UNCLASSIFIED AD NUMBER LIMITATION CHANGES

Size: px

Start display at page:

Download "UNCLASSIFIED AD NUMBER LIMITATION CHANGES"

Debra Kristina Burke
5 years ago
Views:

1 TO: UNCLASSIFIED AD NUMBER AD LIMITATION CHANGES Approved for public release; distribution is unlimited. FROM: Distribution authorized to U.S. Gov't. agencies and their contractors; Administrative/Operational Use; 23 MAY Other requests shall be referred to Office of Naval Research, 875 North Randolph Street, Arlington, VA ONR ltr, 9 Nov 1977 AUTHORITY THIS PAGE IS UNCLASSIFIED

2 THIS REPORT HAS BEEN DELIMITED AND CLEARED FOR PUBLIC RtuEASE UNDER DOD DIRECTIVE AND NO RESTRICTIONS ARE IMPOSED UPON ITS USE AND DISCLOSURE. DISTRIBUTION STATEMENT A APPROVED FOR PUBLIC RELEASE; DISTRIBUTION UNLIMITED, mw'

UNCLASSIFIED Reproduced Armed Services Technical Information Agency ARLINGTON HALL STATION; ARLINGTON 12 VIRGINIA NOTICE: WHEN GOVERNMENT OR OTHER DRAWINGS, SPECIFICATIONS OR OTHER DATA ARE USED

GOVERNMENT THEREBY INCURS NO RESPONSIBILITY, NOR ANY OBLIGATION WHATSOEVER; AND THE FACT THAT THE GOVERNMENT MAY HAVE FORMULATED, FURNISHED, OR IN ANYWAY SUPPLIED THE SAID DRAWINGS,

3 UNCLASSIFIED Reproduced Armed Services Technical Information Agency ARLINGTON HALL STATION; ARLINGTON 12 VIRGINIA NOTICE: WHEN GOVERNMENT OR OTHER DRAWINGS, SPECIFICATIONS OR OTHER DATA ARE USED FOR ANY PURPOSE OTHER THAN IN CONNECTION WITH A DEFINITELY RELATED GOVERNMENT PROCUREMENT OPERATION, THE U. S. GOVERNMENT THEREBY INCURS NO RESPONSIBILITY, NOR ANY OBLIGATION WHATSOEVER; AND THE FACT THAT THE GOVERNMENT MAY HAVE FORMULATED, FURNISHED, OR IN ANYWAY SUPPLIED THE SAID DRAWINGS, SPECIFICATIONS, OR OTHER DATA IS NOT TO BE REGARDED BY IMPLICATION OR OTHERWISE AS IN ANY MANNER LICENSING THE HOLDER OR ANY OTHER PERSON OR CORPORATION, OR CONVEYING ANY RIGHTS OR PERMISSION IJO MANUFACTURE, USE OR SELL ANY PATENTED INVENTION THAT M^Y IN ^NY WAY BE RELATED THERETO. LASSIFIED

$84 MAY 23, 1960 ^ X o < 5 2 ^ -^ > 2 r " 5 «< ^ 3 <A i i c PREPARED UNDER CONTRACT Nonr-225{50)$

4 CONSTRAINT QUALIFICATIONS IN MAXIMIZATION PROBLEMS, II BY KENNETH J. ARROW AND HIROFUMI UZAWA TECHNICAL REPORT NO. 84 MAY 23, 1960 ^ X o < 5 2 ^ -^ > 2 r " 5 «< ^ 3 <A i i c PREPARED UNDER CONTRACT Nonr-225{50) (NR ) FOR OFFICE OF NAVAL RESEARCH A S T M k JUN 8 I960 TIPDR INSTITUTE FOR MATHEMATICAL STUDIES IN THE SOCIAL SCIENCES Applied Mathematics and Statistics Laboratories STANFORD UNIVERSITY Stanford, California

8k May 23, I960 PREPARED UNDER CONTRACT Nonr-225(50) (NR-O47-001+) FOR OFFICE OF NAVAL RESEARCH Reproduction in

