Linear codes ‹#› Linear codes 1 CHAPTER 2: Linear codes ABSTRACT Most of the important codes are special types of so-called linear codes. Linear codes are of very large importance because they have very concise description, very nice properties, very easy encoding And, in principle, easy to describe decoding. IV054 ‹#› Linear codes 2 Linear codes Linear codes Linear codes are special sets of words of the length n over an alphabet {0,..,q -1}, where q is a power of prime. Since now on sets of words Fqn will be considered as vector spaces V(n,q) of vectors of length n with elements from the set {0,..,q -1} and arithmetical operations will be taken modulo q. The set {0,..,q -1} with operations + and · modulo q is called also the Galois field GF(q). Definition A subset C Í V(n,q) is a linear code if (1) u + v Î C for all u, v Î C (2) au Î C for all u Î C, a Î GF(q) Example Codes C1, C2, C3 introduced in Lecture 1 are linear codes. IV054 Lemma A subset C Í V(n,q) is a linear code if one of the following conditions is satisfied (1) C is a subspace of V(n,q) (2) sum of any two codewords from C is in C (for the case q = 2) If C is a k -dimensional subspace of V(n,q), then C is called [n,k] -code. It has qk codewords. If minimal distance of C is d, then it is called [n,k,d] code. Linear codes are also called “group codes“. ‹#› Linear codes 3 Linear codes Exercise Which of the following binary codes are linear? C1 = {00, 01, 10, 11} C2 = {000, 011, 101, 110} C3 = {00000, 01101, 10110, 11011} C5 = {101, 111, 011} C6 = {000, 001, 010, 011} C7 = {0000, 1001, 0110, 1110} IV054 How to create a linear code Notation If S is a set of vectors of a vector space, then let áSń be the set of all linear combinations of vectors from S. Theorem For any subset S of a linear space, áSń is a linear space that consists of the following words: • the zero word, • all words in S, • all sums of two or more words in S. Example S = {0100, 0011, 1100} áSń = {0000, 0100, 0011, 1100, 0111, 1011, 1000, 1111}. ‹#› Linear codes 4 Linear codes Basic properties of linear codes Notation: w(x) (weight of x) denotes the number of non-zero entries of x. Lemma If x, y Î V(n,q), then h(x,y) = w(x - y). Proof x - y has non-zero entries in exactly those positions where x and y differ. IV054 Theorem Let C be a linear code and let weight of C, notation w(C), be the smallest of the weights of non-zero codewords of C. Then h(C) = w(C). Proof There are x, y Î C such that h(C) = h(x,y). Hence h(C) = w(x - y) ł w(C). On the other hand for some x Î C w(C) = w(x) = h(x,0) ł h(C). Consequence • If C is a code with m codewords, then in order to determine h(C) one has to make comparisons in the worth case. • If C is a linear code, then in order to compute h(C) , m - 1 comparisons are enough. ‹#› Linear codes 5 Linear codes Basic properties of linear codes If C is a linear [n,k] -code, then it has a basis consisting of k codewords. Example Code C4 = {0000000, 1111111, 1000101, 1100010, 0110001, 1011000, 0101100, 0010110, 0001011, 0111010, 0011101, 1001110, 0100111, 1010011, 1101001, 1110100} has the basis {1111111, 1000101, 1100010, 0110001}. How many different bases has a linear code? IV054 Theorem A binary linear code of dimension k has bases. ‹#› Linear codes 6 Linear codes Advantages and disadvantages of linear codes I. Advantages - big. 1. Minimal distance h(C) is easy to compute if C is a linear code. 2. Linear codes have simple specifications. • To specify a non-linear code usually all codewords have to be listed. • To specify a linear [n,k] -code it is enough to list k codewords. Definition A k ´ n matrix whose rows form a basis of a linear [n,k] -code (subspace) C is said to be the generator matrix of C. Example The generator matrix of the code and of the code 3. There are simple encoding/decoding procedures for linear codes. IV054 ‹#› Linear codes 7 Linear codes Advantages and disadvantages of linear codes II. Disadvantages of linear codes are small: 1. Linear q -codes are not defined unless q is a prime power. 