Local correctability of expander codes

Transcription

1 Local correctability of expander codes Brett Hemenway Rafail Ostrovsky Mary Wootters IAS April 4, 24

2 The point(s) of this talk Locally decodable codes are codes which admit sublinear time decoding of small pieces of a message. Expander codes are a family of error correcting codes based on expander graphs. In this work, we show that (appropriately instantiated) expander codes are high-rate locally decodable codes. Only two families of codes known in this regime [KSY,GKS 2]. Expander codes (and the corresponding decoding algorithm and analysis) are very different from existing constructions!

3 Outline Local correctability Definitions and notation Example: Reed-Muller codes Previous work and our contribution 2 Expander codes 3 Local correctability of expander codes Requirement for the inner code: smooth reconstruction Decoding algorithm Example instantiation: finite geometry codes 4 Conclusion

4 Outline Local correctability Definitions and notation Example: Reed-Muller codes Previous work and our contribution 2 Expander codes 3 Local correctability of expander codes Requirement for the inner code: smooth reconstruction Decoding algorithm Example instantiation: finite geometry codes 4 Conclusion

5 Outline Local correctability Definitions and notation Example: Reed-Muller codes Previous work and our contribution 2 Expander codes 3 Local correctability of expander codes Requirement for the inner code: smooth reconstruction Decoding algorithm Example instantiation: finite geometry codes 4 Conclusion

6 Error correcting codes Noisy channel Alice Bob

7 Error correcting codes message x Σ k Noisy channel Alice Bob

8 Error correcting codes message x Σ k codeword C(x) Σ N Noisy channel Alice Bob

9 Error correcting codes message x Σ k corrupted codeword w Σ N codeword C(x) Σ N Noisy channel Alice Bob

10 Error correcting codes message x Σ k corrupted codeword w Σ N codeword C(x) Σ N x? Noisy channel Alice Bob

11 Locally decodable codes message x Σ k corrupted codeword w Σ N codeword C(x) Σ N x? Noisy channel Alice Bob

12 Locally decodable codes message x Σ k corrupted codeword w Σ N codeword C(x) Σ N x i? Noisy channel Alice Bob

13 Locally decodable codes message x Σ k corrupted codeword w Σ N codeword C(x) Σ N Bob makes only q queries x i? Noisy channel Alice Bob

14 Locally correctable codes message x Σ k corrupted codeword w Σ N codeword C(x) Σ N Bob makes only q queries C(x) i? Noisy channel Alice Bob

15 Locally correctable codes, sans stick figures Definition C is (q, δ, η)-locally correctable if for all i [N], for all x Σ k, and for all w Σ N with d(w, C(x)) δn, P {Bob correctly guesses C(x) i } η. Bob reads only q positions in the corrupted word, w.

16 Local correctability vs. local decodability When C is linear, local correctability implies local decodability. G x = C(x)

17 Local correctability vs. local decodability When C is linear, local correctability implies local decodability... x. G x = C(x)

18 Before we get too far Some notation For a code C : Σ k Σ N The message length is k, the length of the message. The block length is N, the length of the codeword. The rate is k/n. The locality is q, the number of queries Bob makes. Goal: large rate, small locality.

19 Outline Local correctability Definitions and notation Example: Reed-Muller codes Previous work and our contribution 2 Expander codes 3 Local correctability of expander codes Requirement for the inner code: smooth reconstruction Decoding algorithm Example instantiation: finite geometry codes 4 Conclusion

20 Example: Reed-Muller Codes f (x, y) (f (, ), f (, ), f (, ), f (, )) Alice Message: multivariate polynomial of total degree d, f F q [z,..., z m ]. Codeword: the evaluation of f at points in F m q : C(f ) = {f ( x)} x F m q

21 Locally Correcting Reed Muller Codes Points in F m q message is f F q [z,..., z m ] codeword is {f ( x)} x F m q

22 Locally Correcting Reed Muller Codes Points in F m q We want to correct C(f ) z = f ( z). f ( z) message is f F q [z,..., z m ] codeword is {f ( x)} x F m q

23 Locally Correcting Reed Muller Codes Points in F m q f ( z) f ( v) We want to correct C(f ) z = f ( z). Choose a random line through z, and consider the restriction to that line. g(t) = f ( z + t v) message is f F q [z,..., z m ] codeword is {f ( x)} x F m q

24 Locally Correcting Reed Muller Codes Points in F m q f ( z) f ( v) message is f F q [z,..., z m ] We want to correct C(f ) z = f ( z). Choose a random line through z, and consider the restriction to that line. g(t) = f ( z + t v) This is a univariate polynomial, and g() = f ( z). codeword is {f ( x)} x F m q

25 Locally Correcting Reed Muller Codes Points in F m q f ( z) f ( v) message is f F q [z,..., z m ] codeword is {f ( x)} x F m q We want to correct C(f ) z = f ( z). Choose a random line through z, and consider the restriction to that line. g(t) = f ( z + t v) This is a univariate polynomial, and g() = f ( z). Query all of the points on the line.

