\documentclass[10pt]{article}
\usepackage{amsmath,amssymb}
\newtheorem{define}{Definition}
%\newcommand{\Z}{{\mathbb{Z}}}
%\usepackage{psfig}
\oddsidemargin=0.15in
\evensidemargin=0.15in
\topmargin=-.5in
\textheight=9in
\textwidth=6.25in
\begin{document}
\input{preamble.tex}
%{lecture number}{lecture date}{Ronitt Rubinfeld}{Your name}
\lecture{1}{February 6, 2008}{Ronitt Rubinfeld}{Shubhangi Saraf}
%%%% body goes in here %%%%
\section{Introduction}
Today we'll discuss a classical example of how randomness helps in
algorithms. We'll present a randomized algorithm for the problems of
Polynomial Zero Testing and Polynomial Identity Testing and then
give two very nice applications of these results.
\section{The problems}
The problem of \emph{Polynomial Identity Testing} (PIT) is the
following. Given two polynomials $p$ and $q$ in $n$ variables, we
want to determine if they are the same, i.e.\ is
$p(x_1,x_2,\ldots,x_n) = q(x_1,x_2,\ldots,x_n)$ for all $x_1 \ldots
x_n$. For example, consider
$$(x_1+x_2)(x_3+x_4)^{40}(x_5^2+x_6) {\buildrel ? \over \equiv}
(x_1-x_2)(x_3-x_4)^{40}(x_5^2 + x_6) +
(x_1+2x_2)(x_3-x_4)^{40}(x_5^2+x_6) + x_1x_3x_5^2 +
x_2x_6x_3^{40}.$$
We would like to figure out if the polynomial on the left is
identically equal to the polynomial on the right. The problem with
opening up the brackets and expanding is that we could possibly get
an exponential (in the degree of the polynomial) number of terms.
For instance, if there are $n$ variables and the degree of $d$,
there could be more that $\binom{n}{d}$ terms!
Consider the following related problem of \emph{Polynomial Zero
Testing}. Given a polynomial $p(x_1,\ldots,x_n)$, we want to
determine if it's identically $0$, i.e.\ is $p(x_1,x_2,\ldots,x_n) =
0$ for all $x_1 \ldots x_n$.
Observe that both the above problems are equivalent. Clearly
Polynomial Zero Testing is a special case of Polynomial Identity
Testing when $q$ is the zero polynomial. In the other direction, to
determine if $p(x_1,x_2,\ldots,x_n) = q(x_1,x_2,\ldots,x_n)$,
instead consider $p'(x_1,x_2, \ldots,x_n) \doteq (p-q)(x_1,x_2,
\ldots,x_n)$. Then $p' \equiv 0$ iff $p \equiv q$, and hence we can
instead test if $p'$ is identically $0$.
Before we discuss the above problems, let us clarify some notions.
Assume that the \emph{domain is a field} such as $\mathbb R$ or
$\mathbb Z_p$. For example, if the field is $\mathbb Z_7$, then
$$(x+3)^2 \equiv x^2 + 6x+9 \equiv x^2+6x+2 \qquad \pmod 7.$$
The \emph{degree of a univariate polynomial} is the highest exponent
of a term in the polynomial. For instance, the degree of
$x^{10}+x^3+1$ is $10$. The \emph{total degree of a multivariate
polynomial} is the max over all terms of the sum of degrees in the
term. This is the notion we'll tend to use most of the time unless
specified otherwise. The \emph{maximum degree of a multivariate
polynomial} is the maximum over all terms of the degree of the maximum
degree variable in that term. For example, if the polynomial is
$xy^2 + x^2$, then the total degree is $3$, and the maximum degree is
$2$.
\section{Algorithm for Polynomial Zero Testing}
Assume a polynomial $p$ in $n$ variables and of degree $d$ is given
as a black-box oracle. The oracle takes as input $\bar x =
(x_1,\ldots,x_n)$ and outputs $p(\bar x)$.
\paragraph{Deterministic Algorithm for the Case when $p$ is
Univariate.} Plug in any $d+1$ distinct values to the black-box. If
all are $0$, then output ``${\equiv} 0$''. Else, output ``${\not \equiv}
0$''.
The above algorithm makes $O(d)$ evaluations. It works since a
non-zero polynomial of degree at most $d$ can have at most $d$
roots.
\paragraph{Randomized Algorithm for the Univariate Case.}
The above observation also implies that if the domain field size is
greater than $2d$, then at most $d/|F| \leq 1/2$ of the fraction of
all field elements are zeroes of the polynomial. This suggests the
following randomized algorithm. Pick an element uniformly at random
from a field of large enough size and plug it in. If it outputs $0$,
then output ``${\equiv} 0$''. Else, output ``${\not \equiv}0$''. The
algorithm works in $O(1)$ evaluations. The behavior of the algorithm
is as follows. If $p \equiv 0$, the algorithm outputs $` \equiv 0'$.
If $p \not \equiv 0$, $\Pr[\hbox{algorithm outputs ``${\not \equiv}
0$''}] \geq 1/2$. This same idea is generalized to give an algorithm
that works in the multivariate case.
\paragraph{Randomized Algorithm for the Multivariate Case.}
Observe that a multivariate polynomial can have infinitely many
roots - for example: $p(x,y) = x \cdot y$ over a field of infinite
size has infinitely many roots. However, the same kind of test that
works in the univariate case works in this case too as long as the
field is of large enough size, and a random value is picked for each
$x_i$.
Let us first set up some notation that we'll use for the rest for
the course.
\begin{itemize}
\item $x \in_R S$ denotes ``Pick $x$ uniformly from S''.
\item $x \in_U S$ denotes ``Pick $x$ uniformly from S''.
