Vertex cover problem studied by cavity method: Analytics and population dynamics
Extracted source text ยท arXiv cond-mat/0302289
\documentclass[twocolumn,showpacs,preprintnumbers,amsmath,amssymb]{revtex4}
\input{psfig.sty}
\begin{document}
\title{
Vertex cover problem studied by cavity method:
Analytics and population dynamics
}
\author{
Haijun Zhou
}
\affiliation{
Max-Planck-Institute of Colloids and Interfaces,
14424 Potsdam, Germany
}
\date{February 14, 2003}
\begin{abstract}
We study the vertex cover problem on finite connectivity random graphs
by zero-temperature cavity method. The minimum vertex cover
corresponds to the ground state(s) of a proposed Ising spin model.
When the connectivity $c>e=2.718282$, there is no state for this
system as the reweighting parameter $y$, which takes a similar
role as the inverse temperature $\beta$ in conventional statistical
physics, approaches infinity;
consequently the ground state energy
is obtained at a finite value of $y$ when the free energy function
attains its maximum value. The minimum vertex cover size
at given $c$ is estimated using population dynamics and compared with
known rigorous bounds and numerical results. The backbone size
is also calculated.
\end{abstract}
\pacs{75.10.Nr,89.75.-k,05.20.-y}
\maketitle
\section{\label{sec.I} Introduction}
The statistical physics of spin glass systems on infinitely connected
lattices (e.g., the Sherrington-Kirkpatrick model) has been well
understood \cite{mezard1987} and research interest is now focused on
systems of finite connectivity (FC). For the limiting example of
FC random lattice, development of the cavity method has
been made in recent few years \cite{mezard2001}. In a FC random lattice,
each vertex interacts only with a finite number of randomly chosen
neighbours; and the local
structure of the lattice is tree-like, with the shortest
distance between two randomly chosen vertices diverging as the
system size goes to infinity. This tree-like property makes it
feasible to study the system by iterative cavity method \cite{note13.02.03}.
The cavity method for $n$-dimensional regular lattice systems is
yet to be worked out, and some aspects of such kind of systems has been
understood using the well-developed mathematical tool of
gauge transformation \cite{toulouse1977,nishimori2002}.
The zero-temperature property of random lattice spin glasses is
especially interesting. Here, the cavity formalism could be
greatly simplified because only the minimum energy states
are relevant \cite{mezard2003}.
Furthermore, many combinatorial optimisation problems in
computer sciences could be studied through a mapping into an appropriate
random spin glass model, examples of which include the
K-sat \cite{monasson1997}, the XOR-sat \cite{ricci2001,caracciolo2002},
the vertex covering problem \cite{weigt2000,weigt2001}, the number partition
problem \cite{mertens1998}, and so on.
Recently, the zero-temperature cavity method was used on the random
K-sat problem and the phase diagram for $k=3$ was obtained
\cite{mezard2002,mezard2002b}.
In this work we hope to improve our understanding on the vertex cover
problem. Consider a random graph $G$ composed of $N$ vertices
$V=\{1,2,\cdots,N\}$. Between any two vertices an edge is present
with probability $c/(N-1)$ and absent with probability $1-c/(N-1)$,
so that on average each vertex has $c$ neighbours
(i.e., the average connectivity is $c$). The resulting edge set of
graph $G$ is denoted by $E(G)$. A
vertex cover of this graph consists of a set of vertices $\Lambda=\{i_1,
i_2,\cdots, i_m\}$, with the property that
if edge $(i,j)\in E(G)$, then either $i\in \Lambda$ or
$j\in \Lambda$ or both \cite{weigt2000}. In the general case there
are many different
ways to cover a graph of size $N$; an interesting question is:
Does there exist a vertex cover of size not exceeding $x N$ ($0<x<1$)?
For large systems it was revealed that a sharp threshold value $x_c (c)$
exists. When $x>x_c(c)$, with probability approaching unity a
vertex cover of size $\leq x N$ could be constructed for a given
graph; while when $x< x_c (c)$ this probability approaches zero.
This sharp threshold is also closely related to computational complexity.
The vertex cover problem is NP-complete \cite{garey1979},
and a time growing exponentially with $N$ may be needed to determine
whether a vertex cover of size $\leq x N$ exists or not for a given graph
\cite{weigt2001b}.
In the case of FC random graphs, when $x > x_c(c)$ or $x < x_c(c)$, it is
relatively easy for a heuristic algorithm to check the existence of
such a vertex cover; however, when $x \sim x_c(c)$, search
complexity increases dramatically \cite{weigt2000}.
For the practical purpose of designing better algorithms, it is
important for us to understand the reason of this easy-hard-easy
transition and to obtain a precise estimate of the threshold value
$x_c(c)$.
An rigorous bound exists for $x_c(c)$ \cite{gazmuri1984}:
$x_l (c) < x_c(c) < 1-\ln c /c$,
where $x_l (c)$ is the root of
\begin{eqnarray}
x_l (c) \ln x_l (c) +[1-x_l (c) ] \ln [1-x_l (c)] \nonumber \\
\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\; +(c/2) [1-x_l (c)]^2=0;
\label{eq:gazmuri}
\end{eqnarray}
furthermore, $x_c (c)$ approaches the following asymptotic form at large $c$
\cite{frieze1990}:
\begin{equation}
\label{eq:frieze}
x_c(c)=1-(2/c)[\ln c -\ln \ln c+1-\ln 2]+o(1/c).
\end{equation}
Using replica method of statistical physics, an analytical
expression for $x_c$ was found in \cite{weigt2000}:
\begin{equation}
\label{eq:weigt}
x_c(c)=1-W(c)/c-W^2(c)/ 2 c,
\end{equation}
where $W(c)$ is the Lambert-W-function defined by $W(c)\exp[W(c)]=c$.
Equation (\ref{eq:weigt}) is exact when
$c\leq e=2.718282$ \cite{bauer2001}. For $c >e$
Eq.~(\ref{eq:weigt}) underestimates the true threshold value,
and for $c>20.7$ it is lower than the rigorous lower bound
Eq.~(\ref{eq:gazmuri}).
