CS295: Convex Optimization. Xiaohui Xie Department of Computer Science University of California, Irvine

Transcription

1 CS295: Convex Optimization Xiaohui Xie Department of Computer Science University of California, Irvine

2 Convex set Definition A set C is called convex if x,y C = θx+(1 θ)y C θ [0,1] In other words, a set C is convex if the line segment between any two points in C lies in C.

3 Convex set: examples Figure: Examples of convex and nonconvex sets

4 Convex combination Definition A convex combination of the points x 1,,x k is a point of the form θ 1 x 1 + +θ k x k, where θ 1 + +θ k = 1 and θ i 0 for all i = 1,,k. A set is convex if and only if it contains every convex combinations of the its points.

5 Convex hull Definition The convex hull of a set C, denoted conv C, is the set of all convex combinations of points in C: { k } k convc = θ i x i x i C,θ i 0,i = 1,,k, θ k = 1 i=1 i=1

6 Convex hull Definition The convex hull of a set C, denoted conv C, is the set of all convex combinations of points in C: { k } k convc = θ i x i x i C,θ i 0,i = 1,,k, θ k = 1 i=1 i=1 Properties: A convex hull is always convex conv C is the smallest convex set that contains C, i.e., B C is convex = convc B

7 Convex hull: examples Figure: Examples of convex hulls

8 Convex cone A set C is called a cone if x C = θx C, θ 0.

9 Convex cone A set C is called a cone if x C = θx C, θ 0. A set C is a convex cone if it is convex and a cone, i.e., x 1,x 2 C = θ 1 x 1 +θ 2 x 2 C, θ 1,θ 2 0

10 Convex cone A set C is called a cone if x C = θx C, θ 0. A set C is a convex cone if it is convex and a cone, i.e., x 1,x 2 C = θ 1 x 1 +θ 2 x 2 C, θ 1,θ 2 0 The point k i=1 θ ix i, where θ i 0, i = 1,,k, is called a conic combination of x 1,,x k.

11 Convex cone A set C is called a cone if x C = θx C, θ 0. A set C is a convex cone if it is convex and a cone, i.e., x 1,x 2 C = θ 1 x 1 +θ 2 x 2 C, θ 1,θ 2 0 The point k i=1 θ ix i, where θ i 0, i = 1,,k, is called a conic combination of x 1,,x k. The conic hull of a set C is the set of all conic combinations of points in C.

12 Conic hull: examples Figure: Examples of conic hull

13 Hyperplanes and halfspaces A hyperplane is a set of the form {x R n a T x = b} where a = 0,b R.

14 Hyperplanes and halfspaces A hyperplane is a set of the form {x R n a T x = b} where a = 0,b R. A (closed) halfspace is a set of the form {x R n a T x b} where a = 0,b R. a is the normal vector hyperplanes and halfspaces are convex

15 Euclidean balls and ellipsoids Euclidean ball in R n with center x c and radius r: B(x c,r) = {x x x c 2 r} = {x c +ru u 2 1}

16 Euclidean balls and ellipsoids Euclidean ball in R n with center x c and radius r: B(x c,r) = {x x x c 2 r} = {x c +ru u 2 1} ellipsoid in R n with center x c : { } E = x (x x c ) T P 1 (x x c ) 1 where P S n ++ (i.e., symmetric and positive definite) the lengths of the semi-axes of E are given by λ i, where λ i are the eigenvalues of P. An alternative representation of an ellipsoid: with A = P 1/2 E = {x c +Au u 2 1}

17 Euclidean balls and ellipsoids Euclidean ball in R n with center x c and radius r: B(x c,r) = {x x x c 2 r} = {x c +ru u 2 1} ellipsoid in R n with center x c : { } E = x (x x c ) T P 1 (x x c ) 1 where P S n ++ (i.e., symmetric and positive definite) the lengths of the semi-axes of E are given by λ i, where λ i are the eigenvalues of P. An alternative representation of an ellipsoid: with A = P 1/2 E = {x c +Au u 2 1} Euclidean balls and ellipsoids are convex.

