direct product : decoding & testing
DESCRIPTION
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. - PowerPoint PPT PresentationTRANSCRIPT
![Page 1: Direct Product : Decoding & Testing](https://reader035.vdocuments.net/reader035/viewer/2022081502/56816736550346895ddbe4d0/html5/thumbnails/1.jpg)
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] )
![Page 2: Direct Product : Decoding & Testing](https://reader035.vdocuments.net/reader035/viewer/2022081502/56816736550346895ddbe4d0/html5/thumbnails/2.jpg)
Average-Hardness Amplification
f g
hard on fractionof inputs
hard on 1- fractionof inputs
![Page 3: Direct Product : Decoding & Testing](https://reader035.vdocuments.net/reader035/viewer/2022081502/56816736550346895ddbe4d0/html5/thumbnails/3.jpg)
(Nonuniform) Hardness on Average
f s
f is δ-hard (for size s), if every circuit (of size s) fails to compute f on δ inputs.
2n
{0,1}n
{0,1}n {0,1}n
![Page 4: Direct Product : Decoding & Testing](https://reader035.vdocuments.net/reader035/viewer/2022081502/56816736550346895ddbe4d0/html5/thumbnails/4.jpg)
Amplification via Repetition
Intuition
If on a random x one can compute f(x) on < ( 1- ) fraction of inputs,
then on k independent random ( x1,…, xk ), one can compute all ( f(x1),…, f(xk) ) on < ( 1- )k exp(- k) fraction of inputs.
![Page 5: Direct Product : Decoding & Testing](https://reader035.vdocuments.net/reader035/viewer/2022081502/56816736550346895ddbe4d0/html5/thumbnails/5.jpg)
Direct-Product (DP) Function
For f: U R, its k-wise DP function is fk : Uk Rk where:
fk ( x1, …, xk ) = ( f(x1), …, f(xk) )
![Page 6: Direct Product : Decoding & Testing](https://reader035.vdocuments.net/reader035/viewer/2022081502/56816736550346895ddbe4d0/html5/thumbnails/6.jpg)
Direct Product Theorem [Yao’82, Levin’87, GNW’95, Imp’95, IW’97,…]
If: f is - hard ( for size s )
Then: fk is exp(-k)- hard ( for size s *
poly(,) )
2n
{0,1}n
2n
{0,1}nk
![Page 7: Direct Product : Decoding & Testing](https://reader035.vdocuments.net/reader035/viewer/2022081502/56816736550346895ddbe4d0/html5/thumbnails/7.jpg)
Direct Product as Error-Correcting Code
![Page 8: Direct Product : Decoding & Testing](https://reader035.vdocuments.net/reader035/viewer/2022081502/56816736550346895ddbe4d0/html5/thumbnails/8.jpg)
DP Encoding [Impagliazzo’02, Trevisan’03]
U Uk
ff k
011010
011010110
010
![Page 9: Direct Product : Decoding & Testing](https://reader035.vdocuments.net/reader035/viewer/2022081502/56816736550346895ddbe4d0/html5/thumbnails/9.jpg)
DP Code Parameters U Ukf f
k
• Local encoding• Local approximate decoding
• Poor distance… Distance amplification: “far-away” messages are mapped to “farther-away” codewords
• Superpoly rate… “Derandomized” DP Code with poly rate.
![Page 10: Direct Product : Decoding & Testing](https://reader035.vdocuments.net/reader035/viewer/2022081502/56816736550346895ddbe4d0/html5/thumbnails/10.jpg)
Direct-Product Code:Decoding
![Page 11: Direct Product : Decoding & Testing](https://reader035.vdocuments.net/reader035/viewer/2022081502/56816736550346895ddbe4d0/html5/thumbnails/11.jpg)
DP Theorem: Constructive Proof [Yao’82, Levin’87, GNW’95, Imp’95, IW’97,…]
Given: circuit C ( of size s*poly(,) ) exp(-k)-computing
fk
Construct: circuit C’
( of size s ) ( 1- )-computing
f.
