Advanced Optimization

Transcription

1 Advanced Optimization (10-801: CMU) Lecture 22 Fixed-point theory; nonlinear conic optimization 07 Apr 2014 Suvrit Sra

2 Fixed-point theory Many optimization problems h(x) = 0 x h(x) = 0 (I h)(x) = x 2 / 19

3 Fixed-point theory Many optimization problems h(x) 0 x h(x) 0 (I h)(x) x 2 / 19

4 Fixed-point theory Fixed-point Any x that solves x = f(x) 3 / 19

5 Fixed-point theory Fixed-point Any x that solves x = f(x) Three types of results 1 Geometric: Banach contraction and relatives 3 / 19

6 Fixed-point theory Fixed-point Any x that solves x = f(x) Three types of results 1 Geometric: Banach contraction and relatives 2 Order-theoretic: Knaster-Tarski 3 / 19

7 Fixed-point theory Fixed-point Any x that solves x = f(x) Three types of results 1 Geometric: Banach contraction and relatives 2 Order-theoretic: Knaster-Tarski 3 Topological: Brouwer, Schauder-Leray, etc. 3 / 19

8 Fixed-point theory main concerns existence of a solution uniqueness of a solution stability under small perturbations of parameters structure of solution set (failing uniqueness) algorithms / approximation methods to obtain solutions rate of convergence analyses 4 / 19

9 Fixed-point theory Banach contraction Some conditions under which the nonlinear equation x = T x, x M X, can be solved by iterating x k+1 = T x k, x 0 M, k = 0, 1, / 19

10 Fixed-point theory Banach contraction Theorem (Banach 1922.) Suppose (i) T : M X M; (ii) M is closed, nonempty set in a complete metric space (X, d); (iii) T is q-contractive, i.e., d(t x, T y) qd(x, y), x, y M, constant 0 q < 1. Then, we have the following: (i) T x = x has exactly one solution (T has a unique FP in M) (ii) The sequence {x k } converges to the solution for any x 0 M (iii) A priori error estimate (iv) A posterior error estimate d(x k, x ) q k (1 q) 1 d(x 0, x 1 ) d(x k+1, x ) q(1 q) 1 d(x k, x k+1 ) (v) (Global) linear rate of convergence: d(x k+1, x ) qd(x k, x ) 5 / 19

11 Fixed-point theory Banach contraction If X is a Banach space with distance d = x y T x T y q x y, 0 q < 1 (contraction) 6 / 19

12 Fixed-point theory Banach contraction If X is a Banach space with distance d = x y T x T y q x y, 0 q < 1 (contraction) If inequality holds with q = 1, we call map nonexpansive d(t x, T y) d(x, y) Example: x x + 1 is nonexpansive, but has no fixed points! 6 / 19

13 Fixed-point theory Banach contraction If X is a Banach space with distance d = x y T x T y q x y, 0 q < 1 (contraction) If inequality holds with q = 1, we call map nonexpansive d(t x, T y) d(x, y) Example: x x + 1 is nonexpansive, but has no fixed points! Map is called contractive or weakly-contractive if d(t x, T y) < d(x, y), x, y M. 6 / 19

14 Fixed-point theory Banach contraction If X is a Banach space with distance d = x y T x T y q x y, 0 q < 1 (contraction) If inequality holds with q = 1, we call map nonexpansive d(t x, T y) d(x, y) Example: x x + 1 is nonexpansive, but has no fixed points! Map is called contractive or weakly-contractive if d(t x, T y) < d(x, y), x, y M. Several other variations of maps have been studied for Banach spaces (see e.g., book by Bauschke, Combettes (2012)) 6 / 19

15 Banach contraction proof Blackboard 7 / 19

16 Banach contraction proof Summary: Blackboard d must be positive-definite, i.e, d(x, y) = 0 iff x = y (X, d) must be complete (contain all its Cauchy sequences) T : M M, M must be closed But M need not be compact! Contraction is often a rare luxury; nonexpansive maps are more common (we ve already seen several) 7 / 19

17 More general fixed-point theorems Theorem (Brouwer FP.) Every continuous function from a convex compact subset M R d to M itself has a fixed-point. 8 / 19

18 More general fixed-point theorems Theorem (Brouwer FP.) Every continuous function from a convex compact subset M R d to M itself has a fixed-point. Note: Proves existence only! 8 / 19

19 More general fixed-point theorems Theorem (Brouwer FP.) Every continuous function from a convex compact subset M R d to M itself has a fixed-point. Note: Proves existence only! Generalizes the intermediate-value theorem. 8 / 19

20 More general fixed-point theorems Theorem (Brouwer FP.) Every continuous function from a convex compact subset M R d to M itself has a fixed-point. Note: Proves existence only! Generalizes the intermediate-value theorem. Theorem (Schauder FP.) Every continuous function from a convex compact subset M of a Banach space to M itself has a fixed-point. 8 / 19