5 1 CONSTRAINT QUALIFICATIONS IN MAXIMIZATION PROBLEMS, II by KENNETH J. ARROW and HIROFUMI UZAWA TECHNICAL REPORT NO. 8k May 23, I960 PREPARED UNDER CONTRACT Nonr-225(50) (NR-O ) FOR OFFICE OF NAVAL RESEARCH Reproduction in Whole or in Part is Permitted for Any Purpose of the United States Government INSTITUTE FOR MATHEMATICAL STUDIES IN SOCIAL SCIENCES Applied Mathematics and Statistics Lahoratories STANFORD UNIVERSITY Stanford, California

problems subject to inequality constraints. In the present paper, further results on the subject will be discussed and simplified proofs given. First, the Kuhn-Tucker Constraint Qualification ([5], p.

6 CONSTRAINT QUALIFICATIONS IN MAXIMIZATION PROBUMS, II Kenneth J. Arrow and Hlrofumi Uzawa 1. Introduction In the previous article [h], an investigation was made of the interrelationship between various conditions under which the classical Lagrange method remains valid for maximization problems subject to inequality constraints. In the present paper, further results on the subject will be discussed and simplified proofs given. First, the Kuhn-Tucker Constraint Qualification ([5], p. U83) will be slightlyweakened so that the meaning of the qualification becomes more straightforward. It will be shown in Theorem 1 below that the Lagrangian method can be applied to those constraint maxima for which the present version of the Constraint Qualification is satisfied. Then we show that the Constraint Qualification in the present formulation is the weakest requirement for the Lagrange method to be applicable; namely, in Theorem 2 below, it is proved that if the Lagrange method is justified for any differentiable maximand, then the constraint function satisfies the Constraint Qualification provided the constraint set is convex. Finally, in Section k it will be shown that the Constraint Qualification is implied by the condition that the constraint functions corresponding to a certain set of indices are convex. The latter condition Includes all those given earlier in [3] and [2].

$(i-i> > I )> 'the contained path with origin x and direction Is defined as an n-vector valued function ^(s) satisfying: (2) \ /(e) Is defined and Is continuous for all 0 < 9 < 9 with some positive 9;$

7 2. Definitions and Preliminary Remarks. Let g(x) = [g (x),..., g (x)] be an m-vector valued function defined for n-vectors x = (x., x...., x ). The set of all n-vectors x 1 n for which 6(x) > 0 will he called the constraint set defined hy g(x), denoted by C; I.e., (1) C = (x:g(x) > 0) Given an n-vector x In C and an arbitrary n-vector I = (i-i> > I )> 'the contained path with origin x and direction Is defined as an n-vector valued function ^(s) satisfying: (2) \ /(e) Is defined and Is continuous for all 0 < 9 < 9 with some positive 9; (3) + (0) = x; \ ;(e) e C for all 0 < 9 < 9; (k) i/ib) has a right-hand derivative at 9=0 such that i'(o) =. An n-vector for which there Is a contained path with origin x and direction will be referred to as an attainable direction; and the set of all attainable directions will be denoted by A. The closure of the convex cone spanned by A will be denoted by K ; the elements of K will be referred to as weakly attainable directions.

$tfz\ (5) E = {J:g d (x) = 0], and F is the set of all ineffective indices at x; namely, (6) F = [j : g J C*) >0).$ Let K- he the set of n-vectors defined hy (7) K 2 = u-sl i>o) ; the elements of K may be termed locally constrained directions. constrained. Lemma 1.

Let K- he the set of n-vectors defined hy (7) K 2 = u-sl i>o) ; the elements of K may be termed locally constrained directions. constrained. Lemma 1.

8 r - 3 The set of Indices {1,..., m) Is divided Into two parts E and F. E is the set of all Indices effective Indices at x; namely. tfz\ (5) E = {J:g d (x) = 0], and F is the set of all ineffective indices at x; namely, (6) F = [j : g J C*) >0). Let K- he the set of n-vectors defined hy (7) K 2 = u-sl i>o) ; the elements of K may be termed locally constrained directions. constrained. Lemma 1. Every weakly attainable direction is locally Proof: Let t( e ) lde a contained path with origin x and direction. Then, for any j e E, Hence g J IX0)] = 0 and g J [*(9)] > 0. g^ I > 0. Q.E.D.

