M-S4DLite: Orthogonal Decoupling for Lightweight Time-Series Classification via Diagonal State-Space Models

Abstract Emerging state space models have revolutionized time series forecasting by efficiently capturing long range dependencies, yet their application to classification tasks reveals a fundamental semantic misalignment. Rooted in continuous control theory, standard models suffer from spectral oversmoothing where continuous inductive bias tends to erode the sharp and discrete discriminative motifs critical for classification. To resolve this theoretical conflict, we propose M-S4DLite, a lightweight framework built on the principle of Orthogonal Decoupling. By structurally separating local feature extraction from global evolution, we employ a discrete patching module to preserve high frequency transients and synergize it with a diagonalized S4D backbone for global reasoning. This architecture effectively recalibrates the inductive bias to reconcile discrete local precision with continuous global context while maintaining quasi linear complexity. Extensive experiments demonstrate that M-S4DLite establishes a new Pareto frontier in the balance between efficiency and accuracy, consistently achieving superior performance on diverse benchmarks and exhibiting robustness against resolution degradation compared to pure continuous or convolutional baselines.