EUVIMEDEUVIMED entdecken
Crossref

Spectral-Mamba: Efficient Graph Signal Recovery via Spectral-Serialized State Space Models

Feiyue Zhao, Zhichao Zhang, Yangfan He

IEEE Signal Processing Letters · 2026 · 33 · 1531-1535

DOI: 10.1109/lsp.2026.3682186 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Signal Processing Letters, 2026. Verfasst von Feiyue Zhao und 2 weiteren Autor:innen. Die Arbeit nennt 46 Literaturangaben. Gefördert von: Open Foundation of Hubei Key Laboratory of Applied Mathematics, Hubei University, Foundation of Key Laboratory of System Control and Information Processing.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Graph Signal Recovery (GSR) requires the capture of long-range dependencies on irregular topologies. Traditional Graph Neural Networks (GNNs) often struggle with this because of the oversmoothing pathology inherent in repeated local message passing, whereas graph transformers face quadratic computational bottlenecks. Although emerging State Space Models (SSMs) offer linear-time long-range modeling, adapting their sequential nature to permutation-invariant graphs remains challenging. To address this, we propose Spectral-Mamba, a linear-time framework that inherently bypasses the iterative neighborhood aggregation paradigm. To bridge the gap between irregular topologies and 1-D sequential SSMs, we introduce a spectral serialization strategy based on the Fiedler vector. This mathematically grounded ordering minimizes the total variation of the sequence, injecting a critical topological smoothness prior. Empowered by this structural inductive bias, our framework seamlessly processes irregular graphs as continuous smooth sequences, efficiently capturing long-range diffusion dynamics without the destructive low-pass filtering that is inherent to deep GNNs. The experimental results demonstrate that Spectral-Mamba achieves superior reconstruction accuracy and topological robustness compared to representative baselines, validating its efficacy and efficiency.

Förderung

  • Open Foundation of Hubei Key Laboratory of Applied Mathematics
  • Hubei University
  • Foundation of Key Laboratory of System Control and Information Processing
  • Ministry of Education
  • Nanjing Institute of Technology

Zitieren

APA

Zhao, F., Zhang, Z., & He, Y. (2026). Spectral-Mamba: Efficient Graph Signal Recovery via Spectral-Serialized State Space Models. IEEE Signal Processing Letters, 33, 1531-1535. https://doi.org/10.1109/lsp.2026.3682186

Vancouver

Zhao F, Zhang Z, He Y. Spectral-Mamba: Efficient Graph Signal Recovery via Spectral-Serialized State Space Models. IEEE Signal Processing Letters. 2026;33:1531-1535. doi:10.1109/lsp.2026.3682186

BibTeX
@article{Zhao2026,
  author = {Zhao, Feiyue and Zhang, Zhichao and He, Yangfan},
  title = {Spectral-Mamba: Efficient Graph Signal Recovery via Spectral-Serialized State Space Models},
  journal = {IEEE Signal Processing Letters},
  year = {2026},
  volume = {33},
  pages = {1531--1535},
  doi = {10.1109/lsp.2026.3682186}
}
Kostenlos starten →