26 Resulting parameters Rate is ( m+d m )/q m (we needed d = O(q), so we can decode) Locality is q (the field size) If we choose m constant, we get: Rate is constant, but less than /2. Locality is N /m = N ε.

27 Outline Local correctability Definitions and notation Example: Reed-Muller codes Previous work and our contribution 2 Expander codes 3 Local correctability of expander codes Requirement for the inner code: smooth reconstruction Decoding algorithm Example instantiation: finite geometry codes 4 Conclusion

28 Question: Reed-Muller Codes have locality N ε and constant rate, but rate is less than /2.

29 Question: Reed-Muller Codes have locality N ε and constant rate, but rate is less than /2. Are there locally decodable codes with locality N ε, and rate arbitrarily close to?

30 Previous Work Rate and locality N ε : Multiplicity codes [Kopparty, Saraf, Yekhanin 2] Lifted codes [Guo, Kopparty, Sudan 22]

31 Previous Work Rate and locality N ε : Multiplicity codes [Kopparty, Saraf, Yekhanin 2] Lifted codes [Guo, Kopparty, Sudan 22] These have decoders similar to RM: the queries form a good code.

32 Previous Work Rate and locality N ε : Multiplicity codes [Kopparty, Saraf, Yekhanin 2] Lifted codes [Guo, Kopparty, Sudan 22] These have decoders similar to RM: the queries form a good code. Another regime: Rate bad but locality 3: ( N/2 2O( log(n)) ), Matching vector codes [Yekhanin 28, Efremenko 29,...] These decoders are different: The queries cannot tolerate any errors. There are so few queries that they are probably all correct.

33 Previous Work Rate and locality N ε : Multiplicity codes [Kopparty, Saraf, Yekhanin 2] Lifted codes [Guo, Kopparty, Sudan 22] These have decoders similar to RM: the queries form a good code. Expander codes [H., Ostrovsky, Wootters 23] Decoder is similar in spirit to lowquery decoders. The queries will not form an error correcting code. Another regime: Rate bad but locality 3: ( N/2 2O( log(n)) ), Matching vector codes [Yekhanin 28, Efremenko 29,...] These decoders are different: The queries cannot tolerate any errors. There are so few queries that they are probably all correct.

34 Outline Local correctability Definitions and notation Example: Reed-Muller codes Previous work and our contribution 2 Expander codes 3 Local correctability of expander codes Requirement for the inner code: smooth reconstruction Decoding algorithm Example instantiation: finite geometry codes 4 Conclusion

35 Tanner Codes [Tanner 8] Given: A d-regular graph G with n vertices and N = nd 2 An inner code C with block length d over Σ edges We get a Tanner code C. C has block length N and alphabet Σ. Codewords are labelings of edges of G. A labeling is in C if the labels on each vertex form a codeword of C.

36 Example [Tanner 8] G is K 8, and C is the [7, 4, 3]-Hamming code. N = ( 8 2) = 28 and Σ = {, }

37 Example [Tanner 8] G is K 8, and C is the [7, 4, 3]-Hamming code. A codeword of C is a labeling of edges of G. red blue (,,,,,,,,,,,,,,,,,,,,,,,,,,, ) C {, } 28

38 Example [Tanner 8] G is K 8, and C is the [7, 4, 3]-Hamming code. These edges form a codeword in the Hamming code red blue (,,,,,,,,,,,,,,,,,,,,,,,,,,, ) C {, } 28

39 Encoding Tanner Codes Encoding is Easy!. Generate parity-check matrix Requires: Edge-vertex incidence matrix of graph Parity-check matrix of inner code 2. Calculate a basis for the kernel of the parity-check matrix 3. This basis defines a generator matrix for the linear Tanner Code 4. Encoding is just multiplication by this generator matrix

40 Linearity If the inner code C is linear, so is the Tanner code C C = Ker(H ) for some parity check matrix H. x C H x =