21 More general fixed-point theorems Theorem (Brouwer FP.) Every continuous function from a convex compact subset M R d to M itself has a fixed-point. Note: Proves existence only! Generalizes the intermediate-value theorem. Theorem (Schauder FP.) Every continuous function from a convex compact subset M of a Banach space to M itself has a fixed-point. Remarks: Brouwer FPs very hard. Exponential lower bounds for finding Brouwer fixed points Hirsch, Papadimitriou, Vavasis (1988). 8 / 19

22 More general fixed-point theorems Theorem (Brouwer FP.) Every continuous function from a convex compact subset M R d to M itself has a fixed-point. Note: Proves existence only! Generalizes the intermediate-value theorem. Theorem (Schauder FP.) Every continuous function from a convex compact subset M of a Banach space to M itself has a fixed-point. Remarks: Brouwer FPs very hard. Exponential lower bounds for finding Brouwer fixed points Hirsch, Papadimitriou, Vavasis (1988). Any algorithm for computing a Brouwer FP based on function evaluations only must in the worst case perform a number of function evaluations exponential in both the number of digits of accuracy and the dimension. 8 / 19

23 More general fixed-point theorems Theorem (Brouwer FP.) Every continuous function from a convex compact subset M R d to M itself has a fixed-point. Note: Proves existence only! Generalizes the intermediate-value theorem. Theorem (Schauder FP.) Every continuous function from a convex compact subset M of a Banach space to M itself has a fixed-point. Remarks: Brouwer FPs very hard. Exponential lower bounds for finding Brouwer fixed points Hirsch, Papadimitriou, Vavasis (1988). Any algorithm for computing a Brouwer FP based on function evaluations only must in the worst case perform a number of function evaluations exponential in both the number of digits of accuracy and the dimension. Contrast with n = 1, where bisection yields f(ˆx) f 2 δ in O(δ) 8 / 19

24 Kakutani fixed-point theorem FP theorem for set-valued mappings (recall x (I f)(x)) 9 / 19

25 Kakutani fixed-point theorem FP theorem for set-valued mappings (recall x (I f)(x)) Set-valued map F : M 2 M, x M F (x) 2 M, i.e. F (x) M. Closed-graph {(x, y) y F (x)} is a closed subset ofx X (i.e., x k x, y k y and y k F (x k ) = y F (x)) 9 / 19

26 Kakutani fixed-point theorem FP theorem for set-valued mappings (recall x (I f)(x)) Set-valued map F : M 2 M, x M F (x) 2 M, i.e. F (x) M. Closed-graph {(x, y) y F (x)} is a closed subset ofx X (i.e., x k x, y k y and y k F (x k ) = y F (x)) Theorem (S. Kakutani 1941.) Let M R n be nonempty, convex, compact. Let F : M 2 M be a set-valued map with a closed graph; also for all x M, let F (x) be non-empty and convex. Then, F has a fixed point. Application: See proof of Nash equilibrium on Wikipedia 9 / 19

27 Brouwer FP example Consider a Markov transition matrix A R n n + Column stochastic: a ij 0 and i a ij = 1 for 1 j n 10 / 19

28 Brouwer FP example Consider a Markov transition matrix A R n n + Column stochastic: a ij 0 and i a ij = 1 for 1 j n Claim. There is a probability vector x that is an eigenvector of A. 10 / 19

29 Brouwer FP example Consider a Markov transition matrix A R n n + Column stochastic: a ij 0 and i a ij = 1 for 1 j n Claim. There is a probability vector x that is an eigenvector of A. Prove: x 0, x T 1 = 1 such that Ax = x. 10 / 19

30 Brouwer FP example Consider a Markov transition matrix A R n n + Column stochastic: a ij 0 and i a ij = 1 for 1 j n Claim. There is a probability vector x that is an eigenvector of A. Prove: x 0, x T 1 = 1 such that Ax = x. Let n be probability simplex (compact, convex subset of R n ) 10 / 19

31 Brouwer FP example Consider a Markov transition matrix A R n n + Column stochastic: a ij 0 and i a ij = 1 for 1 j n Claim. There is a probability vector x that is an eigenvector of A. Prove: x 0, x T 1 = 1 such that Ax = x. Let n be probability simplex (compact, convex subset of R n ) Verify that if x n then Ax n 10 / 19

32 Brouwer FP example Consider a Markov transition matrix A R n n + Column stochastic: a ij 0 and i a ij = 1 for 1 j n Claim. There is a probability vector x that is an eigenvector of A. Prove: x 0, x T 1 = 1 such that Ax = x. Let n be probability simplex (compact, convex subset of R n ) Verify that if x n then Ax n Thus, A : n n ; A is obviously continuous 10 / 19

