• Aucun résultat trouvé

Anomalous scaling of moments in a random resistor networks

N/A
N/A
Protected

Academic year: 2021

Partager "Anomalous scaling of moments in a random resistor networks"

Copied!
10
0
0

Texte intégral

(1)

HAL Id: jpa-00210497

https://hal.archives-ouvertes.fr/jpa-00210497

Submitted on 1 Jan 1987

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.

Anomalous scaling of moments in a random resistor networks

G.G. Batrouni, A. Hansen, M. Nelkin

To cite this version:

G.G. Batrouni, A. Hansen, M. Nelkin. Anomalous scaling of moments in a random resistor networks.

Journal de Physique, 1987, 48 (5), pp.771-779. �10.1051/jphys:01987004805077100�. �jpa-00210497�

(2)

Anomalous scaling of moments in

a

random resistor networks

G. G. Batrouni

(1,*),

A. Hansen

(2,**)

and M. Nelkin

(+)

(*) Newman Laboratory of Nuclear Studies, Cornell University, Ithaca, New York 14853, U.S.A.

(**) Department of Physics, Cornell University, Ithaca, New York 14853, U.S.A.

(+) School of Applied and Engineering Physics, Cornell University, Ithaca, New York 14853, U.S.A.

(Reçu le 18 dgcembre 1986, accepté le 26 janvier 1987)

Résumé. 2014 Nous considérons un réseau carré de résistances aléatoires au seuil de percolation de lien

pc = 1/2. Nous calculons la distribution de courant sur l’amas infini naissant au moyen d’une méthode de

gradient conjugué accélérée par transformation de Fourier. Nous calculons le n-ième moment de cette distribution à la fois dans l’ensemble à courant constant et à voltage constant sur des réseaux de taille variant

jusqu’à 256 x 256, et examinons le comportement d’échelle du moment en fonction de la taille. Une analyse

conventionnelle de nos données produit des exposants critiques en accord avec ceux obtenus dans des

publications antérieures. Les auteurs précédents avaient supposé que ces exposants sont déterminés théoriquement dans la limite de n grand par les liens simplement connectés. Ceci correspond à une hypothèse implicite sur l’ordre des limites pour lequel nous ne trouvons aucuene justification théorique. Nous réanalysons nos données en soustrayant la contribution des liens simplement connectés avant de calculer les moments. Dans la limite du réseau infini, ceci ne peut pas faire de différence mais, sur nos réseaux finis, les exposants critiques apparents sont fortement modifiés. Le problème que nous posons ne peut pas être résolu

numériquement. Nous examinons brièvement comment il pourrait être résolu théoriquement.

Abstract. 2014 We consider a random resistor network on a square lattice at the bond percolation threshold

pc = 1/2. We calculate the current distribution on the incipient infinite cluster using a Fourier accelerated

conjugate gradient method. We compute the n-th moment of this distribution in both the constant current and constant voltage ensembles for lattices up to 256 x 256 in size, and examine how these moments scale with lattice size. When analysed in the conventional way, our scaling exponents agree with published results.

Previous authors have assumed that these exponents are theoretically determined in the limit of large n by the singly connected bonds. This makes an implicit assumption about the order of limits for which we find no

theoretical justification. We reanalyse our data by subtracting the contribution of the singly connected bonds before calculating the moments. In the limit of infinite lattice size, this can make no difference, but for our

finite lattices, the apparent scaling exponents are strongly affected. The dilemma that we pose can not be resolved numerically. We discuss briefly how it might be studied theoretically.

Classification

Physics Abstracts

05.40 - 64.60

1. Introduction.

The

study

of fractal

objects

characterized

by

an

infinite number of

scaling

exponents is of consider-

able current interest. This

subject

was first studied in the context of

fully developed

turbulence

[1],

and

has led to some interesting recent

speculations

in

that context

[2, 3].

For more tractable

problems ranging

from

dynamical

systems

[4]

to diffusion

(’)Present

address : Department of Physics, Boston University, Boston, Massachusetts 02215, U.S.A.

(2 )

Present address : Groupe de Physique des Solides,

Ecole Normale Supérieure, 24, rue Lhomond, 75231 Paris Cedex 05, France.

limited aggregation

[5],

to

problems

in

polymer physics [6],

the

subject

has had much recent

activity.

