Tolerant Locally Testable Codes Atri Rudra Qualifying Evaluation Project Presentation Advisor: Venkatesan Guruswami 1 Fake Motivation Elvis Presley is alive! Check DNA Verify this Too much work “Spot Check” Accept Elvis Reject Atri Bruce Campbell ? 2 Outline of the talk Real Motivation Testing Codes Previous work Our Contributions High Level ideas Some Details Open problems 3 Error Correcting Codes C(x) x Encoder y Tester Decoder x Hopeless C(x) x Give up 4 Property testing x Verify a property Oracle access to input Does x have the property ? Make few queries Probabilistic tester T Accepts correct inputs Rejects very bad inputs (whp) 0/1 5 Codes Mapping C : k!n Distance d = min u,v2 (C(u),C(v)) k (¢,¢) is Hamming Distance Rate k/n [n,k,d] d/2 d/2 d 6 Testing Codes x Property x 2? C Make few queries Probabilistic Tester How good is the tester ? Accept x 2 C w.p. 1 Reject x far from C w.p. 2/3 T Hamming Distance Local tester Constant number of queries Sub-linear also interesting 0 w.p. 1 2/3 7 Locally Testable Codes Who Cares ? Heart of PCPs Alternate Characterization of NP X 2? L Proof (x) Verifier checks (x) Makes q queries NP = PCP[ O(log n), O(1)] [ALMSS92]….. 8 Another motivation C(x) x y x Close Far Give up 9 Current Local Testers Reject if y is far Accept if y is close By defn accepts only y2 C Against rationale of codes Close y Far 10 Tolerant Local Testers Dist(y,C) <= c1d/n y Dist(y,C) > c2d/n Accept w.p >= 2/3 Tolerance Reject w.p. >= 2/3 Soundness Close Far q(n) queries (c1,c2,q)- testable Prev work (0,O(1),O(1))-testable Perfect completeness 11 The Holy Grail Constant rate, linear distance Constant Query Complexity Not known even for LTCs Unique decoding radius c1=1/2, c2 ¼ 1/2? d/2 d/2 d 12 Contributions LTCs ! tolerant LTCs Constant rate Sub-linear query complexity [BS04] Constant # queries No generic “complier” Slightly Sub-constant rate [BGHSV04] Constant c1, c2 13 More on Contributions (Constant # queries, Constant Rate) Near uniform queries Sub-constant Rate Partitioned queries Sub-linear # queries Goal: Design codes and tolerant testers 14 Where are we now ? Real Motivation Testing Codes Previous work Our Contributions High Level ideas Some Details Open problems 15 LTC ! tolerant LTC Perfect Completeness Uniform query pattern x c1= O(1/q) by union bound Almost uniform is q is not constant ? T 1 16 Local Tester Revisited x Decision procedure is strict Accept perturbations There is a problem Local View Locally approx correct ) Global approx correct Robustness [BS04] T 1 0 17 What is next ? Constant rate, linear distance Sub-linear query complexity Product of Codes [BS04] 18 Product of Codes C [n,k,d] C2 Any row 2 C Any Column 2 C [n2,k2,d2] Tester ? 2C n C3 n 19 Tester for 2 C row pick row or clm pick j2[n] Rj2 C ? Not known to be robust Big open question True for special cases C is Reed-Solomon C is C’2 n C3? n 20 Larger product of Codes 3 (C ) Similar definition (3D instead of 2D) Same test 2? C2 test Check all n2 pts N2/3 queries N=n3 2 2 2?2CC 2 C2 Robust! [BS04] 21 Formal definition of Robustness v2n r random coin T(v,r)=miny:T(y_r)=1 dist(v,y) T(v)=Er[T(v,r)] T is e-robust 8 v2n, dist(v,C)· e¢T(v) 22 3 C Tolerant test is tolerant LTC Restriction is close to C2? Constant rate N2/3 queries Reduce the # queries Ct (t-Dimension) N2/t queries ¼? C2 23 Tolerance of dist(v,C)· tester n3/3 h f2 C3 closest to v ¸ 2n/3 choices of h 3 C Dist(vh,fh)· n2 Averaging argument If not, for ¸ n/3 h, dist(vh,fh) > n2 ) dist(v,f)> n3/3 dist(vh,C2)·? n2 Similar arguments for other planes v accepted w.p. ¸ 2/3 24 So what do we have now ? Constant rate, linear distance Sublinear query complexity n # queries =2/t C has no local tester but Ct has one 25 What is next ? Slightly sub-constant rate, linear distance n=k¢ exp(logk) for any >0 Constant query complexity Based on PCPs [BGHSV04] 26 PCP of Proximity Variant of PCP introduced in [BGHSV04] CKT-VAL(T)={x:T(x)=1} Verifier VT such that ||=s¢ exp(logs) x2 CKT-VAL(T), 9 , VT(x,)=1 wp 1 x far from CKT-VAL(T), 8 , VT(x,)=1 wp <1/2 #queries in hx,i 9 8 x s=|T| Constant # queries T VT 1 0 27 Local Tester 1.0 Start with good code C0 Use PCPP as an aid Constant rate and linear distance Linear size encoding circuit C1(x)= hC0(x),(x)i There is a problem |x|/|(x)|=o(1) Distance of C1 is bad (x) x x (x) C0 1 0 28 Local Tester 1.1 Increase the “code” part C2(x)=h (C0(x))t,(x) i Choose t such that |(x)|/(t¢|x|)=o(1) Constant query complexity Slightly sub-constant rate, linear distance Not tolerant Just corrupt the proof part Corrupted word still close to C2 (C0(x))t (x) 29 Tolerant Local Tester 1.2 Keep the code and proof parts comparable C3(x)=h(C0(x))k,((x))li k¢|C0(x)|=(l¢|(x)|) Need near uniform queries Constant query complexity Slightly sub-constant rate, Linear distance Used in relaxed LDC in [BGHSV04] 30 To summarize Defined tolerant LTCs Explicit constructions Constant # queries, slightly sub-constant rate Sub-linear # queries, constant rate Both constructions start from some C0 C0 does not have a (tolerant) local tester 31 Open Questions Is “natural” tester for C2 robust ? e-robust for e=O(1) row No lower bounds on n for LTCs Does tolerance make lower bounds easier ? n C3? n 32 Questions ? 33