On Learning Discrete-Time Fractional-Order Dynamical Systems

TL;DR

This paper introduces the first finite-sample complexity guarantees for learning discrete-time fractional-order dynamical systems, demonstrating its effectiveness in modeling and forecasting neurophysiological signals like EEG.

Contribution

It provides the first non-asymptotic sample complexity bounds for identifying DT-FODS parameters, enhancing understanding of data requirements for accurate system modeling.

Findings

Finite-sample guarantees established for DT-FODS identification

Method effectively forecasts intracranial EEG signals

Demonstrates trade-offs between sample size and estimation accuracy

Abstract

Discrete-time fractional-order dynamical systems (DT-FODS) have found innumerable applications in the context of modeling spatiotemporal behaviors associated with long-term memory. Applications include neurophysiological signals such as electroencephalogram (EEG) and electrocorticogram (ECoG). Although learning the spatiotemporal parameters of DT-FODS is not a new problem, when dealing with neurophysiological signals we need to guarantee performance standards. Therefore, we need to understand the trade-offs between sample complexity and estimation accuracy of the system parameters. Simply speaking, we need to address the question of how many measurements we need to collect to identify the system parameters up to an uncertainty level. In this paper, we address the problem of identifying the spatial and temporal parameters of DT-FODS. The main result is the first result on non-asymptotic finite-sample complexity guarantees of identifying DT-FODS. Finally, we provide evidence of the efficacy of our method in the context of forecasting real-life intracranial EEG time series collected from patients undergoing epileptic seizures.