Perhaps the most

straightforward application

is the

distribution of currents among the bonds on the backbone of a random resistor network at the

percolation

threshold

[7, 8].

To be definite we

consider a constant current ensemble where a unit current flows

through

the backbone in each realization. An

important

role is

played

by the

singly

connected bonds which carry all of the current. The number of these bonds has been shown by Coniglio

[9]

to scale as the

1 / v

power of the lattice size. In two dimensions, the correlation

length

exponent v is believed to be exactly 4/3

[10].

Article published online by EDP Sciences and available at http://dx.doi.org/10.1051/jphys:01987004805077100

(3)

772

The moments of the current distribution are

expected to scale as powers of the lattice size for

sufficiently large lattices. We define the

scaling

exponent qn

by

where (in>

is the n-th moment of the current distribution.

The exponent qo is

just

the fractal dimension of the backbone

[11],

the exponent q2 is the resistance exponent and the exponent q4 is related to the

amplitude

of the

1 / f

noise in a

percolating

system

[7].

Since the current is

always

less than or

equal

to

one, qn must be a monotone

non-increasing

function

of n. In any finite lattice, the moments for

large n

will be dominated

by

the

singly

connected bonds for which i = 1. This leads to the

assumption

that

which is 3/4 in two dimensions.

Equation (1.2)

is

strongly

supported

by

all

existing

numerical results

for finite lattices, and is supported

by Migdal-

Kadanoff renormalization group calculations

[12],

and by series

expansions

near six dimensions

[13].

The main theoretical contribution of the present paper is to note that

equation (1.2)

assumes that the

limits n -+ oo and L - oo commute. The numerical support for

equation (1.2),

and the

published

theor-

etical arguments in its favour, both

implicitly

take

the limit n -+ oo first. The definition of qn, however,

from

equation (1.1) clearly

takes the limit L -+ oo

first.

Why might

this make a difference ? Consider the n-th moment

(i n)

as a function of L for

large

but

finite n. The contribution to

(i n)

from the singly

connected bonds with i = 1 grows as the 3/4 power of L. We introduce the finite system

scaling

expo- nents defined

by

where 4 is a finite difference. For any finite n we

expect Zn:> 3/4. Thus for

sufficiently large

lattices

the contribution to

(i n)

from the

singly

connected

bonds will become

negligible compared

to the contri-

bution from the currents i 1. If we define the

exponents qn

by

the

limiting

process of

equation (1.1),

we see that for any finite n, no matter how

large,

the qn’s are not determined by the singly

connected bonds. Thus there is no obvious theoreti- cal reason to expect

equation (1.2)

to be valid.

The exponents qn for large n are determined

by configurations

with currents nearly

equal

to

(but

less

than)

one. These currents appear in bonds that are

connected in parallel to clusters with very

high

resistance ; an example of such a bond is given in

figure

1 where it is marked A. These

high

resistance

Fig. 1. - Several features in the current histograms in figures 2 and 3 are identified in this backbone. A is a bond

carrying a current close to one. B is perfectly balanced,

and thus carries no current. C are singly connected bonds,

or links. D and E are configurations responsible for the peaks marked D and E in the histograms.

clusters must

necessarily

be large. Furthefinore, if

such a cluster is removed from the backbone, the corresponding bond in parallel to it will become singly connected. The appearance of currents close to one are therefore associated with very non-local

properties

of the backbone. We see no a priori

reason why the distribution of these currents close to one, or

equivalently,

the distribution of the large

structures

causing

them, should be

simply

related to

the number of

singly

connected bonds thereby

making equation (1.2)

correct. We return to this

question

in section 4 since it is the essential theoreti- cal

point

raised by the present work.

We have studied this

question numerically

for

finite systems by

introducing

the « blob » moments Bn defined

by

where NL is the average number of

singly

connected

bonds per realization. That means we subtracted from each moment the contribution of the lines which is

unity

for each

singly

connected bond ; thus

we teep

only

the slob contribution. We then intro- duce the finite system

scaling

exponents

In the limit of an infinite lattice the exponent

z,,, y,,

and qn

must agree for any finite n. Our numerical results for L between 16 and 256 show that for the larger lattices, the fits to

equation (1.3)

and

equation (1.5)

are

approximately equally good,

but the exponents Zn

and yn

are

numerically quite

different for

large

n. In

particular

our

numerically determined Yn

seem to

approach

a limiting value

near 1.0 and not 3/4, as n becomes large, which the

Zn form the same simulation

approach

the

expected

limit 3/4.

