HAL Id: jpa-00246643
https://hal.archives-ouvertes.fr/jpa-00246643
Submitted on 1 Jan 1992
HAL is a multi-disciplinary open access archive for the deposit and dissemination of sci- entific research documents, whether they are pub- lished or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés.
Directed walks with complex random weights: phase diagram and replica symmetry breaking
Yadin Goldschmidt, Thomas Blum
To cite this version:
Yadin Goldschmidt, Thomas Blum. Directed walks with complex random weights: phase diagram and replica symmetry breaking. Journal de Physique I, EDP Sciences, 1992, 2 (8), pp.1607-1619.
�10.1051/jp1:1992230�. �jpa-00246643�
Classification Physics Abstracts
05.40 72.20 71.55J
Directed walks with complex random weights: phase diagram
and replica symmetry breaking
Yadin Y. Goldschmidt and Thomas Blum
Department of Physics and Astronomy, University of Pittsburgh, Pittsburgh, PA 15260, U-S-A-
(Received
26 February1992, accepted in final form 21 April1992)
Abstract. We consider the problem of directed walks
(or polymers)
in a random potential with both real and imaginary parts. The correlations of the potential are of short range. The problem is solved using a variational approximation which is expected to become exact in the limit of a large embedding dimension. A rich phase diagram is found which contains five phases.In addition to the simple high temperature phase, there are four ordered phases with diiTer- ent degrees of
replica-symmetry-breaking.
Three of the phases are dominated by interference(quantum)
eiTects, and in two phases a pairing interaction between replicas plays an important role.1 Introduction.
The
problem
of directedpolymers
in a random medium has attracted considerable interestrecently [1-8].
In addition tomodeling
variousphysical phenomena [1-3],
it serves as a prototype of a system whichdisplays
the consequences of local frustration and is somewhatsimpler
than(although
similar in many respectsto)
thespin-glass problem [4-8]. By
"directed"polymers,
one
loosely
means '~nooverhangs"
and moreprecisely
that the walker(polymer) proceeds always
in apositive
directionalong
one chosen coordinate(called "time")
in an(N
+I)-dimensional
space.
Recently,
Mdzard and Parisi(MP)
[6] have found a solution to directedpolymers
in a random realpotential
in the limit oflarge embedding
dimension(N
-
oo);
it exhibits manyproperties
similar to those of the mean-field solution of thespin-glass problem
[9]. Inparticular,
this solution is characterizedby
the existence of many"valleys"
orthermodynamic
"states" a fact that manifests itself in a mathematical property of the solution called"replica
symmetrybreaking" (RSB).
A different
problem,
related to theabove,
is that of directedFeynman paths
which threadthrough
a disorderedmedium, acquiring
randomphases [10-19].
Thisproblem
arises in thestudy
of interference effects in MottVariable-Range Hopping
indoped
insulators and semicon- ductors[10-12].
The transmissionamplitude
for thehopping
of electrons fromone location to another can be described as the
superposition
ofpaths
whichpick
up randomphases
asthey
pass
through impurity
sites. Forstrongly
localized states, the dominant contribution to the1608 JOURNAL DE PHYSIQUE I N°8
transmission
amplitude
is from the shortest and hence directedpaths.
There are severalconjec-
tures and numerical simulations
concerning
thisproblem
in I+ I dimensions[10-15,19].
Here we consider ageneralization
which includes both thedirected-polymer
case and therandom-phase
case, viz. directed walks in a
complex
randompotential, incurring
both randomamplitudes
and random
phases
at each site. Cook and Derrida(CD)
[16] haveinvestigated
aspecial
case of thisgeneral problem by considering
walks on aCayley
tree with randomphases
restricted to 0 or ~r(random signs)
withprobabilities
I p and p,respectively.
In thatwork,
CD haveobtained
expressions
for the "freeenergy" by relating
it to anothermodel,
the Generalized Ran- domEnergy
Model(GREM),
which has been shown to coincide with the model on the tree for the case p= 0
(random amplitudes only)
[4]. CD haveconjectured
that GREM describespolymers
withcomplex weights
on the tree even for p# 0,
when the latter could not be solveddirectly.
