Reproduce “A Stationarity-and-Coupling Criterion for Training-Free Time-Lagged Spectral Embeddings of Multivariate Time Series” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Siddharth Pal, Viktoria Rojkova
A training-free, fixed-length descriptor for multivariate time series, D(τ), is introduced alongside a falsifiable applicability criterion that predicts its suitability for a given task. When the criterion is met, D(τ) achieves competitive accuracy (e.g., 88.5% LOSO on Sleep-EDF with a simple MLP head, comparable to deep methods but 25x faster representation generation) and fails predictably when the criterion is violated.
Key insights
- The effectiveness of the D(τ) descriptor is tied to whether class-discriminative information resides in cross-channel temporal coupling (dynamics) rather than merely marginal per-channel power.
- Approximate stationarity of the time series within analysis windows is a fundamental requirement for meaningful interpretation and application of time-lagged correlation matrices.
- A simple, training-free "pre-flight test" can reliably predict the success or failure of D(τ) for a given task, offering a principled guide for descriptor selection.
- The Marchenko-Pastur law provides a robust, theory-backed method for distinguishing signal from noise in empirical correlation matrices, moving beyond variance-ranked assumptions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.