You are on page 1of 5

a __

_-
BB PHYSICA
1
ELSEVIER Physica A 207 (1994) 208-212

Entropic analysis of random morphologies


C. Andrauda, A. Beghdadib, J. Lafait
Laboratoire doptique des Solides, Universitt Pierre et Marie Curie, C.N.R.S., 4 Place Jussieu,
75252 Paris Cedex 05, France
L. P.M.T.M-C.N.R.S., Institut Galilte, Universitt Paris Nord, 93430 Villetaneuse, France

Abstract

When the random morphology of ramified or percolating clusters exhibit local fluctua-
tions, the framework of the theory of random percolation with its critical exponents and
fractal dimension is still not enough to describe the disorder and the optical properties. We
propose an alternative concept: the configuration entropy, that we compare to the
multifractal analysis on computer simulated morphologies. At the percolation threshold,
the entropy undergoes a maximum and its optimum length a minimum. In contrast with the
multifractal analysis, the configuration entropy gives unambiguous results, relatively
independent of the finite size of the image.

1. Introduction

A large variety of characteristic parameters can be extracted from a given


disordered morphology. The question of the relevance of these parameters
essentially depends on the properties which will be measured on this morphology.
The fractal dimension, based on geometrical concepts is very efficient for
describing the d.c. conductivity of purely random systems composed of metallic
clusters in an insulating matrix [l]. Moreover, the relation to percolation theory
and scaling laws is direct. The multifractal spectrum was then successfully
introduced to account for the l/f noise in such systems presenting local
fluctuations of disorder [2].
A unique characteristic length appears in these approaches: the correlation
length 5, limit size of the scaling behaviour in the medium. More recently, the
optical absorption of granular metal films was shown to undergo a broad
maximum around the percolation threshold [3]. Classical Effective Medium
Theories cannot account for this extra absorption. Two models were proposed
[4,5], both calculating the local optical conductivity on the ramified percolating
medium at a size depending on the frequency of the optical problem. This local

037%4371/94/$07.00 0 1994 Elsevier Science B.V. All rights reserved


SSDI 0378.4371(93)E0548-S
C. Andraud et al. I Physica A 207 (1994) 208-212 209

result is then integrated up to size 5 by assuming the disorder to follow scaling


laws depending on the universal critical exponents of percolation and thus the
fractal dimension [6]. These models fairly well describe the optical properties of
random media near the percolation. We have however examples of granular metal
layers deposited under different conditions (rate of deposition, temperature of the
substrate, etc.) which exhibit notably different optical properties, although having
same metal fraction, the same fractal dimension and a flat multifractal spectrum.
These parameters, metal fraction, fractal dimension (and doubtless multifractal
spectrum) are thus not enough to describe the optical properties. This is the
reason why we have explored another way of description of the disorder related
to thermodynamical concepts and theory of information: the configuration
entropy.

2. Configuration entropy

Shannon defined the entropy [7] of a system as a function of the probability of


occurrence of different states of the system. If a system has n different possible
states with probability of occurrence pi, i = 1,2, . . . , n, then the gain in in-
formation from the occurrence of the event i is defined as
z, = log(l/pj) . (1)

The expected value of such a gain in information is defined as the entropy H of


the system,

H = i piZi or H = - i pi log p, . (2)


1=1 i=l

In the present study the system is a binary image representing a two-phase


material and the elementary source symbol is the pixel intensity which has two
possible values 1 or 0 (black and white representation).
The image is analysed through an elementary sliding cell of size 1 x 1. For a
given position of the cell C,, we count the number of active pixels (pixels having
value 1) contained in Cij. Once the whole image is examined we compute Nk( l),
the number of cells containing k active pixels, and N(Z), the total number of
visited cells. The probability associated with the state of the cell is thus defined as

P,(l) = ~/#)~W) 2 (3)


and the corresponding entropy is

(4)
Obviously, H is a measure of the uncertainty degree [8] in the realisation of a
210 C. Andraud et al. I Physica A 207 (1994) 208-212

given state of the cell of size I x 1. Using the Lagrange multiplier method one can
easily show that H(I) is maximised under the constraint

when the different cell configurations are equally probable, i.e., when

p,(Z) = l/(1 + 1) for every k. (6)


To examine the image structure at different length scales, the cell size is increased
from I = 1 to I = L/2. To compare the entropy values obtained for different 1
values one has to normalise H(1). We thus compute for each 1 value the
normalised entropy defined by

H*(l) = H(l)l&,,(Q > (7)

H,,,(Z) = log@* + 1) . (8)

H*(Z) undergoes a maximum at a characteristic size 1, that we call the entropy


optimum length [9]. The box of size 1, X I,, presents a maximum of disorder and is
the optimum area at which the image must be observed and eventually character-
ised with the aim of optical studies. We have also proved the configuration
entropy increases with the cluster ramification and the system lacunarity.

