1 / 52

Direct Product : Decoding & Testing

Direct Product : Decoding & Testing. Russell Impagliazzo ( IAS & UCSD ) Ragesh Jaiswal ( Columbia U. ) Valentine Kabanets ( IAS & SFU ) Avi Wigderson ( IAS ) ( based on [IJKW’08, IKW’09] ). Average-Hardness Amplification.

hallam
Télécharger la présentation

Direct Product : Decoding & 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. Direct Product :Decoding & Testing Russell Impagliazzo ( IAS & UCSD ) RageshJaiswal ( Columbia U. ) Valentine Kabanets ( IAS & SFU ) AviWigderson ( IAS ) ( based on [IJKW’08, IKW’09] )

  2. Average-Hardness Amplification g f hard on  fraction of inputs hard on 1- fraction of inputs

  3. (Nonuniform) Hardness on Average {0,1}n {0,1}n f s {0,1}n 2n f is δ-hard (for size s), if every circuit (of size s)fails to compute f on δ inputs.

  4. Amplification via Repetition Intuition If on a random x one can compute f(x) on < ( 1- ) fractionof inputs, then on k independent random (x1,…, xk), one can compute all ( f(x1),…, f(xk) ) on < ( 1- )k exp(- k) fractionof inputs.

  5. Direct-Product (DP) Function For f: U  R, its k-wise DP function is fk: UkRkwhere: fk ( x1, …, xk) = ( f(x1), …, f(xk) )

  6. Direct Product Theorem [Yao’82, Levin’87, GNW’95, Imp’95, IW’97,…] Then: fkis exp(-k)- hard ( for size s * poly(,) ) {0,1}nk {0,1}n 2n 2n If: f is - hard ( for size s )

  7. Direct Product as Error-Correcting Code

  8. DP Encoding [Impagliazzo’02, Trevisan’03] U Uk fk f 011 010 0 110 1 1 0 1 0 010

  9. DP Code Parameters U Uk • Local encoding • Local approximate decoding fk f • Poor distance… • Distance amplification: • “far-away” messages are mapped to • “farther-away” codewords • Superpoly rate… • “Derandomized” DP Code with poly rate.

  10. Direct-Product Code:Decoding

  11. DP Theorem: Constructive Proof[Yao’82, Levin’87, GNW’95, Imp’95, IW’97,…] Construct: circuitC’ ( of size s ) ( 1- )-computing f. {0,1}nk {0,1}n 2n   2n Given:circuitC ( of size s*poly(,) ) exp(-k)-computingfk

  12. List-Decoding Lower Bound Theorem: To decode from agreement , need the list size (1/). Proof: Pick L = 1/functions f1, …, fL . Partition inputs into L blocks of size  each. Define C to agree with fik on block i.

  13. DP Decoding: Previous Work X X r1 rk C’ • [GNW, IW97,…]: List-size > exp(1/). • [IJK06]: poly(1/) list size for “large” , but still sub-optimal & complicated. b1 b bk if “enough” bi = f(ri), then output b b

  14. New Decoding Algorithm [IJKW 08] • Features: • list-sizeO(1/)( tight ! ) • simple algorithm (and analysis) • generalizes to Derandomized DP Code

  15. Given C that -computes fk, for  > exp(- k) x On input x, Pick random k-set (A,B2) containing x. B1 B2 Pick random k-set A If C is consistent( C(B1,A)|A = C(A,B2)|A ), output C(A,B2)|x. Else re-sample B2( O((1/) log 1/) times ). Randomly partition: |A|=|B1|=k/2 Freeze these random choices

  16. Given C that -computes fk, for  > exp(- k) AlgoA,B1 (x): On input x, Pick random k-set (A,B2) containing x. Preprocessing Pick random k-set If C is consistent( C(B1,A)|A = C(A,B2)|A ), output C(A,B2)|x. Else re-sample B2( O((1/) log 1/) times ). Randomly partition: |A|=|B1|=k/2

  17. Main Theorem: DP Decoding AlgoA,B1 (x): Preprocessing On input x, Pick random k-set (A,B2) containing x. Pick random k-set Randomly partition: |A|=|B1|=k/2 If C is consistent( C(B1,A)|A = C(A,B2)|A ), output C(A,B2)|x. Else re-sample B2( O((1/) log 1/) times ). Theorem:With probability  (²) over (B1,A), the resulting circuit Algo(1- )-computes f.