2n
{0,1}nk
2n
{0,1}n
![Page 12: Direct Product : Decoding & Testing](https://reader035.vdocuments.net/reader035/viewer/2022081502/56816736550346895ddbe4d0/html5/thumbnails/12.jpg)
List-Decoding Lower Bound
Proof: Pick L = 1/ functions f1, …, fL .
Partition inputs into L blocks of size each.
Define C to agree with fik on block i.
Theorem: To decode from agreement , need the list size (1/).
![Page 13: Direct Product : Decoding & Testing](https://reader035.vdocuments.net/reader035/viewer/2022081502/56816736550346895ddbe4d0/html5/thumbnails/13.jpg)
DP Decoding: Previous Work
r1 X rk
C’
X
b1 b bk
if “enough” bi = f(ri),then output b
b
• [GNW, IW97,…]: List-size > exp(1/).
• [IJK06]: poly(1/) list size for “large” , but still sub-optimal & complicated.
![Page 14: Direct Product : Decoding & Testing](https://reader035.vdocuments.net/reader035/viewer/2022081502/56816736550346895ddbe4d0/html5/thumbnails/14.jpg)
New Decoding Algorithm [IJKW 08]
Features:
• list-size O(1/) ( tight ! )
• simple algorithm (and analysis)
• generalizes to Derandomized DP Code
![Page 15: Direct Product : Decoding & Testing](https://reader035.vdocuments.net/reader035/viewer/2022081502/56816736550346895ddbe4d0/html5/thumbnails/15.jpg)
Given C that -computes fk , for > exp(- k)
B1 B2
APick random k-set
x
Randomly partition: |A|=|B1|=k/2 Freeze these random choices
On input x, Pick random k-set (A,B2) containing x.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 ).
![Page 16: Direct Product : Decoding & Testing](https://reader035.vdocuments.net/reader035/viewer/2022081502/56816736550346895ddbe4d0/html5/thumbnails/16.jpg)
Given C that -computes fk , for > exp(- k)
Pick random k-set
Randomly partition: |A|=|B1|=k/2
On input x, Pick random k-set (A,B2) containing x.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 ).
Preprocessing
AlgoA,B1 (x):
![Page 17: Direct Product : Decoding & Testing](https://reader035.vdocuments.net/reader035/viewer/2022081502/56816736550346895ddbe4d0/html5/thumbnails/17.jpg)
Main Theorem: DP Decoding
Pick random k-setRandomly partition: |A|=|B1|=k/2
PreprocessingOn input x, Pick random k-set (A,B2) containing x.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 ).
AlgoA,B1 (x):
Theorem: With probability (²) over (B1,A), the resulting circuit Algo (1- )-computes f.
![Page 18: Direct Product : Decoding & Testing](https://reader035.vdocuments.net/reader035/viewer/2022081502/56816736550346895ddbe4d0/html5/thumbnails/18.jpg)
Proof Ideas
![Page 19: Direct Product : Decoding & Testing](https://reader035.vdocuments.net/reader035/viewer/2022081502/56816736550346895ddbe4d0/html5/thumbnails/19.jpg)
Flowers, cores, petalsFlower: 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
B
B4
AA B2
B3
B1
B5
![Page 20: Direct Product : Decoding & Testing](https://reader035.vdocuments.net/reader035/viewer/2022081502/56816736550346895ddbe4d0/html5/thumbnails/20.jpg)
Structure (Decoding)
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 ¼ fk )
B
B4
AA B2
B3
B1
B5
Assume: C ²-agrees with fk
Consistency
Correctness
![Page 21: Direct Product : Decoding & Testing](https://reader035.vdocuments.net/reader035/viewer/2022081502/56816736550346895ddbe4d0/html5/thumbnails/21.jpg)
Correctness of Decoding
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 ¼ fk )
Preprocessing likely picks a nice
flower
AlgoA,B likely does not time-out
AlgoA,B (x) likely equals f(x)
![Page 22: Direct Product : Decoding & Testing](https://reader035.vdocuments.net/reader035/viewer/2022081502/56816736550346895ddbe4d0/html5/thumbnails/22.jpg)
Proof of DP Structure: Averaging & Symmetry arguments
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 ¼ fk )
B
B4
AA B2
B3
B1
B5
Assume: C ²-agrees with fk
Averaging
Symmetry
![Page 23: Direct Product : Decoding & Testing](https://reader035.vdocuments.net/reader035/viewer/2022081502/56816736550346895ddbe4d0/html5/thumbnails/23.jpg)
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’ ) ]
![Page 24: Direct Product : Decoding & Testing](https://reader035.vdocuments.net/reader035/viewer/2022081502/56816736550346895ddbe4d0/html5/thumbnails/24.jpg)
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.