The present work focuses on the case of $c> e$. Using zero-temperature
cavity method, we calculate both analytically and numerically the
value of $x_c(c)$ and compare it with Eqs.~(\ref{eq:gazmuri}) and
(\ref{eq:frieze}) and with numerical calculations reported in
\cite{weigt2000}. In Section \ref{sec.II} an energy functional is
introduced. In Section \ref{sec.III} the cavity formalism is
outlined and the free energy expression is given. We investigate
the $y\rightarrow \infty$ situation in Section \ref{sec.IV}
($y$ is a reweighting parameter \cite{mezard2003,mezard2002,mezard2002b},
it plays the role of the inverse
temperature $\beta$ of conventional statistical mechanics).
Section \ref{sec.V} reports the population dynamics results on
the threshold value $x_c(c)$ and on the backbone size.
We conclude the work in Section \ref{sec.VI}.
\section{\label{sec.II} The energy functional}
We attach to each vertex of the random graph $G$ an Ising spin
$\sigma=\{-1,+1\}$. Associated with each spin micro-configuration
is the following energy functional
\begin{equation}
\label{eq:energy}
E[\{\sigma_i\}]=-\sum\limits_{i=1}^N \sigma_i+{\lambda \over 2}
\sum\limits_{(i,j)\in E(G)} (1+\sigma_i)(1+\sigma_j),
\end{equation}
where $\lambda$ is an constant parameter chosen to be greater than unity
\cite{note11.11.02}.
Denote $N_{\rm mvc}(G)$ as the size of the minimum vertex
cover(s) of graph $G$, then the minimum energy
over all the $2^N$ possible spin configurations of graph $G$ is
\begin{equation}
\label{eq:min_energy}
E_{\rm min}=2 N_{\rm mvc} (G) -N.
\end{equation}
Some explanation on Eq.~(\ref{eq:min_energy}). First, $E_{\rm min}$ is
reachable. We denote $\Lambda_{\rm mvc}(G)$ as (one of) the
minimum-sized vertex cover(s), and assign $\sigma=-1$ to vertices in
$\Lambda_{\rm mvc} (G)$ and $\sigma=+1$ to vertices outside. The
energy of this spin configuration is $2 N_{\rm mcv} (G)-N$.
Second, no spin configuration could have lower energy.
To see this, suppose that another spin configuration has lower energy
than Eq.~(\ref{eq:min_energy}). This
spin configuration must contain $N_a <N_{\rm mvc} (G)$ negative spins,
with $N_a+\lambda N_0 < N_{\rm mvc} (G)$, where
$N_0=(1/4)\sum_{(i,j)\in E(G)}(1+\sigma_i) (1+\sigma_j)$.
However, one at most need to change the spin values of $N_0$ vertices
from $+1$ to $-1$ to make
the sum $\sum_{(i,j)\in E(G)}(1+\sigma_i)(1+\sigma_j) =0$, and
the resulting new set of negative spin vertices is a
vertex cover with size $N_a+N_0 <N_{\rm mvc}(G)$. This conflicts with
our original assumption that $N_{\rm mvc} (G)$ is
the minimum vertex cover size.
The problem of finding $x_c (c)$ is converted to finding the average of
$E_{\rm min}/N$ over the random graphs $G$:
\begin{equation}
\label{eq:x_c_cvt}
x_c(c)=(1/2)(1+\overline{E_{\rm min}}/N).
\end{equation}
Here $\overline{(\cdot)}$ means the average of $(\cdot)$ over different
realizations of the random graph. We use cavity method to estimate
$\overline{E_{\rm min}}/N$. In the next section,
the cavity formalism for the present problem is outlined.
The reader is referred to
\cite{mezard1987,mezard2001,mezard2003,mezard2002,mezard2002b} for
more detailed discussion.
\section{\label{sec.III} The zero-temperature cavity formalism}
At zero temperature, only the minimum energy configurations are
relevant. There could be a great many energy local-minima for
Eq.~(\ref{eq:energy}). For very large system size $N$, we group
these configurations into different $``$states''. A state of the
system corresponds to a set of spin micro-configurations. These
spin configurations all have the same energy, which is a
local minimum of Eq.~(\ref{eq:energy}); and two such spin configurations
differ only in a finite number of spin flips.
The average number of states at given density $\epsilon$ of
local minimum energy is assumed to be an exponentially increasing
function of system size $N$, and is characterised by the entropy
density $\Sigma(c,\epsilon)$. We can introduce a reweighting parameter
$y$ and define an zero-temperature free energy density $\Phi(y)$
through the following equation
\begin{equation}
\label{eq:free_energy}
\int d\epsilon \exp[-N y \epsilon+ N \Sigma(c,\epsilon)]
=\exp[-N y\Phi(y)].
\end{equation}
Equation (\ref{eq:free_energy}) has the same form as the conventional
definition of free energy in textbooks of equilibrium statistical
physics. The reweighting parameter $y$ plays the role of inverse
temperature. A large value of $y$ ensures that states with
lower energies will be favoured, provided that such states
exist (i.e., $\Sigma(c,\epsilon)\geq 0$).
Suppose we have a system of $N$ vertices (spins). Now add a
new spin $\sigma_0$ into the system and connect it to $k$ preexisting
spins $\sigma_1,\cdots,\sigma_k$, where $k$ obeys the Poisson distribution
of mean $c$, $P_P(k,c)=e^{-c} c^k /k!$. The energy of the $N$-spin system
at fixed value of the spins $\sigma_1,\cdots,\sigma_k$ is supposed to be
\begin{equation}
\label{eq:e_n}
E^{(N)} (\sigma_1,\cdots,\sigma_k)=A-\sum\limits_{i=1}^k h_i \sigma_i,
\end{equation}
with $A$ being a constant.
In the above equation, $h_i$ is the
local field (called the cavity field) felt by spin $\sigma_i$ in the
absence of $\sigma_0$ in a given macroscopic state. Since the graph is
locally tree-like, as the system size becomes very large, the shortest distance
between two randomly chosen cavity spins also becomes large; therefore,
the cavity field $h_i$ felt by spin $\sigma_i$ becomes independent
of the values of all the other cavity spins \cite{note13.02.03b}.
