Optimizing quasi-dissipative evolution equations with the moment-SOS hierarchy

Saroj Prasad Chhatoi, Didier Henrion, Swann Marx, Nicolas Seguin · 2024

Read Optimizing quasi-dissipative evolution equations with the moment-SOS hierarchy by Saroj Prasad Chhatoi, Didier Henrion, Swann Marx, Nicolas Seguin (2024) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2412-07361.txt.

Introduction

Saroj Prasad Chhatoi, Nicolas Seguin and their co-authors prove a key result for optimizing certain evolution equations in 2024. Their focus is quasi-dissipative nonlinear equations in a Hilbert space, a class that includes reaction-diffusion systems with polynomial terms.

They show there is no relaxation gap between such an equation and its linear reformulation on probability measures. In plain terms, strong and generalized solutions are unique among measure-valued solutions. As a result, non-convex optimization over these partial differential equations can use a moment and sum-of-squares hierarchy with global convergence assured. That opens a rigorous path for control and design problems in this class.

Catalog record

CREATORSaroj Prasad Chhatoi, Didier Henrion, Swann Marx, Nicolas Seguin
YEAR2024
COLLECTIONPAPERS
FORMATTXT
FILE SIZE59 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2412-07361

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Optimizing quasi-dissipative evolution equations with the moment-SOS hierarchy Saroj Prasad Chhatoi, Didier Henrion, Swann Marx, Nicolas Seguin Abstract We prove that there is no relaxation gap between a quasi-dissipative nonlinear evolution equation in a Hilbert space and its linear Liouville equation reformulation on probability measures. In other words, strong and generalized solutions of such equations are unique in the class of measure-valued solutions. As a major consequence, non-convex numerical optimization over these non-linear partial differential equations can be carried out with the infinite-dimensional moment-SOS hierarchy with global convergence guarantees. This covers in particular all reaction-diffusion equations with polynomial nonlinearity. 1 1 footnotetext: CNRS; LAAS; Université de Toulouse, 7 avenue du colonel Roche, F-31400 Toulouse, France. 2 2 footnotetext: Faculty of Electrical Engineering, Czech Technical University in Prague, Technická 2, CZ-16626 Prague, Czechia. 3 3 footnotetext: Nantes Université, Ecole Centrale Nantes, CNRS, LS2N, UMR 6004, F-44000 Nantes, France. 4 4 footnotetext: IMAG, Inria, Université de Montpellier, CNRS, Montpellier, France. D. Henrion’s work was co-funded by the European Union under the ROBOPROX project (reg. no. CZ.02.01.01/00/22 008/0004590). S. Marx’s work was co-funded by the Agence National de la Recherche (ANR) under the ROTATION project (reg. no. ANR-24-CE48-0759). N. Seguin’s work was co-funded by the Agence National de la Recherche (ANR) under the ANR HEAD project ANR-24-CE40-3260. 1 Introduction The moment-SOS 1 1 1 SOS: sum of squares. hierarchy is a mathematical technology that allows to solve numerically with global optimality guarantees a large class of non-convex optimization 2 2 2 Optimization should here be interpreted broadly: finite- and infinite-dimensional, continuous and discrete versions of calculus of variations, optimal control and optimal transport problems on ordinary, stochastic and partial differential equations. problems at the price of solving a family of convex relaxations (typically semidefinite optimization problems) of increasing size. The approach and its applications are described in [ 27 ] . The first step in the approach consists of reformulating a non-convex non-linear optimization problem in a given domain as a linear problem in cones of measures supported on this domain. An archetypal example is the Kantorovich linear reformulation of the non-linear Monge problem of optimal transport [ 48 ] . An important question is whether there is a relaxation gap , i.e. whereas the value of the linear problem on measures differ from the value of the original non-linear problem. In the absence of a relaxation gap, it can be shown (under standard assumptions) that the moment-SOS hierarchy converges, i.e. it generates a sequence of bounds which converges monotonically to the value of the original problems. Then a globally optimal solution can be approximated from the solutions of the successive convex relaxations, see e.g. [ 42 ] for optimal transport. For non-linear partial differential equations (PDEs) and their control, an early attempt to use