An Entropic Factor Model for Robust Portfolio Replication

What the math gives to ML

The paper turns an underdetermined bounded linear inverse problem into a unique convex reconstruction by minimizing separable Fermi-Dirac negative entropy over a box. Its transferable asset is not portfolio-specific: the solution remains within explicit coordinate bounds, avoids extreme allocations when constraints are weak or contradictory, and can be computed through a low-dimensional dual root solve. This suggests a principled replacement for heuristic load balancing in mixture-of-experts routing, where token allocations must satisfy capacity and conservation constraints. The most direct experiment is an entropic routing projection that converts noisy router preferences into capacity-safe expert fractions.