\documentclass[11pt]{article}
\input{../preamble.tex}
\usepackage{amsmath}
\usepackage{amsfonts}
\newcommand{\Log}{{\rm log}}
\newcommand{\ep}{\epsilon}
\newcommand{\de}{\delta}
\newcommand{\bscp}{\mbox{BSC}_p}
\newcommand{\DD}{\mathcal{D}}
\newcommand{\Ex}{\mbox{Exp}}
\newcommand{\Ball}{\mbox{Ball}}
\begin{document}
\lecture{4}{September 16, 2002}{Madhu Sudan}{Amit J. Deshpande}
\section{Introduction}
The main topics included in today's lecture are:
\begin{itemize}
\item Singleton Bound
\item Reed-Solomon Codes
\item Multivariate Polynomial Codes
\end{itemize}
\section{Singleton Bound}
So far we have discussed Shannon's theory, Hamming's metric and
Hamming codes, and Hadamard codes. We are looking for aymptotically
good codes to correct some constant fraction of errors while still
trasmitting the information through the noisy channel at a positive
rate. In our usual notation of $[n, k, d]_q$-codes, Hamming's
construction gives a $[n, n - \Log_2 n, 3]_2$-code while Hadamard
codes were $[n, \Log_2 n, \frac{n}{2}]_2$-codes. But we are looking
for codes where $\frac{k}{n}$ and $\frac{d}{n}$ have a lower bound
independent of $n$ and $q$ is not growing to $\infty$. In this
scenario, we discuss the following simple impossibility result:
\begin{theorem}
For a code $C: \Sigma^k \rightarrow \Sigma^n$ with minimum distance
$d$, $n \geq k + d - 1$.
\end{theorem}
\begin{proof}
In other words, we want to prove that $d \leq n - (k-1)$. Just project
all the codewords on the first $(k-1)$ coordinates. Since there are
$q^k$ different codewords, by pigeon-hole principle at least two of
them should agree on these $(k-1)$ coordinates. But these then
disagree on at most the remaining $n - (k-1)$ coordinates. And hence
the minimum distance of the code $C$, $d \leq n - (k-1)$.
\end{proof}
\section{Reed-Solomon Codes}
\begin{definition}
Let $\Sigma = \F_q$ a finite field and $\alpha_1, \ldots, \alpha_n$ be
distinct elements of $\F_q$. Given $n, k$ and $\F_q$ such that $k \leq
n \leq q$, we define the encoding function for Reed-Solomon codes as:
$C: \Sigma^k \rightarrow \Sigma^n$ which on message $m = (m_0, m_1,
\ldots, m_{(k-1)})$ consider the polynomial $p(X) = \sum_{i=0}^{(k-1)}
m_i X^i$ and $C(m) = \langle p(\alpha_1), p(\alpha_2), \ldots,
p(\alpha_n) \rangle$.
\end{definition}
\begin{theorem}
Reed-Solomon code matches the singleton bound. i.e. it's a $[n, k,
n-(k-1)]_q$-code.
\end{theorem}
\begin{proof}
The proof is based only on the simple fact that a non-zero polynomial
of degree $l$ over a field can have at most $l$ zeroes.
For Reed-Solomon code, two codewords (with corresponding polynomials
$p_1$ and $p_2$) agree at $i$-th coordinate iff $(p_1 - p_2)(\alpha_i)
= 0$. But $(p_1 - p_2)$, from the above fact, can have at most $(k-1)$
zeros which means that the minimum distance $d \geq n - (k-1)$.
\end{proof}
Reed-Solomon codes are linear. This can be easily verified by the fact
that the polynomials of degree $\leq (k-1)$ form a vector space
(i.e. if $p_1, p_2$ are polynomials of degree $\leq (k-1)$ then
similarly are $\beta p_1$ and $p_1 + p_2$). Since the polynomials
$p(X) \equiv 1, p(X) = X, \ldots ,p(X) = X^{(k-1)}$ form the basis for
this vector space, we can also find a generator matrix for
Reed-Solomon codes.
\[ G = \left [ \begin{array}{cccc} 1 & 1 & \ldots & 1 \\ \alpha_1 &
\alpha_2 & \ldots & \alpha_n \\ : & : & : & : \\ {\alpha_1}^{(k-1)} &
{\alpha_2}^{(k-1)} & \ldots & {\alpha_n}^{(k-1)} \end{array} \right ]
\]
One can also prove the theorem about the minimum distance of
Reed-Solomon codes by using the fact that any $k$ columns of $G$ are
linearly independent (because $\alpha_i$'s are distinct and thus the
Vandermonde matrix formed by the $k$ columns is non-singular).
