Property Testing and Affine Invariance Part I Madhu Sudan Harvard University

Size: px

Start display at page:

Download "Property Testing and Affine Invariance Part I Madhu Sudan Harvard University"

Cory McKenzie
5 years ago
Views:

1 Property Testing and Affine Invariance Part I Madhu Sudan Harvard University December 29-30, 2015 IITB: Property Testing & Affine Invariance 1 of 31

2 Goals of these talks Part I Introduce Property Testing (broadly interesting) Philosophy behind Invariance Introduce Algebraic Property Testing Affine-Invariance Part II Structural results about Affine-Invariance Testing & Affine-Invariance December 29-30, 2015 IITB: Property Testing & Affine Invariance 2 of 31

3 Property Testing Broadly: Test if massive data has some global property approximately, quickly. E.g.: 16 th /17 th century astronomy: Do planetary positions have geometric structure? Formalization: Data = Function ff: DD RR (DD, RR finite) Property = PP ff: DD RR Approximately? δδ(ff, PP) small, where δδ ff, PP = min { δδ ff, gg Pr ff xx gg xx } gg PP xx DD Quickly? With l DD queries into ff December 29-30, 2015 IITB: Property Testing & Affine Invariance 3 of 31

4 Ancient Example: Majority Is Majority of Population (roughly) Blue/Red? DD = Population; RR = {Blue, Red} ff = Current preferences PP = gg: Pr xx gg xx = BBBBBBBB 1 2 Test: Pick xx 1,, xx l DD uniformly independently. Accept if more than 1 2 εε 2 Theorem: l vote Blue. ff PP Accept w.p. 1 exp εε 2 l δδ ff, PP εε Accept w.p. exp εε 2 l Emphasis: l independent of DD ; Error acceptable December 29-30, 2015 IITB: Property Testing & Affine Invariance 4 of 31

5 Less Ancient Example: Linearity Is ff xx 1,, xx mm aa ii xx ii ii for some aa 1,, aa nn Abstraction: DD = GG, RR = HH; GG, HH finite groups PP = φφ: GG HH xx, yy φφ xx + yy = φφ xx + φφ(yy)} Test: Pick xx, yy GG uniformly & independently Accept if ff xx + yy = ff xx + ff(yy) Analysis [Blum,Luby,Rubinfeld 90]: ff PP Accept w.p. 1 (by definition) δδ ff, PP δδ Reject w.p. 2 9 δδ (non-trivial) December 29-30, 2015 IITB: Property Testing & Affine Invariance 5 of 31

6 Non-triviality? Example: nn = 3 tt ; GG = Z 3nn ; HH = Z nn ; PP = {xx aaaa mod nn } ; Consider ff xx = xx 3 δδ ff, PP = 1 1 nn Pr AAAAAAAAAAAAAAAAAAAA = 7 9 (Reject iff xx mmmmmm 3 = yy mmmmmm 3 {+1, 1}) Reason for non-triviality: Gap between f usually satisfies P and f usually equals g which always satisfies P Gap invisible in Polling ; gaping in linearity December 29-30, 2015 IITB: Property Testing & Affine Invariance 6 of 31

7 Example 3: Low-degree testing Is ff xx 1,, xx mm gg xx 1,, xx mm with deg gg dd? DD = FF qq mm ; RR = FF qq Test: Is deg ff llllllll dd? (More generally: Is deg ff AA subspace AA?) Locality l = qq vs. DD = qq mm (Example) Analyses: dd for affine αα > 0 s. t. mm, qq, dd qq 2, Pr llllllll RRRRRRRRRRRRRRRRRR ff αα δδ(ff, PP dd) Robust version: ββ > 0 s. t. mm, qq, dd qq 2, EE llllllll δδ(ff llllllll, PP dd ) ββ δδ(ff, PP dd ) December 29-30, 2015 IITB: Property Testing & Affine Invariance 7 of 31

