\documentclass[10pt]{article}
\usepackage{fullpage}
\usepackage{amsfonts}
\usepackage{amssymb}
%\usepackage{epsfig}
\begin{document}
\input{preamble.tex}
\lecture{1}{Feb 7, 2007}{Madhu Sudan}{Brian Jacokes}
%%%% body goes in here %%%%
\section{Administrative Information}
\begin{itemize}
\item Lecturer: Madhu Sudan (madhu@mit.edu)
\item TA: Swastik Kopparty (swastik@mit.edu)
\item Website: \verb+http://theory.csail.mit.edu/~madhu/ST07/+
\end{itemize}
Make sure you're on the mailing list for the course, if you're already
on it you should have received an e-mail on February 1st. The first
problem set is already posted and is due in two weeks. \\
This is an advanced graduate course. It has no text, so we make our
own notes instead. As part of that, every student must scribe at
least once - this is a requirement to pass the course. Sign up for
scribing on the website. Email your scribed notes within 24 hours of
the lecture; they do not need to be too polished. \\
The grading for the class will be based on the following:
\begin{itemize}
\item 3 problemsets - First one is out today
\item 1 project - Work in pairs, read a paper and present it to the
class with any additional progress made. Understanding of the
material is essential.
\item 1 scribing
\item Class participation - Will take place through e-mails between
you and the staff
\end{itemize}
\section{Goals of 6.841}
6.841 assumes that you know the material presented in 6.840. The
first goal of the class will be to identify interesting computational
problems. This is obviously a subjective choice, so our definition of
``interesting'' will be one which is informed by mathematics. For
example, a problem may be interesting if many other problems can be
reduced to one of its instances, or if it is particularly
representative of a larger class of problems. \\
Once we find an interesting problem, we want to find out how much time
and space suffice to solve the problem, and how much are necessary to
solve the problem. If we're unable to find an upper or lower bound
our alternative is to compare the problem to other existing problems,
forming classes and hierarchies of problems. \\
\subsection{Interesting problems}
The following three problems are presented as ``interesting.''
\begin{itemize}
\item \#SAT (``number-SAT''): Given a 3CNF formula $\phi$ on $n$
variables $x_1,\ldots,x_n$ with $m$ clauses $c_1,\ldots,c_m$ (so $\phi
= c_1 \wedge \ldots \wedge c_m$ and each $c_i$ looks something like
$x_{i_1} \vee \bar{x_{i_2}} \vee x_{i_3}$), count the number of
satisfying assignments.
\item CNF minimization: Given a CNF formula $\phi$ with $m$ clauses
and an integer $m' < m$, does there exist a CNF $\psi$ with at most
$m'$ clauses such that $\psi \equiv \phi$? (We say that $\psi \equiv
\phi$ if, for each assignment $a$ of $x_1, \ldots, x_n$, we have
$\psi(a) = \phi(a)$.)
\item Permanent: Given an $n \times n$ matrix $A=\left[a_{ij}\right]$
of integers, compute the \emph{permanent} of $A$. We define $perm(A)
= \Sigma_{\pi:[n]\rightarrow[n]} \left(\Pi_{i=1}^{n} a_{i
\pi(i)}\right)$, where $[n]=\{1,\ldots,n\}$ and $\pi$ represents a
permutation (so $i \neq j$ implies $\pi(i) \neq \pi(j)$). This is
basically the formula for the determinant of a matrix, without the
power of -1 derived from the sign of each $\pi$.
\end{itemize}
We can now see how these problems fit in with our definition of
interesting. To begin with, we can claim that CNF minimization is
interesting. CNF minimization arises naturally in circuit design, for
instance, if one is able to come up with a straightforward (but large)
circuit that solves a problem and then wishes to find a smaller
circuit that would achieve the same goal. It is also symptomatic of a
large class of problems, which is to say that many other problems look
similar or are reducible to CNF minimization. Finally, we may relate
CNF minimization to other problems: if P=NP then it turns out that CNF
minimization would also be in P, although it is intrinsically
different from other satisfiability problems (CNF minimization $\neq$
SAT). \\
As it turns out, the other two example problems are reducible to each
other. The fact that \#SAT $\equiv$ Permanent was proved by Valiant,
who stated that counting is an interesting ``phenomenon.'' To that
end, he created a complexity class called ``Algebraic-NP'' such that
Permanent $\in$ Algebraic-NP-complete. Furthermore, Lipton found that
the worst-case instances of Permanent reduce to random instances,
implying that the worst-case complexity of the problem is at most the
average-case complexity. These problems were also related to CNF
minimization by Toda, who found that CNF minimization $\le$
(\emph{reduces to}) \#SAT. \\
A final note on this topic concerns the paper ``The Complexity of
Theorem Proving Procedures'' by Stephen Cook, which first introduced
the concept of NP-completeness. Despite having a great impact on
today's complexity theory, the paper was almost rejected when it was
first submitted for publication. This demonstrates how recognizing a
concept as interesting is far from an exact science and is quite often
retrospective, as even one of the most important papers ever released
on complexity theory wasn't recognized as interesting at first glance. \\
\section{Technical Definitions}
Now we'll give a few precise definitions that will be used throughout
the course. To begin with, we'll be dealing with what we call
\emph{computational problems}. One definition of a computational
problem is that we are given a function $f$ which maps an input $x$ to
an output $f(x)$, leading to the problem of finding $f(x)$ given $x$.
