Learning Dirac Spectral Transforms for Topological Signals

Leonardo Di Nino, Tiziana Cattai, Sergio Barbarossa, Ginestra Bianconi, Paolo Di LorenzoarXiv 2026 · DOI · arXiv

TL;DR

This paper introduces Dirac-Driven Transform Learning (DDTL), a novel framework for processing signals that live simultaneously on the nodes and edges of a network. By learning a "mass" parameter for each spectral mode, the method adaptively interpolates between coupled (Dirac) and decoupled (Laplacian) representations, achieving superior sparsity and denoising performance compared to traditional fixed-basis methods.

What Problem Does This Paper Address?

In traditional Graph Signal Processing (GSP), we usually analyze signals on nodes (like temperature at weather stations) or edges (like traffic flow on roads) separately using Laplacian operators. However, in many real-world systems—like water distribution or power grids—node and edge signals are intrinsically linked by physical laws (e.g., pressure vs. flow).

The Dirac operator provides a mathematical way to couple these domains, but it assumes a specific, rigid relationship between them. Real-world data is often "messy": some components might be tightly coupled while others act independently. Existing methods force a choice between a Laplacian basis (no coupling) or a Dirac basis (fixed coupling). This paper addresses the need for a flexible, data-driven transform that can discover the optimal level of coupling for any given dataset.

Methods

The authors propose two main contributions to bridge the gap between Laplacian and Dirac representations:

  1. Dirac-Laplacian Frames: They first construct an overcomplete dictionary by concatenating the eigenmodes of both the Dirac and Laplacian operators. This allows a sparse recovery algorithm to "pick and choose" the best representation for different parts of the signal.
  2. Dirac-Driven Transform Learning (DDTL): This is the core innovation. Inspired by the Dirac equation in physics, they introduce a mode-specific mass parameter ($m_i$).
  3. When the mass is zero, the transform behaves like a standard Dirac operator (high coupling).
  4. When the mass is very large, the transform effectively decouples the domains, behaving like a Laplacian operator.
  5. By formulating an optimization problem using the Alternating Direction Method of Multipliers (ADMM), the authors learn these mass parameters directly from data, alongside a sparse representation of the signal.

Datasets

The authors evaluated their approach on two types of data:

  • Synthetic Data: Signals generated over Erdős–Rényi random graphs ($V=40, E=80$). They simulated four regimes: fully coupled, fully decoupled, partially coupled, and a "mixture-of-Dirac" signals.
  • Real-World Data (Anytown Network): A benchmark dataset for water distribution networks consisting of 22 nodes and 41 edges. The signals represent node pressure measurements and pipe flow rates, which are naturally entangled by hydraulic principles.

Key Insights and Ideas

  • The "Topological Spinor": The paper treats the combined node and edge signals as a single mathematical object called a "spinor." This allows for a unified spectral analysis of the entire network structure.
  • Mass as a Coupling Tuner: The most elegant idea is using the physical concept of "mass" as a tunable knob. Instead of a one-size-fits-all coupling, DDTL learns a different knob setting for every frequency mode.
  • Bridging Regimes: The proposed transform is a generalization. It doesn't just compete with Laplacian and Dirac methods; it encompasses them as limiting cases, ensuring it can theoretically perform at least as well as either.
  • Sparsity-Distortion Trade-off: The results highlight that real-world signals are rarely "pure." By being able to adapt the basis, DDTL finds a representation that is both more accurate (lower distortion) and simpler (higher sparsity).

Key Results

  • Sparsity Gains: In synthetic tests, DDTL consistently achieved the lowest Normalized Mean Squared Error (NMSE) across all sparsity levels. Notably, in the "mixture" scenario, it significantly outperformed the joint Dirac-Laplacian frame, proving that learning the parameters is better than just having a bigger fixed dictionary.
  • Denoising Excellence: On the Anytown water network data, DDTL was compared against IDESP (an iterative Dirac filtering framework). DDTL showed superior noise robustness, especially as the Signal-to-Noise Ratio (SNR) increased, demonstrating its ability to learn an informative basis even from noisy real-world measurements.

Limitations and Future Work

  • Computational Complexity: The learning process involves ADMM and manifold constraints (the "oblique manifold"), which may be computationally expensive for very large-scale graphs compared to fixed transforms.
  • Higher-Order Structures: While this paper focuses on nodes and edges (1D simplicial complexes), the authors note that extending this to higher-order structures like triangles (2D) or general hypergraphs is a logical next step.
  • Deep Learning Integration: The current framework is a standalone signal processing tool. The authors suggest that integrating these learnable Dirac transforms into Graph Neural Network (GNN) architectures could lead to more physically-informed and interpretable AI models for topological data.

Summary updated

Abstract

The Dirac operator provides a unified framework for processing signals defined over different order topological domains, such as node and edge signals. Its eigenmodes define a spectral representation that inherently captures cross-domain interactions, in contrast to conventional Hodge-Laplacian eigenmodes that operate within a single topological dimension. In this paper, we compare the two alternatives in terms of the distortion/sparsity trade-off and we show how an overcomplete basis built concatenating the two dictionaries can provide better performance with respect to each approach. Then, we propose a parameterized nonredundant transform whose eigenmodes incorporate a mode-specific mass parameter that captures the interplay between node and edge modes. Interestingly, we show that learning the mass parameters from data makes the proposed transform able to achieve the best distortion-sparsity tradeoff with respect to both complete and overcomplete bases.
BibTeX
@misc{nino2026learning,
  author = {Leonardo Di Nino and Tiziana Cattai and Sergio Barbarossa and Ginestra Bianconi and Paolo Di Lorenzo},
  title = {{Learning Dirac Spectral Transforms for Topological Signals}},
  year = {2026},
  eprint = {2602.14590},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2602.14590},
}

Loading related content...

Figure