2. The restriction to linear codes might be a restriction to weaker codes than sometimes desired. IV054 > ‹#› Linear codes 8 Linear codes Equivalence of linear codes Definition Two linear codes GF(q) are called equivalent if one can be obtained from another by the following operations: (a) permutation of the positions of the code; (b) multiplication of symbols appearing in a fixed position by a non-zero scalar. IV054 Theorem Two k ´ n matrices generate equivalent linear [n,k] -codes over GF(q) if one matrix can be obtained from the other by a sequence of the following operations: (a) permutation of the rows (b) multiplication of a row by a non-zero scalar (c) addition of one row to another (d) permutation of columns (e) multiplication of a column by a non-zero scalar Proof Operations (a) - (c) just replace one basis by another. Last two operations convert a generator matrix to one of an equivalent code. ‹#› Linear codes 9 Linear codes Equivalence of linear codes Theorem Let G be a generator matrix of an [n,k] -code. Rows of G are then linearly independent .By operations (a) - (e) the matrix G can be transformed into the form: [ Ik | A ] where Ik is the k ´ k identity matrix, and A is a k ´ (n - k) matrix. IV054 Example ‹#› Linear codes 10 Linear codes Encoding with a linear code is a vector ´ matrix multiplication Let C be a linear [n,k] -code over GF(q) with a generator matrix G. Theorem C has qk codewords. Proof Theorem follows from the fact that each codeword of C can be expressed uniquely as a linear combination of the basis vectors. Corollary The code C can be used to encode uniquely qk messages. Let us identify messages with elements V(k,q). Encoding of a message u = (u1, … ,uk) with the code C: IV054 Example Let C be a [7,4] -code with the generator matrix A message (u1, u2, u3, u4) is encoded as:??? For example: 0 0 0 0 is encoded as ………………………….. ? 1 0 0 0 is encoded as ………………………….. ? 1 1 1 0 is encoded as ………………………….. ? ‹#› Linear codes 11 Linear codes Uniqueness of encodings with linear codes Theorem If G={wi}i=1k is a generator matrix of a binary linear code C of length n and dimension k, then v = uG ranges over all 2k codewords of C as u ranges over all 2k words of length k. Therefore C = { uG | u Î {0,1}k } Moreover u1G = u2G if and only if u1 = u2. Proof If u1G – u2G=0, then And, therefore, since wi are linearly independent, u1 = u2. IV054 ‹#› Linear codes 12 Linear codes Decoding of linear codes Decoding problem: If a codeword: x = x1 … xn is sent and the word y = y1 … yn is received, then e = y – x = e1 … en is said to be the error vector. The decoder must decide, from y, which x was sent, or, equivalently, which error e occurred. To describe main Decoding method some technicalities have to be introduced Definition Suppose C is an [n,k] -code over GF(q) and u Î V(n,q). Then the set u + C = { u + x | x Î C } is called a coset (u-coset) of C in V(n,q). IV054 Example Let C = {0000, 1011, 0101, 1110} Cosets: 0000 + C = C, 1000 + C = {1000, 0011, 1101, 0110}, 0100 + C = {0100, 1111, 0001, 1010} = 0001+C, 0010 + C = {0010, 1001, 0111, 1100}. Are there some other cosets in this case? Theorem Suppose C is a linear [n,k] -code over GF(q). Then (a) every vector of V(n,k) is in some coset of C, (b) every coset contains exactly qk elements, (c) two cosets are either disjoint or identical. ‹#› Linear codes 13 Linear codes Nearest neighbour decoding scheme: Each vector having minimum weight in a coset is called a coset leader. 