8 Aside: Importance of Low-degree Testing Central element in PCPs (Probabilistically Checkable Proofs). Till [Dinur 06] no proof without (robust) low-degree testing. Since: Best proofs (smallest, tightest parameters etc.) rely on improvements to low-degree tests. Connected to Gowers Norms: [Viola-Wigderson 07]: [AKKLR] Hardness Amplification Yield Locally Testable Codes Best in high-rate regime. [BarakGopalanHåstadMekaRaghavendraSteurer 12]: [BKSSZ 11] Small-set expanders. December 29-30, 2015 IITB: Property Testing & Affine Invariance 8 of 31

9 History of Property Testing (slightly abbreviated) [Blum,Luby,Rubinfeld S 90] Linearity + application to program testing [Babai,Fortnow,Lund F 90] Multilinearity + application to PCPs (MIP). [Rubinfeld+S.] Low-degree testing + Definition [Goldreich,Goldwasser,Ron] Graph property testing + systematic study Since then many developments More graph properties, statistical properties, matrix properties, properties of Boolean functions More algebraic properties December 29-30, 2015 IITB: Property Testing & Affine Invariance 9 of 31

10 What is Property Testing? Algebra Graphs + Regularity Matrices + Linear algebra Statistics + CLT December 29-30, 2015 IITB: Property Testing & Affine Invariance 10 of 31

11 Invariance? Property PP ff: DD RR Property PP invariant under 1-1 ππ: DD DD, if ff PP ff ππ PP Property PP invariant under group GG if ππ GG PP is invariant under ππ. GG is invariance class of PP. Main Observation: Different property tests unified/separated by invariance class. December 29-30, 2015 IITB: Property Testing & Affine Invariance 11 of 31

12 Invariances (contd.) Some examples: Classical statistics: Invariant under all permutations SS DD. Graph properties: Invariant under vertex renaming. Boolean properties: Invariant under variable renaming. Matrix properties: Invariant under mult. by invertible matrix. Algebraic Properties =? Some introspection: Classical statistics only dealt with SS DD Different invariances different techniques. Invariance for algebra? December 29-30, 2015 IITB: Property Testing & Affine Invariance 12 of 31

13 What is Property Testing? Algebra=? SS VVVVVVVVVVVVVVVV SS vvvvvvvvvvvvvvvvvv? SS DDoooooooooo December 29-30, 2015 IITB: Property Testing & Affine Invariance 13 of 31

14 Algebraic Property Testing Property = algebraic Linearity Property (esp. GG = FF mm q ; HH = FF qq ) Low-degree Property. Is there anything else? What is the abstraction? December 29-30, 2015 IITB: Property Testing & Affine Invariance 14 of 31

15 Abstracting algebraic properties [Kaufman+S. 08] Affine Invariance: Range = Small Field FF qq mm Domain = Vector space over extension field FF qq nn Property invariant under affine transformations of domain (xx AA xx + bb) Additional feature: Linearity of Properties: Property = vector space over range. Critical in use in Coding Theory/PCP. December 29-30, 2015 IITB: Property Testing & Affine Invariance 15 of 31

16 Testing Linear Properties R is a field F; P is linear! Universe: {f:d R} P P Must reject Algebraic Property = Code! (usually) December 29-30, 2015 Must accept Don t care IITB: Property Testing & Affine Invariance 16 of 31

17 Connection to Coding Theory Algebraic properties lead to error-correcting codes Low-degree polynomials can t intersect often. True for other classes of algebraic functions. BCH codes, Dual BCH codes (same symmetries) AG/Goppa codes (fewer symmetries) Coding theoretic metrics want Property with: Large pairwise distance Classical Many members ( high rate ) Small locality of tests. New! December 29-30, 2015 IITB: Property Testing & Affine Invariance 17 of 31

