Upload
others
View
0
Download
0
Embed Size (px)
Citation preview
arX
iv:c
ond-
mat
/960
6215
v1 [
cond
-mat
.dis
-nn]
28
Jun
1996
Statistical Mechanics of the Random K-SAT Model
Remi Monasson∗ and Riccardo Zecchina†
∗ Laboratoire de Physique Theorique de l’ENS, 24 rue Lhomond, 75231 Paris cedex 05, France
† Dipartimento di Fisica, Politecnico di Torino, C.so Duca degli Abruzzi 24, I-10129 Torino, Italy
Abstract
The Random K-Satisfiability Problem, consisting in verifying the exis-
tence of an assignment of N Boolean variables that satisfy a set of M = αN
random logical clauses containingK variables each, is studied using the replica
symmetric framework of disordered systems. The detailed structure of the an-
alytical solution is discussed for the different cases of interest K = 2, K ≥ 3
and K ≫ 1. We present an iterative scheme allowing to obtain exact and
systematically improved solutions for the replica symmetric functional or-
der parameter. The caculation of the number of solutions, which allowed us
[Phys. Rev. Lett. 76, 3881 (1996)] to predict a first order jump at the thresh-
old where the Boolean expressions become unsatisfiable with probability one,
is thoroughly displayed. In the case K = 2, the (rigourously known) critical
value (= 1) of the number of clauses per Boolean variable is recovered while
for K ≥ 3 we show that the system exhibits a replica symmetry breaking
transition. The annealed approximation is proven to be exact for large K.
PACS Numbers : 05.20 - 64.60 - 87.10
Typeset using REVTEX
1
I. INTRODUCTION
The emergent collective behaviour observed in a variety of models of statistical mechanics
and in particular in frustrated disordered systems, have been recognized to play a relevant
role also in apparently distant fields such as theoretical computer science, discrete mathe-
matics and complex systems theory [1–3]. Computationally hard problems, characterized by
exponential running time scaling of their algorithms or memory requirements, the so called
NP–complete problems [4], are known to be in one–to–one correspondence with the ground
state properties of spin–glass like models (see [3] and references therein) and the tools and
concepts of statistical physics have shed some new light onto the notion of complexity of
NP-complete problems and have lead to the definition of new search algorithms as the Sim-
ulated Annealing algorithm based on the introduction of an artificial temperature and some
cooling procedures [5].
Very recently, other techniques inspired from statistical mechanics, namely finite size
scaling analysis, have been applied [6] also to the study of universal behaviour in the com-
putational cost (running time) of some classes of algorithms in the course of searching for
solutions of random realizations of the prototype of NP–complete problems, the satisfiability
(SAT) problem we shall discuss. More generally, phase transition concepts are starting to
play a relevant role in theoretical computer science (see [7] and references therein), where the
analysis of general search methods applied to various classes of hard computational prob-
lems characterized by a large number of relevant variables is of crucial importance. In turn,
computer science is a source of highly non–trivial models containing all the paradigms nec-
essary to a deeper undestanding of the physical properties of disordered frustrated systems,
in particular diluted models for which the theoretical framework is still to be completed [8].
One of the major theoretical open questions in this context would be to understand how
algorithmic complexity and spin-glass transition, the so–called replica symmetry breaking
transition [3], are related.
Among the known NP–complete problems, the SAT problem is at the same time the
2
root problem of complexity theory [4] and a prototype model for phase transition in random
combinatorial structures [1,9]. SAT was the first problem proved to be NP–complete by
S. Cook in 1971 [10] and opened the way to the identification of a vast family of other
NP–complete problems for which a polynomial reduction to SAT became available [4]. In
particular the K–satisfiability (K–SAT) problem, a version of SAT we shall discuss in great
detail in what follows, beside playing a central role in NP–completeness proving procedures
[4], is a widespread test for the evaluation of performance of combinatorial search algorithms,
due the intractability concentration phenomena taking place in random samples generated
near criticality.
In a recent work [2], we have shown that the methods of statistical mechanics of random
systems allow to compute some algorithmically relevant quantities such as the typical entropy
of the problem, i.e. the typical number of its solutions, and to clarify the nature of the
threshold behaviour. The scope of this paper is twofold. On the one hand, we aim at
giving a thorough discussion of the analytical derivation of the above results, mainly the
calculation of the entropy jump at the transition. On the other hand, we expose in details
the replica symmetric theory of the K–SAT problem and show how to go beyond the simplest
solution proposed in our previous work [2]. The paper is organized as follows. Section II is
devoted to the presentation of the K–SAT problem and of the known exact results. Sections
III contains an outline of the statistical mechanics approach whereas the replica symmetric
solutions are exposed in Section IV. In the successive sections, fromV toVIII, the outcomes
of the analytical calculations are exposed in detail for the different values of K of interest.
Conclusions and new perspectives will be briefly discussed in Section IX.
II. THE K–SAT PROBLEM AND A BRIEF SURVEY OF KNOWN RESULTS
Given a set of N Boolean variables {xi = 0, 1}i=1,...,N , we first randomly choose K among
the N possible indices i and then, for each of them, a literal zi that is the corresponding xi
or its negation xi with equal probabilities one half. A clause C is the logical OR of the K
3
previously chosen literals, that is C will be true (or satisfied) if and only if at least one literal
is true. Next, we repeat this process to obtain M independently chosen clauses {Cℓ}ℓ=1,...,M
and ask for all of them to be true at the same time, i.e. we take the logical AND of the
M clauses thus obtaining a Boolean expression in the so called Conjunctive Normal Form
(CNF). The resulting K–CNF formula F may be written as
F =M∧
ℓ=1
Cℓ =M∧
ℓ=1
(
K∨
i=1
z(ℓ)i
)
, (1)
where∧
and∨
stand for the logical AND and OR operations respectively.
A logical assignment of the {xi}’s satisfying all clauses, that is evaluating F to true, is
called a solution of the K–satisfiability problem. If no such assignment exists, F is said to
be unsatisfiable.
When the number of clauses becomes of the same order as the number of variables
(M = α N) and in the large N limit – indeed the case of interest also in the fields of
computer science and artificial intelligence [9,11] – the K–SAT problem exhibits a striking
threshold phenomena. Numerical experiments have shown that the probability of finding a
correct Boolean assignment falls abruptly from one down to zero when α crosses a critical
value αc(K) of the number of clauses per variable. Above αc(K), all clauses cannot be
satisfied any longer and one gets interested in minimizing the number of unsatisfiable clauses,
which is the optimization version of K–SAT also referred to as MAX–K–SAT. Moreover, near
αc(K), heuristic search algorithms get stuck in non–optimal solutions and a slow down effect
is observed (intractability concentration). On the contrary, far from criticality heuristic
processes are typically rather efficient [6].
Very schematically, the known results on K–SAT which have been obtained in the frame-
work of complexity theory may be summarized as follows.