1. Design a (Slepian) standard array for an [n,k] -code C - that is a qn - k ´ qk array of the form: IV054 Example A word y is decoded as codeword of the first row of the column in which y occurs. Error vectors which will be corrected are precisely coset leaders! In practice, this decoding method is too slow and requires too much memory. codewords coset leader codeword 2 … codeword 2k coset leader + … + .. + + + coset leader + … + coset leader 0000 1011 0101 1110 1000 0011 1101 0110 0100 1111 0001 1010 0010 1001 0111 1100 ‹#› Linear codes 14 Linear codes Probability of good error correction What is the probability that a received word will be decoded as the codeword sent (for binary linear codes and binary symmetric channel)? Probability of an error in the case of a given error vector of weight i is p i (1 - p)n - i. Therefore, it holds. Theorem Let C be a binary [n,k] -code, and for i = 0,1, … ,n let ai be the number of coset leaders of weight i. The probability Pcorr (C) that a received vector when decoded by means of a standard array is the codeword which was sent is given by IV054 Example For the [4,2] -code of the last example a0 = 1, a1 = 3, a2 = a3 = a4 = 0. Hence Pcorr (C) = (1 - p)4 + 3p(1 - p)3 = (1 - p)3(1 + 2p). If p = 0.01, then Pcorr = 0.9897 ‹#› Linear codes 15 Linear codes Probability of good error detection Suppose a binary linear code is used only for error detection. The decoder will fail to detect errors which have occurred if the received word y is a codeword different from the codeword x which was sent, i. e. if the error vector e = y - x is itself a non-zero codeword. The probability Pundetect (C) that an incorrect codeword is received is given by the following result. Theorem Let C be a binary [n,k] -code and let Ai denote the number of codewords of C of weight i. Then, if C is used for error detection, the probability of an incorrect message being received is IV054 Example In the case of the [4,2] code from the last example A2 = 1 A3 = 2 Pundetect (C) = p2 (1 - p)2 + 2p3 (1 - p) = p2 – p4. For p = 0.01 Pundetect (C) = 0.000099. ‹#› Linear codes 16 Linear codes Dual codes Inner product of two vectors (words) u = u1 … un, v = v1 … vn in V(n,q) is an element of GF(q) defined (using modulo q operations) by u × v = u1v1 + … + unvn. Example In V(4,2): 1001 × 1001 = 0 In V(4,3): 2001 × 1210 = 2 1212 × 2121 = 2 If u × v = 0 then words (vectors) u and v are called orthogonal. Properties If u, v, w Î V(n,q), l, m Î GF(q), then u × v = v × u, (lu + mv) × w = l (u × w) + m (v × w). Given a linear [n,k] -code C, then dual code of C, denoted by C^, is defined by C^ = {v Î V(n,q) | v × u = 0 if u Î C}. Lemma Suppose C is an [n,k] -code having a generator matrix G. Then for v Î V(n,q) v Î C^ <=> vGT = 0, where GT denotes the transpose of the matrix G. Proof Easy. IV054 ‹#› Linear codes 17 Linear codes PARITE CHECKS versus ORTHOGONALITY For understanding of the role the parity checks play for linear codes, it is important to understand relation between orthogonality and special parity checks. If words x and y are orthogonal, then the word y has even number of ones (1’s) in the positions determined by ones (1’s) in the word x. This implies that if words x and y are orthogonal, then x is a parity check word for y and y is a parity check word for x. Exercise: Let the word 100001 be orthogonal to a set S of binary words of length 6. What can we say about the words in S? IV054 > ‹#› Linear codes 18 Linear codes EXAMPLE For the [n,1] -repetition code C, with the generator matrix G = (1,1, … ,1) the dual code C^ is [n,n - 1] -code with the generator matrix G^, described by ^ IV054 ‹#› Linear codes 19 Linear codes Parity check matrices Example If ^ If ^ IV054 Theorem Suppose C is a linear [n,k] -code over GF(q), then the dual code C^ is a linear [n,n - k] -code. Definition A parity-check matrix H for an [n,k] -code C is a generator matrix of C^. ‹#› Linear codes 20 Linear codes Parity check matrices Definition A parity-check matrix H for an [n,k] -code C is a generator matrix of C^. Theorem If H is parity-check matrix of C, then C = {x Î V(n,q) | xHT = 0}, and therefore any linear code is completely specified by a parity-check matrix. IV054 Example Parity-check matrix for and for The rows of a parity check matrix are parity checks on codewords. They say that certain linear combinations of the coordinates of every codeword are zeros. ‹#› Linear codes 21 Linear codes Syndrome decoding Theorem If G = [Ik | A] is the standard form generator matrix of an [n,k] -code C, then a parity