1 / 25

Constraint Satisfaction and Property Testing

Constraint Satisfaction and Property Testing. (and Voting and Double Bubbles). Ryan O'Donnell. Microsoft Theory Group. 7. 8. 12. The Max-Cut problem:. Election recounts? Majority vs. Electoral College?. The 2-Lin(mod 3) problem:. Simultaneously satisfy as many as you can.

lis
Download Presentation

Constraint Satisfaction and Property Testing

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. Constraint Satisfaction and Property Testing (and Voting and Double Bubbles) Ryan O'Donnell Microsoft Theory Group

  2. 7 8 12 TheMax-Cut problem: Election recounts? Majorityvs.Electoral College?

  3. The2-Lin(mod 3) problem: Simultaneously satisfy as many as you can.

  4. 2-Variable Constraint Satisfaction Problems (“2-CSPs”) Variables: x1, x2, x3, … , xn Label Set:  (= allowed values for the variables) Input: Constraints 1, 2, …, Non pairs of variables. Output:Assignment satisfying as many constrs. as possible. “Optimization”, “Approximation Algorithms”

  5. x1 x3 x2 x4 x5 x6 x7 x8 x9 x10 Graph version = { , }               

  6. 2-CSP examples Max-Cut: = {0,1},  ’s of the form “ xi xj ” 2-Lin(mod 3):  = 3,  ’s of the form “xi = xj + c” 2-SAT:  = {0,1},  ’s are Vertex-Cover:Input: A graph.Goal: Select as few vertices as possible s.t. all edges are “covered”. Coloring 3-colorable graphs:Input: a 3-colorable graph. Goal: Legally color it using as few colors as possible. (running example) , , ,

  7. Story of the talk 1. Egg on our face re complexityof algorithms for 2–CSPs. 2. Efficient “Property Testing” algs.) Hardness for CSPs 3. Remarkablyefficient (2–query!) Property Testing algs. exist.

  8. Complexity theory dictum “Essentially every natural algorithmic problem has been shown to be in P (polynomial time) or NP-hard.” (Exceptions: Factoring, Graph-Isomorphism.)

  9. This is a lie. • Given a graph, find a cut achieving at least 90% of the max cut. • Given a 3-colorable graph,color it using at most 100 colors. • Find a vertex cover at most 1.99 times the minimum. • Find a 2-SAT assignment satisfying 95% of the maximum. • Given (1−)-satisfiable 2-Lin(mod p) system, satisfy (1/p)/2 fraction. • Find a cut within factor log log log n of the sparsest cut. • (1 − 1/2k)-approximate Max-k-Cut • … • 1.49-approximate metric TSP • 1.54-approximate minimum Steiner tree • (.1 log n)-approximate asymm. metric TSP • 1.3-approximate minimum multiway cut • 1.51-approximate minimum uncapacitated metric facility location • (log n)-approximate bandwidth of graphs • (log n)1/3-approximating 0-Extension • .92-approximating MAX-E3-Set-Splitting • 1.5-approximate rectangle tiling • O(1)-approximate minimum linear arrangement • O(1)-approximate minimum feedback arc • .51-approximate satisfiable betweenness instances • 

  10. We gotta do something about this! 1. Prove problems are in P. Seems we need a radically new algorithmic idea. Max-Cut: Most recent working algorithmic idea was from ’89–’92… Goemans-Williamson ’94 proved it always finds a cut achieving ¸87 . 8567 % of the optimum. 2. Prove new NP-hardness results. Even after much effort, only some success. Not much for 2-CSPs.(“PCP Theorem” [AS’92, ALMSS’92] + “Parallel Repetition Theorem” [Raz’95])

  11. 2-CSPs? “Unique Games Conjecture” [Khot’02] [O’0?] Max-3-CSP [Khot’02] 2-Lin(mod 2) [ST’06] Max-k-CSP [KR’03] Vertex-Cover [KKMO’04] [KO’06] [DMR’06]([MOO’05]) [MOO’05] Max-Cut-Gain ( 87 . 8567 % ) Coloring 3-Colorable Graphs Max-Cut [KV’05] 2-Lin(mod p) Sparsest Cut 2-SAT A general theory is developing.

  12. Story of the talk 1. Egg on our face re complexityof algorithms for 2–CSPs. 2. Efficient “Property Testing” algs.)Hardness for CSPs 3. Remarkablyefficient (2–query!) Property Testing algs. exist.

  13. Property Testing = “Constant Time Algorithms” = “The art of uninformed decisions” Input: A “huge” object: e.g., truth tablef : m!. Output: YES or NO, depending on whether it has property P. Caveat:You want to answer in constant time. What you can do: • Read f(x) for a few random x, say f(x1), …, f(xk). • Apply a “test”,  (f(x1), …,f(xk))) YES / NO.

