HAL Id: jpa-00247229
https://hal.archives-ouvertes.fr/jpa-00247229
Submitted on 1 Jan 1996
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.
Stability of the Mézard-Parisi Solution for Random Manifolds
D. Carlucci, Cirano de Dominicis, T. Temesvari
To cite this version:
D. Carlucci, Cirano de Dominicis, T. Temesvari. Stability of the Mézard-Parisi Solution for Random Manifolds. Journal de Physique I, EDP Sciences, 1996, 6 (8), pp.1031-1041. �10.1051/jp1:1996114�.
�jpa-00247229�
Stability of the Mézard-Parisi Solution for Random Manifolds
D.M. Carlucci
(~),
C. De Dominicis(~,*)
and T. Temesvari(~)
(~) Scuola Normale
Superiore
di Pisa, Piazza dei Cavalieri, Pisa 56126,Italy
(~) Service dePhysique Théorique,
CESaclay,
91191 Gifsur Yvette, France (~) Institute for Theoretical
Physics,
Eôtvôs University, 1088Budapest, Hungary
(Received
4January
1996, received in final form 16April
1996,accepted
18April 1996)
PACS.05.20.-y Statistical mechanics
PACS.05.50.+q
Latticetheory
and models of chaotic systems PACS.02.50.-r Probabilitytheory,
stochastic processes, and statisticsAbstract. Trie
eigenvalues
of trie Hessian associated with random manifolds are constructed for triegeneral
case of R steps ofreplica
symmetrybreaking.
For trie Parisi limit R~ oJ
(continuum rephca
symmetrybreaking)
which is relevant for trie manifold dimension D < 2, they are shown to be nonnegative.
Résumé. Les valeurs propres de la hessienne, associée avec une variété aléatoire, sont cons-
truites dans le cas
général
de Rétapes
de brisure de lasymétrie
desrépliques.
Dans la limitede Parisi, R
~ oJ
(brisure
continue de lasymétrie
desrépliques)
qui estpertinente
pour la dimension de la variété D < 2, on montrequ'elles
sont nonnégatives.
1. Introduction
The behaviour of
fluctuating
manifoldspinned by quenched
randomimpurities
encompasses arich
variety
ofphysical situations,
from the directedpolymers
m a randompotential (with
amanifold dimension D
=
1)
to interfacial situationsID
= d
-1,
dbeing
the total dimension of thespace).
Trie directedpolymer problem,
forexample,
is itself related to surfacegrowth,
turbulence of the
Burgers equation
andspin glass.
The literatureconcerning
vanousaspects
is wide andflorishing,
agood
deal of itbeing quoted
in some recent reviewsIl,2].
Theapproaches
taken have been
quite
diverse and it is not the purpose of this paper to describe them. This paper is a firststep
m anapproach
to random manifolds that uses field theoretictechniques
m the
replica
formalism. Field theoretictechniques
have been used with success(in
their functional renormalization groupvariety) by
Balents and Fisher [3] to describe thelarge
N(N
= d D is the transversedimension)
for the 4 e dimensional manifold. In the context ofdirected
polymers
Hwa and Fisher [4],using
more conventional fieldtheory,
haveemphasized
the role of
large
but rare fluctuations.Alternately
a unifieddescription
of thegeneral
randommanifold has been
developed by
Mézard and Parisi[si (MP), where,
via the use of thereplica technique, they
have laid down the first step(mean
field and self-consistent Hartree-Focklevel)
of a
systematic
fieldtheory 1IN expansion (~).
This is theapproach
we want to pursue here and this paper may be considered as a firstsequel
to MP.(*)
Author forcorrespondence je-mail: cirano@amoco.saclay.cea.fr)
(~)
The intricacies of trie firstloop
correction werebriefly
considered in further work [6]©
LesÉditions
dePhysique
1996In this paper we
study
theeigenvalues
of the Hessian associated with the1IN expansion
as laid down
by
MP. We do it in ageneral
formby keeping
discrete trie Rsteps
ofrephca symmetry breaking,
and we show that the MP continuous solution(R
= oo, as isappropriate
for D <
2)
does notgenerate negative eigenvalues.
