⚡ electrical engineering

A Unified Fractional Spectral Framework for Spatiotemporal Graph Signals: Bi-Fractional Transform and Geodesic Coupling

This paper proposes a unified framework for spatiotemporal graph signal processing that introduces a two-dimensional graph bi-fractional Fourier transform with independent fractional orders and a geodesic-coupled GFRFT to resolve basis ambiguity, enabling a differentiable, end-to-end learnable Wiener-type filtering approach that outperforms existing state-of-the-art methods.

Original authors: Mingzhi Wang, Manjun Cui, Feiyue Zhao, Yangfan He, Zhichao Zhang

Published 2026-03-03

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Mingzhi Wang, Manjun Cui, Feiyue Zhao, Yangfan He, Zhichao Zhang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: Tuning a Radio for a Noisy World

Imagine you are trying to listen to a radio broadcast, but the signal is coming from two very different places at once: a bustling city street (chaotic, fast-changing) and a quiet library (calm, slow-moving).

In the world of data science, this is like analyzing spatiotemporal graph signals.

The "Graph": Think of it as a map of connections. It could be a map of cities connected by roads, a social network of friends, or even pixels in a video connected to their neighbors.
The "Signal": The data flowing through those connections (like traffic speed, a friend's mood, or the brightness of a pixel).
The Problem: Real-world data is messy. It has noise (static on the radio). To clean it up, we usually use a "Fourier Transform," which is like a radio tuner that separates the clear signal from the noise.

However, traditional tuners have a flaw: they treat the "city" part and the "library" part of your data exactly the same way. If the city needs a fast, aggressive filter to cut through traffic noise, but the library needs a gentle filter to preserve whispers, a single setting ruins one of them.

This paper proposes a new, super-smart tuner that can handle these two different worlds simultaneously.

The Two Main Innovations

The authors built a framework with two clever tricks to solve this problem.

1. The "Bi-Fractional" Tuner (The Two-Handed DJ)

The Old Way: Imagine a DJ with one volume knob that controls both the bass and the treble. If you turn it up to fix the bass, the treble gets too loud. This is what older methods did: they used one "fractional order" (a setting that controls how much of the signal to keep) for the whole map.

The New Way (2D-GBFRFT): The authors gave the DJ two independent knobs.

Knob A controls the spatial map (the cities, the pixels, the friends).
Knob B controls the time dimension (how the data changes over seconds or minutes).

The Analogy: Think of cleaning a muddy window. You have a squeegee for the glass (spatial) and a cloth for the frame (temporal). The old method used one giant rag for both, smearing the frame while cleaning the glass. The new method uses a specialized squeegee for the glass and a soft cloth for the frame, allowing you to clean each part perfectly without messing up the other.

2. The "Geodesic Coupling" (The Smooth Bridge)

The Problem: In the time dimension, there are two ways to analyze data:

Graph-based: Looking at how time flows based on the specific connections of your data (like a winding river).
Standard Math: Using a classic, rigid mathematical formula (like a straight highway).

Both are good, but they speak different languages. If you try to mix them by just averaging them (like blending red and blue paint to get purple), you often lose the "magic" properties that make the math work (specifically, unitarity, which ensures you don't lose energy or data during the process).

The New Way (GC-GFRFT): Instead of mixing them like paint, the authors built a smooth, curved bridge (a geodesic) between the two methods.

The Analogy: Imagine you are walking from a forest (Graph-based) to a city (Standard Math). A straight line might take you through a swamp where you get stuck. A "geodesic" is the shortest, smoothest path along the surface of the earth that keeps you on solid ground.
This bridge allows the system to slide smoothly between the two methods. It guarantees that no matter where you are on the bridge, the math stays perfect and reversible.

How It Works in Practice: The "Smart Filter"

Once they built this new tuner, they used it to create a Learnable Wiener Filter.

The Goal: Remove noise from data (like cleaning up a blurry video or a noisy sensor reading).
The Process:
1. The System looks at the noisy data.
2. The AI (using a method called Gradient Descent) automatically turns the two knobs (the fractional orders) and adjusts the filter settings to find the perfect balance for that specific piece of data.
3. The Bridge (Geodesic Coupling) stays fixed as a safety rail, ensuring the math doesn't break while the AI learns.

Real-World Results:
The authors tested this on:

COVID-19 Data: Cleaning up noisy infection reports across different regions.
Ocean Temperatures: Removing static from sea surface temperature maps.
Dynamic Images: Taking a blurry, shaking video and making it sharp and clear.

The Outcome: Their new method consistently beat the old "one-knob" methods and even competed with massive, complex AI neural networks, but with much less computing power. It was like using a precision scalpel instead of a chainsaw to clean up the data.

Summary

This paper introduces a universal translator for messy, changing data.

It stops treating different parts of data (space vs. time) as if they are the same.
It builds a mathematically perfect bridge between two different ways of analyzing time.
It lets a computer automatically learn the best settings to clean up noise, resulting in clearer images, better sensor readings, and more accurate predictions.

In short: They gave the data a pair of specialized glasses and a smooth path to walk on, so it can see clearly through the noise.

