You are on page 1of 11

CHERD-1558; No.

of Pages 11
ARTICLE IN PRESS
chemical engineering research and design x x x ( 2 0 1 4 ) xxxxxx

Contents lists available at ScienceDirect

Chemical Engineering Research and Design

journal homepage: www.elsevier.com/locate/cherd

A probabilistic analysis of some selected mixing


indices

Zongyu Gu , J.J.J. Chen


Department of Chemical and Materials Engineering, The University of Auckland, Private Bag 92019, Auckland 1142,
New Zealand

a b s t r a c t

There appears to be some confusion in the literature concerning the state of a granular mixture at the end of an
ideal mixing process, and this may have inuenced the basis on which some mixing indices were formulated. As an
attempt to provide clarication, we quantitatively established the three characteristic states of a granular mixture
the completely unmixed, the perfectly ordered, and the randomly mixed states by considering the relevant
probability distributions. We demonstrated how one may derive the expected values of various mixing indices based
on these probability distributions, and veried the results with a simple mixing simulation. Asymptotic arguments
showed the randomly mixed state, not the perfectly ordered state, best describes a well-mixed granular mixture,
while the two states become increasingly indistinguishable as the number of particles per sample increases. We
revealed that unlike the Lacey index, which is normalised based on the randomly mixed state, the normalisation of
the total entropy of mixing uses the perfectly ordered state, making it less than unity at the end of an ideal mixing
operation this has not been sufciently highlighted by the authors proposing such indices. We found it interesting
to analyse the Gini coefcient, a widely accepted measure of income inequality in economics, in the context of
granular mixing. While not advocating its use, we noted that similar to the other indices, its expected values for the
characteristic states are dependent on the choice of the basis of normalisation.
2014 The Institution of Chemical Engineers. Published by Elsevier B.V. All rights reserved.

Keywords: Granular mixing; Mixing index; Lacey index; Entropy of mixing; Gini coefcient

1. Introduction how these indices compare to each other in the context of a


mixing process.
In the study of granular mixing, it is essential to describe In relation to this incomplete understanding of the differ-
the state of a granular mixture as to how homogeneously ent mixing indices, there also appears to be some degree of
the material is mixed using some quantitative measure. confusion in the literature as to what state one should expect a
However, the mixedness of a granular material does not mixture to reach at the end of an ideal mixing process. Thiel
necessarily possess an obvious quantitative sense; conse- et al. (1981) have pointed out that the outcome depends on the
quently, constructing numeric measures to capture the notion properties of the granular mixture: for non-interacting parti-
of mixedness is a process that depends on individual percep- cles that are only distinguishable by colour, mixing proceeds
tions, and involves much freedom and uncertainty. Indeed, via a purely statistical process and complete randomisation
despite the large number of mixing indices found in the liter- can be achieved; for non-identical or interactive particles,
ature, a universally accepted method of evaluating mixedness segregation or the build-up of ordered structure makes the
has not yet been established. The authors who proposed nal state of the mixture less randomised or more ordered,
these indices tend to justify the appropriateness of their own respectively. However, it seems evident that, when formulat-
denitions by arguing for the corresponding physical or math- ing certain mixing indices, some authors tend to neglect the
ematical bases; nevertheless, it is not completely clear as to subtle differences among these states that may be thought


Corresponding author. Tel.: +64 226357788/99238137.
E-mail addresses: zgu231@aucklanduni.ac.nz (Z. Gu), j.chen@auckland.ac.nz (J.J.J. Chen).
http://dx.doi.org/10.1016/j.cherd.2014.04.014
0263-8762/ 2014 The Institution of Chemical Engineers. Published by Elsevier B.V. All rights reserved.

Please cite this article in press as: Gu, Z., Chen, J.J.J., A probabilistic analysis of some selected mixing indices. Chem. Eng. Res. Des. (2014),
http://dx.doi.org/10.1016/j.cherd.2014.04.014
CHERD-1558; No. of Pages 11
ARTICLE IN PRESS
2 chemical engineering research and design x x x ( 2 0 1 4 ) xxxxxx

examine their properties as novel measures of mixedness in


Nomenclature the context of granular mixing.

b number of cells
G Gini coefcient of income inequality in a popu-
lation; Gini coefcient of mixing for a granular
mixture 2. Characteristic states of a granular
H Shannon entropy of a probability distribution mixture
(nat)
M Lacey index 2.1. Representation and description of the
n number of particles in a random sample; characteristic states
number of outcomes in a discrete probability
distribution; number of individuals in a popu- The earliest discussion of the states of a granular mixture can
lation be attributed to Lacey (1943), where he illustrated the likely
N total number of particles in a mixture arrangements of particles at the end of an ideal mixing process
p probabilities in a discrete probability distribu- using schematics as shown in Fig. 1. Lacey argued that after
tion; fractions of a certain type of particles in a thorough mixing, Fig. 1(a) would correspond to what one may
cell hope to obtain, while Fig. 1(b) would be the best one is likely
P overall proportion of black particles in the mix- to achieve. Because these statements use vague expressions
ture such as hope to and likely, questions remain as to, rstly,
s2 variance of sample proportions how well the completely homogeneous state in Fig. 1(a) rep-
Si local entropy of mixing of cell i (nat) resents an actual well-mixed granular mixture, and secondly,
Stot total entropy of mixing (nat) whether it is of any importance to consciously differentiate
XCC cell count random variable between the two states in Fig. 1 when examining a granular
XSP sample proportion random variable mixture.
y income of an individual in a population (dol- We will address these questions by quantitatively estab-
lars) lishing the three characteristic states of a granular mixture. In
y average level of income in a population (dollars) order to facilitate visualisation, we shall represent particles in
a granular mixture as sites on a regular 2D rectangular lattice.
Greek symbols For simplicity, we will consider a binary mixture of black and
 normalised Gini coefcient of mixing based white particles, so that each lattice site has a binary state for
on the perfectly ordered state colour for the particle at that site.
 normalised Gini coefcient of mixing based There are various sampling methods for characterising
on the randomly mixed state the state of a given mixture. One common practice involves
tot normalised total entropy of mixing based on repeatedly taking random samples of a xed size and shape
the perfectly ordered state from the mixture with replacement, and recording the pro-