I k - A function g(x) will be called to satisfy the (Kuhn-Tucker) Constraint Qualification at x if (CQ) K 2 C ^ ^ i.e., every locally constrained direction is weakly attainable.

$Let us define the set K by (8) K = the closure of the set {\(x-x):\ > 0, x e C). Lemma 2.$

9 I k - A function g(x) will be called to satisfy the (Kuhn-Tucker) Constraint Qualification at x if (CQ) K 2 C ^ ^ i.e., every locally constrained direction is weakly attainable. Kuhn and Tucker ([5]^ P> ^83) required that every locally constrained direction be attainable. Let us define the set K by (8) K = the closure of the set {\(x-x):\ > 0, x e C). Lemma 2. If the constraint set C ±s_ convex, then K 3 CK l * Proof; If x e C, then by the convexity of the set C, x + e(x - x) C for all 0 < e < 1 Hence, x - x is attainable and therefore weakly attainable. Since K- is a cone, the conclusion follows.

lagrange Regularity and the Constraint Qualification. The classical lagrange method for constrained maxima (see, e.g. [l], p.

10 T Let B be any set of vectors. The negative polar cone, to be denoted by B", IS defined by B" = {u:ux < 0 for all x e B} We have (see, e.g., Fenchel [k], pp. 8-10), (9) B' IS a closed convex conej (10) B ic- B 2 lm P lies that B i ^B2 j (11) if B is a closed convex cone, B" = B. 3. lagrange Regularity and the Constraint Qualification. The classical lagrange method for constrained maxima (see, e.g. [l], p. 153) has been adapted by Kuhn and Tucker ([5], Theorem 1, p. h6h). (L) If x maximizes a differentlable function f(x) subject to x e C, then there exists a nonnegative m-vector y such that (12) f x + y 8 X = 0, (13) y g(x) = 0.

direction, then Proof; Let i r(9) be a contained path with origin x and Then f[ilf(e)] < fu(0)] = f(x) for all 0 < 6 <.

11 f - 6 An m-vector valued function g(x) will be termed lagrange regular If, for any dlfferentlable function f(x), the condition (L) holds. Lemma 3. If x maximizes f(x) subject to x 6 C, then f x e ^ ' vhere KL - i_s the negative polar cone of JL. direction, then Proof; Let i r(9) be a contained path with origin x and Then f[ilf(e)] < fu(0)] = f(x) for all 0 < 6 <. f x I = f x ' (0) < 0, for any in A and, by continuity and convexity, for any K. Q.E.D. Theorem 1. If g(x) satisfies the Constraint Qualification (CQ), then g(x) is Lagrange regular.

Minkowski-Farkas Lemma we have from (7) and -E -E -E -f = y g for some y > 0. x " D x ^ Define - /-E -F» -F y = (y, y ) with y = 0.

12 Proof; Let f (x) be a dlfferentiable function and x maximize f (x) subject to x C. Then, by Lemma 3> f ^C On the other hand, the condition (CQ) implies, from (lo), that Hence, we have (Ik) f x c K- (Ik), By applying the Minkowski-Farkas Lemma we have from (7) and -E -E -E -f = y g for some y > 0. x " D x ^ Define - /-E -F» -F y = (y, y ) with y = 0. Then, y satisfies conditions (12) and (13). Q.E.D. Theorem 1 is the basic necessity theorem for non-linear programming ([5]> Theorem l) extended to the weaker constraint qualification of this paper.

$Proof: It will be first shown that (15) K 3 CK 2 Let a e K ; then for \ = 1, (16) a(x - x) < 0 for all x c C Then x maximizes$

13 fl Theorem 2. If g(x) l lagrange regular and if the constraint set C defined ^t It Is. a convex set, then g(x) satisfies the Constraint Qualification (CQ). Proof: It will be first shown that (15) K 3 CK 2 Let a e K ; then for \ = 1, (16) a(x - x) < 0 for all x c C Then x maximizes the function f (x) = ax subject to x e C. By the Lagrange regularity of g(x), there is an m-vector y such that (17) a + y g x = 0 ; y > o, and (18) y g(x) = 0. The conditions (17) and (18) imply that (19) _E -E a + y g x = 0, y E > 0

A Sufficient Condition for the Constraint Qualification.

