Abstract
Traditional Graph Signal Processing (GSP) focuses on pairwise interactions, often failing to capture the higher-order dependencies inherent in complex systems like brain networks or social structures. This paper introduces a framework for Topological Signal Processing (TSP) over Weighted Simplicial Complexes (WSCs). By interpreting simplex weights as a metric tensor, the authors develop a method to jointly learn these weights and denoise signals using a weighted Hodge Laplacian. Experimental results on synthetic datasets demonstrate that this joint optimization significantly improves signal recovery and provides consistent metric estimation.
Introduction
While GSP has proven effective for analyzing data on networks, it is restricted to vertices and edges. Many real-world phenomena involve multi-way interactions that are better represented by Simplicial Complexes (SCs), which include triangles and higher-order simplices. However, most TSP methods assume an unweighted or "flat" metric, treating all simplices of the same order equally. This paper addresses the challenge of representing and learning weights within a simplicial complex, proposing that these weights define the underlying geometry of the topological space.
Theoretical Framework
The core of the proposed approach is the Weighted Hodge Laplacian, a generalization of the standard graph Laplacian to higher dimensions.
Metric Interpretation of Weights
The authors prove that assigning weights to a simplicial complex is mathematically equivalent to defining a metric tensor ($G_k$). This tensor modifies the inner product between signals, effectively changing how similarity and "smoothness" are measured on the complex.
Hodge Decomposition
Any flow on a simplicial complex can be decomposed into three orthogonal components:
- Irrotational: Flows originating from nodes (gradient-like).
- Solenoidal: Flows circulating around triangles (curl-like).
- Harmonic: Steady-state flows that are neither irrotational nor solenoidal.
Joint Learning and Metric Estimation
To handle noisy observations, the authors propose Algorithm 1, an alternating minimization strategy. This algorithm iteratively updates signal components and triangle weights ($G_2$) to find a local optimum. By assuming that natural flows exhibit low "circulation" (low Total Variation) around triangles, the framework can reverse-engineer the most likely weights from observed data.
Experimental Results
The framework was validated using synthetic datasets, including random simplicial complexes ($N=40$ nodes, $E=137$ edges, $T=96$ triangles) and random geometric graphs. Edge flows were corrupted with Additive White Gaussian Noise (AWGN) to test denoising capabilities.
Denoising Performance
As shown in Figure 1, the joint learning approach (Algorithm 1) significantly outperforms the standard "flat metric" ($G_2 = I$) approach. The correlation coefficient between the estimated and true signals remains higher for Algorithm 1 even as the noise level ($\sigma$) increases, demonstrating the value of adapting the topology's weights to the signal.
Metric Estimation Accuracy
The consistency of the weight estimator was tested by varying the number of observed signal snapshots ($M$). As illustrated in Figure 2, the Mean Squared Error (MSE) of the learned weights decreases consistently as more data becomes available, confirming that the algorithm accurately recovers the underlying metric of the complex.
Conclusion and Future Work
This work establishes that "weighting" in TSP is a geometric property that can be learned directly from data. By jointly optimizing for signals and weights, the proposed framework achieves superior denoising performance compared to unweighted methods. Future research will focus on scaling these algorithms for larger complexes, learning weights across all orders ($G_0, G_1, G_2$) simultaneously, and validating the approach on real-world biological and social network data.
Summary updated