Regularized unbalanced optimal transport as entropy minimization with respect to branching Brownian motion (Q6945323)

This almost 200 page long research monograph deals with some of the latest technical developments of \emph{optimal transport}, at the crossroads between functional analysis, probability theory and convex optimization. It describes various extensions of the classical problem of optimal transport studied by Monge in the 18th century and revisited by Kantorovich in the 1940s, paving the way to the development of linear programming.\N\NThe lengthy but informative title introduces all the main players:\N\begin{itemize}\N\item the optimal transport problem is \emph{regularized}, which means that a non-linear entropy penalization term is added to the linear functional to be minimized;\N\item the optimal transport problem is \emph{unbalanced}, which means that mass is not only transported, but it can also be created or destroyed during the transport;\N\item the connection is made with \emph{branching Brownian motion} and the Schrödinger bridge problem, a classical problem in probability and statistical mechanics that studies the most likely path a cloud of diffusing particles take to get from one given distribution to another.\N\end{itemize}\N\NThe monograph starts with an enlightening overview (Chapter 1), nicely illustrated with Figure 1 on page 5 which compares visually the classical, regularized and unbalanced versions of optimal transport. The main objects of interest are carefully introduced in Chapter 2. Historically posed in the early 1930s, the Schrödinger bridge problem is now understood as an entropic regularization of the optimal transport problem. The variation studied in this monograph is the branching Brownian motion where a population of particles evolve independently according to Brownian motion. An independent exponential clock is attached to each particle. When the clock rings, the particle dies, and it gives birth to a randomly chosen number of particles, which will evolve similarly. The monograph studies the link between this branching Schrödinger problem and the entropic regularized unbalanced optimal transport. In Chapter 3, equivalence of the values of the two problems is shown with the help of convex analysis and duality. In Chapters 4 and 5, a deeper correspondence is established between the problems, yielding a fine description of laws having a finite entropy with respect to a reference branching Brownian motion. A concluding Chapter 6 collects asymptotic results (small noise limit) and a sketch of numerical algorithms.\N\NInterestingly, the link between the two problems is weaker in the unbalanced case: the constraints in the branching Schrödinger problem are not closed, which makes the problem ill-posed. This is clearly described in Section 3.5.2. The authors convincingly argue that the regularized unbalanced optimal transport formulation is the well-posed lower semi-continuous relaxation of the branching Schrödinger problem.\N\NIn the classical optimal transport problem, entropic regularization allows the development of scalable first order optimization methods, namely the Sinkhorn algorithm (an iterative matrix scaling), making optimal transport possible for deep learning and graphics. The authors however notice in section 6.3.1 that a Sinkhorn like algorithm is unlikely to be adapted for their unbalanced version. Consistently, the numerical scheme described in Section 6.3 follows classical discretization and convex optimization methods used for the Benamou-Brenier dynamic version of optimal transport.\N\NThis very carefully written monograph can serve as an introduction to the latest developments of optimal transport, thanks to a broadly accessible and nicely illustrated introductory chapter. It should also be of interest to the experts in the field, as it collects an impressive amount of technical developments, combining approaches from convex analysis, functional analysis and probability theory.