Abstract:We investigate the approximation of Monge--Kantorovich problems on general compact metric spaces, showing that optimal values, plans and maps can be effectively approximated via a fully discrete method. First we approximate optimal values and plans by solving finite dimensional discretizations of the corresponding Kantorovich problem. Then we approximate optimal maps by means of the usual barycentric projection or by an analogous procedure available in general spaces without a linear structure. We prove the convergence of all these approximants in full generality and show that our convergence results are sharp.
What problem does this paper attempt to address?
This paper aims to solve the problem of discrete approximation of the Optimal Transport (OT) problem on general compact metric spaces. Specifically, the goals of the paper are as follows:
1. **Effective approximation of optimal values, plans and mappings**: The paper studies how to effectively approximate the optimal value, optimal plan and optimal mapping in the Monge - Kantorovich problem by a fully discrete method. This includes:
- Approximating the optimal value and optimal plan by solving the finitely - dimensional discretized Kantorovich problem.
- Approximating the optimal mapping by the usual barycentric projection or a similar method in general spaces without a linear structure.
2. **Proof of convergence**: The paper proves that all these approximation methods can converge in general, and shows that these convergence results are sharp (i.e., optimal). Specifically, the paper focuses on the following two main aspects:
- **Convergence of discrete plans and optimal values**: Proves the convergence of discrete plans and optimal values.
- **Convergence of projection mappings**: Proves the convergence of projection mappings, especially under certain uniqueness assumptions.
3. **Discretization scheme**: The paper proposes a discretization scheme. By partitioning the spaces \(X\) and \(Y\) into finite Borel sets, discrete probability measures \(\mu_h\) and \(\nu_h\) are generated. Then, the Kantorovich problem is approximated by solving a finite linear programming problem.
4. **Construction of projection mappings**: The paper discusses how to construct the projection mapping \(T_h\) from the discrete plan \(\pi_h\). In particular, when \(Y\) is a subset of a linear space, \(T_h\) can be constructed by barycentric projection. For a general space \(Y\), the paper proposes using the weighted geometric median as an alternative method.
5. **Main results**:
- **Theorem A**: Proves that the discrete plan \(\pi_k\) converges to the set \(O_{KP}\) of all optimal solutions of the Kantorovich problem in the \(W_p\) distance, and the optimal value \(I[\pi_k]\) converges to the optimal value \(I^*\).
- **Theorem B**: Under the strong uniqueness assumption (SU), proves that the projection mapping \(T_k\) converges to the optimal transport mapping \(T^*\) in the \(d_p\) distance.
6. **Related work and contributions**: The paper compares the existing research on the discrete approximation of optimal transport, and points out that most of the existing work is limited to Euclidean spaces and the quadratic cost function \(c(x, y)=|x - y|^2\). However, this paper conducts research on more general compact metric spaces and proposes a new method using geometric median projection.
In general, by proposing a new discretization method, this paper solves the challenge of approximating the optimal transport problem on general compact metric spaces and proves the effectiveness and convergence of this method.