The
general
discreteexpression
derived here allows to controlcarefully
thelimiting
processto the MP continuous solution. It will be also of use in further discussions of the solutions for the case D > 2
(and
inparticular
D ~4)
an unclear case to date.In Section 2 we summarize the MP
approach
down to thequadratic
fluctuation termsdefining
trie Hessian
(see
alsoI?i ).
Section 3 is devoted to trie introduction of a discrete Fourier-like transform on triereplica overlaps.
In Sectioù 4 webriefly
discuss theequation
of state at mean field level. In Section 5 we discuss and exhibit thereplicon eigenvalues.
Sections 6-8 are devoted to thetougher
case of thelongitudinal-anomalous eigenvalues.
2.
Summarizing
MPApproach
We thus consider the MP random manifold model defined
by
~7i(°')
"
~Î / d~~ If
~~)j)~
~
+
2v(~> ~'(~))
+l~'~(~)j (1)
~=i where d
= N + D is the dimension of the space, N the dimension of w, i.e. of the transverse space, and D the dimension of the vector ~, the
longitudinal
directions. Thequenched
randompotential
V has a Gaussian distribution of zero mean and correlationV(~, w)V(~', w')
=
-à(~l(~ ~')N f
~~'/~~~ (2)
~~~~
~~~~ 2(1~
+f) ~~ ~ ~~~ ~ ~~~where the function is
regularized
at small argumentsby
the cut-off Ù.Averaging
over the randomquenched
variable V viareplicas replaces il) by
-7irep(Wa)
=
/ d~~
IL L ~)[)~
+J1Wli~)
+PL NI '~'~~~~ ~'~~~~~~ i
a
~i
~a,b
(4)
In
studying (4)
MP have used first a variationalapproach (identical
to the Hartree-Fockapproximation),
which becomes exact in thelarge
N limit. Thenthey
introducedauxihary
fields with the
advantage
that itgenerates
correction toHartree-Fock, namely
(1) The zero
loop
termembodying
the effect ofquadratic fluctuation,
1.e. thelogarithm
of the determinant of the Hessian matrix(ii) 1IN
corrections in asystematic
way(~).
In this work we are interested in (1)
studying
the Hessian and itseigenvalues
around themean
field (Hartree-Fock
orvanational)
solutions. We leave thestudy
of(ii)
for aseparate publication
[8].(~)Note
that in reference [8] trie case1-RSB,
relevant for shortranged correlations,
isextensively
studied both for itsstability
and its1/N
corrections(given
m closed
form).
Following
MP we introduce trien(n +1)/2 component auxihary
fieldsTabi~)
"j~'ai~)~'b(~) là)
and its associated
Lagrange multipliers sab(~).
We need tocompute
triegenerating
function@
=
/ jj d[wa] / jj d[rab]d[sab]
x exp~ ~j / dmab(~) [Nrab(~) wa(~)wb lx)]
a=1 a<b
~
ah
~ ~~~
/
~~( ~
~ÎÎ~~
~
~
~'~~~~l
fl2
d~
£
Nf (raa(~)
+rbb(~) 2rab(~)) (6)
~
ah
now
quadratic
inw's,
hence afterintegrating
over wa@
#
/ fl djTab]djsab]
eXpjNçjT( Sjj
a<b
çjr. s)
=~ £
d~[sab(~)rab(~)
+fl f(raa(~)
+rbb(~) 2rab(x))]
+S[s] (7)
' 2
exp
Sis]
=/ fl d[wa]
exp
~ ~j /
d~[wa(~) ((-V~
+~1)ôab sab
lx)) wb(~))
2
a ah
Assuming
an uniform saddlepoint (for large N)
withTab(~)
" pab +ôrab(~)
3ab(~)
" OEab +ô3ab(~) (8)
one can
expand
around it. MP thus obtainsii)
The mean field contributionÇ(°1(r; s)
(ii)
Thestationarity
condition(vanishing
of linearterms)
that defines p and a(identical
to the variational answer(~)).