  18. Proof Ideas

  19. Flowers, cores, petals Flower:determined by S=(A,B) Core:A Core values:α=C(A,B)A Consistent petals: { (A,B’) | C(A,B’)A = α} [IJKW08]: Flower analysis B1 B B2 A A B3 B5 B4

  20. Structure (Decoding) Assume:C²-agrees with fk • Then: • There are many (²/2) • flowers determined by S=(A,B) • that are nice. • A flower is nice if it has • correct core ( C(S) = fk (S) ), • many (/2) consistent petals. • Also: In a random nice flower, • almost all consistent petals are mostly correct (C ¼fk ) B1 B B2 A A B3 Consistency  Correctness B5 B4

  21. Correctness of Decoding Preprocessing likely picks a nice flower • There are many (²/4) • flowers determined by S=(A,B) • that have: • correct core ( C(S) = fk (S) ), • many (/4) consistent petals, • almost all consistent petals are mostly correct (C ¼fk ) AlgoA,B likely does not time-out AlgoA,B(x) likely equals f(x)

  22. Proof of DP Structure: Averaging & Symmetry arguments Assume:C²-agrees with fk • Then: • There are many (²/2) • flowers determined by S=(A,B) • that are nice. • A flower is nice if it has • correct core ( C(S) = fk (S) ), • many (/2) consistent petals. • Also: In a random nice flower, • almost all consistent petals are mostly correct (C ¼fk ) B1 B B2 A A Averaging B3 B5 Symmetry B4

  23. Proof of DP Structure: Many nice flowers (Averaging) PrA,B [ ( (A,B) correct ) & ( A has/2 correct extensions (A,B’) ) ] = PrA,B [(A,B) correct ] - PrA,B [( (A,B) correct ) & ( A has</2 correct extensions B’ ) ]

  24. Proof of DP Structure: Many nice flowers (Averaging) PrA,B [ ( (A,B) correct ) & ( A has/2 correct extensions (A,B’) ) ]  PrA,B [(A,B) correct ] - PrA,B [( (A,B) correct ) | ( A has</2 correct extensions B’ ) ]  -/2 = /2.

  25. Proof of DP Structure: Consistency implies correctness (Symmetry) Idea: A highly incorrect set S’ can’t be a consistentpetal in a random flower with correct core B S’ A

  26. Proof of DP Structure: Consistency implies correctness Idea: A highly incorrect set S’ can’t be a consistentpetal in a random flower with correct core B S’ A

  27. Proof of DP Structure: Consistency implies correctness Idea: A highly incorrect set S’ can’t be a consistentpetal in a random flower with correct core. • f(A) = C(B,A)A & • C(B,A)A = C(S’)A & • C(S’)A  f(A). • Contradiction ! B S’ A

  28. Direct-Product Testing

  29. Testing C : UkRk IsC = fk, for some f : U  R? Fact:C = fkiff for all pairs of intersectingk-sets (S,S’), withA=SS’, C(S)|A = C(S’)|A

  30. LocalTesting: V-Test[GoldreichSafra, DinurReingold] S S’ C : UkRk A Test for one random pair of intersectingk-sets (S,S’), withA=SS’, if C(S)|A = C(S’)|A

  31. DP Testing: More formally … • (2’) Pr [Testaccepts C ] >  • Cfk on > () of inputs. • - Minimize #queries ( 2 ? 3 ? ) • Analyze small  (  < 1/k ?  < exp(-k) ? ) OnC : UkRkTest makes few queries, and (1) Accepts if C = fk. (2) Rejects if C is “far away” from any fk

  32. DP Testing History * Given C : UkRk, is C =gk? #queries acc.prob. Goldreich-Safra 0020 .99 Dinur-Reingold 06 2 .99 Dinur-Goldenberg 08 2 1/kα Dinur-Goldenberg 08 need > 2 1/k IKW 093 exp(-kα) IKW 09*2 1/kα * Derandomization

  33. Analysis of V-Test

  34. V-Test[GS00,FK00,DR06,DG08] Randomly pick two k-setsS1 =(B1,A) andS2 =(A,B2) (with|A| = k1/2 ). B1 B2 S1 S2 A Accept if C( S1)A = C( S2)A

