HAL Id: jpa-00246480
https://hal.archives-ouvertes.fr/jpa-00246480
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.
Analytic determination of the complex field zeros of REM
C. Moukarzel, N. Parga
To cite this version:
C. Moukarzel, N. Parga. Analytic determination of the complex field zeros of REM. Journal de
Physique I, EDP Sciences, 1992, 2 (3), pp.251-261. �10.1051/jp1:1992141�. �jpa-00246480�
Classification Physics Abstracts
75.10N 05.20
Analytic deterndnation of the complex field
zerosof REM
C.
Moukarzel(*)
and N.Parga
Centro At6mico Bariloche, 8400 S-C- de
Bariloche(**),
Rio Negro, Argentina(Received
5 September 1991, accepted in final form 21November1991)
Abstract The complex field zeros of the Random Energy Model
are analytically determined.
For T < Tc they are distributed in the whole complex plane with a density that decays very fast with the real component of H. For T > Tc a region is found which is free of zeros and is
associated to the paramagnetic phase of the system. This region is separated from the former
by a line of zeros. The real axis -is found to be enclosed in a cloud of zeros in the whole frozen
phase of the system.
1 Intro duction.
The determination of the zeros of the
partition
function allows muchinsight
into the behavior of a model. Since their relation withphase
transitions were clarifiedby
the works ofYang
and Lee[lj
2] and Fisher [3], manyinteresting properties
of zeros distributions have been studied inhomogeneous
systems[4-9].
More
limitedj
on the otherhand,
are the resultsconcerning
zeros distributions of disordered models.Taking
into account that all the information about agiven
model is contained inthe zeros of its
partition function,
it is reasonable to expect that many of theinteresting
properties
[10] of disordered systems will be somehow reflectedby
them.An
example
of this can be found in thecomplex magnetic
field zeros ofspin glasses
j
which are dense near the real axis. This property has been
suggested by
Parisi[I
I] asbeing
a consequence of themultiplicity
of states and wasrecently
illustrated [17]by
a numerical calculation of thecomplex
field zeros distribution in the RandomEhergy
Model(REM) [12].
Here an
analytical
calculation of thecomplex
field zeros of REM isreported.
Thetechnique
is the same as
recently employed by
Derrida [13] to calculate thecomplex fl
zeros of the samemodelj
which had been alsonumerically
estimated [14, 15].In section 2 we describe the basis of the REM and
give
an account of the ideas that allow the calculation of its zeros. Section 3 is devoted to the REM in acomplex
field and the calculation(*
Fellowship
Holder of the Consejo Nacional deInvestigaciones
Cientificas y T4cnicas.(**) Comis16n NacionaJ de Energia At6mica.
JOURNAL DE PHYSIQUEI T 2, N' 3, MARCH 1992
of the dominant contribution to its
partition
function in eachregion
of thecomplex
h =flH plane.
In section 4 these results are used to obtain thedensity
of zeros of REM in thecomplex
field
plane,
while section 5 contains our conclusions.2. The random energy model.
The REM is defined [12] to have 2/~
independent
random energy levelsE;, I=1,..2/~
distributedaccording
toP(E)
=(xN)~~/~ exp(-E~ IN).
Thepartition
function of asample (a sample
isa set of 2/~ numbers
E;),
Z =£(~~ e~P~<
can be written as Z =£~ n(E) e~P~
wheren(E)
is the number of levels between E and E + AE. This
density
of statesn(E)
if found tosatisfy
(n(E))
~- exPN(log
2(()~)
~~~
(n(E)~) (n(E))~
~-
(n(E))
so two different energy
regions
can be identified:If
(E(
<Nec
,
with cc =
(log 2)~/~,
then the number of statesn(E)
is much greater than one and its fluctuations aresmall,
so for atypical sample
we will have [13]n~~~(E)
=In(E))
+nE(n(E))~/~ (2)
where qE is a random number of order I.
On the other
hand,
if(E(
> Nec,
then the number of states is
exponentially
smallmeaning
that a
typical sample
will haveno levels in this
region
+~YP(E)
= °(3)
The
partition
function for atypical sample
will then beztyp
= A + B
(4)
with
A =
£ (n(E))e~~~ (5)
jEj<Ne<
and
B"
L QE(n(E))~~~e~~~ (6)
jEj<Ne<
The calculation of A may be done
by
the steepest descent method [16] forcomplex fl
=pi +ifl2.
