Supplement: Universal Self-Concordant Barrier Functions

Similar documents
Largest dual ellipsoids inscribed in dual cones

Lecture 9 Sequential unconstrained minimization

Primal-Dual Symmetric Interior-Point Methods from SDP to Hyperbolic Cone Programming and Beyond

Lecture 5. The Dual Cone and Dual Problem

Recursive Construction of Optimal Self-Concordant Barriers for Homogeneous Cones

On self-concordant barriers for generalized power cones

Augmented self-concordant barriers and nonlinear optimization problems with finite complexity

L. Vandenberghe EE236C (Spring 2016) 18. Symmetric cones. definition. spectral decomposition. quadratic representation. log-det barrier 18-1

12. Interior-point methods

Barrier Method. Javier Peña Convex Optimization /36-725

Constructing self-concordant barriers for convex cones

On Conically Ordered Convex Programs

Using Schur Complement Theorem to prove convexity of some SOC-functions

Nonsymmetric potential-reduction methods for general cones

Advances in Convex Optimization: Theory, Algorithms, and Applications

Agenda. Interior Point Methods. 1 Barrier functions. 2 Analytic center. 3 Central path. 4 Barrier method. 5 Primal-dual path following algorithms

What can be expressed via Conic Quadratic and Semidefinite Programming?

A Primal-Dual Second-Order Cone Approximations Algorithm For Symmetric Cone Programming

A New Self-Dual Embedding Method for Convex Programming

Determinant maximization with linear. S. Boyd, L. Vandenberghe, S.-P. Wu. Information Systems Laboratory. Stanford University

A Second-Order Path-Following Algorithm for Unconstrained Convex Optimization

1. Gradient method. gradient method, first-order methods. quadratic bounds on convex functions. analysis of gradient method

Convex Feasibility Problems

8. Geometric problems

Primal-Dual Interior-Point Methods. Javier Peña Convex Optimization /36-725

arxiv: v1 [math.fa] 16 Jun 2011

Second-order cone programming

Interior Point Methods: Second-Order Cone Programming and Semidefinite Programming

Euler Equations: local existence

Interior Point Algorithms for Constrained Convex Optimization

3. Convex functions. basic properties and examples. operations that preserve convexity. the conjugate function. quasiconvex functions

10. Unconstrained minimization

EE/ACM Applications of Convex Optimization in Signal Processing and Communications Lecture 17

POLYNOMIAL OPTIMIZATION WITH SUMS-OF-SQUARES INTERPOLANTS

Primal Central Paths and Riemannian Distances for Convex Sets

12. Interior-point methods

8. Geometric problems

3. Convex functions. basic properties and examples. operations that preserve convexity. the conjugate function. quasiconvex functions

PRIMAL-DUAL INTERIOR-POINT METHODS FOR SELF-SCALED CONES

Hessian Riemannian Gradient Flows in Convex Programming

DO NOT OPEN THIS QUESTION BOOKLET UNTIL YOU ARE TOLD TO DO SO

An interior-point trust-region polynomial algorithm for convex programming

Primal Central Paths and Riemannian Distances for Convex Sets

LECTURE 7. k=1 (, v k)u k. Moreover r

Self-Concordant Barrier Functions for Convex Optimization

Convex Optimization. Prof. Nati Srebro. Lecture 12: Infeasible-Start Newton s Method Interior Point Methods

Convex Optimization. Newton s method. ENSAE: Optimisation 1/44

Inner approximation of convex cones via primal-dual ellipsoidal norms

Lecture 15 Newton Method and Self-Concordance. October 23, 2008

Primal-dual IPM with Asymmetric Barrier

ESTIMATES FOR THE MONGE-AMPERE EQUATION

On well definedness of the Central Path

Lecture 5. Theorems of Alternatives and Self-Dual Embedding

Central Paths in Semidefinite Programming, Generalized Proximal Point Method and Cauchy Trajectories in Riemannian Manifolds

An Optimal Affine Invariant Smooth Minimization Algorithm.

Convex analysis and profit/cost/support functions

Matrix Support Functional and its Applications

Key words. Complementarity set, Lyapunov rank, Bishop-Phelps cone, Irreducible cone

Lecture 5: Hyperbolic polynomials

New stopping criteria for detecting infeasibility in conic optimization

15. Conic optimization

Convex Optimization M2

Einstein-Hessian barriers on convex cones

On John type ellipsoids

A LOCALIZATION PROPERTY AT THE BOUNDARY FOR MONGE-AMPERE EQUATION

5. Duality. Lagrangian

INTERIOR-POINT ALGORITHMS FOR CONVEX OPTIMIZATION BASED ON PRIMAL-DUAL METRICS

Applications of Linear Programming

Convex Functions. Daniel P. Palomar. Hong Kong University of Science and Technology (HKUST)

Lecture: Duality of LP, SOCP and SDP

Convex Functions. Pontus Giselsson

Lecture: Smoothing.

