Nonlinear Programming Algorithms Handout

Transcription

1 Nonlinear Programming Algorithms Handout Michael C. Ferris Computer Sciences Department University of Wisconsin Madison, Wisconsin 5376 September 9 1 Eigenvalues The eigenvalues of a matrix A C n n are the roots of the characteristic polynomial φ(λ) := det(λi A). The set of all eigenvalues is called the spectrum and is denoted by λ(a). Eigenvectors are vectors x satisfying Ax = λx for some λ λ(a). If A R n n is symmetric then n A = QΛQ T = λ i q i qi T where Q R n n is orthogonal (basis of eigenvectors) and Λ R n n is diagonal (each entry called an eigenvalue). Note that in this case the eigenvalues and eigenvectors may be taken as real. The above factorization is called a spectral decomposition. For real symmetric matrices, the eigenvalues are assumed to be ordered using λ 1 λ... λ n. Singular Values Orthogonal matrices satisfy U T U = UU T = I. If A R m n then there exist orthogonal matrices U = [u 1,..., u m ] R m m and V = [v 1,..., v n ] R n n such that U T AV = diag(σ 1,..., σ p ) R m n where p = min{m, n} and σ 1 σ... σ p. This is often written as A = USV T and is termed the singular value decomposition. Singular values are non-negative; σ i (A) denote the ith largest singular value of A. If we define r by σ 1... σ r > σ r+1 =... = σ p = 1

2 then rank(a) = r, ker(a) = span{v r+1,..., v n }, im(a) = span{u 1,..., u r } and r A = σ i u i vi T. How are eigenvalues and singular values related? σ(a) = + λ(aa T ) Definition 1 A matrix A is positive definite if x T Ax > for all x. A matrix A is positive semidefinite if x T Ax for all x. When A is a symmetric positive definite matrix, then singular values and eigenvalues coincide (take U = V = Q and S = Λ), and thus σ 1 (A) = largest eigenvalue of A If A is square, then σ n (A) = smallest eigenvalue of A σ n (A) min i λ i max λ i σ 1 (A) i Lemma For a symmetric matrix A R n n, the following are equivalent 1. A is positive definite. The (n) leading principal subdeterminants of A are strictly positive 3. All principal subdeterminants of A are strictly positive 4. The eigenvalues of A are strictly positive The matrix [ shows symmetry to be necessary for this result. ] 3 Matrix Norms Suppose A R m n. A α,β := sup x Ax β x α In particular, when α = β = p, this is called the p-norm of A. The Frobenius norm is defined by A F = m nj=1 A ij. Various properties of matrix norms detail in GVL (pp 54). Interesting facts: A = σ 1

3 p A F = σi, p = min(m, n) max i,j A A F n A m A 1 = max A ij 1 j n n A = max 1 i m j=1 A ij A ij A mn max A ij i,j A A 1 A min x Ax x = σ n Lemma 3 Let A be a real symmetric matrix. For each x R n λ n x x T Ax λ 1 x If A is also positive definite, then A = λ 1. Proof Since A = QΛQ T (Q invertible), we let x = Qy for some y. Then Also Thus, n x T Ax = y T Q T AQy = y T Λy = λ i yi. y = y T y = x T QQ T x = x. λ n x n n = λ n yi λ i yi (= x T n Ax) λ 1 yi = λ 1 x Finally, Ax = QΛQ T Qy = QΛy, so n Ax = y T ΛQ T QΛy = λ i yi λ 1 y = λ 1 x, the inequality following from the positive definite assumption. Hence Ax x λ 1 so A = sup x Ax x λ 1. But the supremum is attained at an eigenvector of A corresponding to λ 1. Definition 4 Suppose A R n n is nonsingular. The condition number of A, κ(a), is defined to be A A 1. 3