14 The condition (19) Implies that ag < 0 for all such that g > 0 ; namely, a e K Hence, we have the relation (15). Then, by (lo) and (ll). (20) K 3 3K 2 Applying Lemma 2, ^ 3)K 3 DK 2 Q.E.D. k. A Sufficient Condition for the Constraint Qualification. Let E' be a subset of the effective Indices defined by (21) E' ={J:jeE, ug =0, for some u > 0 with u > 0) x = J Theorem 3- If g (x) Is convex for every J e E', then g(x) satisfies the Constraint Qualification (CQ) at x.

Hence, for any J e E" = E - E', there exists J such that (23) '4 I* > 0 ' ü J > 0 for all k e E, k ^ J, * Let i2k)

15 10 - Proof: By the Mlnkowskl-Parkas Lemma, an Index J belongs to E' If and only If (22) ^ = 0 for a11 6 such that 8 k 6 ^ 0 > k e E, k / J. Hence, for any J e E" = E - E', there exists J such that (23) '4 I* > 0 ' ü J > 0 for all k e E, k ^ J, * Let i2k) I* = I l J JeE" Then, by (23) we have (25) -E" * g I > 0. By (22) and the definition (7) of K, we have Hence, -E 1 g x l=o for all I e K (26) gxl =0 for all I e L, where L is the linear space spanned by K. 1

For any positive a > 0, define (28) *(e) = 5 + ( + ag*) e.

16 11 - Since g J (x) is convex for J e E 1, the condition (26) implies that gj d (x + ) has its minimum in L at = 0; hence. (27) g E, (x + ) > 0 for all I e L, since E 1 C E^ so that g^ (x) =0. We shall now prove that IC is contained in KL. Let 6 be any element of IC. For any positive a > 0, define (28) *(e) = 5 + ( + ag*) e. From (27), (29) g [t(e)] > 0 for all 9 > 0 From (25) and the definition (7) of K, dg E "u(e)] de e=o -E"/. *\ E" _E" * = g x (I + a ) = g x 6 + a gj I > 0 Since E' C E, we have g^'t^o)] = 0; hence. (30) g*" [ + (9)] > 0 for 9 > 0 and sufficiently small.

$. Hence ${6) is a contained path, and + 0^ Is an attainable direction for all a > 0. Therefore, being a limit of + a as a tends to zero, is weakly attainable. Q.E.D.$ Obviously, the conclusion of Theorem 3 will hold if either g J (x) is convex for all j or E' is the null set.

Obviously, the conclusion of Theorem 3 will hold if either g J (x) is convex for all j or E' is the null set.

17 12 Finally, g [+(0)] > 0, so that (31) ^[^(e)] > 0 for 6 > 0 and sufficiently small. The relations (29-31) imply that g[+(e)] > 0 for 9 sufficiently small.. Hence ${6) is a contained path, and + 0^ Is an attainable direction for all a > 0. Therefore, being a limit of + a as a tends to zero, is weakly attainable. Q.E.D. Obviously, the conclusion of Theorem 3 will hold if either g J (x) is convex for all j or E' is the null set. Corollary 1: If g J (x) ls_ convex (in particular, linear) for every J, then g(x) satisfies the Constraint Qualification. Corollary 1 extends Theorem 2 of [3]. Corollary 2; ([3], Theorem 3). Suppose g(x) i concave, and g (x ) > 0 for some x. Then g(x) satisfies the Constraint Qualification. I

' This condition was used by Slater [6] for f(x) concave. Corollary 3; ([2], Theorem 2, p. 18).