check matrix for C is H = [-AT | In-k]. Example Definition Suppose H is a parity-check matrix of an [n,k] -code C. Then for any y Î V(n,q) the following word is called the syndrome of y: S(y) = yHT. Lemma Two words have the same syndrom iff they are in the same coset. Syndrom decoding Assume that a standard array of a code C is given and, in addition, let in the last two columns the syndrom for each coset be given. When a word y is received, compute S(y) = yHT, locate S(y) in the “syndrom column”, and then locate y in the same row and decode y as the codeword in the same column and in the first row. IV054 ‹#› Linear codes 22 Linear codes KEY OBSERVATION for SYNDROM COMPUTATION When preparing a ”syndrome decoding'' it is sufficient to store only two columns: one for coset leaders and one for syndromes. Example coset leaders syndromes l(z) z 0000 00 1000 11 0100 01 0010 10 Decoding procedure • Step 1 Given y compute S(y). • Step 2 Locate z = S(y) in the syndrome column. • Step 3 Decode y as y - l(z). IV054 Example If y = 1111, then S(y) = 01 and the above decoding procedure produces 1111 – 0100 = 1011. Syndrom decoding is much fatser than searching for a nearest codeword to a received word. However, for large codes it is still too inefficient to be practical. In general, the problem of finding the nearest neighbour in a linear code is NP-complete. Fortunately, there are important linear codes with really efficient decoding. ‹#› Linear codes 23 Linear codes Hamming codes An important family of simple linear codes that are easy to encode and decode, are so-called Hamming codes. Definition Let r be an integer and H be an r ´ (2r - 1) matrix columns of which are non-zero distinct words from V(r,2). The code having H as its parity-check matrix is called binary Hamming code and denoted by Ham(r,2). Example IV054 Theorem Hamming code Ham(r,2) • is [2r - 1, 2r – 1 - r] -code, • has minimum distance 3, • is a perfect code. Properties of binary Hamming coes Coset leaders are precisely words of weight Ł 1. The syndrome of the word 0…010…0 with 1 in j -th position and 0 otherwise is the transpose of the j -th column of H. ‹#› Linear codes 24 Linear codes Hamming codes - decoding Decoding algorithm for the case the columns of H are arranged in the order of increasing binary numbers the columns represent. • Step 1 Given y compute syndrome S(y) = yHT. • Step 2 If S(y) = 0, then y is assumed to be the codeword sent. • Step 3 If S(y) ą 0, then assuming a single error, S(y) gives the binary position of the error. IV054 > ‹#› Linear codes 25 Linear codes Example For the Hamming code given by the parity-check matrix and the received word y = 110 1011, we get syndrome S(y) = 110 and therefore the error is in the sixth position. Hamming code was discovered by Hamming (1950), Golay (1950). 1 It was conjectured for some time that Hamming codes and two so called Golay codes are the only non-trivial perfect codes. Comment Hamming codes were originally used to deal with errors in long-distance telephon calls. IV054 ‹#› Linear codes 26 Linear codes ADVANTAGES of HAMMING CODES Let a binary symmetric channel is used which with probability q correctly transfers a binary symbol. If a 4-bit message is transmitted through such a channel, then correct transmission of the message occurs with probability q4. If Hamming (7,4,3) code is used to transmit a 4-bit message, then probability of correct decoding is q7 + 7(1 - q)q6. In case q = 0.9 the probability of correct transmission is 0.651 in the case no error correction is used and 0.8503 in the case Hamming code is used - an essential improvement. IV054 > ‹#› Linear codes 27 Linear codes IMPORTANT CODES • Hamming (7,4,3) -code. It has 16 codewords of length 7. It can be used to send 27 = 128 messages and can be used to correct 1 error. IV054 • Golay (23,12,7) -code. It has 4 096 codewords. It can be used to transmit 8 388 608 messages and can correct 3 errors. • Quadratic residue (47,24,11) -code. It has 16 777 216 codewords and can be used to