Theresulting phase diagram
consists of threephases
ahigh-temperature phase,
alow-temperature
frozenphase,
and a newphase (as compared
to therandom-amplitude model)
in which interference effects become
important.
In this
work,
our aim is toapply
thereplica
method to thisgeneralized
model in order to obtain more detailed information about the structure of the various "states" in the differentphases
thatmight
occur. We have used a variational(Hartree-type) approximation
which isexpected
to become exact in the limit of alarge embedding
dimension [6]. To oursurprise,
thephase
structure we obtain is richer than that of CD[16],
in the sense that we find two newphases
which are characterizedby
a"pairini'
ofreplicas (See Fig. I).
Medina et al. [12] haveargued
thatpairing
ofreplicas
isexpected
in therandom-phase
model and that "doublets"of
replicas
willplay
a role similar to that ofsinglets
in therandom-amplitude
case.Thus,
in addition to theregular
scheme ofRSB,
we allow for an extrapairing
interaction whensearching
for the best variational solution. The
"simple" high-temperature phase (I)
isseparated
from two of the otherphases (II
andIII) by
first order transitionlines,
similar to what has been foundby
CD. These twophases
areseparated
from each other and from the"pairing" phases (IV
andV) by
second order transition lines. Details will begiven
below. We have obtained thewave functions and pattern of RSB in each
phase
which shedslight
on theirphysical properties.
It seems
plausible
to us that the reason that CD did not find the two newphases (IV
andV)
is because the GREM(or REM)
allowsonly
for a tree-like structure of RSB and doesnot accommodate the extra
pairing
interaction. Since it has never been shown that GREMactually
describes the random walk on a tree for the case of randomsigns (or phases),
it may describefaithfully only
part of thephases.
Forcompleteness,
we should mention that it ispossible
inprinciple
that thephase diagram
of a walk on aregular
lattice in the limit oflarge embedding
dimension may be different than that of a walk on theCayley
tree.Furthermore,
there is
always
thepossibility
that a solution to the saddlepoint equations
with a different orhigher
level of RSB(I.e.
a differentparametrization
or more variationalparameters)
than theone we have obtained exists.
Lastly,
there ispossibly
a difference between thephase diagram
of the
random-sign
model and a model with a continuous distribution of randomimaginary phases,
but this is not verylikely according
to CD(see
p. 984 of Ref.[16]).
Our method differs from that of MP in the sense that we use the Hamiltonian
approach
rather than theLagrangian (Green's function) approach
thatthey
use, but the results are identical for the case of a realpotential
whichthey
have considered.2.
Description
of the model.In the continuum
limit,
the model can be describedby
apath (x(t), t)
in an(N+ I)-dimensional
space-time,
with x an N-dimensional vector and t the "time" coordinate whichalways
increaseso
°'~
iii iv
i,o
i
is
~
2.O
2.5
O.5 1-o 1.5 2.O 2.5
I
Fig, I. The phase diagram containing five phases.for a directed walk. Each
path
has aweight
whichdepends
on the local value of the randompotential (both
itsamplitude
andphase).