– For K = 2, 2–SAT belongs to the class P of polynomial problems [12]. P is defined as
the set of computational problems whose best solving algorithms have running times
increasing polynomially with the number of relevant variables [4]. For α > αc, MAX–
2–SAT is in NP–complete [4] : NP–complete problems are the hardest nondeterministic
4
polynomial problems, whose solutions may be found by the exhaustive inspection of a
decision tree of logical depth growing in a polynomial way with the number of relevant
variables; it is generally thought that the running times of their best solving algorithms
scale exponentially with the number of relevant variables [4]. The mapping of 2–SAT
on directed graph theory [13] allows to derive rigorously the threshold value αc = 1
and an explicit 2–SAT polynomial algorithm working for α < αc has been developed
[12].
– For K ≥ 3, both K–SAT and MAX–K–SAT belong to the NP–complete class. Only
upper and lower bounds on αc(K) are known from a rigorous point of view. Remark-
able application of finite size scaling techniques has, recently, allowed to find precise
numerical values of αc for K = 3, 4, 5, 6 [1].
– For K ≫ 1, clauses become decoupled and the asymptotic expression αc ≃ 2K ln 2 can
be easily found.
For brevity, we do not discuss here the results concerning the algorithmic approaches to
K–SAT and MAX–K–SAT [12,14,15]. We just mention that MAX–K–SAT belongs to the
subclass of NP–complete problems which allows for a Polynomial Approximation Scheme
for quasi–optimal solutions [14]. A recent numerical study of the critical behaviour in the
computational cost of satisfiability testing can be found in [6].
For α = MN
> 0, K–SAT can be cast in the framework of statistical mechanics of random
diluted systems by the identification of an energy–cost function E(K,α) equal to the number
of violated clauses [9,2]. The study of its ground state allows then to address the optimization
version of the K–SAT problem as well as to characterize the space of solutions by its typical
entropy, i.e. the degeneracy of the ground state. The vanishing condition on the ground
state energy for a given K, corresponds to the existence of a solution to the K–SAT problem
and thus identifies a critical value αc(K) of α below which random formulas are satisfiable
with probability one. For α > αc(K), the ground state energy becomes non zero and
5
gives information on the maximum number of satisfiable clauses, i.e. on the MAX–K–
SAT problem. Previous works on the statistical mechanics of combinatorial optimization
problems - like Traveling Salesman, Graph Partitioning or Matching problems [3,16–18] -
focused mainly on the comparison between the typical cost of optimal configurations and
the algorithmic results. The issues arising in K–SAT are of different nature, and the key
quantity to be discussed [19] is rather the typical number of existing solutions, i.e. the
ground state typical entropy SK(α).
A crucial rigorous result on which the whole statistical mechanics approach is founded
concernes the self-averageness taking place in MAX–K–SAT. For any K, independently of
the particular but randomly chosen sample of M clauses, the minimal fraction of violated
clauses is narrowly peaked around its mean value when N → ∞ at fixed α [15].
III. STATISTICAL MECHANICS OF THE K–SAT AND MAX–K–SAT COST
FUNCTION
As discussed above, we map the random SAT problem onto a diluted spin energy–cost
function through the introduction of spin variables, Si = 1 if the Boolean variable xi is true,
Si = −1 if xi is false. The clauses structure is taken into account by a M × N quenched
random matrix ∆ where ∆ℓ,i = −1 (respectively +1) if clause Cl contains xi (resp. xi), 0
otherwise. Then the function
E[∆, S] =M∑
ℓ=1
δ
[
N∑
i=1
∆ℓ,i Si;−K
]
(2)
where δ[i; j] denotes the Kronecker symbol, turns out to be equal to the number of violated
clauses in that the quantity∑N
i=1 ∆ℓ,i Si equals −K if and only if all Boolean variables in
the ℓ–th clause take the values opposite to the desired ones, i.e. iff the clause itself is false.
Moreover, we impose the constraints
N∑
i=1
∆2ℓ,i = K, ∀ℓ = 1, ..., M (3)
6
to ensure that the number of Boolean variables in any clause is exactly equal toK. Then, the
ground state (GS) properties of the cost function (2) will reflect those of K–SAT (EGS = 0)
and MAX–K–SAT (EGS > 0). In (2), one may interpret K as the number of “neighbours”
to which each spin is coupled inside a clause. To study the ground state properties fo the
cost function (2), we follow the replica approach in the framework of diluted models which is
indeed much more complicated than that of long–range fully connected disordered models.
As we shall see below, replica theory must be formulated in a functional form involving not
only interactions between pairs of replicas but all multi–replicas overlaps.
To compute the GS energy, we first introduce a fictitious temperature 1/β to regularize
all mathematical expressions and send β → ∞ at the end of the calculation. Note that the
introduction of a finite temperature also greatly helps to understand the physical proper-
ties of the model. We procede by computing the model “free–energy” density at inverse
temperature β, averaged over the clauses distribution
F (β) = − 1
βNlnZ[∆] , (4)
where Z[∆] is the partition function
Z[∆] =∑
{Si}
exp (−βE[∆, S]) . (5)
As already mentioned, the energy (2) is self-averaging and should therefore be obtained from
the above free–energy. The overline denotes the average over the random clauses matrices
satisfying the constraint (3) and is performed using the replica trick lnZ = limn→0Zn−1
n,
starting from integer values of n. The typical properties of the ground state, i.e. the internal
energy and the entropy, will then be recovered in the β → ∞ limit.
The averaged integer moments of the partition function are given by the following formula
Z[∆]n =∫ n/2∏
r=1
∏
a1<...<a2r
dQa1,...,a2rdQa1,...,a2reNF [Q,Q] , (6)
where
7
F [Q, Q] =∑
r,{ai}
Qa1,...,a2rQa1,...,a2r + ln
∑
{Sa=±1}
exp
−∑
r,{ai}
Qa1,...,a2rSa1 . . . Sa2r
+ α ln∑
{σa}
∫ 2π
0
n∏
a=1
dσa
2πcos(
∑
a
σaσa)∏
a
(
1 + (e−β − 1)δ[σa;−K])
×
∏
a
cos(σa)K
1 +∑
r,{ai}
(−1)rQa1,...,a2r2r∏
i=1
tan(σai)
K
, (7)
in which∑
r,{ai} ≡∑n/2
r=1
∑
a1<...<a2r and where the sums over all the σa’s run over the integers
between −K and K. The order parameters Qa1,...,a2r are the typical overlaps between spins
configurations belonging to different replicas
Qa1,...,a2r =1
N
N∑
i=1
Sa1i . . . Sa2r
i (8)
and the Qa1,...,a2r are their conjugated Lagrange multipliers.
In the large N,M limit (with fixed α = M/N), the partition function (6) may be
evaluated by taking the saddle-point over all order parameters Q, Q. Since the function
F [Q, Q] is invariant under permutation of replicas, a possible natural saddle-point can be
sought within the so called replica symmetric (RS) Ansatz [17,18]
Qa1,...,a2r = Qr , Qa1,...,a2r = Qr , (9)
which preserves invariance.