+ec A
~
2~
deexp
-N(e~
+fle) (7)
e<
The
integration
contour is stretched toinfinity through paths
of steepest descent for the realpart of the exponent of the
integrand.
When theintegration
limitslay
onopposite
sides of the saddlepoint
in e =-fl/2.we
can write/+ec
-co +co +ec
# + +
(8)
-ec
~ec ~co ~co
and A has contributions both from the saddle
point
and from theintegration
limits forlarge
N. This
happens
if[pi
< flc = 2ec "2(log2)~/~
and we getA ~- exP
lNflecl
+ exPlN(log
2 +(fl/2)~) l (9)
whereas if
)pi
> flc the two steepest descentpaths
run from theintegration
limits down toinfinity
on the same side of the saddlepoint
so it does not contribute and we have/+ec
+co ~E<
" +
(1°)
-ec
~ec ~cc
so
A
~
exp(Nflec) (11)
It has to be mentioned [13] that the lowest
lying
states in REM have order one fluctuations fromsample
tosample
[12] so the contributionscoming
from theintegration limits,
both in(9)
and(11),
are known up to afluctuating complex (if fl
iscomplex)
factor of order one. We will return to thispoint
later on because it results crucial for our calculation.The contribution B of the fluctuations can be calculated for
large
N.Owing
to the ran- domness of qE, B is not theintegral
of ananalytic
function so the steepest descent method isno
longer applicable.
On the otherhand,
the qE are uncorrelated for different values of E so the term with thelargest
modulus will dominate the sum in(6).
We can then estimate themodulus of B as
(B( ~
_~~fl~~~~ ll~(E))~/~~~~~ (12)
It is
easily
seen that whenlpi
<flc/2
the maximum is in between the limits solBl
~- exPI(('Og
2 +Pi)) (13)
while if
lpi
>flc/2
one of the limits dominates and(B(
~- exPIN (Pi lee)1 (14)
We will now use these results to calculate Z for
a
typical sample
of REM when acomplex
field H acts on it. We will take
fl
real from now onj as well as hi and h2non-negative
(h
=flH
= hi +ih2)
which can be done without any loss ofgenerality.
3. REM i11 a
complex
field.When a
magnetic
field H acts on the system, theenergies
E; of the states nowhave,
in addition to the internalenergies (coming
from interactions amongspins)
a contribution from the interaction with thefield, E;
= U;HM;j
where theM;
are themagnetizations
of thestates. The
partition
function Z =£;
e~P~< is then~N
z(p h) £ ~-pU,+hM, (~~)
,
i=I
The internal
energies
U; have a Gaussian distribution which isindependent
of thefield,
sotaking
into account that among the2~
states of the system,gm = gM =
N~M
havemagnetization
M =Nm,
we can write 2Z(fl, h)
=£ e~~'~
Zm(fl) (16)
-iimii
where
gm
Zm(fl)
=
L e~~~' (17)
is the
partition
function of asample
of REM with gm instead of2~
states. Ifwe write gm =
(qm)~j
then all the results of last sectionconcerning
Z are valid for Zm with the solecondition of
replacing log
2by
log
qm =log
2 [(1 +m) log(I
+m)
+(I m) log(I m)]
(for large N).
These restricted
partition
functions Zm can be written asZm(fl)
=L n(tL, m) e~~~" (18)
where
n(u,m)
are the numbers of states with internal energy NV andmagnetization
Nm.They
can be seen tosatisfy
(n(lt, m))
~- exPN('°g
qmtt~)
(19) In(U, m)~) ln(U, m))~
~-
ln(U, m))
Then for a
typical sample
we will have(YP(U, ml
~-
(n(Ui ml
+ n(tL,m)(n(tL, m)1~/~
,
if Iv <
tLc(m)
~~o~
n YP(tL,
m)
~- °,
if (U( >
tLc(m)
where the
q(u,m)
are random numbersdescribing
thetypical fluctuations,
and the energy thresholduc(m)
is nowm-dependent.