Using Reed-Solomon codes with $k = \frac{n}{2}$ we can get a $[n,
\frac{n}{2}, \frac{(n+2)}{2}]_q$-code which means that we can correct
much more erros than before. Typically Reed-Solomon codes are used for
storage of information in CD's because they are robust against bursty
errors that come in contiguous manner, unlike the random error model
studied by Shannon. Also if some information is erased in the
corrupted encoding, we can still retrieve the original message by
interpolating the polynomial on the remaining values we get.
A way to visualize Reed-Solomon code on binary alphabet is to consider
$q = n = 2^m$. This gives $C: \Sigma^k = {\F_2}^{k \Log_2 n}
\rightarrow \Sigma^n = {\F_2}^{n \Log_2 n}$ a $[n \Log_2 n, k \Log_2
n, n - (k-1)]_2$-code. And if you put $n \Log_2 n = N$ and $n - (k-1)
= 5$ then this gives a $[N, N - 4 \Log_2 N, 5]_2$-code. As we have
already seen, Hamming's construction gave a $[N, N - \Log_2 N,
3]_2$-code and Hamming's impossibility result said that for a $[N, k,
2t+1]_2$-code, $k \geq N - t \Log_2 N$. Reed-Solomon code don't
achieve this but give a fairly closer $k \geq N - 2t \Log N$ (We will
discuss later about BCH codes which match this bound).
\section{Multivariate Polynomial Codes}
Here instead of considering polynomials over one variable (like in
Reed-Solomon codes), we will consider multivariate polynomials. For
example, let's see the following encoding similar to Reed-Solomon
codes but using bivariate polynomials.
\begin{definition}
Let $\Sigma = \F_q$ and let $k = l^2$ and $n = q^2$. A message is
typically $m = (m_{00}, m_{01}, \ldots, m_{ll})$ and is treated as the
coefficients of a bivariate polynomial $p(X,Y) = \sum_{i=0}^{l}
\sum_{j=0}^{l} m_{ij} X^i Y^j$ which has degree $l$ in each variable.
The encoding is just evaluating the polynomial over all the elements
of $\F_q \times \F_q$. i.e. $C(m) = \langle p(x,y) \rangle_{(x,y) \in
\F_q \times \F_q}$.
\end{definition}
But what is the minimum distance of this code ? We will use the
following lemma to find it.
\begin{lemma}{\bf (Schwartz-Zippel lemma)}
A multivariate polynomial $Q(X_1, \ldots, X_m)$ (not identically $0$)
of total degree $L$ is non-zero on at least $(1 - \frac{L}{|S|})$
fraction of points in $S^m$, where $S \subseteq \F_q$.
\end{lemma}
\begin{proof-of-lemma}{:}
We will prove it by induction on the number of variables. For $m=1$,
it's easy as we know that $Q(X_1, \ldots, X_m)$ can have at most $L$
zeros over the field $\F_q$. Now assume that the induction hypothesis
is true for the a multivariate polynomial with upto $(m-1)$ variables,
for $m>1$. Consider
\[ Q(X_1, \ldots, X_m) = \sum_{i=0}^{t} {X_1}^{i} Q_{i} (X_2, \ldots,
X_m) \]
where $t \leq L$ is the largest exponent of $X_1$ in $Q(X_1, \ldots,
X_m)$. So the total degree of $Q_t (X_2, \ldots, X_m)$ is at most
$(L-t)$.
By induction hypothesis it implies that $Q_t (X_2, \ldots, X_m) \neq
0$ on at least $(1 - \frac{(L-t)}{|S|})$ points in $S^{(m-1)}$. But
suppose $Q_t (s_2, \ldots, s_{m}) \neq 0$ then $Q(X_1, s_2, \ldots,
s_{m})$ is a not-identically-zero polynomial of degree $t$ in $X_1$,
and therefore is non-zero on at least $(1 - \frac{t}{|S|})$ fraction
of choices for $X_1$.
So putting it all together, $Q(X_1, \ldots, X_m)$ is non-zero on at
least $(1 - \frac{(L-t)}{|S|})(1 - \frac{t}{|S|}) \geq (1 -
\frac{L}{|S|})$ points in $S^m$.
\end{proof-of-lemma}
Using this, for $m=2$ and $S=\F_q$, we get that the above bivariate
polynomial code is a $[q^2, l^2, (1-\frac{2l}{q})
q^2]_q$-code.
Some interesting cases of multivariate polynomial codes include
Reed-Solomon code ($l=k$ and $m=1$), Bivariate polynomial code
($l=\sqrt{k}$ and $m=2$) and Hadamard code ($l=1$, $m=k$ and $q=2$) !
\end{document}