tot normalised total entropy of mixing based on portion of black particles within each sample; given that a
the randomly mixed state sufcient number of samples are taken, the distribution of
these sample proportions (denoted using the random variable
Subscripts XSP ) will characterise the state of the mixture. Alternatively,
0 expected value of the mixing index for the com- one can divide the entire domain of the mixture into a nite
pletely unmixed state number of xed cells, with the number of black particles in
p expected value of the mixing index for the per- each cell counted to give cell counts (denoted using the ran-
fectly ordered state dom variable XCC ); the state of the mixture can hence also
r expected value of the mixing index for the ran- be described by the distribution of these cell counts this
domly mixed state approach is particularly straightforward if the cells all con-
tain an equal number of particles. Other techniques, though
not considered in this work, include, for instance, counting
the contact points between different types of particles (Akao
of as ideal outcomes of mixing the implication is that the
et al., 1976).
resulting indices may be susceptible to incorrect interpreta-
In Sections 2.22.4, we will discuss the three character-
tion if one does not clearly understand the bases of their
istic states of a granular mixture and recapitulate existing
denitions. Therefore, we nd it important to differentiate the
understanding of these states in terms of the probability dis-
characteristic states of a granular mixture in quantitative and
tributions of sample proportions and cell counts.
probabilistic terms, and link such differences among the states
to the differences among the formulations of mixing indices
found in the literature.
It is also interesting to note that the concept of mixedness
has also been applied extensively in disciplines other than 2.2. The completely unmixed state
chemical engineering. Mixedness is frequently referred to by
alternative terminologies, such as species diversity in ecol- Prior to mixing, we normally expect the collection of particles
ogy (Lloyd et al., 1968), income inequality in economics (Sen, to be in a highly segregated state. Many possible congura-
1973), and racial segregation in sociology (Theil and Finizza, tions, such as those shown in Fig. 2, may be used to exemplify
1971), just to name a few. It is thus possible and of interest such a state, where particles of each colour cluster in a partic-
to adopt the mixedness concept in these other sciences and ular region of the lattice.

Please cite this article in press as: Gu, Z., Chen, J.J.J., A probabilistic analysis of some selected mixing indices. Chem. Eng. Res. Des. (2014),
http://dx.doi.org/10.1016/j.cherd.2014.04.014
CHERD-1558; No. of Pages 11
ARTICLE IN PRESS
chemical engineering research and design x x x ( 2 0 1 4 ) xxxxxx 3

Fig. 1 Pictorial representations of the likely arrangements of particles in a granular mixture at the end of an ideal mixing
operation used by Lacey (1943).

Fig. 2 Instances of the completely unmixed state represented on a 2D lattice.

We may dene the completely unmixed state using the ide- which says that the sample proportion is always equal to the
alised case where a binary outcome can be associated with XSP , overall proportion of black particles in the mixture; that is to
which leads to a Bernoulli distribution for XSP , given by say, the outcome of random sampling is deterministic. We
can similarly expect the random variable for the cell counts
Pr(XSP = x) = Px (1 P)
1x
for x = 0, 1 (1) to follow
 NP

where P is the overall proportion of black particles in the mix- Pr XCC = =1 (4)
b
ture. According to this distribution, any sample taken from the
mixture will either have a sample proportion of zero, when all As Thiel et al. (1981) have noted, the perfectly ordered state is
the particles in the sample are white, or unity, when all the only attainable in very special cases, and for the mixing of non-
particles in the sample are black, while the relative likelihood interacting particles that are only distinguishable by colour,
of either outcome is proportional to the overall proportion of
the respective type of particles.
By similar arguments, one can show that the distribution of
cell counts for a completely unmixed material should approx-
imately follow

N
Pr(XCC = x) = Pxb/N (1 P)
1xb/N
for x = 0, (2)
b

where N is the total number of particles in the mixture, and b


is the number of cells.

2.3. The perfectly ordered state

The perfectly ordered state of a granular mixture is shown in


Fig. 3, or equally well by Fig. 1(a) as was used by Lacey (1943).
We may dene the perfectly ordered state using the prob-
ability statement

Fig. 3 The perfectly ordered state represented on a 2D


Pr(XSP = P) = 1 (3) lattice.

