econometrics Turnpike Properties of Optimal Control Systems by: Alexander J. Zaslavski 1. Introduction the turnpike theory we study the structure of solutions when an objective function (an optimality criterion) is In this paper we discuss recent progress in the turnpike fixed while T1,T2 and the data vary. To have turnpike theory which is one of our primary areas of research. properties means that the solutions of a problem are Turnpike properties are well known in mathematical determined mainly by the objective function (optimality economics. The term was first coined by Samuelson (see criterion), and are essentially independent of the choice [1]) who showed that an efficient expanding economy of time interval and data, except in regions close to the would for most of the time be in the vicinity of a balanced endpoints of the time interval. If a real number t does not equilibrium path. These properties were studied by many belong to these regions, then the value of a solution at researches for optimal paths of models of economic the point t is closed to a “turnpike” - a trajectory (path) dynamics determined by set-valued mappings with which is defined on the infinite time interval and depends convex graphs. In our recent book [5] we present a only on the objective function (optimality criterion). This number of turnpike results in the calculus of variations, phenomenon has the following interpretation. If one optimal control, the game theory and in economic wishes to reach a point A from a point B by a car in an dynamics obtained by the author. The results collected optimal way, then one should enter onto a turnpike, spend in [5] demonstrate that the turnpike properties are a most of one’s time on it and then leave the turnpike to general phenomenon which holds for various classes reach the required point. of variational problems and optimal control problems P.A. Samuelson discovered the turnpike phenomenon arising in engineering and in models of economic growth. in a specific situation in 1948. In further studies Turnpike properties are studied for optimal control turnpike results were obtained under certain rather problems on finite time intervals [T1,T2] such that T1 < strong assumptions on an objective function (optimality T2. Here T1,T2 are real numbers in the case of continuouscriterion). Usually it was assumed that an objective time problems and are integers in the case of discretefunction is convex, as a function of all its variables and time problems. Solutions of such problems (trajectories does not depend on the time variable t. In this case it or paths) depend on an optimality criterion determined was shown that the “turnpike“ is a stationary trajectory by an objective function (integrand), the time interval (a singleton). [T1,T2], and on data which is some initial conditions. In Since convexity assumptions usually hold for models of economic growth, turnpike theory has many applications in mathematical Alexander J. Zaslavski economics. There are several turnpike Alexander J. Zaslavski received his doctorate at the results for nonconvex (noncocave) Institute of Mathematics of the Siberian branch of problems but for these problems the Soviet Academy of Sciences in Novosibirsk. convexity (concavity) was replaced by He is a senior researcher at the Department of other restrictive assumptions which Mathematics of the Technion - Israel Institute hold for narrow classes of problems. of Technology. His research interests are the Therefore experts considered the optimization theory, the calculus of variations, turnpike phenomenon as an interesting optimal control, the dynamical systems theory, and important property of some very nonlinear analysis, the game theory and models particular optimal control systems with of economic dynamics. He is an author of 400 origin in mathematical economics and research papers and three monographs. for which a “turnpike” was usually a singleton or a half-ray. This situation has changed in the last period of time which 36 AENORM vol. 20 (77) December 2012 econometrics began at 1995 when the works [2, 3, 4] appeared. In [2] we studied a general class of unconstrained discrete-time optimal control problems and established that a turnpike property holds for a typical (generic) problem and that a turnpike is a set which is not necessarily a singleton. For this class of problems the turnpike can be a singleton but rather seldom. In [3, 4] we studied the turnpike properties of extremals of one-dimensional second order variational problems arising in the theory of thermodynamical equilibrium for materials. We showed that for this class of problems the turnpike is a periodic curve which is not necessarily a singleton. In our book [5] we collected turnpike results which were obtained after 1995 and for which we do not need convexity of an objective function and its time independence. The results of [5] allow us today to think about turnpike properties as a general phenomenon which holds for various