41 Linearity If the inner code C is linear, so is the Tanner code C C = Ker(H ) for some parity check matrix H. x C H x = So codewords of the Tanner code C also are defined by linear constraints: v y y C v G, H = y Γ(v)

42 Example: vertex edge incidence matrix of K 8 row for each vertex column for each edge Columns have weight 2 (Each edge hits two vertices) Rows have weight 7 (Each vertex has degree seven)

43 Example: parity-check matrix of a Tanner code K 8 and the [7, 4, 3]-Hamming code Parity-check of Hamming code Edge-vertex incidence matrix of K 8

44 Example: parity-check matrix of a Tanner code K 8 and the [7, 4, 3]-Hamming code Vertex

45 Example: parity-check matrix of a Tanner code K 8 and the [7, 4, 3]-Hamming code

46 Example: parity-check matrix of a Tanner code K 8 and the [7, 4, 3]-Hamming code

47 Example: parity-check matrix of a Tanner code K 8 and the [7, 4, 3]-Hamming code

48 Example: parity-check matrix of a Tanner code K 8 and the [7, 4, 3]-Hamming code

49 If the inner code has good rate, so does the outer code Say that C is linear If C has rate r, it satisfies ( r )d linear constraints. Each of the n vertices of G must satisfy these constraints.

50 If the inner code has good rate, so does the outer code Say that C is linear If C has rate r, it satisfies ( r )d linear constraints. Each of the n vertices of G must satisfy these constraints. C is defined by at most n ( r )d constraints.

51 If the inner code has good rate, so does the outer code Say that C is linear If C has rate r, it satisfies ( r )d linear constraints. Each of the n vertices of G must satisfy these constraints. C is defined by at most n ( r )d constraints. Length of C = N = # edges = nd/2

52 If the inner code has good rate, so does the outer code Say that C is linear If C has rate r, it satisfies ( r )d linear constraints. Each of the n vertices of G must satisfy these constraints. C is defined by at most n ( r )d constraints. Length of C = N = # edges = nd/2 The rate of C is R = k N N n ( r )d N = 2r.

53 Better rate bounds? The lower bound R > 2r is independent of the ordering of edges around a vertex Tanner already noticed that order matters. Let G be the complete bipartite graph with 7 vertices per side Let C be the [7, 4, 3] hamming code Then different natural orderings achieve a Tanner code with [49, 6, 9] ( ) [49, 2, 6] ( ) [49, 7, 7] ( ) Meets lower bound of 2 7

54 Expander codes When the underlying graph is an expander graph, the Tanner code is a expander code. Expander codes admit very fast decoding algorithms [Sipser and Spielman 996] Further improvements in [Sipser 96, Zemor, Barg and Zemor 2, 5, 6]

55 Outline Local correctability Definitions and notation Example: Reed-Muller codes Previous work and our contribution 2 Expander codes 3 Local correctability of expander codes Requirement for the inner code: smooth reconstruction Decoding algorithm Example instantiation: finite geometry codes 4 Conclusion

56 Outline Local correctability Definitions and notation Example: Reed-Muller codes Previous work and our contribution 2 Expander codes 3 Local correctability of expander codes Requirement for the inner code: smooth reconstruction Decoding algorithm Example instantiation: finite geometry codes 4 Conclusion

57 Main Result Given: a d-regular expander graph; an inner code of length d with smooth reconstruction. Then: We will give a local-correcting procedure for this expander code.

58 Smooth Reconstruction codeword c Σ N Bob makes q queries c i? Bob

59 Smooth Reconstruction codeword c Σ N Bob makes q queries Suppose that: Each Bob s q queries is (close to) uniformly distributed (they don t need to be independent!) c i? Bob

60 Smooth Reconstruction codeword c Σ N Bob makes q queries Suppose that: Each Bob s q queries is (close to) uniformly distributed (they don t need to be independent!) c i? Bob

61 Smooth Reconstruction codeword c Σ N Bob makes q queries Suppose that: Each Bob s q queries is (close to) uniformly distributed (they don t need to be independent!) c i? Bob

62 Smooth Reconstruction codeword c Σ N Bob makes q queries c i! Suppose that: Each Bob s q queries is (close to) uniformly distributed (they don t need to be independent!) From the (uncorrupted) queries, he can always recover c i. Bob

63 Smooth Reconstruction codeword c Σ N Bob makes q queries c i Suppose that: Each Bob s q queries is (close to) uniformly distributed (they don t need to be independent!) From the (uncorrupted) queries, he can always recover c i. But! He doesn t need to tolerate any errors. Bob

64 Smooth Reconstruction codeword c Σ N Bob Bob makes q queries c i Suppose that: Each Bob s q queries is (close to) uniformly distributed (they don t need to be independent!) From the (uncorrupted) queries, he can always recover c i. But! He doesn t need to tolerate any errors. Then: We say that the code has a smooth reconstruction algorithm.

