AM 205: lecture 18. Last time: optimization methods Today: conditions for optimality

Transcription

1 AM 205: lecture 18 Last time: optimization methods Today: conditions for optimality

2 Existence of Global Minimum For example: f (x, y) = x 2 + y 2 is coercive on R 2 (global min. at (0, 0)) f (x) = x 3 is not coercive on R (f for x ) f (x) = e x is not coercive on R (f 0 for x )

3 Convexity An important concept for uniqueness is convexity A set S R n is convex if it contains the line segment between any two of its points That is, S is convex if for any x, y S, we have {θx + (1 θ)y : θ [0, 1]} S

4 Convexity Similarly, we define convexity of a function f : S R n R f is convex if its graph along any line segment in S is on or below the chord connecting the function values i.e. f is convex if for any x, y S and any θ (0, 1), we have Also, if then f is strictly convex f (θx + (1 θ)y) θf (x) + (1 θ)f (y) f (θx + (1 θ)y) < θf (x) + (1 θ)f (y)

5 Convexity Strictly convex

6 Convexity Non-convex

7 Convexity Convex (not strictly convex)

8 Convexity If f is a convex function on a convex set S, then any local minimum of f must be a global minimum 1 Proof: Suppose x is a local minimum, i.e. f (x) f (y) for y B(x, ɛ) (where B(x, ɛ) {y S : y x ɛ}) Suppose that x is not a global minimum, i.e. that there exists w S such that f (w) < f (x) (Then we will show that this gives a contradiction) 1 A global minimum is defined as a point z such that f (z) f (x) for all x S. Note that a global minimum may not be unique, e.g. if f (x) = cos x then 0 and 2π are both global minima.

9 Convexity Proof (continued...): For θ [0, 1] we have f (θw + (1 θ)x) θf (w) + (1 θ)f (x) Let σ (0, 1] be sufficiently small so that z σw + (1 σ) x B(x, ɛ) Then f (z) σf (w) + (1 σ) f (x) < σf (x) + (1 σ) f (x) = f (x), i.e. f (z) < f (x), which contradicts that f (x) is a local minimum! Hence we cannot have w S such that f (w) < f (x)

10 Convexity Note that convexity does not guarantee uniqueness of global minimum e.g. a convex function can clearly have a horizontal section (see earlier plot) If f is a strictly convex function on a convex set S, then a local minimum of f is the unique global minimum Optimization of convex functions over convex sets is called convex optimization, which is an important subfield of optimization

11 Optimality Conditions We have discussed existence and uniqueness of minima, but haven t considered how to find a minimum The familiar optimization idea from calculus in one dimension is: set derivative to zero, check the sign of the second derivative This can be generalized to R n

12 Optimality Conditions If f : R n R is differentiable, then the gradient vector f : R n R n is f (x) f (x) x 1 f (x) x 2. f (x) x n The importance of the gradient is that f points uphill, i.e. towards points with larger values than f (x) And similarly f points downhill

13 Optimality Conditions This follows from Taylor s theorem for f : R n R Recall that f (x + δ) = f (x) + f (x) T δ + H.O.T. Let δ ɛ f (x) for ɛ > 0 and suppose that f (x) 0, then: f (x ɛ f (x)) f (x) ɛ f (x) T f (x) < f (x) Also, we see from Cauchy Schwarz that f (x) is the steepest descent direction

14 Optimality Conditions Similarly, we see that a necessary condition for a local minimum at x S is that f (x ) = 0 In this case there is no downhill direction at x The condition f (x ) = 0 is called a first-order necessary condition for optimality, since it only involves first derivatives

15 Optimality Conditions x S that satisfies the first-order optimality condition is called a critical point of f But of course a critical point can be a local min., local max., or saddle point (Recall that a saddle point is where some directions are downhill and others are uphill, e.g. (x, y) = (0, 0) for f (x, y) = x 2 y 2 )

16 Optimality Conditions As in the one-dimensional case, we can look to second derivatives to classify critical points If f : R n R is twice differentiable, then the Hessian is the matrix-valued function H f : R n R n n H f (x) 2 f (x) x1 2 2 f (x) x 2 x 1. 2 f (x) x nx 1 2 f (x) x 1 x 2 2 f (x) x f (x) x nx 2 2 f (x) x 1 x n 2 f (x) x 2 x n f (x) x 2 n The Hessian is the Jacobian matrix of the gradient f : R n R n If the second partial derivatives of f are continuous, then 2 f / x i x j = 2 f / x j x i, and H f is symmetric