4 The condition number of a matrix depends on the underlying norm. If A is singular, then κ(a) =. Note that κ (A) = A A 1 = σ 1(A) σ n A. Corollary 5 If A is real symmetric positive definite, then A 1 = λ 1 n and κ(a) = λ 1 λ n. λ 1 1 x x T A 1 x λ 1 n x Proof The eigenvalues of A 1 are simply λ 1 i. Definition 6 Suppose A C n n. The spectral radius of A, ρ(a), is defined as the maximum of λ 1,..., λ n, where λ 1,..., λ n are the eigenvalues of A. Note that if A C n n and λ is any eigenvalue of A with eigenvector u, then Au = λ u, so that λ A. Hence ρ(a) A. Lemma 7 Let A C n n. Then lim k A k = if and only if ρ(a) < 1. Lemma 8 (Neumann) Let E R n n and ρ(e) < 1. Then I E is nonsingular and I E 1 = lim k E i k i= Proof Since ρ(e) < 1, the maximum eigenvalue of E must be less than 1 in modulus and hence I E is invertible. Furthermore so that (I E)(I + + E k 1 ) = I E k I + E + + E k 1 = (I E) 1 (I E) 1 E k The result now follows in the limit as k from the above lemma. 4 Results from Matrix Theory Lemma 9 (Debreu) Suppose H R n n and A R m n. The following are equivalent: (a) Az = and z implies z, Hz > (b) There exists γ such that H + γa T A is positive definite 4

5 Remark If (b) holds and γ γ then for any z, z, (H + γa T A)z = z, (H + γa T A)z + (γ γ) Az z, (H + γa T A)z so that H + γa T A is also positive definite. Proof (b) (a) If Az = and z, then < z, Hz = z, (H + γa T A)z. (a) (b) If (b) were false, then there would exist z k of norm 1, with z k, (H + ka T A)z k Without loss of generality, we can assume that z k z. Now for each k z k, (k 1 H + A T A)z k Az so Az =. But z k, Hz k + k Az k z k, Hz k z, Hz and this contradicts (a), since z = 1. Lemma 1 (Schur) Suppose that ( A B M = C D ) (1) with D nonsingular. Then M is nonsingular if and only if S = A BD 1 C is nonsingular, and in that case ( M 1 S = 1 S 1 BD 1 ) D 1 CS 1 D 1 + D 1 CS 1 BD 1 () Remark S is the Schur complement of D in M. We can also show that det M = det S det D since ( ) (( ) ( ) ( )) A B A B I I det M = det = det C D C D C D 1 DC D ( A BD = det 1 C BD 1 ) ( ) I det I DC D = det A BD 1 C det D 5

6 Proof If S is nonsingular, then just multiply the expressions given in (1) and () to obtain I. Thus assume M is nonsingular. Write its inverse as ( ) E F M 1 = G H Since MM 1 = ( I I ) we have I = AE + BG (3) = AF + BH (4) = CE + DG (5) I = CF + DH (6) From (5) we have G = D 1 CE, and substituting this in (3) yields (A BD 1 C)E = I. This implies that S = A BD 1 C is nonsingular and that E = S 1. Using (5) again we have that G = D 1 CS 1. From (6) we have H = D 1 (I CF ), and putting this in (4) yields AF + BD 1 BD 1 CF = that is which implies F = S 1 BD 1. 5 Matrix Norms SF + BD 1 = Suppose A R m n. A α,β := sup x Ax β x α In particular, when α = β = p, this is called the p-norm of A. The Frobenius norm is defined by A F = m nj=1 A ij. Various properties of matrix norms detaile in GVL (pp 54). Interesting facts: The two norm of A is the square root of the largest eigenvalue of A T A. max i,j A A F n A m A 1 = max A ij 1 j n n A = max A ij 1 i m j=1 A ij A mn max A ij A i,j A 1 A The condition number of a matrix depends on the underlying norm, but is defined for a square matrix by κ(a) := A A 1 6

7 If A is singular, then κ(a) =. Note that κ (A) = A inva = σ 1(A) σ n A where σ i (A) is the ith largest singular value of A. Singular values are non-negative. How are they related to eigenvalues? p A F = σi, p = min(m, n) A = σ 1 min x Ax x = σ n 6 Positive (Semi)-Definite Matrices Definition 11 A matrix A is positive definite if x T Ax > for all x. A matrix A is positive semidefinite if x T Ax for all x. Theorem 1 See Bertsekas 7 Strong Convexity Let f: Ω R, h: Ω R where Ω is an open convex set. Definition 13 f is strongly convex (ρ) on Ω if ρ > such that x, y Ω, λ [, 1] f((1 λ)x + λy) (1 λ)f(x) + λf(y) ρ λ(1 λ) x y Definition 14 h is strongly monotone (ρ) on Ω if ρ > such that x, y Ω h(x) h(y), x y ρ x y Theorem 15 (Strong convexity) If f is continuously differentiable on Ω then the following are equivalent: (a) f is strongly convex (ρ) on Ω (b) For all x, y Ω, f(y) f(x) + f(x), y x + (ρ/) x y (c) f is strongly monotone (ρ) on Ω If f is twice continuously differentiable on Ω, then (d) For all x,y,z Ω, x y, f(z)(x y) ρ x y 7