![Page 25: Direct Product : Decoding & Testing](https://reader035.vdocuments.net/reader035/viewer/2022081502/56816736550346895ddbe4d0/html5/thumbnails/25.jpg)
Proof of DP Structure: Consistency implies correctness (Symmetry)
Idea: A highly incorrect set S’ can’t be a consistent petal in a random flower with correct core
B S’A
![Page 26: Direct Product : Decoding & Testing](https://reader035.vdocuments.net/reader035/viewer/2022081502/56816736550346895ddbe4d0/html5/thumbnails/26.jpg)
Proof of DP Structure: Consistency implies correctness
Idea: A highly incorrect set S’ can’t be a consistent petal in a random flower with correct core
B S’A
![Page 27: Direct Product : Decoding & Testing](https://reader035.vdocuments.net/reader035/viewer/2022081502/56816736550346895ddbe4d0/html5/thumbnails/27.jpg)
Proof of DP Structure: Consistency implies correctness
Idea: A highly incorrect set S’ can’t be a consistent petal in a random flower with correct core.
B S’A
• f(A) = C(B,A)A &• C(B,A)A = C(S’)A &• C(S’)A f(A).
Contradiction !
![Page 28: Direct Product : Decoding & Testing](https://reader035.vdocuments.net/reader035/viewer/2022081502/56816736550346895ddbe4d0/html5/thumbnails/28.jpg)
Direct-Product Testing
![Page 29: Direct Product : Decoding & Testing](https://reader035.vdocuments.net/reader035/viewer/2022081502/56816736550346895ddbe4d0/html5/thumbnails/29.jpg)
Testing
C : Uk Rk
Is C = fk, for some f : U R ?
Fact: C = fk iff for all pairs of intersecting k-sets (S,S’), with A=SS’, C(S)|A = C(S’)|A
![Page 30: Direct Product : Decoding & Testing](https://reader035.vdocuments.net/reader035/viewer/2022081502/56816736550346895ddbe4d0/html5/thumbnails/30.jpg)
Local Testing: V-Test [Goldreich Safra, Dinur Reingold]
C : Uk Rk
Test for one random pair of intersecting k-sets (S,S’), with A=SS’, if C(S)|A = C(S’)|A
AS S’
![Page 31: Direct Product : Decoding & Testing](https://reader035.vdocuments.net/reader035/viewer/2022081502/56816736550346895ddbe4d0/html5/thumbnails/31.jpg)
DP Testing: More formally …On C : Uk Rk Test makes few queries,
and(1) Accepts if C = fk. (2) Rejects if C is “far away” from any
fk(2’) Pr [ Test accepts C ] > C fk on > () of inputs.