Very schematically, the computation of all the three terms appearing on the right hand
side of equation (7) procedes as follows.
i) First term :
n/2∑
r=1
∑
a1<...<a2r
Qa1,...,a2rQa1,...,a2r =n/2∑
r=1
(
n
2r
)
Qr Qr , (10)
ii) Second term : we compute the trace over the set {Sa} by means of the following
relation
(
n∑
a=1
Sa
)2r
=r∑
j=0
C(j, r, n)∑
a1<...<a2j
Sa1 . . . Sa2j , (11)
8
where C(j, r, n) is given by
C(j, r, n) =∂2r
∂x2rcoshn(x) tanh2j(x)|x=0 . (12)
We then find
∑
{Sa=±1}
exp
−∑
r,{ai}
Qa1,...,a2rSa1 . . . Sa2r
= exp(−D0)× (13)
∫ ∞
−∞
dxdx
2πcos(xx) (2 cosh(x))n exp
−n/2∑
i=1
(−1)iDix2i
(14)
where D0 = −∑nj=1C(0, j, n)Dj and all order parameters Dj (j = 1, . . . , n/2) are
solutions of the linear equations
Qr =n/2∑
j=r
C(r, j, n)Dj , ∀ 1 ≤ r ≤ n/2 (15)
iii) Third term : we rewrite the argument of the logarithm as
1
(2π)K
∫ 2π
0dx1 . . . dxK
n/2∑
r1=0
. . .n/2∑
rK=0
Qr1 . . . QrK exp
(
4iK∑
l=1
xlrl
)
×(
1 + (e−β − 1)K∏
l=1
e−ixl cos(xl)
)n
(16)
where Q0 ≡ 1. In the limit n → 0, we expand the logarithm stemming from the last
factor (elevated to the power n) in the above expression.
Putting all three terms together, the RS value of the free–energy reads
1
nF [{Qr, Qr}] =
∞∑
j=1
ajQj −∞∑
i,j=1
bijQiDj − α∞∑
l=1
(1− e−β)l
l 2Kl
1 +l/2∑
i=1
Qi
(
l
2i
)
K
+
∫ ∞
−∞
dxdx
2πcos(xx) ln (2 cosh(x)) exp
(
−∞∑
i=1
(−1)iDix2i
)
, (17)
where
aj = (−1)j∂2j
∂x2jln(cosh(x))|x=0 , bij =
1
(2i)
∂2j
∂x2jtanh2i(x)|x=0 . (18)
9
Within the replica symmetric Ansatz, equations (8) and (9) teach us that Qi is the disorder–
averaged 2ith-moment of the local magnetization of spin Si : Qi = 〈S〉2 i . It results conve-
nient to introduce the probability distribution P (m) of the Boolean magnetization, m = 〈S〉,
such that
Qi =∫ 1
−1dm P (m) m2 i , (19)
together with a similar function P (y) relative to the conjugate parameters
P (y) =∞∑
i=1
Di∂2iδ
∂y2i(y) . (20)
The free energy then reads
− βF (β) = ln 2 +∫ ∞
−∞
dudv
2πcos(uv) ln (cosh v) exp
(
−∫ ∞
−∞dyP (y) cos(uy)
)
+
∫ 1
−1dx∫ ∞
−∞dyP (x)P (y) ln (1 + x tanh y) +
∫ ∞
−∞dyP (y) ln (cosh y) +
α∫ 1
−1
K∏
ℓ=1
dxℓP (xℓ) ln
[
1 + (e−β − 1)K∏
ℓ=1
(
1 + xℓ
2
)
]
(21)
where by definition P (x) and P (y) are even functions (more precisely distributions) satisfying
the saddle point functional equations
P (x) =1
1− x2
∫ ∞
−∞
du
2πcos
[
u
2ln(
1 + x
1− x
)]
exp[
−∫
dyP (y) cos(uy)]
, (22)
and
P (y) = αKδ(y)− αK∫ 1
−1
K−1∏
ℓ=1
dxℓP (xℓ)1
2
[
δ(
y +1
2lnA(K−1)
)
+ δ(
y − 1
2lnA(K−1)
)]
,
(23)
with
A(K−1) ≡ A(K−1)({xℓ} , β) = 1 + (e−β − 1)K−1∏
ℓ=1
(
1 + xℓ
2
)
. (24)
Eliminating P (y) leads to a self–consistent equation for the magnetization distribution
P (x) =1
1− x2
∫ ∞
−∞du cos
[
u
2ln(
1 + x
1− x
)]
×
exp
[
−αK + αK∫ 1
−1
K−1∏
ℓ=1
dxℓP (xℓ) cos(
u
2lnA(K−1)
)
]
(25)
10
and to the following expression of the free–energy
− βF (β) = ln 2 + α(1−K)∫ 1
−1
K∏
ℓ=1
dxℓP (xℓ) lnA(K) +
αK
2
∫ 1
−1
K−1∏
ℓ=1
dxℓP (xℓ) lnA(K−1) −1
2
∫ 1
−1dxP (x) ln(1− x2) . (26)
Note that in eq.(26) A(K) si given by a formula similar to (24), where the upper bound of
the product is replaced by K. To end with, let us remark that equation (25) can in turn be
transformed into an integro–differential equation
∂P (x)
∂α= −KP (x) +K
∫ 1
−1
K−1∏
ℓ=1
dxℓ
[
P (x1) + α(K − 1)∂P (x1)
∂α
]
P (x2) . . . P (xK−1)×
1
2
[
∂η(x)
∂xP (η(x)) +
∂η(−x)
∂xP (η(−x))
]
(27)
where η(x) = [(x + 1)A(K−1) + x − 1]/[(1 + x)A(K−1) + 1 − x] and for which the boundary
condition is given by the solution of (25) in α = 0:
P (x)|α=0 = δ(x) . (28)
IV. A TOY MODEL : THE K = 1 CASE
The K = 1 case can be solved either by a direct combinatorial method or within our
statistical mechanics approach. Though this particular case does not present any critical
behaviour, its study will turn out to be useful in understanding the K > 1 models in which
we are interested. Moreover, the K = 1 toy model allows to check the correctness of the
statistical mechanics results.
In this case, a sample of M clauses is completely described by giving directly the num-
bers ti and fi of clauses imposing that a certain Boolean variable Si must be true or false
respectively. Therefore the partition function corresponding to a given sample reads
Z[{t, f}] =N∏
i=1
(e−βti + e−βfi) , (29)
and the average over the disorder gives
11
1
NlnZ[{t, f}] = 1
N
∑
{ti,fi}
M !∏N
i=1(ti!fi!)lnZ[{t, f}]
= ln 2− αβ
2+
∞∑
l=−∞
e−αIl(α) ln
(
cosh
(
βl
2
))
, (30)
where Il denotes the lth modified Bessel function. The zero temperature limit gives the
ground state energy
EGS(α) =α
2[1− e−αI0(α)− e−αI1(α)] (31)
and the ground state entropy
SGS(α) = e−αI0(α) ln 2 . (32)
One may notice that for any α > 0, the ground–state energy is positive. Therefore, the
clauses are never satisfiable all together and the overall function (1) is false with probability
one. Nonetheless, the entropy is finite, implying an exponential degeneracy of the ground–
state describing the minimum number N.EGS(α) of unsatisfiable clauses. Such a degeneracy
is due to the presence of a finite fraction of variables e−αI0(α) which are subject to equal
opposite constraints ti = fi, and whose corresponding spins may be chosen up or down
indifferently without changing the energy.