uc(m)
=(log qm)~/~ (21)
The
partition
function of atypical sample
under field will then beZ~~~(fl, h)
"Le~~~Z$~~(fl) (22)
m
We now use the results of last section and write
zjYP(p)
=
A~(p)
+B~( p) (23)
with
Am
= pm exp(Nfluc(m))
jif
fl
>2uc(m)
(24) Am
= pm exp(Nfluc(m))
+ exp(N(log
qm +(fl/2)~))
,
if
fl
<2uc(m)
and
Bm
= nm explNfltLc(m)I
i
if
fl
>tLc(m)
Bm = nm exp
I)('Og
qm +fl~)I
,
if
fl
<tLc(m)
~~~~
Here qm is
a random number of order one which describes the fluctuations in the
density
of states as
already
mentioned. Theorigin
of pm is somewhat different.They
are due tofluctuations in the values of the lowest
lying
states: Theground
state energy(internal
energyin this
case)
Uo has order onesample
tosample
fluctuations[12,
13] soU)YP(m)
~N(uc(m)
+(IN)
with(
an order one randomnumb<r
soeP~°
~
pe~P"<(~')
j
where p is
apositive fluctuating
number.
We see in this way that both in A and B the
ground
state contributions arenon-analytic (due
todisorder),
and thereforenon-integrable by
the steepest descent method. Theonly
analytic
contribution is thesaddle-point
term in(24).
Let us now turn to the evaluation of the sum in
(22).
We defineflc(m)
asflc(m)
=21tc(m) (26)
and its inverse
mc(fl)
such that satisfiesi 1/2
fl/2
=log
2
[(1+ mc) log(I
+mc)
+(I mc) log(I
-m
)]j (27)
2
The relation
hc(fl)
=atanh(mc(fl))
defines the critical fieldhc
which is the limit between the frozen andparamagnetic phases
of REM with field [12]. This critical field satisfiesfl/2
=Jog
2 +log(cash(hc)) hc tanh(hc)]~/~ (28) Using (26)
we can rewrite the conditionfl
>2uc(m)
as (m( >mc(fl)
and the conditionfl
>uc(m)
as (m( >mc(2fl).
The sum in(22)
can now be writtenZ~YP
= A+ B
(29)
with
A =
£
Am(fl)e~~'~ (30)
m
and
B =
£ Bm(fl)e~~'~ (31)
We then have
A = A~ + A~ =
£ p~eNlPu<(m)+mhl
+£ eNj(p/2)?+iogq~+mhj
~~~~
-ismsi jmj<m~(p)
B = B~ +
B~
=£ ~~eNlPu<(m)+mhl
+£ ~~e+lP~+'Ogqm+2mhj
~~~~lml>mc(2p) jmj<m~(2p)
The
only analytic
contribution isA2,
which will be estimatedby
means of the steepest descent method. We willignore
Bi because it isalready
included inAI
Let us first discuss theevaluation of the
fluctuating
contributions Al and 82. Thenon-analyticity
of theintegrand
makes it
impossible
to use the steepest descent method. On the otherhand,
asalready
discussed in section II, the term with the greatest modulus will dominate the sums so we have(AI
~-_n~~~
exPlNlfl(logqm)~/~
+mhill (34)
and
(82
'~ mm exp( ) ~d~(log qm)
+2mhi] ) (35)
(m( <m<(2p) It is worth
noticing
that(82(
= 0 iffl
>flc/2 (36)
On the other
hand,
iffl
<flc/2,then mc(2fl)
> 0 and there exist two ranges ofhi
in which 82gives
different contributions:The maximum of
[82(
is at iii=
tanh(2hi)
when(l(
<mc(2fl)
and at+mc(2fl)
when [iii) >mc(2fl).
This can bewrittenj using (27)
and(28)
asexp (
~
~d~ +
log
2 + log(cosh(2hi))])
,
if hi <
hc(2fl) [82(
'~~ ~
(37)
exp
(N
§Y~ +[hi (mc(2fl)])
,
if
hi
>hc (2fl)
Now we discuss the evaluation of
AI
The maximization condition is satisfied at fh =tanh(I)
where
I
isa function of
hi
andfl
that satisfies(
=
(log
2 +log cosh(I) I tanh(I)j
~~~(38)
If we put
I
=
hc($)
with$
=
all
,
now the unknown is
a(fl, hi)- Replacing
this in(38)
we findI
=
ahi,
so thefollowing
set ofequations:
1#/2 #h
=Jog
2 +log cosh(I) I tanh(I)]
~/~ ~~g~=
PI
is
equivalent
to(38).