  14. Testing “Dictatorships” For CSP hardness reductions, relevant P is being a “Dictatorship”: f(x) = xi f a Dictatorship) test outputs YES with prob. ¸pYES f “very non-Dictatorial”) test outputs YES with prob. ·pNO

  15. Testing “Dictatorships” “ k-query, -based, (pYES , pNO) Dictatorship test ” (for an unknown f : m!) 1.x1 , x2, …, xkchosen at random (somehow) from m 2. (f(x1), f(x2), …,f(xk))is output, either YES or NO Requirement: f is a Dictatorship) Pr[output YES] ¸ pYES f “very non-Dictatorial” ) Pr[output YES] ·pNO

  16. CSP hardness Rule of Thumb NP-hardness (or “Unique Games Conjecture”) reduction for: “Satisfying a fraction of the optimum, given a k-CSP instance withconstraints.” “ k-query, -based, (pYES , pNO) Dictatorship test ”  pNO pYES Why? Remark: This idea is old: from [BGS’95]. Novelty: 2-query Dictatorship tests exist!

  17. Rule of Thumb example: Max-Cut Max-Cut: 2-CSP over {0,1} with constraints of form “ xixj ”. 2-query, “  ”-based, (90%, 80%) Dictator test for f : {0,1}m! {0,1} )“Assuming UGC, it is NP-hard to find cuts that achieve 88.888 % of the optimal cut.” 1. Pick x, y2 {0,1}m in some clever random way. 2. Query f(x), f(y) and output YES iff f(x)f(y). ) Dictatorships pass w.p. ¸90%, “Totally not Dictatorships”pass w.p. ·80%.

  18. Story of the talk 1. Egg on our face re complexityof algorithms for 2–CSPs. 2. Efficient “Property Testing” algs.) Hardness for CSPs 3. Remarkablyefficient (2–query!) Property Testing algs. exist.

  19. 2-query, “  ”-based Dictatorship test? f : {0,1}m! {0,1} Voting: 0 & 1 are two parties. m voters. f is voting rule. winner f m voters

  20. 2-query, “  ”-based Dictatorship test? [KKMO’04] suggestion: Election #1: Each voter flips a coin. Election #2: Each voter, with probability 90%, reverses their vote. Test: Winner #1  Winner #2. Prob[ Dictatorship passes ]: Prob[ Majoritypasses ]: Prob[ Electoral College passes ]: 90% ¼ 79.5% ¼ 70.1%

  21. Majority Is The Highest [KKMO’04] conjectured, [MOO’05] proved: “Majority is the non-Dictator passing the test with highest probability.” Hence: “ ”-based, ( , Hence: UGC-hardness of finding cut within ● [GW’94] is optimal Max Cut alg., assuming UGC ● Resolvesconjectures [Kalai’03,’04] in theory of voting, also problems [ADFS ’04] in combinatorics. ● Result can be used to prove (sometimes improve) essentially allknown UGC reductions. ) Dictatorship test. ) Dictatorship test. 90% 79.5 % 79.5% of Max Cut. ¼ 88.4 % 90% Consequences:

  22. 2-CSPs? “Unique Games Conjecture” [Khot’02] [O’0?] Max-3-CSP [Khot’02] 2-Lin(mod 2) [ST’06] Max-k-CSP [KR’03] Vertex-Cover [KKMO’04] [KO’06] [DMR’06]([MOO’05]) [MOO’05] Max-Cut-Gain ( 87 . 8567 % ) Coloring 3-Colorable Graphs Max-Cut [KV’05] 2-Lin(mod p) Sparsest Cut 2-SAT A general theory is developing. [MOO’05]

  23. The proof that Majority is the highest 1. Generalize Central Limit Theorem. “Sums of random 0’s and 1’s ! Gaussians.” “Polynomials of random 0’s and 1’s ! polynomials of Gaussians.” 2. m Gaussians is like uniform distribution on m-dim. sphere. Problem becomes a cut problem on the sphere.Specifically “Min-Bisection”. 3. For small noise params (angles), essentially similar to finding the blob of half-volume w/ smallest perimeter. (Connections to Double Bubble problem.)

  24. Open problems I’m thinking about 1. Prove Unique Games Conjecture. ([FKO]: trying to give reduction fromMax-Cut hardness.) 2. Analyze various other constant-query Dictatorship tests. 3. Change from the “Dictatorship testf : {0,1}m! {0,1}” paradigm. ([KO’06] has some partial work on this.)

More Related