The above results are recovered in our approach leading to the conclusion that the RS
Ansatz is exact for all β and α when K = 1. Equation (25) can be explicitly solved at any
temperature 1/β and the solution reads
P (x) =∞∑
ℓ=−∞
e−αIℓ(α) δ
(
x− tanh
(
βℓ
2
))
, (33)
which, in the limit of physical interest β → ∞, becomes
P (x) = e−αI0(α)δ(x) +1
2(1− e−αI0(α)) (δ(x− 1) + δ(x+ 1)) . (34)
The finite value of the ground state entropy may be ascribed to the existence of unfrozen
spins whose fractional number is simply the weight of the δ–function in x = 0. At the same
time, it appears that the non zero value of the ground state energy is due to the presence
12
of completely frozen spins of magnetizations x = ±1. This is an important feature of the
problem which remains valid for any K, as we shall see in the following. In Fig. 1 we report
the plots of the above energy and entropy at zero temperature.
V. REPLICA SYMMETRIC SOLUTIONS FOR ALL K
A relevant general mechanism for the comprehension of the overcoming critical behaviour
in K–SAT is the accumulation of Boolean magnetizations 〈S〉 = ±(1−O(e−|z|β)), (z = O(1)),
in the limit of zero temperature and for α → αc. The emergence of Dirac peaks in x = ±1
signals that a freezing process has just occurred and that a further increase of α beyond αc
would cause the appearance of unsatisfiable clauses. This scenario – which can be verified
by inspection of eq. (33) for K = 1 – is true also for K > 1. In fact, by computing the
fraction of violated clauses through
E = − 1
N
∂
∂βlnZ[∆] , (35)
at temperature 1/β, one sees that the ground state energy depends only upon the magne-
tizations of order ±(1 − O(e−|z|β)), if any, and that such contributions can be described by
the introduction of the new rescaled function
R(z) = limβ→∞
[
P
(
tanh
(
βz
2
))
∂
∂ztanh
(
βz
2
)]
, (36)
which, from (25), fulfills the saddle–point equation
R(z) =∫ ∞
−∞
du
2πcos(uz) exp
[
− αK
2K−1+ αK
∫ ∞
0
K−1∏
ℓ=1
dzℓR(zℓ) cos(u min(1, z1, . . . , zK−1))
]
.
(37)
The corresponding ground state energy reads, see (26) and (36),
EGS(α) = α(1−K)∫ ∞
0
K∏
ℓ=1
dzℓR(zℓ)min(1, z1, . . . , zK) +
αK
2
∫ ∞
0
K−1∏
ℓ=1
dzℓR(zℓ)min(1, z1, . . . , zK−1)−∫ ∞
0dzR(z)z . (38)
13
It is easy to see that the saddle–point equation (37) is in fact a self–consistent identity for
R(z) in the range z ∈ [0, 1] only. Outside this interval, equation (37) is merely a definition
of the functional order parameter R. This remark will be useful in the following.
To start with, R(z) = δ(z) is obviously a solution of (37) for all values of α and K, giving
a zero ground state energy since no spin are frozen with magnetizations ±1. Let us assume
that R(z) includes another Dirac peak in 0 < z0 ≤ 1. Then, inserting this distribution in the
exponential term on the r.h.s. of (37), we find that R(z) on the l.h.s. necessarily includes all
Dirac peaks centered in kz0, where k = 0,±1,±2,±3, . . .. We now plunge again the whole
series in the r.h.s. of (37). For large enough k, kz0 is larger than one and the exponentiated
term includes a cosu contribution, which causes the presence of Dirac distributions centered
in all (positive and negative) integers. Therefore, as soon as R(z) is different from δ(z),
it contains an infinite set of Dirac functions peaked around all integer numbers. Clearly,
the simplest self–consistent solution to (37) will be obtained for z0 = 1 since the process
described above closes after one iteration. This solution reads [2]
R(z) =∞∑
ℓ=−∞
e−γ1Iℓ(γ1)δ(z − ℓ) , (39)
where γ1 depends on K and α and fulfills the implicit equation
γ1 = αK
[
1− e−γ1I0(γ1)
2
]K−1
. (40)
The physical meaning of γ1 may be understood by looking at the definition of the rescaled
function order parameter (36). Turning back to the magnetization distribution, we indeed
find in the zero temperature limit
P (x) = e−γ1I0(γ1)Pr(x) +1
2(1− e−γ1I0(γ1)) (δ(x− 1) + δ(x+ 1)) , (41)
where Pr(x) is a regular (i.e. without Dirac peaks in x = ±1) magnetization distribution
normalized to unity. The above identity is a straigthforward extension of the expression (34)
(when K = 1, γ1 = α from (40) and Pr(x) = δ(x)) to any value of K. Inserting eq.(39) in
(38) gives t the value of the cost–energy
14
EGS(α) =γ12K
(
1− e−γ1I0(γ1)−Ke−γ1I1(γ1))
. (42)
It is therefore clear that, in the RS context, the SAT to UNSAT transition corresponds to
the emergence of peaks centered in x = ±1 with finite weights, that is to a transition from
γ1 = 0 to γ1 > 0. This simplest solution was presented in ref. [2].
In addition to (39), there exist other RS solutions to the saddle–point equations. For
instance, if we choose z0 =12, the insertion process ends up after two iterations and generates
Dirac peaks centered in all integer and half–integer numbers. More generally, for any integer
p ≥ 1, we may define the solution to (37)
R(z) =∞∑
l=−∞
rℓ δ
(
z − ℓ
p
)
, (43)
having exactly p peaks in the interval [0, 1[, whose centers are zℓ =ℓp, ℓ = 0, . . . , p− 1. The
coefficients rℓ of these distributions are self–consistently found through
rℓ =∫ 2π
0
dθ
2πcos(ℓθ) exp
p∑
j=1
γj(cos(jθ)− 1)
(44)
for all ℓ = 0, . . . , p− 1 where
γj = αK
1
2− r0
2−
j−1∑
ℓ=1
rℓ
K−1
−
1
2− r0
2−
j∑
ℓ=1
rℓ
K−1
, ∀j = 1, . . . , p− 1
γp = αK
1
2− r0
2−
p−1∑
ℓ=1
rℓ
K−1
. (45)
The corresponding energy reads, from (43) and (38),
EGS =α(1−K)
p
(
1− r02
)K
+p−1∑
j=1
1− r02
−j∑
l=1
rl
K
+
αK
2p
(
1− r02
)K−1
+p−1∑
j=1
1− r02
−j∑
l=1
rl
K−1
−p∑
j=1
j
pγj
r02+
rj2+
j−1∑
l=1
rl
. (46)
Note that the last term of (46) includes the coefficient rp, which may be computed using
identity (44). It is easy to check that the first non trivial solution (39) corresponds to p = 1.