It is easy to see that when hi =
hc(fl)
we obtain a= I while a < I if
hi
>hc(fl)
and a > I ifhi
<hc(fl),
so(
willalways
be betweenhi
andhc(fl).
The term
AI
will then have thefollowing
formjAi1
~ eXpjNj(
+ hitanh(i)jj (40)
The calculation of A2 may be done
by
the steepest descent method(see
theappendix
fordetails)
forlarge
N. It follows thatagain
two differentexpressions
arefound,
nowdepending
on the values of h and
fl.
A smooth arc is found in thecomplex
hplane
which divides these twopossibilities.
This arc touches the real axis athc(fl)
and theimaginary
axis at h =ix/2.
In the inner part of this arc we find
A2
'~ expN( ~
+hmc(fl))
+ expN( ~
~ ~~ +log(cosh(h))) (41)
and in the outer side
A2
~ eXp
Nj(
+
hmc(fl)j (42)
It is clear that Z will in
general
be the sum of several terms of the form e~W with~9 a
complex function,
so forlarge
Njust
that term with the greatest modulus will be relevant. We then have todetermine,
in eachregion
of thecomplex
hplane,
the term with the greatest real part of ~9. This willgive
rise to the appearance of aphase diagram
for the model incomplex
h.In what follows we will be
only
interested in the determination of the modulus of Z,which,
aswe shall see, is
enough
to calculate thedensity
of zeros of the model. If we write Z[=
expN#,
the different contributions we have to compare are
11
" ~~~
+
log cosh(h)
forfl
< flc and[hi
<h(~~(h2, fl)
4~
= +[hi jm~(p)
forp
< p~ andvhi
#~
= +hi tanh(I)
Vfl andVhi
#4
" +)
+
log(cash(2hi))
forfl
< and[hi
<(hc(2fl)
#5
= fl~ +[hi tanh(hc(2fl))
forfl
<'
and[hi
>h~(2fl)
Three different ranges of temperature can be identified:
a)
Iffl
> flc thenonly
#3 exists so ~°~ ~=
#3 ,Vh.
b)
If flc >fl
>flc/2
we get contributionsNfrom11, 42
and#3.
In this case two differentregions
are found in the
complex
hplane (Fig, I,
upperportion), according
to what term is the mostimportant.
Theseregions
areseparated by
an arc which goes from(hc(fl),0)
on the real axis to(0, h[(fl))
on theimaginary
axis. In the inside of this arcii
dominates while in the outside it is#3
which is the mostimportant. #2
is never relevant.c)
Iffl
<flc/2
then also#4
and#5 contribute,
andthey
have to becompared
toii
and#3
in theirrespective
zones. It is found that#5
is neverrelevant,
I-e-(11 45)
and(#3 Is)
arenever
negative.
On the other band#4
isgreater
than#3 (for hi
<hc(2fl)
and it is also2
greater
thanii
in the upperportion
of the zonewhere11
>#3,
so the line where#4 equals ii
is lower than that where#3 equals ii
The situation in case
c)
is thatdepicted
in the lowerportion
offigure
I. The arcseparating
#i
and#4
now touches theimaginary
axis at apoint (0, h((fl))
which results fromequating #4
to
ii
forhi
" 0 and satisfiesp2 ~p2
CDS(hl(fl))
= e~~ ~ 8(43)
so we can see that
h)(fl
=
0)
=x/4.
w/2
~~
fl = 0.75 flc
«/4 ~
@~
w/2
4,
fl= o.30 j
4l3
«/4
#~ hi
o-o
o-o 1.o 2.o 3.o
Fig. 1. Shown are the phases of the model in the complex field plane for two different values of the temperature. For Tc < T < 2Tc
(upper figure)
two distinct zones appear separated by an arc of zeros.Above this arc
(phase 3)
densely distributed zeros are found whfle the region below it(phase I)
is freeof zeros. For T > 2Tc the outer part of the arc appears divided in two zones by a vertical line, which
is not a line of
zeros. The densities of zeros are different to the left
(phase
4) andright (phase
3) ofthis vertical line.
