Optimal Long Test with One Free Bit Nikhil Bansal (IBM) Subhash Khot (NYU) 1/17 Min Vertex Cover Vertex Cover: Given G=(V,E), a set S of vertices s.t. each edge has at least one end point in S. Vertex Cover Hardness: 1.36 assuming P NP [Dinur-Safra 02] 2 - assuming UGC [Khot-Regev 03] S is vertex cover iff V\S is independent set. [Khot-Regev 03]: Even if graph has independent set of size (1/2 - ) |V| cannot find one of size |V|. 2/17 Our results Thm: Assuming UGC, it is NP-Hard to get 2- approx. for vertex cover, even for essentially bipartite graphs. (even though min-VC is easy for bipartite graphs). Independent Independent (1/2-) n (1/2-) n Vertex Cover Equivalently, hard to find independent set of size n even if graph has two disjoint independent sets of size (½-) n 3/17 Additional Features 1. Previous results on VC use biased long code. (most naturally viewed as combinatorial constructions) 2. Our result is most naturally viewed as a PCP. (in fact, assuming UGC we show a PCP with 1 free bit, and completeness = 1- , soundness = ) 3. Very natural dictatorship test with simple analysis. 4. Unlike [Khot-Regev], do not need a (equivalent) version of UGC with special properties. 4/17 1|prec|j wjCj Problem Given n jobs, arbitrary weights and sizes (job j: wt. wj, size pj) Precedence constraints: DAG (edge (i,j) ) j cannot start before i finishes) A valid schedule Precedence Graph Goal: Schedule jobs to minimize weighted completion time. Various 2-approximations: Potts linear ordering formulation Completion time Formation Time indexed Formulation Sidney’s decomposition Since the 70’s and continuing until recently 5/17 Scheduling Problem No (1+) approximation [Ambuhl-Mastrolilli-Svensson 07] (assuming SAT cannot be solved in sub-exponential time) Special case of Vertex Cover [Chudak-Hochbaum 99, Correa-Schulz 05, …] Belief that < 2 might be possible. Thm: Hardness of 2- assuming a variant of UGC. Our dictatorship test inspired by certain hard instances proposed by Woeginger [W 03] (do not discuss this connection in talk) 6/17 Outline • • • • Introduction Free Bit Complexity of PCPs Background (UGC, Influences, …) Dictatorship Test & Proof 7/17 PCP Theorem [AS, ALMSS] X 2 SAT can be verified by writing proof of length poly(n) And querying only O(1) positions in the proof. PCP Thm: NP ½ PCPc,s(O(log n), q=O(1)) c: Completeness, prob. of accepting correct proof s: Soundness, prob. of accepting wrong proof q: Number of queries Another parameter: number of free bits. 8/17 Free bit Complexity f = log2 (number of accepting configurations for query) Eg: Hastad’s test: Accept if x © y © z = 0 3 queries, but accepts on 4 answers: (0,0,0), (0,1,1), (1,0,1), (1,1,0) Free bits (f) = log2 4 =2 Thm [FGLSS, BGS]: PCPc,s with f=0 is equivalent to Reduction: 3-SAT to graph on V vertices, s.t. NP-Hard to tell if independent Set has size at least c |V| or at most s |V|. (f=0 means verifier expects unique answer for each query.) Corollary: Implies (1-s)/(1-c) hardness for vertex cover. 