An exact reformulation, a nonconvex algorithm
Transport through an orthogonal matrix
Square the entries of an orthogonal matrix to obtain a coupling. Follow its descent—and compare it with the exact assignment.
Energy Exact optimum Gap Orthogonality error
Preparing the point clouds…
What is computed and what is only drawn
Both clouds contain n distinct points of mass 1/n. The dashed pairing is an exact Hungarian assignment for the squared Euclidean cost. The moving solid segments represent πᵢⱼ = Qᵢⱼ²/n. To keep the picture readable, only the five largest masses in each row are drawn; their line widths encode mass. All n² entries contribute to the energy and every numerical update.
The projected gradient is taken on the orthogonal group, with the Frobenius metric divided by n. Cayley retractions and a decreasing-energy line search approximate the continuous flow; the time coordinate adds the accepted step sizes. The time slider selects stored orthogonal matrices, without blending them. The energy graph includes the whole computed trajectory: the darker segment is the part already reached. Changing either cloud control resets time and computes a new trajectory in a background worker.
The minimum over orthostochastic couplings equals the Kantorovich optimum, but this does not mean gradient descent finds it. The local-minimum example starts near a strict nonglobal local minimum for a three-point problem: the energy can stop above the dashed level. The experiment is illustrative, not a global optimization certificate.