4.
Density
ofcomplex
field zeros of REM.Having
obtained#
= ~°~ ~ in the variousphases
of themodel,
we can calculate thedensity
Nof zeros
p(h)
of Zby
means of the formula [13]1
32 32
~~~~
2x~~~
2x
3h]
~
3hj) ~~~~'~~~
~~~~which is a consequence of the electrostatic
analogy
[2] that identifiesp(h)
with acharge density
and
#
with theresulting
electrostaticpotential.
The
density
of zerosassociated
toii
is null(because
the real part of ananalytic
function has avanishing Laplacian).
The densities of zeroscorresponding
to#3
and#4
arerespectively
p2
p~ =
-i7~#3
=
a(I tanh~(ahi))
~ ~
(45)
2x fl2 +
2hi(1
tanh(ahi))
p4 "
i7~#4
"2(1- tanh~(2hi)) (46)
where a
=
ilhi
asalready
defined.In addition to these distributed zerosj there exist other zeros located
right
on the lines divid-ing
the different zones, and whichdensity
isproportional
to thediscontinuity
in the normal component of the '~electric field" across that line. On a lineseparating
twogeneric phases
aand b we then have
Pt~
=) it(4a -16) (47)
The
density
of zeros on line 1-3 is found to be~~~
x~~~~~~~~~ osh(i~~~~~s(2h2)~~
~
cosh(i~~~~~s(2h2)~~~
~~~~
On line 1-4 the
expression
is the same with theonly
difference thatI (I
=
ahi)
isreplaced by 2hi
Their zero-densities areequal
where lines 1-3 and 1-4 meet.The line
3-4,
which is the vertical athi
"jhc(2fl)
is not a line of zeros, I.e. the electricfield is found to be continuous there. On the other band the
density
of distributed zerosjumps
as one crosses line 3-4
by
a factor(1+ 2(hi/fl)~(l tanh~(2hi))).
The
density
of zeros on the line 1-3 is found to be null at thepoint
where that line touches the real haxis,
inagreement
with the second order character of the field-driven transition of the model.5 Conclusions.
We have estimated the behavior of the
partition
function of REM incomplex
field. Three differentphases
appear:For T < Tc there is
only
onephase
which we calledphase
3 and is the continuation tocomplex
field of the frozenphase
of the model. For Tc < T theparamagnetic phase
is also present, which we callphase
I(Fig. I,
upperportion). Finally
for T > 2Tc what we calledphase
4 also appears atcomplex
values of h(Fig. I,
lowerportion).
There is nophysical (I.e.
for real values of
h)
counterpart for thisphase.
We have also calculated the densities of zeros in the different
phases.
Aparticular
structure has been revealed: The zeros are dense near the real axis in the frozenphase.
This characteristicof the zeros distributions is
intimately
related to theproperties
ofspin glasses
and has beendiscussed
by
ParisiIll], although
noexplicit
calculation had been done before. A numerical estimate of thecomplex
field zeros of REM has been done [17] and showsgood
coincidence [18] with the results herereported.
The existence of distributed zeros is in this model related to disorder in
a very clear way:
The
fluctuating
factors pm and qm areresponsible
for the appearance, inZ,
ofnonanalyticities
of the sort which are necessary to obtain a non-zero
Laplacian
andconsequently
distributedzeros.
Acknowledgements.
We thank B. Derrida for
sending
us his resultsprior
topublication.
Appendix.
Here we describe the evaluation ofA2
by
means of the steepest descent method [16].Starting
from
A2
"~~~~~
dm
e~'(~') (Al)
-mc(P) with
~(m)
= mh +log
2 [(1 +m) log(I
+m)
+(I m) log(I m)]
+~ (A2)
2
~
we see that
A2
is non-zeroonly
forfl
<flc,
otherwisemc(fl)
= 0 and this term does notcontribute.
Following
the usualprocedure
fore~,aluating
this kind ofintegrals
in thecomplex plane,
westretch the
integration
contour toinfinity following paths
of steepest descent for the real part of it Theimaginary
part of theintegrand
will in this way be constantalong
the wholepathj
and may be taken out from theintegral.