- Minimize # queries ( 2 ? 3 ? )- Analyze small ( < 1/k ? <
exp(-k) ? )
![Page 32: Direct Product : Decoding & Testing](https://reader035.vdocuments.net/reader035/viewer/2022081502/56816736550346895ddbe4d0/html5/thumbnails/32.jpg)
DP Testing History Given C : Uk Rk, is C = gk ? #queries
acc.prob.Goldreich-Safra 00 20 .99Dinur-Reingold 06 2 .99Dinur-Goldenberg 08 2 1/kα
Dinur-Goldenberg 08 need > 2 1/kIKW 09 3 exp(-kα)IKW 09* 2 1/kα
* Derandomization
*
![Page 33: Direct Product : Decoding & Testing](https://reader035.vdocuments.net/reader035/viewer/2022081502/56816736550346895ddbe4d0/html5/thumbnails/33.jpg)
Analysis of V-Test
![Page 34: Direct Product : Decoding & Testing](https://reader035.vdocuments.net/reader035/viewer/2022081502/56816736550346895ddbe4d0/html5/thumbnails/34.jpg)
V-Test [GS00,FK00,DR06,DG08]
Randomly pick two k-sets S1 =(B1,A) and S2 =(A,B2)
(with |A| = k1/2 ).
B1 B2
AS1 S2
Accept if C( S1 )A = C( S2 )A
![Page 35: Direct Product : Decoding & Testing](https://reader035.vdocuments.net/reader035/viewer/2022081502/56816736550346895ddbe4d0/html5/thumbnails/35.jpg)
Flowers, cores, petalsFlower: determined by S=(A,B)
Core: A
Core values: α=C(A,B)A
Consistent petals: { (A,B’) | C(A,B’)A = α }
B
B4
AA B2
B3
B1
B5
![Page 36: Direct Product : Decoding & Testing](https://reader035.vdocuments.net/reader035/viewer/2022081502/56816736550346895ddbe4d0/html5/thumbnails/36.jpg)
Structure (Testing)
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).
B
B4
AA B2
B3
B1
B5
Assume: V-Test accepts with prob ²
“Locally” C is DP
![Page 37: Direct Product : Decoding & Testing](https://reader035.vdocuments.net/reader035/viewer/2022081502/56816736550346895ddbe4d0/html5/thumbnails/37.jpg)
Harmonious Flowers
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) ).
B
B4
AA B2
B3
B1
B5
E
C(A, B1 )E ¼ C(A, B2 )E, with |E| = |A|
Assume: V-Test accepts with prob ²
Proof by symmetry arguments
(as in Decoding)
![Page 38: Direct Product : Decoding & Testing](https://reader035.vdocuments.net/reader035/viewer/2022081502/56816736550346895ddbe4d0/html5/thumbnails/38.jpg)
Harmony DP structureHarmonious flower:• many (/2) consistent
petals,• on consistent petals, V-test
accepts almost certainly ( 1-poly(²) ).
B
B4
AA B2
B3
B1
B5Main Lemma: Define g(x) = PLURALITY { C( S’ )x | consistent petals S’ , x S’ }.Then C(S’) ¼ gk (S’) for almost all (1-poly(²)) consistent petals S’.
![Page 39: Direct Product : Decoding & Testing](https://reader035.vdocuments.net/reader035/viewer/2022081502/56816736550346895ddbe4d0/html5/thumbnails/39.jpg)
Assume otherwise.A random B1 in Cons has many “minority” elements x where C(B1)x g(x).
A random E ½ B1 has many “minority” elements [Chernoff]
A random B2=(E,D2) is likely s.t. C(B2)E ¼ g(E) [def of g]
Then C(B1)E C(B2)E, Hence no harmony !
BA D2
D1
E
B1
Proof Sketch of Main Lemma
B2
![Page 40: Direct Product : Decoding & Testing](https://reader035.vdocuments.net/reader035/viewer/2022081502/56816736550346895ddbe4d0/html5/thumbnails/40.jpg)
Decoding vs. Testing
![Page 41: Direct Product : Decoding & Testing](https://reader035.vdocuments.net/reader035/viewer/2022081502/56816736550346895ddbe4d0/html5/thumbnails/41.jpg)
Decoding Testing
There are many harmonious flowers with:• many consistent petals,• restricted to consistent
petals, V-Test accepts almost surely.
V-Test ²-accepts C
Conclude: C(S’) ¼ gk (S’) for almost all consistent petals S’ of the flower.