33 Brouwer FP example Consider a Markov transition matrix A R n n + Column stochastic: a ij 0 and i a ij = 1 for 1 j n Claim. There is a probability vector x that is an eigenvector of A. Prove: x 0, x T 1 = 1 such that Ax = x. Let n be probability simplex (compact, convex subset of R n ) Verify that if x n then Ax n Thus, A : n n ; A is obviously continuous Hence by Brouwer FP: there is an x n such that Ax = x 10 / 19

34 Brouwer FP example Consider a Markov transition matrix A R n n + Column stochastic: a ij 0 and i a ij = 1 for 1 j n Claim. There is a probability vector x that is an eigenvector of A. Prove: x 0, x T 1 = 1 such that Ax = x. Let n be probability simplex (compact, convex subset of R n ) Verify that if x n then Ax n Thus, A : n n ; A is obviously continuous Hence by Brouwer FP: there is an x n such that Ax = x How to compute such an x? 10 / 19

35 Conic optimization 11 / 19

36 Some definitions Let K be a cone in a real vector space V 12 / 19

37 Some definitions Let K be a cone in a real vector space V Let y K and x V. We say y dominates x if αy K x K βy, for some α, β R. 12 / 19

38 Some definitions Let K be a cone in a real vector space V Let y K and x V. We say y dominates x if αy K x K βy, for some α, β R. Max-min gauges M K (x/y) := inf {β R x βy} m K (x/y) := sup {α R αy x}. Shorthand: K 12 / 19

39 Some definitions Let K be a cone in a real vector space V Let y K and x V. We say y dominates x if αy K x K βy, for some α, β R. Max-min gauges M K (x/y) := inf {β R x βy} m K (x/y) := sup {α R αy x}. Shorthand: K Parts: We have an equivalence relation x K y on K if x dominates y and vice versa. 12 / 19

40 Some definitions Let K be a cone in a real vector space V Let y K and x V. We say y dominates x if αy K x K βy, for some α, β R. Max-min gauges M K (x/y) := inf {β R x βy} m K (x/y) := sup {α R αy x}. Shorthand: K Parts: We have an equivalence relation x K y on K if x dominates y and vice versa. The equivalence classes are called parts of the cone. 12 / 19

41 Hilbert projective metric If x K y with y 0, then α, β > 0 s.t. αy x βy. 13 / 19

42 Hilbert projective metric If x K y with y 0, then α, β > 0 s.t. αy x βy. Def. (Hilbert metric.) Let x K y and y 0. Then, d H (x, y) := log M(x/y) m(x/y) 13 / 19

43 Hilbert projective metric If x K y with y 0, then α, β > 0 s.t. αy x βy. Def. (Hilbert metric.) Let x K y and y 0. Then, d H (x, y) := log M(x/y) m(x/y) Proposition. Let K be a cone in V ; (K, d H ) satisfies: d H (x, y) 0, and d H (x, y) = d H (y, x) for all x, y K d H (x, z) d H (x, y) + d H (y, z) for all x K y K z, and d H (αx, βy) = d H (x, y) for all α, β > 0 and x, y K. If K is closed, then d H (x, y) = 0 iff x = λy for some λ > 0. In this case, if X K satisfies that for each x K \ {0} there is a unique λ > 0 such that λx X and P is a part of K, then (P X, d H ) is a genuine metric space. Proof: on blackboard 13 / 19

44 Nonexpansive maps with d H Def. (OPSH maps.) Let K V and K V be closed cones. The f : K K is called order preserving if for x K y, f(x) K f(y). It is homogeneous of degree r if f(λx) = λ r f(x) for all x K and λ > 0. It is subhomogeneous if λf(x) f(λx) for all x K and 0 < λ < / 19

45 Nonexpansive maps with d H Def. (OPSH maps.) Let K V and K V be closed cones. The f : K K is called order preserving if for x K y, f(x) K f(y). It is homogeneous of degree r if f(λx) = λ r f(x) for all x K and λ > 0. It is subhomogeneous if λf(x) f(λx) for all x K and 0 < λ < 1. Exercise: Prove that if f : K K is OPH of degree r > 0 then d H (f(x), f(y)) rd H (x, y). 14 / 19

46 Nonexpansive maps with d H Def. (OPSH maps.) Let K V and K V be closed cones. The f : K K is called order preserving if for x K y, f(x) K f(y). It is homogeneous of degree r if f(λx) = λ r f(x) for all x K and λ > 0. It is subhomogeneous if λf(x) f(λx) for all x K and 0 < λ < 1. Exercise: Prove that if f : K K is OPH of degree r > 0 then d H (f(x), f(y)) rd H (x, y). In particular, if r = 1, then f is nonexpansive (in d H ) 14 / 19

