SpectraLDS: Provable Distillation for Linear Dynamical Systems
This provides a provable and efficient solution for sequence prediction tasks like language modeling, though it is incremental as it builds on prior spectral transformation work.
The paper tackles the problem of identifying symmetric linear dynamical systems (LDS) with accuracy guarantees independent of state dimension or memory, resulting in a method that preserves predictive accuracy while enabling constant-time and constant-space inference per token.
We present the first provable method for identifying symmetric linear dynamical systems (LDS) with accuracy guarantees that are independent of the systems' state dimension or effective memory. Our approach builds upon recent work that represents symmetric LDSs as convolutions learnable via fixed spectral transformations. We show how to invert this representation, thereby recovering an LDS model from its spectral transform and yielding an end-to-end convex optimization procedure. This distillation preserves predictive accuracy while enabling constant-time and constant-space inference per token, independent of sequence length. We evaluate our method, SpectraLDS, as a component in sequence prediction architectures and demonstrate that accuracy is preserved while inference efficiency is improved on tasks such as language modeling.