The sum over allpaths,
eachweighted appropriately,
is
given by
aFeynman path integral (which
onemight
call apartition function, though
it is notnecessarily real)
of the form:Z(y, T)
=~~~~~
~XXpl- f
dt~°
(~~)
~
+
~X~
+iy8(X, t)
+flV(X, )j
). (2,I)
(o o) 2 fit 2
The first term in the
weight,
which resemblesa kinetic energy, has a coefficient no which we call the tension. It is a soft version of the constraint found in lattice versions which allows a
walker to move from a
given
siteonly
to its immediateneighbors and,
in the case ofpolymers,
is related to their
flexibility
orrigidity. Depending
on themodel,
onemight
choose no to be1610 JOURNAL DE PHYSIQUE I N°8
temperature
dependent,
but here we do not make any suchspecial
choice. The mass term(+-
p)
serves as aregulator
and willeventually
be taken to zero. Theterm17e(x,t)
representsthe random
phase,
whileflV(x, t)
denotes the randomamplitude
at each site. We consider thecase in which the variables fl and V have normal
(Gaussian)
distributions with zero means and thefollowing
correlations:v(x,i)v(x>,i>)
=
-J(t-t>)Ni~[(xjx')~j (2.2)
e(x, t)e(x', t')
=
-6(t t')Nf~ ~~
j~'~~ j, (2.3)
where
g(z)
indicatesaveraging g(z)
with respect to thequenched
random variables. The N-dependence
has been chosen(following MP)
such that thelarge
N limit of thetheory
is well defined. Forlarge
distances we assume that bothf~
andf~
vanish likef(z~)
«
-((z~)~+; (2.4)
whereas,
for short distances thef's
areregular
at theorigin (see below).
This form follows from dimensional considerations in which one seeks a "smoother"representation
of the correlation than an N-dimensional delta function. [6]We use the
replica
method [9] to carry out thequenched
averages over the randompotent1als.
Upon replicating
andaveraging,
one obtains:(Z*Z)
~((ya),
T) =~~~~~~~~ ~xi ~x2n
xpl- f
dtit ~j (~~
~+ ~
~j xi
(joj,o) 2
~
2
~
(
2£ eoepN f~ (j(xa p)~)
~,p
+
£
Niv (j(xa xp)~)j ), (2.5)
~ p
WheTe
+I, if a =
I,
, n;
(2.6)
~« "
-l, if a = n
+1,.. ,2n.
Note that one
replicates (Z*Z)
in thecomplex
case; this choice is apparent in the context of electronhopping conductivity
where the quantum mechanicalprobability
is the square of the absolute value of the transmissionamplitude. Consequently,
there are 2nreplicas:
the first set of n from theZ*'s,
and the second set from the Z's. Theaveraging
results in interactionsamong the
replicas.
The interactionsstemming
from therandom-amplitude averaging
areall
attractive; while,
thosearising
from therandom-phase averaging
are"charge-dependent".
Their
sign depends
on whether the tworeplicas
inquestion belong
to the same or differentsets.
Finally,
one can extract thefollowing
associatedN-dimensional, (2n)-body
Hamiltonian:fi2
~2
1~
2
~ [
0zf~
2( ~°~~~~~ Nno
~~° ~~~j
+i~jxa~
+Nf~ (£(xa p)~)
,
(2.7)
2
~ ~ p Ko
where xa's have been rescaled
by n)~/~
and wherep has been rescaled
by
no.Note that G
=
(Z*Z)"
ofequation (2.5)
satisfies theSchr6dinger-like equation
[5]:)
= -HG.
(2.8)
In the limit of
large
T,G "
'vt(1011'v0(iYal)~~~°~, (2.9)
where
Eo
is theground-state
energy of H and ivo is thecorresponding
wave function. Thus~ljn~ -(In f dyi dynG(yi,
, Yn Yi,
, Yn T) "
(2.1°)
is the 'free
energy" density.
3. The variational
approximation.
In their
study
of manifolds in a random(real) potential,
MP have useda variational
approach
to surpass standard
perturbation theory. They
haveemployed
ageneral quadratic
action with tunable parameters to serve as a variationalapproximation
to the exact action and haveargued
that the
result,
which isequivalent
in this case to aHartree-type approximation,
becomes exact forlarge
N [6]. Theirproof
involvesrepresenting
the action in terms ofauxiliary
fields andusing
a steepest descentapproximation.
Such anapproach
isequivalent
to the Hartree methodand becomes exact when
N,
whichmultiplies
theaction,
becomesinfinitely large.
We have been able togeneralize
theirproof
to thecomplex potential
case.For the one-dimensional manifolds of interest
here,
one can recover the results of MP(the
case 7 =
0) by choosing
a Gaussian wave function as a variational candidate for theground
state of the Hamiltonian
(Eq. (2.7)) [20].
