\documentclass{article}
\input{preamble.tex}
\newcommand{\TIME}[1]{\ensuremath{\mbox{TIME}(#1)}}
\newcommand{\SPACE}[1]{\ensuremath{\mbox{SPACE}(#1)}}
\newcommand{\setof}[1]{\ensuremath{\left\{ #1\right\}}}
\newcommand{\sizeof}[1]{\ensuremath{\left| #1\right|}}
\newcommand{\Sizeof}[1]{\ensuremath{\Big| #1 \Big|}}
\newcommand{\zo}{\ensuremath{\{0,1\}}}
\newcommand{\F}{\ensuremath{\mathbf{F}}}
\begin{document}
\lecture{19}{April 22, 1999}{Daniel A. Spielman}{Chris Peikert}
\section*{Loose Ends}
First, a quick proof that MA$(k) \subseteq$ AM: in the homework, we
proved that MA $\subseteq$ AM, except it was in the form of $\Sigma
\cdot \mbox{BP} \cdot \mbox{P} \subseteq \mbox{BP} \cdot \Sigma \cdot
\mbox{P}$. We can use essentially the same proof to show that MAM
$\subseteq$ AM. Now consider a language like MAMAM. Since MAM
$\subseteq$ AM, MAMAM $\subseteq$ MAAM = MAM $\subseteq$ AM. This
process can be done for arbitrarily long alternating strings, which
proves the claim.
See last year's notes for a proof that MA$(k(n)) \subseteq$ MA$(k(n)/2
+ 1)$.
\section*{New Topic: PSPACE $\subseteq$ IP(poly)}
As a warmup, we prove that \#P $\subseteq$ IP, and in particular, \#SAT
$\in$ IP, where we remain formal by treating IP as a function class
(instead of a language).
\subsection*{Arithmetization}
If we have a boolean formula $\phi(x_1,\ldots,x_n)$, then normally
each $x_i$ is either 0 or 1. We want to extend the meaning of the
formula to allow $x_i$ to take on arbitrary values from some field
$\F$. In the original formula, we have three operations: $\wedge,
\vee, \neg$. We make the following replacements in the formula:
\begin{eqnarray*}
a \wedge b & \rightarrow & a \cdot b \\
a \vee b & \rightarrow & 1 - (1-a)(1-b) \\
\neg a & \rightarrow & 1 - a
\end{eqnarray*}
Note that these replacements preserve De Morgan's laws. In addition,
if $a,b \in \zo$, then arithmetization does not change the value of
the function. Finally, notice that if we arithmetize a formula
$\phi(x_1,\ldots,x_n)$ to a polynomial $p(x_1,\ldots,x_n)$, then the
degree of $p$ is at most the number of leaves in $\phi$, when
interpreted as a tree.
Given a 3CNF $\phi$, define \#$\phi$ to be the number of satisfying
assignments of $\phi$, which equals
\[
\sum_{x_1 \in \zo} \cdots \sum_{x_n \in \zo} p(x_1,\ldots,x_n).
\]
Now given $a_1,\ldots,a_n \in \F$, a verifier can compute
$p(a_1,\ldots,a_n)$. In our interactive proof, the prover will assert
what \#$\phi$ is, and the verifier will end up checking a single value
$p(a_1,\ldots,a_n)$, using intermediate subassertions.
To make this notion of ``intermediate subassertions'' formal, we
define an $\epsilon$-step to be a sub-protocol/proof such that,
beginning with an assertion $A$,
\begin{itemize}
\item either V rejects, or
\item we end with an assertion $B$ having the following properties:
\begin{eqnarray*}
A \mbox{ is true} & \Rightarrow &
\exists \mbox{ a prover s.t. } B \mbox{ is true} \\
A \mbox{ is false} & \Rightarrow &
\Pr{B \mbox{ is false, given V didn't reject}} > 1 - \epsilon.
\end{eqnarray*}
\end{itemize}
In other words, if V ``believes'' $B$ with probability $\alpha \approx
1$, then V believes $A$ with probability $\alpha - \epsilon$.
Now assume that we begin with the following assertion:
\begin{equation}
\sum_{x_j \in \zo} \cdots \sum_{x_n \in \zo}
p(a_1,\ldots,a_{j-1},x_j,\ldots,x_n) = \alpha \label{one}
\end{equation}
The prover sends a polynomial $q(x_j)$ in $x_j$ only, with $\deg(q)
\leq \deg(p)$, which is supposedly
\begin{equation}
\sum_{x_{j+1} \in \zo} \cdots \sum_{x_n \in \zo}
p(a_1,\ldots,a_{j-1},x_j,x_{j+1},\ldots,x_n). \label{two}
\end{equation}
Then V rejects automatically if $q(0) + q(1) \neq \alpha$, otherwise V
chooses $a_j \in \F$ at random, lets $\beta = q(a_j)$, and sends $a_j$
to the prover. The new assertion is then
\begin{equation}
\sum_{x_{j+1} \in \zo} \cdots \sum_{x_n \in \zo}
p(a_1,\ldots,a_j,x_{j+1},\ldots,x_n) = \beta. \label{three}
\end{equation}
We claim that this is a $\frac{\deg(p)}{|\F|}$-step. To prove this,
note that three assertions are involved (they are marked above). If
(\ref{one}) is true, then there exists some $q(x_j)$ with proper
degree and some prover that sends it, so (\ref{two}) is true. Of
course, if (\ref{one}) and (\ref{two}) are true, then so is
(\ref{three}).