After the addition of spin $\sigma_0$, the
minimum energy of the $(N+1)$-spin system at fixed value of $\sigma_0$ is
\begin{equation}
\label{eq:e_n+1}
E(\sigma_0)=\left\{\begin{array}{ll}
A-\sigma_0, &\;\;{\rm if}\;\;k=0 \\
A-\sum_{i=1}^k \hat{w}(h_i) & \; \\
\;\;\;\; -[1+\sum_{i=1}^k
\hat{u}(h_i)] \sigma_0, &\;\;{\rm if}
\;\;k\geq 1.
\end{array}
\right.
\end{equation}
Here,
\begin{equation}
\label{eq:wu}
\begin{array}{llll}
\hat{w}(h) = 0 & \;\;{\rm if}\;\;h=1 &\;\;\;{\rm and}
\;\;=|h| &\;\;{\rm if}\;\;{h\leq 0}, \\
\hat{u}(h) = -1 &\;\;{\rm if}\;\;h=1 &\;\;\;{\rm and}\;\;=0
&\;\;{\rm if}\;\;{h\leq 0}.
\end{array}
\end{equation}
(We have used the fact that the cavity fields at zero temperature are
integer-valued and do not exceed unity \cite{note13.02.03b}.) The energy
shift caused by the addition of spin $\sigma_0$ is
$\triangle E_1=-1$ (if $k=0$) and $\triangle E_1=-\sum_{i=1}^k
[\hat{w}(h_i)-|h_i|]-|1+\sum_{i=1}^k \hat{u}(h_i)|$ (if $k\geq 1$).
Equation (\ref{eq:e_n+1}) indicates that the cavity field at
spin $\sigma_0$ is $h_0=1$ (if $k=0$) or $h_0=1+\sum_{i=1}^k \hat{u}(h_i)$
(if $k\geq 1$). By definition \cite{note13.02.03b}, the cavity field
of a spin at each state has unique value; however, its value may
be different for different states. Denote the probability distribution of
the cavity field at vertex $i$ among different states as
$P_i(h)$ (it is called the h-survey in Ref.~\cite{mezard2002b}).
Because different vertices have different local
environments, the h-surveys are different for different vertices.
With the introduction of the reweighting parameter $y$ which favours
lower-energy states, the h-survey at spin $\sigma_0$
is related to those of the cavity spins by
\begin{widetext}
\begin{equation}
\label{eq:h_survey}
P_0(h)=\delta_k^0 \delta(h-1)+
[1-\delta_k^0] C\int \prod\limits_{i=1}^k [P_i(h_i) d h_i ]
\delta [h-1-\sum\limits_{i=1}^k \hat{u}(h_i)] \exp(-y \triangle E_1),
\end{equation}
where $C$ is an normalisation constant.
Eq.~(\ref{eq:h_survey}) is actually a self-consistent equation
for the h-survey. A careful analysis of Eq.~(\ref{eq:h_survey}) leads to
the following expression
\begin{equation}
\label{eq:ph_form}
P(h)=\left\{
\begin{array}{ll}
\sum_{l=0}^\infty \zeta_l \delta(h+l),
& \;\;\;\;\;\;{\rm with}\;\;{\rm probability}\;\;p_1 \\
\delta(h-1), & \;\;\;\;\;\;{\rm with}\;\;{\rm probability}\;\;p_2 \\
\alpha \sum_{l=0}^\infty \zeta_l \delta(h+l) +(1-\alpha) \delta(h-1), &
\;\;\;\;\;\;{\rm with}\;\;{\rm probability}\;\;p_3=1-p_1-p_2
\end{array}\right.
\end{equation}
\end{widetext}
where $\zeta_l\geq 0$ and $\sum_l \zeta_l=1$;
$\alpha\in (0,1)$ is determined by certain probability distribution
$\rho(\alpha)$, and
\begin{eqnarray}
\label{eq:p1p2}
p_1&=& 1-\exp(-c p_2), \nonumber \\
p_2&=& \exp(-c (1-p_1)).
\end{eqnarray}
For very large system size $N$, Eq.~(\ref{eq:free_energy}) suggests
that at fixed value of $y$,
$\Phi(y)=\epsilon-\Sigma(c,\epsilon)/y$, with $\epsilon$ being
implicitly determined by $\partial \Sigma(c,\epsilon)/\partial \epsilon
=y$. An explicit expression for the free energy density could be
obtained by the following way:
After the addition of spin $\sigma_0$, the averaged number of states
of the $(N+1)$-system is
\begin{widetext}
\begin{equation}
\exp[(N+1) \Sigma({2 E(G(N))+2 k \over N+1},{E \over N+1})]
=
\int d \triangle E_1 P^{(1)}(\triangle E_1)
\exp[N \Sigma({2 E(G(N))\over N}, {E-\triangle E_1 \over N})],
\label{eq:entropy}
\end{equation}
where $P^{(1)}(\triangle E_1)$ is the probability distribution
function of the energy shift $\triangle E_1$:
\begin{equation}
\label{eq:p_add}
P^{(1)}(\triangle E_1)=
\delta_k^0 \delta(\triangle E_1 +1)
+(1-\delta_k^0) \int
\prod\limits_{i=1}^k [P_i(h_i) d h_i]
\delta[\triangle E_1+\sum\limits_{i=1}^k (\hat{w}(h_i)-|h_i|)
+|1+\sum\limits_{i=1}^k \hat{u}(h_i)|].
\end{equation}
A logarithm operation is performed on Eq.~(\ref{eq:entropy}),
and the resulting equation is averaged over all the
possible realizations of the cavity fields
and $k$ (this average operation is denoted by an overbar in the
following equation). We obtain that
\begin{equation}
-y \Phi(y)=
\Sigma(c,\epsilon)-y\epsilon
=-c {\partial \Sigma(c,\epsilon) \over \partial c}
+\overline{
\ln \int d \triangle E_1 P^{(1)} (\triangle E_1)
\exp(-y \triangle E_1)}.
