1 / 31

Linear Programming

Linear Programming. (2) Combinatorial rounding Pipage rounding. Weighted Vertex Cover. V. LP-Relaxation. Rounding. A. Performance Ratio : 2 ( x i < 2 x i *). Half-integerality. Theorem. Proof. Min-2Sat.

bao
Download Presentation

Linear Programming

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Linear Programming (2) Combinatorial rounding Pipage rounding

  2. Weighted Vertex Cover V

  3. LP-Relaxation Rounding A Performance Ratio: 2 (xi< 2 xi*)

  4. Half-integerality Theorem Proof

  5. Min-2Sat • Given a 2-CNF Boolean formula, find a satisfying assignment (if exists) with the min # of true variables. • Vertex Cover (unweighted) is a subproblem of Min-2Sat: For every edge {vi, vj} of the vertex cover problem, define a clause (xi + xj) for the Min-2Sat problem.

  6. Integer Linear Program

  7. Relax to Linear Program Rounding

  8. What about xi* = ½ ? • For instance, what if the input has both (xi + xj) and (xi + xj)? • Note: If xi* = ½, then xj* > ½. So, after the initial rounding, only those clauses with xi* = xj* = ½ are left. • These clauses form a new formula for which any assignment has p.r. = 2. • There is a simple polynomial-time algorithm for 2Sat. - -

  9. Algorithm for 2-Sat • Convert the 2-Sat formula F with variables x1, … , xn to a digraph: V = {x1, x2, … , xn}, For each (yi + yj), generate two edges (yi, yj), (yj, yi). • If xi and xi are strongly connected, then F is not satisfiable. • O/w, if xixi, then assign all y reachable from xi as TRUE. • Likewise for the case xixi _ _ _ _ _

  10. Example _ _ x1 TRUE x2TRUE x3 TRUE x1TRUE _ x1TRUE

  11. Scheduling on Unrelated Parallel Machines Integer program

  12. LP-Relaxation

  13. Combinatorial Property of x* Bipartite graph H

  14. Rounding

  15. Prune t < T

  16. How to find T*?

  17. Proof of the Combinatorial Property Case 1: H is connected. Fact: H has at most m+n edges.

  18. Proof • x* has mn components • x* is an extreme point (a vertex), and so must satisfy exactly mn equations (active constraints). • For equations xij = 0 or xij = 1, xij must be integral. • At most m+n active constraints allow xij to be fractional. • So, H has at most m+nedges.

  19. Each job has degree > 2, and so must have a child.

  20. Case 2.H is not connected

  21. Max Weighted Hitting • Input: E ={1, 2, … , n}, C = {S1, S2, … , Sm}, Si subset of E,w(Si), p > 0. • Find A subset of E, |A|=p, with the max weight {w(B) | B subset of E, A hits B}.

  22. Associated Integer Program not linear

  23. Relaxation:

  24. Add new variables to make it linear

  25. Pipage Rounding IDEA of Rounding: 1) Choose two variables xjand xk and round at least one of them to 0 or 1. 2) Balance the total = p (e.g., if xjis rounded down to 0, then xk is rounded up by xj). 3) Make sure the new value of L(x) does not decrease ?? Use a companion function F(x) instead of L(x)

  26. Use an alternative integer program also not linear

  27. (over x {0,1}) Relation: F(x) = L(x) F(x) > (1 – 1/e) L(x) (over 0 < x < 1) Proof Assume | Sj| = k. geom-mean arith-mean ineq satisfies Note and f (1) > 1 – 1/e

  28. Pipage Rounding Algorithm

  29. Lemma Proof For a specific term about Sl, if it contains zero or one of xi and xk then it is a constant or linear wrt . If it contains both xiand xk, then it is of the form with a> 0, and so it is convex.

  30. So, In other words, F(x) is nondecreasing during the rounding processes. It means the final approx. sol x satisfies and p.r. is at most e/(e-1) ~1.58. A

More Related