.
Stationarity
withrespect
tosab(x):
Pab "
j ~jGab(P) (9)
(~) Only in trie
large
N limit. If N isfinite,
m trie vanational answerf(y)
isreplaced by /(y)
=
ù
fo°°doEa'~/~~~e~" J(2ay/N)
withf
-
J
as N - ce..
Stationarity
withrespect
torab(~):
°ab "
2flf' (paa
+ pbb2pab) (ii)
°aa "
~
OEaà(12)
b(~a)
(iii)
The Hessian matrix of thequadratic
form inôsab (after integration
of thequadratic
terms môrab
which inparticular
says thatôsaa
=
£~
~~ sab, 1-e-
W
~ exP
lNç°lP; ail / fl dlôsaài
exp
fl fl ô~~à(p)M~~>~~(p)ôscà(-p) (13)
a<b
a<j
p~~~~~~~~~
2
j"jpaa Î~Î
2pab) ~ ~~~~~~~
~~~~~~~ ~~~~~~ ~~~~~~~
~
lGac(P ~)
+Gbd(P ~) Gad(l'~ ~) Gbc IF ~)l (14)
3. Discrete Fourier Transform
We shall work in a formalism where the level R of
replica symmetry breaking
iskept
discrete(R
= 0replica symmetric,
R= one
step
ofbreaking,
R= oo Parisi
breaking,
1-e- continuousoverlaps).
In theend,
if necessary,we shall let R ~ oo.
For an observable aap, with aap = ar for an
overlap
onfl
= r, define the discrete Fourier transform as
R+1
âk
"~j
Pr(ar ar-1) (là)
r=k
and its inverse
ar =
f (âk âk+i) l~~~
~=o
Pk
Here the
pr's
are the size of the successive Pansiboxes,
po e n, pR+i + 1. The above transform has theproperties
ar ar-1 "
(âr âr+1) (17)
~ a°7b7P
r" ~°P
(18)
7
âklk
" ék
(19)
In
particular
(a ~)k
"
(~°)
k
All this is but a discrete version of the
operations
introduced in MP for the continuum(R
~ oo).
Note that in the above all
quantities
with indices outside the range(0,1,.
,
R
+1)
are to be takenidentically
null. If the limit R ~ oo is understood asâr
~â(u)
ar ~
a(u) (21)
then the connexion with MP
is,
m theirnotations,
à(u)
= àla) lai(u)
lai(U)
=/~ dta(t)
+Ua(U) (22)
4. Trie
Equation
of State With the above definitions we havePr =
j L GTIP) (23)
p
Ôk(P)
"
~ ~
(24)
and the
equation
of state~~ =
2p Il
~fl jGR+i(q) Gr(q)]lr
=0,1,.
, R(25)
fl
~ R+1
"
~j
pt(at at-1) (26)
t=0
The last
equation,
identical to(12),
also writesR+1
à~
=fl p~(a~ a~_~)
= 0
(27)
r=o
To solve for the
equation
of state one uses the transform that allows to writei~r(~) ~r-1(~)i
"
) [Ôr(~) Ôr+1(~)j (~~)
and
Yr +
j fl IGR+i (q) Gr(q)1
"
j fl ) lâk âk+i )Ôk (q)Ôk+i (q)
+ÔR+i q)1
~ ~ ~~~~~ ~
j29)
The
equation
of state then writes, withf(vr)
" 16 +
vr]~'
°r "
2fl/~iYr) (~°)
1-e-
~fl k
~~~
~~~~
~~~~ ~
~ ~~~ k
~~~~~~~ ~~~~~~~~~~
~~~ ~~~~~
~~~(31)
Here we have
Ôk(q)
"
~ ~
(32)
q ~ ak
~°k ~
ΰÎ(~)
p~ ~ ~
(in
the Parisigauge) easily restituting
the MP solution as given for the case of noise withlong
range correlations.
