Dynamic Universal Approximation via Signature Controlled Differential Equations

What the math gives to ML

The paper provides a principled way to turn history-dependent dynamics into ordinary controlled differential equations whose state is a path signature, rather than relying on an ad hoc recurrent memory. Its transferable asset is the separation between the control path, the accumulated signature state, and a vector field acting on that state; this gives a continuous-time memory module with explicit truncation and stability knobs. A practical neural adaptation is a truncated signature-controlled CDE in which the signature state summarizes all iterated interactions of the input history up to degree N, while a learned vector field predicts the hidden-state derivative. The main experiment should test whether increasing signature depth improves long-range-memory tasks at comparable parameter count and whether the resulting dynamics are more stable than a standard neural CDE or RNN.