  35. Flowers, cores, petals Flower:determined by S=(A,B) Core:A Core values:α=C(A,B)A Consistent petals: { (A,B’) | C(A,B’)A = α} B1 B B2 A A B3 B5 B4

  36. Structure (Testing) Assume: V-Test accepts with prob² There are many(²/2) flowers determined by S=(A,B) such that: There is g : U  R so that on almost all consistent petals Bi , C (Bi) gk (Bi). B1 B B2 A A B3 B5 B4 “Locally” C is DP

  37. Harmonious Flowers C(A, B1 )E¼ C(A, B2 )E, with |E| = |A| Assume: V-Test accepts with prob² • There are many(²/2) harmonious flowers determined by S=(A,B). • Harmonious flower: • many (/2) consistent petals, • on consistent petals, V-test accepts almost certainly ( 1-poly(²) ). B1 B B2 E A A B3 B5 B4 Proof by symmetry arguments (as in Decoding)

  38. Harmony  DP structure • Harmonious flower: • many (/2) consistent petals, • on consistent petals, V-test accepts almost certainly ( 1-poly(²) ). B1 B B2 A A Main Lemma: Define g(x) = PLURALITY { C( S’ )x| consistent petals S’ , x S’ }. Then C(S’) ¼gk (S’) for almost all (1-poly(²)) consistent petals S’. B3 B5 B4

  39. Proof Sketch of Main Lemma Assume otherwise. A random B1inConshas many “minority” elements xwhere C(B1)x g(x). A random E ½ B1has many “minority” elements [Chernoff] A random B2=(E,D2) is likely s.t. C(B2)E¼ g(E) [def of g] Then C(B1)E C(B2)E, Hence no harmony ! B1 D1 B D2 E A B2

  40. Decoding vs. Testing

  41. Decoding Testing C ²-computes fk V-Test²-accepts C • There are manynice flowers with: • correct core, • many consistent petals. • There are many harmonious flowers with: • many consistent petals, • restricted to consistent petals, V-Test accepts almost surely. Consistency  Correctness Harmony  DP Define: g(x) = PLURALITY { C( S )x } consistent petals S : x2 S Conclude:C(S’) ¼gk (S’) for almost all consistent petals S’ of the flower. Conclude: g(x) = f(x) for almost all inputsx.

  42. DP Testing: The Z-Test

  43. Local DP structure Field of flowers (Ai,Bi) Each with its own Local DP functiongi Global g ? B1 B3 B2 Bi B A A A A A A A A A A

  44. Is there GLOBAL DP function g ? • Yes, if ² > 1/ka[DG08] [we re-prove it] • ( can “glue together” many flowers ) • No, if² < 1/k [DG08] • But,with one extra query, get ² = exp( - ka) !

  45. Z-Test S1 Randomly pick k-setsS1 =(B1,A1), S2=(A1,B2), S3=(B2,A2) ( |A1| = |A2| = m = k1/2). B1 A1 S2 B2 A2 S3 Accept if C( S1)A1= C( S2)A1andC( S2 )B2 = C( S3)B2

  46. Derandomization

  47. Subspace DP Code [IJKW 08] T Uk T = { 8-dim affine subspacesof U } ( k = q8 ) U = ( Fq )m • Same list-decoding algo • (from  = 1/poly(k) agreement) • Same DP Test (V-Test) • ( for  = 1/poly(k) acc prob ) Corollary: Polynomial-rate locally (approximately) list-decodable and locally testable code.

  48. Independent vs. Subspace DP Code All k-sets All d-dim subspaces • -approx list-decodable from agreement: • exp ( - k ) 1/poly(  k ) • 2-query testable, acc prob > 1/poly(  k ). • 3-query testable, • acc prob > exp (-k1/2 ) • Analysis:sampling properties of DP graphs • ChernoffChebyshev • (full independence) (2-wise independence)

  49. Summary ( Derandomized ) DP Code Decoding and Testing Analysis of V-Test : Either reject C, or verify that “locally” C = gk( for some g ), and get g(x1), …, g(xk) for independent random xi‘s. Application to 2-Query PCP: Parallel k-repetition for restricted games.

  50. Other Results Yao’s XOR binary ECC: fk (x1, …, xk) = f(x1)  …  f(xk) - approximately locally list-decodable from agreement ½ + ,  > exp(- k), with list size O(1/2) (tight)

More Related