Unconstrained minimization: assumptions

Interior Point Methods for Linear Programming: Motivation & Theory

Primal-dual path-following algorithms for circular programming

Lecture 6: Conic Optimization September 8

Unconstrained minimization

Second Order Elliptic PDE

INTERIOR-POINT ALGORITHMS FOR CONVEX OPTIMIZATION BASED ON PRIMAL-DUAL METRICS

The local equicontinuity of a maximal monotone operator

MATH 31BH Homework 5 Solutions

Self-equilibrated Functions in Dual Vector Spaces: a Boundedness Criterion

Bandit Convex Optimization

Second-Order Cone Program (SOCP) Detection and Transformation Algorithms for Optimization Software

Sobolev regularity for the Monge-Ampère equation, with application to the semigeostrophic equations

Extreme Abridgment of Boyd and Vandenberghe s Convex Optimization

Dedicated to Michel Théra in honor of his 70th birthday

An Infeasible Interior Proximal Method for Convex Programming Problems with Linear Constraints 1

Optimal Transportation. Nonlinear Partial Differential Equations

On duality gap in linear conic problems

Math 4263 Homework Set 1

A polynomial-time interior-point method for conic optimization, with inexact barrier evaluations

Deterministic Dynamic Programming

Lecture No 1 Introduction to Diffusion equations The heat equat

Advanced Analysis Qualifying Examination Department of Mathematics and Statistics University of Massachusetts. Monday, August 26, 2013

Lecture 1: Background on Convex Analysis

Optimization and Optimal Control in Banach Spaces

Change Of Variable Theorem: Multiple Dimensions

Linear DifferentiaL Equation

1. Introduction Boundary estimates for the second derivatives of the solution to the Dirichlet problem for the Monge-Ampere equation

Transcription:

IE 8534 1 Supplement: Universal Self-Concordant Barrier Functions

IE 8534 2 Recall that a self-concordant barrier function for K is a barrier function satisfying 3 F (x)[h, h, h] 2( 2 F (x)[h, h]) 3/2, F (x)[h] θ( 2 F (x)[h, h]) 1/2, for any x int K and any direction h R n. Furthermore, we call a barrier function F (x) to be ν-logarithmically homogeneous if for all x int K and t > 0. F (tx) = F (x) ν ln t

IE 8534 3 A fundamentally important result due to Nesterov and Nemirovski is the following: Theorem 1 Any closed convex cone admits a self-concordant, logarithmically homogeneous barrier function where θ = O( n). In this note we shall show why this is the case. First, let us note some basic facts. Proposition 1 Suppose that F (x) is a smooth ν-logarithmically homogeneous barrier function for K. Then the following identities hold where x int K and t > 0: F (tx) = 1 F (x); t 2 F (tx) = 1 t 2 2 F (x); 2 F (x)x = F (x); ( F (x)) T x = ν.

IE 8534 4 For an ν-logarithmically homogeneous function, the second property of the self-concordance always holds F (x)[h] = h T F (x) = h T 2 F (x)x = [( 2 F (x)) 1/2 h] T [( 2 F (x)) 1/2 x] h T 2 F (x)h x T 2 F (x)x = ν( 2 F (x)[h, h]) 1/2, where we used the Cauchy-Schwartz inequality.

IE 8534 5 There are intimate relationships between a convex cone and its barrier function; the dual cone and the conjugate of the barrier funciton. Recall that the conjugate of a function f is: f (s) = sup{( s) T x f(x) x dom f}, where dom f stands for the domain of the function f. Suppose that f is a barrier function of K. Then f is a barrier function of K. Moreover, for x int dom f and s int dom f the following three statements are equivalent s = f(x); x = f (s); x T s = f(x) + f (s).

IE 8534 6 For a convex cone K, the following function is known as its characteristic function ϕ(x) = e x,y dy. K Let F (x) = ln ϕ(x). One can show the following properties of F : F (x) is a barrier function for K; F (x) is strictly convex; F (x) is n-logarithmically homogeneous.

IE 8534 7 Let us consider the last property. Consider a linear transformation A that keeps K invariant, i.e. AK = K; in that case, A T K = K. Then, ϕ(ax) = e Ax,y dy K = e x,a T y dy K }{{} = dy / det A y :=A T y K e x,y = ϕ(x)/ det A. Now, for t > 0, the transformation A = ti keeps K invariant. Therefore, ϕ(tx) = ϕ(x)/t n, and so F (tx) = F (x) nt.

IE 8534 8 Moreover, A keeps K invariant, because F (Ax) = F (x)/ det A we have F (Ax) = F (x) ln det A. Differentiating the above identity yields for any k 1. k F (Ax)[Ah, Ah,..., Ah] = k F (x)[h, h,..., h] }{{}}{{} k k In order to show that the first property of the self-concordance is also satisfied, we fix a given point ι int K. For any x int K, there is a mapping A x such that ι = A x x. Let h = A x h. It will be sufficient to prove that there exists a constant C > 0 such that 3 F (ι)[h, h, h ] C( 2 F (ι)[h, h ]) 3/2 h R n which is true because 2 F (ι) is positive definite.

IE 8534 9 The characteristic function exists for any convex cone. The barrier function defined that way is also known as the universal barrier function. As we have just seen, it is also a self-concordant barrier. One problem, however, is that the universal barrier function is not easily computable. For special cases, one may be able to compute the associated universal barrier function. For instance, it is easy to see that the characteristic function for R n + is: R n + e x 1y 1 x n y n dy = 1/ n x i, and so the universal barrier function is simply the logarithmic barrier function n i=1 ln x i. i=1

IE 8534 10 Interestingly, there is also a very nice geometric interpretation of the universal barrier function. Let S be a convex set. Introduce the set function S (x) := ( x + S). Let u(x) := ln(vol(s (x))). Nesterov and Nemirovski showed that u(x) is a self-concordant barrier function for S with complexity parameter θ = O( n). Güler proved that if S is a convex cone (S = K) then the following identity holds F (x) = u(x) + ln n!.

IE 8534 11 Key References: Yu. Nesterov and A. Nemirovski, Interior-Point Polynomial Algorithms in Convex Programming, SIAM Studies in Applied Mathematics, 13, 1994. O. Güler, Barrier Functions in Interior Point Methods, Mathematics of Operations Research, 21, pp. 860 885, 1996.