Combinatorial Interpretations of Dual Fitting and Primal Fitting Ari Freund Cesarea Rothschild Institute, University of Haifa Dror Rawitz Department of Computer Science, Technion Approximation Using LP Duality Minimization problem LP-relaxation and dual: (P) min wT x s.t. Ax b x0 (D) max bT y s.t. AT x w y0 Find xZn and y such that wTx r · bTy wTx r · bTy r · Opt(P) r · Opt Question: How do we find such solutions? 2 Primal-Dual Schema x and y are constructed simultaneously In each iteration: y is updated such that relaxed dual complementary slackness conditions are satisfied Primal complementary slackness conditions are obeyed Used extensively in the last decade (e.g., [GW95,BT98]) 3 A Combinatorial Approach: The Local Ratio Technique Based on weight manipulation Primal-Dual Schema Local Ratio Technique [BR01] Dual update Weight subtraction Local Ratio Technique is more intuitive Breakthrough results were achieved due to local ratio (e.g., FVS [BBF99,BG96], Max [BBFNS01]) Conclusion: combinatorial approach is beneficial 4 Metric Uncapacitated Facility Location Problem (MUFL) Non-Standard Applications: 3-approximation algorithm that relaxes primal comp. slackness conditions [JV01] 1.861 and 1.61-approximation algorithms both using dual fitting [JMMSV03] Motivation: combinatorial interpretations of both non standard applications 5 Dual Fitting Construct an infeasible dual y and a feasible primal x such that wTx bTy Find r s.t. y/r is feasible wTx bTy = r · bT(y/r) r · Opt Problem: finding the smallest r s.t. (for all input instances) y/r is feasible. y y/8 6 This Work Two new approximation frameworks: Combinatorial Based on weight manipulation (in the spirit of local ratio) Framework 1st Dual Fitting 2nd Primal Fitting* Examples MUFL [JMMSV03] MUFL [JV01] Set Cover [Chv79] Disk Cover [Chu] * Defined in this paper 7 An Example: Set Cover Input: C = {S1,…,Sm}, Si U, w : C R+ Solution: C’ C s.t. S U Measure: w( S ) SC ' SC ' Algorithm Greedy: 1. While instance is not empty do: 2. k argmini{w(Si) / |Si|} 3. Add Sk to the solution 4. Remove the elements in Sk and discard empty sets Approximation ratio is H n n i 1 1 i 8 Combinatorial Interpretation Uses weight manipulation A new weight function: w$ = r · w Opt$ = r · Opt w(Solution) Opt$ Performance ratio r In this case r = Hn 9 Combinatorial Interpretation In each iteration: Uncovered elements issue checks Bookkeeping is performed by adjusting weights A weight function is subtracted from w (and from w$) A zero-weight set Sk is added to the solution Elements covered by Sk retract checks that were given to other sets $ Checks are not retracted with respect to w 10 Example u1 =0 =2 =2 =4 =6 u2 u3 u4 S1 w1=4 S2 w2=10 4 8 4 10 S3 w3=12 6 8 0 0 11 Analysis - w Consider an element u u is covered by S(u) in iteration j(u) u pays j u j 1 j for S(u) j (u ) w(Sol) = U uU j 1 j j j j 12 Analysis – $ w In the j’th iteration: Opt$ decreases by at least |Uj| · j (“Local Ratio” argument: one check from each element must be cached) Deletion of elements may further decrease Opt$ Also, Opt$ = 0 at termination $ U Opt r Opt j j j Solution is r-approximate Assumption: w$ 0 throughout execution 13 Analysis (same as [JMMSV03]) Problem: find min{r | w$ 0 at all times} zi - amount paid by ui For fixed d and w(S) 1 d max zi w( S ) i 1 s.t. zi zi 1 i d i 1zi w( S ) i zi 0 i d 1 zi d 1i 1 zi H d w( S ) i 1 u z1 1 S u2 z2 u3 z3 . . . ud zd Approx ratio Hn 14 Combinatorial Interpretation of Dual Fitting Dual Fitting Combinatorial Framework Increasing a dual variable Dividing y by r (inflated dual) Subtracting a weight function Defining w$ = r · w (inflated weights) wTx bTy wTx Opt$ Find r s.t. y/r is feasible Find r s.t. w$ 0 Remark: problem of finding the best r can be formulated using LP, but LP-theory is not used in its solution. 15 Primal Fitting Construct an infeasible primal solution x and a (feasible) dual solution y such that wTx bTy The primal solution is non integral r s.t. r ·x is a feasible integral solution wT(r ·x)= r · wTx r · bTy r · Opt Can be used to analyze: 3-approx algorithm for MUFL [JV01] 9-approx algorithm for a disk cover [Chu] Both were originally designed using primal-dual 16 Combinatorial Interpretation of Primal Fitting Primal Fitting Combinatorial Framework Increasing a dual variable Subtracting a weight function Multiplying x by r (deflated primal) Defining w$ = w/r (deflated weights) wTx bTy (w$)T(r ·x) Opt Value of primal variable Fraction of weight left 17 The End 18