Stability for the mailing problem (Q2314025)

If \(Y\) is a locally compact, separable metric space, then \(\mathscr{M}_+(Y)\) is the set of positive finite Radon measures in \(Y\), that is, the set of positive measures on the \(\sigma\)-algebra of Borel sets of \(Y\) that are finite and inner regular. The set of probability measures, i.e., measures \(\mu\) satisfying \(\mu(Y)=1\), is denoted \(\mathscr{P}(Y)\). The support of a measure \(\mu\) is the set \(\mathrm{supp}(\mu)=\bigcap\{C\subset Y\}\) such that \(C=\overline C\) and \(\mu(Y\backslash C)=0\). It is said that \(\mu\) is supported on a Borel set \(E\) if \(\mu(Y\backslash E)=0\). The total variation, or a mass of \(\mu\) is defined as \(|\mu|=\mu(Y)\). If \(X\subset\mathbb{R}^d\) is a compact, convex set, then \(\mathrm{Lip}_1\) is the space of 1-Lipschitz curves \(\gamma:\mathbb{R}^+\to X\) endowed with the metric of uniform convergence on compact subsets of \(\mathbb{R}^+=[0,\infty)\). For every \(\gamma\in\mathrm{Lip}_1\) the stopping time is defined as \(T(\gamma)=\inf\{t\}\subset\mathbb{R}^+\) such that \(\dot\gamma(s)=0\) for \(s\ge t\). On the set \(\{\gamma\}\subset\mathrm{Lip}_1(X)\) such that \(T(\gamma)<\infty\), the evaluations at time \(0\) and at time \(\infty\) are \(e_0(\gamma)=\gamma(0)\) and \(e_\infty=\gamma(T(\gamma))\), respectively. The set \(\mathbf{TP}=\{\mathbf{P}\in\mathscr{M}_+(\mathrm{Lip}_1(X))\}\) such that \(\mathbf{P}\) is supported on \(\{\gamma;\ T(\gamma)<\infty\}\) is called the space of traffic plans with values in \(X\). Every traffic plan \(\mathbf{P}\) naturally identifies a coupling between an irrigating and an irrigated probability measure on \(\mathbb{R}^d\). The measure \(\pi_{\mathbf{P}}=(e_0,e_\infty)_\sharp\mathbf{P}\) in \(\mathscr{M}_+(X\times X)\) is called the coupling of \(\mathbf{P}\). For every \(\pi\in\mathscr{M}_+(X\times X)\), the set \(\mathbf{TP}(\pi)=\{\mathbf{P\in TP}(X)\}\) such that \(\pi_{\mathbf{P}}=\pi\) is called the set of traffic plans with coupling \(\pi\). For a traffic plan \(\mathbf{P}\) and a point \(x\in\mathbb{R}^d\), the multiplicity of \(\mathbf{P\in TP}(X)\) at \(x\) is defined as \(|x|_{\mathbf{P}}=\mathbf{P}(\{\gamma\in\mathrm{Lip}_1(X)\})\) such that \(\gamma(t)=x\) for some \(t\in\mathbb{R}^+\). For any \(\alpha\in[0,1)\), a functional \(\mathscr{E}^\alpha:\mathbf{TP}(X)\to[0,1)\) defined as \(\mathscr{E}^\alpha(\mathbf{P})=\int\limits_{\mathrm{Lip}_1}\int\limits_{\mathbb{R}^+}|\gamma(t)|^{\alpha-1}_{\mathbf{P}}|\dot\gamma(t)|dt\,d\mathbf{P}(\gamma)\) is called \(\alpha\)-energy. The set of optimal traffic plans with coupling \(\pi\in\mathscr{M}_+(X\times X)\) is defined as \(\mathbf{OTP}(\pi)=\{\mathbf{P\in TP}(\pi);\ \mathscr{E}^\alpha(\mathbf{P})\le\mathscr{E}^\alpha(\mathbf{P'})\}\) for all \(\mathbf{P'\in TP}(\pi)\). The class \(\mathbf{TP}_C(X)\) is the set \(\mathbf{TP}_C(X)=\left\{\mathbf{P\in TP}(X);\ \int\limits_{\mathrm{Lip}_1}T(\gamma)d\,\mathbf{P}\le C\right\}\). The topology on \(\mathrm{Lip}_1\) is induced by the uniform convergence on compact subsets of \([0,\infty)\). If \(\{\mathbf{P}_n\}\) is a sequence in \(\mathscr{M}(\mathrm{Lip}_1)\), then it is said that \(\mathbf{P}_n\overset{*}{\rightharpoonup}\mathbf{P}\) if and only if \(\int\limits_{\mathrm{Lip}_1}d\mathbf{P}_n(\gamma)\to\int\limits_{\mathrm{Lip}_1}d\mathbf{P}(\gamma)\) for every \(f:\mathrm{Lip}_1\to\mathbb{R}\) continuous and compactly supported with respect to the considered topology in \(\mathrm{Lip}_1\). The optimal branched transportation is a variant of the Monge-Kantorovich optimal transportation of mass, in which the mass particles are assumed to interact, rather than traveling independently, while moving from a source to a target distribution. In particular there is a gain in the cost of the transportation whenever some mass is transported in a grouped way. A consequence of this assumption is that the particles' paths form a one-dimensional network which develops branched structures. In order to translate in mathematical terms the convenience of grouping mass during the transportation, one considers a cost functional obtained integrating along the network created by the particles' trajectories a subadditive function of the intensity of the flow. In this paper, the authors consider the problem called mailing problem, also known as who goes where problem. This is a version of the branched transportation problem in which not only the initial and target mass distributions are given, but also a coupling between the two measures, that is, one knows a priori where each mass particle should be moved, and the only unknown is an optimal transportation network realizing such coupling. The authors address an open problem on the stability of optimal traffic plans for the mailing problem, with respect to variations of the given mass distribution. They show that if \(\alpha > 1-\frac1d\), \(C > 0\), \(\pi\in\mathscr{P}(X\times X)\) is a probability measure, \(\{\pi_n\}_{n\in\mathbb{N}}\subset\mathscr{P}(X\times X)\)) is such that \(\pi_n\overset{*}{\rightharpoonup}\pi\) weakly* in the sense of measures, \(\mathbf{P}_n\in\mathbf{OTP}(\pi_n)\) is an optimal traffic plan for the mailing problem, \(\mathbf{P}_n\subset\mathbf{TP}_C\), and there exists a traffic plan \(\mathbf{P}\in\mathscr{P}(\mathrm{Lip}_1)\) satisfying \(\mathbf{P}_n\overset{*}{\rightharpoonup}\mathbf{P}\) as \(n\to\infty\), then \(\mathbf{P}\) is optimal, namely \(\mathbf{P}\in\mathbf{OTP}(\pi)\) and \(\mathscr(E)^\alpha(\mathbf{P})_n\to\mathscr(E)^\alpha(\mathbf{P})\).