A geometric theory of control of systems with distributed parameters (Q677381)

Every reviewer likes to review articles like this one. It is written by one of the highly respected authors, a coauthor of original Pontryagin's theory, and a leading expert on geometric aspects of controlled differential systems (CDS). The author offers first a systematic, highly original and deep insight into geometric interpretation of systems with distributed parameters (SDP). The review of geometric aspects of control theory touches on a variety of points of view. He begins his discourse with remarks on control theory for SDP-s. Clearly, almost any differential equation or variational problem may be restated as a problem of control theory. For the autonomous system of equations \(dx/dt=f(x)\), \(x(0)=x_0\) we may regard time \(t\) as a control of the observation \(x(t)\). For any dynamic system, one can pose inverse problems that could be restated in control theoretic terms. Neither is there a clear distinction between systems with lumped parameters (SLP) and SDP-s. The author remarks the fact that several classes of SDP problems can be represented in finite-dimensional spaces. One has to recall the method of characteristics described in several textbooks [such as \textit{I. G. Petrovskij}, Lectures on partial differential equations, 3rd ed. (Russian) (1961; Zbl 0115.08103)], allowing to reduce systems of partial differential equations to systems of ordinary differential equations. Also the class SPD is so large that the author decided to restrict it to \((m,n)\)SDPs, where \(m\) denotes the number of independent variables, while \(n\) is the number of variables needed to describe the state of the system. The main purpose of the paper under review is to study the phase portraits, thus the geometry of the controlled differential systems (CDS). The author quotes a well-known observation that our intuition geared to Euclidean geometry in finite-dimensional spaces leads us astray in spaces of infinite dimensionality, and in non-Euclidean geometries. However if a system can be converted to a finite-dimensional one, or approximated by one, geometric intuition can greatly help us to understand problems arising in control theory. Monge structures, attainability-controllability are examples of mathematical structures (also called sets of relations, or mappings between mathematical objects). The author refers to the famous volumes of Bourbaki, as an example of systematic search for mathematical structures that are basic to all branches of mathematics. A methodological step was suggested in the Felix Klein Erlangen Program, where Klein classified different metrics, that is different geometries by invariance with respect to the action of different types of transformations. According to Emmy Noether each conservation law corresponds to some symmetry of the system. Thus one could study symmetries for control systems, for example for the \((m,n)\)CDSs. This extension of Klein's idea was attempted by the reviewer in a specific area of control theory called design theory [see for example, \textit{V. Komkov}, Z. Angew. Math. Mech. 61, 75-80 (1981; Zbl 0461.73076)], and by many authors in control theory. The problem consists of finding groups of transformations keeping the essential structure of the system invariant. The idea of lamination was introduced by the author and his collaborators, where layers represent trajectories. Thus control becomes identified with connectedness. Speaking broadly, connectedness is a rule for shifting from one layer to another. The author points out analogy with physics, and efforts to construct a unified field theory. The analogy can be made between calibration fields and control. ``Laws of nature'' supposedly ignore our units of measurement. There is an important consequence in global transformations. Scaling should commute with derivatives so that the basic structures are preserved. Should this rule be violated, new terms will appear that must be compensated for. This is the principle of local calibrational invariance, which is now extended to control theory. The author also mentions a number of extremely difficult problems of the TOKOMAK type, such as a heavy liquid flowing over a lighter liquid. Some of these problems where the number of degrees of freedom is finite have been successfully engineered. As an example we could point to the problem of stabilizing an arrow-shaped missile with the thrust applied at the end, pushing the arrow forward. The stabilizing controls anticipate small deviations from the intended flight path, constantly counteracting small perturbations. Following such a broad introduction, the author introduces a system of partial differential equations: \(\partial Q^\beta/\partial x^\alpha= f^\beta_\alpha(x,Q,u)\), \(\alpha= 1,2,\dots,m\); \(\beta=1,2,\dots,n\). Here \(u\) denotes admissible control. The summation convention is valid throughout this article. The author makes a careful distinction between the local states, and global states, which is the difference between Eulerian and Lagrangian approach to a motion of a continuous medium. In Euler's representation the geometry of state space is infinite-dimensional, while in Lagrange's representation the geometry is that of \(n+m\) space. Such \(n+m\) state space is now analyzed. Suppose that \(S^m\) is a trajectory. The corresponding tangent space \(T_yS^m\) at the point \(y\), is a subspace of the tangent space \(T_yD^{m+n}(y)\). Regarding \(y\) as a new variable (recalling that this is a Lagrangian