DC Programming: A brief tutorial. Andres Munoz Medina Courant Institute of Mathematical Sciences

Transcription

1 DC Programming: A brief tutorial. Andres Munoz Medina Courant Institute of Mathematical Sciences munoz@cims.nyu.edu

2 Difference of Convex Functions Definition: A function exists f is said to be DC if there g, h convex functions such that f = g h Definition: A function is locally DC if for every there exists a neighborhood U such that f U is DC. x

3 Contents Motivation. DC functions and properties. DC programming and DCA. CCCP and convergence results. Global optimization algorithms.

4 Motivation Not all machine learning problems are convex anymore Transductive SVM s [Wang et. al 2003] Kernel learning [Argyriou et. al 2004] Structure prediction [Narasinham 2012] Auction mechanism design [MM and Muñoz]

5 Notation Let g be a convex function. The conjugate of g is defined as g g (y) =supy, x g(x) x For >0, g(x 0 ) denotes the subdifferential of g at, i.e. x 0 g(x 0 )={v R n g(x) g(x 0 )+x x 0,v g(x 0 ) will denote the exact subdifferential.

6 DC functions A function f : R R is DC iff is the integral of a function of bounded variation. [Hartman 59] A locally DC function is globally DC. [Hartman 59] All twice continuously differentiable functions are DC. Closed under sum, negation, supremum and products.

7 DC programming DC programming refers to optimization problems of the form. min x g(x) h(x) More generally, for f i (x) DC functions min x g(x) h(x) subject to f i (x) 0

8 Global optimality conditions. A point x is a global solution if and only if h(x ) g(x ) Let w =infg(x) h(x), then a point x is a global solution if and only if x 0=inf x { h(x)+t g(x) t w }

9 Global optimality conditions. A point x is a global solution if and only if h(x ) g(x ) Let w =infg(x) h(x), then a point x is a global solution if and only if x 0=inf x { h(x)+t g(x) t w }

10 Global optimality conditions. A point x is a global solution if and only if h(x ) g(x ) Let w =infg(x) h(x), then a point x is a global solution if and only if x 0=inf x { h(x)+t g(x) t w }

11 Local optimality conditions If x verifies h(x ) int g(x ), then is a strict local minimizer of g h

12 DC duality By definition of conjugate function inf x g(x) h(x) =inf x =inf x =inf y g(x) (supx, y h (y)) inf y This is the dual of the original problem y h (y)+g(x) x, y h (y) g (y)

13 DC algorithm. We want to find a sequence x k that decreases the function at every step. Use duality. If y h(x 0 ) then h (y) g (y) =x 0,y h(x 0 ) sup(x, y g(x)) x =inf x g(x) h(x)+x 0 x, y g(x 0 ) h(x 0 )

14 DC algorithm Solve the partial problems S(x ) inf{h (y) g (y) :y h(x )} T (y ) inf{g(x) h(x) :x g (y )} Choose y and. k S(x k ) x k+1 T (y k ) Solve concave minimization problems. Simplified DCA y k h(x k ) x k g (y k )

15 DCA as CCCP If the function is differentiable, the simplified DCA becomes y k = h(x k ) and x k+1 g (y k ) Equivalent to

16 DCA as CCCP If the function is differentiable, the simplified DCA becomes y k = h(x k ) and x k+1 g (y k ) Equivalent to

17 DCA as CCCP If the function is differentiable, the simplified DCA becomes y k = h(x k ) and x k+1 g (y k ) x k+1 argmin g(x) x, h(x k ) Equivalent to

18 DCA as CCCP If the function is differentiable, the simplified DCA becomes y k = h(x k ) and x k+1 g (y k ) x k+1 argmin g(x) x, h(x k ) Equivalent to x k+1 argmin g(x) h(x k ) x x k, h(x k )

19 CCCP as a majorization minimization algorithm To minimize f, MM algorithms build a majorization function F such that f(x) F (x, y) x, y f(x) =F (x, x) x Do coordinate descent on F In our scenario F (x, y) =g(x) h(y) x y, h(y)

20 Convergence results Unconstrained DC functions: Convergence to a local minimum (no rate of convergence). Bound depends on moduli of convexity. [PD Tao, LT Hoai An 97] Unconstrained smooth optimization: Linear or almost quadratic convergence depending on curvature [Roweis et. al 03] Constrained smooth optimization: Convergence without rate using Zangwill s theory. [Lanckriet, Sriperumbudur 09]

21 Global convergence NP-hard in general: Minimizing a quadratic function with one negative eigenvalue with linear constraints. [Pardalos 91] Mostly branch and bound methods and cutting plane methods [H. Tuy 03, Horst and Thoai 99] Some successful results with low rank functions.

22 Cutting plane algorithm[h. Tuy 03] All limit points of this algorithm are global minimizers Finite convergence for piecewise linear functions (Conjecture). Keeps track of exponentially many vertices.

23 Cutting plane algorithm[h. Tuy 03] All limit points of this algorithm are global minimizers Finite convergence for piecewise linear functions (Conjecture). Keeps track of exponentially many vertices.

24 Cutting plane algorithm[h. Tuy 03] All limit points of this algorithm are global minimizers Finite convergence for piecewise linear functions (Conjecture). Keeps track of exponentially many vertices.

25 Cutting plane algorithm[h. Tuy 03] All limit points of this algorithm are global minimizers Finite convergence for piecewise linear functions (x 0,t 0 ) (Conjecture). Keeps track of exponentially many vertices.

