Code Design for Dependable Systems

Chapter 9.3.2.2 - Design for SEC-Sb/pxbEL Codes: 2. Codes Designed by Odd / Even-Weight Column Square Matrices —

2. Codes Designed by Odd / Even-Weight Column Square Matrices —
Codes II—

Here the SEC-Sb/p×bEL codes designed by another method are presented and called type
II codes.

Figure 9.3 Check-bit lengths compared with information-bit lengths of the SEC-S4=p4EL type I codes.


Preliminaries

Definition 9.2   Let an odd-weight column square matrix be a nonsingular b × b matrix
whose columns are odd weight. Let an even-weight column square matrix be a b × b matrix
whose columns are b copies of an even-weight vector (including the zero vector).                 

Because there are 2b-1 even-weight column vectors having dimension b, there exist
2b-1 even-weight column square matrices.

Here, we show an example design method of nonsingular odd-weight column b × b
square matrices.

Definition 9.3  The matrix Mb shown below is defined as the matrix having b rows and
2b-1 odd-weight columns:

 

In Mb, α is a root of the (b −1)-th degree binary primitive polynomial g(x), is a co-
efficient vector of xi mod g(x), and pi Î {0, 1} is a bit determined to make the column
vector ai, i = 0, 1, × × × , 2b-1, be odd weight.                                                                 

Lemma 9.1     The following shows a nonsingular odd-weight column square matrix generated
from any consecutive b column vectors in the matrix Mb:

 

where ij i + j mod 2b-1, and 0jb1.

From this lemma we obtain 2b-1 nonsingular odd-weight column square matrices.

Design of the Parity-Check Matrices  Let A and B be the sets of the odd-weight
column square matrices and the even-weight column square matrices, respectively. The
following lemma provides the basic idea of the code design method.

Lemma 9.2     Consider two different vectors each having degree r (i.e., Q and Q') and
each being constructed of elements from the even-weight column square matrices and the
odd-weight column square matrices (i.e., Qj,Q'j Î A È B for j = 0, 1,× × × , r
− 1). In each
vector there exists at least one matrix included in A.

 

Assume that the i-th elements in Q and Q' (i.e., Qi and Q'i), respectively, are different
square matrices. In other words, Qi is an odd-weight column square matrix and Q'i is an
even-weight column square matrix, and vice versa. Then any summation of binary column
vectors in Q produces a different result than that given by any summation of binary column
vectors in Q'.


Proof   Let V and V', each having r b-tuples, be the results of the summation of column
vectors in Q and Q', respectively:

 

Without loss of generality, Qi and Q'i can be regarded as the odd-weight column square
matrix and the even-weight column square matrix, respectively. Assume vi = v'i. Then V is
the vector resulting from the summation of an even number of columns in Q, and V' is that
resulting from the summation of odd number of columns in Q'. There exists an odd-weight
column matrix in Q', say Q'l, in the l-th row for li. So v'i is an odd-weight b-tuple and vl
has even weight. Therefore, V V'.                                                                 Q.E.D.

Below we provide an example expressed as a submatrix Hi corresponding to the i-th
block. In other words, if we use the matrix from B in the first row, then we write in this
place, and so on:

 

In order to distinguish single-bit errors from single-byte errors, excluding single-bit errors,
every submatrix has at least one b × b identity matrix, which is included in A. In the
submatrix Hi, for example, this can be expressed as follows:

Theorem 9.8   The code described by the following matrix H is an SEC- Sb/p×bEL code
with check-bit length R = br and codeword length N bits:

 

where A and B are row vectors in which odd-weight square matrices and even-weight
square matrices are used, respectively, and the top A in each column is the row vector
including all b × b identity matrices. In this matrix each submatrix Hi includes different
pattern of A’s and B’s and has at least one A.

The code length (in bits) of the code above is expressed by the following equation:

 

where r is the number of check-bytes.

Proof   It is apparent that the code satisfies condition 1 of Theorem 9.4 for any single-bit
errors and any single-byte errors. The binary columns in H are distinct. Therefore condition
2 of Theorem 9.4 is satisfied as well. If ij, then matrices Hi and Hj have different
patterns of A’s and B’s. By Lemma 9.2, condition 3 of Theorem 9.4 is satisfied. Since
every submatrix has at least one identity matrix, condition 4 of Theorem 9.4 is satisfied.
Hence the code described by H is an SEC-Sb/p×bEL code.

The number of blocks is equal to that of the nonzero integers expressed by r-digit binary
numbers, meaning 2r − 1. The block length B is determined by the number of distinct
column vectors using r matrices of the odd-weight column square matrices and the even-
weight column square matrices. B is also determined by the condition that every Hi includes
at least one row of identity matrices. Based on this, the maximum block length B bits can be
expressed as b(2b-1)r-1 = b2(b-1)(r-1). Therefore Eq. (9.4) is valid.                            Q.E.D.

Example 9.2   [FUJI94]

For b = 4 and R = 8 the (96, 88) SEC-S4/8×4EL code is given by the following
matrix H:

 

Figure 9.4 shows this matrix expressed in binary form.

Figure 9.4 Binary formof the (96, 88) SEC-S4=84EL codes. Source: [FUJI94].  1994 IEEE.

Expanding the Code Length   The code design method of Theorem 9.8 says that
the check-bit length should be a multiple of the byte length b. This condition can be
relaxed by taking any check-bit length R > 2b, as shown by the following theorem.

Theorem 9.9   Let the following matrix H be a parity-check matrix of an SEC-Sb/p×bEL
code with a code length in bits b
(2r − 1)2(b-1)(r-1) and a check-bit length br:

 

where Hi, i = 0, 1,× × × , n − 1, is a submatrix of H. The code described by the following
matrix H' is an SEC-Sb/p×bEL code with check-bit length br + 1:

 

If this procedure is performed c times on H, then the code length in bits of the expanded
code with check-bit length R = br + c, 0 ≤ c < b can be expressed as follows:

 

Figure 9.5 illustrates the relation between the information-bit length and the check-bit length of the
SEC-Sb/p×bEL code determined by Theorems 9.8 and 9.9 for b = 4.

Figure 9.5 Check-bit lengths compared with information-bit lengths of the SEC-Sb=pbEL type II codes for b ¼ 4. Source: [FUJI94]. 1994 IEEE.

 

UNLIMITED FREE
ACCESS
TO THE WORLD'S BEST IDEAS

SUBMIT
Already a GlobalSpec user? Log in.

This is embarrasing...

An error occurred while processing the form. Please try again in a few minutes.

Customize Your GlobalSpec Experience

Category: ISO Containers
Finish!
Privacy Policy

This is embarrasing...

An error occurred while processing the form. Please try again in a few minutes.