Alternatively, if (\ref{one}) is false and V does not reject, then
(\ref{two}) must be false. In this case, because $q$ and $p$ have
degree at most $\deg(p)$, there are at most $\deg(p)$ choices for
$a_j$ such that (\ref{three}) is true (by Schwartz's lemma), so the
probability of choosing one of these values is at most
$\frac{\deg(p)}{|\F|}$.
Now we present the full interactive proof for \#3SAT: the prover
begins with the assertion
\[
\sum_{x_1 \in \zo} \cdots \sum_{x_n \in \zo}
p(x_1,\ldots,x_n) = \beta_0
\]
which, via a $\frac{\deg(p)}{|\F|}$-step, is reduced to the assertion
\[
\sum_{x_2 \in \zo} \cdots \sum_{x_n \in \zo}
p(a_1,x_2,\ldots,x_n) = \beta_1
\]
which is further reduced, over $n$ similar steps, to the assertion
\[
p(a_1,\ldots,a_n) = \beta_n.
\]
The verifier takes polynomial time to check this assertion, and
accepts if it is true. The entire protocol is an $\frac{n \cdot
\deg(p)}{|\F|}$-step, by a probability bound of going from a false
assertion to a true one.
For the entire protocol, we can use $|\F| \approx 2^n$, and $\deg(p)$
polynomial in $n$.
For the proof that PSPACE $\subseteq$ IP(poly), note that V need not truly
evaluate $p$; it just needs to be convinced of the value. Recall the
outline of the proof that PSPACE $\subseteq$ AP: let M run
in polynomial space,
and see that the graph of configurations of M on $w$ has at
most $2^{n^k}$ nodes, for some $k$, and every node has outdegree 1
since M is deterministic. Then M accepts $w \iff
\mbox{FromTo}(\mbox{start},\mbox{accept},2^{n^k}) = 1$, where the
definition of FromTo is
\begin{equation}
\mbox{FromTo}(x,y,2^j) = \bigvee_z (\mbox{FromTo}(x,z,2^{j-1}) \wedge
\mbox{FromTo}(z,y,2^{j-1})) \label{ft}
\end{equation}
and where FromTo$(x,y,1)$ has a polynomial-sized formula.
Now we arithmetize: let $P_1(x_1,\ldots,x_{n^k},y_1,\ldots,y_{n^k})$
be an arithmetization of FromTo$(x,y,1)$. Furthermore, recursively
define
\[
P_k(\overline{x},\overline{y}) =
\sum_{z_1 \in \zo} \cdots \sum_{z_{n^k} \in \zo}
P_{k-1}(\overline{x},\overline{z}) \cdot
P_{k-1}(\overline{z},\overline{y}).
\]
We claim that for $x,y \in \zo^{n^k}$, $P_k(\overline{x},\overline{y})
= \mbox{FromTo}(\overline{x},\overline{y},2^k)$. We've arithmetized
the $\wedge$ of FromTos in (\ref{ft}) to multiplication, but we've
only replaced the $\vee_z$ by summation. However, this is OK because
of determinism - in the $\vee_z$, there is a {\em unique} node $z$ for
which both FromTos are true, if we define FromTo to require {\em
exact} lengths of paths, and make a loop from the accept node to
itself. Therefore the sum is either zero or one, depending on whether
there is a path of proper length to the accept node.
Now we outline the interactive proof of any PSPACE problem. The first
assertion is $P_{n^k}(\mbox{start},\mbox{accept}) = 1$. The general
assertion is of the form $P_j(x_1,\ldots,x_n,y_1,\ldots,y_n) = c$ with
$x_i,y_i \in \F$.
First, $P_{n^k}(\mbox{start},\mbox{accept}) = 1$ is written out a sum
over $z_i$'s. We interactively apply a $\frac{n \cdot
\deg(p)}{|\F|}$-step (as above) to get an assertion like
\[
P_{j-1}(x_1,\ldots,x_{n^k},a_1,\ldots,a_{n^k}) \cdot
P_{j-1}(a_1,\ldots,a_{n^k},y_1,\ldots,y_{n^k}) = c'
\]
which can be converted into another assertion
$P_{j-1}(\overline{u},\overline{v}) = c''$ by another $\epsilon$-step
(as we will show). We then repeat the process until we get an
assertion about $P_1$, which the verifier can then check directly.
To complete the proof, we need only show that an assertion about a
product $P_k \cdot P_k$ can be converted to an assertion about a
single value of $P_k$. Define a function that takes an element from
the field $\F$ and returns a vector,
\[
T(\alpha) = \alpha(x_1,\ldots,x_{n^k},a_1,\ldots,a_{n^k}) +
(1-\alpha)(a_1,\ldots,a_{n^k},y_1,\ldots,y_{n^k}).
\]
This can be viewed as ``linear interpolation'' (but in the field $\F$)
between the two vectors in the original product of $P_k$'s. Then the
prover sends a function $q(\alpha)$, which is supposedly the function
$P_k(T(\alpha))$. The verifier immediately rejects if $q(0) \cdot
q(1) \neq c'$, otherwise it chooses $b \in \F$ at random, and computes
$c'' = q(b)$. The new assertion is now $P_k(T(b)) = c''$. This is an
$\epsilon$-step, and the analysis is similar to what appears above.
This completes the proof.
\end{document}