A coupling of Brownian motions in the \(\mathcal{L}_0\)-geometry (Q1751959)

The authors improve a known result on the monotonicity of the transportation cost by \textit{J. Lott} [Calc. Var. Partial Differ. Equ. 36, No. 1, 49--84 (2009; Zbl 1171.53318)], by means of a probabilistic proof which allows them in particular to require much weaker regularity. Let \((M,(g_t))\) be a connected manifold endowed with a Ricci flow, such that \((M,g_t)\) is complete for each \(t\in[0,T]\). The ``\(\mathcal{L}_0\)-functional'' is given on any piecewise smooth curve \(\gamma: [t_0,t_1]\to M\) by \[ \mathcal{L}_0(\gamma) := {1\over 2} \int_{t_0}^{t_1} \Big(|\dot\gamma(t)|^2_{g(t)} + R_{g(t)}(\gamma(t))\Big) dt. \] The associated distance \(L_0\) on \([0,T]\times M\) is then defined in the natural way. Denote by \(P(M)\) the space of Borel probability measures on \(M\), and by \(C_0\) the optimal \(L_0\)-transportation cost on \([0,T]\times P(M)\), also defined in the natural way. Fix \(0<S<T\) and \(0\leq t_0<t'_0\leq T-S\), and for \(0\leq s\leq S\) let \(\sigma:= t_0+S-s, \sigma':= t'_0+S-s\). The authors call ``\(g(\sigma)\)-Brownian motion'' a time-inhomogeneous diffusion process \((X_s)_{0\leq s\leq S}\) on \(M\) which is associated with \(\Delta_{g(\sigma)}\). They assume that \(\inf_{[0,T]\times T^1M} \mathrm{Ricci}_g > -\infty\). Then their main result is the existence, for any \((m,m')\in M^2\), of a \(g(\sigma)\)-Brownian motion \(X\) started from \(m\) and of a \(g(\sigma')\)-Brownian motion \(X'\) started from \(m'\), coupled in such a way that \( L_0\big((\sigma,\text{law of }X_s), (\sigma',\text{law of }X_s')\big)_{0\leq s\leq S}\) be a supermartingale. Of course, this result at once entails the \(C_0\)-monotonicity previously established by Lott [loc. cit.], moreover only under the above-mentioned lower bound on the Ricci tensor and, as well, with \(L_0\) replaced by \(\varphi\circ L_0\), for any non-decreasing concave function \(\varphi\). The difficulties due to the \(L_0\)-cut locus and to the weak regularity assumption are overcome by the use of geodesic random walk approximations of \(X,X'\).