26 Cutting plane algorithm[h. Tuy 03] All limit points of this algorithm are global minimizers Finite convergence for piecewise linear functions (x 0,t 0 ) (Conjecture). Keeps track of exponentially many vertices.

27 Cutting plane algorithm[h. Tuy 03] All limit points of this algorithm are global minimizers Finite convergence for piecewise linear functions (x 0,t 0 ) (Conjecture). Keeps track of exponentially many vertices.

28 Cutting plane algorithm[h. Tuy 03] All limit points of this algorithm are global minimizers Finite convergence for piecewise linear functions (Conjecture). Keeps track of exponentially many vertices.

29 Cutting plane algorithm[h. Tuy 03] All limit points of this algorithm are global minimizers Finite convergence for piecewise linear functions (x 1,t 1 ) (Conjecture). Keeps track of exponentially many vertices.

30 Cutting plane algorithm[h. Tuy 03] All limit points of this algorithm are global minimizers Finite convergence for piecewise linear functions (x 1,t 1 ) (Conjecture). Keeps track of exponentially many vertices.

31 Cutting plane algorithm[h. Tuy 03] All limit points of this algorithm are global minimizers Finite convergence for piecewise linear functions (x 1,t 1 ) (Conjecture). Keeps track of exponentially many vertices.

32 Cutting plane algorithm[h. Tuy 03] All limit points of this algorithm are global minimizers Finite convergence for piecewise linear functions (Conjecture). Keeps track of exponentially many vertices.

33 Cutting plane algorithm[h. Tuy 03] All limit points of this algorithm are global minimizers Finite convergence for piecewise linear functions (x 2,t 2 ) (Conjecture). Keeps track of exponentially many vertices.

34 Cutting plane algorithm[h. Tuy 03] All limit points of this algorithm are global minimizers Finite convergence for piecewise linear functions (x 2,t 2 ) (Conjecture). Keeps track of exponentially many vertices.

35 Cutting plane algorithm[h. Tuy 03] All limit points of this algorithm are global minimizers Finite convergence for piecewise linear functions (x 2,t 2 ) (Conjecture). Keeps track of exponentially many vertices.

36 Cutting plane algorithm[h. Tuy 03] All limit points of this algorithm are global minimizers Finite convergence for piecewise linear functions (Conjecture). Keeps track of exponentially many vertices.

37 Cutting plane algorithm[h. Tuy 03] All limit points of this algorithm are global minimizers Finite convergence for piecewise linear functions (x 3,t 3 ) (Conjecture). Keeps track of exponentially many vertices.

38 Cutting plane algorithm[h. Tuy 03] All limit points of this algorithm are global minimizers Finite convergence for piecewise linear functions (x 3,t 3 ) (Conjecture). Keeps track of exponentially many vertices.

39 Cutting plane algorithm[h. Tuy 03] All limit points of this algorithm are global minimizers Finite convergence for piecewise linear functions (x 3,t 3 ) (Conjecture). Keeps track of exponentially many vertices.

40 Branch and bound.[hoarst and Thoai 99] Main idea is to keep upper and lower bounds of g h on simplices. S k Upper bound: Evaluate g(x k ) h(x k ) for x k S k. Use DCA as subroutine for better bounds. Lower bound: If (v i ) n+1 n+1 and. Solve x = i=1 α i v i are the vertices of i=1 S k min α g i α i v i α i h(v i )

41 Low rank optimization and polynomial guarantees[goyal and Ravi 08] A function f : R n R has rank k n if there exists g : R k R and α 1,...,α k R n such that f(x) =g(α 1 x,..., α k x) Most examples in Economy literature. For a quasi-concave function to solve min f(x). x C f we want Can always transform DC programs to this type of problem.

42 Algorithm Let g satisfy the following conditions. The gradient g(y) 0 g(λy) λ c g(y) for all λ>1 and some c α i x>0 for all x P There is an algorithm that finds with f(x) (1 + )f(x. ) in O c k k x

43 Further reading Farkas type results and duality for DC programs with convex constraints. [Dinh et. al 13] On DC functions and mappings [Duda et. al 01]

44 Open problems Local rate of convergence for constrained DC programs. Is there a condition under which DCA finds global optima. For instance g h might not be convex but h g might. Finite convergence of cutting plane methods.

45 References Dinh, Nguyen, Guy Vallet, and T. T. A. Nghia. "Farkas-type results and duality for DC programs with convex constraints." (2006). Goyal, Vineet, and R. Ravi. "An FPTAS for minimizing a class of lowrank quasi-concave functions over a convex set." Operations Research Letters 41.2 (2013): Tao, Pham Dinh, and Le Thi Hoai An. "Convex analysis approach to DC programming: theory, algorithms and applications." Acta Mathematica Vietnamica 22.1 (1997): Horst, R., and Ng V. Thoai. "DC programming: overview." Journal of Optimization Theory and Applications (1999): Tuy, H. "On global optimality conditions and cutting plane algorithms." Journal of optimization theory and applications (2003): Yuille, Alan L., and Anand Rangarajan. "The concave-convex procedure."neural Computation 15.4 (2003):

46 References Salakhutdinov, Ruslan, Sam Roweis, and Zoubin Ghahramani. "On the convergence of bound optimization algorithms." Proceedings of the Nineteenth conference on Uncertainty in Artificial Intelligence. Morgan Kaufmann Publishers Inc., Hartman, Philip. "On functions representable as a difference of convex functions." Pacific J. Math 9.3 (1959): Hiriart-Urruty, J-B. "Generalized Differentiability/Duality and Optimization for Problems Dealing with Differences of Convex Functions." Convexity and duality in optimization. Springer Berlin Heidelberg,