\label{eq:add}
\end{equation}
To compute $\partial \Sigma/\partial c$, we setup an edge between
two cavity spins. The energy shift, $\triangle E_2$,
caused by this new edge obeys the following distribution
\begin{equation}
P^{(2)}(\triangle E_2)=\int \prod\limits_{i=1}^2 [d h_i P_i(h_i)]
\delta[\triangle E_2-({\rm min}_{\sigma_1,\sigma_2}
[{\lambda \over 2}(1+\sigma_1)(1+\sigma_2)-h_1 \sigma_1 -h_2 \sigma_2]
+|h_1|+|h_2|)].
\end{equation}
The averaged number of states of the new system is
\begin{equation}
\exp[N \Sigma({2E(G(N))+2 \over N},{E \over N})]=
\int d \triangle E_2 P^{(2)}(\triangle E_2)
\exp[N \Sigma({2 E(G(N)) \over N}, {E -\triangle E_2 \over N})].
\end{equation}
After performing the same procedure as mentioned below
Eq.~(\ref{eq:p_add}), we find that
\begin{equation}
\label{eq:link}
{\partial \Sigma(c,\epsilon) \over \partial c}
=
(1/2)\overline{
\ln \int d \triangle E_2 P^{(2)}(\triangle E_2)
\exp(-y \triangle E_2)}.
\end{equation}
The free energy density expression could be obtained from
Eqs.~(\ref{eq:add}) and (\ref{eq:link}). After taking into
consideration Eq.~(\ref{eq:ph_form}), we arrive at the
following expression for the free energy density:
\begin{eqnarray}
\label{eq:fe_1}
\Phi(y)&=&2 p_1-1-c p_2^2+{c p_2 p_3 \over y} \int d \alpha
\rho(\alpha) \ln (\alpha+(1-\alpha) e^{-2 y}) \nonumber \\
& &-{p_3 \over y} \sum\limits_{m=1}^\infty
{P_P(m,c p_3) \over 1-e^{-c h}}
\int \prod\limits_{i=1}^m [\rho (\alpha_i) d \alpha_i]
\ln (e^{-2 y} +(1-e^{-2 y})\prod\limits_{i=1}^m \alpha_i) \nonumber \\
& &
+{c p_3^2 \over 2 y} \int \prod\limits_{i=1}^2 [ d \alpha_i \rho(\alpha_i)]
\ln (1-(1-e^{-2 y}) \prod\limits_{i=1}^2 (1-\alpha_i)).
\end{eqnarray}
At given $y$, the energy density and the entropy density are calculated by
$\epsilon = \Phi(y)-y{d \Phi(y)/d y}$ and
$\Sigma = y^2 {d \Phi(y)/ d y}$, respectively.
\end{widetext}
\section{\label{sec.IV} The $y\rightarrow \infty$ limit}
At this stage, it is helpful for us to introduce an auxiliary probability
distribution function called the u-survey \cite{mezard2002b}:
%
\begin{equation}
Q_i(u)=C\int d h P (h) \delta(u-\hat{u}(h)) \exp(y (\hat{w}(h)-|h|)).
\end{equation}
%
Based on Eq.~(\ref{eq:ph_form}) we know that
$Q_i(u)=\delta(u)$ (with probability $p_1$) or $Q_i(u)=
\delta(u+1)$ (with probability $p_2$), or
%
\begin{equation}
\label{eq:u_hybrid}
Q_i(u)=\eta \delta(u)+ [1-\eta] \delta(u+1).
\end{equation}
%
The hybrid h-survey and u-survey at a given vertex are related by
$\alpha=\eta/(\eta+(1-\eta) e^y)$. The distribution of the $\eta$
value in the hybrid u-survey Eq.~(\ref{eq:u_hybrid}) is
governed by
%
\begin{widetext}
\begin{equation}
\label{eq:eta}
\rho(\eta)=\sum\limits_{m=1}^{\infty} {P_P(m,c p_3) \over
1-e^{-c p_3}} \int \prod\limits_{i=1}^n [\rho(\eta_i) d \eta_i ]
\delta(\eta-1+{e^{y} \prod\limits_{l=1}^n \eta_l \over
\prod\limits_{l=1}^n [\eta_l+e^{y} (1-\eta_l)]+(e^{y}-1)
\prod\limits_{l=1}^n \eta_l}).
\end{equation}
%
The free energy expression of Eq.~(\ref{eq:fe_1})
could be rewritten in the following form
%
%
\begin{eqnarray}
\label{eq:fe_2}
\Phi(y) &=& p_1-p_2-{c\over 2} [(1-p_1)^2+(p_2)^2]
+{p_3 (1+c p_2) \over y} \overline{\ln [\eta+e^{-y} (1-\eta)]} \nonumber \\
&-& {p_3 [1+c(1-p_1)] \over y} \overline{\ln [1-\eta+e^{-y} \eta]}
+{c p_3^2 \over 2 y} \overline{\ln [e^{-y}+(1-e^{-y})
(\eta_1+\eta_2-2 \eta_1 \eta_2)]}.
\end{eqnarray}
\end{widetext}
%
%
The overbars in Eq.~(\ref{eq:fe_2}) denote
the average over the $\eta$ distribution given by
Eq.~(\ref{eq:eta}).
When $c\leq e$, the only solution of Eq.~(\ref{eq:p1p2}) is that
$p_1=1-p_2$, $p_2=\exp(-c p_2)=W(c)/c$, and $p_3=0$. In this case we
recover Eq.~(\ref{eq:weigt}). $\Sigma(c,\epsilon) =0$ for this solution,
indicating that there is only one state.
It was found that for $c\leq e$, all the vertices of the random graph
could be removed by application of an leaf-removal algorithm \cite{bauer2001}.
When $c >e$ the above-mentioned solution becomes unstable,
and a new solution appears with $p_2=\exp[-c \exp(-c p_2)]
<\exp(-c p_2)$, $p_1=1-\exp(-c p_2)$, and $p_3 >0$. The analysis of
Bauer and Golinelli \cite{bauer2001} reveals that a core with size
proportional to $N$ remains after application of the
leaf-removal algorithm. The core is a strongly connected subgraph, in which
each vertex is connected at least to two other vertices.