For later use, let us take a one
step
difference m theequation
of state°r °r-1 "
) (Ôr àr+1)
"
2fl1/~(Yr) J'iYr-1)1 (33)
with
Pr-i Yr =
j L lGr(q) Gr-i(q)1
=
j L
[Ôr(q) Ôr+i
(q)] (34)
q q
1-e-
"
~
~~ ~~~~)
flÔr(~)Ôr+ll~) (3à)
r ~
For small yr-i Yr
(infinitesimal
in thecontinuum)
onegets
1 =
-4£Ôr(q)Ôr+1(q)Î"(Yr) (~~)
q
a formula used below.
5. Hessian
Eigenvalues:
TheReplicon
SectorThe Hessian
(14)
written m terms ofoverlaps
becomesM~~'~~(P)
"MÎ(Î (P) (37)
with
r=anb
,
s=cnd
u =
max(a
n c,
a n
d) (38)
u=max(bnc
,
bnd)
In ultrametric space,
only
three of the fouroverlaps
are distinct. In thereplicon subspace
r = s, and we have two
independent overlaps
u,u > r +1. In thelongitudinal
anomalous sector, if r#
s, one cankeep max(u, u)
toparametrize
M.For the
replicon
sector, and ongeneral ground [9-11],
theeigenvalues
aregiven by
the double discrete Fourier transform~~~~
~ ~~Î
~~
Î
~"
~~~~~ ~~~i~ ~~~~-i
+Miii,~ ii j~~~
fork,1>r+1.
Using
the Hessianexpression
Mjjj
#~~j~Î~~~~ £ jG~(q)
+Gu(q) 2Gr(~)l lG~(P ~)
+ GV(P ~) 2Gr(P ~)l
~ ~
j40) Hence,
for the double transformÀplr; k,1)
=-) L
[Ôklq)Ôilp q)
+Ôilq)Ôklp q)j 141)
and
using (36),
valid inparticular
in thecontinuum,
ÀPI~Î k, i)
"
fl (2Ôr(~)Ôr+1(~) (Ôk (~)Ôl (P ~)
+Ôl (~)Ôk (P ~)j
i,k à r~+1 j42)
[rhe
propagators Ôr(q)
aredecreasing
functions of (q( and of r(since -âr
is anmcreasing function).
Hence the mostdangerous eigenvalue
isÀPI~,
~ +l,
r +Î)
" 2
L (àri~)Ôr+li~) Ôr+li~)Ôr+liP ~)) 1~3)
or m the continuum limit
>Pi~.
x~)
- 2
~l~D / dq~ q~
+
~
à~~~
~p q~~/
~
à~~~ lq~
+~ à~~~ 144)
1?or D <
4,
one can set the ultraviolet cut-off toinfinity
and rewrite thedangerous rephcon eigenvalues.
m the continuumlimit,
as>~j~;
~,~)
m
fl [à~jq) à~jp )j~ j4à)
where the
marginal stability
isobviously displayed (~)
Ii. Hessian
Eigenvalues:
TheLongitudinal
Anomalous Sector'fhe
analysis
oflongitudinal
anomalous sector isnotoriously
more diflicult. It has been con-siderably simplified by
the introduction[10-12]
ofkernels,
1-e-by going
to ablock-diagonal iepresentation
of the ultrametric matrices. Take in(37)
acomponent
of thelongitudinal
<~nomalous sector
M[1],
r#
s,which, given
thatonly
threeoverlaps
can bedistinct,
can bewritten
(5)
~M)"~
t =max(u; u)
'To
M)"~
is associated a kernel(in
fact a Fouriertransform) K["~,
where k =0,1, 2,...,
R+1,
k = 0corresponding
to thelongitudinal
sector. Likewise for the inverseM~~
of the mass(~) It is worth mentionmg at this
point
that trie hmit D- 4 is delicate to take. Indeed
solving
for triet~quation
of state(30-32),
one finds that triebreakpoint
xi[[a](x)
= 0 for ~ >xii
crosses itsboundary
value xi
" 1, at some intermediate dimension
DOIT),
2 < Do < 4. This delicate point is left out forfurther consideration. See also [13].