Please cite this article in press as: Gu, Z., Chen, J.J.J., A probabilistic analysis of some selected mixing indices. Chem. Eng. Res. Des. (2014),
http://dx.doi.org/10.1016/j.cherd.2014.04.014
CHERD-1558; No. of Pages 11
ARTICLE IN PRESS
4 chemical engineering research and design x x x ( 2 0 1 4 ) xxxxxx

Fig. 4 Instances of the randomly mixed state represented on a 2D lattice.

the perfectly ordered state does not represent the nal state continuously measure sample proportions or cell counts and
of the mixture at the steady state. examine their distributions during mixing to track changes in
the mixedness of a granular material. It is often more conve-
2.4. The randomly mixed state nient, however, to adopt a certain numerical property of the
probability distributions as an effective measure of mixedness
Non-interacting particles that are identical except colour at rather than considering the specic distributions at all times.
the end of an ideal mixing operation will reach the randomly This leads to the creation of many of the mixing indices avail-
mixed state (Thiel et al., 1981). On a 2D lattice, this state is able in the literature: states of a granular mixture condense
represented by Figs. 4 and 1(b), the latter used by Lacey (1943). into probability distributions by the process of sampling, and
These randomly mixed states can be generated by completely the probability distributions further condense into mixing
randomising the arrangement of black and white particles on indices upon taking certain properties of the distributions.
the lattice, where the number of each type is set by a pre- Note how this process represents a cascade of information
scribed value of the overall proportion of black particles in the reduction, from the exact arrangement of particles in a gran-
mixture. ular mixture, eventually to the value of the mixing index.
Unlike the perfectly ordered state, random samples taken In this section, we will demonstrate the principle and the
from a granular mixture in the randomly mixed state will methods of forming measures of mixedness from appropriate
likely contain a different number of black particles each. Given probability distributions by considering both existing and new
a sample size of n, the distribution of sample proportions is mixing indices. The aim is to link our established understand-
approximated by ing of the states of a granular mixture to the expected values
  and other distinctive features of various mixing indices.
n n(1x) 1 2
Pr(XSP = x) = Pnx (1 P) for x = 0, , , . . ., 1 (5)
nx n n
3.1. Simulation of mixing
 
a Prior to presenting our analysis of the different mixing indices,
where denotes a binomial coefcient. The mean and
b it is useful to consider how we might verify the results we are
variance of this distribution can be shown to be P and [P(1 about to derive for these indices. One way is to simulate a sim-
P)/n], respectively (Lacey, 1943, 1954). Note that the variance of ple granular mixing process and compute the mixing indices
the distribution of sample proportions for the randomly mixed as mixing takes place it is convenient to carry out such a
state asymptotically approaches zero in the limit of n : mixing process directly on the lattice that we have been using
this implies that the two states become increasingly indistin- throughout this paper for the illustration of ideas. As to how
guishable as the number of particles per sample increases. we will induce mixing on the lattice, note that we are not too
However, for a granular material, unlike a uid, there are only concerned with the mixing method, namely, the set of rules by
a nite number of particles in any sample taken; thus, the per- which the mixture gets mixed progressively. This is because
fectly ordered state and the randomly mixed state are indeed the characteristic states of primary interest are all close to the
two dissimilar states. beginning or the end of the mixing operation, largely inde-
Finally, the probability mass function of XCC for the ran- pendent of the path of changes that the mixture follows. Thus,
domly mixed state can be written as we simply need a random process that continuously increases
  the mixedness of particles on the lattice many methods are
N/b N/bx N possible (e.g., Too et al., 1979), but we will hereinafter limit
Pr(XCC = x) = Px (1 P) for x = 0, 1, 2, . . ., (6)
x b ourselves to one method that is adequate for the purpose of
generating the mixing data.
whose mean and variance are (NP/b) and [NP(1 P)/b], respec- Black and white particles were initially arranged on a 100-
tively. by-100 square lattice as shown in Fig. 5(a), representing an
instance of the completely unmixed state this state corre-
3. Measures of mixedness sponded to step 1 of the mixing process. The overall proportion
of black particles was chosen to be P = 0.25. An algorithm was
We have demonstrated how probability distributions can be then applied recursively to the lattice to cause mixing in dis-
used to characterise states of a granular mixture with differ- crete steps, each of which corresponded to a number of similar
ent degrees of mixedness; using such relationships, one may actions, where a single action randomised the distribution of

Please cite this article in press as: Gu, Z., Chen, J.J.J., A probabilistic analysis of some selected mixing indices. Chem. Eng. Res. Des. (2014),
http://dx.doi.org/10.1016/j.cherd.2014.04.014
CHERD-1558; No. of Pages 11
ARTICLE IN PRESS
chemical engineering research and design x x x ( 2 0 1 4 ) xxxxxx 5

Fig. 5 Arrangement of particles at selected steps from the simulated mixing process.