To estimate the value of $x_c(c)$ for $c >e$,
let we first consider the limiting situation of $y\rightarrow \infty$.
Equation (\ref{eq:free_energy}) ensures that the minimum $\epsilon$
corresponds to $y=\infty$, provided that the configurational entropy
is non-negative at this limit.
\begin{figure}
\psfig{file=figure01.ps,angle=270,width=1.\linewidth}
\caption{\label{fig:fig01}
The $\eta$ value distribution Eq.~(\ref{eq:eta}) obtained
by population dynamics at $c=6.0$. The $\rho(\eta)$ distribution
strongly depends on the value of the reweighting parameter $y$.
In this figure, three curves at $y=0.1$ (dotted line),
$y=y^*=1.5545$ (corresponding to the free energy maximum, dashed
line), and $y=20$ (solid line) are shown.}
\end{figure}
From Eq.~(\ref{eq:eta}) we know that as $y\rightarrow \infty$,
\begin{equation}
\label{eq:eta_limit}
\rho(\eta)=r_1 \delta(\eta-0^+) +r_2 \delta(\eta-1^-)+r_3 \rho^* (\eta),
\end{equation}
where $\rho^* (\eta)$ is a uniform distribution
over $(0,1)$. Equation~(\ref{eq:eta_limit}) is confirmed
by the population dynamics calculation (Fig.~\ref{fig:fig01}).
It is easy to verify that
\begin{equation}
\label{rrr}
r_1={1-c p_2 \over c p_3},\;\;\;r_2={c+c p_1-\ln c \over c p_3},\;\;\;
r_3={\ln c-1 \over c p_3}.
\end{equation}
Consequently, we obtain from Eq.~(\ref{eq:fe_2}) that
\begin{eqnarray}
\label{eq:1rsb}
\epsilon_{\infty} & = & 1-{\ln^2(c) /2 c}-{\ln (c) / c}
-{3 / 2 c}, \nonumber \\
\Sigma_\infty & = & -{\pi^2 (\ln c -1)^2 / 16 c}.
\end{eqnarray}
The minimum energy value given in the above equation
is an improved lower-bound
of the true average minimum energy. It was obtained also in
\cite{weigt2001} by replica method. However, at $y=\infty$
the entropy density $\Sigma_\infty$ is negative, suggesting that there is no
state at this energy density. Indeed the $x_c(c)$ value
[Eq.~(\ref{eq:x_c_cvt})] of this solution
also exceeds the rigorous lower-bound when $c>27.3$.
The true energy density must be higher than the value
given in Eq.~(\ref{eq:1rsb}).
A better estimate of the minimum energy density could be
obtained by calculating the maximum value of $\Phi(y)$ with respect
to the reweighting parameter $y$ \cite{note11.11.02b}.
This is done in the next
section with population dynamics \cite{mezard2001}.
\section{\label{sec.V} Population dynamics at finite $y$}
To obtain the value of $\Phi(y)$ at any given $y$, the technique of
population dynamics is used \cite{mezard2001}.
A large population of $\eta$'s is generated and this population then
evolves according to Eq.~(\ref{eq:eta}). The averages in
Eq.~(\ref{eq:fe_2}) are calculated numerically. The resulting
estimates of $x_c(c)$ are shown in Fig.~\ref{fig:fig02} and
listed in Table~\ref{tab:tab01}.
\begin{figure}
\psfig{file=figure02.ps,angle=270,width=1.0\linewidth}
\caption{\label{fig:fig02}
The minimum vertex cover size
$x_c(c)$ estimated by population dynamics (solid line and squares)
and its comparison with the exact numerical enumeration values of
\cite{weigt2000} (circles), the replica-symmetric estimation
Eq.~(\ref{eq:weigt}) (dotted line), and the
estimation given by Eq.~(\ref{eq:1rsb}) (dashed line).
In the inset, the $x_c(c)$ value obtained by population dynamics
(solid line) is compared with the rigorous lower-bound Eq.~(\ref{eq:gazmuri})
(dashed line) and the asymptotic value Eq.~(\ref{eq:frieze}) (long dashed
line).
}
\end{figure}
\begin{table}
\caption{\label{tab:tab01}
The $x_c(c)$ value obtained by the population dynamics (column 2)
and its comparison with the exact numerical enumeration value
reported in \cite{weigt2000} (column 4).
The reweighting parameter value $y^*$ of column 3 at
given $c$ corresponds to the maximum of the free energy density
Eq.~(\ref{eq:fe_2}). To determine $y^*$,
Eq.~(\ref{eq:fe_2}) is fitted with two adjustable parameters:
$\Phi(y)=\epsilon_\infty+C_1 /y+C_2 \exp(-y)/y$, where
$\epsilon_\infty$ is given in Eq.~(\ref{eq:1rsb}). The
population size adopted in the present work is $20,000$.
}
\begin{tabular}{cccc}
$c$ & $x_c$ (cavity) & $y^*$ &$x_c$ (enumeration) \\ \hline
$4.0$ & $.5194$ & $1.3093$ & $.523\pm .001$ \\
$5.0$ & $.5603$ & $1.5758$ & \\
$6.0$ & $.5934$ & $1.5545$ & $.599\pm .001$ \\
$7.0$ & $.6210$ & $1.5260$ & \\
$8.0$ & $.6443$ & $1.5259$ & $.656\pm .003$ \\
$9.0$ & $.6643$ & $1.5326$ & \\
$10.0$ & $.6819$ & $1.5434$ & $.697\pm .003$ \\ \hline
\end{tabular}
\end{table}
The threshold curve $x_c (c)$ obtained by the present method lies within the
rigorous bound given by Eq.~(\ref{eq:gazmuri}). It is therefore an improved
estimate compared with Eq.~(\ref{eq:weigt}) and Eq.~(\ref{eq:1rsb}), both
of which exceed the rigorous lower-bound when $c$ is larger than certain
value. However, the values of $x_c$ estimated by the cavity method is
systematically smaller than enumeration results on
finite systems (Fig.~\ref{fig:fig02}).