18 Norms A function f : R n R is called a norm, denoted x, if nonegative: f(x) 0, for all x R n definite: f(x) = 0 only if x = 0 homogeneous: f(tx) = t f(x), for all x R n and t R satisfies the triangle inequality: f(x +y) f(x)+f(y) notation: denotes a general norm; symb denotes a specific norm Distance: dist(x,y) = x y between x,y R n.

19 Examples of norms l p -norm on R n : x p = ( x 1 p + + x n p ) 1/p l1 -norm: x 1 = i x i l -norm: x = max i x i Quadratic norms: For P S n ++, define the P-quadratic norm as x P = (x T Px) 1/2 = P 1/2 x 2

20 Equivalence of norms Let a and b be norms on R n. Then α,β > 0 such that x R n, α x a x b β x a. Norms on any finite-dimensional vector space are equivalent (define the same set of open subsets, the same set of convergent sequences, etc.)

21 Dual norm Let be a norm on R n. The associated dual norm, denoted, is defined as z = sup {z T x x 1}.

22 Dual norm Let be a norm on R n. The associated dual norm, denoted, is defined as z = sup {z T x x 1}. z T x x z for all x,z R n x = x for all x R n The dual of the Euclidean norm is the Euclidean norm (Cauchy-Schwartz inequality)

23 Dual norm Let be a norm on R n. The associated dual norm, denoted, is defined as z = sup {z T x x 1}. z T x x z for all x,z R n x = x for all x R n The dual of the Euclidean norm is the Euclidean norm (Cauchy-Schwartz inequality) The dual of the l p -norm is the l q -norm, where 1/p +1/q = 1 (Holder s inequality) The dual of the l norm is the l 1 norm The dual of the l 2 -norm on R m n is the nuclear norm, Z 2 = sup {tr(z T X) X 2 1} where r = rank Z. = σ 1 (Z)+ +σ r (Z) = tr(z T Z) 1/2,

24 Norm balls and norm cones norm ball with center x c and radius r: {x x x c r} norm cone: C = {(x,t) x t} R n+1 the second-order cone is the norm cone for the Euclidean norm norm balls and cones are convex

25 Polyhedra A polyhedron is defined as the solution set of a finite number of linear equalities and inequalities: P = {x Ax b,cx = d} where A R m n, A R p n, and denotes vector inequality or componentwise inequality. A polyhedron is the intersection of finite number of halfspaces and hyperplanes.

26 Simplexes The simplex determined by k + 1 affinely independent points v 0,,v k R n is { } C = conv{v 0,,v k } = θ 0 v 0 + +θ k v k θ ર 0,1 T θ = 1 The affine dimension of this simplex is k, so it is often called k-dimensional simplex in R n. Some common simplexes: let e 1,,e n be the unit vectors in R n. unit simplex: conv{0,e 1,,e n } = {x x ર 0,1 T θ 1} probability simplex: conv{e 1,,e n } = {x x ર 0,1 T θ = 1}

27 Positive semidefinite cone notation: S n : the set of symmetric n n matrices S n + = {X S n X ર 0}: symmetric positive semidefinite matrices S n ++ = {X S n X 0} symmetric positive definite matrices S n + is a convex cone, called positive semidefinte cone. Sn ++ comprise the cone interior; all singular positive semidefinite matrices reside on the cone boundary.

28 Positive semidefinite cone: example X = [ ] x y S+ 2 x 0,z 0,xz y 2 y z Figure: Positive semidefinite cone: S 2 +

29 Operations that preserve complexity intersection affine function perspective function linear-fractional functions

30 Intersection If S 1 and S 2 are convex. then S 1 S2 is convex.

31 Intersection If S 1 and S 2 are convex. then S 1 S2 is convex. If S α is convex for every α A, then α A S α is convex.

