Dana Moshkovitz MIT
Take-home message: Prove hardness results assuming The Projection Games Conjecture (PGC)! Theorem: If SAT requires exponential time, and the PGC holds, then (1- )lnN approximating Set-Cover instances of size N requires time 2 N ( ) (tight; stronger than Feige’s).
AB AA BB e : A B An edge e=(a,b) E is satisfied by assignments f A :A A, f B :B B, if e (f A (a))=f B (b). PG: Given a game G=(G=(A,B,E), A, B, { e } e ), Find f A :A A, f B :B B maximizing P e (e satisfied).
There exists c>0, such that for any 1/n c, a Boolean formula , | |=n can be efficiently transformed to a projection game with graph size n 1+o(1) poly(1/ ) and alphabet size poly(1/ ): ◦ If is satisfiable, then there exist assignments to the projection game that satisfy all edges. ◦ If is not satisfiable, then any assignments satisfy at most fraction of the edges.
Parallel repetition [Raz94]: graph size n (log(1/ )) and alphabet size poly(1/ ). [M-Raz08]: graph size n 1+o(1) poly(1/ ) and alphabet size exp(1/ ).
1.Greedy: lnN-lnlnN+O(1)- approx in poly time [C79,S96] 2.Linear programming: lnN–approx; poly time [S99] 3.Sub-exponential: (1- )lnN-approx in 2 N O( ) -time [CKW09]
Who proved?Approx factorTime assuming ETH Comments Lund-Yannakakis, 93log N/42 2 logN (1) Bellare-Goldwasser- Lund-Russell 93 log N/C2 N (1/loglogN) Feige 96(1- )lnN2 N ( /loglogN) Raz-Safra/Arora- Sudan 97 log N/C2 N (1) Alon-M-Safra 06~0.22lnN2 N (1) Current work(1- )lnN2N()2N() Assuming PGC
All the results use the scheme of Lund- Yannakakis that has two components: 1. Projection game with =1/(logn) Combinatorial set-cover gadget. PCP time lower bound. Combinatorics approximation factor.
sub-universe per b B subset per a A, A For every e=(a,b) E, covers a subset of the sub-universe of b that is associated with e ( ). Combinatorial gadget: the only ways to cover a sub-universe: 1.Pick B and use subsets associated with . 2.Pick more than lnN subsets.
In the non-sat case, for any f A :A A, for (1- ) fraction of the b B, all b’s neighbors a A have different projections (a,b) (f A (a)) AB AA BB 1 4 2
We show how to transform any projection game to a Feige/rainbow game (without parallel repetition!) … Standard projection game: At most fraction of neighbors agree … Feige/rainbow game: All neighbors disagree
A bipartite graph (N,C,E); for every coloring of the N vertices, where no color set is of fraction > , at least (1- ) fraction of the C vertices have all their neighbors from different colors. The construction: incidence graph of lines and points NC
Prove hardness results assuming the projection games conjecture, e.g., for Clique, Chromatic Number, Shortest-Vector-Problem (SVP), Group-Steiner-Tree…