We thus start with a Gaussian variational wave function of the form:'V = fit exP
I-j L£'~~O,p
z;o
z;p), (3.i)
a,fl "
where1i1 is a 2n x 2n matrix. It is the exact
ground
state of aquadratic
Hamiltonian of the form:~ "
~)[[(
+
'[~°~
+)[~"P~"
~P' ~~'~~where the matrices 1i1 and
fir
are related
by:
1i1 =
(pi
+fir)~/~, (3.3)
where
I
is theidentity
matrix. Note that even for the best Gaussian wave function theground-
state energy of this Hamiltonian differs from that of the exact Hamiltonian. In order to have
some
physical
intuition into themeaning
of the variationalparameters
asrepresenting
effective"interactions" among
replicas
and also in order to make a connection withMP,
we will start with theparametrization
offir (and
ih will then begiven by (3.3)), although
in ourapproach
it is
technically
unnecessary to introduce h(3.2)
at all.For
practical
reasons, it isimpossible
to choose the mostgeneral
matrixAi (since
onerequires
anexpression
forgeneral
n to enable continuation to n =0);
rather it is necessary1612 JOURNAL DE PHYSIQUE I N°8.
to select an
appropriate parametrization
ofk.
For theshort-range
case, it is reasonable tobelieve, following
MP for a realpotential,
that afinite-step
RSB scheme isappropriate.
Sucha choice is further
supported by
the finite-n GREM consideredby
CD. Ourparametrization
scheme first breaks the 2n x 2n matrixli
into twon x n matrices as follows:
fir
=
(~ (~
,
(3.4)
2
where
ii
and 1j2are shown in
figure
2. The matrixii
represents the interactions among the set of nequally charged "particles"
and 1j2 the interactions amongoppositely charged
particles (recau Eq. (2.6)).
This choice ofli
breaks the 2nreplicas
into ~ groups of size2k,
k
each
containing equal
numbers ofoppositely charged particles.
Agiven particle
has non-zero interactiononly
with the(2k-1) particles
within its group. Thestrength
of interaction betweenequally charged replicas
is ae, and that betweenoppositely charged replicas
is ao.Furthermore,
there can be an additional interaction between a
given replica
and oneparticular replica
of anopposite charge;
thestrength
of this extra interaction is adFinally,
the center-of-mass of each group is free.Z -ae
ae~
-ae Z -ae
0 0
-ae ~ae Z
~ ~°e ~ae
-ae Z -ae
0 0
-ae ~ae ~
~ -ae ~ae
-ae ~ -ae
0 0
-ae ~ae ~
Fig.
2. The matrixii
isan n x n
(hierarchical)
matrix with one-stepbreaking.
It consists ofn/k
matrices
along
the diagonal. Thediagonal
elements of these k x k matrices are ~=
(k I)(«o
+«e)
+ ad, this choice ensures translational invariance. The matrix12
is obtained byletting
ae - ao,which parametrizes interactions between oppositely-charged particles, and ~ - -«o ad where ad parametrizes the
(extra)
pairing interaction.Along
with the aforementioned basis foradopting
afinite-step
BSBscheme,
this choice has been further motivatedby
thefollowing
scenario. Thecharged
interaction inducesoppositely charged replicas
to formpairs (hence ad)(
while the overall attractive interaction elicits aclustering
ofpairs (hence
ma andae).
Thisclustering
amounts to the fact thatreplica
symmetryis broken in the n - 0 limit
(which
can bethought
of asresulting
from the existence of many"states"
energetically
close to each other and unrelatedby symmetry) [6,9].
Some additional variational parameters have been considered but
proved inconsequential.