When the average connectivity $c$ is large,
the estimated $x_c(c)$ value approaches the asymptotic value but
it lies below the asymptotic curve (Fig.~\ref{fig:fig02} inset).
These discrepancies suggest that the $x_c(c)$ threshold value
obtained by the cavity method is not exact; it could
serve as an improved lower-bound of the real threshold curve.
The reason for the failure of the cavity method to obtain exact
threshold values for $c>e$ are discussed in the next section.
\begin{figure}
\psfig{file=figure03.ps,angle=270,width=1.0\linewidth}
\caption{\label{fig:fig03}
Fraction of frozen spins calculated by population dynamics
\cite{note11.11.02c} (squares) and
its comparison with numerical result reported in \cite{weigt2001} and
the replica-symmetric result \cite{weigt2001} (thick solid line).
}
\end{figure}
The cavity field distribution Eq.~(\ref{eq:ph_form}) contains information
on the spin state at a given vertex. A spin $\sigma_i$ will be fixed
at $\sigma_i=+1$ if the local cavity field is distributed according to
$P(h_i)=\delta(h-1)$ or $P(h_i)=\alpha \sum_{l=0}^\infty
\zeta_l \delta(h+l)+(1-\alpha)\delta(h-1)$ but with
$\alpha\rightarrow 0^{+}$ \cite{note11.11.02c}; on the other hand, $\sigma_i$ will
be fixed at $\sigma_i=-1$ with probability $\sum_{l=1}^\infty \zeta_l$
if $P(h_i)=\sum_{l=0}^\infty \zeta_l \delta(h+l)$ or
$P(h_i)=\alpha \sum_{l=0}^\infty \zeta_l \delta(h+l)+(1-\alpha)\delta(h-1)$
but with $\alpha\rightarrow 1^-$ \cite{note11.11.02c}. For the minimum vertex covers
at $y=y^*$, the probability for a randomly
chosen vertex to have fixed spin value is calculated by
population dynamics (see Fig.~\ref{fig:fig03}).
The fraction of frozen spins calculated by the population dynamics
is much lower that that obtained by numerical enumeration \cite{weigt2001}.
This discrepancy may be further indication that the full hierarchy of
replica symmetry breaking is needed to completely describe the
property of the minimum vertex covers. We have noticed
that the fraction of frozen spins strongly depends
on the value of the reweighting parameter $y$. When $y\rightarrow \infty$
the replica-symmetric result of \cite{weigt2001} is recovered;
while for $y\sim 2 y^*$--$3 y^*$ the data obtained by numerical
enumeration is approached.
\section{\label{sec.VI} Discussion}
In this work, we estimate the average minimum vertex cover size $x_c(c)$
for random graphs of finite connectivity $c>e$ in the large $N$ limit.
The obtained $x_c(c)$ curve lies within the rigorous bound
\cite{gazmuri1984} and approaches the asymptotic curve
Eq.~(\ref{eq:frieze}) at large $c$ values. It could be regarded as an
improved lower-bound of the real threshold $x_c(c)$ value.
When $c>e$ the threshold $x_c$ estimated by the cavity method is
not exact. The reason may be the following: The cavity method
is equivalent to one-step replica-symmetry-breaking
\cite{mezard1987,mezard2001,mezard2003}, and the cavity
fields on different vertices are considered as uncorrelated. Because
of the core percolation beyond $c=e$ \cite{bauer2001} in the random
graph, the cavity fields of different vertices may actually be
correlated strongly. To partly account for this effect, one possibility
is to consider also non-integer cavity fields. We hope to return to
this point in a latter work.
The cavity method has inspired very efficient algorithms to
tackle the random K-sat problem.
In the random K-sat problem, there
exists a glassy phase at $y\rightarrow \infty$ for certain range of
the connectivity $c$ \cite{mezard2002,mezard2002b}. In this glassy phase,
the minimum energy is still located at $y=\infty$ but the
complexity is positive, and the algorithm based on
the idea of cavity field \cite{mezard2002b} works
well in this phase. For the vertex cover problem, the present
work suggests that such a $y=\infty$ glass phase does not exist
(this conclusion seems also true for the vertex cover problem on
random hyper-graphs where each $''$edge" is a triangle \cite{weigt2003}).
This indicates that vertex covers of size $N x_c(c)+O(1)$ are extremely
few. It remains to
be seen whether or not algorithms based on cavity method could efficiently
find vertex covers of size slightly beyond $N x_c(c)$ for a given
random graph.
\section*{Acknowledgment}
This work is made possible by a post-doctoral fellowship of
the Max Planck Society.
I thank M. Weigt for suggestions concerned with earlier
versions of this manuscript, and A. K. Hartmann and M. Weigt
for sharing numerical data. I also thank Professor R. Lipowsky and
professor YU Lu for support.
%\bibliography{/usr/people/fornax/zhou/Bibliography/bibliography}
\begin{thebibliography}{24}
\expandafter\ifx\csname natexlab\endcsname\relax\def\natexlab#1{#1}\fi
\expandafter\ifx\csname bibnamefont\endcsname\relax
\def\bibnamefont#1{#1}\fi
\expandafter\ifx\csname bibfnamefont\endcsname\relax
\def\bibfnamefont#1{#1}\fi
\expandafter\ifx\csname citenamefont\endcsname\relax
\def\citenamefont#1{#1}\fi
\expandafter\ifx\csname url\endcsname\relax
\def\url#1{\texttt{#1}}\fi
\expandafter\ifx\csname urlprefix\endcsname\relax\def\urlprefix{URL }\fi
\providecommand{\bibinfo}[2]{#2}
\providecommand{\eprint}[2][]{\url{#2}}
\bibitem[{\citenamefont{M\'{e}zard et~al.}(1987)\citenamefont{M\'{e}zard,
Parisi, and Virasoro}}]{mezard1987}
\bibinfo{author}{\bibfnamefont{M.}~\bibnamefont{M\'{e}zard}},
\bibinfo{author}{\bibfnamefont{G.}~\bibnamefont{Parisi}}, \bibnamefont{and}
\bibinfo{author}{\bibfnamefont{M.~A.} \bibnamefont{Virasoro}},
\emph{\bibinfo{title}{Spin Glass Theory and Beyond}}
(\bibinfo{publisher}{World Scientific}, \bibinfo{address}{Singapore},
\bibinfo{year}{1987}).