18 13 Proof; Since g (x) Is concave. g (x ) - g (x) < g x (x - x) E ** E But g (x ) > 0 and g (x) = 0; hence. i^(x* - x) > 0 If ug x = 0, then ug x (x - x) = 0, which, for u > 0, Is only possible for u = 0. Then E 1 is null. Q.E.D. ' This condition was used by Slater [6] for f(x) concave. Corollary 3; ([2], Theorem 2, p. 18). Suppose (32) g(x) l quasi-concave; (33) g (x ) > 0, for some x ; and (3^) for each J e E, either g J (x) concave or ß^ ^ 0. Then the Constraint Qualification is satisfied. Proof: If g (x) is concave, for j E, then by (33) (35) 6^/0

If, for some j e E, <e -0 for all I e JC Then g^i =0 for all? j hence -4.0, contradicting (35). Therefore, the set E' is null. Q.E.D. Corollary k: (non-degeneracy).

19 - Ik. Hence, by assumption (3^), the relation (35) holds for all J e E. Since the set C Is convex, x - x Is attainable for all x e C (see Lemma 2). Since x is an interior point of C, x - x is an Interior point of A; hence, the set L, the space spanned by K_, is the entire space. If, for some j e E, <e -0 for all I e JC Then g^i =0 for all? j hence -4.0, contradicting (35). Therefore, the set E' is null. Q.E.D. Corollary k: (non-degeneracy). If g has a rank equal to the number of effective constraints, then g(x) satisfies the Constraint Qualification. Proof: By the assumption of the maximum rank -E ug =0 implies u = 0 j hence, the set E' is null. Q.E.D. 1

20 1^^* 15 Remark: The case of nonlinear equality contraints Is not handled hy Theorem 3 or Corollaries 1-3- Indeed, It could easily happen In that case there Is no linear contained path. Corollary h does remain valid if some or all constraints are equalities; see [3]> Appendix 1. -

ii - 16 - REFERENCES [l] Apostol, T. M. Mathematical Analysis; A Modern Approach to Advanced Calculus. Reading, Mass.: Add ison-wesley, 1957- [2] Arrow, K. J., and A. C. Enthoven.

" Proceedings of the Third Berkeley Symposium on Mathematical Statistics and Probability, Berkeley and Los Angeles: University of California Press, 1956, Volume V, pp. 1-20. [h] Arrow, K. J., L.

21 ii REFERENCES [l] Apostol, T. M. Mathematical Analysis; A Modern Approach to Advanced Calculus. Reading, Mass.: Add ison-wesley, [2] Arrow, K. J., and A. C. Enthoven. "Quasi-Concave Programming". The RAND Corporation, P-1847, December 16, [3] Arrow, K. J., and L. Hurwicz. "Reduction of Constrained Maxima to Saddle-Point Problems." Proceedings of the Third Berkeley Symposium on Mathematical Statistics and Probability, Berkeley and Los Angeles: University of California Press, 1956, Volume V, pp [h] Arrow, K. J., L. Hurwicz, and H. Uzawa. "Constraint Qualification in Maximization Problems," Naval Research Logistics Quarterly, (forthcoming). [5] Fenchel, W. Convex Cones, Sets and Functions, Princeton University, 1953 (hectographed). [6] Kuhn, H. W., and A. W. Tucker. "Nonlinear Programming", In J. Neyman (ed) Proceedings of the Second Berkeley Symposium on Mathematical Statistics and Probability. Berkeley and Los Angeles: University of California Press, 1951, pp. 481-^92. [7] Slater, M. "lägrange Multipliers Revisited; A Contribution to Non-linear Prograinmlng." Cowles Commission Discussion Paper, Math. 1+03, November, V

UNCLASSIFIED. .n ARMED SERVICES TECHNICAL INFORMATION AGENCY ARLINGTON HALL STATION ARLINGTON 12, VIRGINIA UNCLASSIFIED

UNCLASSIFIED. .n ARMED SERVICES TECHNICAL INFORMATION AGENCY ARLINGTON HALL STATION ARLINGTON 12, VIRGINIA UNCLASSIFIED UNCLASSIFIED.n 260610 ARMED SERVICES TECHNICAL INFORMATION AGENCY ARLINGTON HALL STATION ARLINGTON 12, VIRGINIA UNCLASSIFIED NOTICE: When government or other drawings, specifications or other data are