Error Correction Review

Transcription

1 Error Correction Review A single overall parity-check equation detects single errors. Hamming codes used m equations to correct one error in 2 m 1 bits. We can use nonbinary equations if we create symbols from sequences of bits. E.g., four bits can represent 0,1,...,15. Nonbinary check equations can use more advanced arithmetic, such as mod 16.

2 Nonbinary Single Error Correcting Code The single check equation c 1 +c 2 + +c n = 0 allows detection of a single symbol error in a received n-tuple. Furthermore, the syndrome s defined by s = r 1 +r 2 + +r n indicates the magnitude of the error. If the error is in location i and the incorrect symbol is r i = c i +e i, then s = r 1 +r 2 + +r n = c 1 + +(c i +e i )+ +c n = e i. The syndrome tells exactly what should be subtracted from the incorrect symbol in order to obtain a codeword. What is not known is where the error is which symbol is wrong.

3 More Equations Needed A second equation is needed to identify the error location. The effect on syndrome of an error magnitude should be different for each location. A reasonable choice for this second equation: 1 c c n c n = 0. Now every valid codeword satisfies two equations: 1 c c c n = 0 1 c c n c n = 0 We can derive encoding equations to express c 1,c 2 in terms of c 3,...,c n. Example: let symbols be 4-bit values with addition modulo 16. For n = 15, [ ] H = is check matrix for a code that can almost correct single symbols errors.

4 Decoding Procedure Suppose there is a single error of magnitude e i 0 in location i. The syndrome s = [s 0 s 1 ] can be expressed in terms of unknowns i and e i : n n s 0 = r j = e i + c j = e i s 1 = j=1 j=1 n jr j = ie i + j=1 n jc j = ie i j=1 We can determine e i and i from the syndrome equations: e i = s 0 i = ie i = s 1 e i s 0 Sadly, division is not always defined for modulo 16 arithmetic. E.g., suppose s 0 = 4, s 1 = 8. Then s 1 = is 0 mod 16 has four solutions: 2, 6, 10, 14. We cannot be certain where the single error is located.

5 Galois Field Arithmetic To make error location work, we define a new arithmetic for bit vectors. Addition is exclusive-or (binary addition without carry) Multiplication is determined multiplying by 2: bit shift with feedback. Example: multiplication in GF(16). 2 = = 0100 = = 1000 = = 0011 = 3 Fact: every nonzero 4-bit sequence is a power of 2.

6 2EC BCH code Let α = 2 and define two equations using α and α 3. The parity-check matrix is the following: [ ] 1 α α 2 α n 1 H = 1 α 3 α 6 α 3(n 1) Codewords defined by H are the polynomials c(x) with zeroes α and α 3. So every codeword is a multiple of the minimal polynomials of α and α 3. Therefore the number of check bits satisfies n k 2m.

7 Syndromes for 2EC BCH code Consider any error pattern of weight 2: e(x) = x i 1 +x i 2, 0 i 1 < i 2 < n Syndrome of r(x) has two partial syndrome components, S 3 : = r(α) = e(α) = α i 1 +α i 2 S 3 = r(α 3 ) = e(α 3 ) = α 3i 1 +α 3i 2 = (α i 1 ) 3 +(α i 2 ) 3, S 3 are known quantities in this system of equations; they are computed at run time. Partial syndromes can be computed by multiply-accumulate circuits. α α 3 m m r 0,...,r n 1 r 0,...,r n 1 S 3 Unknowns i 1 and i 2 appear in the exponent equations are transcendental.

8 2EC BCH decoding: change of variable Let us simplify the problem by a change of variables: X 1 = α i 1, X 2 = α i 2 This yields a system of algebraic equations of degree 3: = α i 1 +α i 2 = X 1 + X 2 S 3 = α 3i 1 +α 3i 2 = X1 3 + X3 2 X 1 and X 2 are called the error locators. Error locations are obtained from X 1,X 2 by taking logarithms to base α: i 1 = log α X 1 i 2 = log α X 2 The logarithms can be calculated using lookup tables or sequential search.

9 2EC BCH decoding: degree reduction We can replace the equation of degree 3: S 3 = X3 1 +X3 2 X 1 +X 2 = X 2 1 +X 1 X 2 +X 2 2 = (X 1 +X 2 ) 2 +X 1 X 2 = S 2 1 +X 1X 2 Next rewrite the two equations for X 1,X 2 to obtain a system of degree 2: X 1 +X 2 = X 1 X 2 = S 3 +S 2 1 = S 3 +S 3 1 In other words, from the partial syndromes,s 3 (known quantities) we can compute the sum and product of the unknowns X 1,X 2. We have obtained the coefficients of a quadratic polynomial whose zeroes are X 1,X 2.

10 2EC BCH decoding: polynomial equation Given the sum and product of X 1,X 2, we can construct the quadratic polynomial Λ R (x) whose zeroes are X 1 and X 2 : Λ R (x) = (x X 1 )(x X 2 ) = x 2 (X 1 +X 2 )x+x 1 X 2 = x 2 + x+ S 3 +S 3 1 = x 2 +Λ 1 x+λ 2 If there are two errors, then X 1 X 2. Therefore and the coefficient Λ 1 = = X 1 +X 2 0 Λ 2 = S 3 +S 3 1 is well defined and nonzero because X 1 0 and X 2 0.

11 2EC BCH decoding: other cases When there are two errors, the error locators are the zeroes of Λ R (x). Bit error locations can be found by evaluating Λ R (α i ) for i = 0,1,...,n 1. Quadratic formula cannot be used in GF2 m. However, there are efficient methods for factoring quadratic polynomials over GF2 m. What if fewer than two errors? If one error, the error pattern e(x) = x i 1 has syndrome components: = e(α) = α i 1 = X 1 S 3 = e(α 3 ) = α 3i 1 = X 3 1 In this case, S 3 = S 3 1 and Λ 2 = 0. X 1 is zero of Λ R (x) = x. When there is no error, = S 3 = 0. Λ R (x) = 1 has no zeroes.

12 2EC BCH code: decoding procedure summarized 1. Compute partial syndromes: = r(α), S 3 = r(α 3 ). 2. If = 0 and S 3 = 0 then assume no error most plausible assumption. 3. If 0 and S 3 = S1 3 then assume 1 error. Error locator is X 1 =. Error location is i 1 satisfying X 1 = α i If 0 and S 3 S1 3 then assume 2 errors. Error locators are the zeroes in of Λ R (x), where Λ R (x) = x 2 + x + S 3 +S 3 1. If X 1 = α i 1, X 2 = α i 2 are zeroes of Λ R (x), errors are in locations i 1,i If Λ R (x) does not have 2 zeroes, or if = 0 and S 3 0, then 3 errors have occurred detectable but uncorrectable.

13 Reed-Solomon Codes Reed-Solomon codes use 2t equations to correct t errors. 0 = c 0 +c 1 α+c 2 α 2 + +c n 1 α n 1 0 = c 0 +c 1 α 2 +c 2 α 4 + +c n 1 α 2(n 1). 0 = c 0 +c 1 α 2t +c 2 α 4t + +c n 1 α 2t(n 1) The codeword symbols c i and α are m bits quantities, and n = 2 m 1. Multiplication and addition in these equations use Galois field arithmetic: every nonzero element has a reciprocal, so decoding equations work. Reed-Solomon codes are used in CDs, DVDs, and almost all disk drives.