1. Problem Statement

Graph Signal Processing (GSP) extends spectral analysis to data on irregular domains (graphs). However, existing methods for spatiotemporal graph signals (signals evolving over time on a graph) face two primary limitations:

Lack of Adaptivity in Heterogeneous Dimensions: Traditional 2D Graph Fractional Fourier Transforms (2D-GFRFT) apply a single shared fractional order to both the spatial and temporal dimensions. This fails to capture scenarios where spatial and temporal dynamics have distinct spectral characteristics (e.g., smooth spatial diffusion vs. rapid temporal fluctuations).
Basis Ambiguity in Temporal Analysis: When analyzing time-varying signals, there is a choice between using a graph-induced temporal basis (derived from a path graph) or a classical Discrete Fractional Fourier Transform (DFRFT) basis. Both are unitary but induce different spectral coordinate systems. Naive interpolation between them (e.g., linear mixing) breaks unitarity, leading to energy loss and unstable reconstruction.

2. Methodology

The authors propose a unified framework comprising two novel transforms and a learnable filtering strategy.

A. Two-Dimensional Graph Bi-Fractional Fourier Transform (2D-GBFRFT)

Concept: Instead of a shared order, this transform assigns two independent fractional orders ( $\alpha_1, \alpha_2$ ) to the two factor graphs of a Cartesian product graph ( $G_1 \square G_2$ ).
Mechanism: It performs a separable transform where the spatial factor graph uses order $\alpha_1$ and the temporal factor graph uses order $\alpha_2$ .
Properties: It retains the desirable properties of the standard GFRFT, including unitarity, invertibility, and index additivity, while allowing dimension-wise spectral control.

B. Geodesic-Coupled GFRFT (GC-GFRFT)

Concept: To resolve the ambiguity between the graph-induced temporal basis and the DFRFT basis, the authors construct a coupling path on the unitary manifold ( $U(N)$ ).
Mechanism:
1. Define a relative change-of-basis operator $W_t$ between the two temporal bases.
2. Construct a principal geodesic connecting the two endpoints using the matrix logarithm and exponential: $F^{(\lambda; \beta)}_{t,GC} = F^\beta_{G2} \exp(\lambda \log(W_t))$ .
3. The parameter $\lambda \in [0, 1]$ controls the interpolation.
Key Advantage: Unlike Euclidean interpolation, this geodesic path guarantees unitarity for all $\lambda$ $λ$ , ensuring energy preservation and a closed-form inverse.
- $\lambda = 0$ : Reduces to the graph-induced basis (2D-GBFRFT).
- $\lambda = 1$ : Reduces to the DFRFT basis (JFRFT).

C. Differentiable Wiener Filtering Framework

Optimization Strategy: The authors develop a hybrid optimization approach for denoising and restoration:
- Learnable Parameters: The fractional orders ( $\alpha, \beta$ ) and diagonal spectral filter coefficients ( $h$ ) are optimized end-to-end using Gradient Descent (GD) to minimize Mean Squared Error (MSE).
- Fixed Hyperparameter: The coupling parameter $\lambda$ is treated as a structural regularizer. It is selected via a coarse grid search and kept fixed during gradient updates to maintain structural stability.
Efficiency: The framework avoids forming large Kronecker matrices explicitly, utilizing separable matrix multiplications and cached eigenvectors to maintain computational efficiency ( $O(N^3)$ for offline preprocessing, $O(N^2)$ for online transforms).

3. Key Contributions

2D-GBFRFT: A novel separable transform that decouples fractional orders for spatial and temporal dimensions, enhancing modeling fidelity for heterogeneous spatiotemporal data.
GC-GFRFT: A mathematically rigorous method to unify graph-induced and discrete temporal bases via geodesic coupling on the unitary manifold, preserving unitarity and invertibility.
Learnable Filtering: A differentiable Wiener-type filtering scheme that learns optimal spectral parameters and filters simultaneously, validated on real-world datasets.

4. Experimental Results

The framework was evaluated on time-varying graph signal denoising (COVID-19, Sea Surface Temperature, PM-2.5) and dynamic image restoration (deblurring and denoising on the REDS dataset).

Denoising Performance:
- 2D-GBFRFT consistently outperformed the shared-order 2D-GFRFT, proving the benefit of independent orders.
- GC-GFRFT achieved the best or tied-best performance across most configurations, surpassing classical filters (SGWT, Median, ARMA) and competing with or exceeding Graph Neural Networks (GCN, GAT, ChebyNet) in many scenarios.
Dynamic Image Restoration:
- In deblurring and denoising tasks, GC-GFRFT achieved the lowest MSE and highest PSNR/SSIM among fractional transform baselines.
- Qualitative results showed sharper edges and fewer artifacts compared to classical spatial filters and competitive performance against deep learning baselines (Restormer, NAFNet), despite the latter being trained on large datasets while the proposed method uses instance-adaptive optimization.
Computational Complexity: The method maintains a complexity comparable to other fractional transforms, with the added cost of geodesic coupling being negligible in typical spatiotemporal settings where the temporal dimension is small.

5. Significance

This work bridges the gap between graph-based spectral analysis and classical time-frequency analysis. By introducing bi-fractional orders, it addresses the heterogeneity of real-world spatiotemporal data. The geodesic coupling mechanism provides a principled solution to the basis selection problem, ensuring mathematical stability (unitarity) where naive methods fail. The proposed framework offers a lightweight, interpretable, and highly effective alternative to heavy deep learning models for signal processing on graphs, particularly in scenarios requiring adaptability to varying signal structures without massive training data.

Future Directions: The authors suggest extending the geodesic coupling to support spectral extrapolation (beyond interpolation) and integrating these differentiable modules as lightweight layers into deep graph neural networks.