classes of optimal control problems. It was my great pleasure to receive on October 2000 the following letter from Paul A. Samuelson, the discoverer of the turnpike phenomenon. Here T is a real number, y and z are points of the space Rn and an integrand is strictly convex and differentiable function such that If a reader is not familiar with absolutely continuous functions, it is possible to assume that functions v are continuously differentiable or at least piecewise C1 functions. We intend to study the behavior of extremals of the problem (P0) when the points y,z and the real number T vary and T is sufficiently large. In order to meet our goal let us consider the following auxiliary minimization problem: (P1) By the strict convexity of f and the growth condition, the problem (P1) possesses a unique solution . It is easy to see that Dear Professor Zaslavski: Define I note with interest your long paper “The Turnpike Property ...Functions” in Nonlinear Analysis 42 (2000), 1465-98. It may be of interest to report that this property and name originated just over half a century ago when, as a Guggenheim Fellow on a 1948-49 sabbatical leave from MIT, I conjectured it in a memo written at the RAND Corporation in Santa Monica, California. In The Collected Scientific Papers of Paul A. Samuelson, MIT Press, 1966, 1972, 1977, 1986, it is reproduced. R. Dorfman, P.A. Samuelson, R.M. Solow, Linear Programming and Economic Analysis, McGraw-Hill, 1958 gives a pre-Roy Radner exposition. I believe that somewhere Lionel McKenzie has given a nice survey of the relevant mathematical-economics literature. With admiration, Paul A. Samuelson It is easy to see that the integrand is a differentiable and strictly convex function such that Since the functions have and L are both strictly convex we and Consider an auxiliary variational problem 2. Problems with convex integrands Let | . | be the Euclidean norm in the n-dimensional Euclidean space Rn and let < . , . > be the scalar product in Rn. We consider the variational problem (P0) (P2) is an absolutely continuous function such that v(0) = y, v(T) = z; where T > 0 and . Clearly, for any positive number T and any absolutely continuous function we have is an absolutely continuous function such that AENORM vol. 20 (77) December 2012 37 econometrics It follows from the equations above that the problems (P0) and (P2) are equivalent. Namely, a function is a solution of the problem (P0 ) if and only if it is a solution of the problem (P2). We claim that the integrand possesses the following property (C): Clearly, the integrals do not exceed a positive constant c0(|y|,|z|) which depends only on the norms |y|,|z| and does not depend on T. Therefore then Assume that a sequence satisfies = 0: The growth condition implies that the sequence is bounded. Let (y, z) be its limit point. Then, This implies that , as claimed. Assume that y,z are points of the space Rn, T > 2 is a real number and that an absolutely continuous function is an optimal solution of the problem (P0). Since the problems (P0) and (P2) are equivalent the function is also an optimal solution of the problem (P2). We claim that where a positive constants c0(|y|,|z|) depends only on |y| and |z|. Consider an absolutely continuous function defined by In the sequel we denote by mes(E) the Lebesgue measure of a Lebesgue measurable set . If a reader is not familiar with the Lebesgue measure theory, we can say, roughly speaking, that a set of real numbers E is Lebesgue measurable if it is (in some sense) the limit of a sequence of sets of real numbers such that each Ei is a finite union of open intervals with no intersections. In this case where the Lebesgue measure of a set which is a finite union of open intervals with no intersections is the sum of their lengths. Now let be given. It follows from the property (C) that there exists a positive number such that for each point which satisfies In view of the choice of the constant and the inequality we have and By the definition of the functions 38 AENORM vol. 20 (77) and x, we have December 2012 It is easy now to see that the optimal solution spends most of the time in an - neighborhood of . By the inequality above, the Lebesgue measure of the set of all real numbers t, such that (t) does not belong to this -neighborhood, does not exceed the constant which depends only on |y|,|z| and and does not depend on T. Following the tradition, the point is called the turnpike. Moreover we can show that the set econometrics is contained in the union , where of two intervals 3. Nonconvex nonautonomous integrands We showed in the previous section that the structure of optimal solutions of the problem (P0), under the assumptions posed on f, is rather simple and the turnpike is calculated easily as a solution of the problem (P1). Nevertheless, the convexity of the integrand f and its time independence are very essential for the proof of this turnpike result. In order to obtain a turnpike result for essentially larger classes