It’s been a while since the genesis of this blog and its first post. Quite a shame I couldn’t keep up with the weekly blog commitment since the first week. At this point, I have just started my first research rotation with Professor Netanel Raviv to explore whether there exist better means of coding a perceptron with improved minimum and relative distance metrics than a parity code. To help my understanding of the topic which is rooted in Coding Theory, I read through several chapters of Ron Roth’s textbook “Introduction to Coding Theory” and in this blog I will note what I found most interesting and relevant to my research rotation.
The goal is to better grasp linear codes. But before I cover linear codes, I will first cover some basic concepts of channel coding. Figure 1.1 is a graphic of the overview of communications system that transmits a message from a Source to a Destination. The Source Encoder serves two roles: (1) translating the output of the source to the input to the channel and (2) compressing the source output, if needed. The compression determines whether a channel is lossy or lossless. What’s to note is that due to physical limitations of communications systems in real life, the message received may (usually) differ from the message transmitted due to noise.
Channels will be modeled as a discrete probabilistic channel where the channel encoder takes in an information word (message) and outputs a codeword which the channel outputs as a received word, then decoded by the channel decoder as a decoded codeword and decoded information. The goal of the system is to make sure that the decoded information word matches the information word and the decoded codeword matches the codeword.
We assume that it is more likely to less errors than more errors and errors are introduced to each element of a message independently and with a fixed probability p < 1/2. The Hamming distance is the metric of choice to measure to “number of errors”. The Hamming distance is defined as the number of different elements between two words. The distance
d of a
(n,M) code is the minimum Hamming distance between any two of its codewords. The Hamming weight of a codeword is the number of non-zero entries. This is equivalent as the Hamming distance of a codeword with the zero codeword.
Appropriately, the code is called an
(n,M,d) code where the
n, M parameters are the code length and code size, respectively. By nearest neighbor decoding, received words are decoded to its closest codeword (codeword of the minimum Hamming distance with the received word). Thus, the larger the minimum distance of a code, the better.
Now on to the topic of linear codes. Linear codes are defined as a code
C where a linear combination of any two codewords is also a codeword in
C. In short, a linear code is a vector subspace. The dual code of
C is the orthogonal complement of
C in its alphabet subspace. A linear code is written as a
[n,k,d] code where
k is the dimension of the code. A linear code has a nonunique
k x n generator matrix
G whose rows form a basis of the code. The parity check matrix of
C is a generator matrix for the dual code of
C. A systematic generator matrix is a
(I | A) matrix and these exist when the first
k columns of any generator matrix of
C is linearly independent. Systematic generator matrices are what simplify encoding and decoding of linear codes.
The source encoder maps an information word
u to a codeword
c by matrix multiplication with the code’s generator matrix
c = uG. For a systematic generator matrix, this is equivalent to
c = (u | uA) where
A is a
k x (n-k) matrix.
Decoding methods rely on a standard array for a code which is a 2-D array in which the first row consists of the codewords of
C starting with the zero codeword. Subsequent rows (cosets) start with a word of smallest Hamming weight
e followed by words
e + c. Decoding takes the following steps: (1) Find the row that contains received word
y. (2) The codeword
c is the first entry in the column containing
y. Step 2 is effectively computing
c = y - e.
One decoding method is syndrome decoding. A syndrome
s of a word
y is defined as
s = HyT. It is important to note that since a code
C is a kernel of
H, the syndrome of any codeword in
C is the zero vector. Syndrome decoding takes two steps: (1) Finding the syndrome
s of the received word
y. (2) Finding the minimum-weight word
e such that
s = HeT. This is computationally intractable and requires a look-up table unless the redundancy
n-k is small.
One of the advantages of working with linear codes is that everything boils down to linear algebra. A code can be describe using its basis, the code’s distance is its weight, and these together make the mapping of a message to code (and vice-versa) simple. This is the main reason why linear codes are the most prevalent codes used today.