(4)

In section 2, we present our numerical results for the statistics of the current distribution. In section 3,

we give the detailed results for the scaling of the

moments and the « blob » moments. We conclude in section 4 with a brief discussion of the theoretical

significance

of our results.

2. Numerical methods and results.

We study a two-dimensional square lattice of size L with two types of bonds. They either have a resist-

ance R with

probability

p or infinite resistance with

probability (1- p ).

In all of our calculations we

have taken p = 1/2, which is the threshold for bond

percolation in the infinite system or finite system with self

duality.

We look for an

incipient

infinite

cluster which carries current from one side of this lattice to the

opposite

side. We start from the lower

left-hand corner,

looking

for a connected cluster

reaching

across the lattice from the lower edge to the

upper edge. If no such cluster is found, a new lattice

is generated. If such a cluster is found, we then start

from the upper

right-hand

corner

searching

for the

first node on the upper edge connected to this incipient infinite cluster. This node at the upper

edge

of the lattice and the leftmost node at the lower

edge belonging

to the

incipient

infinite cluster are then chosen as the two contact points between the cluster and an external current source. In our constant current ensemble, a unit current flows

through

the incipient infinite cluster in each realization.

There are three classes of bonds in the cluster : those which will carry a current,

perfectly

balanced

bonds and bonds

belonging

to

dangling

ends. The

first two classes of bonds form the backbone

[14].

In figure 1 the bond marked B is

perfectly

balanced.

We

identify

the resistors

belonging

to the backbone by using the «

burning

»

algorithm

of Herrmann, Hong and

Stanley [15].

It is

quite

costly in computer time to

identify

the backbone at this

point

in the

calculation since the computer time needed grows as the number of

loops

in the backbone. An alternative way would be to solve the Kirchhoff

equations

for

the entire cluster, and then

identify

the backbone

from the currents that the resistors carry. As our iterative methods for

solving

the Kirchhoff

equations

improved, it turned out that this alternative method would have been more efficient for the larger

lattices. We only discovered this after we had

completed the calculations

reported

here.

To solve the Kirchhoff

equations

on the backbone,

we used the conjugate gradient method for lattices of sizes 50 and 200, and the Fourier accelerated

conjugate

gradient method for lattices of size 16, 32, 64, 128 and 256. We have described this method in a recent publication

[16].

For the largest lattices, the implementation of Fourier acceleration made the

conjugate gradient

method run

approximately

5 times faster in cpu time. We generated 1000

backbones for L = 16, 32, 50, and 64, 1 250 for

L = 128, 350 for L = 200, and 160 for L = 256. The currents were determined to within an accuracy of

10- 8,

which is the maximum obtainable

precision

on

the FPS 264 we used. The error criterion we used is discussed in reference

[16].

In figures 2 and 3 we

plot histograms

of the current distributions for L = 64 and L = 256. The statistical errors in these histo-

Fig. 2. - Relative number of bonds carrying a current i, N (i,

L )/ E

N (i, L ), versus loglo i for 1 000 lattices of

i

size L = 64. The leftmost bin in the histogram contains all bonds carrying a current i __ 10- 8.

Fig. 3. - A histogram of the same quantities as in figure 2,

but based on 160 lattices of size L = 256.

(5)

774

grams are comparable, and of the order of 2 % of the

height

of the last bin for the L = 64

histogram,

and

4 % of the height of the corresponding bin for the

L = 256 histogram. Note the finite size effects at

certain currents. These are associated with the features marked C, D and E in figure 1. C denotes

the « links » or

singly

connected bonds which carry a

current i = 1. If one of these were cut, no current would flow in the backbone. D denotes a configura-

tion with four bonds, each

carrying

a current

i = 1/2, and E a

configuration

with three bonds

carrying

a current i = 1/4 and one bond

carrying

a

current i = 3I4. These features are all more

import-

ant for L = 64 than for L = 256 as can be seen in

figures

2 and 3. A curious feature in these

histograms

is that the number of

perfectly

balanced bonds, marked B, relative to the total number of bonds in the backbone seems

approximately equal

to 0.5 %,

independent

of lattice size. In these

histograms,

which are normalized to one, the horizontal axis is

logarithmic

with ten bins per octave. This was a

strategic error on our part. Had we realized from the

beginning

that we were

going

to

emphasize

the

results for large currents so heavily, we would have

stored our current distributions in a different form.

