Smooth Learning with Hard Constraints via Legendre-Regularized Policies
arXiv:2607.24007
2026
Architecture
1 ideas extracted · analyzed Aug 30, 2026
What the math gives to ML
The paper's central transferable construction is a policy layer whose output is the solution of a strictly convex, Legendre-regularized optimization problem over the original feasible set. Unlike penalty-based constraint handling, this gives exact feasibility while retaining a single-valued differentiable map from network logits to decisions; the inverse regularizer Hessian supplies an explicit Jacobian for backpropagation. This is particularly useful for neural policies with polyhedral resource, budget, flow, or allocation constraints, where projection layers can be nondifferentiable at active-set changes. The most practical adaptation is a trainable constrained decision head using a Legendre or log-barrier layer, with the paper's approximation result motivating a sufficiently expressive context encoder upstream.
Ideas from this paper
△ Mechanism confirmed, baseline not beaten
2026
Replace an unconstrained neural decision head followed by clipping, projection, or penalty losses with a Legendre-regularized optimizer layer that always returns a feasible decision. The layer maps network-produced latent scores smoothly into the relative interior of a convex feasible region and can be differentiated using the inverse Hessian of the regularizer.
Useful8/10
Difficulty5/10
Novelty6/10