A TEXT FOR NONLINEAR PROGRAMMING Thomas W. Reiland Department of Statistics North Carolina State University Raleigh, NC 27695-8203 Email: reiland@ncsu.edu Table of Contents Chapter I: Optimality Conditions..................................................Error! Bookmark not defined. § 1.1 Differentiability.................................................................Error! Bookmark not defined. § 1.2. Unconstrained Optimization ............................................................................................. 0 § 1.3 Equality Constrained Optimization...................................Error! Bookmark not defined. § 1.3.1 Interpretation of Lagrange Multipliers..........................Error! Bookmark not defined. § 1.3.2 Second order Conditions – Equality Constraints ..........Error! Bookmark not defined. § 1.3.3 The General Case ..........................................................Error! Bookmark not defined. § 1.4 Inequality Constrained Optimization ...............................Error! Bookmark not defined. § 1.5 Constraint Qualifications (CQ) ........................................Error! Bookmark not defined. § 1.6 Second-order Optimality Conditions ...............................Error! Bookmark not defined. § 1.7 Constraint Qualifications and Relationships Among Constraint Qualifications ..... Error! Bookmark not defined. Chapter II: Convexity ...................................................................Error! Bookmark not defined. § 2.1 Convex Sets ......................................................................Error! Bookmark not defined. § 2.2 Convex Functions ............................................................Error! Bookmark not defined. § 2.3 Subgradients and differentiable convex functions ............Error! Bookmark not defined. 1.2 Unconstrained Optimization 1.2-1 § 1.2. Unconstrained Optimization Consider the following problem: Max f ( x), x E n , f has continuous 1st order partial derivatives (i.e. differentiable) Theorem 1.1 Suppose there is a direction d E n such that f ( x0 )T d 0 . Then there exists 0 such that for t , 0 t , f ( x0 td ) f ( x0 ) . f ( x 0 td ) f ( x 0 ) 0 t 0 t By the definition of a limit, there exists 0 for 0 t f ( x 0 td ) f ( x 0 ) 0 (eventually this is positive) t Hence for any t, 0 t , then f ( x0 td ) f ( x0 ) 0 . QED. Proof: f ( x 0 )T d lim Corollary 1.1 If x is a finite local maximum of f ( x) then f ( x ) 0 (1) Proof: Suppose f ( x ) 0 ; choose d f ( x ) . Then f ( x )T f ( x ) 0 Theorem 1.1 implies that there exists 0 such that f ( x d ) f ( x ) for 0 # This contradicts the fact that x is a local maximum. QED. Remark: Moreover, (1) holds if x is any local optimum of f ( x) . (1) is a necessary condition to identify candidate points and is not sufficient to guarantee optimality. These candidates are often called critical (stationary) points and include local max/min points, global max/min points, and saddle points. Example 1-3: f ( x) x3 , f ( x) 3x 2 , 3x 2 0 x 0 but this is only an inflection point. Theorem 1.2 Let the function f ( x), x E n , have continuous 2nd order partial derivatives. f ( x ) 0 If (2) n And zT 2 f ( x ) z 0 , z E , z 0 ( 2 f negative definite at x ) (3) Then x is a local maximum point of f. If the inequality is reversed in (3), then x is a local minimum of f. Together, (2) and (3) are sufficient conditions. Proof: Since f is twice differentiable at x , for each x E n 1 2 f ( x) f ( x ) f ( x )T ( x x ) ( x x )T 2 f ( x )( x x ) x x ( x ; x) () 2 where ( x ; x) 0 as x x Suppose x is not a local maximum point 1.2 Unconstrained Optimization For 1.2-2 1 1 0 , k is a positive integer, there exists a vector xk such that x xk and k k f ( xk ) f ( x ) Consider the sequence {xk } , k 1, 2,..., ; since f ( x ) 0 and f ( xk ) f ( x ) , and x x denoting d k k Then () implies the following: xk x () f ( xk ) f ( x ) xk x 2 (x x ) 1 ( xk x )T f ( x ) k ( x ; xk ) 2 xk x xk x Since f ( xk ) f ( x ) , then this must be positive. f ( xk ) f ( x ) 1 T d k f ( x )d k ( x ; xk ) 0 () 2 2 xk x But d k 1 for every k and so there exists