the linear measure formulation was reported in [ 47 ] for semi-linear elliptic equations. More recent attempts using the moment-SOS approach can be found in [ 6 , 34 , 38 ] for linear PDEs, and more recently in [ 23 , 16 ] for non-linear PDEs. A fully general non-linear setup including optimization, calculus of variations and PDEs was described in [ 37 ] . As far as we know, in these references, there is neither convergence guarantee of the proposed hierarchy, nor a proof of no relaxation gap. In the context of scalar hyperbolic conservation laws, a particular class of nonlinear PDEs, no relaxation gap was ensured by introducing entropy inequalities [ 39 ] (see also [ 11 ] for a similar result for conservation laws admitting parameters). In [ 36 ] it was shown that there is no relaxation gap for scalar problems (i.e. when then dimension of the domain or the dimension of the codomain is equal to one), both for calculus of variations and for optimal control problems. Conversely, an example of a variational problem with relaxation gap is provided when the dimensions of the domain and of the codomain are greater than one. It is also shown that in the presence of integral constraints, a relaxation gap may occur at any dimension of the domain and of the codomain. More recently, it was shown [ 28 ] that under convexity assumptions, there is no relaxation gap for a broad class of variational and optimal control problems on non-linear PDEs. The question of the absence of a relaxation gap for specific classes of non-linear PDEs remains however widely open. These results were obtained for measures supported on finite-dimensional subsets of the time, space, function and function gradient domains. There is however the possibility of reformulating optimization problems over nonlinear PDEs as linear problems on measures supported on infinite-dimensional functional spaces. This is the point of view adopted e.g. in [ 17 , Part III] for optimization with relaxed controls. Measures on infinite-dimensional spaces and statistical measure-valued solutions are also prominently appearing in [ 19 ] and in the subsequent work [ 46 ] on Navier-Stokes equations. This notion of statistical solutions allows for sound numerical implementations, see e.g. [ 18 ] . These solutions seem to be related with the measure-valued solutions studied in [ 25 ] and approximated numerically in the context of the moment-SOS hierarchy. This motivated us to focus on reformulations with measures on infinite-dimensional spaces and to address the question of relaxation gap in this infinite-dimensional setup. In this paper, we show that there is no relaxation gap for a broad class of non-linear PDEs defined by evolution equations on infinite-dimensional functional spaces, with an operator satisfying a quasi-dissipativity condition also known as quasi-monotonicity or m -accretive property, explained in details in [ 14 , 49 , 32 ] . The key ingredient in the proof is differential calculus with probability measures [ 1 ] . No relaxation gap for this class of PDEs implies that we can use the infinite-dimensional moment-SOS hierarchy to solve optimization and optimal control problems on these equations, with convergence guarantees. It is of interest to mention [ 3 ] which also provides an alternative proof in the case of Hamiltonian PDEs, but without providing a numerical scheme nor making a clear link with the semigroup literature. To the best of our knowledge, this link is disclosed in our paper for the first time. Existing numerical methods for non-linear PDEs focus on specific classes of PDEs. Except conservation laws for which sophisticated methods such as the wave front tracking is available [ 21 ] , these methods focus on semi-linear equations and provide a numerical scheme for the linear part which is proved to work with the nonlinear terms of the equation by using some fixed-point arguments. Classical numerical schemes for PDEs are based on space(-time) discretizations include the finite-difference method [ 9 ] , the finite-element method [ 7 ] , the finite-volume method [ 15 ] or discontinuous Galerkin methods [ 8 ] . We refer the interested reader to [ 50 ] for a general overview of numerical schemes for nonlinear PDEs. Existing methods being based on fixed-point arguments and discretization (including implicit schemes), only local convergence can be guaranteed. Furthermore, when facing optimization problems involving PDEs, only local minima can be obtained with such methods. See e.g. [ 44 ]

Read the complete text