For
example,
we haveperformed
calculations which allow for non-zero interactions betweenreplicas belonging
to distinct groups with differentstrengths
forequally-
andoppositely-charged pairs,
I-e-, we have put in two parameters inplace
of all the zeros iniii
and1j2, (with only
an overall free
center-of-mass). However,
these additional parameters become zero under vari- ation. The samephenomenon
occurs in the realpotential
case consideredby
MP [6]. In aseparate
calculation,
we have allowed for thepossibility
of groupscontaining unequal
numbers ofoppositely charged particles,
but the variationalprocedure
leads to"charge
neutral" groupswhenever 7
#
0.The matrix
li
can be written as follows:
fir
=
[k(«o
+ me) +2«d]
12 lPij
©ik -(ae-«o) i2lPij©fk
-«o
J219
ii ©
fj,
gnifies the directCalculating he
of he Eq.
(2.7) with
the ave iV (Eq.
.I))
equires
owing
the
elements of
fli
and 1i1~~ heinteraction
erms,
be
omputedin
theollowing manner:l~j(xa -xp)~j) _
~ f~i) j~~ ~ ~~>jnoN
~
_ jj/Ni"
~~ >
1614 JOURNAL DE PHYSIQUE I N°8
The matrix fli and its inverse have the same form as
k;
in terms of itseigenvalues,
fli can be written:fli=
Ii I24lI~©Ik
+
)(12-li) i2lPij4lfk
+
j(13 Ii) f2
4lij
©ik
+
(14 13
+Ii 12) )2
4li~
©fki (3.9)
and the inverse fli~~ is obtained
simply by replacing
the l's withl~~'s.
Note that theregulator
p is
required
toperform
this inversion.Calculating (H)w
leads to thefollowing expression:
)$ (~~~ ~4k~
~~ ~/k~~
~~4k~
~~ '~~~~~~ ~
~~~~~~
~'~~~~ ~
k
~~
~ j~o~ ~~~~ ~
k~~
~~~
~~
~~~~~ ~
~o
~~o
~ ~~~~~~~
~
~o
~~o
~ ~~ ~~~~~~
~~2k~~
~o
~~o
~~o
~ ~~ ~~~~~"
~~2k~~
~o
~~o
~~o I'
~~'~~~We will focus on
short-range correlations; therefore,
any interactionresultin~
in a term of the form/[£a;1/~
+ p~ii
has been setequal
to zero in the p - 0 limit(f[oo]
=0).
This step is where the difference between short andlong
range sets in. Infact, henceforth,
we willconcentrate on a
particular short-ranged
correlation:/v[z]
=
([xl
=
(f°
~fl~'
~~°~~~)~~
(3,ll)
0, otherwise,
where
lo
isnegative. Higher
powers of z, if present, will notchange
any of thephysics
and may have a very minor effect on the location of thephase
boundaries. Since the l'sare
independent
functions of the
a's,
one can find thestationarity
conditionsby extremizing
with respect to the l's and k(instead
of a's andk).
The variable koriginally
lies between I and n: I < k < n, which in the n- 0 limit becomes 0 < k < I.
When
looking
for the best variational parameters, one must restrict the variations to remain within thephysical
range. In thiswork,
there are twoimportant
constraints on the allowed values of the l's andk,
besides the usualrequirement
that1;
> 0 which is necessary for thenormalizability
of the wave function. The first constraint is:(k
kll~_i
1 ~1~_~
k ~ ~ ~'
(3.12)
Table I. Listed here are
expressions for l's, a's,
F and k that resultfrom e~tremizing
F.Phase l's a's F and k
li"o «o=o
I
l~=0 «e=0 F=7~-fl~
13 " o ad
" o
Ii
=2kl fl
ma = 2fl~ +272
F =4klfl 6kfl2
II 12 =
2ki(fl2
+7~)I
me=
2fl2 27~
l~
=2kl fl
ad= 0
-4ki
fl2 +(2k
+2)fl
=(fl2
+72)1
Ii
= I ma =k/2
F =(1+ y2 3fl2)/2
III
12
= 2 ae =-k/4
13=1
ad=0k=(3fl2-7~+1)/4fl~
Ii
=(I k)/2k~
ao=
1/4k4
F =(1 8k3fl2)/4k2
IV 12 =
1/2k2
ae=
(k 1)/4k4
13
=1/2k
ad =(1- 2k)/8k4 4k2(72
+fl2 fl~/k)I
= I
Ii
= ma =2fl2
+272
F= +
(2 2(2
V 12 =
27/(2 k)
I me=
2fl2 272
>3 =
2k~fl
ad=
~~~ ~~~~~
2kfl~
~ + ~~ll~~ 4fl~
= 0Thh condition follows from the
requirement
that the arguments of thel's
must bepositive,
since these functions are
only
defined forpositive
arguments.Taking
into account that md > 0(I.e.,
thepairing
isalways attractive),
one establishes thatIi
> 13 in allphases,
and thusequation (3.12)
is sufficient to ensure that allarguments
ofl's
are
positive.