17 Optimality Conditions Suppose we have found a critical point x, so that f (x ) = 0 From Taylor s Theorem, for δ R n, we have f (x + δ) = f (x ) + f (x ) T δ δt H f (x + ηδ)δ = f (x ) δt H f (x + ηδ)δ for some η (0, 1)

18 Optimality Conditions Recall positive definiteness: A is positive definite if x T Ax > 0 Suppose H f (x ) is positive definite Then (by continuity) H f (x + ηδ) is also positive definite for δ sufficiently small, so that: δ T H f (x + ηδ)δ > 0 Hence, we have f (x + δ) > f (x ) for δ sufficiently small, i.e. f (x ) is a local minimum Hence, in general, positive definiteness of H f at a critical point x is a second-order sufficient condition for a local minimum

19 Optimality Conditions A matrix can also be negative definite: x T Ax < 0 for all x 0 Or indefinite: There exists x, y such that x T Ax < 0 < y T Ay Then we can classify critical points as follows: H f (x ) positive definite = x is a local minimum H f (x ) negative definite = x is a local maximum H f (x ) indefinite = x is a saddle point

20 Optimality Conditions Also, positive definiteness of the Hessian is closely related to convexity of f If H f (x) is positive definite, then f is convex on some convex neighborhood of x If H f (x) is positive definite for all x S, where S is a convex set, then f is convex on S Question: How do we test for positive definiteness?

21 Optimality Conditions Answer: A is positive (resp. negative) definite if and only if all eigenvalues of A are positive (resp. negative) 2 Also, a matrix with positive and negative eigenvalues is indefinite Hence we can compute all the eigenvalues of A and check their signs 2 This is related to the Rayleigh quotient, see Unit V

22 Heath Example 6.5 Consider f (x) = 2x x x 1 x 2 + 3x 2 2 6x Then f (x) = [ 6x x x 2 12x 1 + 6x 2 6 ] We set f (x) = 0 to find critical points 3 [1, 1] T and [2, 3] T 3 In general solving f (x) = 0 requires an iterative method

23 Heath Example 6.5, continued... The Hessian is and hence H f (1, 1) = H f (2, 3) = H f (x) = [ [ [ 12x ] ], which has eigenvalues 25.4, 1.4 ], which has eigenvalues 35.0, 1.0 Hence [2, 3] T is a local min. whereas [1, 1] T is a saddle point

24 Optimality Conditions: Equality Constrained Case So far we have ignored constraints Let us now consider equality constrained optimization min f (x) subject to g(x) = 0, x Rn where f : R n R and g : R n R m, with m n Since g maps to R m, we have m constraints This situation is treated with Lagrange mutlipliers

25 Optimality Conditions: Equality Constrained Case We illustrate the concept of Lagrange multipliers for f, g : R 2 R Let f (x, y) = x + y and g(x, y) = 2x 2 + y g is normal to S: 4 at any x S we must move in direction ( g(x)) (tangent direction) to remain in S 4 This follows from Taylor s Theorem: g(x + δ) g(x) + g(x) T δ

26 Optimality Conditions: Equality Constrained Case Also, change in f due to infinitesimal step in direction ( g(x)) is f (x ± ɛ( g(x)) ) = f (x) ± ɛ f (x) T ( g(x)) + H.O.T. Hence stationary point x S if f (x ) T ( g(x )) = 0, or f (x ) = λ g(x ), for some λ R

27 Optimality Conditions: Equality Constrained Case This shows that for a stationary point with m = 1 constraints, f cannot have any component in the tangent direction to S Now, consider the case with m > 1 equality constraints Then g : R n R m and we now have a set of constraint gradient vectors, g i, i = 1,..., m Then we have S = {x R n : g i (x) = 0, i = 1,..., m} Any tangent direction at x S must be orthogonal to all gradient vectors { g i (x), i = 1,..., m} to remain in S

28 Optimality Conditions: Equality Constrained Case Let T (x) {v R n : g i (x) T v = 0, i = 1, 2,..., m} denote the orthogonal complement of { g i (x), i = 1,..., m} Then, for δ T (x) and ɛ R >0, ɛδ is a step in a tangent direction of S at x Since we have f (x + ɛδ) = f (x ) + ɛ f (x ) T δ + H.O.T. it follows that for a stationary point we need f (x ) T δ = 0 for all δ T (x )

29 Optimality Conditions: Equality Constrained Case Hence, we require that at a stationary point x S we have f (x ) span{ g i (x ), i = 1,..., m} This can be written succinctly as a linear system f (x ) = (J g (x )) T λ for some λ R m, where (J g (x )) T R n m This follows because the columns of (J g (x )) T are the vectors { g i (x ), i = 1,..., m}

