Uncertainty Principle for Vertex-Time Graph Signal Processing

We present an uncertainty principle for graph signals in the vertex-time domain, unifying the classical time-frequency and graph uncertainty principles within a single framework. By defining vertex-time and spectral-frequency spreads, we quantify signal localization across these domains. Our framework identifies a class of signals that achieve maximum concentration in both the spatial and temporal domains. These signals serve as fundamental atoms for a new vertex-time dictionary, enhancing signal reconstruction under practical constraints, such as intermittent data commonly encountered in sensor and social networks. Furthermore, we introduce a novel graph topology inference method leveraging the uncertainty principle. Numerical experiments on synthetic and real datasets validate the effectiveness of our approach, demonstrating improved reconstruction accuracy, greater robustness to noise, and enhanced graph learning performance compared to existing methods.

💡 Research Summary

This paper introduces a unified uncertainty principle that simultaneously addresses signal localization in the vertex‑time domain and the spectral‑frequency domain, thereby bridging classical time‑frequency uncertainty with recent graph‑signal uncertainty results. The authors model a vertex‑time signal f(v,t) as an element of L²(V×T) and employ the Joint Fourier Transform (JFT) to map it onto the combined spectral (graph eigenvalues λ) and frequency (temporal ω) space. Two orthogonal projection operators are defined: Π_VT, which projects onto a closed subspace of L²(V×T) (including vertex‑time limiting, graph‑heat, and Gaussian‑time atoms), and Π_SF, which projects the JFT of the signal onto a subspace of L²(Λ×Ω) defined by a measurable region Σ in the (λ,ω) plane. By measuring the energy concentration (spread) of a signal after each projection, the authors derive a lower bound on the product of the vertex‑time spread and the spectral‑frequency spread. This bound generalizes earlier results for pure graph signals and for continuous‑time signals, and it reveals that, unlike the classical case, there exist signals that can be perfectly localized in both domains when the time axis is treated continuously.

The paper identifies a family of optimal atoms that achieve equality in the bound. Each atom is a tensor product of a graph heat kernel e^{‑τ_v L}δ_{v₀} (centered at vertex v₀ with spatial diffusion scale τ_v) and a temporal Gaussian kernel (centered at t₀ with temporal diffusion τ_t). By adjusting τ_v and τ_t, these atoms can be tuned to concentrate energy on arbitrary vertex‑time subsets while remaining bandlimited in a prescribed spectral‑frequency region.

Two concrete applications are built on this theory. First, a vertex‑time dictionary is constructed by selecting a set of such atoms that maximally concentrate energy on a target support S⊂V×T. The dictionary learning problem is formulated as a sparse coding task with an L₁ regularizer, solved via alternating minimization. Experiments on synthetic graphs and a real COVID‑19 case dataset (58 California counties over two years) demonstrate that the proposed dictionary outperforms existing short‑time vertex Fourier and wavelet transforms, achieving an average reconstruction gain of about 4 dB even under severe intermittency and noise levels down to 0 dB.

Second, the uncertainty principle is incorporated into graph topology inference. The authors propose an optimization that maximizes spectral‑frequency energy concentration while enforcing the vertex‑spectral spread constraint derived from the uncertainty principle. This yields a graph Laplacian estimate that respects the inherent space‑time trade‑off of the data. Empirical results show a 4.7 % improvement in edge‑recovery accuracy over state‑of‑the‑art graph learning methods, together with enhanced robustness to measurement noise.

Overall, the contributions are threefold: (1) a rigorous, subspace‑based uncertainty principle for vertex‑time signals; (2) a principled design of energy‑concentrated vertex‑time atoms and a corresponding dictionary for signal reconstruction under irregular sampling; and (3) a novel graph learning framework that explicitly accounts for the space‑time localization trade‑off. The work advances graph signal processing by providing both theoretical foundations and practical tools for handling non‑uniform, asynchronous spatio‑temporal data, and it opens avenues for multi‑scale extensions, non‑linear diffusion kernels, and online adaptive implementations.