Though there might be continuous solutions to (37), we believe they can be reasonably
15
approximated by the large p solutions we have presented here. In the following sections,
we shall therefore analyze which are the physical implications of the above solutions in the
different cases of interest, K = 2, K ≥ 3 and K >> 1.
VI. THE K = 2 CASE
The case K = 2 is the first relevant instance of K–SAT. Graph theory has allowed [13]
to show that for α = αc(2) = 1 the problem undergoes a satisfiability transition which can
be also viewed as a P/ NP–complete transition, from 2–SAT to MAX–2–SAT.
Let us first consider the simplest p = 1 RS solution [2]. Self–consistency equation (40)
leads to the solution γ1 = 0 for any α. However, for α > 1 one finds another solution
γ1(α) > 0 which maximizes the free–energy (and EGS) and therefore must be chosen (this is
a well known peculiar aspect of the replicas formalism [3]). When approaching the threshold
from upside, we indeed find
EGS(α|p = 1) =4
27(α− 1)3 +O
(
(α− 1)4)
≃ 0.1481 (α− 1)3 . (47)
As expected, the p = 1 RS theory predicts EGS = 0 for α ≤ 1 and EGS > 0 when α > 1,
giving back the rigorous result αc(2) = 1 : for α > 1 the fraction of violated clauses becomes
finite and the corresponding CNF formulas turn out to be false with probability one. The
transition taking place at αc is of second order with respect to the order parameter γ1 and
is accompanied by the progressive appearance of two Dirac peaks for P (x) in x = ±1 with
equal amplitudes (1− e−γ1I0(γ1))/2.
It is straightforward to verify that RS solutions with p ≥ 2 are not present below α = 1.
However, above the threshold, one has to check whether their ground state energy are larger
than the one of the p = 1 solution, that is if they can be relevant for MAX–K–SAT. For
p = 2, resolution of equations (44) and (45) close to αc(2) leads to (discarding the choice
r1 = 0 which amounts to the p = 1 solution)
r0 = 1− 8 + 2√2
7(α− 1) +O
(
(α− 1)2)
16
r1 =3−
√2
7(α− 1) +O
(
(α− 1)2)
(48)
for the coefficients of the Dirac peaks in z = 0 and z = 12respectively. Inserting these
expansion into the energy (46), one finds
EGS(α|p = 2) =9 + 4
√2
98(α− 1)3 +O
(
(α− 1)4)
≃ 0.1496 (α− 1)3 , (49)
which is slightly larger than the p = 1 result (47). Numerical calculations for higher values
of p ≥ 3 confirm that the energy increases very slowly with p. We have found that for
large p’s the ground state energy is almost stationary, so that the p = 10 solution can be
considered as a very fair approximation of the optimal p → ∞ RS solution. The coefficients
rℓ of the distributions present in the order parameter R(z) (43) are displayed Fig. 2 for
different values of α and in the cases p = 1, p = 5 and p = 10.
The ground state energy predicted by the p = 10 RS solution is compared to numerical
exhaustive simulations carried out for small sized systems on Fig 3. For α > αc = 1, the
theoretical estimate of EGS seems to sligthly deviates from the numerical findings, which
would signal the occurrence of a Replica Symmetry Breaking (RSB) transition at the thresh-
old. If it were so, the situation would be reminiscent of the case of neural networks with
continuous weights, where RS theory is able to localize the storage capacity but not to
predict the minimal fraction of errors beyond the transition [20,21]. The extrapolation of
the simulations results for finite size systems to N → ∞ is shown Fig. 4 for the particular
choice α = 3. Data seem in favor of RSB but one cannot exclude that 1/N2 effects could
make coincide both numerics and theory. Moreover, for α ≫ 1, the exact asymptotic scaling
of the ground state energy EGS ≃ α/4 [15] is compatible with the RS prediction. In this
work we are primarily concerned with the K–SAT problem and therefore we shall leave the
detailed analysis of the validity of RS theory above αc for further studies [22].
From the above discussion, it is however reasonable to believe that RS theory is exact
in the region 0 ≤ α ≤ αc = 1 at least [23]. As already mentioned, the key quantity to study
in this range is the typical number of solution to the problem, i.e. the typical ground state
17
entropy SGS(α) given by eq.(26) in the β → ∞ limit. Notice that a simpler expression of
the ground state entropy, more precisely of its derivative, may be obtained by differentiating
(26) with respect to α and using the saddle-point equation (27). The result reads
∂SGS
∂α(α) =
∫ 1
−1
K∏
ℓ=1
dxℓ P (xℓ) ln
[
1−K∏
ℓ=1
(
1 + xℓ
2
)
]
, (50)
and is valid for any α and K. Using the initial value SGS|α=0 = ln 2 and the above equation
(50), one can in principle compute the ground state entropy for any value of α. However,
due to the difficulty of finding a solution of the integral equation (25), it turns out to be
convenient to develop a systematic expansion of the entropy in the parameter α. We now
briefly present the procedure to be employed for a generic value of K.
Inserting P (x)|α=0 = δ(x) into formula (50), we obtain the slope of the entropy at the
origin
∂SGS
∂α
∣
∣
∣
∣
∣
α=0
= ln(
1− 1
2K
)
, (51)
which coincides with the annealed result [1,19]. Then, we use eq.(27) to compute the first
derivative of the magnetizations distribution in α = 0,
∂P (x)
∂α
∣
∣
∣
∣
∣
α=0
= −αKδ(x) +αK
2δ(
x+1
2K − 1
)
+αK
2δ(
x− 1
2K − 1
)
. (52)
Now, we differentiate eq.(50) with respect to α and inject the above result, which is needed
to obtain the second derivative of the ground state entropy at α = 0,
∂2SGS
∂α2
∣
∣
∣
∣
∣
α=0
= −K2 ln(
1− 1
2K
)
+K2
2ln(
1− 1
2K − 1
)
+K2
2ln
(
1− 2K−1 − 1
2K−1(2K − 1)
)
,
(53)
which is negative as required since the entropy is expected to be a concave function of α.
The whole procedure, consisting in successive differentiations of eqs.(27) and (50) can then
be iterated to compute symbolically all the derivatives of P (x) and SGS(α) with respect to
α in α = 0.