30 Optimality Conditions: Equality Constrained Case We can write equality constrained optimization problems more succinctly by introducing the Lagrangian function, L : R n+m R, Then we have, L(x, λ) f (x) + λ T g(x) = f (x) + λ 1 g 1 (x) + + λ m g m (x) L(x,λ) x i = f (x) x i + λ 1 g 1 (x) x i + + λ n g n(x) x i, i = 1,..., n L(x,λ) λ i = g i (x), i = 1,..., m

31 Optimality Conditions: Equality Constrained Case Hence L(x, λ) = [ x L(x, λ) λ L(x, λ) ] [ f (x) + Jg (x) = T λ g(x) ], so that the first order necessary condition for optimality for the constrained problem can be written as a nonlinear system: 5 [ f (x) + Jg (x) L(x, λ) = T ] λ = 0 g(x) (As before, stationary points can be classified by considering the Hessian, though we will not consider this here...) 5 n + m variables, n + m equations

32 Optimality Conditions: Equality Constrained Case See Lecture: Constrained optimization of cylinder surface area

33 Optimality Conditions: Equality Constrained Case As another example of equality constrained optimization, recall our underdetermined linear least squares problem from I.3 min f (b) subject to g(b) = 0, b Rn where f (b) b T b, g(b) Ab y and A R m n with m < n

34 Optimality Conditions: Equality Constrained Case Introducing Lagrange multipliers gives where b R n and λ R m L(b, λ) b T b + λ T (Ab y) Hence L(b, λ) = 0 implies [ f (b) + Jg (b) T ] λ = g(b) [ 2b + A T λ Ab y ] = 0 R n+m

35 Optimality Conditions: Equality Constrained Case Hence, we obtain the (n + m) (n + m) square linear system [ ] [ ] [ ] 2I A T b 0 = A 0 λ y which we can solve for [ b λ ] R n+m

36 Optimality Conditions: Equality Constrained Case We have b = 1 2 AT λ from the first block row Subsituting into Ab = y (the second block row ) yields λ = 2(AA T ) 1 y And hence b = 1 2 AT λ = A T (AA T ) 1 y which was the solution we introduced (but didn t derive) in I.3

37 Optimality Conditions: Inequality Constrained Case Similar Lagrange multiplier methods can be developed for the more difficult case of inequality constrained optimization

38 Steepest Descent We first consider the simpler case of unconstrained optimization (as opposed to constrained optimization) Perhaps the simplest method for unconstrained optimization is steepest descent Key idea: The negative gradient f (x) points in the steepest downhill direction for f at x Hence an iterative method for minimizing f is obtained by following f (x k ) at each step Question: How far should we go in the direction of f (x k )?

39 Steepest Descent We can try to find the best step size via a subsidiary (and easier!) optimization problem For a direction s R n, let φ : R R be given by φ(η) = f (x + ηs) Then minimizing f along s corresponds to minimizing the one-dimensional function φ This process of minimizing f along a line is called a line search 6 6 The line search can itself be performed via Newton s method, as described for f : R n R shortly, or via a built-in function

40 Steepest Descent Putting these pieces together leads to the steepest descent method: 1: choose initial guess x 0 2: for k = 0, 1, 2,... do 3: s k = f (x k ) 4: choose η k to minimize f (x k + η k s k ) 5: x k+1 = x k + η k s k 6: end for However, steepest descent often converges very slowly Convergence rate is linear, and scaling factor can be arbitrarily close to 1 (Steepest descent will be covered on Assignment 5)

41 Newton s Method We can get faster convergence by using more information about f Note that f (x ) = 0 is a system of nonlinear equations, hence we can solve it with quadratic convergence via Newton s method 7 The Jacobian matrix of f (x) is H f (x) and hence Newton s method for unconstrained optimization is: 1: choose initial guess x 0 2: for k = 0, 1, 2,... do 3: solve H f (x k )s k = f (x k ) 4: x k+1 = x k + s k 5: end for 7 Note that in its simplest form this algorithm searches for stationary points, not necessarily minima

42 Newton s Method We can also interpret Newton s method as seeking stationary point based on a sequence of local quadratic approximations Recall that for small δ f (x + δ) f (x) + f (x) T δ δt H f (x)δ q(δ) where q(δ) is quadratic in δ (for a fixed x) We find stationary point of q in the usual way: 8 q(δ) = f (x) + H f (x)δ = 0 This leads to H f (x)δ = f (x), as in the previous slide 8 Recall I.4 for differentiation of δ T H f (x)δ