Safe screening rules for portfolio optimization with linear and cardinality constraints

What the math gives to ML

The paper develops optimality-preserving screening for cardinality-constrained convex quadratic programs by combining a perspective relaxation, Fenchel duality, and primal upper and lower bounds. The transferable asset is the ability to certify that a variable cannot or must participate before solving a full sparse optimization problem. A direct neural-network use is layerwise structured pruning: fit a convex reconstruction surrogate for a layer under a channel budget, compute dual screening bounds, and permanently remove channels only when the bounds certify the decision. The guarantee applies to the layerwise surrogate rather than automatically to the original end-to-end nonconvex network, so this should be tested as a safe compression or initialization procedure.