a convergent subsequence of {d k } , say {d k } , that converges to d where d 1 . Since all d k ’s are in the unit ball in E n which is compact; every sequence in a compact set S has a convergent subsequence which converges to an element in S. Since ( x ; x) 0 as k , implies the following: lim k k 1 T 1 d k f ( x )d k ( x ; xk ) d T f ( x )d 0 # 2 2 This contradicts (3) so x must be a local max. QED. Note: “ 0 ” in (3) not adequate. Example 1-4: f ( x) x3 f ( x) 3x 2 , 2 f ( x) 6 x At x 0 , f ( x) 0 and zT f ( x ) z 0 but x is not a local maximum. Example 1-5: max f ( x) x 4 f ( x) 4 x3 0 x 0 is a candidate point. 2 f ( x) 12 x 2 2 f (0) 0 So zT 2 f (0) z 0 and the sufficient conditions from Theorem 1.2 are not satisfied (too strong in this case) even though 0 is a maximum point. Example 1-6: min f ( x1 , x2 ) (2 x1 x2 ) 2 ( x2 1) 2 1.2 Unconstrained Optimization 1.2-3 8 x1 4 x2 2(2 x1 x2 )2 f ( x) 2(2 x1 x2 )(1) 2(x 2 1) 4 x1 4 x2 2 8 4 2 f ( x) 4 4 1 x1* 8 x1 4 x2 Critical Points f ( x) 2 0 * 4 x1 4 x2 2 x2 1 1 f ,1 0 2 8 4 z1 2 2 z2 z 8 z1 4 z2 8 z1 z2 0 z 0 4 4 2 * Therefore, x is a local (and global) minimum of f. z1 Example 1-7: min f ( x1 , x2 ) ( x1 x2 ) 2 2( x1 x2 ) f ( x) 0 x1 x2 2( x1 x2 ) 2 2 1 so z T 2 f ( x) z 2 z12 4 z1 z2 2 z22 0 ; this equals 0 at z 2 f ( x) 2 2 1 So sufficient conditions are not satisfied even though x1 x2 provides minimum. Special Cases: 1) One Variable: Given f : E1 E1 , x is an unconstrained local maximum (minimum) of f ( x ) 2 f ( x ) 0 and 0 (> 0). x x 2 Suppose f ( x ) 0 i. If the first nonzero derivative of f is of odd degree, then x is neither a local maximum or minimum. f ( x ) if ii. If the first nonzero derivative of f is positive and of even degree, then x is a local minimum. iii. If the first nonzero derivative of f is negative and of even degree, then x is a local maximum. 2) Two Variables: Given f : E 2 E1 , x is an unconstrained local maximum (minimum) of f ( x ) f ( x ) 2 f ( x ) 2 f ( x ) 2 f ( x ) 2 f ( x ) 0, f ( x ) if 0 (> 0), and 0 x1 x2 x12 x12 x22 x1x2 Theorem: The symmetric matrix S ( si j ) nn is positive definite if and only if s11 0 and all leading minors have determinant greater than zero. e.g. s11 s21 s12 0, s22 , S 0 1.2 Unconstrained Optimization 1.2-4 Theorem: The symmetric matrix S ( si j ) nn is negative definite if and only if s11 0 and s11 s12 s21 s22 s11 s12 s13 0, s21 s31 s22 s32 s23 0, s33 , (1)n S 0 . Theorem 1.3 2nd Order Conditions If f is twice continuously differentiable and if x is a local maximum of f, then (4) f ( x ) 0 T 2 And z f ( x ) z 0 z (5) If f ( x ) 0 (6) And z f ( x) z 0 (7) T 2 For every x in some open ball BS ( x ) around x and z , then x is a local maximum. If the inequalities in (5) and (7) are reversed, then the theorem applies to a local minimum. Conditions (4) and (5) constitute necessary conditions while (6) and (7) describe sufficient conditions. Theorem 1.4 Let f be twice continuously differentiable. If f ( x ) 0 (8) And z f ( x) z 0, z, z 0, x x in open ball around x (9) Then f has a strict local maximum at x . Reversing the inequality in (9) implies x is a strict local minimum of f. T 2 Example 1-8: f ( x) x 2 p , p is some positive integer f ( x) 2 px(2 p 1) 0 at x 0 , satisfying necessary conditions in Corollary 1.1. 2 f ( x) 2 p(2 p 1) x 2( p 1) For p 1 : 2 f ( x) 2 x0 2 0 , at x 0 , sufficient conditions met from Theorem 1.2 for local minimum. For p 1 : 2 f ( x) 0 so sufficient conditions in Theorem 1.2 for a local minimum at x 0 are not satisfied, yet f has a minimum at x 0 . Using Theorem 1.3, at x 0 , all conditions for a local minimum (both necessary and sufficient) from Theorem 1.3 are satisfied. Necessary Conditions: f ( x ) 2 p( x ) 2 p 1 0 at x 0 satisfies (4) 2 f ( x ) 2 p(2 p 1)( x )2( p 1) 0 at x 0 so zT 2 f (0) z 0 z satisfies (5) Sufficient Conditions: f ( x ) 0 at x 0 satisfies (6) 1.2 Unconstrained Optimization 1.2-5 2 f ( x) 2 p(2 p 1)( x) 2( p 1) so zT 2 f ( x) z 0 in any neighborhood of x 0 which satisfies (7) In fact, by Theorem 1.4, x 0 is a strict local minimum since zT 2 f ( x) z 0 x 0 .