particles inside a 4-by-4 square region. A single step of mix- probability distributions. For the completely unmixed state,
ing then required this mixing action to be repeated until the the distribution of sample proportions, XSP , is given by Eq. (1);
region had assumed all 952 possible locations it can occupy thus, the expected value of the variance of sample proportions
in the mixture exactly once, following a random sequence of for the completely unmixed state is
these locations. A total of 300 of these steps of mixing were
performed. Fig. 5 shows the arrangement of particles on the 2 2
s20 = E(XSP
2
) [E(XSP )] = (0 + P) (0 + P) = P(1 P) (7)
lattice at selected steps.
Several mixing indices were calculated at every step of
The distribution XSP for the randomly mixed state is given by
mixing from the distributions of sample proportions and cell
Eq. (5); we have noted that the variance of this distribution is
counts. This requires either obtaining random samples from
equal to
the mixture or dividing the domain into cells. A total of 5000
square samples, 5-by-5 in size (hence n = 25), were taken at
P(1 P)
random locations within the mixture at every step, with the s2r = (8)
n
proportions of black particles in the samples noted; mean-
while, the domain was also divided into 400 xed square cells Eqs. (7) and (8) simply reproduced the results of Lacey (1954).
(as indicated by the regions enclosed by the lines in Fig. 5(g)), For the perfectly ordered state as described by Eq. (3), the vari-
also 5-by-5 in size, with the count of black particles in the cells ance of sample proportions for the perfectly ordered state is
noted. The results of the calculated mixing indices are to be given by the trivial result
presented in Sections 3.23.4.
It should be noted that when one takes random samples, s2p = 0 (9)
the mixing indices calculated for even the same state of the
mixture would involve variability as a result of sampling. Since
We will now consider the well-known Lacey index (Lacey,
our work focuses on the expected values of the mixing indices,
1954), dened as
as many as 5000 random samples were taken at every step
of mixing to calculate the mixing indices, in order to min-
s20 s2
imise any statistical noise due to sampling. Taking such a large M= (10)
number of samples is rarely feasible in practice, so it would s20 s2r
be plausible to appeal to statistical inference in practical sit-
uations when one calculates mixing indices from a limited Since s20 and s2r are xed with given P and n, it is useful to think
number of samples. of the Lacey index as a normalised version of the variance
of sample proportion. Because the normalisation is based on
the completely unmixed and the randomly mixed states, the
3.2. Variance of sample proportions; the Lacey index expected values of the Lacey index for these two states are
M0 = 0 and Mr = 1, respectively. As the mixture evolves from
A large portion of the mixing indices available, especially those the completely unmixed state to the randomly mixed state
found in the early literature (e.g., Lacey, 1943, 1954; Beaudry, during a typical mixing process, the variance correspondingly
1948; Blumberg and Maritz, 1953), are based on or related to decreases from s20 and converges to s2r , while the Lacey index
the variance of sample proportions, commonly denoted as s2 . increases from zero to unity.
The expected values of the variance of sample proportions It is important to note that the above process does not
for the characteristic states can be derived from appropriate involve the perfectly ordered state. The expected value of the

Please cite this article in press as: Gu, Z., Chen, J.J.J., A probabilistic analysis of some selected mixing indices. Chem. Eng. Res. Des. (2014),
http://dx.doi.org/10.1016/j.cherd.2014.04.014
CHERD-1558; No. of Pages 11
ARTICLE IN PRESS
6 chemical engineering research and design x x x ( 2 0 1 4 ) xxxxxx

Fig. 7 Local entropy as a function of the fraction of black


particles in the cell for a binary mixture.

as a general measure of dividedness with applications in many


different elds (Lloyd et al., 1968; Theil and Finizza, 1971; Theil,
1972; Sen, 1973; Ogawa, 2006).
Fig. 6 The Lacey index from the simulated mixing
The possibility of applying information theory in granu-
process.
lar mixing research was probably rst pointed out by Inoue
Lacey index for this state may be evaluated by substituting Eq. and Yamaguchi (1968); their theory was then extended to a
(9) into Eq. (10), and using Eqs. (7) and (8), we have more generally applicable one by Ogawa and Ito (1973). Many
other authors (Lai and Fan, 1975; Schutyser et al., 2001; Finnie
s20 0 P(1 P) 0 n et al., 2005; Alemaskin et al., 2005; Masiuk and Rakoczy, 2006;
Mp = = = (11) Guida et al., 2010) have since proposed slightly different de-
s20 s2r P(1 P) P(1 P)/n n1
nitions of entropy of mixing. The majority of these measures
The signicance of Eq. (11) lies in the fact that Mp is identically are calculated from cell counts.
greater than unity for all n > 1. As we normally work within a Because these authors have adopted different notations
range between zero and unity, the normalisation of the Lacey and terminologies, we will now reformulate an entropy-based
index has clearly excluded the perfectly ordered state outside mixing index that encompasses features common to most of
of this range. Note that this distinction between the randomly these treatments. Consider a binary mixture containing N par-
mixed state and the perfectly ordered state can only be made ticles where the overall proportion of black particles is given
when the sample size is nite because s2r s2p and Mp Mr by P; we divide the domain of the mixture into a total of b
as n . equally sized cells, whose locations are xed. Let pi,black be the
Fig. 6 shows the changes in the Lacey index from the simu- fraction of black particles in cell i, then pi,white = 1 pi,black is
lated mixing process, where the expected values of the index that of white particles. Noting the denition of the Shannon
for the three characteristic states are marked with horizontal entropy given by Eq. (12), we dene the local entropy of mixing
lines. The normalisation of the Lacey index is such that dur- of cell i as
ing mixing, it takes values roughly within the range of zero to 
unity, which corresponds to the expected values of the index Si = pi,j ln pi,j = pi ln pi (1 pi ) ln(1 pi ) (13)
for the completely unmixed and the randomly mixed states j {black,white}
respectively. Given that n = 25, the expected value of the Lacey
index for the perfectly ordered state can be found from Eq. (11) where the index i = 1, 2, . . ., b, and pi have replaced pi,black for
to be 25/(25 1) 1.042 this is not reached at the steady state simplicity of notation. The total entropy of mixing of the entire
of mixing. mixture is hence dened as the arithmetic mean of the local
entropies, namely
3.3. Entropy of mixing

