Graph Signal Wiener Filtering in the Linear Canonical Domain: Theory and Method Design

Basic Information

• Paper ID: 2510.10512
• Title: Graph Signal Wiener Filtering in the Linear Canonical Domain: Theory and Method Design
• Authors: Xiaopeng Cheng, Zhichao Zhang
• Category: eess.SP (Signal Processing)
• Publication Date: October 12, 2025
• Paper Link: https://arxiv.org/abs/2510.10512

Abstract

Filtering methods based on graph linear canonical transform (GLCT) typically optimize transformation parameters and filters separately, resulting in high computational costs and limited stability. To address this issue, this paper proposes a trainable joint optimization framework that combines GLCT parameters and Wiener filtering into an end-to-end learning process, achieving synergistic optimization between transform domain construction and filtering operations. The method not only eliminates the tedious grid search required by traditional strategies but also significantly enhances the flexibility and training stability of the filtering system. Experimental results on real graph data demonstrate that the proposed method outperforms existing approaches in denoising tasks, with superior denoising performance, higher robustness, and lower computational complexity.

Research Background and Motivation

Problem Definition

In irregular structures such as social networks, transportation systems, and biological molecular networks, data typically resides on non-Euclidean grids, making classical signal processing methods inapplicable. Graph signal processing (GSP) emerges to address this challenge by modeling irregular structure data as graphs, where nodes represent data entities, edges encode their relationships, and signal values are attached to nodes.

Core Challenges

1. Noise Interference: Graph signals inevitably suffer from noise during acquisition, transmission, and storage
2. Filtering Theory Adaptability: Classical linear filtering is based on Euclidean space properties and cannot be directly transferred to graph structures representing non-Euclidean spaces
3. Parameter Optimization Complexity: Existing GLCT filtering methods typically optimize transformation parameters and filters separately, leading to high computational costs and limited stability

Research Motivation

The limitations of existing methods are primarily manifested in:

• Low Computational Efficiency: Grid search strategies require enumerating numerous candidate parameter combinations
• Uncoordinated Optimization: Separation of transform domain selection and filter design leads to suboptimal results
• Poor Scalability: Grid search in high-dimensional parameter spaces faces combinatorial explosion

Core Contributions

1. Novel GLCT Definition: Proposes Lap-CM-CC-CM-GLCT based on Laplacian eigenbasis, filling gaps in existing CM-CC-CM-GLCT formulations and organizing the CDDHFs-GLCT and CM-CC-CM-GLCT frameworks
2. Differentiability Theory: Proves the differentiability of GLCT core modules under weighted adjacency matrices and Laplacian matrices, providing theoretical support for end-to-end optimization of transformation parameters and filter coefficients
3. Joint Optimization Framework: Constructs the GLCT-GWF framework, achieving end-to-end joint optimization of GLCT parameters and filter coefficients, with effectiveness and robustness verified on real graph signal denoising tasks

Method Details

Task Definition

Given the observation model: $\tilde{f} = Gf + n$, where $G$ is a known perturbation matrix, $f$ is a smooth signal, and $n$ is an additive noise term. The objective is to design an optimal filtering method to recover the original signal $f$ in the transformed spectral domain with minimum mean square error (MSE).

Core Technical Components

1. Graph Linear Canonical Transform (GLCT)

GLCT is determined by a 2×2 matrix $M = (a,b;c,d)$ where $ad-bc=1$. This matrix can be decomposed as: $\begin{bmatrix} a & b \\ c & d \end{bmatrix} = \begin{bmatrix} 1 & 0 \\ \xi_1 & 1 \end{bmatrix} \begin{bmatrix} 1 & b \\ 0 & 1 \end{bmatrix} \begin{bmatrix} 1 & 0 \\ \xi_3 & 1 \end{bmatrix}$

where $\xi_1 = \frac{d-1}{b}$, $\xi_2 = -b$, $\xi_3 = \frac{a-1}{b}$.

2. Lap-CM-CC-CM-GLCT Definition

The proposed CM-CC-CM-GLCT definition based on Laplacian matrices is defined as: $\hat{f}^{IV}_M = F^M_L f = CM^{\xi_1}_L U_L CM^{\xi_2}_L U^{-1}_L CM^{\xi_3}_L f$

3. Joint Optimization Objective

Traditional two-step optimization process:

1. Fix transformation parameters $(a,b,d)$ and solve for optimal filter $H^*$
2. Grid search $(a,b,d)$ to minimize MSE

Proposed joint optimization: $\min_{a,b,d,H} E\{\|F^{M^{-1}}_L HF^M_L \tilde{f} - f\|^2_2\}$

Algorithm Framework

Algorithm 1: Traditional Grid Search Method

Input: Graph signal f, target signal f̃, parameter grids A,B,D
Output: Optimal parameters (a*,b*,d*), optimal filter H*
1. Precompute spectral basis (eigendecomposition)
2. for a ∈ A, b ∈ B, d ∈ D:
   - Construct GLCT operators F^M and F^{M^{-1}}
   - Solve Wiener-Hopf equation: h = T^{-1}q
   - Evaluate loss MSE(H,a,b,d)
   - Update optimal solution

Algorithm 2: Adam Joint Optimization Method

Input: Graph signal f, target signal f̃, learning rate ε
Output: Learned filter and GLCT parameters
1. Initialize parameters (a₀,b₀,d₀) and filter H₀
2. while stopping condition not met:
   - Compute gradients ∇_{H,a,b,d}MSE
   - Update parameters: H ← H - ε∇_H MSE
   - Update GLCT parameters: a,b,d ← a,b,d - ε∇_{a,b,d}MSE