Since we

computed

the moments of the distribution

directly

from the solution of Kirchhoff’s laws and not from the

histograms,

the

principal

results of this

paper are not affected.

3. Moments of the current distribution.

We computed the n-th moment of the current

distribution defined

by

where the sum over k goes over all bonds on the backbone for each realization, and the sum over r

goes over each of the

NR

realizations. These sums were computed

directly

and stored for each realiza- tion for n = 0 to 9.5 in intervals of 0.5. The contribution to each of these moments from the

singly

connected bonds

(links)

is the average number of links per realization. In table I we list our data for the cases n = 0, 2, 4, 6, and 8. The column to the far

right gives

the average number of links per realization. In each entry the top line is the full n-th moment, and the bottom line is the « blob » moment with the contribution from the links subtracted.

It is

easily

seen that, over the range of L studied,

n = 0 is dominated by the blobs, n = 2 has compar- able contribution from links and blobs, and n = 4, 6,

and 8 are dominated

by

the links. On the other hand

it is

equally

easily seen that the blob contributions to the

higher

moments are growing faster than the link contributions with

increasing

system size. This is a natural consequence of the links

forming

a

vanishing

subset of the backbone as L - oo ; the blob contri-

Table 1. - Shows the current moments in > and « blob » moments

Bn

= in > -

NL for

n = 0, 2, 4, 6, and 8, and

lattice sizes

L = 16, 32,50,64, 128,200, and 256. The top entry in each line is the full moment, and the lower entry is the« blob » moment. The column to

the far

right is the number

oflinks, NLI for

each lattice size.

(6)

bution to the higher moments has to grow faster than the total moment with

increasing

L in order to

« catch up with » the link contribution and eventually

dominate the moments. This leads to y,, -- zn

[17].

The

question

which of the finite size exponents y, or z, better approximate the « true » exponents

qn cannot be determined from this observation :

ignoring

all other finite size effects than the presence of the links, there are two possible situations that are

possible : A)

The full moments scale with z,, = qn, and the « blob » moments approach the

straight

lines of the full moments from below in

log-log plots

versus L.

B)

The blob-moments scale with y,, = zn, and the full moments approach the straight

lines

of the « blob » moments from above in log-log

plots

versus L. Both of these scenarios obey the

inequality

above.

It would seem as this

point

that the only way to

answer the

above-posed question

is

by

determining

whether the full moments or the « blob » moments scale better for the lattice sizes that can be studied

numerically.

However, the so-far

ignored

effects of

the boundaries will restrict the sizes of both the full moments and the « blob » moments for the smaller

lattices, thus

increasing

the values of both y,, and z,,. Now, if scenario A is correct, we expect that both

the full moments and the « blob » moments to

approach

the

asymptotic straight

lines from below in

log-log plots

versus L, and with the « blob » mo-

ments

having

more curvature than the full moments.

If, on the other hand, scenario B is correct, we expect that the « blob » moments will

pick

up some

Fig. 4. - The Oth current moment

(i °)

and the corre-

sponding « blob » moment Bo =

(i °) - NL

as a function

of lattice size. The straight lines are based on least squares fits. The fit to the full moment is based on all the data

points. The fit to the « blob » moments neglects the L = 16 point.

curvature in these plots and

again approach

the

asymptotic

lines from below. The full moments

however, will have the effect of the links and the

boundary

effects

pulling

in

opposite

directions, and

the net effect is that the full moments will reduce their curvature. Thus, it may appear that the log-log

plots

of the full moments are closer to

straight

lines

also in this scenario.

To determine the scaling exponents z,, and

y,,, we carry out a least squares fit to

equations (1.3)

and

(1.5) respectively.

In

figures

4, 5 and 6 we give

Fig. 5. - The second current moment

i2 >

and the

corresponding « blob » moment B2 as a function of lattice size.

Fig. 6. - The 8th current moment

i 8>

and the corre-

sponding « blob » moment B8 as a function of lattice size.

(7)

776

Table II. - Shows the scaling exponents Zn based on the

current moments i" ), and the scaling exponents y.

based on the « blob » moments

Bn

= in ) -

NL

for

n = 0, 2, 4, 6, and 8.

log-log plots

whose

slopes

lead to those exponents for n = 0, 2 and 8. In table II we

give

the

resulting

exponents based on lattice sizes 32 to 256, and in

figure

7 we

plot

these exponents as a function of n.