1
b
We will rst state the mathematical denition of the Shan-
Stot = Si (14)
non entropy in information theory (Shannon, 1948): consider a b
i=1
discrete probability distribution with n outcomes, correspond-
ing to the events E1 , . . ., En with probabilities p1 , . . ., pn ; given
It is apparent from Eq. (13) that, for a binary mixture, the
exactly one of the n events is bound to occur, the Shannon
local entropy is simply a function of the fraction of black parti-
entropy of the distribution, H, is dened as
cles in the cell. This functional relationship is shown in Fig. 7,
and is a frequently quoted result in information theory. We

n
H= pi ln pi (12) observe that the local entropy is maximised when pi = 0.5 in
i=1 a cell, which corresponds to a local entropy of ln 2.
As was done for the variance, we will derive the expected
When the natural logarithm is used in the denition in Eq. (12), values of the total entropy for the characteristic states. For the
the Shannon entropy has the unit nat. The Shannon entropy completely unmixed state, the distribution of cell counts, XCC ,
has a number of interesting properties and can be interpreted given by Eq. (2), suggests that the fraction of black particles in

Please cite this article in press as: Gu, Z., Chen, J.J.J., A probabilistic analysis of some selected mixing indices. Chem. Eng. Res. Des. (2014),
http://dx.doi.org/10.1016/j.cherd.2014.04.014
CHERD-1558; No. of Pages 11
ARTICLE IN PRESS
chemical engineering research and design x x x ( 2 0 1 4 ) xxxxxx 7

Fig. 8 Total entropy of mixing for the randomly mixed


state (Stot,r ) as a function of (N/b) and P for a priori
predictions as shown by the solid curves, compared with
Fig. 9 The normalised total entropy of mixing in Eq. (18)
expected values for the perfectly ordered state (Stot,r ) as
from the simulated mixing process.
shown by the dotted lines.

study of uid mixing as a result of the large number of par-


any cell is either zero or unity, and hence the local entropy of
ticles per cell; however, for granular materials, the expected
any cell must be zero, which then gives
values for the two states can be signicantly different with
nite cell sizes, and as a result, the precise value of Stot,r has
Stot,0 = 0 (15)
to be computed from Eq. (17) or obtained from plots such as
Fig. 8.
For the perfectly ordered state, Eq. (4) indicates that all the
Even with devices such as Fig. 8, it is still relatively hard
local entropies, and hence also the total entropy, must be equal
to compute the total entropy for the randomly mixed state,
to
Stot,r , not to mention that the case is further complicated if
the cell sizes are variable; in contrast, the total entropy for the
Stot,p = P ln P (1 P) ln(1 P) (16) perfectly ordered state, Stot,p , is very easy to calculate from Eq.
(16) without such limitations. Therefore, when it comes to nor-
Since Eq. (16) is similar in form to Eq. (13), note the convenient malising the total entropy of mixing, it is not surprising that all
fact that we can obtain values of Stot,p using Fig. 7 by replacing the authors have chosen Stot,p , rather than Stot,r as the basis of
pi with P and Si with Stot,p . Next, using Eq. (6), the entropy of normalisation for their entropy-based indices, expressing the
the randomly mixed state can be expressed as a discrete sum normalised total entropy in a form similar to
of weighted local entropies, which writes
Stot,0 Stot 0 Stot Stot
N/b  
tot = = = (18)
 N/b N/bx xb Stot,0 Stot,p 0 Stot,p Stot,p
Stot,r = Px (1 P)
x N
x=0 where Stot,p is dependent on P according to Eq. (16). By com-
xb