The level curves of it aretopologically
similar to those of z~. Two steepest descent lines cross each other at the saddlepoint
which is located atmo "
tanh(h):
one of them goes down to thevalleys
to the left andright
of thesaddle-point
and will be called AA'while the other climbs towards the mountain tops and is the line LL'.
The way in which the
original integration path
has to be deformeddepends
on the relativepositions
of +mc and LL'.Depending
onfl
andh,
two situations arepossible:
a)
When theintegration
limits+mc(fl) lay
onopposite
sides ofLL',
thereare contributions from three
paths:
the one that runsthrough
thesaddle-point (path AA')
j
and those
ending
at+m~(fl).
b)
When bothintegration
limits are on the same side ofLL',
thepath
thatjoins
thevalleys through
thesaddle-point
does not contribute to theintegral.
The limit condition for this two cases is that
mc(fl) lays exactly
onLL',
which in turn isequivalent
to the condition thatImag(it(mc)) =Imag(it(mo)).
This lastidentity
results from the fact thatLL', being
a steepest descent for the real part ofit,
is a level curve for theimaginary
part of it. This limit condition is then writtent~nh(hi)
"~~~~~~~~~~ ~~~~~~~~~
~~~~
and for
fl
< flc it defines a smooth arc in thecomplex
hplane (with hi
and h2Positive)
thatjoins
thepoints (hc(fl), 0)
and(0, x/2).
Theshape
and size of this arcdepend
onfl.
In itsinner part situation
a)
is valid so forlarge
hi2 +~ ex
~ ~
~ ~ ~ ~
~'~C(fl))
+ expNj fl
+fll
4 ~
'°g(cosh(h)))
while in the outer side of this arc
b)
holds andA2
~ exp
N( ~~
+hmc(fl)) (A5)
References
[Ii
Yang C. N. and Lee T. D., Phys. Rev. 87(1952)
404.[2]Yang C. N. and Lee T. D., Phys. Rev. 87
(1952)
410.[3] Fisher M., Lectures in Theoretical Physics 7C (University of Colorado Press, Boulder, Colorado,
1965).
[4] Jones G., J. Math. Phys. 7
(1966)
2000.[5] van Saarloos W. and Kurtze D, J. Phys. A : Jfath. Gem. 17
(1984)
1301;Stephenson J. and Couzens R., Physica A 129
(1984)
201.[6] Suzuki M., Prcg. Theor. Phys. 38
(1967)
1243,1255;Ono S. et al., J. Phys. Sac. Jpn 26s
(1969)
96;Abe lt., Frog. Theor. Phys. 38
(1967)
72 322.[7] Itzykson C. et al., Nuc. Phys. 13220 [FSB]
(1983)
415.[8] Giasser M. L. et al., Phys. Rev. B 35
(1987)
1841.[9] Bhanot G. et al., Phys. Rev. Left. 59
(1987)
803;Marinari E., Nuc. Phys. 13235
[FSI Ii (1984)
123;Alves N. et al., Phys. Rev. Lett. 64
(1990)
3107.[10]Spin Glasses and Beyond, M. hiezard, G. Parisi and M. Virasoro
(World Scientific)
Lecture Notesin Phys. Vol 9.
[ll]
Parisi G., in Disordered systems and Localization, Springer Lecture Notes in Phys., Vol. 149.[12] Derrida B., Phys. Rev. B 24
(1981)
2613;Derrida B., Phys. Rep. 67
(1980)
29.[13] Derrida B., Physica A 177
(1991)
31.[14] Moukarzel C. and Parga N., in Proceedings of the Second Latin American Workshop on Nonlinear Phenomena, Santiago de Chile, September 1990
(North
Holland,1991).
[15] hioukarzel C. and Parga N., Physica A 177
(1991)
24.[16] Asymptotic Expansions, E. T. Copson Ed.
(Cambridge
Univ. Press,1971).
[17] Moukarzel C. and Parga N., Complex Field Zeros of REM: A numerical estimation submitted to J. Phys. A
(1991).
[18] Moukarzel C. and Parga N., The ltE&I Zeros in the Complex Temperature and Magnetic Field Planes, Proceedings of "International Conference on Complexity: Fractals, Spin Glasses and
Neural Networks", ICTP, Trieste 2-6 July 1991, to appear in Physica A