47 Birkhoff s theorem Let L be a linear operator on a cone K (L : K K) 15 / 19

48 Birkhoff s theorem Let L be a linear operator on a cone K (L : K K) Contraction ratio κ(l) := inf {λ 0 d H (Lx, Ly) λd H (x, y) for all x K y in K}. 15 / 19

49 Birkhoff s theorem Let L be a linear operator on a cone K (L : K K) Contraction ratio κ(l) := inf {λ 0 d H (Lx, Ly) λd H (x, y) for all x K y in K}. Theorem (Birkhoff.) Let (L) := sup {d H (Lx, Ly) Lx K Ly} be the projective diameter of L. Then κ(l) = tanh( 1 4 (L)) 15 / 19

50 Birkhoff s theorem Let L be a linear operator on a cone K (L : K K) Contraction ratio κ(l) := inf {λ 0 d H (Lx, Ly) λd H (x, y) for all x K y in K}. Theorem (Birkhoff.) Let (L) := sup {d H (Lx, Ly) Lx K Ly} be the projective diameter of L. Then κ(l) = tanh( 1 4 (L)) If (L) <, then we have a strict contraction! 15 / 19

51 Application to Pagerank eigenvector Markov transition matrix A R n n + Column stochastic: a ij 0 and i a ij = 1 for 1 j n 16 / 19

52 Application to Pagerank eigenvector Markov transition matrix A R n n + Column stochastic: a ij 0 and i a ij = 1 for 1 j n Consider cone K R n + Suppose (A) < (next slide) 16 / 19

53 Application to Pagerank eigenvector Markov transition matrix A R n n + Column stochastic: a ij 0 and i a ij = 1 for 1 j n Consider cone K R n + Suppose (A) < (next slide) Then d H (Ax, Ay) κ(a)d H (x, y) strict contraction 16 / 19

54 Application to Pagerank eigenvector Markov transition matrix A R n n + Column stochastic: a ij 0 and i a ij = 1 for 1 j n Consider cone K R n + Suppose (A) < (next slide) Then d H (Ax, Ay) κ(a)d H (x, y) strict contraction Need to argue that ( n, d H ) is a complete metric space 16 / 19

55 Application to Pagerank eigenvector Markov transition matrix A R n n + Column stochastic: a ij 0 and i a ij = 1 for 1 j n Consider cone K R n + Suppose (A) < (next slide) Then d H (Ax, Ay) κ(a)d H (x, y) strict contraction Need to argue that ( n, d H ) is a complete metric space Invoke Banach contraction theorem. 16 / 19

56 Application to Pagerank eigenvector Markov transition matrix A R n n + Column stochastic: a ij 0 and i a ij = 1 for 1 j n Consider cone K R n + Suppose (A) < (next slide) Then d H (Ax, Ay) κ(a)d H (x, y) strict contraction Need to argue that ( n, d H ) is a complete metric space Invoke Banach contraction theorem. Linear rate of convergence 16 / 19

57 Application to Pagerank eigenvector Let K = R n + and K = R m +, and A R m n 17 / 19

58 Application to Pagerank eigenvector Let K = R n + and K = R m +, and A R m n A(K) K iff a ij 0 17 / 19

59 Application to Pagerank eigenvector Let K = R n + and K = R m +, and A R m n A(K) K iff a ij 0 x K y is equivalent to I x := {i x i > 0} = {i y i > 0} 17 / 19

60 Application to Pagerank eigenvector Let K = R n + and K = R m +, and A R m n A(K) K iff a ij 0 x K y is equivalent to I x := {i x i > 0} = {i y i > 0} In this case, we obtain d H (x, y) = log ( ) x i y j max i,j I x x j y i 17 / 19

61 Application to Pagerank eigenvector Let K = R n + and K = R m +, and A R m n A(K) K iff a ij 0 x K y is equivalent to I x := {i x i > 0} = {i y i > 0} In this case, we obtain d H (x, y) = log ( ) x i y j max i,j I x x j y i Lemma If A R m n +. If there exists J [n] s.t. Ae i K Ae j for all i, j J, and Ae i = 0 for all i J then the projective diameter (A) = max i,j J d H(Ae i, Ae j ) <. 17 / 19

62 More applications Geometric optimization on the psd cone Sra, Hosseini (2013). Conic geometric optimisation on the manifold of positive definite matrices. arxiv: MDPs, Stochastic games, Nonlinear eigenvalue problems, etc. 18 / 19

63 References Nonlinear functional analysis Vol.1 (Fixed-point theorems). E. Zeidler. Nonlinear Perron-Frobenius theory. Lemmens, Nussbaum (2013). 19 / 19