Technical Innovations

1. Differentiability Guarantee: Proves the differentiability of loss functions with respect to transformation parameters and filter coefficients, enabling end-to-end optimization
2. Computational Complexity Optimization:
   • Grid search complexity: $O(n_a n_b n_d N^4)$
   • Adam joint optimization complexity: $O(KN^2)$
3. Theoretical Properties: The proposed Lap-CM-CC-CM-GLCT satisfies important properties including linearity, zero rotation, additivity, invertibility, and unitarity

Experimental Setup

Datasets

Synthetic Data

1. 5-nn Graph: 15 nodes, each connected to 5 nearest neighbors
2. Swiss Roll Graph: 30-node manifold structure
3. Sensor Graph: 20-node sensor network

Real Data

1. Sea Surface Temperature (SST): 50 nodes, k∈{2,6,10}
2. PM2.5: Air quality data, 50 nodes
3. COVID-19: Epidemic spread data, 50 nodes

Evaluation Metrics

Mean square error (MSE) is used as the primary evaluation metric to assess denoising performance.

Comparison Methods

• GFRFT_W: Graph fractional Fourier transform based on weighted adjacency matrix
• GFRFT_L: Graph fractional Fourier transform based on Laplacian matrix
• Various GLCT Variants: wAdj-CDDHFs-GLCT, Lap-CDDHFs-GLCT, etc.

Implementation Details

• Grid Search: Parameter range 0,2, step size 0.1
• Adam Optimization: Learning rate 0.005, 5000 iterations
• Noise Settings: Gaussian noise, standard deviation s∈{0.5,1.0,1.5} (synthetic data), s∈{0.5,0.6,0.7} (real data)

Experimental Results

Main Results

Synthetic Data Results

Experiments on three synthetic graphs demonstrate that GLCT methods consistently outperform GFRFT methods:

Real Data Results

On the SST dataset, wAdj-CDDHFs-GLCT achieves the lowest MSE value of 1.442 under k=2, s=0.5 settings, representing approximately 25% improvement compared to traditional GFRFT methods.

Performance Analysis

1. Denoising Performance: GLCT methods demonstrate superior denoising performance under all test conditions
2. Robustness: As noise intensity increases, GLCT methods show more graceful performance degradation
3. Computational Efficiency: Adam joint optimization significantly reduces computational complexity

Ablation Study

By comparing different GLCT variants, the importance of each component is verified:

• CDDHFs-GLCT performs best in most cases
• CM-CC-CM-GLCT shows advantages in certain specific scenarios
• Laplacian basis and adjacency matrix basis each have applicable scenarios

Related Work

Graph Signal Processing Foundations

• Graph Fourier Transform (GFT): Extends classical Fourier transform to graph-structured data
• Graph Fractional Fourier Transform (GFRFT): Introduces fractional-order parameters, interpolating between identity transform and complete GFT

Linear Canonical Transform Development

• Classical LCT: Generalizes rotation to affine transformation, extending FT and FRFT
• Graph Domain Extension: Zhang et al. define GLCT based on CDDHFs; Li et al. propose CM-CC-CM implementation

Wiener Filtering Extension

Traditional graph Wiener filtering operates in fixed transform domains; this paper extends it to learnable transform domain selection.

Conclusions and Discussion

Main Conclusions

1. The joint optimization framework effectively addresses computational complexity and stability issues of traditional methods
2. Rich parameterization of GLCT combined with joint learning achieves better signal-noise separation
3. End-to-end optimization avoids suboptimal results that may arise from sequential optimization

Limitations

1. Non-convex Optimization: The joint optimization problem is inherently non-convex, potentially leading to local optima
2. Parameter Initialization Sensitivity: Adam optimization performance may depend on initialization strategies
3. Theoretical Convergence Guarantees: Lacks rigorous global convergence theoretical analysis

Future Directions

1. Develop stronger theoretical convergence guarantees
2. Explore adaptive parameter initialization strategies
3. Extend to time-varying and dynamic graph signal processing

In-Depth Evaluation

Strengths

1. Solid Theoretical Contributions: Provides complete differentiability proofs and property analysis
2. Strong Method Innovation: First to achieve joint optimization of GLCT parameters and filters
3. Comprehensive Experimental Validation: Verifies method effectiveness across multiple datasets and settings
4. Significant Computational Efficiency Improvement: Complexity reduction from $O(N^4)$ to $O(N^2)$

Weaknesses

1. Insufficient Convergence Analysis: Lacks in-depth theoretical analysis of non-convex optimization convergence
2. Parameter Sensitivity: Limited analysis of sensitivity to learning rates and initialization
3. Application Scope Limitations: Primarily focuses on denoising tasks; applicability to other graph signal processing tasks remains to be verified

Impact

1. Academic Value: Provides new optimization paradigm for graph signal processing field
2. Practical Value: Significantly reduced computational complexity enables large-scale applications
3. Reproducibility: Provides detailed algorithm descriptions and experimental settings

Applicable Scenarios

• Large-scale graph signal denoising tasks
• Graph signal applications requiring real-time processing
• Embedded systems with strict computational efficiency requirements
• Multi-scale graph structure analysis

References

The paper cites 49 relevant references covering core areas including graph signal processing, linear canonical transforms, and Wiener filtering, providing solid theoretical foundation for the research.

Overall Assessment: This paper makes important contributions to the graph signal processing field. Through a joint optimization framework, it effectively addresses computational complexity issues of traditional methods. With sufficient theoretical analysis and comprehensive experimental validation, it demonstrates high academic and practical value.