis graph-spectral analysis.

4. Algorithmic Implementation and Computational Considerations

The following table summarizes core computational regimes and their costs:

Method	Main SVD Targets	Cost Scaling
Global SVD	$L$, $L_\square$, or $P^{\alpha,k}_{\boxtimes}$	$O(N^3)$ where $N = N_1 N_2$
Kronecker SVD	$L_1$, $L_2$ or $P_1$, $P_2$	$O(N_1^3 + N_2^3)$ (plus Kronecker assembly)
Fractional SVD-GFT	$(L_1^\alpha, L_2^\alpha)$	Same as above
Learnable SVD-GFT	$\mathbf{A}$ from graph shift op ($K$-NN)	Dominated by SVD; but can be integrated into DNN pipelines (Wang et al., 1 Oct 2025)

Direct product-global SVD is only tractable for modest graph sizes; Kronecker approaches dramatically reduce runtime and storage at the cost of only axis-disjoint spectral bases. For neural speech and real-time DSP, learnable SVD-GFTs replace explicit inverses by 1-D convolutions, maintaining stability and avoiding ill-conditioning from small singular values (Wang et al., 1 Oct 2025).

5. Theoretical Properties and Guarantees

SVD-based GFTs always yield real and biorthogonal (but not necessarily orthogonal) frequency bases, supporting perfect reconstruction ($L_2$-norm preservation), and, when built from the Laplacian, the zero singular value always corresponds to the constant vector. The variational (min-max) characterization of singular values relates them to the extremizers of graph-signal variation.

Specializations:

• For undirected graphs, SVD-based GFT reduces exactly to classical (real-symmetric) GFT.
• For directed circulant graphs, careful realification and permutation show that SVD-based GFT coincides with the DFT structure, up to known linear adjustments (Chen et al., 2022).
• The spectrum $\{\sigma_\ell^{(\alpha,k)}\}$ is monotonic nondecreasing in $\alpha$ for fixed $k$, a property established via operator inequalities (Xie et al., 12 Oct 2025).

6. Practical Applications and Empirical Performance

Denoising, Compression, and Interpolation

Bandlimiting in the SVD-GFT domain (retaining coefficients below a spectral cutoff $M$) has proven highly effective for denoising correlated spatio-temporal data, such as weather station measurements on product time-space graphs (Cheng et al., 2022, Li et al., 4 Jun 2025, Xie et al., 12 Oct 2025). In these cases, SVD-GFTs demonstrate superior energy compaction, bandwidth efficiency, and SNR improvements (often 1–2 dB over alternative GFTs; eg., SNR $\approx$ 17.0 dB for Brest data with $M=32/768$ retained (Cheng et al., 2022)). Fractional SVD-GFTs (with $\alpha<1$) further improve denoising compared to classical settings (Li et al., 4 Jun 2025).

Deep Neural Network Representations for Speech Enhancement

In neural speech enhancement, SVD-based GFT yields strictly real-valued, invertible time–graph frequency representations, resolving the amplitude/phase misalignment inherent to complex GFT or STFT pipelines. Masking-based DNNs can operate on a single real stream, exhibiting improved intelligibility and perceptual quality over traditional STFT or eigenvector-GFT methods (Wang et al., 2024). Learnable SVD-GFT frameworks further adapt the topology and basis, enabling end-to-end training, reduced numerical instability, and consistent improvement over baseline metrics (e.g., W-PESQ, STOI) (Wang et al., 1 Oct 2025).

Adaptive and Tunable Graph-Spectral Analysis

Parametric UGRM-GFTs and fractional SVD-GFTs enable adaptive tuning to underlying graph dynamics, with empirical gains in SNR, lower approximation error under bandwidth constraint, and robustness to noise over classical fixed-matrix approaches (Xie et al., 12 Oct 2025, Li et al., 4 Jun 2025).

7. Limitations and Open Challenges

While SVD-based GFTs provide broad generalization and computational flexibility, several challenges remain:

• For massive graphs, global SVD scaling remains prohibitive; scalable incremental SVDs or randomized solvers may be needed.
• Kronecker and separable approaches, while efficient, may not capture all global couplings in highly nonseparable product graphs.
• In learnable SVD-GFTs, avoiding trivial representations (e.g., degenerate graphs) requires regularization or architectural constraints.
• SVD-based GFTs yield strictly real analysis; extension to complex or quaternionic domains for further signal feature extraction is an active direction.

Theoretical and practical extensions—e.g., higher-order products, graph filters in the SVD-GFT domain, and direct neural modulation—continue to advance the expressive power and application range of the SVD-based GFT paradigm.

---

Key References:

• "Graph Fourier transforms on directed product graphs" (Cheng et al., 2022)
• "Graph Fourier transform based on singular value decomposition of directed Laplacian" (Chen et al., 2022)
• "SVD-based ugmt-gft on directed product graphs" (Xie et al., 12 Oct 2025)
• "Time-Graph Frequency Representation with Singular Value Decomposition for Neural Speech Enhancement" (Wang et al., 2024)
• "Learning Time-Graph Frequency Representation for Monaural Speech Enhancement" (Wang et al., 1 Oct 2025)
• "SVD-Based Graph Fractional Fourier Transform on Directed Graphs and Its Application" (Li et al., 4 Jun 2025)