\item $ x \in \mathcal D$ denotes ``Pick $x$ according to distribution
$\mathcal D$''.
\item $x \in_{\mathcal D} S$ denotes ``Pick
$x$ according to distribution $\mathcal D$ on set $S$''.
\end{itemize}
\noindent The following is the algorithm that works in the
multivariate case (needs $|F| \geq 2d$):
\begin{enumerate}
\item Pick $S \subseteq F$ arbitrarily such that $|S| \geq 2d$.
\item Pick $x_1,x_2, \dots,x_n \in_R S$.
\item If $p(x_1, \ldots, x_n) = 0$, output ``${\equiv} 0$''; else output
``${\not\equiv}0$''.
\end{enumerate}
Again, the algorithm needs just $O(1)$ evaluations of $p$ on numbers
of size $O(\log d)$.This is shown be the below claim, which can be proved by induction on $n$, with the univariate case as the base case.
\begin{claim}
\noindent If the polynomial $p \not \equiv 0$, then
$\Pr [p(x_1,\ldots,x_n) = 0] \leq \frac{d}{|S|}$.
\end{claim}
The behavior of the algorithm is as follows. If $p \equiv 0$, the
algorithm outputs ``${\equiv} 0$''. If $p \not \equiv 0$,
$\Pr[\hbox{algorithm outputs ``${\not \equiv} 0$''}] \geq 1/2$.
This probability can be improved by either repeating the algorithm multiple times, or
picking a bigger field size.
\section{Can we derandomize PIT?}
Assume that the polynomial is given as an arithmetic circuit.
Kabanets and Impagliazzo proved that if we can derandomize
polynomial identity testing, then one of the two statements given
below must be true:
\begin{enumerate}
\item NEXP is not contained in P/poly.
\item Permanant is not computable by polynomial-size arithmetic circuits.
\end{enumerate}
Hence derandomizing PIT would be a very powerful result.
\section{Applications of PIT}
\subsection{The man on the moon problem}
Assume Alice on the earth has a string $\bar a = a_1a_2\ldots a_n$
and Bob on the moon has a string $\bar b = b_1b_2 \ldots b_n$. They
want to decide if $\bar a = \bar b$.
One way of doing this would be for Alice to transmit $\bar a$ all
the way to the moon. This would take $n$ bits of transmission.
Alternatively, Alice could view $\bar a$ as the coefficients of a
degree $n$ univariate polynomial $p = \sum a_ix^i$ over a field of
size at least $2n$. Then, pick arbitrarily $S \subseteq F$ such that
$|S| \geq 2n$, pick $x \in_R S$, and send $x$ and $\bar a(x)$ to Bob
(the man on the moon). He checks if $\bar a (x) = \bar b (x)$. If
yes, they conclude that $\bar a = \bar b$. Else they're different.
By our previous analysis, the above algorithm indeed works. If $\bar
a = \bar b$ then they will always conclude that they are equal, and
if $\bar a \neq \bar b$, then with probability at least a half, they
would find that out. In the above randomized algorithm, only $O(\log
n)$ bits are needed to be transmitted.
\subsection{Bipartite Matching}
We'll first introduce some preliminaries. A \emph{bipartite graph}
$G = (V,E)$ is one in which the set of vertices $V$ can be
partitioned into two sets $S$ and $T$ such that all the edges in $G$
go between $S$ and $T$, and there are no edges with both endpoints
lying in the same set. A \emph{matching} $M$ is a subset of the set
of edges $E$ such that no two edges in $M$ share a vertex. A
\emph{perfect matching} is one that has an edge adjacent to each vertex.
A natural question that arises in graph theory is that given a
graph, decide if it has a perfect matching. The problem of actually
finding a perfect matching in a graph that has one, can be solved
using network flows. The result we'll show here is not as strong,
but is a first step in the direction of getting the best algorithm
for finding a bipartite matching. We'll show how to use Polynomial
Zero Testing to give a randomized algorithm for determining if a
given graph has a perfect matching.
Given a graph $G = (V,E)$, the \emph{Frobenius matrix} $A_G =
[a_{ij}]$ of the graph is a matrix whose entries are variables or
zeroes. In particular, $a_{ij} = x_{ij}$ if $(i,j) \in E$, and
$a_{ij} = 0$ otherwise.
\begin{claim}
A graph $G$ has a perfect matching iff
$\det(A_G) \neq 0$.
\end{claim}
\begin{proof}
$$\det(A_G) = \sum_{\sigma \in S_n} {\rm sgn}(\sigma) \prod_{i = 1}^n
a_{i{\sigma_i}}$$
Here $S_n$ denotes the set of all permutations of $1$ through $n$.
Observe that every permutation represents a possible matching
between the two vertex sets.
Also, $\prod_{i = 1}^n a_{i{\sigma_i}} \not\equiv 0$ if and only if all the edges corresponding the
the perfect matching represented by the permutation $\sigma$ are
present in $G$. Since every other term has a different combination
of variables, the non-zero terms don't get canceled out. This
completes the proof of the above claim.
\end{proof}
Since the determinant of $A_G$ is a multivariable polynomial of
degree $n$, it can be tested by the Polynomial Zero Testing
algorithm to see if it's identically zero or not, and hence to
determine if the graph $G$ contains a perfect matching.
The above idea was used by Lov\'asz to determine if a perfect matching
exists in a graph. Since then, it has been shown that similar
approaches can be extended to the non-bipartite case as well.
For instance, Mucha and Sankowski (FOCS 2004)
showed that one can find a maximum matching
in general graphs in time $O(n^{\omega})$,
which is the time required for matrix multiplication.
Nick Harvey at MIT (FOCS 2006) showed a simpler algorithm
with the same running time.
\end{document}