Succinct representation of codes with applications to testing Elena Grigorescu Tali Kaufman Madhu Sudan Outline ► Testing membership in error correcting codes ► Sufficient conditions for testing algebraic codes ► Possible promising perspective: rich group of symmetries of code ► Our result: affine/cyclic invariant, sparse codes can be described succinctly by a single, short codeword ► Implies locally testability results ► Proof sketch ► Conclusions Locally testable codes f : FN F2 , N 2 n 0 1 1 … 0 CC q queries C satisfies Code: ( f , g ) 0 Linear: f , g C f C -Accept w.p 1 if -Reject w.p. ε if ( f , C ) 1 ( q independent of n) Testing linear codes via duality ► dual C { y { 0, 1} : y, x 0, x C } N ► [BHR] Test for linear properties are essentially of the form: 1. Given x, pick 2. Accept iff y, x 0 of test: weight ( y) | {i | yi 0} | ► Dual-distance: smallest weight of a codeword in dual-C ► Locality Sufficient conditions for testing Necessary condition for local testing (linear codes): - small “dual distance” - not sufficient( [BHR] show random LDPC not locally testable) Sufficient conditions - Possible approach: nice symmetries of code C is invariant under permutation : [ N ] [ N ] iff x ( x1 , x2 ,..., x N ) C x ( x (1) , x ( 2) ,..., x ( N ) ) C Aut (C ) { : x C , x C} Symmetries and testing Many known testable codes have somewhat large symmetry groups: Eg. Linearity: invariance under general linear group Low degree, Reed-Muller, BCH: invariance under affine group Specific sufficient condition: [KS] affine invariance + ‘local characterization’ imply testing AKKLR Conjecture: 2 transitivity + small dual distance Falsified in general [GKS] Modified AKKLR Question: What if dual code is generated by single low-weight codeword and its shifts under some group G (“SingleOrbit Property under G”) Are these codes testable (for some group G? for all groups G?) Single orbit property under affine invariant/cyclic groups ► ► Affine group: Cyclic group: { : F2n F2n | ( x) ax b} { : F2n F2n | ( x) ax} C has single orbit under cyclic group: w=01001 then B={01001, 10100, 01010, 00101, 10010} is a basis for C ► ► Formally, C has k-single orbit under G ( included in Aut(C) ) if c C , wt(c) k s.t. C Span ({c | G}) Our work ► Study “Single-Orbit Property” of common codes. ► Def: C is sparse if it contains a poly number of codewords ► Duals of binary sparse + affine invariant codes have the single-orbit property under affine group - under some block-length restriction: n prime - [KS’08] Single-orbit codes under affine group are testable. ► Duals of binary sparse + cyclic invariant codes have the single-orbit property under cyclic group - under more block-length restrictions: n, N-1 primes - No testing implications Related works ► ► Sparse, large distance codes are testable [KL, KS] ( tests are coarse, unstructured) Affine/linear invariant + “characterization” imply testing ► Here: sparse affine invariance ► [KL] dual-e-BCH codes are testable (unstructured tests) e-BCH are spanned by shortest codewords ► ► large distance “characterization” (explicit tests) Here: dual-e-BCH are spanned by a single, short codeword (explicit basis / tests) Toward an explicit description of binary affine invariant codes : F2 F2 ( x) ax b ► Affine invariance: ► Any function f : F n F2 is of the form 2 n n f ( x) Trace ( p( x)) ► The Trace function: Trace ( x) x x 2 x 2 x 2 2 n 1 Trace : F2n F2 Trace( x y) Trace( x) Trace( y) Trace(x ) Trace(x ) ... Trace(x i 2i 2n -1 i ) Explicit description of sparse affine families f ( x) Trace( x x 7 x12 x19 ) ► Let ► - What aff inv families does f belong to? Consider the binary rep of degrees: 1, 111, 1100, 10011 Trace( ) C Then Trace( x101 ) C Trace( x 011 ) C Trace( x10001) C Trace(x10001 x 011 ) C , etc... ► ► ► In general: if degree d occurs then its shadow occurs Sparsity translates= into few monomials Shadow(10011) {10011,10010,10001,10000,11,10,1} Affine/Cyclic codes are described by a small set of degrees Proof ingredients Strong number theoretic result of Bourgain implies high weight of functions of the form Trace(a1 x ... a k x g ( x)) dk d1 few degs > Degs inside trace 2n / 2 0 deg< 2 n / 2 2n / 2 2 n(1 ) 2n 1 ? Weil bounds Bourgain Proof ingredients (contd) MacWilliams type counting estimates - fourier transform between the functions that represent number of codewords for each weight in C and in dual- C, respectively For sparse codes of length N and of high distance obtain: Nk # codew. weight k in dual - C |C| Proof sketch Want: exists codew. c with wt < k s.t. Span(aff(c))=Dual-C Dual-C C’ C(a) w C Dual-C’ ► C described by set of degrees D Let dual-C’= Span( aff(w) ) If C’ C then there exists a D ► Trace( x a ) C ' / C Let C(a) Trace( x a ) C ► ► ► ► ► ► weight<k Associate C(a) to codew. w Does every wt<k codew. belong to a dual of some C(a) ? New goal: exists w that does not belong to the dual of any C(a), for all a We show | (dual - C) k | | ( dual - C(a) ) k | Proof Sketch ► ► C, C(a): sparse, high dist (Bourgain) (assuming N-1 and n are primes) How many codew of wt k in dual-C? Nk Nt ► How many codew of wt k in dual-C(a) ? Nk N t 1 ► Total number of degrees a to consider: N/n ► Therefore, there exists codew. of wt<k in dual-C that whose orbit generates C Specifics of the affine case proof ► Here only assume n prime- Bourgain doesn’t hold for all monomials ► Need codes C(a) to have deg a < ► N 1 Use shadow property Trace ( x 11101111) C' Trace(x 10000101) C' ► Show that enough to consider a in the set {2 1 | i [n] } i Cyclic codes under: : F2n F2n ( x) ax ► Punctured affine invariant codes are cyclic ► Cyclic codes are described by generator polynomial (or its roots in the field) ► Alternatively described by function families of the form ► Invariant C {Trace( ai x i )} i 0 ► Degrees can be arbitrary Single orbit: affine vs cyclic codes ► Affine (length N= 2 n ) n prime degrees of monomials are shadow closed |Aut(C)|= N 2 “single orbit” implies testing ► Cyclic (length N-1) n, N-1 primes degrees of monomials are arbitrary |Aut(C)|=N not known if “single orbit” implies testing Open Questions 2 1 ► Do same results hold for non-prime n, ► Single orbit under what other groups imply testing? How large does the Aut group should be to imply testing? ► Small weight basis + invariance implies testing? ► Examples of families where the tests are not the “expected” ones (I.e. not the ones suggested by the description of Aut group) n ? Thank you