The second constraint we have found necessary toimplement
isIi l~
+13
2 0.(3.13)
This constraint ensures that the kinetic energy, and hence the total energy, remains bounded from above as k
-
0,
andsubsequently
the extremum isalways
aglobal
maximum as a function of k. We have found that thisrequirement
is necessary in order to finda consistent solution
throughout
the entirefl
7plane.
When these conditions are nolonger satisfied,
we enforce the constraints asequalities (either
or both asnecessary)
and search for the extremum of the1616 JOURNAL DE PHYSIQUE I N°8
free energy
subject
to the constraints. We also check eposteriori
that the arguments of/
are
always
less than~~~,
asrequired by equation (3.ll).
Ithappens
that the extremalpoint
isii
actually
a saddlepoint,
since the best wave function results from a delicate balance between the attractive andrepulsive
interactions. In allphases,
the energy is maximal with respect to the variation in k. It is also a maximum with respect to Ii and13 (when they
are not related tol~ through
aconstraint)
and a minimum with respect tol~.
Table II. The
e~pressions for
thephase
bovndelies.Phase Boundaries
I II
(6fl~
+ 467~)1
+54fl(fl~
+72)I
+ 8187~ 90fl~
= 01- III
y~+p~
= I
II III y
=
lip
II V
4fl~
1~fi 4fl
+fl£fi
= 0
(P?+72) (P?+7?)
III Iv y~
p~
= i
Iv v
(p2
+72)
'=
8vfp4
4. The
phase diagram.
Before
proceeding
to theresults,
let us rescale variables as follows:>;
-lfol-~fini~>;,
7-
lfol-~f/Ki17,
andp
-
j/ol~~f)Kilfl.
Furthermore,
let~
~i~~ ~~~
~~'~~Extremizing
F(Eq. (3,10)) subject
to the constraints(Eqs. (3.12)
and(3.13)) yields
fivephases
as shown infigure
I. Theresulting
values forl's,
a's and F's areprovided
in tableI, along
with theaccompanying stationarity equation
for k.The
phase
boundaries are found in table II. Phases I,II,
and III meet at thepoint (7
"
~,fl
=
~);
whilephases
II, III,IV,
and V meet at thepoint (y
=~, fl
=
~).
The2 2 2 2
transitions between
phase
I andphases
II and III are of firstorder,
with theexception
of thepoint
at 7 = 0. All otherphase
boundaries are continuous transitions. One can understandsome of the
properties
of thesephases
in terms of three order parameters: ma(the "glass"
order
parameter),
ao+2ae +2ad(the
"thermal" orderparameter)
and ad(the "pairing"
orderparameter).
The"glass"
order parameter is zero inphase
I(the high-temperature,
weak- interactionphase)
and non-zero in all otherphases.
The "thermal" order parameter is zero inphases I,
III andIV,
which are allhigh-temperature phases.
Notice that in these threephases,
the
kinetic-energy
constraint(Eq. (3.13))
holds as anequality.
The"pairing"
order parameter is zero inphases I,
II andIII,
in which thecharged
interaction is weak. This time the otherconstraint
(Eq. (3.12))
holds as anequality
inphases
IV andV,
where the'§~airini'
order parameter is non-zero. PhasesIII,
IV and V do not exist for 7= 0
(no
randomphases),
and thus we can say that interference effects are strong in thesephases,
and wemight
call them"quantum" phases.