xb
  xb
 paring Eqs. (10) and (18), it is clear that the total entropy is
ln + 1 ln 1 (17) normalised differently from the Lacey index as the former uses
N N N
Stot,0 and Stot,p as its bases of normalisation whereas the lat-
The expression for Stot,r in Eq. (17) does not readily simplify ter uses s20 and s2r instead. This means that the expected value
into a more condensed form. It is still possible to correlate of the normalised total entropy for the completely unmixed
Stot,r with the two independent variables in Eq. (17), namely state and the perfectly ordered state are tot,0 = 0 and tot,p =
(N/b) and P, for making a priori predictions of the total entropy 1, respectively, and thus tot,r is identically less than unity
for the randomly mixed state. This correlation for Stot,r as for all nite values of (N/b); that is to say, when a granular
described by Eq. (17) is represented graphically using the black mixture reaches the randomly mixed state at the end of an
solid lines in Fig. 8; we have also plotted the value of Stot,p for ideal mixing operation, tot will remain less than unity. One
the perfectly ordered state as red dashed horizontal lines. It is will nd many cases in the literature where the normalised
also worth pointing out that the trend for any P > 0.5 is iden- entropy-based measures, proposed by different authors and
tical with that for (1 P). As P tends towards 0.5 from either are hence slightly different in form, do not approach unity
side, Stot,r increases at any particular (N/b) > 1; Stot,p increases (e.g., Schutyser et al., 2001; Finnie et al., 2005; Guida et al.,
regardless of (N/b), reaching a maximum of ln 2 0.6931 when 2010). This is illustrated in Fig. 9 for based on the normalised
P = 0.5, which is in accordance with Fig. 7. total entropy calculated from our simulation.
We observe that as (N/b) , Stot,r Stot,p for any given Little attention has been drawn to this difference in the
value of P, again demonstrating the asymptotic equivalence of bases of normalisation of the different indices. Authors of
the randomly mixed state and the perfectly ordered state in these entropy-based indices have not offered a reason for
the limit of very large number of particles per cell. This implies this different normalisation when presenting their denitions,
that Stot,p may be an acceptable approximation of Stot,r in the nor attempted to discriminate between the perfectly ordered

Please cite this article in press as: Gu, Z., Chen, J.J.J., A probabilistic analysis of some selected mixing indices. Chem. Eng. Res. Des. (2014),
http://dx.doi.org/10.1016/j.cherd.2014.04.014
CHERD-1558; No. of Pages 11
ARTICLE IN PRESS
8 chemical engineering research and design x x x ( 2 0 1 4 ) xxxxxx

Fig. 11 The Lorenz curve construction for computing the


Gini coefcient (Sen, 1973).

where n is the number of people in the population consid-


ered, yi is the income of person i, and y is the average level of
income. This algebraic denition of the Gini coefcient shows
Fig. 10 The normalised total entropy of mixing in Eq. (19)
that the measure accounts for differences between all pairs of
from the simulated mixing process.
observed incomes.
Another equivalent denition of the Gini coefcient can
state and the randomly mixed state for granular materials.
be perceived graphically using the Lorenz curve (Sen, 1973):
Schutyser et al. (2001) did report that tot did not approach
for a particular poorest fraction x of the population, we can
unity if (N/b) was too small, but did not offer a rationalisation
nd the cumulated income earned by this bottom fraction of
as we have done with Fig. 8. Some (e.g., Finnie et al., 2005) even
the population, which is a fraction y of the total income of
mistakenly referred to Stot,p as the value of the total entropy
the entire population. Plotting ordered pairs (x, y) gives the
at the steady state of mixing.
Lorenz curve in Fig. 11. The Lorenz curve always lies below
It is necessary to point out that there is nothing fundamen-
the y = x line (termed the line of absolute equality) unless all
tally wrong with normalising Stot using Stot,p , and we are not
incomes are exactly equal. The Gini coefcient is dened as
implying that Stot,r is a superior basis of normalisation. If Stot,r
the ratio of the area between y = x and the Lorenz curve, to
was used for normalisation, we get
the right-angled triangular region beneath y = x. By geometric
Stot,0 Stot 0 Stot Stot arguments, it is clear that the Gini coefcient is always in the

tot = = = (19)
Stot,0 Stot,r 0 Stot,r Stot,r range 0 G 1.
The Gini coefcient is rarely applied in chemical engineer-
instead, which is an alternative denition of the normalised ing. We found an instance where the Gini coefcient was used
total entropy. The changes in tot from the simulated mixing as a measure of performance of stirrers (Trivellato, 2011), but
process are shown in Fig. 10. We would simply like to remind have not seen any analysis on its possible use in the study of
the readers of the danger of not noting the different bases of granular mixing. Thus, it is of interest to dene the Gini coef-
normalisations used for various indices, or more fundamen- cient in the context of granular mixing so that it can be used
tally, the different natures of the various states of a granular as a measure of mixedness.
mixture, especially when it comes to interpreting the values One possible denition uses the distribution of cell counts,
of these indices. Those who use entropy-based indices nor- as in the case of the entropy of mixing in Section 3.3. We draw
malised in a form similar to that in Eq. (18), for example, an analogy between a binary granular mixture and the popu-
should be informed that their expected values for the ran- lations in a society by treating each cell as a person possessing
domly mixed state at the end of an ideal mixing process could an income that is equal to the number of black particles in the
be below unity. cell; hence, the inequality among the incomes of the cells
is related to the degree of mixedness of the material, which
3.4. The Gini coefcient of mixing can hence be characterised using the Gini coefcient. Substi-
tuting relevant parameters in our analogy into Eq. (20) yields
The Gini coefcient is a well-established measure of income our denition of the Gini coefcient of mixing
inequality within a population in the eld of economics. The
intention of including the Gini coefcient here is to show how
1  b b
analyses in other disciplines have parallels in chemical engi- G= |XCC,i XCC,j |
neering, and the authors would like to note here that we are 2b (NP/b)
2
i=1 j=1
not necessarily advocating its use. Algebraically, the Gini coef-
1 
b b
cient can be dened as a half of the relative mean absolute
= |XCC,i XCC,j | (21)
difference between all pairs of incomes (Sen, 1973), or in equa- 2bNP
i=1 j=1
tion form:

