Paper detail

The Lie Detector

How many free variables do we really need to build a credible model of a physical system? Currently there is no systematic approach; we appeal to some physical principles, tune free variables by comparing with canonical cases, and hope our real-world applications interpolate between them. In this work we combine two pioneering and entirely disparate pieces of mathematics: the century-old techniques of Sophus Lie for solving differential equtions and recent work initiated by Field's medallist Terence Tao on converting NP-complete combinatorical problems into neighbouring convex optimisations. We present a novel and fully systematic procedure for designing models of physical systems with necessary and just-sufficient complexity, in marked contrast with the approach to function approximation taken by neural networks and other current approaches to machine learning. Our methodology replaces the ad-hoc development of models to recover structure and understanding from observational, experimental or simulated data. At its core, our method seeks to find invariant properties of differential equations known as Lie symmetries, and for this reason we have called our algorithm the Lie Detector.