\bibitem[{\citenamefont{M\'{e}zard and Parisi}(2001)}]{mezard2001}
\bibinfo{author}{\bibfnamefont{M.}~\bibnamefont{M\'{e}zard}} \bibnamefont{and}
\bibinfo{author}{\bibfnamefont{G.}~\bibnamefont{Parisi}},
\bibinfo{journal}{Eur.\ Phys.\ J.\ B} \textbf{\bibinfo{volume}{20}},
\bibinfo{pages}{217} (\bibinfo{year}{2001}).
\bibitem[{not({\natexlab{a}})}]{note13.02.03}
\bibinfo{note}{The cavity method as outlined in Ref.~\cite{mezard2001} is exact
for a tree. If loops exist in the graph, this method is an approximation and
is equivalent to the first-step replica symmetry breaking (RSB) solution. For
the vertex cover problem, the present work suggests that one needs to go
beyond this first-step RSB solution. To Extend the cavity formalism is both
challenging and important.}
\bibitem[{\citenamefont{Toulouse}(1977)}]{toulouse1977}
\bibinfo{author}{\bibfnamefont{G.}~\bibnamefont{Toulouse}},
\bibinfo{journal}{Comm.\ Phys.} \textbf{\bibinfo{volume}{2}},
\bibinfo{pages}{115} (\bibinfo{year}{1977}).
\bibitem[{\citenamefont{Nishimori}(2002)}]{nishimori2002}
\bibinfo{author}{\bibfnamefont{H.}~\bibnamefont{Nishimori}},
\bibinfo{journal}{Physica A} \textbf{\bibinfo{volume}{306}},
\bibinfo{pages}{68} (\bibinfo{year}{2002}).
\bibitem[{\citenamefont{M\'{e}zard and Parisi}(2003)}]{mezard2003}
\bibinfo{author}{\bibfnamefont{M.}~\bibnamefont{M\'{e}zard}} \bibnamefont{and}
\bibinfo{author}{\bibfnamefont{G.}~\bibnamefont{Parisi}},
\bibinfo{journal}{J. Stat. Phys.} \textbf{\bibinfo{volume}{111}},
\bibinfo{pages}{1} (\bibinfo{year}{2003}).
\bibitem[{\citenamefont{Monasson and Zecchina}(1997)}]{monasson1997}
\bibinfo{author}{\bibfnamefont{R.}~\bibnamefont{Monasson}} \bibnamefont{and}
\bibinfo{author}{\bibfnamefont{R.}~\bibnamefont{Zecchina}},
\bibinfo{journal}{Phys.\ Rev.\ E} \textbf{\bibinfo{volume}{56}},
\bibinfo{pages}{1357} (\bibinfo{year}{1997}).
\bibitem[{\citenamefont{Ricci-Tersenghi
et~al.}(2001)\citenamefont{Ricci-Tersenghi, Weigt, and Zecchina}}]{ricci2001}
\bibinfo{author}{\bibfnamefont{F.}~\bibnamefont{Ricci-Tersenghi}},
\bibinfo{author}{\bibfnamefont{M.}~\bibnamefont{Weigt}}, \bibnamefont{and}
\bibinfo{author}{\bibfnamefont{R.}~\bibnamefont{Zecchina}},
\bibinfo{journal}{Phys.\ Rev.\ E} \textbf{\bibinfo{volume}{63}},
\bibinfo{pages}{026702} (\bibinfo{year}{2001}).
\bibitem[{\citenamefont{Caracciolo and Sportiello}(2002)}]{caracciolo2002}
\bibinfo{author}{\bibfnamefont{S.}~\bibnamefont{Caracciolo}} \bibnamefont{and}
\bibinfo{author}{\bibfnamefont{A.}~\bibnamefont{Sportiello}},
\bibinfo{journal}{J.\ Phys.\ A:\ Math.\ Gen.} \textbf{\bibinfo{volume}{35}},
\bibinfo{pages}{7661} (\bibinfo{year}{2002}).
\bibitem[{\citenamefont{Weigt and Hartmann}(2000)}]{weigt2000}
\bibinfo{author}{\bibfnamefont{M.}~\bibnamefont{Weigt}} \bibnamefont{and}
\bibinfo{author}{\bibfnamefont{A.~K.} \bibnamefont{Hartmann}},
\bibinfo{journal}{Phys.\ Rev.\ Lett.} \textbf{\bibinfo{volume}{84}},
\bibinfo{pages}{6118} (\bibinfo{year}{2000}).
\bibitem[{\citenamefont{Weigt and Hartmann}(2001{\natexlab{a}})}]{weigt2001}
\bibinfo{author}{\bibfnamefont{M.}~\bibnamefont{Weigt}} \bibnamefont{and}
\bibinfo{author}{\bibfnamefont{A.~K.} \bibnamefont{Hartmann}},
\bibinfo{journal}{Phys.\ Rev.\ E} \textbf{\bibinfo{volume}{63}},
\bibinfo{pages}{056127} (\bibinfo{year}{2001}{\natexlab{a}}).
\bibitem[{\citenamefont{Mertens}(1998)}]{mertens1998}
\bibinfo{author}{\bibfnamefont{S.}~\bibnamefont{Mertens}},
\bibinfo{journal}{Phys.\ Rev.\ Lett.} \textbf{\bibinfo{volume}{81}},
\bibinfo{pages}{4281} (\bibinfo{year}{1998}).
\bibitem[{\citenamefont{M\'{e}zard et~al.}(2002)\citenamefont{M\'{e}zard,
Parisi, and Zecchina}}]{mezard2002}
\bibinfo{author}{\bibfnamefont{M.}~\bibnamefont{M\'{e}zard}},
\bibinfo{author}{\bibfnamefont{G.}~\bibnamefont{Parisi}}, \bibnamefont{and}
\bibinfo{author}{\bibfnamefont{R.}~\bibnamefont{Zecchina}},
\bibinfo{journal}{Science} \textbf{\bibinfo{volume}{297}},
\bibinfo{pages}{812} (\bibinfo{year}{2002}).