1 
n n
where XCC,i and XCC,j are independent random variables hav-
G= yi yj (20) ing a probability distribution identical to XCC . Note we can also
2n2 y
i=1 j=1 calculate the Gini coefcient of mixing using the Lorenz curve,

Please cite this article in press as: Gu, Z., Chen, J.J.J., A probabilistic analysis of some selected mixing indices. Chem. Eng. Res. Des. (2014),
http://dx.doi.org/10.1016/j.cherd.2014.04.014
CHERD-1558; No. of Pages 11
ARTICLE IN PRESS
chemical engineering research and design x x x ( 2 0 1 4 ) xxxxxx 9

for which we would need to plot the cumulated fraction of


black particles in cells against the fraction of cells.
We will now derive expressions for the expected values of
the Gini coefcient of mixing for the characteristic states. For
the perfectly ordered state, since all cell counts are equal as
suggested by Eq. (4), the expected value of the Gini coefcient
for the perfectly ordered state is simply

Gp = 0 (22)

which, in the language of economics, implies absolute equality


of the incomes of all cells.
For the completely unmixed state, the distribution of XCC
is given by Eq. (2). It can be shown (Gu and Chen, 2013) that
the expected value of the Gini coefcient of mixing for the
completely unmixed state is

G0 = 1 P (23)

For the randomly mixed state, we will approximate XCC using Fig. 12 The Gini coefcient of mixing from the simulated
a normal distribution where the mean and variance are iden- mixing process.
tical to those of the binomial distribution given by Eq. (6), i.e.,
XCC N(NP/b, NP(1 P)/b). Subsequently, we can derive (Gu and
Chen, 2013) the expected value of the Gini coefcient of mixing
for the randomly mixed state, yielding


b(1 P) In a typical granular mixing process where the binary mix-
Gr = (24)
NP ture changes state from completely unmixed to a randomly
mixed, the Gini coefcient of mixing will start from G0 , which
To summarise, approximate values of the Gini coefcient of is always below unity, and approach Gr , which is always above
mixing for the characteristic mixtures can be determined from zero, as long as the number of particles per cell is nite. For the
the analytical expressions given by Eqs. (22), (23), and (24). simulated mixing process, the changes in the Gini coefcient
Once again, note that as (N/b) , Gr Gp , implying that are shown in Fig. 12. Akin to how the normalised total entropy
the randomly mixed state asymptotically converges to the of mixing does not converge to unity at the steady state of mix-
perfectly ordered state in the limit of very large cells. ing, the Gini coefcient of mixing similarly does not converge
We will now consider the problem of normalisation of the to zero as such is its expected value for the perfectly ordered
Gini coefcient as we did for the previous indices. Interest- state.
ingly, the Gini coefcient is a naturally normalised index and One may make an attempt at normalising the Gini coef-
is bounded below and above by zero and unity as entailed by cient with different bases so that it behaves more similarly
its denition we have noted this fact when discussing the to the established mixing indices like the normalised total
Lorenz curve earlier in this section. entropy, whose value starts at zero for the completely

Table 1 Summary of some mixing indices considered in this work.


Denition Dened in . . . Expected value for the . . . Features of normalisation

Completely Randomly Perfectly


unmixed state mixed state ordered state
s20 s2 n
M= Eq. (10) 0 1 M = 0 when completely unmixed,
s20 s2r n1
M = 1 when randomly mixed
Stot,0 Stot Stot,r
tot = Eq. (18) 0 a
1 tot = 0 when completely
Stot,0 Stot,p Stot,p
unmixed,
tot = 1 when perfectly ordered
 Stot,0 Stot Stot,p 
tot = Eq. (19) 0 1 a
tot = 0 when completely
Stot,0 Stot,r Stot,r
unmixed,

tot = 1 when randomly mixed
G0 G
 =
G0 Gp
Eq. (25) 0 
1 1  = 0 when completely unmixed,
b  = 1 when perfectly ordered
NP(1 P)
  1
G0 G
 b
 = Eq. (26) 0 1 1   = 0 when completely unmixed,
G0 Gr NP(1 P)
  = 1 when randomly mixed

a
The values of Stot,r and Stot,p can be evaluated from Eqs. (17) and (16), respectively.

Please cite this article in press as: Gu, Z., Chen, J.J.J., A probabilistic analysis of some selected mixing indices. Chem. Eng. Res. Des. (2014),
http://dx.doi.org/10.1016/j.cherd.2014.04.014
CHERD-1558; No. of Pages 11
ARTICLE IN PRESS
10 chemical engineering research and design x x x ( 2 0 1 4 ) xxxxxx