18 Why Study Affine-Invariance Unify known testability results? [Kaufman+S 08]: Unified [BLR 90], [RS 92], [AKKLR 03], [JPRZ 04], [KR 04]. [Haramaty+RonZewi+S 13]: extends [BKSSZ 10], [HSS 11] [Guo+Haramaty+S 14]: strengthens robust low-degree tests [ALMSS 92,Raz-Safra 96] What leads to testability? Negative results: Counterexamples to AKKLR conjecture Positive results: Restrictions within Affine-invariance. New Codes and Implications: Lifted codes December 29-30, 2015 IITB: Property Testing & Affine Invariance 18 of 31

19 Rest of talk (including tomorrow) AKKLR Conjecture Motivation, Counterexample, Lessons Lifted Codes Intriguing generalization of polynomials! Ideas behind analyses of local tests Role of the tensor product December 29-30, 2015 IITB: Property Testing & Affine Invariance 19 of 31

20 AKKLR Conjecture [Alon,Kaufman,Krivelevich,Litsyn,Ron 03] Extended low-degree testing to case of dd qq. Proof extended that of BLR. Conjectured that testing should apply to symmetric codes with local constraints. Symmetric =? 2-transitive invariance class 2-transitivity supports local decoding Constraints =? December 29-30, 2015 IITB: Property Testing & Affine Invariance 20 of 31

21 Constraints, Characterization, Testing Testing Constraints Example: Can not test degree dd polynomials with locality l dd + 1 No local constraints! SS FF qq, SS dd + 1, pp SS pp ff SS ff, where pp = random deg. dd poly, ff = random function. Constraint= SS, VV : SS DD, SS l; VV h: SS RR. Is ff SS VV? Testing Characterizations Characterization= CC 1,, CC MM ; CC jj =constraint. ff PP jj, ff satisfies CC jj December 29-30, 2015 IITB: Property Testing & Affine Invariance 21 of 31

22 Constraints/Characterizations suffice? [Ben-Sasson,Harsha,Raskhodnikova 04]: No! even characterizations don t. AKKLR: Perhaps symmetry suffices? Strong form: Constraint + 2-transitivity suffices Does above imply characterization? Weak form: Characterization + 2-transitivity Both forms false: [Grigorescu,Kaufman,S 08]: Constraint + 2-transitivity Characterization [Ben-Sasson,Maatouk,Shpilka,S 11]: Characterization + 2-transitivity Testing December 29-30, 2015 IITB: Property Testing & Affine Invariance 22 of 31

23 Structure of Affine-Invariant Properties PP gg: FF QQ mm FF qq ; QQ = qq nn ; PP linear, affine-invariant. TTTT xx xx + xx qq + + xx qqnn 1. DD = Deg PP Monomials xx 1,, xx mm s.t. PP = TTTT cc MM MM MM DD Closure properties of the degree set Deg(PP) : xx 1 ii xx 2 jj MM Deg PP xx 1 ii+jj MM (mod xx 1 QQ xx 1 ) Deg(PP); xx 1 ii MM Deg PP xx 1 qqqq MM mod xx 1 QQ xx 1 Deg(PP) xx 1 ii MM Deg(PP) & jj pp ii xx 1 jj MM, xx 1 jj xx 2 ii jj MM Deg PP jj pp ii jj tt ii tt tt, jj = jj tt pp tt ; ii = ii tt pp tt tt ; qq = pp ss Any set closed wrt all three above is a degree set December 29-30, 2015 IITB: Property Testing & Affine Invariance 23 of 31

24 Known Testable Properties Focus on univariate properties PP ff: FF qq nn FF qq Basic locally testable univ. properties Reed-Muller Deg(RRRR ww ) = xx dd dd = tt dd tt pp tt ; tt dd tt ww} (locality l 2 ww+1 ) Sparse: Deg PP tt ; locality l l tt nn, qq [KaufmanLitsyn,GrigorescuKS,KLovett,BenSassonRonZewiS] Operations: Let PP 1 be l 1 -locally testable and PP 2 be l 2 -locally testable; l = l l 1, l 2, qq s.t. PP 1 PP 2, and PP 1 + PP 2 are l-locally testable. [BGMatoukShpilkaS,GuoS] and one more operation (to come later) December 29-30, 2015 IITB: Property Testing & Affine Invariance 24 of 31