(~)
Except for trie componentsM[j[
that also mcorporate arephcon
contribution.operator
matrix we bave a kernelF["~
Therelationship
between them is theDyson's equation [10,11]
Éj,S
=~r,S f
~r>t /lk(t)
/ht;s
~~~~ 2
~_~
~
Ak(t)
~Éj'~
=-Àkir)Fj'~Àkis) 147)
Here the
Ak(r)
arespecial replicon eigenvalues
definedby
lÀ(r;k,r+1)
k>r+1Ak(r)
=
j48)
À(r;r+1,r+1)
k<r+1with as in
(39)
andÎ(Pt Pt+i)
t < k 1/lk(t)
=
)(pt 2pt+1)
t = k 1(49)
(Pt Pt+i)
t > k iThe kernel
K["~
itself is a transformed of the massoperator
matrixM)"~
It writes as ageneralized
discrete Fourier transformlr
s R+1Ki~
=L
+2L
+4L
p~
imi~ Mil[) iso)
t=k t=r+1 t=s+1
where we have
displayed
the case k < r < s(for
differentorderings
the summation£~~~
isonly kept
in the allowedsummands). Knowing K["~,
one thencomputes F["~
via(46, 47)
and(M~l)["~
via an inverse transform. That is if one is able to solve theintegral equation (46)
which is
only possible
withespecially simple
kernelsK["~.
This turned out to be the case whenstudying
the barepropagator
of the standard spmglass [10]
where~
K["~
=Ak(min(r; s)) (si
4
The same functional form also appears for the
spin glass
model introducedby
Nieuwenhuizen[14].
For the system studiedhere,
let us construct the kernelK["~
with the matrixM)"~
asgiven by
the Hessian(14) manifestly symmetric
in r, s. Let us chooser < s and we then find
R+i
jKi'~ Iv)
=L L
pt
iiGtlq)Gt (p q) Gt-i (q)Gt-i (p q)i
q t=max(s+i,k)
Gs lq) iGtlP q) Gt-i (P q)i iGt (q) Gt-i lq)i G~ iv q)j (52)
which also rewrite as
j~r;s
lBk(S)
r < 3 < ~ i~ ~
Bs+i(S)
s / k 1(53)
where
Bkls)
=
~ (Glq)S q))
~
Gs (q)Ôk (P q) Ô~ (q)G~(p q) (54)
The
important point
is that the kernelonly depends
onmax(r, s)
jKj>s
mB~(max(r; s)) (àà)
tbus
rendering
soluble theintegral equation (46).
7. Solution of trie
Dyson Equation:
ThePropagator
KernelIi the case of
(K["~
=Ak(min(r; s))
the solution of theDyson equation (46)
has beengiven (in
the continuumlimit)
m[10].
We wish to extend it here to the case(K["~
=Bk(max(r; s))
Ii this section we work for a given value of
k,
1.e. we remain in a sub-block of thelongitudinal
anomalous sector
(k
= 0 for thelongitudinal).
The
propagator
kernel ofequations (46, 47)
is obtained asi~ Ak(r)/lkAk(s)
~~~~where
/lk
is a Wronskian andil
theregular
andirregular
solution of a discrete Sturm-Liouvilleequation
andAk
thespecial replicon eigenvalues
of(48). Namely
we have for the Wronskian/~~
=~~ ~~~~~~Î~~~ ~Î~~~~~~~
~~~D~Bk(s) (57)
with
1~~
j~s~
ej~
~~
j~ (à8)
The functions
#*
aregiven by
lit js)
m i +
f jB~(t) B~(s)j
/~k~~~ij(t) (à9)
2
tmo
Àkit)
jij(s)
m
( jB~(t) B~js)j (kjjj ii (t) B~(s)c~ (60)
with
Bk(s)
as in(53).
Notethat, contrary
to the solution obtained for(si ),
whereCk
"
1,
here, #j
and#)
remamcoupled
viaCk
~~
=( ~k(t)
~+
~~~Ak(t)
kIn terms of
#)
the Wronskian can also be written as/~k
"4Ck
1 +~j ~k(S) )~))) Î(S)1.