of variational problems and optimal control problems we need other methods and ideas. The following example helps to understand what happens if the integrand f is nonconvex and nonautonomous and what kind of turnpike we have for general nonconvex nonautonomous integrands. Consider an integrand defined by and the family of the problems of the calculus of variations (P3): where y, z, T1, T2 are real numbers and T2 > T1. It is clear that the functions f depends on the time variable t, for each real number t, the function f(t, . , . ) : is convex, and for each pair of numbers the function is nonconvex. Hence the function is also nonconvex and depends on t. Let y, z, T1, T2 be real numbers, T2 > T1 +2 and let a function R1 be an optimal solution of the problem (P3). Note that the problem (P3) possesses a solution since the integrand f is a continuous function and the function is convex and grows superlinearly at infinity for each point Consider a function defined by Clearly, and Therefore where It is easy to see that for any real number following inequality holds: the Since the constant c1(|y|,|z|) does not depend on T2 and T1 it follows from the inequality above if the length of the interval T2 - T1 is sufficiently large, then the function is equal to cos(t) up to for most . Again, as in the case considered in the previous section, we can show that where > 0 is a constant which depends only on , |y| and |z|. This example demonstrates that there are nonconvex time dependent integrands for which the turnpike property holds with the same type of convergence as in the case of convex autonomous variational problems, but with the the turnpike which is an absolutely continuous time dependent function defined on the infinite interval . This leads us to the following definition of the turnpike property for general integrands. Consider the problem of the calculus of variations (P) AENORM vol. 20 (77) December 2012 39 econometrics is an absolutely continuous function such that v(T1) = y, v(T2) = z. Here T1 < T2 are real numbers, and the function Rn is continuous. We say that the integrand f possesses the turnpike property if there exists a locally absolutely continuous function (called the “turnpike”) which depends only on f such that the following condition holds: For each bounded subset K of the space Rn and each positive number there exists a positive constant T(K, ) such that for each pair of real numbers T1 0 and T2 T1+ 2T(K, ), each pair of points and each optimal solution of the problem (P), we have The turnpike property is very important for applications. Assume that the integrand f possesses the turnpike property, the bounded set K and a small positive number are given, and we know a finite number of “approximate” solutions of the problem (P). Then we know the turnpike Xf , or at least its approximation, and the positive constant T(K, ) which is an estimate for the time period required to reach the turnpike. We can use this information if we need to find an “approximate” solution of the problem (P) with a new time interval [T1, T2] and the new values at the end points T1 and T2. More precisely, instead of solving this new problem on the “large” interval [T1, T2] we can find an “approximate” solution of problem (P) on the “small” interval [T1, T1 + T(K; )] with the values y, Xf (T1 + T(K; )) at the end points and an “approximate” solution of problem (P) on the “small” interval [T2 - T(K; ), T2] with the values Xf (T2 - T(K; )), z at the end points. Then the concatenation of the first solution, the function Xf :[T1 + T(K; ), T2 + T(K; )] and the second solution is an “approximate” solution of problem (P) on the interval [T1, T2] with the values y, z at the end points. In Chapter 2 of [5] we consider a general space of continuous integrands which is endowed with a natural complete metric. We establish the existence of a set which is a countable intersection of open everywhere dense sets in such that for each the turnpike property holds. Moreover we show that the turnpike property holds for approximate solutions of variational problems with a integrand and that the turnpike phenomenon is stable under small perturbations of f. 40 AENORM vol. 20 (77) December 2012 References P. A. Samuelson, “A catenary turnpike theorem involving consumption and the golden rule”, American Economic Review 55 (1965), 486-496 A. J. Zaslavski, “Optimal programs on infinite horizon 1,2“, SIAM Journal on Control and Optimization 33 (1995), 1643-1686 A. J. Zaslavski, “The existence and structure of extremals for a class of second order infinite horizon variational problems”, “Journal of Mathematical Analysis and Applications 194 (1995), 459-476 A. J. Zaslavski, “Structure of extremals for onedimensional variational problems arising in continuum mechanics“, Journal of mathematical Analysis and Applications 198 (1996), 893-921 A. J. Zaslavski, “Turnpike properties in the calculus of variations and optimal control”, Springer, New York, 2006
© Copyright 2026 Paperzz