25 Constraint + 2-transitivity Characterization Counterexample univariate wlog. Idea: Remove basis elements from RRMM 2 so resulting property is not Reed-Muller or sparse, but satisfies closure. Specifically Deg PP = xx 2ii +2 jj ii jj nn 3 xx2ii ii xx 0 Thm: For PP as above, l = Ω(nn) Key Lemma:αα 1,, αα kk FF 2 nn lin. ind. over FF 2, 2 αα αα 1 kk is non-singular 2 αα kk 1 2 αα kk kk December 29-30, 2015 IITB: Property Testing & Affine Invariance 25 of 31

Towards Counterexample to weak form Idea: Start with PP 1,, PP kk : l-locally testable properties and let PP = ii PP ii By construction PP is l-locally characterized Hope: locality of testing as kk

26 Towards Counterexample to weak form Idea: Start with PP 1,, PP kk : l-locally testable properties and let PP = ii PP ii By construction PP is l-locally characterized Hope: locality of testing as kk Unfortunately: Sparse Anything = Sparse RM RM = RM (RM+Sparse) (RM+Sparse) = RM+Sparse Need non RM+Sparse locally testable property. Idea lift sparse properties to non-sparse ones! December 29-30, 2015 IITB: Property Testing & Affine Invariance 26 of 31

27 Lifting Base Code BB bb: FF QQ FF qq affine-invariant Lifted Code LL mm (BB) = ff: FF QQ mm FF qq llllllll ff llllllll BB ; LL mm (BB) gg: FF QQ mm FF qq gg: FF QQ mm FF qq Lift of Sparse Sparse ; Lift of RM RM [BMSS] Use lifts and intersections to show Characterization + 2-transitivity Testing December 29-30, 2015 IITB: Property Testing & Affine Invariance 27 of 31

28 Characterizations + 2-transitivity Testable PP ff: FF 2 nn FF 2 ; nn = nn 1 nn kk ; nn ii distinct primes BB ii bb: FF 2 nn ii FF 2 ; Deg BB ii = xx 2jj +2 jj+1 jj 1, xx, xx 2, xx 4 PP ii = LL nn BB ii ff: FF 2 nn FF 2 ; PP = ii PP ii nn ii Lemma: Test-locality PP as kk Proof Steps: Let PP ii = jj ii PP jj ; Understand Deg(PP ii ) Find YY ii Deg(PP ii ) with nice recursive structure. Extract Matrices MM ii such that constraint lies in its kernel. Prove ker MM ii ker(mm ii 1 ) December 29-30, 2015 IITB: Property Testing & Affine Invariance 28 of 31

29 Characterizing testability Conjecture: To be OO(1) locally testable, code must be obtained from RRRR, SSSSSSSSSSSS by finite #composition steps using LL mm, +, Conjecture implies: tt kk prime nn, SS 1,, nn, αα 1, αα 2,, αα kk FF 2 nn lin. ind. 2 the matrix MM = αα ss ii has rank tt ii kk,ss SS Implication Open! In general, few techniques to lower bound rank of matrix over finite fields December 29-30, 2015 IITB: Property Testing & Affine Invariance 29 of 31

30 Next Lecture Nice Lifted Properties Surprising implications in incidence geometry Testability of Affine-invariant codes Some ideas December 29-30, 2015 IITB: Property Testing & Affine Invariance 30 of 31

31 Thank You December 29-30, 2015 IITB: Property Testing & Affine Invariance 31 of 31

Two Decades of Property Testing Madhu Sudan Harvard

Two Decades of Property Testing Madhu Sudan Harvard April 8, 2016 Two Decades of Property Testing 1 of 29 Kepler s Big Data Problem Tycho Brahe (~1550-1600): Wished to measure planetary motion accurately.