transmit 140 737 488 355 238 messages and correct 5 errors. • Hamming and Golay codes are the only non-trivial perfect codes. ‹#› Linear codes 28 Linear codes GOLAY CODES - DESCRIPTION Golay codes G24 and G23 were used by Voyager I and Voyager II to transmit color pictures of Jupiter and Saturn. Generation matrix for G24 has the form G24 is (24,12,8) –code and the weights of all codewords are multiples of 4. G23 is obtained from G24 by deleting last symbols of each codeword of G24. G23 is (23,12,7) –code. IV054 ‹#› Linear codes 29 Linear codes GOLAY CODES - CONSTRUCTION Matrix G for Golay code G24 has actually a simple and regular construction. The first 12 columns are formed by a unitary matrix I12, next column has all 1’s. Rows of the last 11 columns are cyclic permutations of the first row which has 1 at those positions that are squares modulo 11, that is 0, 1, 3, 4, 5, 9. IV054 > ‹#› Linear codes 30 Linear codes SINGLETON BOUND If C is a linear [n,k,d] -code, then n - k ł d - 1 (Singleton bound). To show the above bound we can use the following lemma. Lemma If u is a codeword of a linear code C of weight s,then there is a dependence relation among s columns of any parity check matrix of C, and conversely, any dependence relation among s columns of a parity check matrix of C yields a codeword of weight s in C. Proof Let H be a parity check matrix of C. Since u is orthogonal to each row of H, the s components in u that are nonzero are the coefficients of the dependence relation of the s columns of H corresponding to the s nonzero components. The converse holds by the same reasoning. IV054 Corollary If C is a linear code, then C has minimum weight d if d is the largest number so that every d - 1 columns of any parity check matrix of C are independent. Corollary For a linear [n,k,d] it holds n - k ł d - 1. A linear [n,k,d] -code is called maximum distance separable (MDS code) if d = n –k + 1. MDS codes are codes with maximal possible minimum weight. ‹#› Linear codes 31 Linear codes REED-MULLER CODES Reed-Muller codes form a family of codes defined recursively with interesting properties and easy decoding. If D1 is a binary [n,k1,d1] -code and D2 is a binary [n,k2,d2] -code, a binary code C of length 2n is defined as follows C = { u | u + v |, where u Î D1, v Î D2}. Lemma C is [2n,k1 + k2, min{2d1,d2}] -code and if Gi is a generator matrix for Di, i = 1, 2, then is a generator matrix for C. Reed-Muller codes R(r,m), with 0 Ł r Ł m are binary codes of length n = 2m. R(m,m) is the whole set of words of length n, R(0,m) is the repetition code. If 0 < r < m, then R(r + 1,m + 1) is obtained from codes R(r + 1,m) and R(r,m) by the above construction. Theorem The dimension of R(r,m) equals The minimum weight of R(r,m) equals 2m - r. Codes R(m - r - 1,m) and R(r,m) are dual codes. IV054 ‹#› Linear codes 32 Linear codes Singleton Bound IV054 Singleton bound: Let C be a q-ary (n, M, d)-code. Then M Ł q n-d+1 . Proof Take some d − 1 coordinates and project all codewords to the resulting coordinates. The resulting codewords are all different and therefore M cannot be larger than the number of q-ary words of length n−d−1. Codes for which M = q n−d+1 are called MDS-codes (Maximum Distance Separable). Corollary: If C is a q-ary linear [n, k, d]-code, then k + d Ł n + 1. > ‹#› Linear codes 33 Linear codes Shortening and puncturing of linear codes IV054 Let C be a q-ary linear [n, k, d]-code. Let D = {(x1, ... , xn-1) | (x1, ... , xn-1, 0)ÎC}. Then D is a linear [n-1, k-1, d]-code – a shortening of the code C. Corollary: If there is a q-ary [n, k, d]-code, then shortening yields a q-ary [n−1, k−1, d]-code. Let C be a q-ary [n, k, d]-code. Let E = {(x1, ... , xn-1) | (x1, ... , xn-1, x)ÎC, for some x Ł q}, then E is a linear [n-1, k, d-1]-code – a puncturing of the code C. Corollary: If there is a q-ary [n, k, d]-code with d >1, then there is a q-ary [n−1, k, d-1]-code. > ‹#› Linear codes 34 Linear codes Lengthening of Codes – Constructions X and XX IV054 Construction X Let