We see that the full moments give better fits than the

« blob » moments, but except for 1 = 16, the differ-

ences are small. However, the

resulting

effective scaling exponents yn and zn are very different.

The exponents zn appear

large compared

to previous- ly published values

[7, 8].

We attribute this to our

choice of boundary conditions ; our backbones are

connected to the current source

through only

two

nodes on

opposite

edges of the lattice, while in other calculations the backbone has been connected

through

bars along each of these edges. For

n = 0 the difference between yo and zo is small since the links do not play an important role in the

effective fractal dimension of the backbone, even for

small lattices. For n = 2, our exponent y2 is consider- ably

larger

than Z2, and this implies a substantial

Fig. 7. - The scaling exponent z,, based on the full current

moment, i’>, and the scaling exponent yn based on the

« blob » moment Bn = in) - NL as a function of n.

change

in the resistance exponent

[18].

A similar

change

would have been found by other authors,

had they

analysed

their data in the way that we did., For larger values of n, all of the moments for all L up to L = 256 are dominated by the links, but the

contribution of the blobs is still

accurately computed.

In light of the previous discussion of the effect of the boundaries on the

scaling

behaviour of the full moments and the « blob » moments, we find it impossible to

distinguish

between scenario A or B,

or

equivalently,

determine

whether yn

or zn are closer to the « true » exponents qn.

Finally

we consider the distinction between our

constant current ensemble and the constant voltage

ensemble considered

by

other authors

[8].

Instead of

holding

the current in each backbone fixed, we could have

kept

the

voltage

across the backbone fixed.

The current in the links would then have been

proportional

to the conductance of that backbone rather than

being

fixed at i = 1. The current in each

of the other bonds would also have scaled pro-

portional

to the conductance. This would smear out all of the finite size effects in the current distribution,

making

the

corresponding histograms

of the currents

distribution smooth. Especially, there would be no

trace of the distinctive

peak

due to the links. In reference

[8],

large lattices were analysed in a

constant

voltage

ensemble with the scaling assump- tion

Under reasonable

assumptions

of statistical

indepen-

dence, it was suggested in reference

[8]

that

Since we stored all of the moments for each realization, we could

analyse

our data for the

moments in a constant voltage as well as a constant current ensemble. We found

good

agreement with

equation (3.3).

The statistical

independence

assump- tions

leading

to

equation (3.3)

are not obvious. The

n-th current moment in the constant-voltage ensem-

ble can be written

as Gn in)

in the constant-current

ensemble, where G is the conductance of the back- bones.

Equation (3.3)

assumes that this moment scales

as G) n i n).

To check this assumption, we

measured the correlation between the conductance and the number of links, since we expect the current associated with the links to be more correlated to the conductance than the currents associated with the bonds making up the blobs. We found a correlation coefficient of

approximately

0.2, and we found that

this correlation coefficient did not vary

appreciably

with lattice size.

4. Theoretical considerations.

When

analysing

a

probability

distribution which contains an infinite number of

independent

scaling

Références

Documents relatifs

voltage characteristic I(V) of an electrical network where each bond is a non linear resistor character- ized by a threshold value vg for the the potential drop..

The scattering function and radius of gyration of an ideal polymer network are calculated depending on the strength of the bonds that form the crosslinks..

Tadaki considered for a given prefix-free Turing machine and some natural number a the set of all strings that are compressed by this machine by at least a bits relative to their

Sparse denotes the ratio of null coefficients in the dense representation for the total degree d. For example “Sparse = 66% “ means that 66% of the coefficients are equal to zero in

178 - Oliviero Forti ed altri, l'immigrazione clandestina in Italia, la traduzione del cielo all'alba, l'arte di Bwomezia per la stampa, Roma, Italia, dicembre 2005.. 179 -

3: simulated temperature against time of a silver bar, with various applied DC voltages as indicated, dotted line indicates temperature of 89 K, where the bar temperature

Contrairement aux publications de la langue anglaise, qui restreignent l'IMF aux cas de sepsis documentés par la positivité d'un prélèvement central, hémoculture ou culture du

K EY WORDS: Centripetal force, Gravitational constant -G, Gravitational Force, Hydrostatic equilibrium, Linear acceleration, Mass, Orbital and escape velocity,