Answers are grounded in this paper and cite their sources.
Data-driven discovery of Lie point symmetries usually identifies the equation or needs its order. We find an evolution equation’s symmetries from any exact or learned solution operator, whose Jacobian transports its characteristics; over a polynomial class of generators, discovery becomes a generalised eigenvalue problem. Through learned operators, the method misses 0.01–0.44 dimensions of the finite Burgers, heat and Korteweg–de Vries algebras, 0.018 on Benjamin–Ono and 0.000 on Kuramoto–Sivashinsky. Learned-symmetry baselines and a model-free jet-space method miss at least 0.87; another jet-space method, LieNLSD, does better only on the full heat class. Through exact solvers, a weak form free of data derivatives recovers the non-trivial generators at noise up to 10% of the data’s standard deviation (3% on Benjamin–Ono). When a sparse library contains the equation, identifying it first is exact on 176 of 180 runs, and our method at best matches it. Outside the library (nonlocal Benjamin–Ono, fractional Burgers), identified surrogates have wrong algebras, while the weak form estimates the scaling exponent within 0.04 through exact solvers on clean data and at 1% noise. As data augmentation, discovered generators that include the Galilean boost cut a neural emulator’s error 2.7–25-fold on states with means outside its training range.
No takes yet. Share an insight, caveat, or question.
Markov et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: