\documentclass[12pt]{article}
\usepackage{epsfig}

 \textwidth 6in
\textheight 8.5in
\topmargin 0in
\oddsidemargin 0.25in
\evensidemargin 0.25in

%======================== CROSS-REFERENCING ==================
 \newcommand{\app}[1]{Appendix~\ref{sect.#1}}
 \newcommand{\sect}[1]{\S\ref{sect.#1}}      % Ref. to section or subsection
                                        %   to appear as, say, $2.37
 \newcommand{\baresect}[1]{\ref{sect.#1}}      % Ref. to one more section or
                                 %    subsection to appear as, say, 2.37
 \newcommand{\eq}[1]{Eq.~(\ref{eq.#1})}	% Ref. to equation
 \newcommand{\bareeq}[1]{(\ref{eq.#1})}	% Ref. to equation
 \newcommand{\fig}[1]{Fig.~\ref{fig.#1}}
 \newcommand{\barefig}[1]{\ref{fig.#1}}

 \newcommand{\sectlabel}[1]{\label{sect.#1}}
 \newcommand{\eqlabel}[1]{\label{eq.#1}}



% common number for a group of equations
% a better way would be for mathletters to work directly with all
%   equation, eqnarray, etc. environments rather than needing equationGroup...
\newcounter{eqletter} \setcounter{eqletter}{0}

\newenvironment{mathletters}{
\setcounter{eqletter}{0}
\refstepcounter{equation}% increment eqn number, make this ref for label
\renewcommand{\theequation}{\arabic{equation}\alph{eqletter}}
}{}

% within mathletters, use equationGroup to group equations with same number
\newenvironment{equationGroup}{
\addtocounter{equation}{-1} % undo increment from equation environment
\stepcounter{eqletter}
\begin{equation}
}{\end{equation}}


%======================== FIGURES ========================

\newcommand{\figdef}[3]{% label, body, caption
\begin{figure}[!htb]
 \centering\leavevmode#2%
 \caption{\small #3}
 \label{fig.#1}
\end{figure}                 }


%======================== COMMON EXPRESSIONS ========================
% boolean values -- command absorbs any following space, so must use e.g. \true\ ...
\newcommand{\true}{\mbox{\sc true}}
\newcommand{\false}{\mbox{\sc false}}


% number of 1-bits in an assignment; inner braces required e.g. for a^\ones{x y}
\newcommand{\ones}[1]{{| #1 |}}

% bitwise and operation
\newcommand{\bitA}{\wedge}

% W matrix with hat: unnormalized version of the matrix
\newcommand{\W}{\hat{W}}

% expectation value
\newcommand{\expect}[1]{% body
\left\langle #1 \right\rangle
}

% state vectors
\newcommand{\stateVec}[2]{ \pmatrix{ #1 \cr #2 \cr } }
\newcommand{\stateVecV}[2]{ \pmatrix{ #1 \cr \vdots \cr #2 \cr } }
\newcommand{\stateVecB}[4]{ \pmatrix{ #1 \cr #2 \cr #3 \cr #4 \cr } }

% ket
\newcommand{\ket}[1]{ { \left| #1 \right> }}

\newcommand{\muCrit} { {\mu_{\rm crit}} }
\newcommand{\cAvg} { { c_{\rm avg} }}
%\newcommand{\mMax} { { m_{\rm max} }}
\newcommand{\mMax} { M }
\newcommand{\Nproblems}{ { N_{\rm problems} }}
\newcommand{\Psoln}{ { P_{\rm soln} }}
\newcommand{\Psoluble}{ { P_{\rm soluble} }}

\newcommand{\Ns} { {N_s} }
\newcommand{\Nss} { {N_{s'}} }
\newcommand{\Nboth} { {N_{\rm both}} }
\newcommand{\Nother} { {N_{\rm other}} }



% scaled summation variables
\newcommand{\s}[1]{ \hat{#1} }

\newcommand{\sNs} { {\s{N}_s} }
\newcommand{\sNss} { {\s{N}_{s'}} }
\newcommand{\sNboth} { {\s{N}_{\rm both}} }
\newcommand{\sNother} { {\s{N}_{\rm other}} }

%=============================================
\begin{document}
\title{Single-Step Quantum Search Using Problem Structure}

\author{
Tad Hogg \\
Xerox Palo Alto Research Center \\
3333 Coyote Hill Road, Palo Alto, CA 94304 \\
hogg@parc.xerox.com
}

\maketitle

\begin{abstract}
The structure of satisfiability problems is used to improve search
algorithms for quantum computers and reduce their required coherence
times. The asymptotic average behavior of the these algorithms is
determined exactly, and used to identify the best algorithm from among
a class of methods that use only a single coherent evaluation of
problem properties. The resulting algorithm improves on previous
quantum algorithms for most random $k$-SAT problems, but remains
exponential for hard problem instances.  Compared to good classical
methods, the algorithm performs better, on average, for weakly and
highly constrained problems but worse for hard cases, indicating the
need to include additional problem structure in quantum algorithms.
\end{abstract}


\section{Introduction}

Quantum
computers~\cite{benioff82,bernstein92,deutsch85,deutsch89,divincenzo95,feynman86,lloyd93,rieffel98}
offer the possibility of faster combinatorial search by operating
simultaneously on all search states. For instance, quantum computers
can rapidly factor integers~\cite{shor94}, a problem thought to be
intractable for classical machines. A number of algorithms have been
proposed for combinatorial search
problems~\cite{boyer96,brassard98,cerf98,cerny93,grover96,grover97a,hogg97,terhal97}. In
particular, without making use of problem structure, quantum computers
result in a square root improvment in performance. This unstructured
search is the quantum equivalent of the classical generate-and-test
procedure and is the best improvement possible in this case. Further
improvement in performance requires using problem structure in the
search, just as with classical heuristics.

A major challenge in realizing the potential of quantum computers
is maintaining coherence over many computational
steps~\cite{landauer94a,unruh94,haroche96,monroe96a}. While this
difficulty seems unlikely to be a fundamental
limitation~\cite{berthiaume94,shor95,knill98} and small quantum
computations have been implemented~\cite{chuang98,chuang98a}, the
technical challenge of maintaining coherence can be reduced by
developing quantum algorithms that minimize the required coherence
time.

Thus an important open question is the extent to which problem
structure can not only improve the performance of quantum search but
also reduce the number of steps, and hence the required coherence
time. As an extreme case, we consider the performance of single-step
quantum search, i.e., algorithms that make use of only a single
evaluation of structure associated with a problem. Single-step search
is very effective for highly constrained problems~\cite{hogg98},
outperforming both unstructured quantum search and classical
heuristics in these cases.  By requiring only one step, it also needs
less coherence time than the exponentially many steps required by the
unstructured algorithm, and hence should be easier to implement. Can
the technique used for highly constrained problems be extended to the
more challenging case of hard search problems with an intermediate
number of constraints~\cite{cheeseman91,hogg95e}? Conversely, to what
extent does the restriction to a single step limit the extent to which
the capabilities of quantum computers can be used?

Addressing these questions requires determining the asymptotic
behavior of the algorithm. Algorithms that use problem structure, such
as classical heuristics, are often difficult to analyze theoretically
due to the complicated dependencies introduced by the characteristics
of the particular problem instance. Thus one is often forced to use
empirical evaluation, which is a major limitation for evaluating
quantum algorithms since classical simulations of quantum algorithms
have an exponential slowdown. By limiting the use of such
dependencies, a single-step search algorithm can be simpler to analyze
than multistep quantum searches.

This paper, demonstrates the potential of this approach for algorithms
that use the number of conflicts in each search state. Furthermore, we
show how to evaluate the asymptotic average scaling behavior of the
algorithm directly, rather than relying on empirical evaluation. This
evaluation is used to optimize the algorithm, and also demonstrates a
general technique for studying the average behavior of quantum search
algorithms.

In the remainder of this paper, we first summarize the satisfiability
search problem and then describe a class of one-step quantum
algorithms for this problem. This class includes both the previous
unstructured and highly constrained methods as special cases. In
\sect{analysis}, the average behavior of these algorithms is
determined. This result is then used to identify the best performing
algorithms for problems with differing degrees of constraint in the
following two sections. Finally we discuss the precision requirements
on the parameters used in the algorithms and some additional
properties of the search behavior.


\section{Satisfiability}\sectlabel{sat}

One well-studied compbinatorial search problem~\cite{garey79} is
satisfiability (SAT), which consists of a logical propositional
formula in $n$ variables $V_1,\ldots,V_n$ and the requirement to find
a value (\true\ or \false) for each variable that makes the formula
true. This problem has $N=2^n$ assignments. For $k$-SAT, the formula
consists of a conjunction of clauses and each clause is a disjunction
of $k$ variables, any of which may be negated. For $k \geq 3$ these
problems are NP-complete. An example of such a clause for $k=3$, with
the third variable negated, is $V_1$ OR $V_2$ OR (NOT $V_3$), which is
false for exactly one assignment for these variables: $\{V_1=\false,
V_2=\false, V_3=\true\}$. A clause with $k$ variables is false for
exactly one assignment to those variables, and true for the other
$2^k-1$ choices.  Since the formula is a conjunction of clauses, a
solution must satisfy every clause. We say an assignment conflicts
with a particular clause when the values the assignment gives to the
variables in the clause make the clause false.  For example, in a four
variable problem, the assignment $$\{V_1=\false, V_2=\false,
V_3=\true, V_4=\true\}$$ conflicts with the $k=3$ clause given above,
while $$\{V_1=\false, V_2=\false, V_3=\false, V_4=\true\}$$ does
not. Thus each clause is a constraint that adds one conflict to all
assignments that conflict with it. The number of distinct clauses $m$
is then the number of constraints in the problem.

The assignments for SAT can also be viewed as bit-strings with the
correspondence that the $i^{th}$ bit is 0 or 1 according to whether
$V_i$ is assigned the value false or true, respectively. In turn,
these bit-strings are the binary representation of integers, ranging
from 0 to $2^n-1$.  For definiteness, we arbitrarily order the bits so
the values of $V_1$ and $V_n$ correspond, respectively, to the least
and most significant bits of the integer. For example, the assignment
$$\{V_1=\false, V_2=\false, V_3=\true, V_4=\false\}$$ corresponds to
the integer whose binary representation is 0100, i.e., the number 4.

For bit-strings $r$ and $s$, let $\ones{s}$ be the number of 1-bits in
$s$ and $r \bitA s$ the bitwise AND operation on $r$ and $s$. Thus
$\ones{r \bitA s}$ counts the number of 1-bits both assignments have
in common. We also use $d(r,s)$ as the Hamming distance between $r$
and $s$, i.e., the number of positions at which they have different
values. These quantities are related by
\begin{equation}\eqlabel{hamming}
d(r,s) = \ones{r} + \ones{s} - 2 \ones{r \bitA s}
\end{equation}
Let $c(s)$ be the number of conflicts for assignment $s$ in a given
SAT problem.

An example 1-SAT problem with $n=2$ is the propositional formula (NOT
$V_1$) AND (NOT $V_2$).  This problem has a unique solution:
$\{V_1=\false, V_2=\false\}$, an assignment with the bit
representation 00. The remaining assignments for this problem have bit
representations 01, 10, and 11.

Theoretically, search algorithms are often evaluated for the worst
possible case. However, in practice, search problems are often found
to be considerably easier than suggested by these worst case
analyses~\cite{hogg95e}.  This observation leads to examining the
typical behavior of search algorithms with respect to a specified {\em
ensemble} of problems, i.e., a class of problems and a probability for
each to occur. A commonly studied ensemble is random $k$-SAT,
specified by the number of variables $n$, the size of the clauses $k$
and the number of distinct\footnote{This ensemble differs slightly
from other studies where the clauses are not required to be distinct.}
clauses $m$. A problem instance is created by randomly selecting $m$
distinct clauses from the set of all possible
clauses~\cite{nijenhuis78}. When $n$ is large, the typical behavior of
random $k$-SAT is determined by $\mu = m/n$, the ratio of clauses to
variables. In particular, for each $k$ there is a threshold value
$\muCrit$ on $\mu$ below which most random $k$-SAT problems are
soluble and above which most have no
solutions~\cite{crawford95,selman95}. For $k=3$, this value is
approximately $\muCrit=4.2$.

The quantum searches considered here are incomplete methods, i.e.,
they can find a solution if one exists but can never guarantee no
solution exists. For studying such algorithms, the ensembles would
ideally contain only instances with a solution. For example, we could
consider the ensemble of random {\em soluble} $k$-SAT, in which each
instance with at least one solution is equally likely to appear.
Unfortunately, this ensemble does not have a simple expression for the
number of problems as required for the analytic performance evaluation
given below. Instead, for $\mu < \muCrit$, most random problems indeed
have a solution so random $k$-SAT is useful for studying incomplete
search methods for underconstrained problems.

Randomly selected overconstrained problems usually have no solutions
so random SAT is not a useful ensemble when $\mu>\muCrit$. An
alternative with simple analytic properties is the ensemble with a
prespecified solution. In this case, a particular assignment is
selected to be a solution. Then the $m$ clauses are selected from
among those that do not conflict with the prespecified
solution. Compared to random selection among soluble problems, using a
prespecified solution is more likely to pick problems with many
solutions, resulting in somewhat easier search problems, on average.

Each clause in a $k$-SAT formula conflicts with exactly one of the
$2^k$ possible assignments for the variables that appear in the
clause. Thus the average number of conflicts in an assignment is
$\cAvg = m/2^k$. While this average is the same for {\em all} SAT problems
with given $m$ and $k$, the variance in the number of conflicts varies
from problem to problem. For the random ensemble, the average value of
the variance is $\cAvg (1-2^{-k})$.  Thus when $m \gg 1$, the relative
deviation decreases as $O(1/\sqrt{m})$ and hence most assignments have a
number of conflicts very close to the average.



\section{Quantum Search Algorithms}

Quantum computers use physical devices whose full quantum state can be
controlled. For example~\cite{divincenzo95}, an atom in its ground
state could represent a bit set to 0, and an excited state for 1. The
atom can be switched between these states and also be placed in a
uniquely quantum mechanical {\em superposition} of these values, which
can be denoted as a vector $\stateVec{\psi_0}{\psi_1}$, with a
component (called an {\em amplitude}) for each of the corresponding
classical states for the system. These amplitudes are complex numbers.

A quantum machine with $n$ quantum bits exists in a superposition of
the $2^n$ classical states for $n$ bits.  The amplitudes have a
physical interpretation: when the computer's state is measured, the
superposition randomly changes to one of the classical states with
$|\psi_s|^2$ being the probability to obtain the state $s$. Thus
amplitudes satisfy the normalization condition $\sum_s |\psi_s|^2 =
1$. This measurement operation is used to obtain definite results from
a quantum computation.

Quantum algorithms use operations that rapidly manipulate the
amplitudes in a superposition. Because quantum mechanics is linear and
the normalization condition must always be satisfied, these operations
are limited to unitary linear operators. That is, a state vector
$\psi$ can only change to a new vector $\psi^\prime$ related to the
original one by a unitary transformation, i.e., $\psi' = U \psi$ where
$U$ is a unitary matrix\footnote{A complex matrix $U$ is unitary when
$U^\dagger U = I$, where $U^\dagger$ is the transpose of $U$ with all
elements changed to their complex conjugates. Examples include
permutations, rotations and multiplication by phases (complex numbers
whose magnitude is one).} of dimension $2^n \times 2^n$. In spite of
the exponential size of the matrix, in many cases the operation can be
performed in a time that grows only as a polynomial in $n$ by quantum
computers~\cite{boyer96,hoyer97}. Importantly, the quantum computer
does not explicitly form, or store, the matrix $U$. Rather it performs
a series of elementary operations whose net effect is to produce the
new state vector $\psi'$. The components of the new vector are not
directly accessible: rather they determine the probabilities of
obtaining various results when the state is measured.

Search algorithms for SAT problems can make use of readily computed
properties of individual assignments, e.g., a test of whether a given
assignment is a solution. With quantum computers, these properties can
be evaluated simutaneously for all assignments. In this paper we focus
on algorithms that make use of this simutaneous evaluation just once.

\subsection{Single-Step Search}

Single-step methods could be implemented in a variety of ways. One
simple approach starts with an equal superposition of all the
assignments, adjusts the phases based on the number of conflicts in
each of the assignments, and then mixes the amplitudes from different
assignments. This algorithm requires only a single testing of the
assignments, corresponding to a single classical search step.

For a $k$-SAT problem with $n$ variables and $m$ clauses, the
algorithm takes the following form.  The initial state has amplitude
$\psi_s=2^{-n/2}$ for each of the $2^n$ assignments $s$, and the final
state vector is $\phi = U P \psi $ where the matrices $P$ and $U$ are
defined as follows. The matrix $P$ is diagonal with $P_{ss} =
p_{c(s)}$ depending on the number of conflicts $c$ in the assignment
$s$, ranging from 0 to $m$.

The mixing matrix is defined in terms of two simpler operations: $U =
W T W$.  The Walsh transform $W$ has entries
\begin{equation}\eqlabel{W}
W_{rs} = 2^{-n/2} (-1)^\ones{r \bitA s}
\end{equation}
for assignments $r$ and $s$ and can be implemented
efficiently~\cite{boyer96,grover96}.  The matrix $T$ is diagonal with
elements $T_{rr}=t_{\ones{r}}$ depending only on the number of 1-bits
in each assignment, ranging from 0 to $n$.  These definitions for $W$
and $T$ lead to a mixing matrix $U$ whose elements $U_{rs}=u_{d(r,s)}$
depend only on the Hamming distance between the assignments $r$ and
$s$~\cite{hogg98}, with
\begin{equation}\eqlabel{u}
u_d = 2^{-n} \sum_{h z} (-1)^z {d \choose z} {n-d \choose h-z} t_h
\end{equation}

Unlike previous algorithms, where the phase choices are restricted to
$\pm 1$ or powers of $i$, this algorithm potentially uses a different
phase choice for each number of conflicts and each number of 1-bits in
an assignment. Because the number of conflicts in a given assignment
is efficiently computable for SAT problems, these phase choices can be
efficiently implemented~\cite{hogg98b}.

This procedure defines a class of algorithms. A particular choice of
the phases $p_c$ and $t_h$ is required to complete the algorithm's
specification. For example, the choices $p_0=1$, $t_0=1$ and the
remaining phases set to $-1$ gives a single step of the unstructured
search algorithm~\cite{grover96}. Similarly, $p_c=i^c$ and $t_h=i^h$
gives an algorithm appropriate for maximally constrained 1-SAT
problems~\cite{hogg98}.


\subsection{Selecting Phase Values}\sectlabel{scaling}

To motivate the choice of phases appropriate for large $k$-SAT
problems, note that the number of assignments with $h$ 1-bits is $n
\choose h$. So for large $n$, most states have $h$ close to
$n/2$. Similarly, the number of assignments at Hamming distance $d$
from a given solution is $n \choose d$, which is also concentrated
near $d=n/2$. Because the algorithm starts with an equal amplitude in
all states, achieving significant amplitude in a given solution with
the mixing matrix $U$ requires significant contribution from the vast
majority of states whose distance from that solution, $d$, is close to
$n/2$. Thus for large $n$, we can expect the behavior of the algorithm
to be dominated by the behavior of $u_d$ near $d=n/2$. In turn, this
leads to a focus on the behavior of the phases $t_h$ for $h$ near
$n/2$.

Similarly, for the phases $p_c$ based on the number of conflicts in
the states, provided the number of clauses is large, i.e., $m \gg 1$,
most assignments will have nearly the average number of conflicts
$\cAvg= m/2^k$. Thus the behavior of the algorithm will be dominated
by the phase choices near this average value.

Asymptotically, as $n$ increases we can expect the behavior of the
phases near the average values will be the only important choices
influencing the algorithm's behavior.  A constant value for the phases
just gives an overall phase factor for the resulting amplitude, which
has no effect on the probability to find a solution but can be chosen
to slightly simplify the analysis. Thus we are led to consider a
linear variation in the phase values. In this case, the phase choices
are described by two constants $\rho$ and $\tau$ so
\begin{equation}\eqlabel{rho}
p_c = e^{i \pi \rho (c-\cAvg)}
\end{equation}
and
\begin{equation}\eqlabel{th}
t_h = e^{i \pi \tau (h-n/2)}
\end{equation}
Because $c$ and $h$ are integers, it is sufficient to consider values
for $\rho$ and $\tau$ in the range $-1$ to 1. Furthermore, changing
the sign of both $\rho$ and $\tau$ simply changes the resulting
amplitude to its complex conjugate.  Thus we can further restrict
consideration to $\tau$ in the range 0 to 1.

From \eq{u}, this choice for $t_h$ and the binomial theorem give
\begin{equation}\eqlabel{ud}
u_d = \cos^n\left(  \frac{\pi \tau}{2} \right) \tan^d\left(  \frac{\pi \tau}{2} \right) (-i)^d
\end{equation}
For example, when $\tau=1/2$, $u_d = 2^{-n/2} (-i)^d$ as used for
solving 1-SAT problems~\cite{hogg98}. When $1/2 < \tau <1$, the
magnitude of $u_d$ increases with $d$, leading to relatively little
connection between solutions and states at distances less than $n/2$
from them. For $k$-SAT problems, states with relatively few conflicts
tend to be clustered. So we can expect it will be advantageous to
emphasize the mixing at distances less than $n/2$ by selecting $\tau <
1/2$. As shown below, the best values for $\tau$ do indeed satisfy
this criterion.

Completing the algorithm requires particular choices for $\rho$ and
$\tau$.  The asymptotic analysis given below shows how these choices can
optimize the performance based on the values of $k$ and $m/n$.


\section{Analysis for Random $k$-SAT}\sectlabel{analysis}

After completing the algorithm, the amplitude in assignment $r$ is
\begin{equation}\eqlabel{phi}
\phi_r = \sum_s U_{rs} P_s \frac{1}{2^{n/2}} = \frac{1}{2^{n/2}} \sum_{s} u_{d(r,s)} \sum_c p_c \chi(s,c)
\end{equation}
where $\chi(s,c)$ is 1 if the assignment $s$ has $c$ conflicts, and
otherwise $\chi(s,c)=0$. The probability to find a solution when this
state is measured is
\begin{equation}\eqlabel{pSoln}
\Psoln = \sum_{\{r | \mbox{\scriptsize $r$ is a solution}\}} |\phi_r|^2 = \sum_r |\phi_r|^2 \chi(r,0)
\end{equation}

When the phase choices are particularly simple, as with the
unstructured search algorithm~\cite{grover96}, or the problem has a
simple relation between Hamming distance from a solution and number of
conflicts, as in 1-SAT problems~\cite{hogg98}, this expression for
$\Psoln$ can be evaluated to give a simple analytic result. In more
general cases, the solution probability can only be evaluated with a
classical simulation, limiting the study of more complex algorithms or
problem structures to relatively small sizes. A third approach to
evaluating $\Psoln$ is to average over an ensemble of problems. This
approach is developed in the remainder of this section and shows how
the structure of search problem ensembles can be used to analyze the
asymptotic behavior of the algorithm, and hence used to select the
best phase values to use.


\subsection{Average Behavior}

Averaging over a simple ensemble of problems reduces the evaluation of
the algorithm's behavior to that of counting the likelihood of various
types of problems in the ensemble. Using \eq{phi} and \bareeq{pSoln},
the average probability of finding a solution is
\begin{equation}\eqlabel{pSolnAvg}
\expect{\Psoln}=\frac{1}{2^n} \sum_r \sum_{ss'} u_{d(r,s)} u_{d(r,s')}^*  
	\sum_{cc'} p_c p_{c'}^* \expect{\chi(s,c) \chi(s',c')\chi(r,0)}
\end{equation}

The expected value $\expect{\chi(s,c) \chi(s',c')\chi(r,0)}$ is just
the fraction of problems for which $r$ is a solution and $s$ and $s'$
have, respectively, $c$ and $c'$ conflicts. To evaluate this sum, let
$a$ be the number of conflicts $s$ and $s'$ have in common, and let
$b=c-a$ and $b'=c'-a$ be their respective numbers of distinct
conflicts. With \eq{rho}, the inner sum over $c$ and $c'$ in
\eq{pSolnAvg} becomes
\begin{equation}\eqlabel{cSum}
\sum_{bb'} e^{i \pi \rho (b-b')} \sum_a \expect{\chi(s,b+a) \chi(s',b'+a)\chi(r,0)}
\end{equation}
The sum over $a$ just gives the fraction of problems
$\Nproblems(b,b')/\Nproblems$ for which $r$ is a solution and $s$ and
$s'$ have, respectively, $b$ and $b'$ unique conflicts.
Here $\Nproblems$ is the number of problems in the ensemble.

This expectation value is readily determined for the ensemble of
random $k$-SAT with $m$ clauses which has $\mMax = {n
\choose k} 2^k $ possible clauses to select from and
\begin{equation}
\Nproblems = {\mMax \choose m}
\end{equation}
possible problems, each of which is equally likely to be selected.
$\Nproblems(b,b')$ is the number of ways $m$ clauses can be selected
from the $\mMax$ available to satisfy the conditions on $r$, $s$ and $s'$.

\figdef{venn}{
\epsfig{file=venn.eps,height=1.5in}
}{Clause selection. The figure shows how the $\mMax$ possible clauses
are grouped. The dark gray region represents clauses that conflict
with $r$ and so cannot be selected. The white regions represent
clauses that conflict only with $s$ or $s'$. The selected clauses must
include $b$ and $b'$ conflicting only with $s$ and $s'$,
respectively. The remaining $m-b-b'$ clauses must be selected from the
light gray region: conflicting with both $s$ and $s'$ or with
neither.}

The possible clause selection is illustrated in \fig{venn}. For given
assignments $r$, $s$ and $s'$, let $\Ns$ and $\Nss$ be the number of
clauses that conflict only with $s$ and $s'$, respectively, and
$\Nother$ the number that do not conflict with $r$ and conflict with
both $s$ and $s'$ or neither (the light gray region in \fig{venn}).
Then we have
\begin{equation}\eqlabel{problems}
\Nproblems(b,b') = {\Ns \choose b} {\Nss \choose b'} {\Nother \choose m-b-b'}
\end{equation}
as the number of problems for which $s$ and $s'$ have, respectively,
$b$ and $b'$ unique conflicts, and $r$ is a solution. 

\figdef{wxyz}{
\begin{picture}(140,100)
	\put(85,100){\makebox(0,0){$ \overbrace{\makebox(120,0){}}^n $}}

	\put(20,70){
		\begin{picture}(120,20)
		\put(0,0){\framebox(120,20){}}
		\end{picture}
	}

	\put(20,45) {
	\begin{picture}(120,20)
		\put(0,0){\framebox(60,20){}}
		\put(60,0){\framebox(60,20){\rule{60pt}{20pt}}}
	\end{picture}
	}

	\put(20,20) {
	\begin{picture}(120,20)
		\put(0,0){\framebox(30,20){}}
		\put(30,0){\framebox(60,20){\rule{60pt}{20pt}}}
		\put(90,0){\framebox(30,20){}}
	\end{picture}
	}

	\put(35,10){\makebox(0,0){$w$}}
	\put(65,10){\makebox(0,0){$x$}}
	\put(95,10){\makebox(0,0){$y$}}
	\put(125,10){\makebox(0,0){$z$}}

	\put(10,80){\makebox(0,0){$r$}}
	\put(10,55){\makebox(0,0){$s$}}
	\put(10,30){\makebox(0,0){$s'$}}
\end{picture}
}{Grouping of variables based on assigned values in $r$, $s$ and $s'$,
each shown as a horizontal box schematically indicating values
assigned to each of the $n$ variables. In each assignment, the value
given in $r$ to a variable is shown as white, while black indicates
the opposite value. In this diagram, variables are grouped according
to the differences in values they are given in the three
assignments. For instance, the first group, consisting of $w$
variables, has those variables assigned the same value in all three
assignments.}

To determine values for $\Ns$, $\Nss$ and $\Nother$,
consider the $n$ variables in four mutually exclusive groups based on
the values they are assigned in $r$, $s$ and $s'$, as illustrated in
\fig{wxyz}:
\begin{enumerate}
\item the $w$ variables with the same values in all three assignments
\item the $x$ variables with the same value in $r$ and $s$, but opposite value in $s'$
\item the $y$ variables with the same value in $s$ and $s'$, opposite that of $r$
\item the $z$ variables with the same value in $r$ and $s'$, but opposite value in $s$
\end{enumerate}
Assignment $r$ conflicts with $n \choose k$ clauses leaving
$\mMax - {n \choose k}$ clauses available for selection.
The principle of inclusion and exclusion~\cite{palmer85} gives the
number of available clauses that conflict with
\begin{mathletters}\eqlabel{clauses}  % clauses refers to the group of equations
\begin{itemize}
\item both $s$ and $s'$
is the number that conflict with both $s$ and $s'$ minus the number of
those that also conflict with $r$:
\begin{equationGroup}
\Nboth = {w + y \choose k} - {w \choose k}
\end{equationGroup}

\item $s$ only is
\begin{equationGroup}
\Ns = {n \choose k} - {w+x \choose k} - \Nboth
\end{equationGroup}

\item $s'$ only is
\begin{equationGroup}
\Nss = {n \choose k} - {w+z \choose k} - \Nboth
\end{equationGroup}

\item both $s$ and $s'$ or with neither is
\begin{equationGroup}
\Nother = {n \choose k}(2^k-1) - \Ns - \Nss
\end{equationGroup}

\end{itemize}
\end{mathletters} % can add \eqlabel to individual group members if needed

Through the expressions of \eq{clauses}, $\Nproblems(b,b')$ given in
\eq{problems} depends on $w$, $x$, $y$ and $z$, but otherwise is
independent of the choice of assignments $r$, $s$ and
$s'$. Furthermore, $n=w+x+y+z$, $d(r,s)=y+z$ and $d(r,s')=x+y$.  Thus,
in \eq{pSolnAvg} the sum over the assignments $s$ and $s'$ can be rewritten
as a sum over $x$, $y$ and $z$ (with $w=n-x-y-z$) times the number of
ways to pick $s$ and $s'$ with assigned values matching each other
and those of $r$ as specified by the values of $w$, $x$, $y$ and $z$. This
latter quantity is just the multinomial coefficient $n \choose
w,x,y,z$.  Finally, because the quantities in the sum depend on $w$,
$x$, $y$ and $z$ but not the specific choice of the assignment $r$,
the sum $2^{-n} \sum_r$ appearing in \eq{pSolnAvg} just gives
one. Thus for the ensemble of random $k$-SAT, \eq{pSolnAvg} becomes
\begin{equation}\eqlabel{pSolnRandom}
\expect{\Psoln}= \sum_{xyz} {n \choose w,x,y,z} u_{y+z} u_{x+y}^* \sum_{bb'}  e^{i \pi \rho (b-b')} \frac{\Nproblems(b,b')}{\Nproblems}  
\end{equation}


\subsection{Asymptotic Behavior}

For the random $k$-SAT ensemble, \eq{pSolnRandom} gives the exact
value for $\expect{\Psoln}$. Its asymptotic behavior as the number of
variables and clauses increases can be determined using Stirling's
formula for the binomial coefficients~\cite{abramowitz65} and then
expressing the sum as an integral that can be readily analyzed.

The discussion in \sect{scaling} indicates the main contribution to
the sums of \eq{pSolnRandom} is from assignments with close to the
average number of conflicts, $\cAvg= m/2^k$. That is, terms for which
$b$ and $b'$ are $O(m)$. We thus use the scaled values $\s{b} = b/m$
and $\s{b'}=b'/m$ to simplify the analysis. Similarly the main
contribution in the outer sum comes from values of $d(r,s)=y+z$ and
$d(r,s')=x+y$ close to $n/2$. This suggests the scaling $\s{w} =
w/n$,...,$\s{z} = z/n$. As we will see below, these are indeed the
appropriate scaling behaviors for the dominant contributions to the
sum.

\subsubsection{Sum over Conflicts}

With $w,x,y,z$ scaling as $O(n)$, the number of possible clauses
$\mMax$ and each value in \eq{clauses} scale as ${n \choose k} =
O(n^k)$. This value is much larger than the actual number of clauses
$m$ that appear in hard problems, for which $m = O(n)$. We consider
the scaling for $m$ of $1 \ll m \ll n^k$. A convenient scaling for the
numbers of available clauses is $\s{N}_{\ldots} = N_{\ldots}/\mMax$ so
that
\begin{eqnarray}\eqlabel{Nscaled}
\sNboth 	&=&	\frac{(\s{w} + \s{y})^k - \s{w}^k}{2^k}  \\
\sNs 	&=&	\frac{1 - (\s{w}+\s{x})^k}{2^k} - \sNboth	\nonumber \\
\sNss 	&=&	\frac{1 - (\s{w}+\s{z})^k}{2^k} - \sNboth	\nonumber \\
\sNother &=&	1-2^{-k} - \sNs - \sNss				\nonumber
\end{eqnarray}
with corrections of order $1/n$.

When $R=O(n^k)$ and $S=O(m) \ll \sqrt{R}$, $R \choose S$ scales as $e^{S +
S \log(R/S)} /\sqrt{2\pi S}$. Thus $\Nproblems(b,b')/\Nproblems$ scales
as $\exp(m X) m^{-1} Y$
where
\begin{equation}
X = 	  \s{b}  \log \frac{\sNs}{\s{b}} 
	+ \s{b}' \log \frac{\sNss}{\s{b}'}
	+ (1-\s{b}-\s{b}') \log \frac{\sNother}{1-\s{b}-\s{b}'}
\end{equation}
and
\begin{equation}
Y = \frac{1}{2 \pi \sqrt{\s{b} \s{b}' (1-\s{b}-\s{b}')}}
\end{equation}

For the inner sum of \eq{pSolnRandom}, this quantity is multiplied by
$\exp(i \pi m \rho (\s{b} - \s{b}'))$ and summed over $b$ and $b'$. When
$m$ is large, this sum can be approximated by an integral over the scaled
variables $\s{b}$ and $\s{b}'$. This change to using scaled variables
introduces a power of $m$ for each variable, so the sum is asymptotic to
\begin{equation}
m \int d\s{b} \,d\s{b}' \,Y \exp(m (X + i \pi \rho (\s{b} - \s{b}')))
\end{equation}

The asymptotic behavior of this integral as $m \rightarrow \infty$ is
readily evaluated by the method of steepest descents~\cite{bender78}.
This involves considering complex values for the integration variables
and noting that the value of the integral is dominated by its behavior
around a stationary point, i.e., values for $\s{b}$ and $\s{b}'$ for
which $X + i \pi \rho (\s{b} - \s{b}')$ has zero derivatives with
respect to $\s{b}$ and $\s{b}'$. Specifically, the integral is
asymptotic to the value of the integrand at the stationary point
multiplied by $2 \pi m^{-1} /\sqrt{-\det D}$ where $D$ is the matrix
of 2nd derivatives of $X + i \pi \rho (\s{b} - \s{b}')$ evaluated at
the stationary point, and $\det D$ is its determinant. Evaluating
these derivatives then shows the inner sum is asymptotic to $\exp(m I)$
with
\begin{equation}
I = \log \left( 
	e^{i \pi \rho} \sNs
	+ e^{-i \pi \rho} \sNss
	+ \sNother 
\right)
\end{equation}
which depends on $\s{w},\ldots,\s{z}$ through \eq{Nscaled}.

As presented above, the expansion of the binomials involved in this
sum required assuming $m \ll \sqrt{n^k}$. When $m$ is larger than
this, some of the corrections to the expansion no longer go to zero as
$m$ increases. However, if $\rho$ is small, specifically of order
$n/m$, these additional corrections do not change the final asymptotic
result. As described below, this behavior of $\rho$ is the appropriate
choice for $m \gg n$ so we use this result over the full set of
scaling behaviors for $m$.


\subsubsection{Sum Over Variable Groupings}

For the remaining sum in \eq{pSolnRandom}, over $x$, $y$ and $z$, the
multinomial scales as
\begin{equation}
{n \choose w,x,y,z} \sim \exp(n H) n^{-3/2} \sqrt{\frac{1}{(2 \pi)^3 \s{w} \s{x} \s{y} \s{z}}}
\end{equation}
with
\begin{equation}\eqlabel{H}
H = -\s{w} \log \s{w}-\ldots-\s{z} \log \s{z}
\end{equation}
and $\s{w} = 1 - \s{x}-\s{y}-\s{z}$.

The $u_{y+z} u_{x+y}^*$ factors equal $\exp(n U)$ with
\begin{equation}
U = 2 \beta_C + \beta (\s{x}+2\s{y}+\s{z}) + i \pi (\s{x}-\s{z})/2
\end{equation}
where, from \eq{ud},
\begin{eqnarray}
\beta_C & = & \log(\cos(\pi \tau/2)) \\
\beta  & = & \log(\tan(\pi \tau/2)) \nonumber
\end{eqnarray}

Combining these values with the result from the inner sum again gives
a sum that can be approximated by an integral. After changing to scaled
variables this becomes
\begin{equation}\eqlabel{integral}
\expect{\Psoln} \sim  n^{3/2} \int d\s{x} \,d\s{y} \,d\s{z} \,
	\sqrt{\frac{1}{(2 \pi)^3 \s{w} \s{x} \s{y} \s{z}}} \exp(n (H+U) + m I)
\end{equation}
with $\s{w}=1-\s{x}-\s{y}-\s{z}$.  The method of steepest descents
applies to this integral. Thus, its asymptotic behavior is determined
by the stationary point, namely the values of $\s{x}$, $\s{y}$ and
$\s{z}$ for which the derivatives of $n (H+U) + m I$ with respect to
these three variables are zero. Let $\Delta$ be the corresponding $3
\times 3$ matrix of 2nd derivatives and $A$ the value of $-(H + U + I
m/n)$, both evaluated at this point. The asymptotic behavior is then
\begin{equation}\eqlabel{saddle}
\expect{\Psoln} \sim 
	\sqrt{\frac{-1}{\s{w} \s{x} \s{y} \s{z} \det \Delta}} \exp(-n A)
\end{equation}
evaluated at the stationary point.
These quantitites depend on the parameters $k$, $\mu=m/n$,
$\rho$ and $\tau$.

Because of the powers of $k$ introduced with the numbers of available
clauses in \eq{Nscaled}, the stationary point has no simple closed
form. However, it can be readily evaluated numerically. This
evaluation can be used to select values for $\rho$ and $\tau$ to
minimize the decay rate $A$. In this way the asymptotic analysis based
on the ensemble average can be used to optimize the search
algorithm. In the following sections we describe the resulting
behavior of these optimal choices for problems with various scalings
of the number of clauses $m$.


\section{Solving Hard Problems}

For random $k$-SAT, the hardest problem instances are concentrated at
a threshold value of $\mu=m/n$ depending on $k$. For 3-SAT, this
threshold is at $\mu=4.2$~\cite{crawford95}. Thus we examine the
behavior of the single step algorithm when $\mu$ is constant.  In this
case, the minimum decay rate $A$ is shown in \fig{optimal}.  The
corresponding best choices for $\rho$ and $\tau$ are shown in
\fig{parameters}. As one caveat, these values were obtained by numerical
minimization of $A$ so could be local minima. If so, other choices for
$\rho$ and $\tau$ would give even better performance than the values
reported here.

For comparison, \fig{optimal} also shows the scaling of random
selection, $\expect{S}/2^n$, where $S$ is the number of solutions. The
expected number of solutions, and its exponential scaling, is readily
computed for random $k$-SAT~\cite{williams92}. The unstructured search
requires multiple steps, with a cost on the order of $\sqrt{S/2^n}$
for a problem with $S$ solutions. As an approximate indication of the
average cost for a problem ensemble in \fig{optimal}, we use
$\sqrt{\expect{S/2^n}}$ instead of the exact value
$\expect{\sqrt{S/2^n}}$ which cannot be analytically computed.

\figdef{optimal}{
\epsffile{optimal.eps}
}{Smallest exponential decay rate $A$ for $\expect{\Psoln}$ as a function of
$\mu=m/n$ for random 3-SAT. For comparison, the gray and dashed curves
show the scaling for random selection and, approximately, for
unstructured search.}


\begin{table}[ht]
\caption{Best parameter values and scaling behavior for single-step search of
3-SAT problems for random and prespecified solution
ensembles.}\label{values}
\begin{center}
\begin{tabular}{r|ccc||ccc}
 	& \multicolumn{3}{c}{random} & \multicolumn{3}{c}{prespecified} \\
$\mu$	& $\tau$	& $\rho$ 	& $A$ & $\tau$	& $\rho$  & $A$ \\ \hline

1 & 	0.238 &  0.348 &  $0.027$ & 	0.239 &  0.344 &  $0.026$ \\
2 & 	0.260 &  0.291 &  $0.094$ &  	0.262 &  0.286 &  $0.088$ \\
3 & 	0.275 &  0.249 &  $0.181$ &  	0.278 &  0.242 &  $0.162$ \\
4 & 	0.286 &  0.218 &  $0.280$ & 	0.293 &  0.211 &  $0.231$ \\
5 &	0.295 &  0.195 &  $0.386$ & 	0.304 &  0.188 &  $0.281$ \\
6 &	0.303 &  0.176 &  $0.497$ &	0.311 &  0.172 &  $0.309$
\end{tabular}
\end{center}
\end{table}


\figdef{parameters}{
\epsffile{parameters.eps}
}{Optimal choices of $\tau$ (black) and $\rho$ (gray) as a function of
$\mu=m/n$. Solid curves are for the random 3-SAT ensemble and the dashed
curves are for the ensemble with a prespecified solution.}

An important observation from these results is even the best use of
problem structure based only on the number of conflicts cannot remove
the exponential search cost with a single step algorithm of the type
described here. Thus for problems with $m=O(n)$, any possibility for
further improvements in search performance requires more sophisticated
algorithms, such as using multiple steps or incorporating additional
problem structure. For example, using neighborhood information is
crucial for the performance of classical heuristics. As one
comparison, the cost of a classical heuristic method is empirically
observed~\cite{crawford95} to scale as $2^{n/19.5}$ for random 3-SAT
problems near $\mu=4.2$, corresponding to a value of $A$ equal to
$\log(2)/19.5=0.036$. This scaling is better than the single-step
quantum algorithm.

A second observation from \fig{optimal} is, for $\mu$ less than about
3.5, the optimal choices of $\rho$ and $\tau$ give exponentially
better performance, on average, than the unstructured search
algorithm, which also requires multiple steps. Thus this analysis
demonstrates how the structure of simple search ensembles can be
exploited to improve quantum search performance and simultaneously
reduce the required coherence time. Moreover, by giving the actual
asymptotic scaling this result is more definitive than prior empirical
studies of algorithms based on classical simulations of small
problems~\cite{hogg97}.

Beyond $\mu=4.2$, the fraction of soluble problems in the random
3-SAT ensemble, $\Psoluble$, drops to zero as $n$ increases. Since
this analysis averages over {\em all}\/ problems in the ensemble,
$\expect{\Psoln} \leq \Psoluble$. The performance of this algorithm
for overconstrained soluble problems is given instead by
$\expect{\Psoln}/\Psoluble$. Unfortunately, this ensemble does not
have a simple expression for $\Psoluble$, or even just its leading
exponential scaling, making this evaluation difficult.  Instead one
could use empirical classical search to evaluate $\Psoluble$ and
estimate its scaling behavior as a function of $\mu$. However, this
technique is increasingly difficult as $\mu$ increases due to the
rapidly decreasing fraction of soluble problems in the ensemble.

An alternate approach is to examine an ensemble where all problems are
soluble and that is analytically simple, e.g., the ensemble with a
prespecified solution. The evaluation of $\expect{\Psoln}$ proceeds as
shown above, with the addition of needing to keep track of the
distances of the sets $r$, $s$ and $s'$ to the prespecified solution
(which affects the available number of clauses that can be selected to
produce the required numbers of conflicts).  The resulting optimal
behavior is shown in \fig{prespecified}, with the corresponding best
values for $\rho$ and $\tau$ given in \fig{parameters}. The decay rate
decreases as $\mu$ increases past 7, i.e., problems become easier as
the number of clauses increases. Presumably a similar behavior would
be seen for the soluble cases in the random ensemble as well because
these two ensembles are fairly similar when $\mu$ is large. On the
other hand, random selection and the unstructured algorithm do not
improve as $\mu$ increases: they do not take advantage of the
structure of highly constrained problems.


\figdef{prespecified}{
\epsffile{prespecified.eps}
}{Smallest exponential decay rate $A$ for $\expect{\Psoln}$ as a function of
$\mu=m/n$ for 3-SAT with prespecified solution. For comparison, the
gray and dashed curves show the scaling for random selection and,
approximately, for unstructured search.}


Since classical simulations of quantum algorithms are limited to few
variables, this asymptotic analysis can also indicate of the extent to
which these simulations match the asymptotic behavior. An example is
\fig{scaling} showing that the behavior matches that from
Table~\ref{values} for $\mu=2$ and $\mu=4$ even with a small number of
variables.

\figdef{scaling}{
\epsffile{scaling.eps}
}{Optimal asymptotic behavior of $\expect{\Psoln}$ for 3-SAT with
$\mu=2$ (gray) and 4 (black) on a log-scale vs.~$n$. The points show
the exact values of $\expect{\Psoln}$.}



\section{Limiting Cases}

The behavior of the minimum decay rate shown in \fig{optimal} and
\barefig{prespecified} suggests that $A$ decreases toward zero for
small and large values of $\mu$ for soluble problems. This is
confirmed in \fig{limits} which shows the behavior of both ensembles
over a larger range of values.

As $\mu \rightarrow 0$, the figure shows the minimum decay
rate is nearly a straight line with slope 2 on this log-log plot,
indicating $A = O(\mu^2)$ in this limit. With this limiting
behavior $\expect{\Psoln}$ decays as $\exp(-A n)=\exp(O(m^2/n))$.
In particular, if $m$ grows no faster than $\sqrt{n}$, 
$\expect{\Psoln}$ will remain $O(1)$ as $n$ increases.

Similarly, as $\mu \rightarrow \infty$, \fig{limits} shows $A$
decreasing in a straight line with slope $-1$, indicating $A =
O(1/\mu)$. Correspondingly, $\expect{\Psoln}$ then decays as
$\exp(O(n^2/m))$. So if $m$ grows at least as fast as $n^2$, the
probability to find a solution, on average, will remain $O(1)$ as $n$
increases.

\figdef{limits}{
\epsffile{limits.eps}
}{Limiting behaviors, on a log-log plot, for exponential decay rate
$A$. Behavior for random 3-SAT (black curve, up to $\mu=6$),
prespecified solution 3-SAT (gray curve, for $\mu$ between 0.1 and 40)
and the upper bound based on probability to find the prespecified
solution (black curve, for $\mu \ge 20$). For comparison the dashed
lines show the limiting behaviors for small and large $\mu$, which
correspond very closely to the exact values for $\mu<0.3$ and
$\mu>1000$.}

In the remainder of this section, we confirm these limiting
observations from \fig{limits}. While such weakly and highly
constrained problems are fairly simple, the single-step algorithm
outperforms the best classical heuristic methods for these cases,
which require evaluating $O(n)$ assignments on average.

%This analysis also improves on previous work~\cite{hogg97c} showing $O(1)$ behavior for highly constrained problems, but only when $m$ grew faster than a particular multiple of $n^2$ (equal to 17.3 for $k=3$).


\subsection{Solving Weakly Constrained Problems}

When $1 \ll m \ll n$, the decay rate $A = -(H + U + I m/n)$,
can be treated through an expansion in the small quantity $\mu=m/n$.
Specifically, the location of the stationary point for the variables
$\s{x}$, $\s{y}$ and $\s{z}$ is determined, to $O(1)$, by setting
the derivatives of $H+U$ to zero. The contribution from the sum
over conflicts, $\mu I$, only introduces corrections of $O(\mu)$.

The expression $H+U$ evaluated at the $O(1)$ values for the stationary
point is zero, for {\em any}\/ choice of the parameters $\rho$ and
$\tau$. The $O(\mu)$ values and the contribution from $\mu I$ then
give a scaling for $\expect{\Psoln}$ of $\exp(O(m))$.  However, for an
appropriate choice of $\tau$ and $\rho$, the coefficient of this
$O(m)$ term can be set to zero, so that the actual scaling is
dominated by the $O(\mu^2)$ correction, i.e., $\exp(O(n \mu^2))$
corresponding to the behavior seen in \fig{limits}.

While the final result of this analysis does not have a simple closed
form for the parameter values that eliminate the $O(m)$ decay, they
are determined by trigonometric equations. These arise from setting to
zero the derivatives of the $O(m)$ contribution with respect to $\tau$
and $\rho$. The optimal choice for $\tau$ is determined by
\begin{equation}
2 \cos^k \left( \frac{\pi \tau}{2} \right) \cos\left(k \frac{\pi \tau}{2} \right) = 1
\end{equation}
With this value for $\tau$, $\rho$ must then satisfy $\sin(\pi (\rho +
k \tau))=0$. Among the many possible solutions for $\rho$ and $\tau$,
we select the one in the range 0 to 1 for definiteness. For $k=3$
these equations give $\tau = 0.201389$ and $\rho=0.395832$, which
correspond to the limiting values in \fig{parameters} as $m/n
\rightarrow 0$.

These values of the parameters and the corresponding stationary point
in \eq{saddle} give $\expect{\Psoln} \sim e^{-\alpha_{\rm weak}
m^2/n}$. The decay rate is thus $A = \alpha_{\rm weak} \mu^2$. When
$m$ grows no faster than $\sqrt{n}$, $\expect{\Psoln}=O(1)$.  This
performance is better than any classical heuristic, and also improves
on the unstructured quantum search since solutions are still rare.

As an example, when $k=3$, $\alpha_{\rm weak}=0.029405$. The
corresponding decay rate $A$ is shown as the dashed line in
\fig{limits} for the $\mu \rightarrow 0$ limit.

% expected number of solutions: 2^n Binomial[Binomial[n,k] (2^k-1),m]/Binomial[Binomial[n,k] 2^k,m]


\begin{table}[ht]
\caption{Scaling of $\expect{\Psoln}$ from \eq{pSolnRandom} and
$\expect{S}/2^n$ for weakly constrained random 3-SAT with $m=2
\protect\sqrt{n}$.}\label{table.weak}
% \protect required for \sqrt in caption ????
\begin{center}
\begin{tabular}{cccc}
$n$ 	& $m$ 	& $\expect{\Psoln}$ & $\expect{S}/2^n$ \\ \hline
4	& 4	& 0.908 	& 0.569 \\
9	& 6	& 0.897 	& 0.447 \\
16	& 8	& 0.894 	& 0.343 \\
25	& 10	& 0.893 	& 0.263 \\
36	& 12	& 0.892		& 0.201
\end{tabular}
\end{center}
\end{table}
% values from best/c/data/optUR-2sqrt.dat

An example of the approach to the asymptotic behavior is given in
Table~\ref{table.weak} for $m = 2 \sqrt{n}$. In this case, the
asymptotic scaling limit is $\exp(-4 \alpha_{\rm weak})=0.889$. This
behavior compares with the still rapid decrease in expected fraction
of solutions which scales as $\expect{S}/2^n = (1-2^{-k})^m$ or
$(7/8)^m$ for $k=3$. Thus the unstructured search scaling for this
case is $(7/8)^{\sqrt{n}}$ which still decreases faster than
polynomially.


\subsection{Solving Highly Constrained Problems}

For $m \gg n$, we focus on the ensemble of problems with a
prespecified solution. In fact, with this many clauses, there are
relatively few solutions in addition to the prespecified one. Thus a
simpler evaluation considers the probability that the quantum search
finds the prespecified solution, rather than {\em any}\/
solution. This quantity is a lower bound on $\expect{\Psoln}$, and is
a tight bound when $m$ is large.

This lower bound is given by setting assignment $r$ in \eq{pSolnAvg}
to be the prespecified solution rather than summing over all possible
assignments.  The derivation leading to \eq{pSolnRandom} proceeds as
before except for two changes. First, eliminating the sum over $r$
gives an additional factor of $2^{-n}$. Second, the number of possible
problems $\Nproblems$ replaced by
\begin{equation}
\mMax - {n \choose k} \choose m
\end{equation}
reflecting the smaller number of problems with a prespecified solution.

The asymptotic analysis gives an additional overall factor
of $ 2^{-n} (1-2^{-k})^{-m}$. The resulting decay rate for
$\expect{\Psoln}$ then has an upper bound given by the value of $-(H +
U - \log 2 + (I - \log (1-2^{-k}) )) m/n)$ evaluated at the
corresponding stationary point.

For $m \gg n$, we can expand the stationary point evaluation
in powers of $1/\mu$. Following the behavior for the optimal value
of $\rho$ suggested by \fig{parameters} and the values obtained in
connection with \fig{limits}, we take $\rho$ to be proportional to
$1/\mu$. The $O(1)$ values for the stationary point in this case are
simply $\s{x}=\s{y}=\s{z}=1/4$. Because $\rho \rightarrow 0$, the
leading behavior for $I$ is just $\log(1-2^{-k})$ so the exponential
scaling of $O(m)$ is exactly zero. The contributions to the scaling of
$O(n)$ can be made equal to zero by selecting $\tau=1/2$ and $\rho =
2^{k-2}(2^k-1)/(k \mu)$. The simple form for the $O(1)$ stationary point
values also allows evaluating the overall $O(1)$ asymptotic behavior.
Specifically, with these parameters, the scaling of the probability to
find the prespecified solution is
\begin{equation}
\frac{4}{\sqrt{16 + (k-1)^2 \pi^2}} \exp \left(-\frac{(2^k-1)^3 \pi^2}{16 k^2}  \frac{n^2}{m} \right)
\end{equation}
The corresponding value for the decay rate, $A=(2^k-1)^3 \pi^2/(16 k^2
\mu)$, is shown in \fig{limits}.

In particular, when $m$ is at least $O(n^2)$, we have $O(1)$
performance, duplicating the behavior seen with highly constrained
$k$-SAT problems~\cite{hogg98}.



\section{Required Precision}

We have presented the scaling behavior of $\expect{\Psoln}$ for
appropriate choices of $\rho$ and $\tau$. In particular, for
sufficiently weakly or highly constrained problems $\expect{\Psoln}$
is $O(1)$, which is better performance than possible with
classical methods for these easy problems.

One important question in realizing such scaling is that of the
required precision in implementing the parameters. Fortunately, these
requirements are not particularly strict because the parameters appear
in the exponents. In particular, an error of $\epsilon$ in $\rho$ or
$\tau$ will give error $O(\epsilon^2)$ in the exponent. E.g., for
weakly constrained case, this gives exponential decay of
$\exp(O(\epsilon^2) m )$. To maintain $O(1)$ behavior, $\epsilon$ will
need to be small enough so that $\epsilon^2 m = O(1)$, i.e., the
precision requirement is $\epsilon = O(1/\sqrt{m})$. Similarly, for $m
\gg n$, such parameter errors give exponential decay of
$\exp(O(\epsilon^2) n )$, so the precision requirement is $\epsilon =
O(1/\sqrt{n})$.

Thus only a square root precision in the implementation of the values
of $\rho$ and $\tau$ is required. While by no means trivial, this
shows the algorithm does not require exponentially precise parameter
values to achieve the scaling.

The precise scaling is determined by the size of the second
derivatives of the exponent around the maximum. For $k=3$, a
root-mean-square combined error of $\epsilon$ in $\rho$ and $\tau$
introduces at worst a factor $\exp(-3.18 \epsilon^2 m)$. This is
greater than 1/2 provided $\epsilon < 0.46/\sqrt{m}$. For example,
with $m = 2 \sqrt{n}$, this requirement is satisfied for a 10,000
variable problem provided $\epsilon < 0.033$, which allows, roughly, a
10\% error in the values.
% underconstrained.nb (precision requirement example for k=3)




\section{Problem Search Costs}

The ensemble average leading to \eq{pSolnRandom} provides a direct
analysis of $\expect{\Psoln}$. This technique can also be generalized
to quantities involving positive integer powers of $\Psoln$, such as
the variance. However, there remain a number of additional properties
whose ensemble behavior cannot be expressed in simple form.  One such
quantity is the expected solution cost which, for any particular
problem, is $1/\Psoln$. Since ensembles that include even one problem
with no solutions will have $\expect{1/\Psoln} = \infty$, a more
useful quantity is its median value or the behavior of its
distribution to identify the extent to which the algorithm performs
well for many or most problems. Such quantities require classical
simulation to evaluate. A comparison of these quantities is given in
Table~\ref{cost}. We see that $\frac{1}{\expect{\Psoln}}$ underestimates
the median search cost, but is a better estimate than $\expect{1/\Psoln}$
even when the latter is restricted to soluble problems.

\begin{table}[ht]
\caption{Comparison of search cost estimates based on 1000 soluble random 3-SAT problems.}\label{cost}
\begin{center}
\begin{tabular}{r|ccc||ccc}
        & \multicolumn{3}{c}{$\mu=2$} & \multicolumn{3}{c}{$\mu=4$} \\
$n$	& $\frac{1}{\expect{\Psoln}}$	& ${\rm median}\left( \frac{1}{\Psoln} \right)$ 	& $\expect{\frac{1}{\Psoln}}$ & 

	 $\frac{1}{\expect{\Psoln}}$	& ${\rm median}\left( \frac{1}{\Psoln} \right)$ 	& $\expect{\frac{1}{\Psoln}}$ \\ \hline

~ & & & & & & \\

10 & 	2.6 &  2.6 &  2.8 & 	15 &  17 &  25 \\
20 & 	6.6 &  6.8 &  7.4 &  	228 &  352 &  705
\end{tabular}
\end{center}
\end{table}

Quantum search works by using correlation between distance from
solution and number of conflicts. This is very different from
properties used by classical searches. An interesting question is the
extent to which these properties are different, e.g., to what extent
are some problems hard for quantum search but easy for classical ones
and vice versa. An indication is given in \fig{classical}.
Specifically, the expected quantum search cost for a single instance
is given by $1/\Psoln$. For a classical comparison, each problem was
solved repeatedly with the GSAT local search
method~\cite{selman92}. The expected search cost was computed by the
ratio of the total number of GSAT steps to the number of solutions
found by these repeated searches. The figure shows a general
correlation between search difficulty for the two methods, but with
substantial variation.

The figure also shows the absolute number of steps required for the
quantum method is considerably larger than the classical heuristic for
these problems. This contrasts with sufficiently weakly or highly
constrained problems where the quantum method requires $O(1)$ steps
while the classical search uses $O(n)$. Achieving further improvement
for the quantum search may require incorporating into the phase
choices the neighborhood information used in the classical heuristic.

\figdef{classical}{
\epsffile{classical.eps}
}{Comparison of quantum and classical search for random 3-SAT with $n=20$ and $m=80$. Each point shows the expected search cost for a single problem instance.}



%Show distribution of solution probabilities. ApEn does not appear to identify atypical cases (1/Psoln vs. ApEn) (see compare.m in repair/c/approxEnt)




\section{Discussion}

We have shown how an analysis based on ensemble averages can be used
to design quantum search algorithms. The resulting algorithm was
evaluated for satisfiability problems with various numbers of
constraints, including those in the hard region as well as the easier
weakly and highly constrained cases. Compared to unstructured search,
this gives average behavior that, in most cases, is exponentially
better. Moreover, this performance is obtained with only a single
evaluation of the assignment properties rather than the exponentially
large number of repeated evaluations required by the unstructured
method. Thus this single-step algorithm requires much less coherence
time for the quantum operations.  

The new quantum algorithm introduced in this paper does not achieve
the performance of existing classical heuristics for hard problems,
but is better for sufficiently weakly and highly constrained cases in
terms of the number of search steps required. Classical and quantum
search steps both require about the same number of elementary
operations. However, the extent to which differences in number of
search steps translate into differences in actual execution time
depends on details of implementation and developments of quantum
compiler technology for optimizing the choice of basic
operations. 

Classical heuristics use more properties of the problem than the
algorithm described here. These properties include the difference
between the number of conflicts in an assignment and those of its
neighbors, and the conflicts associated with partial
assignments. Incorporating such information in the quantum algorithm,
through additional variation in the phase values, may improve quantum
search.

An important advantage of basing the algorithm on ensembles is the use
of average quantities of ensemble rather than requiring knowledge of
the quantities for an individual search problem. This contrasts with
the unstructured search method which requires knowledge of the number
of solutions for a particular problem, or various values must be tried
repeatedly~\cite{boyer96}.  An interesting open question is whether
the algorithm, e.g., the choice of $\rho$ and $\tau$, could be
improved by adjusting the values based on readily computed
characteristics of an individual problem instance. In effect this
would amount to using a more specific ensemble whose instances are
more likely to be similar to the given instance than random problems.
This observation raises the question of identifying a reasonable
ensemble for a given problem.

The technique, introduced in this paper, of evaluating asymptotic
behavior for a problem ensemble could be useful for evaluating a
variety of quantum algorithms that are not easily addressed
theoretically and hence would otherwise require slow classical
simulation. This evaluation requires only that the properties
of assignments used by the algorithm and the nature of the ensemble
allow for an explicit determination of the ensemble averages,
in analogy with \eq{pSolnRandom}. Furthermore, in many respects
this analysis is simpler than that for heuristic classical methods.
This is because classical searches introduce correlations in their
path through a search space based on a series of heuristic choices.
Accounting for the variation due to these choices is difficult
to model theoretically. By contrast, the quantum search, by in effect
exploring all search paths simultaneously, avoids this difficulty
thereby giving relatively simple analytic expressions for the
average behavior.

The power of this algorithm comes from exploiting the correlation
between the number of conflicts and the distance from a solution among
the large number of assignments that have close to the average number
of conflicts and differ by about $n/2$ values from a solution.  The
resulting behavior of the scaling rate as a function of $\mu$ gives
another perspective on the nature of the phase transition and the hard
problems. That is, near the transition region, this correlation is
weakest, reducing the amount of oscillatory behavior that can be
removed from the sum of $\expect{\Psoln}$ by appropriate choices of
the phase parameters.  This correlation is perfect for maximally
constrained problems, but enough remains for harder problems to
substantially improve the search performance.

More generally, the significance for quantum search of such
correlations within individual problems provide an alternate
perspective on the distinction between easy and hard problem
instances.  This perspective is unlike classical heuristics that rely
on behavior of states near solutions as guides, and can become stuck
in local minima or among large collections of assignments with the
same number of conflicts~\cite{frank97}. For the quantum algorithm,
local minima are not an issue: instead it is the limited correlation
between distance and conflicts for states {\em far}\/ from solutions
that prevents efficient search. Because of these very different
characteristics, an interesting direction for future work is
identifying individual problems or problem ensembles where the
correlations are stronger even though the local minima for states
relatively near solutions remain. In such cases, quantum algorithms
could perform much better than classical heuristics.

\section*{Acknowledgments}
I have benefited from discussions with Carlos Mochon, Wolf Polak and
Eleanor Rieffel. The On-Line Encyclopedia of Integer
Sequences~\cite{sloane73} was helpful in identifying the exact form of
the optimal parameters for highly constrained problems.


% for submission insert result of bibtex directly into this file
%\bibliography{physrev,received}
%\bibliographystyle{plain}

\begin{thebibliography}{10}

\bibitem{abramowitz65}
M.~Abramowitz and I.~Stegun, editors.
\newblock {\em Handbook of Mathematical Functions}.
\newblock Dover, New York, 1965.

\bibitem{bender78}
Carl~M. Bender and Steven~A. Orszag.
\newblock {\em Advanced Mathematical Methods for Scientists and Engineers}.
\newblock McGraw Hill, NY, 1978.

\bibitem{benioff82}
P.~Benioff.
\newblock Quantum mechanical hamiltonian models of {Turing} machines.
\newblock {\em J. Stat. Phys.}, 29:515--546, 1982.

\bibitem{bernstein92}
Ethan Bernstein and Umesh Vazirani.
\newblock Quantum complexity theory.
\newblock In {\em Proc. 25th ACM Symp. on Theory of Computation}, pages 11--20,
  1993.

\bibitem{berthiaume94}
Andre Berthiaume, David Deutsch, and Richard Jozsa.
\newblock The stabilization of quantum computations.
\newblock In {\em Proc. of the Workshop on Physics and Computation
  (PhysComp94)}, pages 60--62, Los Alamitos, CA, 1994. IEEE Press.

\bibitem{boyer96}
Michel Boyer, Gilles Brassard, Peter Hoyer, and Alain Tapp.
\newblock Tight bounds on quantum searching.
\newblock In T.~Toffoli et~al., editors, {\em Proc. of the Workshop on Physics
  and Computation (PhysComp96)}, pages 36--43, Cambridge, MA, 1996. New England
  Complex Systems Institute.

\bibitem{brassard98}
Gilles Brassard, Peter Hoyer, and Alain Tapp.
\newblock Quantum counting.
\newblock {Los Alamos} preprint quant-ph/9805082, May 27 1998.

\bibitem{cerf98}
Nicolas~J. Cerf, Lov~K. Grover, and Colin~P. Williams.
\newblock Nested quantum search and {NP}-complete problems.
\newblock {Los Alamos} preprint quant-ph/9806078, June 23 1998.

\bibitem{cerny93}
Vladimir Cerny.
\newblock Quantum computers and intractable ({NP}-complete) computing problems.
\newblock {\em Physical Review A}, 48:116--119, 1993.

\bibitem{cheeseman91}
Peter Cheeseman, Bob Kanefsky, and William~M. Taylor.
\newblock Where the really hard problems are.
\newblock In J.~Mylopoulos and R.~Reiter, editors, {\em Proceedings of
  IJCAI91}, pages 331--337, San Mateo, CA, 1991. Morgan Kaufmann.

\bibitem{chuang98a}
Isaac~L. Chuang, Neil Gershenfeld, and Mark Kubinec.
\newblock Experimental implementation of fast quantum searching.
\newblock {\em Physical Review Letters}, 80:3408--3411, 1998.

\bibitem{chuang98}
Isaac~L. Chuang, Lieven M.~K. Vandersypen, Xinlan Zhou, Debbie~W. Leung, and
  Seth Lloyd.
\newblock Experimental realization of a quantum algorithm.
\newblock {\em Nature}, 393:143--146, 1998.
\newblock Los Alamos preprint quant-ph/9801037.

\bibitem{crawford95}
James~M. Crawford and Larry~D. Auton.
\newblock Experimental results on the crossover point in random {3SAT}.
\newblock {\em Artificial Intelligence}, 81:31--57, 1996.

\bibitem{deutsch85}
D.~Deutsch.
\newblock Quantum theory, the {Church-Turing} principle and the universal
  quantum computer.
\newblock {\em Proc. R. Soc. London A}, 400:97--117, 1985.

\bibitem{deutsch89}
D.~Deutsch.
\newblock Quantum computational networks.
\newblock {\em Proc. R. Soc. Lond.}, A425:73--90, 1989.

\bibitem{divincenzo95}
David~P. DiVincenzo.
\newblock Quantum computation.
\newblock {\em Science}, 270:255--261, 1995.

\bibitem{feynman86}
R.~P. Feynman.
\newblock Quantum mechanical computers.
\newblock {\em Foundations of Physics}, 16:507--531, 1986.

\bibitem{frank97}
J.~Frank, P.~Cheeseman, and J.~Stutz.
\newblock When gravity fails: Local search topology.
\newblock {\em J. of Artificial Intelligence Research}, 7:249--281, 1997.

\bibitem{garey79}
M.~R. Garey and D.~S. Johnson.
\newblock {\em Computers and Intractability: A Guide to the Theory of
  NP-Completeness}.
\newblock W. H. Freeman, San Francisco, 1979.

\bibitem{grover97a}
Lov~K. Grover.
\newblock Quantum computers can search arbitrarily large databases by a single
  query.
\newblock {Los Alamos} preprint quant-ph/9706005, Bell Labs, 1997.

\bibitem{grover96}
Lov~K. Grover.
\newblock Quantum mechanics helps in searching for a needle in a haystack.
\newblock {\em Physical Review Letters}, 78:325--328, 1997.

\bibitem{haroche96}
Serge Haroche and Jean-Michel Raimond.
\newblock Quantum computing: Dream or nightmare?
\newblock {\em Physics Today}, 49:51--52, August 1996.

\bibitem{hogg97}
Tad Hogg.
\newblock A framework for structured quantum search.
\newblock {\em Physica D}, 120:102--116, 1998.
\newblock Los Alamos preprint quant-ph/9701013.

\bibitem{hogg98}
Tad Hogg.
\newblock Highly structured searches with quantum computers.
\newblock {\em Physical Review Letters}, 80:2473--2476, 1998.
\newblock Preprint at
  pub\-lish.aps.org/eprint/gate\-way/eplist/aps1997oct30\_002.

\bibitem{hogg95e}
Tad Hogg, Bernardo~A. Huberman, and Colin Williams.
\newblock Phase transitions and the search problem.
\newblock {\em Artificial Intelligence}, 81:1--15, 1996.

\bibitem{hogg98b}
Tad Hogg, Carlos Mochon, Eleanor Rieffel, and Wolfgang Polak.
\newblock Tools for quantum algorithms.
\newblock {Los Alamos} preprint quant-ph/9811073, Xerox, Sept. 1998.

\bibitem{hoyer97}
Peter Hoyer.
\newblock Efficient quantum transforms.
\newblock {Los Alamos} preprint quant-ph/9702028, February 1997.

\bibitem{knill98}
Emanuel Knill, Raymond Laflamme, and Wojciech~H. Zurek.
\newblock Resilient quantum computation.
\newblock {\em Science}, 279:342--345, 1998.

\bibitem{landauer94a}
Rolf Landauer.
\newblock Is quantum mechanically coherent computation useful?
\newblock In D.~H. Feng and B-L. Hu, editors, {\em Proc. of the Drexel-4
  Symposium on Quantum Nonintegrability}, Boston, 1994. International Press.

\bibitem{lloyd93}
Seth Lloyd.
\newblock A potentially realizable quantum computer.
\newblock {\em Science}, 261:1569--1571, 1993.

\bibitem{monroe96a}
Christopher Monroe and David Wineland.
\newblock Future of quantum computing proves to be debatable.
\newblock {\em Physics Today}, 49:107--108, November 1996.

\bibitem{nijenhuis78}
A.~Nijenhuis and H.~S. Wilf.
\newblock {\em Combinatorial Algorithms for Computers and Calculators}.
\newblock Academic Press, New York, 2nd edition, 1978.

\bibitem{palmer85}
E.~M. Palmer.
\newblock {\em Graphical Evolution: An Introduction to the Theory of Random
  Graphs}.
\newblock Wiley Interscience, NY, 1985.

\bibitem{rieffel98}
Eleanor~G. Rieffel and Wolfgang Polak.
\newblock An introduction to quantum computing for non-physicists.
\newblock {Los Alamos} preprint quant-ph/9809016 TR-98-044, FXPAL, Sept. 8
  1998.

\bibitem{selman95}
Bart Selman and Scott Kirkpatrick.
\newblock Critical behavior in the computational cost of satisfiability
  testing.
\newblock {\em Artificial Intelligence}, 81:273--295, 1996.

\bibitem{selman92}
Bart Selman, Hector Levesque, and David Mitchell.
\newblock A new method for solving hard satisfiability problems.
\newblock In {\em Proc. of the 10th Natl. Conf. on Artificial Intelligence
  (AAAI92)}, pages 440--446, Menlo Park, CA, 1992. AAAI Press.

\bibitem{shor95}
P.~Shor.
\newblock Scheme for reducing decoherence in quantum computer memory.
\newblock {\em Physical Review A}, 52:2493--2496, 1995.

\bibitem{shor94}
Peter~W. Shor.
\newblock Algorithms for quantum computation: Discrete logarithms and
  factoring.
\newblock In S.~Goldwasser, editor, {\em Proc. of the 35th Symposium on
  Foundations of Computer Science}, pages 124--134, Los Alamitos, CA, November
  1994. IEEE Press.

\bibitem{sloane73}
Neil Sloane.
\newblock {\em A Handbook of Integer Sequences}.
\newblock Academic Press, NY, 1973.
\newblock On-line version available at
  http://www.research.att.com/$\sim$njas/sequences.

\bibitem{terhal97}
Barbara~M. Terhal and John~A. Smolin.
\newblock Single quantum querying of a database.
\newblock {Los Alamos} preprint quant-ph/9705041 v4, IBM, Nov. 14 1997.

\bibitem{unruh94}
W.~G. Unruh.
\newblock Maintaining coherence in quantum computers.
\newblock {\em Physical Review A}, 51:992--997, 1995.

\bibitem{williams92}
Colin~P. Williams and Tad Hogg.
\newblock Exploiting the deep structure of constraint problems.
\newblock {\em Artificial Intelligence}, 70:73--117, 1994.

\end{thebibliography}

\end{document}