Generalized characteristics of first order PDEs. Applications in optimal control and differential games

Publication date 9 July 1998

calculus of variations differential games dynamic programming first order PDE generalized characteristics optimal control value function viscosity solution

Mathematics Subject Classification ID

Research exposition (monographs, survey articles) pertaining to partial differential equations (35-02) General first-order partial differential equations and systems of first-order partial differential equations (35Fxx) Existence theories for optimal control problems involving partial differential equations (49J20) Optimality conditions for problems involving partial differential equations (49K20) Differential games (aspects of game theory) (91A23)

Review text

The main subject of this book is the local study of certain types of generalized solutions of first order PDEs of the form: \[ F(x,u(x),Du(x))=0, \;x\in D\subset\mathbb{R}^n, \;u(x)=w(x), \;x\in M\subset D, \;(x,u,p)\in A\subset\mathbb{R}^{2n+1} \tag{1} \] illustrated by applications to some problems in optimal control and differential games. In comparison with two other books recently published by the same publisher on the same subject [\textit{A. I. Subbotin}, ``Generalized solutions of first-order PDEs, Birkhäuser, Boston (1994; Zbl 0820.35003); \textit{M. Bardi} and \textit{I. Capuzzo-Dolcetta}, ``Optimal control and viscosity solutions of Hamilton-Jacobi-Bellman equations, Birkhäuser, Boston (1997; Zbl 0890.49011)], the present one may be characterized by the priority given to Cauchy's method of characteristics. A global characterization of the content of the book is provided by the author himself in the Preface in the following terms: ``Traditionally, only the (characteristic) field on the surface of codimension one was associated with the first order PDEs, the nonsmoothness of the solution and/or of the Hamiltonian being considered as an obstacle for the application of the classical MC (i.e. ``Method of characteristics). The main finding of this book says that one can overcome this obstacle using the same characteristics, but defined on the appropriate surface of codimension 3,5,\dots{} . More precisely, this main idea of the author seems to be the following: assuming that a ``locus of nonsmooth points either of the ``extended Hamiltonian \(F(.,.,.)\) in (1) or of the solution \(u(.)\) is a differentiable manifold of the form: \[ W_{2m+1} =\{z=(x,u,p)\in\mathbb{R}^{2n+1}; \;F_i(z)=0, \;i=-m,-m+1,... \;m \} \tag{2} \] then (under certain hypotheses) one looks for some smooth functions, \(\lambda _i(.)\) such that if the ``singular Hamiltonian is defined by: \( H(x,u,p):=\sum_{i=-m}^{m}\lambda_i(z)F_i(z)\), \(z=(x,u,p) \) then the ``singular characteristic vector field defined by: \[ \xi_H(x,u,p):=(H_p(x,u,p),\langle p,H_p(x,u,p)\rangle,- H_x(x,u,p)-pH_u(x,u,p)) \tag{3} \] is tangent to the manifold in (2). From this point of view, the author does not seem to be aware of the reviewer's more general approach in [\textit{Şt. Mirică}, Rend. Sem. Mat. Univ. Padova 77, 317-350 (1987; Zbl 0627.49015)] where to a countably \(C^1\)- stratified extended Hamiltonian \(F(.,.,.)\) has been associated a ``stratified characteristic orientor field: \[ \xi_F^S(x,u,p):=\{(x',\langle p,x'\rangle,p')\in T_{(x,u,p)}A; \;DF(x,u,p).(y,\langle p,y\rangle, q)=\langle x',q\rangle-\langle p',y\rangle \] \[ \forall \;(y,\langle p,y\rangle,q)\in T_{(x,u,p)}A\} \] which, under the same hypotheses, coincides with the one in (3) on the strata of the form in (2). The largest part of the book consists in the use of the ``regular and of the above ``singular characteristics to characterize the local behaviour of solutions of the problem in (1) in neighbourhoods of points of the manifolds of the type (2). From this point of view, this book looks similar, though at a much higher level of generality and mathematical rigour, to the well known book of \textit{R. Isaacs}' book on ``Differential games [Wiley, N.Y. (1965; Zbl 0125.38001)]. Thus, the author develop a theory of ``echivocal, ``dispersal, ``switching, ``universal, ``focal (regular, semiuniversal, irregular, etc.), ``geometrically indifferent , etc., surfaces that may be associated to a nonsmooth solution of the problem in (1) especially in optimal control and the theory of differential games. Pursuing a rather elaborate local analysis of possible solutions of problems of the form in (1), the author does not mention the fact that, at least in optimal control and differential games, a study ``in the large, as R. Isaacs calls it or, more precisely, the use of a ``verification theorem, in the terminology of [\textit{W. H. Fleming} and \textit{R. Rishel}, ``Deterministic and Stochastic Optimal Control, Springer, New York (1975; Zbl 0323.49001)], remains the decisive argument proving the optimality of the trajectories and the coincidence of the solution of (1) with the value function of the original problem; after all, in a neighbourhood of the point under analysis may ``arrive trajectories from other neighbourhoods that may realize better performances. Among several other things, as it is the ``heuristic derivation of the Euler-Lagrange equation and of the Hamilton-Jacobi-Bellman equation in calculus of variations and optimal control in sections 3.1-3.3, one may note two other rather ``hazardous statements in Section 3.4: a) In ``Variation Calculus (terminology of the author) the Hamiltonian is smooth and convex in \(p\); b) In ``Variation Calculus, optimal control and differential games the Hamilton-Jacobi-Bellman-Isaacs equation possess a unique viscosity solution and moreover, this coincides with the value function of the problem. From the multitude of well known examples contradicting both statements [e.g., \textit{L. Cesari}, ``Optimization -- theory and applications, Springer, New York (1983; Zbl 0506.49001)] let us mention only the most famous ones: 1) Bernoulli's Brachistochrone problem in parametric form has the Hamiltonian: \[ H(x,p):=0 \text{ if }\|p\|\leq 1/ \sqrt {x_2+k}, \;H(x,p):=-\infty \text{ otherwise, }(x,p)\in\mathbb{R}^2\times\mathbb{R}^2 \] which, obviously is not smooth, and the corresponding Hamilton-Jacobi-Bellman equation: \[ H(x,Du(x))=0, \;u(0,0)=0 \tag{4} \] has two distinct classical (hence also viscosity) solutions: \(u_1(x)\equiv 0 \neq u_2 (.)=\)value function; 2) Plateau's problem for surfaces of revolution in parametric form has the Hamiltonian: \[ H(x,p):=0 \text{ if }\|p\|\leq x_2, \;H(x,p):=-\infty \text{ otherwise, } (x,p)\in \mathbb{R}^2 \times \mathbb{R}^2 \] which is obviously not smooth, and the corresponding Hamilton-Jacobi-Bellman equation in (4) has also (at least) two distinct viscosity solutions. However, whatever the shortcomings, many of the results and computations in this book may turn out to be very useful in the study of certain particular problems in Optimal control and differential games. Reviewer's note on terminology: Throughout this book it is used the rather strange term ``variation calculus and occasionally that of ``variational problem while the classics in the field (e.g. O. Bolza, G. Bliss , C. Carathéodory, etc.) use the term ``Calculus of variations; on the other hand, in some recent books and articles one uses the term ``Variational calculus which seems to be ``dangerously close to that of ``Variational methods which is something different if not just the opposite (Calculus of variations reduces an optimization problem to a functional equation while the variational method reduces a functional equation to an optimization problem).

Recommendations

Cited in

(54)

This page was built for publication: Generalized characteristics of first order PDEs. Applications in optimal control and differential games

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q4397692)