A slightly broader definition deals with problems where more than one
output is possible for each input: given a relation $R \subseteq
\{0,1\}^* \times \{0,1\}^*$ and an input $x$, produce an output $y$
such that $(x, y) \in R$. \\
We also consider \emph{languages}, from which \emph{decision problems}
arise. A language is defined by $L \subseteq \{0,1\}^*$ and the
associated decision problem is that we are given $x$ and wish to
decide if $x \in L$. Such problems will be the primary focus of our
course for a few main reasons. First of all, it is much easier to
find comparisons and reductions between decision problems than between
computational problems. For example, it is not clear at all how to
reduce an arbitary counting problem to another unless they are
sufficiently similar, whereas many seemingly unrelated decision
problems often end up being reducible to each other. In addition,
many comparison problems have analogous languages, which sometimes
allows us to use our findings in decision problems to form broader
results about computational problems. \\
Our principle way of comparing problems will be through
\emph{reductions}. We write that $\pi_1 \leq \pi_2$ to mean that the
problem $\pi_1$ is ``no harder than'' the problem $\pi_2$. This is
done by using a reduction from $\pi_1$ to $\pi_2$, although there is
still some freedom as to how we define such a reduction. In the
specific setting where $\pi_1$ and $\pi_2$ are decision problems,
characterized by languages $L_1$ and $L_2$ respectively, we can use
what is called a ``Karp reduction'' or ``many-one reduction.'' A Karp
reduction is written as $L_1 \leq _{Karp} L_2$ and is given by a
function $A : \{0,1\}^* \rightarrow \{0,1\}^*$ which is both time- and
space-efficient to compute, and for which $x \in L_1$ if and only if
$A(x) \in L_2$. \\
The Karp reduction is very specific, so more generally we can use a
Turing reduction, which can be thought of as adaptable computation.
Such a reduction can solve $\pi_1$ by using a subroutine which solves
$\pi_2$, but also performing computations before, after, and between
the subroutine calls. We can see that all Karp reductions also fall
into the category of being Turing reductions, but one example of the
advantage of Turing reductions over Karp reductions is that they can
equate (at least if P=NP) the SAT and $\overline{\mbox{SAT}}$
problems, whereas this is not possible using a Karp reduction. \\
A final concept which is used in the course is that of the
``complexity zoo.'' This is generally drawn as a circle with lines
demarcating different complexity classes [\emph{Scribe's note}: I'm
horribly afraid to try this using my touchpad mouse], and an arrow
from one class to another indicating that the first class is no harder
than the second class. Examples of complexity classes include L
(log-space problems, important in streaming), which is a subset of P
(poly-time problems), which is a (proper?) subset of NP
(non-deterministic poly-time problems), and so on. Some problems,
including the Permanent and CNF minimization problems we discussed
earlier, don't have very tight positions within this figure. For
instance, the Permanent problem is in PSPACE (poly-space problems) but
not in NP or coNP, even though there is a very large gap between these
classes of problems. \\
\section{Course Overview}
The first three weeks of the course will be spent on provable lower
bounds. We won't be spending too much time on this topic because
there is an embarassingly small number of good lower bounds on
problems. \\
Although it's easy to prove time and space hierachies
(e.g. $\mbox{TIME}(n^2) \subsetneq \mbox{TIME}(n^{10})$ or
$\mbox{SPACE}(n^3) \subsetneq \mbox{SPACE}(n^{6})$), it's difficult to
compare the two resources. For instance, we know that $\mbox{TIME}(t)
\subseteq \mbox{SPACE}(t) \subseteq \mbox{TIME}(2^t)$, but although
the first and last classes are not equal we don't know which
containment is proper. Likewise we can't prove proper containments in
the relations $\mbox{L} \subseteq \mbox{P} \subseteq \mbox{PSPACE}$ or
$\mbox{TIME}(t) \subseteq \mbox{NTIME}(t) \subseteq \mbox{TIME}(2^t)$,
even though they are there. Along this line, it has been shown that
at least one of the following statements is true: SAT $\not \in$
Linear-time (which is defined as $\bigcup_{c > 0} \mbox{TIME}(n
(\log{n})^c)$) or SAT $\not \in$ LOGSPACE. A tool which is helpful in
this problem is a resource called \emph{alternation}, which will also
help us with CNF minimization. \\
We'll use alternation along with several other resources in our study
of complexity theory. Over the course of the term we'll deal with
interaction, randomness, proofs, counting problems, and average case
complexity. To complete the term we'll investigate quantum computing,
which goes against the Turing thesis and is therefore very relevant to
our study of computation. \\
\end{document}