The Lie algebra generated by gradient vector fields

What the math gives to ML

The paper proves a strong expressivity statement: on a compact manifold with any smooth non-degenerate symmetric metric, arbitrary smooth vector fields lie in the finite span of iterated Lie brackets of metric-gradient fields. This suggests parameterizing neural dynamics using scalar potentials as primitive objects while recovering rotational and non-conservative behavior through learned commutator layers. The most direct transfer is a Lie-bracket neural ODE or state-space block, implemented with automatic differentiation of potential networks or short compositions of their flows. The theorem is qualitative and gives no approximation rate or constructive depth bound, so empirical validation must compare this structured parameterization with ordinary vector-field MLPs.