Partial regularity for optimal transport maps (Q2354708)

The authors study some problems related to the regularity of optimal transportation maps in optimal transport theory. They show that, even without any condition on the cost function or on the supports of the densities, optimal transport maps are always smooth outside a closed singular set of measure zero. There are some basic assumptions on the cost functions which are needed to ensure existence and uniqueness of optimal maps. Let \(X,Y\) be two open sets in \(\mathbb R^n\). Then, the basic assumptions are: (C0) the cost function \(c:X\times Y\to \mathbb{R}\) is of class \(C^2\) with \(\|c\|_{C^2(X\times Y)}<\infty\), (C1) for any \(x\in X\), the map \(Y\ni y\mapsto -D_xc(x,y)\in \mathbb{R}^n\) is injective, (C2) for any \(y\in Y\), the map \(X\ni x \mapsto -D_yc(x,y)\in \mathbb{R}^n\) is injective, (C3) det\((D_{xy})c(x,y)\neq 0\) for all \((x,y)\in X\times Y\). The main results are: Theorem 1.3. Let \(X,Y\subset \mathbb{R}^n\) be two bounded open sets, and let \(f:X\to \mathbb{R}^+\) and \(g:Y\to \mathbb{R}^+\) be two continuous probability densities, respectively, bounded away from zero and infinity on \(X\) and \(Y\). Assume that \(c:X\times Y\to \mathbb{R}\) satisfies (C0)-(C3), and denote by \(T:X\to Y\) the unique optimal transport map sending \(f\) to \(g\). Then, there exist two relatively closed sets \(\Sigma_X\subset X, \;\Sigma_Y\subset Y\) of measure zero such that \(T:X\setminus \Sigma_X\to Y\setminus \Sigma_y\) is a homeomorphism of class \(C_{loc}^{0,\beta}\) for any \(\beta <1\). In addition, if \(c\in C_{loc}^{k+2,\alpha}(X\times Y)\) and \(g\in C_{loc}^{k,\alpha}(Y)\) for some \(k\geq 0\), and \(\alpha\in (0,1)\), then \(T:X\setminus \Sigma_X\to Y\setminus \Sigma_y\) is a diffeomorphism of class \(C_{loc}^{k+1,\alpha}\). In Theorem 1.4 the corresponding result is obtained for two continuous probability densities \(f,g\) defined on a smooth Riemannian manifold \(M\) and the optimal transport map \(T:M\to M\) for the cost \(c=d^2/2\), where \(d\) is the Riemannian distance.