C and D be q-nary linear codes with parameters [n, K, d] and [n, k, D], where D > d, and K > k. Assume also that there exists a q-nary code E with parameters [l, K − k, δ ]. Then there is a ”longer” q-nary code with parameters [n + l, K, min(d + δ, D)]. The lengthening of C is constructed by appending φ(x) to each word x ∈ C, where φ : C/D → E is a bijection – a well known application of this construction is the addition of the parity bit in binary codes. Construction XX Let the following q-ary codes be given: a code C with parameters [n, k, d]; its sub-codes Ci , i = 1,2 with parameters [n, k − ki , di] and with C1 ∩ C2 of minimum distance ≥ D; auxiliary q-nary codes Ei , i = 1,2 with parameters [li , ki , δi]. Then there is a q-ary code with parameters [n + l1 + l2 , k, min{D, d2 + δ1, d1 + δ2 , d + δ1 + δ2}]. > ‹#› Linear codes 35 Linear codes Strength of Codes IV054 •Strength of codes is another important parameter of codes. It is defined through the concept of the strength of so-called orthogonal arrays - an important concepts of combinatorics. •An orthogonal array QAλ(t, n, q) is an array of n columns, λq t rows with elements from Fq and the property that in the projection onto any set of t columns each possible t-tuple occurs the same number λ of times. t is called strength of such an orthogonal array. •For a code C, let t(C) be the strength of C - if C is taken as an orthogonal array. •Importance of the concept of strength follows also from the following Principle of duality: For any code C its minimum distance and the strength of C^ are closely related. Namely d(C) = t(C^) + 1. > ‹#› Linear codes 36 Linear codes Dimension of Dual Linear Codes IV054 If C is an [n, k]-code, then its dual code C⊥ is [n, n − k] code. A binary linear [n, 1] repetition code with codewords of length n has two codewords: all-0 codeword and all-1 codeword. Dual code to [n, 1] repetition code is so-called sum zero code of all binary n-bit words whose entries sum to zero (modulo 2). It is a code of dimension n − 1 and it is a linear [n, n − 1, 2] code > ‹#› Linear codes 37 Linear codes Reed-Solomon Codes An important example of MDS-codes are q-ary Reed-Solomon codes RSC(k, q), for k ≤ q. They are codes generator matrix of which has rows labelled by polynomials X i, 0 ≤ i ≤ k − 1, columns by elements 0, 1, . . . , q − 1 and the element in a row labelled by a polynomial p and in a column labelled by an element u is p(u). RSC(k, q) code is [q, k, q − k + 1] code. Example Generator matrix for RSC(3, 5) code is Interesting property of Reed-Solomon codes: RSC(k, q)⊥ = RSC(q − k, q). Reed-Solomon codes are used in digital television, satellite communication, wireless communication, barcodes, compact discs, DVD,... They are very good to correct burst errors - such as ones caused by solar energy. IV054 > ‹#› Linear codes 38 Linear codes Trace and Subfield Codes • Let p be a prime and r an integer. A trace tr is mapping from Fpr into Fp defined by tr(x) = • Trace is additive (tr(x1 + x2) = tr(x1) + tr(x2)) and Fp-linear (tr(λx) = λtr(x)). • If C is a linear code over Fpr and tr is a trace mapping from Fpr to Fp, then trace code tr(C) is a code over Fp defined by (tr(x1), tr(x2), . . . , tr(xn)) where (x1, x2, . . . , xn) ÎÎ C. • If FnprÉ C is a linear code of strength t, then strength of tr(C) is at least t. • Let C be a linear code. The subfield code CFp consists of those codewords of C all of whose entries are in Fp. • Delsarte theorem If C is a linear code. Then tr(C)^ = (C^)Fp . IV054 > ‹#› Linear codes 39 Linear codes Soccer Games Betting System IV054 Ternary Golay code with parameters (11, 729, 5) can be used to bet for results of 11 soccer games with potential outcomes 1 (if home team wins), 2 (if guests win) and 3 (in case of a draw). If 729 bets are made, then at least one bet has at least 9 results correctly guessed. In case one has to bet for 13 games, then one can usually have two games with pretty sure outcomes and for the rest one can use the above ternary Golay code. >