The Problem: The vertex cover problem VC: Min Sum wi xi s.t forall edge ij xi+xj>=1. Effective progress: To get 1-effective approximation you need tight constreint. E.g. Say you know that x1+x2+2x3=2 you can change the objective function from COST = w1x1+ w2x2+ w3x3+ w4x4+w5x5+...+wnxn to COST' = (w1-1)x1+(w2-1)x2+(w3-2)x3+w4x4+w5x5+...+wnxn Oviously COST-COST'=2 Notice that multiplying tight constreint is valid and therefore youcould have multiply it by Min(w1,w2,w3/2) and this whay one of the coaficions would vanist in the new objective function. Getting tight constreint: First atempt. Add a variable xij, wij. Change constreint to tight xi+xj+xij=2 Problems: if wij starts with 0, no progress can be made if wij>0 than if wij>=wi,wj than take all vertices. else after LR wij=0 and we have the same problem Second (and last) atempt: Add a variable xij, wij. Change constreint to tight xi+xj-xij=1 //interpetation xij=overcover we stard wij=0 and define COST = w1x1+...+wnxn + w12x12+...+wixj+.. Example: before: 7---0---3 After: 4--3--0 and new cost COST'=COST-(3x1+3x1-3x12)=COST-3 In general: if 12 is an edge and w1,w2>0 we select w'=Min{w1,w2} and w1-=w1; w2-=w'; w12+=w'; SO that COST'=COST-w1 After we done: Good news1: every edge has zero weight endpoint Good news2: Total COST is reduced. Bad news1: some edges has positive weight Bad news2: We are not sure if COST reduced to zerow. More tigh constreints: We new kind of tight constreint (TC). Idea: linear combination of TCs gives TC. Example1: 1--12--2--23--3 gives TC123=TC12-TC23 gives x1-x12+x23-x3=0 Good news: even if w2=0 we may use this TC. Bad news: This TC gives us not progress in COST. Example2: 1--12--2--23--3--34--4 gives TC1234=TC12-TC23+TC34 gives x1-x12+x23-x34+x4=1 Good news: If found make a progress in COST Bad news: We need +--?--?--+--?--?--+, what if not found? Generalization: Odd path: 1--12--2--23--3--34--4--...--2k gives TC1234..2k=TC12-TC23+TC34-T45+...+T2k-1_2k gives x1-x12+x23-x34+x4-...+x2k=1 Good news: If found make a progress in COST Bad news: We need +--?--?--+--?--?--+..., Alternation path. Q: is this sufficient? A: Almost. Just generalized not non simple paths. Claim: If no alternation exists then COST=0. Define A0 = all positive vertices A1 = N(A0), A2= N+(A1) ... i.e. A_odd = N(A_previous) A_even= N+(A_previous) Claim1: No edges in A2i Claim2: No positive edges in A2i+1 Define L= union of A2i M union of A2i+1 R all the rest Claims: No edges in L all edges in E(M)-E(L) are zeros; L={+...+++00...000} M={00...0000} R={00...000} E(L,L)={} W(E(M,M))=0 W(E(M,R))=0 The following is feasible fractional solution: xi=1 for all i in L xi=0 for all i in M; xij=0 for ij in MxM xi=1/2 for all i in R; xij=1/2 for ij in MxM xi=0 COST = 0; The following is partial feasible integral solution: xi=1 for all i in L xi=0 for all i in M; xij=0 for ij in MxM Proof: W(OPT(L,M))>=W(E(L,M)) and every integral solution C can be replaced C-L+M HALL: M in E is a perfect matching if G(M)=G B(X,Y,E) is perfect if it has a perfect matching X' in X is defect if |N(x)| < |X| B is defected if it contain a defect Hall: B is prefect <===> B is not defected. Proof: ===> is trivial <=== Run VC(0,1/2,1). No need for 1/2 and all wij is 0/1. if ++ then L is defect and therefore B is not perfect. Generalized: G, C Perfect if existt wij s.t. ci is homogenious w.r.t. w. X i.s. is defect if C(N(X)) < (C(X) G is defected if... Hall: B is prefect <===> B is not defected. ===> is trivial <==== Run VC(0,1/2,1) if ++ the L is defect.