For y
=
0,
we have checked that our results lead to the samestationarity
conditions as derivedby
MP(note
thatthey
have made the choice no=
fl)
[6].Actually,
because we have two sets ofreplicas,
we finddegenerate
solutions at 7 =0;
infact,
we find an even greaterdegeneracy
when ourparametrization
does not assume"charge neutrality." However,
as soonas 7
# 0, only
thecharge-neutral
solution survives. On thefl
= 0axis,
there issimilarly
ahigh degeneracy
of the solutions with respect to k. It follows from the absence of attraction between boundpairs
whenfl
= 0. We believe that this behavior is a feature of the variationalapproximation
used here and is true in the N- oo limit and that this
degeneracy
wouldprobably
be lifted in finite dimensions. Thisdegeneracy
isagain
lifted as one turns onfl,
I-e-, the real part of the randompotent1al.
In that case boundpairs
are attracted to each other and tend to form groups of 2kreplicas.
It is
tempting
toidentify
ourphases I,
II and III with thecorresponding phases
foundby
CD
[16].
To make thecomparison,
we first subtracted the free energy ofphase
I from that ofphases
II and III both in ourexpressions
and in those of CD.Expanding
toleading
order in p, we have found that the energy ofphase
III becomes identical to that foundby
CD if weidentify
p=
7~/4 (where
p is theprobability
for anegative sign)
and also takeIn(I()
= I for
their tree constant
(I.e.,
a tree withbranching
ratio ofe).
Inphase II,
our free energy is very close to that of CD forfl
not toolarge (we
have checked thatby comparing
contourplots
of thecorresponding
freeenergies).
We attribute the difference to the fact that their model is ona lattice and ours in the
continuum,
and one may need to tune the tension no to make the two modelstruly
coincide. Note further that had weplotted
thephase diagram
in the variables Tversus
7~,
the I-IIphase boundary
wouldapproach
the y = 0 axistangentially
as in CD. Seefigure
3 for a detailedcomparison
of the free energy between the results of CD and those of this paper.We should
emphasize
that it isimpossible
for ourphase
III to extend into thehigher
7region,
since we have found that first the constraint(3.12)
is violated and further into thelarge
7
region (when 7~
> 1+3fl~),
k becomesnegative
which is alsounacceptable.
It turns out that on the transition line betweenphases
III andIV,
k=
),
and this is where the constraint(3.12)
has to beimplemented.
5. Discussion.
In this paper, we have calculated the
phase diagram
for directed walks in a randompotential having
both real andimaginary
parts. We have used a variationalapproach
that should become exact in the limit oflarge embedding
dimension. Theadvantage
of the variational method is that one canidentify
the different parameters in the trial wave function as effective interactionsamong
replicas (or "particles"), providing
an intuitivephysical picture
for theproperties
of the differentphases,
likegrouping, pairing, interference,
etc. Since we have not proven that there1618 JOURNAL DE PHYSIQUE I N°8
o.o o-o
I-O I-O
fi
f k
~ f k n o
d
2.O ~ 2.O
a a
o-O I.O 2.o o-O I-o 2 o
y 7
a) b)
Fig.
3. A comparison of contourplots of the functionf(fl,7)
"F(fl,7) FI(fl,7)
between the results of CD(a)
and the present paper 16). Fi stands for the free energy in phase I. Infigure
3a weuse the ad hoc correspondence
7~/2
=
-In(1 2p)
in order to plot CD's results in the(fl, 7)
Plane.The lines stand for contours of fixed value. These are
(a)
0.3(b)
0.2(c)
0.I(d)
0.0(e)
-o.05(f)
-o.15(g)
-0.25(h)
-0A (I) -0.6(j)
-0.8(k)
-1.2 (1) -1.6(m)
-2.0in)
-2.8(o)
-3.6. In phase I, f isidentically zero.
cannot be other solutions to the saddle
point equations,
we cannot rule out thatultimately
a more elaborate variational solution may be found. We believethough
that any such solution will retain the essentialphysical
features of the variousphases
which have been found in thiswork.