65 Smooth reconstruction, sans stick figures Definition A code C Σ d has a q-query smooth reconstruction algorithm if, for all i [d] and for all codewords c C : Bob can always determine c i from a set of queries c i,..., c iq Each c ij is (close to) uniformly distributed in [d].

66 Outline Local correctability Definitions and notation Example: Reed-Muller codes Previous work and our contribution 2 Expander codes 3 Local correctability of expander codes Requirement for the inner code: smooth reconstruction Decoding algorithm Example instantiation: finite geometry codes 4 Conclusion

67 Main Result Given: a d-regular expander graph; an inner code of length d with smooth reconstruction. Then: We will give a local-correcting procedure for this expander code.

68 Decoding algorithm: main idea

69 Decoding algorithm: main idea Want to correct the label on this edge For this diagram q = 2

70 Decoding algorithm: main idea Want to correct the label on this edge For this diagram q = 2

71 Decoding algorithm: main idea Want to correct the label on this edge For this diagram q = 2

72 Decoding algorithm: main idea Want to correct the label on this edge For this diagram q = 2

73 Decoding algorithm: main idea Want to correct the label on this edge For this diagram q = 2

74 The expander walk as a tree Want to correct the label on this edge O(log n) q-ary tree (inner code has q-query reconstruction)

75 The expander walk as a tree O(log n) True Statements: The symbols on the leaves determine the symbol on the root. There are q O(log(n)) N ε leaves. The leaves are (nearly) uniformly distributed in G. q-ary tree

76 The expander walk as a tree O(log n) True Statements: The symbols on the leaves determine the symbol on the root. There are q O(log(n)) N ε leaves. The leaves are (nearly) uniformly distributed in G. Idea: Query the leaves! q-ary tree

77 The expander walk as a tree O(log n) True Statements: The symbols on the leaves determine the symbol on the root. There are q O(log(n)) N ε leaves. The leaves are (nearly) uniformly distributed in G. Idea: Query the leaves! Problems: q-ary tree

78 The expander walk as a tree O(log n) q-ary tree True Statements: The symbols on the leaves determine the symbol on the root. There are q O(log(n)) N ε leaves. The leaves are (nearly) uniformly distributed in G. Idea: Query the leaves! Problems: There are errors on the leaves.

79 The expander walk as a tree O(log n) q-ary tree True Statements: The symbols on the leaves determine the symbol on the root. There are q O(log(n)) N ε leaves. The leaves are (nearly) uniformly distributed in G. Idea: Query the leaves! Problems: There are errors on the leaves.

80 The expander walk as a tree O(log n) q-ary tree True Statements: The symbols on the leaves determine the symbol on the root. There are q O(log(n)) N ε leaves. The leaves are (nearly) uniformly distributed in G. Idea: Query the leaves! Problems: There are errors on the leaves. Errors on the leaves propagate.

81 The expander walk as a tree O(log n) q-ary tree True Statements: The symbols on the leaves determine the symbol on the root. There are q O(log(n)) N ε leaves. The leaves are (nearly) uniformly distributed in G. Idea: Query the leaves! Problems: There are errors on the leaves. Errors on the leaves propagate.

82 The expander walk as a tree O(log n) q-ary tree True Statements: The symbols on the leaves determine the symbol on the root. There are q O(log(n)) N ε leaves. The leaves are (nearly) uniformly distributed in G. Idea: Query the leaves! Problems: There are errors on the leaves. Errors on the leaves propagate.

83 The expander walk as a tree O(log n) q-ary tree True Statements: The symbols on the leaves determine the symbol on the root. There are q O(log(n)) N ε leaves. The leaves are (nearly) uniformly distributed in G. Idea: Query the leaves! Problems: There are errors on the leaves. Errors on the leaves propagate.

84 The expander walk as a tree O(log n)! q-ary tree True Statements: The symbols on the leaves determine the symbol on the root. There are q O(log(n)) N ε leaves. The leaves are (nearly) uniformly distributed in G. Idea: Query the leaves! Problems: There are errors on the leaves. Errors on the leaves propagate.