There are many nice flowers with:• correct core,• many consistent
petals.
C ²-computes fk
Define: g(x) = PLURALITY { C( S )x }consistent petals S :
x2 S
Conclude: g(x) = f(x) for almost all inputs x.
Consistency
Correctness Harmony DP
![Page 42: Direct Product : Decoding & Testing](https://reader035.vdocuments.net/reader035/viewer/2022081502/56816736550346895ddbe4d0/html5/thumbnails/42.jpg)
DP Testing: The Z-Test
![Page 43: Direct Product : Decoding & Testing](https://reader035.vdocuments.net/reader035/viewer/2022081502/56816736550346895ddbe4d0/html5/thumbnails/43.jpg)
Local DP structureField of flowers (Ai,Bi)
Each with its ownLocal DP function gi
Global g ? BAA
B2 AA B
i AAB3 AA
B1 AA
![Page 44: Direct Product : Decoding & Testing](https://reader035.vdocuments.net/reader035/viewer/2022081502/56816736550346895ddbe4d0/html5/thumbnails/44.jpg)
Is there GLOBAL DP function g ?
Yes, if ² > 1/ka [DG08] [we re-prove it] ( can “glue together” many flowers )
No, if ² < 1/k [DG08] But, with one extra query, get ² = exp( - ka) !
![Page 45: Direct Product : Decoding & Testing](https://reader035.vdocuments.net/reader035/viewer/2022081502/56816736550346895ddbe4d0/html5/thumbnails/45.jpg)
Z-TestRandomly pick k-sets S1 =(B1,A1), S2=(A1,B2),
S3=(B2,A2) ( |A1| = |A2| = m = k1/2). B1
B2
A1
A2
S1
S2
S3
Accept if C( S1 )A1= C( S2 )A1
and C( S2 )B2 =
C( S3 )B2
![Page 46: Direct Product : Decoding & Testing](https://reader035.vdocuments.net/reader035/viewer/2022081502/56816736550346895ddbe4d0/html5/thumbnails/46.jpg)
Derandomization
![Page 47: Direct Product : Decoding & Testing](https://reader035.vdocuments.net/reader035/viewer/2022081502/56816736550346895ddbe4d0/html5/thumbnails/47.jpg)
Subspace DP Code [IJKW 08]
U = ( Fq )m T Uk T = { 8-dim affine
subspaces of U } ( k = q8 )
• 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.
![Page 48: Direct Product : Decoding & Testing](https://reader035.vdocuments.net/reader035/viewer/2022081502/56816736550346895ddbe4d0/html5/thumbnails/48.jpg)
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
Chernoff Chebyshev (full independence) (2-wise independence)
![Page 49: Direct Product : Decoding & Testing](https://reader035.vdocuments.net/reader035/viewer/2022081502/56816736550346895ddbe4d0/html5/thumbnails/49.jpg)
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.
![Page 50: Direct Product : Decoding & Testing](https://reader035.vdocuments.net/reader035/viewer/2022081502/56816736550346895ddbe4d0/html5/thumbnails/50.jpg)
Other ResultsYao’s XOR binary ECC:
fk (x1, …, xk) = f(x1) … f(xk)
- approximately locally list-decodable from agreement ½ + , > exp(- k), with list size O(1/2) (tight)
![Page 51: Direct Product : Decoding & Testing](https://reader035.vdocuments.net/reader035/viewer/2022081502/56816736550346895ddbe4d0/html5/thumbnails/51.jpg)
Open Questions Other examples of derandomized DP
Codes ? Better parameters ???
Derandomized 2-PCP : Re-proving / improving [Moshkovitz-Raz’08, Dinur-Harsha’09] using similar techniques.
( recent progress [Dinur-Meir’09] )
![Page 52: Direct Product : Decoding & Testing](https://reader035.vdocuments.net/reader035/viewer/2022081502/56816736550346895ddbe4d0/html5/thumbnails/52.jpg)