representation), a system of equations is derived \(\partial y^\gamma/\partial x^\alpha= F^\gamma_\alpha(y,u)\), where \(F\) is an \((m+n)\times m\) matrix formed by adjoining the identity matrix \(I_m\) to the \(f^\beta_\alpha\) matrix. Integrability equations for this system are the usual compatibility equations of mechanics, implying that order of differentiation can be interchanged: \(\partial/\partial x^i(\partial Q^\beta/\partial x^\alpha)= \partial/\partial x^\alpha(\partial Q^\beta/\partial x^i)\). Instead of dealing with Cartesian tensors, the author prefers to define \(m\)-vectors, or multivectors, which is a set of \(m\) usual vectors \(\{v_1,v_2,\dots,v_m\}\) each vector \(v_i\in\mathbb{R}^{n+m}\). This turns out to be a wise decision, since later it permits the author to convert his CDS to a Pfaffian form. The inner product VW is given by \([\text{det}(v_i,w_j)]^{1/2}\). The indicator set for the system of equations \(\partial y^\gamma/\partial x^\alpha= F^\gamma_\alpha(y,u)\) is represented as a differential inclusion: \(\partial y/\partial x\in I(y)\), \(y\in D^{m+n}(y)\). Let \(K\) denote the field of cones of admissible directions. Thus multivector \(V\) is admissible if and only if there exists a \(\lambda>0\) and \(W\in I(y)\), \(y\in D^{m+n}(y)\), such that \(V=\lambda W\). The indicator set is non-empty and satisfies the inequality \(L(y,V)\leq 0\), while on the boundary \(L(y,V)= 0\). \(L\) is a smooth real function of all of its arguments. A familiar operation of homogenization is introduced, and an \((m,n)\)CDS system is rewritten as a Pfaffian form: \(\varphi^\beta_\gamma(y,u(y))dy^\gamma=0\). An existence condition for an admissible trajectory is that this Pfaffian system must be completely integrable. Advantages of Pfaffian notation are reasonably well-known to differential geometers, in particular the properties of cancellation of identical terms. Of course a Pfaffian form becomes an ordinary differential equation after division by \(d\tau\), where \(\tau\) is a parameter. The system of equations is: \(dx^\alpha/d\tau= w^\alpha\), \(dQ^\beta/d\tau= f^\beta_\alpha(x,Q,u)w^\alpha\), where \(w^\alpha\) is regarded as another control. This is a \((1,m+n)\)CDS, which can be converted to an \((m,n)\)CDS system of the form: \(\partial Q^\beta/\partial x^\alpha= f^\beta_\alpha(x,Q,u)\). Admissible trajectories are invariant manifolds of smallest dimension. It is obvious that this equation is closely related to the equation \(\partial G/\partial x^\alpha+ (\partial G/\partial Q^\beta)f^\beta_\alpha(x,Q,u(x,Q))= 0\) (just insert \(G=Q^\beta\)). In terms of \(y\), it can be rewritten as \(F^\gamma_\alpha(y,u(y))\partial G/\partial y^\gamma= 0\), \(\alpha=1,\dots,m\). Commutation algebra of vector fields in \(D^{m+n}(y)\) satisfies the completeness condition for fields \(F_\alpha\). Now a metric is introduced for \((m,n)\)CDS: \(L_0(y,V_m)= \{\inf(\lambda^{-1}),\lambda>0;\lambda V_m\in I(y)\}\). For \(m=1\) this is exactly the Minkowski metric. For any positive integer \(m\) it is a norm of the \(m\)-vector \(V:|V|= L_0(y,V_m)\). \(L_0\) can also be regarded as a measure, since it defines the volume spanned by \(V_m\) at the point \(y\). It is also an infimum for a multidimensional integral, solving a variational problem. Thus it could be called the Lagrangian for the corresponding optimal control. The author proposes to start an optimal control problem by defining the geometry, particularly the unit sphere \(L_0=1\). Examples follow this statement. Much of what follows is a rewriting of now classical Hamiltonian systems, with Pontryagin's Hamiltonian assuming its maximum value along an optimal trajectory. The flows on manifold formalism is used consistently in a detailed discussion of 1-CDS . The calibration factor turns out to be the density of the flowing substance. A global calibration invariance turns out to be the maximum principle providing necessary condition for optimization. The interesting aspect of this theory is broadening of the usual approach (for example Bellman's dynamic programming), whereby pieces of optimal trajectories are again optimal. Such pieces now do not have to preserve the original dimension, and it is not necessary to admit constructions having a uniform dimension. The author compares this aspect of optimal control theory to the spontaneous breaking of symmetry in modern physics. An unexpected bonus is a completely new principle of light propagation, following from a new version of a multidimensional maximality principle. This principle is valid within the framework of accuracy of geometric optics. It states that light is propagated after it leaves its source at a specific instant so that at any subsequent time it occupies the region of illumination which is the greatest possible volume compared to the volume it could reach without violation of its velocity at any point and in any direction. Moreover, the rate of increase of the volume of the region of illumination is also maximized. Since laws of geometric optics are invoked, the region of illumination is the region bounded by Huygens' wave front. Of course any other disturbance can be substituted for propagation of light, and this principle remains valid for the ``region of excitation''. The article closes with an example of reduction of a system of partial differential equations to the ``normal form'' of first-order equations with left-side independent of control, an example of a special form of indicatrix, and comments on the calibration-compensation approach to optimal control theory.