85 Correcting the last layer

86 Correcting the last layer

87 Correcting the last layer O(log n) O(log n) Edges to get us to uniform locations in the graph (not read) Edge we want to learn (not read) Edges for error correction (read)

88 Why should this help? Now the queries can tolerate a few errors. O(log n) q-ary tree

89 Why should this help? False statement: Now the queries can tolerate a few errors. O(log n) q-ary tree

90 Why should this help?! False statement: Now the queries can tolerate a few errors. O(log n) q-ary tree

91 Why should this help? O(log n)! q-ary tree False statement: Now the queries can tolerate a few errors. True statements: This is basically the only thing that can go wrong. Because everything in sight is (nearly) uniform, it probably won t go wrong.

92 Decoding algorithm

93 Decoding algorithm Each leaf edge queries its symbol

94 Decoding algorithm If my correct value were, there would be some path below me with error. If my correct value were, there would be some path below me with errors. Each leaf edge thinks to itself...

95 Decoding algorithm Local corrector: (, ) (, ) (, ) (, ) Each second-level edge reads its symbol and thinks to itself...

96 Decoding algorithm Local corrector: (, ) (, ) (, ) (, ) If I were... path with two errors... Each second-level edge reads its symbol and thinks to itself...

97 Decoding algorithm Local corrector: (, ) (, ) (, ) (, ) If I were... or path with no errors. Each second-level edge reads its symbol and thinks to itself...

98 Decoding algorithm Local corrector: (, ) (, ) (, ) (, ) If I were... path with one error... Each second-level edge reads its symbol and thinks to itself...

99 Decoding algorithm Local corrector: (, ) (, ) (, ) (, ) If I were... or path with two errors. Each second-level edge reads its symbol and thinks to itself...

100 Decoding algorithm If my correct value were, there would be some path below me with 2 errors. If my correct value were, there would be some path below me with errors. Each second-level edge reads its symbol and thinks to itself...

101 Decoding algorithm If my correct value were, there would be some path below me with Ω(log(n)) errors. If my correct value were, there would be some path below me with 7 errors. etc.

102 Decoding algorithm If my correct value were, there would be some path below me with Ω(log(n)) errors. If my correct value were, there would be some path below me with 7 errors. TRIUMPHANTLY RETURN! This only fails if there exist a path that is heavily corrupted. Heavily corrupted paths occur with exponentially small probability.

103 Outline Local correctability Definitions and notation Example: Reed-Muller codes Previous work and our contribution 2 Expander codes 3 Local correctability of expander codes Requirement for the inner code: smooth reconstruction Decoding algorithm Example instantiation: finite geometry codes 4 Conclusion

104 One choice for inner code: based on affine geometry See [Assmus, Key 94, 98] for a nice overview Let L,..., L t be the r-dimensional affine subspaces of F m q, and consider the code with parity-check matrix H: x F m q t L i H H i, x = { x L i x L i q m

105 One choice for inner code: based on affine geometry See [Assmus, Key 94, 98] for a nice overview Let L,..., L t be the r-dimensional affine subspaces of F m q, and consider the code with parity-check matrix H: query the q r nonzeros in this row x F m q L { i x L i t H i, x = x L H i q m Smooth reconstruction: To learn a coordinate indexed by x F m q : pick a random r-flat, Li, containing x. query all of the points in Li.

106 One choice for inner code: based on affine geometry See [Assmus, Key 94, 98] for a nice overview Let L,..., L t be the r-dimensional affine subspaces of F m q, and consider the code with parity-check matrix H: query the q r nonzeros in this row x F m q L { i x L i t H i, x = x L H i q m Smooth reconstruction: To learn a coordinate indexed by x F m q : pick a random r-flat, Li, containing x. query all of the points in Li. Observe: This is not a very good LCC!

107 One good instantiation Graph: Ramanujan graph Inner code: Finite geometry code Results: For any α, ɛ >, for infinitely many N, we get a code with block length N, which has rate α has locality (N/d) ɛ tolerates constant error rate

108 Outline Local correctability Definitions and notation Example: Reed-Muller codes Previous work and our contribution 2 Expander codes 3 Local correctability of expander codes Requirement for the inner code: smooth reconstruction Decoding algorithm Example instantiation: finite geometry codes 4 Conclusion

109 Summary When the inner code has smooth reconstruction, we give a local-decoding procedure for expander codes. This gives a new (and yet old!) family of linear locally correctable codes of rate approaching.

110 Open questions Can we use expander codes to achieve local correctability with lower query complexity? Can we use inner codes with rate < /2?

111 The end Thanks!