9/17 Our Result Thm: Assuming UGC, there is a PCP with 1 free bit, completeness = 1-, soundness = , Cor: PCP with 0 free bits, completeness = ½ - , soundness = Pf: Take PCP with 1 free bit (has 2 good answers per query), verifier can choose one of these answers randomly. (1-s)/(1-c) = 2- hardness for V.C. (+ almost-bipartite property) Using usual UGC techniques, suffices to give a related dictatorship test on the boolean hypercube. (Dictatorship test with 1 free bit, completeness c and soundness s will translate to PCP with same properties) 10/17 Dictatorship tests (1,1,0) (1,0,0) x1 (1,1,1) (1,0,1) (0,1,0) (0,1,1) x2 x3 (0,0,0) Vertices are x = (x1,…,xn) 2 {0,1}n Labeling: f ! {0,1}n ! {0,1} (0,0,1) Dictator (co-ordinate) labeling: f(x) = xi 0 1 0 0 1 0 1 Dictatorship Test: 1. Completeness: Accept any dictator labeling with prob ¸ c 2. Soundness: Accept any function “far” from dictator with prob. · s Far from dictator = All variables have small degree-k influences 11/17 1 Influences 1 0 Infli(f) = Pr[ f(x) f(x© i) ] Infli(f) = S: i 2 S (S)2 0 0 0 1 1 1 i-th coordinate Eg: 1. Dictator f(x) = xi has influence 1 for co-ordinate i, 0 for others 2. Majority function f(x) has small influences ((1/n1/2)) Soundness: For any f s.t. Infli(f) · for all co-ordinates i. Test must fail (pass with prob. · s) Actually: Deg-k influence Inflki(f) = S:i2S, |S|·k (S)2 (for list-decoding purposes) 12/17 Dictatorship Test (1,1,0) Pick a random sub-cube on n co-ordinates f(x1, *, x3, *, …., xn) Accept if mono-chromatic (all 0’s or all 1’s). (1,0,0) (1,1,1) (1,0,1) (0,1,0) (0,0,0) (0,1,1) (0,0,1) Huge number of queries (2n) , but 1 free bit! Completeness: For any dictator function f(x1,…,xn) = xi Random sub-cube on n co-ordinates is mono-chromatic with probability 1- 1 0 0 0 1 0 1 1 13/17 Soundness Soundness: Low influences ) · fraction subcubes mono-chromatic (not true say if f=0 everywhere or f=1 everywhere) Folding Trick: Consider subcube C at x, and subcube at Accept if both monochromatic and have different colors. Soundness: If 1/3 · E[f] · 2/3, for any > 0, 9 k, = O(1) s.t. if Inflki(f) · for each i, then · fraction of subcubes monochromatic. Proof follows from [Mossel-O’donnell-Oleszkiewicz’05 ] Invariance Principle: Low deg-k influence ) f is “random-like” [MOO’05] Ain’t Over Till It’s Over (proposed by Freidgut-Kalai) 14/17 Alternate Proof Soundness: Small influences, then · fraction of subcubes monochromatic. Will Show: Random subcube contains a 1 with prob. ¸ 1- /2 (symmetric argument implies it has 0 also with prob. ¸ 1-/2) Hence non-monochromatic with prob. ¸ 1- Pf: Random subcube (x,S)= Pick random x, n random coordinates S Define fS(x) = max (y : yi = xi for all i 2 [n]\S } Subcube (x, S) contains a 1 iff fS(x) = 1 Claim: For random S, with prob. 1 - /10 E[fS] ¸ 1- /10 15/17 Proof (continued) To show: For 1-/10 fraction of S, Proof: Define f = f0 , f1, …., fk = fS For i =i1,…,ik 2 S, fj(x) = max ( fj-1(x) , fj-1(x © ij)) E[fj] = E[fj-1] + (fj-1)/2 E[fS] ¸ 1-/10 k = |S| = n 1 1 0 0 10 1 0 1 1 If total influence >> 1/, if choose fraction of co-ordinates, expect to add up 1 unit of influences. Freidgut, KKL: If f is balanced, and has all influences small (· ), then total influence is high ¸ (1/log()) (Technical issue: Only f had high sum of influences, not fj-1.) 16/17 Concluding Remarks Assuming UGC, min-VC is 2- hard even in almost bipartite graphs New: k- hardness of vertex cover in k-uniform hypergraphs that are almost k-partite. Implies optimum hardness for some scheduling problems. Could be useful in other contexts (such as coloring ?) Min-VC: 1.36 hardness still best assuming P NP 17/17 Thanks ! 18/17 Influences Infli(f) = Pr[ f(x) f(x© i) ] Infli(f) = S: i 2 S (S)2 Deg-k influence Inflki(f) = S: i 2 S, |S| · k (S)2 Crucial point: i Inflki(f) · k Far from dictator: For all i, Inflki(f) · If dictatorship test accepts with prob ¸ s + some i s.t. Inflki(f) ¸ Can list decode 19/17