8 is equivalent to the above. Proof We show (a) (b) (c). (a) (b) The hypothesis gives f(x + λ(y x)) f(x) λ so taking the limit as λ (b) (c) Applying (b) twice gives f(y) f(x) ρ (1 λ) x y f(x), y x f(y) f(x) ρ x y f(y) f(x) + f(x), y x + ρ x y f(x) f(y) + f(y), x y + ρ x y Adding these inequalities gives f(y) + f(x) f(x) + f(y) + f(x) f(y), y x + ρ x y from where the result follows. (c) (b) The hypothesis gives f(x + t(y x)) f(x), y x dt which implies the result. (b) (a) Letting y = u and x = (1 λ)u + λv in (b) gives ρt x y dt f(u) f((1 λ)u + λv) + f((1 λ)u + λv), λ(u v) + ρ λ(u v) (7) Also letting y = v and x = (1 λ)u + λv in (b) implies f(v) f((1 λ)u + λv) + f((1 λ)u + λv), (1 λ)(v u) + ρ (1 λ)(v u) (8) Adding (1 λ) times (7) to λ times (8) gives the required result. To complete the proof, we assume that f is twice continuously differentiable on Ω. (d) (c) This follows from the hypothesis since f(x) f(y), x y = f(y + t(x y))(x y)dt, x y ρ x y (c) (d) Let x, y, z Ω. Then z + λ(x y) Ω for sufficiently small λ, so x y, f(z + λ(x y))(x y) = The result follows in the limit as λ. x y, f(z + λ(x y)) f(z) + o(1) λ ρ x y + o(1) 8

9 8 Lipschitz Continuity Definition 16 f is Lipschitz continuous (ρ) on Ω if ρ > such that x, y Ω f(y) f(x) ρ y x Lemma 17 (Lipschitz continuity) Consider the following statements: (a) f is Lipschitz continuous on Ω with constant ρ (b) For all x, y Ω, f(y) f(x) + f(x), y x + (ρ/) x y (c) For all x, y Ω, f(x) f(y), x y ρ x y Then (a) implies (b) implies (c). If f is twice continuously differentiable, then (c) implies (d) For all x, y Ω, y x, f(z)(y x) ρ y x If x y, f(x)(y x) γ y x for some γ (not necessarily positive), then (d) implies (a), possibly with a different constant ρ. Proof (a) (b) f(y) f(x) f(x), y x = (b) (c) Invoking (b) twice gives f(x + t(y x)) f(x), y x dt y x ρ y x tdt = ρ y x f(y) f(x) + f(x), y x + ρ x y f(x) f(y) + f(y), x y + ρ x y f(x + t(y x)) f(x) dt Adding these inequalities gives f(y) + f(x) f(x) + f(y) + f(x) f(y), y x + ρ x y from where (c) follows. (c) (d) It follows from (c) that f(x + λ(y x)) f(x), λ(y x) ρλ y x If we divide both sides by λ, (d) then follows in the limit as λ. 9

10 (d) (a) Let δ := max{ γ, } + ρ. We first show that f(z) δ. Note that However, f(z) = sup x, f(z)y = 1 y =1 f(z)y = sup x, f(z)y x =1, y =1 x y, f(z)(y x) + x, f(z)x + y, f(z)y 1 { γ x y + ρ x + y } 1 {max{ γ, }( x + x y + y ) + ρ( x + y )} Hence, f(z) max{ γ, } + ρ, as required. The Lipschitz continuity now follows easily since f(y) f(x) = f(x + t(y x))(y x)dt δ y x 1