We have found a rich
phase diagram
with fivephases
asdepicted
infigure I,
with allphases
except Idisplaying
somedegree
of RSB. We have verified that the value of the exponent v[12,
19] is in allphases,
similar to what has been foundby
MP in the realpotential
case [6]. Our2
phase diagram
includes two newphases
ascompared
to theCayley-tree
calculation of CD[16].
We would like to comment about the Monte Carlo simulations done
by
CD.They
have checked for the location of the first order transitions betweenphases
I and II and betweenphases
Iand III
(Fig.
6 in Ref.[16]),
and also verified that there is a transition betweenphases
II and III(Fig.
2c in [16] shows onepoint
on theboundary).
These results are also in agreement with ours and are notenough
to rule out or establish thepossibility
of other transition lines asdisplayed
infigure
I. It is achallenge
for future research toverify
thephase diagram
offigure
I
by
additional simulations.A number of open issues
concerning
the behavior in lowerembedding
dimensionsremain;
for instance: In what dimension do some of these
phases
obtaineddisappear?
And how does the exponent vdepend
on dimension? Some of thesequestions
maypossibly
be addressedby
a
I/N expansion
[7].Acknowledgements.
We thank NSF grant DMR-9016907 for
support.
YYG thanks M. Mdzard for useful dhcussions and hishospitality
at ENS Paris. TB thanks Y.Shapir
and D-S- Koltun.References
[ii
Huse D-A- and Henley C-L-, Phys. Rev. Lett. 54(1985)
2708.[2] Kardar M. and Zhang Y-C-, Phys. Rev. Lett. 58
(1987)
2087.[3] Kardar M., NucJ. Phys. B 290
(1987)
582.[4] Derrida B. and Spohn H., J. Stat. Phys. 51
(1988)
817.[5] Parisi G., J. Phys. France 51
(1990)
1595;M6zard M., J. Phys. France 51
(1990)
1831.[6] M6zard M. and Parisi G., J. Phys. France Ii
(1991)
809; J. Phys. A 23(1990)
L1229.[ii
Cook J. and Derrida B., Europhys. Lett. lo(1988)
195; J. Phys. A 23(1990)
1523.[8] Bouchaud J-P- and Orland H., J. Stat. Phys. 61
(1990)
877.[9] M6zard M., Parisi G. and Virasoro M-A-,
Spin
Glass Theory and Beyond(World
Scientific, Singapore,1987);
Binder K. and
Young
A-P-, Rev. Mod. Phys. 58(1986)
801.[10] Nguyen Vi., Spivak B-Z- and Shklovskii B-I-, JETP Lett. 41
(1985)
42; Sov. Phys. JETP 62(1985)
lo21.[ill
Shapir Y. and Wang X-R-, Europhys. Lett. 4(1987)
10.[12] Medina E., Kardar M., Shapir Y. and Wang X-R-, Phys. Rev. Lett. 62
(1989)
941.[13] Zhang Y-C-, Phys. Rev. Lett. 62
(1989)
979.[14] Zhang Y-C-, Europhys. Lett. 9
(1989)
l13.[15] Zhang Y-C-, J. Stat. Phys. 57
(1989)
l123.[16] Cook J, and Derrida B., J. Stat. Phys. 61
(1990)
961.[17] Medina E., Kardar M. and Spohn H., Phys. Rev. Lett. 66
(1991)
2176.[18] Blum T., Shapir Y, and Koltun D-S-, J. Phys. I France1
(1991)
613.[19]Blum
T, and Goldschmidt Y-Y-, Directed Paths with Random Phases, preprint(1991),
Nucl.Phys.
B[FS]
to be published.[20] Shakhnovich E-I- and Gutin A-M-, J. Phys. A Math. Gen. 22