~~
~~~~k
8.
Stability
Discussion in trie LA SectorGiven the
explicit
expressions(53)
ofBk(s)
in terms ofÔt (or Gt)
itself apositive
definitefunction, decreasing
withincreasing
tDtGt
"Ptli ~#~~ Gj~~GÎ~~ ~~~/~ (62)
one can
easily
show that(1)
B(r)
< °(63)
(ù) DrB(r)
eB(r
+1) B(r)
> 0(64)
Under these conditions we can now
give
a iower bound for thelongitudinal
anomalouseigen-
values. We can wnte out the determmant of the
eigenvalues
restricted to the sub-block k:det
(M))) ÀI)
=jj [Ak(r)
À]l
+
( Bk(t)
~
)~~~~ #)(t)j
(65)
r=0 t=0 ~
which is
directly
related to the Wronskian(apart
from theCk factor).
Note
that,
fromequation (59),
one also obtainsDs#t là)
=
-2DsBk(3) É ()~~
~
#t(t). (66)
With the
boundary
value#)(0)
= 2 and
D~B(s)
>0,
one infersthat,
in case of <Ak(t)
for allt,
4t(S)
> o(67)
D~#)(s)
> 0(68)
Consider
now in
equation (65)
a small value of such that <Ak(t)
for ail t For that value the determinant of(65)
cannotuanish,
it isalways positive
on account of(63, 67).
Hence the spectrum of thelongitudinal
anomalous sector iscertainly
boundedby
the bottom values of therephcon
spectrumÀp=o(r;
r+1,
r +1).
This establishesstability,
amarginal stability
asm the continuum limit where
Àp=o(~;
~,~)
= 0.
9. Conclusions
We thus have set up for random manifolds the formalism to discuss the
stability
of mean field solutions. For the well established continuous(R
=
oo) solution, adequate
for D <2,
we areable to prove its
marginal stability.
The formalism used here allows a discussion of anytype
(discrete,
continuous ormixed)
of mean field solutions.It is worth
noticing that,
whendiscussing
thestability,
theonly place
whereproperties
of the noise correlation function mayplay
a direct role(outside
of thefully
continuous R= oo
solutions)
is m(33-36).
There theequation
of state is used to find the alternativeanalytical
form forf",
that cast in evidence thenon-negative
character of therephcon eigenvalues.
Acknowledgments
One of us
(CD) gladly acknowledges
useful discussions with M. Mézard.References
[ii
Natterman T. and VillainJ.,
Phase Trans. Il(1988)
5.[2]
Halpin-Healey
T. andZhang Y.C., Phys. Rep.
254(1995)
215.[3] Balents L. and Fisher
D.S., Phys.
Reu. B 48(1993)
5949.[4] Hwa T. and Fisher
D.S., Phys.
Reu. B 49(1994)
3136.~
[si
Mézard M. and ParisiG.,
J.Phys.
I France 1(1991)
809.[6] Mézard M. and Parisi
G.,
J.Phys.
I France 2(1992)
2231.[7] Goldschmidt
Y.Y.,
Nucl.Phys.
8393(1993) 507;
J.Phys.
I France 4(1994)
87.[8] De Dominicis
C.,
Temesvari T. and CarlucciD.M.,
mpreparation.
[9] Kondor I. and De Dominicis
C., Europhys.
Lett. 2(1986)
617.[10]
De DommicisC.,
Kondor I. and TemesvariT.,
J.Phys.
I France 4(1994)
1287.[Il]
TemesvariT.,
De Dommicis C. and KondorI.,
J.Phys.
A Math. Gen. 27(1994)
7569.[12] De Dominicis C. and Kondor
I.,
J.Phys.
Lett. France 46(1985)
L1037.[13] Balents
L.,
Bouchaud J-Ph. and Mézard M. J.Phys
I France 6(1996)
1007.[14] Nieuwenhuizen