Regularity of optimal transport in curved geometry: the nonfocal case (Q961478)

The authors study the extension problem of smoothness of optimal transport maps for the quadratic cost function in \({\mathbb R}^n\). They consider only the important case when the cost is the squared geodesic distance on a Riemannian manifold which has many applications in Riemannian geometry. Let \(M\) be a smooth connected complete Riemannian manifold and for \((x, y) \in M\times M \setminus {\text{cut}}(M)\), where \({\text{cut}}(M)\) denotes the cut locus of \(M\), take coordinate systems \((x_i)\) and \((y_j)\) around \(x\) and \(y\), respectively. Set \(c(x',y') = d(x',y')^2/2\), where \(d\) is the geodesic distance on \(M\), and note that \(c\) is \(C^\infty\) near \((x, y)\). For \(\xi\in T_xM,\, \eta \in T_yM\), define \[ \mathfrak S(x, y) \cdot(\xi, \eta) = \frac{3}{2}\sum_{i,j,k,l,r,s} \left(c_{ij, r}c^{r,s} - c_{ij, kl}\right)\xi^i \xi^j \eta^k\eta^l, \] where \(c_i\) (resp., \(c_{,j}\)) denotes the partial derivative with respect to \(x_i\) (resp., \(y_j\)), evaluated at \((x, y)\), and \(c_{i,j}\) is the mixed second derivative with respect to \(x_i\) and \(y_j\), and so forth; and \((c^{i,j})\) denotes the inverse of \((c_{i,j})\), always evaluated at \((x, y)\). It is known that this formula defines a covariant tensor, called Ma-Trudinger-Wang tensor (MTW tensor in short). Moreover, if \(\xi\) and \(\eta\) are orthogonal unit vectors in \(T_xM\), then \(\mathfrak S(x, y) \cdot(\xi, \eta) \) coincides with the sectional curvature at \(x\) along the plane generated by \(\xi\) and \(\eta\). We assume that \[ \mathfrak S(x, y) \geq K |\xi|^2 |\eta|^2\quad \text{whenever}\,\,\, \sum_{i,j} c_{i,j}\xi^i\eta^j = 0, \] where \(K\) is a positive constant (strong MTW condition) or \(K = 0\) (weak MTW condition). This condition implies that the sectional curvature of \(M\) is bounded below by \(K\). In [Optimal transport. Old and new. Grundlehren 338. Berlin: Springer (2009; Zbl 1156.53003)], \textit{C. Villani} proved a necessary condition for the regularity of optimal transport. In this paper, the authors prove a sufficient condition for regularity under a simplifying nonfocality assumption. For any \(\xi \in T_xM,\, |\xi|=1\), let \(t_C(\xi)\) be the first time \(t\) such that \(({\text{exp}}_x(s\xi))_{0 \leq s \leq t'}\) is not minimizing for \(t' > t\), and let \(t_F(\xi) \geq t_C(\xi)\) be the first time \(t\) such that \(d_{t\xi}{\text{exp}}_x\) is not one-to-one. Define \({\text{TCL}}(x) = \{t\xi\,:\, t = t_C(\xi)\,\}\) and \({\text{TCL}}(M) = \cup_x {\text{TCL}}(x)\). Set also \({\text{TFL}}(x) = \{t\xi\,:\, t = t_F(\xi)\,\}\) and \({\text{TFL}}(M) = \cup_x {\text{TFL}}(x)\). We say that the cut locus of \(M\) is nonfocal if \({\text{TCL}}(M) \cap {\text{TFL}}(M) = \emptyset\). With these assumptions, the authors prove that for any \(C^\infty\) positive densities \(f\) and \(g\) on \(M\), the optimal transport map from \(\mu(dx) = f(x) \text{vol}(dx)\) to \(\nu(dy) = g(y) \text{vol}(dy)\), with cost function \(c = d^2\), is smooth.