32 Intersection If S 1 and S 2 are convex. then S 1 S2 is convex.

33 Intersection If S 1 and S 2 are convex. then S 1 S2 is convex. If S α is convex for every α A, then α A S α is convex.

34 Intersection: example 1 Show that the positive semidefinite cone S n + is convex. Proof. S n + can be expressed as S n + = z =0 { } X S n z T Xz 0. Since the set { } X S n z T Xz 0 is a halfspace in S n, it is convex. S n + is the intersection of an infinite number of halfspaces, so it is convex.

35 Intersection: example 2 The set S = {x R m m x k coskt 1, t π/3} k=1 is convex, since it can be expressed as S = t π/3 S t, where S t = {x R m 1 (cost,,cosmt) T x 1}. Figure: The set S for m = 2.

36 Affine function Theorem Suppose f : R n R m is an affine function (i.e., f(x) = Ax +b). Then the image of a convex set under f is convex S R n is convex = f(s) = {f(x) x S} is convex the inverse image of a convex set under f is convex B R m is convex = f 1 (B) = {x f(x) B} is convex

37 Affine function: example 1 Show that the ellipsoid { } E = x (x x c ) T P 1 (x x c ) 1 where P S n ++ is convex. Proof. Let S = {u R n u 2 1} denote the unit ball in R n. Define an affine function f(u) = P 1/2 u +x c E is the image of S under f, so is convex.

38 Affine function: example 2 Show that the solution set of linear matrix inequality (LMI) S = {x R n x 1 A 1 + +x n A n ર B}, where B,A i S m, is convex. Proof. Define an affine function f : R n S m given by f(x) = B (x 1 A 1 + +x n A n ). The solution set S is the inverse image of the positive semidefinite cone S m +, so is convex.

39 Affine function: example 3 Show that the hyperbolic cone where P S+ n, is convex. S = {x R n x T Px (c T x) 2,c T x 0}, Proof. Define an affine function f : R n S n+1 given by f(x) = (P 1/2 x,c T x). The S is the inverse image of the second-order cone, so is convex. {(z,t) z 2 t,t 0},

40 Perspective and linear-fractional function perspective function P : R n+1 R n : P(x,t) = x, dom P = {(x,t) t > 0} t images and inverse images of convex sets under P are convex. linear-fractional function P : R n R m : f(x) = Ax +b c T x +d, dom f = {x ct x +d > 0} images and inverse images of convex sets under f are convex.

41 Generalized inequalities: proper cone Definition A cone K R n is called a proper cone if K is convex K is closed K is solid, which means it has nonempty interior K is pointed, which means that it contains no line (i.e., x K, x K = x = 0)

42 Generalized inequalities: proper cone Definition A cone K R n is called a proper cone if K is convex K is closed K is solid, which means it has nonempty interior K is pointed, which means that it contains no line (i.e., x K, x K = x = 0) Examples: nonnegative orthant K = R n + = {x Rn x i 0, i} positive semidifinite cone K = S n +; how about S n ++? nonnegative polynomials on [0, 1]: K = {x R n x 1 +x 2 t + +x n t n 1 0, t [0,1]}

43 Generalized inequalities: definition A proper cone K can be used to define a generalized inequality, a partial ordering on R n, x K y y x K x K y y x intk where the latter is called a strict generalized inequality.

44 Generalized inequalities: definition A proper cone K can be used to define a generalized inequality, a partial ordering on R n, x K y y x K x K y y x intk where the latter is called a strict generalized inequality. Examples: componentwise inequality (K = R n + ) x R n + y x i y k, i = 1,,n matrix inequality (K = S n +) x S n + y Y X is positive semidefinite

45 Generalized inequalities: properties Many properties of K are similar to on R: transitive: x K y, y K z = x K z reflexive: x K x antisymmetric: x K y, y K x = x = y preserved under addition: x K y, u K v = x +u K y +v preserved under nonnegative scaling: x K y, α 0 = αx K αy preserved under limits: suppose limx i = x, limy i = y. Then x i K y i, i = x K y

46 Minimum and minimal elements K is not in general a linear ordering: we can have x K y y K x x S is called the minimum element of S with respect to K if y S = x K y x S is called the minimal element of S with respect to K if y S, y K x = y = x

47 Minimum and minimal elements K is not in general a linear ordering: we can have x K y y K x x S is called the minimum element of S with respect to K if y S = x K y x S is called the minimal element of S with respect to K if Example: y S, y K x = y = x Figure: K = R 2 +. x 1 is the minimum element of S 1. x 2 is the minimal element of S 2.