18
In the K = 2 case, we have calculated this way the power expansion of SGS(α) up to the
seventh order in α (which implies an uncertainty less than one percent with respect to the
sixth order Taylor expansion on the range α ∈ [0; 1]). The result reads
SGS(α) = ln 2− 0.28768207 α− 0.01242252 α2 − 0.0048241588 α3 − 0.0023958362 α4 −
0.0013119155 α5 − 0.00081617226 α6 − 0.00053068034 α7 − . . . , (54)
in which, for simplicity, we have reported only few significant digits of the coefficients. The
latter are computed symbollically and have the form of a logarithm of rational number. At
the transition we find SGS(αc) ≃ 0.38 which is indeed very high as compared to SGS(0) = ln 2.
A plot of the entropy versus α is shown Fig. 3. For completness, we stress that the ground
state entropy and the logarithm of the number of solutions, which coincide below αc, have
different meanings (and values) above the treshold. In this region, the latter equals to −∞
since all solutions have disappeared while the former quantity reflects the degeneracy of the
lowest state (with strictly positive energy) and is continuous at the transition as shown by
simulations.
Since, for α > αc, there do not exist anymore sets of Si’s such that the energy (2) remains
non extensive, the vanishing of the exponentially large number of solutions that were present
below the threshold is surprisingly abrupt. We then conclude that the transition itself is due
to the appearance, with probability one, of contradictory logical loops in all the solutions
and not to a progressive disappearance of the number of these solutions down to zero.
This perfectly agrees with the graph–theoretical derivation of the critical α which is indeed
based on a probabilistic calculation of appearance of contradictory cycles in oriented random
graphs representing Boolean formulas.
VII. THE K ≥ 3 CASE
The K = 3 case is the first NP–complete instance of K–SAT. The resolution of the RS
equations leads to a scenario different from the previous K = 2 case. We shall see below
19
that RS theory does not allow to derive the value of the threshold αc(3) ≃ 4.17 ± 0.05,
which was estimated by means of finite–size scaling techniques [1]. This is due to the fact
that the calculation of αc(3) requires the introduction of Replica Symmetry Breaking (RSB),
leading to very complicated equations we have not yet succeeded in solving. However, it is
a remarkable fact that, in the relevant region for 3–SAT, i.e. for α ranging from zero up to
αc(3), the ground state entropy computed using RS theory seems to be exact.
Let us start with the p = 1 RS solution (39). Solving eq. (40) leads to the following
scenario (see Fig. 5). For α < αm(3) ≃ 4.667, there exists the solution γ1 = 0 only. At
αm(3), a non zero solution γ1(α) 6= 0 discontinuously appears. The corresponding ground
state energy is negative in the range αm(3) ≤ α < αs(3) = 5.181, meaning that the new
solution is metastable and that EGS = 0 up to αs(3). For α > αs(3) the γ1(α) 6= 0 solution
becomes thermodynamically stable. ¿From the above scheme one is tempted to conclude
that αs(3) corresponds to the desired threshold αc(3). However, this prediction is wrong
since the experimental value αc(3) = 4.17 ± 0.05 is lower than both αm(3) and αs(3). The
failure of the above p = 1 RS prediction is also confirmed by the large K limit. One finds
αm(K) ∼ K2K/16/π and αs(K) ∼ K2K/4/π which are larger than the exact asymptotic
value αc(K) ∼ 2K ln 2. It is worth noticing that though the scaling of αc(K) for large K
is wrong within the p = 1 RS Ansatz, the asymptotic value for large α (and any K) of the
ground state energy for MAX–K–SAT is correctly predicted : EGS(α) ∼ α/2K [15].
We now turn to improved RS solutions by looking at larger values of p. When p = 2, the
previous transition scenario remains qualitatively unaltered, but the precise values of the
spinodal and the threshold points are quantitatively modified. One finds, see Fig. 5. that
αm(3|p = 2) ≃ 4.45 while αs(3|p = 2) ≃ 4.82. The ground state energy curve is similar to
the p = 1 curve but is shifted to the right. Though still incorrect, the p = 2 prediction is
thus closer to the real threshold value. For larger integers p, we have found that αm(3|p) and
αs(3|p) still decrease but quickly converge to the values 4.428 and 4.605 respectively (the
maximum value of p we have tried is p = 20, but in practice convergence was already reached
for p = 10). In Fig. 6, we have plotted the values of the coefficients rℓ (ℓ = 0, . . . , p − 1)
20
entering (43) for p = 1, p = 5 and p = 10. The departure of the coefficient curves for
p = 5 from the p = 10 curves displaying r0, r2, r4, r6 and r8 is clearly visible as soon as the
remaining coefficients of the p = 10 solution, namely r1, r3, r5, r7 and r9 which are implicitly
set to zero in the p = 5 solution, acquire a non negligible value.
Therefore, we may conclude from the above analysis that RS theory is unable to correctly
predict the value of the transition threshold. When crossing the latter, a first order Replica
Symmetry Breaking transition presumably takes place. Note that the relevance of RSB for
the random SAT problem has already been suggested from a simple evaluation of the first
few moments of the partition function [24]. The calculation of the threshold value would
require the introduction of a Replica Symmetry Broken Ansatz to replace (9). However,
the issue of RSB in diluted models is largely an open one [8], due to the complex structure
of the saddle–point equations involved, and we shall not attempt here at pursuing in this
direction.
In the following, we shall rather show that RS theory still provide a consistent and very
precise analysis of the behaviour of the random K–SAT problem below its threshold. This
requires the inspection of the ground state entropy in the region where R(z) = δ(z). Using
the method exposed in the previous Section, we have computed SGS to the 8th order in α
and found that
SGS(α) = ln 2− 0.13353139 α− 0.00093730474 α2 − 0.00011458425 α3 −
0.000016252451 α4 − 2.4481877 10−6 α5 − 3.9910735 10−7 α6 −
6.5447303 10−8 α7 − 1.167915 10−8 α8 − . . . , (55)
in which, again, we have reported only few sufficient digits of the (exactly known) coefficients.
The entropy curve is displayed Fig. 7 in the range 0 ≤ α ≤ αc(3). By computing the zero
entropy points (αze) given by the ℓ − th order entropy expansion, one finds a convergent
succession of values toward αze(3) = 4.75 (within one percent of precision), definitely outside
the range of validity 0 ≤ α ≤ αs(3|p → ∞) ≃ 4.605 of the expansion (55). Notice that
αze|ℓ=1(3) = 5.1909 corresponds to the annealed theory. A similar calculation for the cases
21
K = 4, 5, 6 yields qualitatively similar results which show an even quicker convergence
towards a zero entropy point such that αc(K) < αze(K) (see next Section for the analysis
of the large K limit where both values coincide).