3. Application to randpm morphologies

Our method is applied to simulated images of size 400 x 400 pixels. These
images are composed of penetrating disks of the same size, 11 pixels (see Fig. l),
obtained by random trial. The black pixel fraction is increased from 0.10 to 0.95
with a step 0.05. The percolation occurs at p, = 0.62. For a given black pixel
fraction (p = 0.20, see Fig. 1) the configuration entropy undergoes as expected a
maximum value (Hz,, = 0.867) for an optimum length (I, = 33 pixels) of the
analysis block size. Fig. 2 displays the variation of the configuration entropy
maximum and the optimum entropy length versus the black pixel fraction. They
both undergo an extremum (max of Hz,, and min of I,) around the percolation
critical value p,. This result is coherent with the evolution of the morphology
which becomes more and more ramified and lacunar up to p,, where the
maximum of disorder is expected to be reached. Around pC, the morphology
needs thus to be described down to a very small size (I, = 12 pixels) roughly
corresponding to the size of its elementary components (diameter of the disks is
11 pixels). In fact the minimum of I, is a plateau roughly centered at p, = 0.60,
whereas the maximum of HzaX is reached at a lower value, p = 0.53. This
discrepancy could be due to the fact that we did not take into account enough
C. Andraud et al. I Physica A 207 (1994) 208-212 211

50 100
0.2 0.4 0.6 0.8
ANALYSIS BLOCK SIZE (in pixels) CONCENTRATION

Fig. 1. Fig 2.
Fig. 1. Computer simulated random morphology (400 x 400 pixels) obtained by random throw of
overlapping disks of size 11 pixels, and corresponding entropy. Black phase surface fraction: 0.20,
H max
* = 0.866, lo = 33 pixels.

Fig. 2. Evolution of the optimum length (in pixels) and of the configuration entropy maximum with
the black phase surface fraction. The entropy maximum is 0.985 at p = 0.53 and the optimum length
minimum is 12 pixels centered at p = 0.60, p, (percolation) is 0.62.

constraints in the definition of the normalisation entropy H,,,(I). We are


improving this point.
We have shown elsewhere [9] that the configuration entropy is much more
sensitive to small variations in the disorder than other characteristic parameters of
the disorder deduced from fractal and multifractal analysis. Moreover Hi,, and I,
are much less sensitive to the finite size of the images. The configuration entropy
is thus a new sensitive and efficient tool for describing the disorder. It has now to
be systematically related to the optical properties of disordered morphologies and
eventually be employed in analytical models of the effective dielectric function of
these media.

References

[l] A. Kapitulnik and G. Deutscher, Phys. Rev. Lett. 49 (1982) 1444.


[2] R. Rammal, C. Tannous, P. Breton and A.N.S. Tremblay, Phys. Rev. Lett. 54 (1985) 1718.
[3] P. Gadenne, A. Beghdadi and J. Lafait, Opt. Commun. 65 (1988) 17.
[4] Y. Yagil, M. Yosefin, D.J. Bergmann, G. Deutscher and P. Gadenne, Phys. Rev. B 43 (1991)
1342.
[5] T. Robin and B. Souillard, Physica A 193 (1993) 79.
[6] D. Stauffer, Introduction to Percolation Theory (Taylor and Francis, London, 1985).
212 C. Andraud et al. I Physica A 207 (1994) 208-212

[7] C.E. Shannon, Bell. Syst. Tech. J. 210 (1948) 623.


[S] L. Brillouin, Sciences and Information Theory (Academic Press, New York, 1956).
[9] A. Beghdadi, C. Andraud, J. Lafait, J. Peiro and M. Perreau, Entropic and multifractal analysis
of disorder morphologies, Fractals 1, No. 3 (1993), to appear.

You might also like