48 Separating hyperplane theorem Theorem Suppose C and D are two convex sets that do not intersect, i.e., C D =. Then there exist a = 0 and b such that a T x b for x C, and a x b b for x D The hyperplane {x a x = b} is called a separating hyperplane for C and D. Figure: Examples of convex and nonconvex sets

49 Supporting hyperplane theorem supporting hyperplane to set C at boundary point x 0 {x a x = a T x 0 } where a = 0 and satisfies a T x a T x 0 for all x C. Theorem (supporting hyperplane theorem) If C is convex, then there exists a supporting hyperplane at every boundary point of C. Figure: Examples of convex and nonconvex sets

50 Dual cones Definition (dual cones) Let K be a cone. The set is called the dual cone of K. Property: K = {y x T y 0 x K} K is always convex, even when the original cone K is not (why? intersection of convex sets) y K if and only if y is the normal of a hyperplane that supports K at the origin

51 Dual cones : examples Examples: K = R+: n K = R+ n K = S+ n : K = S+ n K = {(x,t) x 2 t}: K = {(x,t) x 2 t} K = {(x,t) x t}: K = {(x,t) x t} the first three examples are self-dual cones

52 Dual of positive semidefinite cone Theorem The positive semidefinite cone S n + is self-dual, i.e., given Y Sn, tr(xy) 0 X S n + Y Sn +

53 Dual of positive semidefinite cone Theorem The positive semidefinite cone S n + is self-dual, i.e., given Y Sn, tr(xy) 0 X S n + Y Sn + Proof. To prove =, suppose Y / S n +. Then q with q T Yq = tr(qq T Y) < 0, which contradicts the lefthand condition.

54 Dual of positive semidefinite cone Theorem The positive semidefinite cone S n + is self-dual, i.e., given Y Sn, tr(xy) 0 X S n + Y Sn + Proof. To prove =, suppose Y / S n +. Then q with q T Yq = tr(qq T Y) < 0, which contradicts the lefthand condition. To prove =, since X ર 0, write X = n i=1 λ iq i q T i, where λ i 0 for all i. Then because Y ર 0. tr(xy) = tr(y n λ i q i qi T ) = i=1 n λ i qi T Yq i 0, i=1

55 Dual of a norm cone Theorem The dual of the cone K = {(x,t) R n+1 x t} associated with a norm in R n is the cone defined by the dual norm, K = {(u,s) R n+1 u s}, where the dual norm is given by u = sup{u T x x 1}.

56 Dual of a norm cone Theorem The dual of the cone K = {(x,t) R n+1 x t} associated with a norm in R n is the cone defined by the dual norm, K = {(u,s) R n+1 u s}, where the dual norm is given by u = sup{u T x x 1}. Proof. We need to show x T u +ts 0 x t u s The = direction follows from the definition of the dual norm. To prove =, suppose u > s. Then by the definition of dual norm, x with x 1 and x T u s. Taking t = 1, we have u T ( x)+v < 0, which is a contradiction.

57 Dual cones and generalized inequalities Properties of dual cones: let K be the dual of a convex cone K. K is a convex cone (intersection of a set of homogeneous halfspaces) K 1 K 2 = K 2 K 1 K is closed (intersection of a set of closed sets) K is the closure of K (if K is closed, then K = K) dual cones of proper cones are proper, hence define generalized inequalities: y ર K 0 y T x 0 for all x ર K 0