Therefore, SGS is always positive below αs(3|p → ∞). In contradistinction with the
p = 1 RS solution [2], the large p RS solution cannot be ruled out by a simple inspection
of their corresponding entropy. A more important consequence of the previous calculation
of the entropy is that, at the threshold αc, the RS entropy is still extensive. The crucial
point is now to understand whether such value of the entropy is exact up to αc or whether
Replica Symmetry Breaking (RSB) effects have come into play. This issue may be clarified
by resorting to exhaustive numerical simulation. As reported in [2], simulations in the
range N = 12, ..., 28 lead to the conclusion that not only the entropy is indeed finite at the
transition but also that our analytical solution appears exact up to αc. In particular the
1/N extrapolation of the entropy value at α = 4.17 shows a remarkable agreement between
the numerical trend and the RS prediction SGS(αc) ≃ 0.1 (see the inset of the figure in
[2]). RSB corrections to the RS theory seem thus to be absent below αc, which leads us to
conjecture that the RSB transition could occur at αc exactly. In this sense the situation
would be partially similar to the binary network case [25] : the RS entropy would be exact
up to αc (but does not vanish) that would also coincide with the symmetry breaking point.
VIII. THE ASYMPTOTIC CASE OF LARGE K
In the large K limit, the saddle point equations lead to a closed form for the probability
distribution P (x). In fact, in terms of the quantity
Q(A) =∫ 1
−1
K−1∏
ℓ=1
dxℓP (xℓ)δ(
A− A(K−1)
)
, (56)
the differential equation (27) reads
∂P (x)
∂α= −KP (x) +K
∫ ∞
−∞dA
(
Q(A) + α∂Q(A)
∂α
)
×
1
2
[
∂η(x)
∂xP (η(x)) +
∂η(−x)
∂xP (η(−x))
]
, (57)
22
where η(x) has been defined in Section III. For K ≫ 1 , we may expand Q(A) as
Q(A) ≃ δ(A− 1) +1
2K−1δ′(A− 1) +
1
2
(
1
4+
1
4
∫ 1
−1dxP (x)x2
)K−1
δ′′(A− 1) + . . . (58)
Under the changes of variables G(y, α) = (1− tanh2 y)P (tanh y) and
V (α) = αK(
1
4+
1
4
∫ 1
−1dxP (x)x2
)K−1
, (59)
equations (57) and (58) simplify into the celebrated heat equation
∂G(y, V )
∂V= 2
∂2G(y, V )
∂y2. (60)
whose normalized solution is G(y, V ) = exp(−y2/2V )/√2πV . Turning back to P (x), we
find
P (x) ≃ 1√
2πV (α)(1− x2)exp
(
− 1
8V (α)ln2
(
1 + x
1− x
)
)
, (K ≫ 1) , (61)
where V (α) is given by the self–consistency equation (59). The latter may be easily estimated
for large K : V (α) ≃ αK/4K−1. Therefore, when α < αc(K) ≃ 2K ln 2, V (α) is vanishingly
small, that is P (x) → δ(x), proving that the replicas become uncoupled in the large K
limit [1]. In addition, it can be checked that the zero entropy point αze(K) reaches the
threshold αc(K) from above. Another way of looking at the entropy is provided by equation
(53) : it is a simple check the fact that αc(K)2 ∂2SGS
∂α2 |α=0 → 0 for large K. We may then
conclude that the annealed approximation becomes exact when K ≫ 1. As said above, K
may be understood as the connectivity of our model and, in the asymptotic regime K ≫ 1,
RS theory includes only Gaussian interactions as in long–range spin–glasses models [25]. In
Fig. 8 we report some instances of the probability distribution, calculated for different values
of K and α. Notice that since the critical point coincides, in this large K limit, with the
zero entropy point (which is far below the point where the RS energy becomes positive -
see previous Section), the probability distribution of the Boolean magnetization is far from
being concentrated in ±1.
23
IX. CONCLUSION AND PERSPECTIVES
In this paper, we have presented the Replica Symmetric theory of the random K–SAT
problem. We have shown that, in this context, the natural quantity emerging from the
analytical study is the distribution of the average values of the Boolean variables, indicating
to what extent the latters are determined by the constraints imposed by the clauses. The
knowledge of this probability distribution requires the resolution of a functional saddle–point
equation, for which we have presented an iterative sequence of improved solutions. The most
surprising result we have derived is the fact that the entropy is finite just below the transition,
i.e. that the latter is characterized by an abrupt disappearance of all exponentially numerous
solutions due to the emergence of contradictory loops.
Some numerical simulations we have performed for K = 2 as well as in the K = 3 case
are in remarkable quantitative agreement with our RS calculations of the entropy jump at
the threshold [2]. It seems therefore reasonable to infer that RS is correct up to the critical
ratio of clauses per Boolean variable. Would it be so the physical picture of the space of
solutions would not necessarily be simple. Replica Symmetry can indeed hide a non trivial
structure of the solutions, as has been shown for long range spin-glasses [26] models and
in the (closer to K–SAT) case of neural networks [27]. This issue is probably of crucial
importance to understand the performances of local search algorithms.
As for the values of the critical thresholds themselves, RS gives the correct prediction
αc = 1 for K = 2 but fails in estimating the critical αc for K ≥ 3. The study of the
NP–complete instances K ≥ 3 of the K–SAT problem requires to break replica symmetry.
As a consequence, their direct study will surely reveal far from being easy and will require
non trivial analytical efforts.
Another route which one can follow to reach a better understanding of the K > 2 case
consists in starting from the relatively well understood 2–SAT case and modifying it to
get closer to the 3–SAT problem. Such a pertubative approach can be implemented by
considering a mixed model, which one may refer to as (2 + ǫ)–SAT model, composed of
24
(1− ǫ)M clauses of length three and ǫM clauses of length two. As a consequence, we have
at our disposal a model interpolating smoothly between 2–SAT (ǫ = 0) and 3–SAT (ǫ = 1),
which one can attempt at solving for small ǫ at least. Preliminary investigations suggest
that, both from analytical and numerical points of view, the above model could be of interest
for exploring the connection between the order of the RS to RSB phase transition and the
computational cost transition observed numerically near criticality. Work is in progress
along these lines [22].
Acknowledgments : We thank S. Kirkpatrick and B. Selman for useful discussions.
25
REFERENCES
∗ Email: [email protected] ; preprint LPTENS 96/35
LPTENS is a unite propre du CNRS, associee a l’Ecole Normale Superieure et a
l’Universite de Paris–Sud.
† Email: [email protected]; preprint POLFIS-TH 96/15
also at INFN sez. Torino.
