Adam M. Oberman,Yuanlong Ruan
Abstract:An efficient method for computing solutions to the Optimal Transportation (OT) problem with a wide class of cost functions is presented. The standard linear programming (LP) discretization of the continuous problem becomes intractible for moderate grid sizes. A grid refinement method results in a linear cost algorithm. Weak convergence of solutions is stablished. Barycentric projection of transference plans is used to improve the accuracy of solutions. The method is applied to more general problems, including partial optimal transportation, and barycenter problems. Computational examples validate the accuracy and efficiency of the method. Optimal maps between nonconvex domains, partial OT free boundaries, and high accuracy barycenters are presented.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is: **Develop an efficient linear programming method applicable to a wide range of cost functions to solve the Optimal Transportation (OT) problem**. Specifically, the standard linear programming discretization method becomes intractable for medium - sized grids when dealing with continuous OT problems. Therefore, the authors propose a grid refinement method, which makes the cost of the algorithm linear, and improves the accuracy of the solution through barycentric projection.
### Main problems and challenges
1. **High computational complexity**: Traditional linear programming methods will rapidly increase in computational complexity when dealing with larger - scale problems, resulting in excessive consumption of computational resources.
2. **Limited scope of application**: Existing numerical methods are either inaccurate, or have too high a computational cost (polynomial time), or are too strict in their requirements for cost functions and measures, and cannot be widely applied to practical problems.
3. **Requirements of modern applications**: Modern applications need to handle large - scale problems or a large number of OT problems, and existing methods cannot meet these requirements.
### Solutions
The authors introduce an efficient linear programming method with the following characteristics:
- **Linear cost**: It is linear in both time and memory requirements and can handle problems with up to 500,000 variables on an ordinary laptop.
- **Wide range of cost functions**: It is not limited to quadratic cost functions and can handle a large class of cost functions.
- **Grid refinement method**: Through a multi - scale solution process, the number of variables is reduced, from quadratic dependence to linear dependence.
- **Barycentric projection**: It is used to recover the mapping from the transport plan, significantly improving the accuracy of the solution.
### Mathematical formulas
The objective of the optimal transport problem is to minimize the total workload, corresponding to a mapping \(T\), such that:
\[
\text{Minimize } I[T]=\int_X c(x, T(x)) d\mu(x)
\]
where \(c(x, y)\) is the cost function, and \(\mu\) and \(\nu\) are two probability measures, defined on \(X\) and \(Y\) respectively.
The Kantorovich formula transforms the optimal transport problem into an infinite - dimensional linear programming:
\[
\text{(KLP) Minimize } I[\pi]=\int_{X\times Y} c(x, y) d\pi(x, y), \quad \text{for } \pi\in\Pi(\mu, \nu)
\]
where \(\Pi(\mu, \nu)\) is the set of all transport plans, that is:
\[
\Pi(\mu, \nu)=\{\pi\in P(X\times Y)\mid \text{marginals of } \pi \text{ are } \mu \text{ and } \nu\}
\]
### Conclusion
This method not only performs well in terms of efficiency and accuracy, but can also reveal solution characteristics that cannot be obtained by other methods, such as optimal mappings between non - convex sets and cases of non - quadratic cost functions. In addition, this method is easy to generalize to related problems, such as partial optimal transport and high - resolution barycentric calculations.