Lecture 3: Convex Sets and Functions 3.1 Convex Sets

Lecture 3: Convex Sets and Functions 3.1 Convex Sets

EE 227A: Convex Optimization and Applications January 24, 2012 Lecture 3: Convex Sets and Functions Lecturer: Laurent El Ghaoui Reading assignment: Chapters 2 (except x2.6) and sections 3.1, 3.2, 3.3 of BV. You can also look at section 3.1 of the web textbook. 3.1 Convex Sets Definition. A subset C of Rn is convex if and only if it contains the line segment between any two points in it: 8 x1; x2 2 C; 8 θ1 ≥ 0; θ2 ≥ 0; θ1 + θ2 = 1 : θ1x1 + θ2x2 2 C: Some important special cases of convex sets are the following. • The set is said to be an affine subspace if it contains the entire line passing through any two points. This corresponds to the condition above, with θ1; θ2 arbitrary. Subspaces and affine subspaces are convex. • The set is said to be a convex cone if the condition above holds, but with the restriction θ1 + θ2 = 1 removed. Examples. • The convex hull of a set of points fx1; : : : ; xmg is defined as ( m m ) X m X Co(x1; : : : ; xm) := λixi : λ 2 R+ ; λi = 1 ; i=1 i=1 and is convex. The conic hull: ( m ) X m λixi : λ 2 R+ i=1 is a convex cone. • For a 2 Rn, and b 2 R, the hyperplane H = fx : aT x = bg is affine. The half-space fx : aT x ≤ bg is convex. 3-1 EE 227A Lecture 3 | January 24, 2012 Sp'12 n×n n • For a square, non-singular matrix R 2 R , and xc 2 R , the ellipsoid fxc + Ru : kuk2 ≤ 1g is convex. (The center of the epllipsoid is xc, and you can think of R as the "radius".) With P = RRT , and assuming R is full rank, we can describe the ellipsoid as T −1 x :(x − xc) P (x − xc) ≤ 1 : • A polyhedron is a set described by a finite number of affine inequalities and equalities: P = fx : Ax ≤ b; Cx = dg ; where A; C are matrices, b; d are vectors, and inequalities are understood component- wise. Sometimes bounded polyhedra are referred to as polytopes. • The probability simplex ( n ) n X p 2 R+ : pi = 1 i=1 is a special case of a polyhedron, and is useful to describe discrete probabilities. • The second-order cone n+1 (x; t) 2 R : t ≥ kxk2 (3.1) is a convex cone. It is sometimes called \ice-cream cone", for obvious reasons. (We will prove the convexity of this set later.) • The positive semi-definite cone n T n×n S+ := X = X 2 R : X 0 is a convex cone. (Again, we will prove the convexity of this set later.) Support and indicator functions. For a given set S, the function T φS(x) := max x u u2S is called the support function of S. If S is the unit ball for some norm: S = fu : kuk ≤ 1g, then the support function of S is the dual norm. Another important function associated with S is the indicator function 0 x 2 S; I (x) = S +1 x 62 S: 3-2 EE 227A Lecture 3 | January 24, 2012 Sp'12 Operations that preserve convexity. Two important operations that preserve convexity are: • Intersection: the intersection of a (possibly infinite) family of convex sets is convex. n We can use this property to prove that the semi-definite cone S+ is convex, since n T n×n n T S+ = X = X 2 R : 8 z 2 R ; z Xz ≥ 0 ; from which we see that the set is the intersection of the subspace of symmetric matrices with a set described by an infinite number of linear inequalities of the form zT Xz ≥ 0, indexed by z 2 Rn. Likewise, the second-order cone defined in (3.1) is convex, since the T condition t ≥ kxk2 is equivalent to the infinite number of affine inequalities t ≥ u x, kuk2 ≤ 1. • Affine transformation: If a function is affine (that is, it is the sum of a linear function and a constant), and C is convex, then the set f(C) := ff(x): x 2 Cg is convex. A particular example is projection on a subspace, which preserves convexity. Separation theorems. There are many versions of separation theorems in convex analy- sis. One of them is the separating hyperplane theorem: Theorem 1 (Separating hyperplane). If C, D are two convex subsets of Rn that do not intersect, then there is an hyperplane that separates them, that is, there exit a 2 Rn, a 6= 0, and b 2 R, such that aT x ≤ b for every x 2 C, and aT x ≥ b for every x 2 D. Another result involves the separation of a set from a point on its boundary: Theorem 2 (Supporting hyperplane). If C ⊆ Rn is convex and non-empty, then for any x0 at the boundary of C, there exist a supporting hyperplane to C at x0, meaning that there n T exist a 2 R , a 6= 0, such that a (x − x0) ≤ 0 for every x 2 C. 3.2 Convex Functions Domain of a function. The domain of a function f : Rn ! R is the set domf ⊆ Rn over which f is well-defined, in other words: dom f := fx 2 Rn : −∞ < f(x) < +1g: Here are some examples: • The function with values f(x) = log(x) has domain dom f = R++. n • The function with values f(X) = log det(X) has domain dom f = S++ (the set of positive-definite matrices). 