The dynamical Schrödinger problem in abstract metric spaces (Q6103491)

The authors consider the (dynamic) geometric Schrödinger problem: \(\mathcal{A}_{\varepsilon}(q)=\frac{1}{2}\int_{0}^{1}\left\vert \frac{dq_{t}}{dt}\right\vert ^{2}dt+\frac{\varepsilon ^{2}}{2}\int_{0}^{1}\left\vert \nabla V\right\vert ^{2}(q_{t})dt-\rightarrow \min\), under the constraint.\(q\in C([0,1],M)\), with endpoints \(q_{0}\),\(q_{1}\in M\), where \(\varepsilon >0\) is a given small temperature parameter, \(M\) a Riemannian manifold with scalar product \(\left\langle \cdot,\cdot \right\rangle_{q}\) at a point \(q\in M\) and induced Riemannian distance \(d\), and \(V:M\rightarrow \mathbb{R}\) a given potential which is globally bounded from below on \(M\). The authors also introduce the semi-flow \(\Phi (s,q_{0})\) corresponding to the autonomous \(V\)-gradient flow: \(\frac{d}{ds}\Phi (s,q_{0})=\nabla V(\Phi (s,q_{0}))\), \(s\geq 0\), with the initial condition: \(\Phi (0,q_{0})=q_{0}\in M\). The authors assume that the gradient flow satisfies the contractivity estimate: \(d(\Phi (s,p_{0}),\Phi (s,p_{0}^{\prime}))\leq e^{-\lambda s}d(p_{0},p_{0}^{\prime})\), \(\forall s\geq 0\), \(p_{0},p_{0}^{\prime}\in M\). The first main result proves that for any \(q_{0}\),\(q_{1}\in M\), the Schrödinger functional \(\mathcal{A}_{\varepsilon}\) \(\Gamma\)-converges as \(\varepsilon \rightarrow 0\) to \(\mathcal{A}(q)=\frac{1}{2}\int_{0}^{1}\left \vert \frac{dq_{t}}{dt}\right\vert ^{2}dt\), for the uniform convergence on the space of curves with fixed endpoints \(q_{0}\),\(q_{1}\). For the verification of the \(\limsup\) property, the authors build a regularized curve \(q^{\varepsilon}\) which uniformly converges to \(q\) and such that \(\limsup_{\varepsilon \downarrow 0}\mathcal{A}_{\varepsilon}(q_{\varepsilon})\leq \mathcal{A}(q)\). This is based on an inequality (Evolution Variational Inequality, EVI) concerning the perturbed flow: \(\widetilde{q}_{t}=\Phi (h(t),qt)\), \(t\in \lbrack 0,1]\), where \(h\) is any nonnegative function satisfying: \(h(0)=h(1)=0\). For the \(\liminf\) property, the authors use the lower semicontinuity property of \(\mathcal{A}(q)\) for the uniform convergence. The second main result proves that \(V\) is \(\lambda\)-geodesically convex, that is satisfies: \(V(q_{\theta})\leq (1-\theta)V(q_{0})+\theta V(q_{1})-\frac{\lambda}{2}\theta (1-\theta)d^{2}(q_{0},q_{1})\), \(\forall \theta \in (0,1)\), for any geodesic \((q_{\theta})_{\theta \in \lbrack 0,1]}\) in \(M\). The authors here again use the inequality concerning the above-indicated perturbed flow. In the main part of their paper, the authors describe the metric framework in which they extend the above-indicated estimate. Given a metric space \((X,d)\), they define the space \(AC([0,1],(X,d))\) of absolutely continuous curves on \([0,1]\). Given a functional \(E:X\rightarrow \mathbb{R}\cup \{+\infty \}\), they define a gradient flow of \(E\) in the EVI\(_{\lambda}\) sense (with \(\lambda \in \mathbb{R}\)) as a curve \((\gamma_{t})_{t>0}\subset X\) such that \((\gamma_{t})\in AC_{loc}((0,\infty),X)\) and \(\frac{1}{2}\frac{d}{dt} d^{2}(\gamma_{t},y)+\frac{\lambda}{2}d^{2}(\gamma_{t},y)+E(\gamma_{t})\leq E(y)\), \(\forall y\in X\), a.e. \(t>0\). They prove the \(\Gamma\)-convergence result in this metric context: under appropriate assumptions, if \(x,y\in X\) are such that \(E(x),E(y)<\infty\), then \(\Gamma -\lim_{\varepsilon \rightarrow 0}\mathcal{A}_{\varepsilon}+\iota_{01}= \mathcal{A}+\iota_{01}\), for the uniform convergence on \(C([0,1],X)\), where \(\mathcal{A}_{\varepsilon}(\gamma)=\frac{1}{2}\int_{0}^{1}\left\vert \overset{.}{\gamma}_{t}\right\vert ^{2}dt+\varepsilon ^{2}\frac{1}{2} \int_{0}^{1}\left\vert \partial E(\gamma_{t})\right\vert ^{2}dt\), for all \((\gamma_{t})\in C([0,1],X)\), \(\iota_{01}(\gamma)=0\) if \(\gamma_{0}=x\) and \(\gamma_{1}=y\), and \(\iota_{01}(\gamma)=+\infty\) otherwise, and \(\mathcal{A}(\gamma)\) is the first term of \(\mathcal{A}_{\varepsilon}(\gamma)\), the second integral of \(\mathcal{A}_{\varepsilon}(\gamma)\) being the Fisher information. The authors give a new proof of the result showing that the potential \(E\) is \(\lambda\)-convex along any geodesic. They then analyze the dependence of the optimal entropic cost on the temperature parameter \(\varepsilon >0\). Computing the first and second derivatives of this cost function, that provide a second order expansion with respect to \(\varepsilon\). The paper ends with the description of examples and applications: Hadamard spaces, Boltzmann-Shannon relative entropy on an associated Wasserstein space, entropy functions on this Wasserstein space, the mean-field Schrödinger problem, the non-linear mobility Wasserstein distance. In each case, the authors describe the setting, they present the related literature and they show how their results can be applied.