[1] S. Kirkpatrick and B. Selman, Science 264, 1297 (1994)
[2] R. Monasson and R. Zecchina, Phys. Rev. Lett. 76, 3881 (1996)
[3] M. Mezard, G. Parisi, M.A. Virasoro Spin Glass Theory and Beyond, World Scientific,
Singapore, 1987
[4] M. Garey and D.S. Johnson, Computers and Intractability; A guide to the theory of
NP–completness, W.H. Freeman and Co., San Francisco, 1979
[5] S. Kirkpatrick, C.D. Gelatt Jr. and M.P. Vecchi, Science 220, 339 (1983)
[6] B. Selman and S. Kirkpatrick, Artificial Intelligence 81, 273 (1996)
[7] Issue 1–2, Artificial Intelligence 81, T. Hogg, B.A. Huberman and C. Williams Eds.,
(1996)
[8] C. De Dominicis and P. Mottishaw, J. Phys. A: Math. Gen.20, L1267 (1987)
K.Y.M. Wong and D. Sherrington, J. Phys. A: Math. Gen.21, L459 (1988)
[9] S. Kirkpatrick, G. Gyorgyi, N. Tishby and L. Troyansky, in Advances in Neural Infor-
mation Processing Systems 6, J. Cowan, G. Tesauro and J. Alspector Eds., (1994)
[10] S.A. Cook, The complexity of theorem–proving procedures, in Proc. 3rd Ann. ACM
Symp. on Theory of Computing, Assoc. Comput. MAch., New York, 151 (1971)
[11] A. Kamath, R. Motwani, K. Palem and P. Spirakis, Random Structures and Algorithms
26
7, 59 (1995)
[12] B. Aspvall, M.F. Plass and R.E. Tarjan, Inf. Process. Lett.8, 121 (1979)
[13] A. Goerdt, in Proc. 7th Int. Symp. on Mathematical Foundations of Computer Science,
264 (1992)
V. Chvatal and B. Reed, in Proc. 33rd IEEE Symp. on Foundations of Computer Sci-
ence, 620 (1992)
[14] M. Yannakakis, in Proc. 4th Annual ACM–SIAM Symp. on Discrete Algorithms, 1
(1993).
[15] A.Z. Broder, A.M. Frieze and E. Upfal, in Proc. 4th Annual ACM–SIAM Symp. on
Discrete Algorithms, 322 (1993)
[16] G. Baskaran, Y. Fu, P.W. Anderson, J. Stat. Phys. 45, 1 (1986)
[17] M. Mezard, G. Parisi, J. Physique Lett. 46, L771 (1985); J. Physique 47, 1285 (1986)
[18] H. Orland, J. Physique 46, L763 (1985)
[19] O. Dubois, Theoretical Computer Science 81, 49 (1991)
O. Dubois and J. Carlier, Theoretical Computer Science 81, 65 (1991)
[20] J.A. Hertz, A. Krogh, R.G. Palmer, Introduction to the theory of neural computation,
Addison–Wesley, Redwood City (CA), 1991
[21] E. Gardner, J. Phys. A 21, 257 (1988)
E. Gardner and B. Derrida, J. Phys. A 21, 271 (1988)
[22] S. Kirkpatrick, R. Monasson, B. Selmann and R. Zecchina, in preparation
[23] It would be worth studying the local stability of the RS saddle–point. This is however
a difficult task, that can not exclude the occurrence of first order RSB. See :
M. Mezard and G. Parisi, J. Physique 48, 1451 (1987)
27
[24] L. Troyansky, N. Tishby, proceedings of PhysComp96, T. Toffoli, M. Biafore, J. Leao
eds., Boston (1996)
[25] W. Krauth and M. Mezard, J. Physique 50 , 3057 (1989)
[26] G. Parisi and M. Virasoro, J. Phys. (Paris) 50, 3317 (1986); T. R. Kirkpatrick and D.
Thirumalai, Phys. Rev. B 36, 5388 (1987);
[27] R. Monasson and R. Zecchina, Phys. Rev. Lett. 75, 2432 (1995); (Erratum) 76, 2205
(1996)
28
FIGURES
FIG. 1. Ground state cost–energy (bold line), or fraction of violated clauses, and entropy (thin
line) versus α for K = 1.
FIG. 2. Order parameters ri (i = 0, ..., p− 1) corresponding to the different RS solutions p = 1
(dashed line), p = 5 (dashed–dotted lines) and p = 10 (continuous lines), for K = 2 and α ∈ [1, 3].
The curves representing ri for i = 1..., p − 1 overlap and therefore are hardly distinguishable.
FIG. 3. RS ground state entropy (left curve, left scale) and RS ground state energy (right scale)
for p = 1, 10 versus α for K = 2. The dashed lines interpolate the numerical data of exhaustive
simulations on systems of size N = 16, 20, 24 and averaged over 15000, 7500, 2500 samples respec-
tively. Errors bars are within 10% for the entropy and even smaller for the energy and thus not
reported explicitly.
FIG. 4. 1/N extrapolation of the minimal fraction of violated clauses (i.e. ground state
cost–energy) for α = 3 and N = 18, 20, 22, 24, 26 averaged over 20000, 15000, 10000, 7500 and
5000 samples respectively. The extrapolated value appears to be different from the value 0.14472
toward which the RS solutions with increasing p rapidly converge.
FIG. 5. RS ground state energy for K = 3 (continuous lines) computed for p = 1, ..., 10 and
compared with the results of numerical simulations on systems of size N = 16, 20, 24 and averaged
over 15000, 7500, 2500 samples respectively (error bars are of the order of the size of the dots).
The RS ground state energy becomes positive (for p >> 1) at αs ≃ 4.605 whereas the value at
which the unstable solution appears is αm ≃ 4.428. Scope of the dashed line is to help the eye in
following the expected, yet unkown, RSB behaviour.
FIG. 6. Order parameters ri (i = 0, ..., p− 1) corresponding to the different RS solutions p = 1
(dashed line), p = 5 (dashed–dotted lines) and p = 10 (continuous lines), for K = 3.
29
FIG. 7. RS entropy for K = 3 (continuous line) compared with the results of exhaustive
numerical simulations for N = 16, 20, 24 and averaged over 15000, 7500, 2500 samples respectively
(see also ref.[4]). Errors bars are within 10% and not reported explicitly.
FIG. 8. Probability distributions P (x) as functions of the magnetization x, calculated for
α = 2K ln 2 (critical threshold in the K >> 1 limit) and for K = 10, 12, 14, 16, 18.
30
0.25 0.5 0.75 1 1.25 1.5 1.75 2
M/N
0
0.1
0.2
0.3
0.4
0.5
0.6
1.25 1.5 1.75 2 2.25 2.5 2.75 3
M/N
0
0.2
0.4
0.6
0.8
1
p=10p=5
p=1
0.5 1 1.5 2 2.5 3
M/N
0
0.2
0.4
0.6
0.8
1
0.04
0.08
0.12
Ln(2)
p=1,10
0.01 0.02 0.03 0.04 0.05 0.06
1/N
0.13
0.14
0.15
0.16
0.17
K=2, M/N=3, N=18,20,22,24,26
p=1
p>>1 RS-values
4.5 5 5.5 6 6.5 7 7.5 8
M/N
0
0.025
0.05
0.075
0.1
0.125
p=1
p=10
4.6 4.8 5 5.2 5.4 5.6 5.8 6
M/N
0
0.1
0.2
0.3
0.4
0.5
p=10
p=5
p=1
1 2 3 4 5 6
M/N
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Ln(2)
-0.75 -0.5 -0.25 0 0.25 0.5 0.75 1x
0
5
10
15
20
25
K=10
K=12
K=14
K=16
K=18
M/N=Log(2) 2^K