Learning High-Dimensional Transient Neural Dynamics for Zero-Shot Cross-Subject Reconstruction

Accurate and scalable reconstruction of high-dimensional neural time series remains a central challenge in dynamical systems modeling. In this work, we study long-horizon reconstruction of hippocampal local field potentials (LFPs) using a PCA-DMD framework, with Dynamic Mode Decomposition (DMD) as a data-driven approximation to Koopman spectral analysis. We extend prior analyses by scaling to sequences of up to $300,000$ samples and introducing a cross-subject zero-shot evaluation protocol to assess the generalizability of learned latent representations across four subjects. PCA-DMD consistently achieves strong reconstruction quality under both long-horizon and zero-shot transfer settings, with correlation exceeding 0.95 and very low reconstruction errors. In contrast, standard DMD variants degrade substantially as signal length increases and show limited cross-subject robustness. These results demonstrate that PCA-DMD provides a scalable and transferable framework for reconstructing high-dimensional neural signals, highlighting its potential for modeling long-range neural dynamics in challenging multi-subject settings.