\bibitem[{\citenamefont{M\'{e}zard and Zecchina}(2002)}]{mezard2002b}
\bibinfo{author}{\bibfnamefont{M.}~\bibnamefont{M\'{e}zard}} \bibnamefont{and}
\bibinfo{author}{\bibfnamefont{R.}~\bibnamefont{Zecchina}},
\bibinfo{journal}{Phys. Rev. E} \textbf{\bibinfo{volume}{66}},
\bibinfo{pages}{056126} (\bibinfo{year}{2002}).
\bibitem[{\citenamefont{Garey and Johnson}(1979)}]{garey1979}
\bibinfo{author}{\bibfnamefont{M.}~\bibnamefont{Garey}} \bibnamefont{and}
\bibinfo{author}{\bibfnamefont{D.~S.} \bibnamefont{Johnson}},
\emph{\bibinfo{title}{Computers and Intractability: A Guide to the Theory of
NP-Completeness}} (\bibinfo{publisher}{Freeman}, \bibinfo{address}{San
Francisco}, \bibinfo{year}{1979}).
\bibitem[{\citenamefont{Weigt and Hartmann}(2001{\natexlab{b}})}]{weigt2001b}
\bibinfo{author}{\bibfnamefont{M.}~\bibnamefont{Weigt}} \bibnamefont{and}
\bibinfo{author}{\bibfnamefont{A.~K.} \bibnamefont{Hartmann}},
\bibinfo{journal}{Phys.\ Rev.\ Lett.} \textbf{\bibinfo{volume}{86}},
\bibinfo{pages}{1658} (\bibinfo{year}{2001}{\natexlab{b}}).
\bibitem[{\citenamefont{Gazmuri}(1984)}]{gazmuri1984}
\bibinfo{author}{\bibfnamefont{P.~G.} \bibnamefont{Gazmuri}},
\bibinfo{journal}{Network} \textbf{\bibinfo{volume}{14}},
\bibinfo{pages}{367} (\bibinfo{year}{1984}).
\bibitem[{\citenamefont{Frieze}(1990)}]{frieze1990}
\bibinfo{author}{\bibfnamefont{A.~M.} \bibnamefont{Frieze}},
\bibinfo{journal}{Discrete Math.} \textbf{\bibinfo{volume}{81}},
\bibinfo{pages}{171} (\bibinfo{year}{1990}).
\bibitem[{\citenamefont{Bauer and Golinelli}(2001)}]{bauer2001}
\bibinfo{author}{\bibfnamefont{M.}~\bibnamefont{Bauer}} \bibnamefont{and}
\bibinfo{author}{\bibfnamefont{O.}~\bibnamefont{Golinelli}},
\bibinfo{journal}{Eur.\ Phys.\ J.\ B} \textbf{\bibinfo{volume}{24}},
\bibinfo{pages}{339} (\bibinfo{year}{2001}).
\bibitem[{not({\natexlab{b}})}]{note11.11.02}
\bibinfo{note}{The value of $\lambda$ is irrelevant as long as $\lambda > 1$:
The bond energy, $(\lambda/2)(1+\sigma_i)(1+\sigma_j)$, equals $0$ or
$2\lambda$; if it is $2 \lambda$, one can choose either $\sigma_i=-1$ or
$\sigma_j=-1$ and the total energy will be lowered. However, choosing
$\lambda\simeq 1$ might facilitate a search algorithm in finding the true
energy minimum.}
\bibitem[{not({\natexlab{c}})}]{note13.02.03b}
\bibinfo{note}{Some more discussion about the cavity fields. At zero
temperature, each state of the system corresponds to an ensemble of spin
microscopic configurations. All these configurations have the same energy
(corresponding to a minimum value of the energy functional
Eq.~(\ref{eq:energy})), and only finite number of spin flips is needed to
transit from one to another configuration of them. The spins on some of the
vertices are fixed at $\sigma=+1$ or $\sigma=-1$ in all these configurations,
while the spins on the remaining vertices are not fixed. We say that those
spins whose value fluctuates from one configuration to another do not feel
any magnetic field in this state ($h=0$). A spin whose value is fixed feels a
non-zero magnetic field. The magnitude of this non-zero cavity field is
defined as half the total increase in minimum energy associated with flipping
this spin. For our present problem, an energy local minimum has value $2
N_{\downarrow} -N$, where $N_{\downarrow}$ is the total number of negatively
valued spins. Therefore, we conclude that (1) all the cavity fields must be
integer-valued, and (2) the cavity fields could not exceed unity ($h\leq
1$).}
\bibitem[{not({\natexlab{d}})}]{note11.11.02b}
\bibinfo{note}{This is understood by the following consideration: According to
Eq.~(\ref{eq:free_energy}), to obtain the energy absolute minimum the
reweighting parameter $y$ should be set to its largest possible value. On the
other hand, $\Sigma=y^2 d\Phi (y)/d y$ changes sign from positive to negative
as $y$ is increased beyond some value. Therefore the largest $y$ value is
obtained by the equation $d \Phi(y)/ d y=0$, which corresponds to the maximum
of $\Phi(y)$.}
\bibitem[{not({\natexlab{e}})}]{note11.11.02c}
\bibinfo{note}{In the numerical calculation, $\alpha\rightarrow 0^+$ is
replaced by the condition $\alpha\leq \alpha_c$ and $\alpha\rightarrow 1^-$
is replaced by $\alpha\geq 1-\alpha_c$ with $\alpha_c$ chosen to be $0.01$.
It turns out that the calculated fraction of frozen spins is sensitive to the
$\alpha_c$ value. A more stringent criterion (say $\alpha_c=0.001$) leads to
even smaller fraction, while a looser criterion (say $\alpha_c=0.05$) leads
to much bigger value.}
\bibitem[{wei()}]{weigt2003}
\bibinfo{note}{M. Weigt and H. Zhou (unpublished).}
\end{thebibliography}
\end{document}