3-3 EE 227A Lecture 3 | January 24, 2012 Sp'12 Definition of convexity. A function f : Rn ! R is convex if i) dom f is convex; ii) 8 x; y 2 dom f and 8 λ 2 [0; 1], f (λx + (1 − λ)y) ≤ λf(x) + (1 − λ)f(y). Note that the convexity of the domain is required. For example, the function f : R ! R defined as x if x 62 [−1; 1] f(x) = +1 otherwise is not convex, although is it linear (hence, convex) on its domain ] − 1; −1) [ (1; +1[. We say that a function is concave if −f is convex. Here are some examples: • The support function of any set is convex. • The indicator function of a set is convex if and only if the set is convex. T T n • The quadratic function f(x) = x P x + 2q x + r, with P 2 S++, is convex. (For a proof, see later.) • The function f : R ! R defined as f(x) = 1=x for x > 0 and f(x) = +1 is convex. Alternate characterizations of convexity. Let f : Rn ! R. The following are equiva- lent conditions for f to be convex. • Epigraph condition: f is convex if and only if its epigraph epif := (x; t) 2 Rn+1 : t ≥ f(x) is convex. We can us this result to prove for example, that the largest eigenvalue n function λmax : S ! R, which to a given n × n symmetric matrix X associates its largest eigenvalue, is convex, since the condition λmax(X) ≤ t is equivalent to the n condition that tI − X 2 S+. • Restriction to a line: The function f is convex if and only if its restriction to any line n n is convex, meaning that for every x0 2 R , and v 2 R , the function g(t) := f(x0 + tv) is convex. For example, the function f(X) = log det X is convex. (Prove this as an exercise.) You can also use this to prove that the quadratic function f(x) = xT P x + 2qT x + r is convex if and only if P 0. • First-order condition: If f is differentiable (that is, domf is open and the gradient exists everywhere on the domain), then f is convex if and only if 8 x; y : f(y) ≥ f(x) + rf(x)T (y − x): The geometric interpretation is that the graph of f is bounded below everywhere by anyone of its tangents. 3-4 EE 227A Lecture 3 | January 24, 2012 Sp'12 • Second-order condition: If f is twice differentiable, then it is convex if and only if its Hessian r2f is positive semi-definite everywhere. This is perhaps the most commonly known characterization of convexity. For example, the function f(x; t) = xT x=t with domain f(x; t): t > 0g, is convex. Pn (Check this!) Other examples include the log-sum-exp function, f(x) = log i=1 exp xi, and the quadratic function alluded to above. Operations that preserve convexity. • The nonnegative weighted sum of convex functions is convex. • The composition with an affine function preserves convexity: if A 2 Rm×n, b 2 Rm and f : Rm ! R is convex, then the function g : Rn ! R with values g(x) = f(Ax + b) is convex. • The pointwise maximum of a family of convex functions is convex: if (fα)α2A is a family of convex functions index by α, then the function f(x) := max fα(x) α2A is convex. For example, the dual norm x ! max yT x y :kyk≤1 is convex, as the maximum of convex (in fact, linear) functions (indexed by the vector y). Another example is the largest singular value of a matrix A: f(A) = σmax(A) = n maxx : kxk2=1 kAxk2. Here, each function (indexed by x 2 R ) A ! kAxk2 is convex, since it is the composition of the Euclidean norm (a convex function) with an affine function A ! Ax. Also, this can be used to prove convexity of the function we introduced in lecture 2, k n X T X n kxk1;k := jxj[i] = max u jxj : ui = k; u 2 f0; 1g ; u i=1 i=1 where we use the fact that for any u feasible for the maximization problem, the function x ! uT jxj is convex (since u ≥ 0).

View Full Text

Details

  • File Type
    pdf
  • Upload Time
    -
  • Content Languages
    English
  • Upload User
    Anonymous/Not logged-in
  • File Pages
    8 Page
  • File Size
    -

Download

Channel Download Status
Express Download Enable

Copyright

We respect the copyrights and intellectual property rights of all users. All uploaded documents are either original works of the uploader or authorized works of the rightful owners.

  • Not to be reproduced or distributed without explicit permission.
  • Not used for commercial purposes outside of approved use cases.
  • Not used to infringe on the rights of the original creators.
  • If you believe any content infringes your copyright, please contact us immediately.

Support

For help with questions, suggestions, or problems, please contact us