unmixed state and increases with a higher degree of mixed- of the characteristic states used for the normalisation indices
ness. A possible normalisation of such kind is given by such as the Lacey index (M) are normalised based on the ran-
domly mixed state, while indices like the normalised total
G0 G G0 G G entropy of mixing (tot ) are normalised based on the per-
 = = =1 (25)
G0 Gp G0 0 G0 fectly ordered state. As a granular mixture tends towards the
randomly mixed state as the equilibrium mixture of an ideal
which is analogous to the normalisation of the total entropy
mixing process is achieved, the Lacey index will approach
of mixing in Eq. (18). Another possible normalisation is
unity while the normalised total entropy will stabilise at a
G0 G value below unity. Being aware of such features of the mix-
 = (26) ing indices when one uses them is an essential prerequisite
G0 Gr
for correctly interpreting them.
which makes the normalisation akin to that used for the Lacey We have examined the Gini coefcient from the study of
index found in Eq. (10). One can choose from either of the income inequality in economics, dened the Gini coefcient
two normalised indices as long as the consequences that arise of mixing for granular materials, and highlighted its properties
from the normalisation used are noted. as a possible mixing index.
A key achievement from our analysis of the Gini coef-
cient of mixing is that we have illustrated how the formulation
References
of a measure of mixedness in another eld can inuence its
interpretation when adopted as a mixing index for study-
Akao, Y., Shindo, H., Yagi, N., 1976. Estimation of mixing index
ing granular materials. Though not advocating the index, we
and contact number by spot sampling. Powder Technol. 15 (2),
should note that if the Gini coefcient of mixing were to be 207214.
used, one should at least be familiar with its bases of normal- Alemaskin, K., Manas-Zloczower, I., Kaufman, M., 2005. Entropic
isation in order to correctly relate the values of the index to analysis of color homogeneity. Polym. Eng. Sci. 45 (7),
the states of mixing. 10311038.
Beaudry, J.P., 1948. Blender efciency. Chem. Eng. 55 (7), 112113.
Blumberg, R., Maritz, J.S., 1953. Mixing of solid particles. Chem.
3.5. Summary of indices discussed; other mixing
Eng. Sci. 2 (6), 240246.
indices Finnie, G.J., Kruyt, N.P., Ye, M., Zeilstra, C., Kuipers, J.A.M., 2005.
Longitudinal and transverse mixing in rotary kilns: a discrete
Denitions, key expected values, and features of normalisa- element method approach. Chem. Eng. Sci. 60 (15), 4083
tion of the mixing indices discussed in Sections 3.23.4 are 4091.
summarised in Table 1. It is important to understand such Gu, Z., Chen, J.J.J., 2013. Measures of mixedness for the mixing of
basic features of any mixing index in order to interpret it granular materials: a probabilistic approach. Honours project
report, Department of Chemical and Materials Engineering,
correctly.
University of Auckland, Unpublished manuscript.
We note that it is possible to apply the same principle of Guida, A., Nienow, A.W., Barigou, M., 2010. Shannon entropy for
analysis to other types of mixing indices in the literature. For local and global description of mixing by Lagrangian particle
example, Marigo et al. (2012) presented a similar analysis of tracking. Chem. Eng. Sci. 65 (10), 28652883.
their segregation index based on contact points between Inoue, I., Yamaguchi, K., 1968. (Analysis of
different types of particles where they derived the expected solid mixing and stochastic process). Res. Assoc. Powder.
Technol. 5 (1), 10101018 (in Japanese).
values of the index for the three characteristics states that we
Lacey, P.M.C., 1943. The mixing of solid particles. Trans. Inst.
considered in this work.
Chem. Eng. 21, 5359.
Lacey, P.M.C., 1954. Developments in the theory of particle
4. Conclusions mixing. J. Appl. Chem. 4 (5), 257268.
Lai, F.S., Fan, L.T., 1975. Application of a discrete mixing model to
We have demonstrated the process of matching the charac- the study of mixing of multicomponent solid particles. Ind.
teristic states of a granular mixture to the expected values Eng. Chem. Process. Des. Dev. 14 (4), 403411.
Lloyd, M., Zar, J.H., Karr, J.R., 1968. On the calculation of
of various mixing indices through analysing the probability
information-theoretical measures of diversity. Am. Midland
distributions of sample proportions and cell counts the infor- Nat. 79 (2), 257272.
mation is converted from the exact arrangement of particles Marigo, M., Cairns, D.L., Davies, M., Ingram, A., Stitt, E.H., 2012. A
in the mixture, to the probability distributions, and ultimately numerical comparison of mixing efciencies of solids in a
to the value of the mixing indices. cylindrical vessel subject to a range of motions. Powder
We have also noted that a quantitative understanding of Technol. 217, 540547.
Masiuk, S., Rakoczy, R., 2006. The entropy criterion for the
the states of a granular mixture using probability distribu-
homogenisation process in a multi-ribbon blender. Chem.
tions helps avoid any confusion of the states specically,
Eng. Process. 45 (6), 500506.
we have shown that at the end of an ideal mixing opera- Ogawa, K., 2006. Quantitative index for anxiety/expectation and
tion, the granular mixture is best represented by the randomly its applications. J. Chem. Eng. Jpn. 39 (1), 102110.
mixed state, not the perfectly ordered state, and thus one Ogawa, K., Ito, S., 1973. A denition of quality of mixedness. J.
should discriminate between these two states in the study Chem. Eng. Jpn. 8 (2), 148151.
of granular mixing; these two states merge into one in the Schutyser, M.A.I., Padding, J.T., Weber, F.J., Briels, W.J., Rinzema,
A., Boom, R., 2001. Discrete particle simulations predicting
limit of very large sample size or cell size, which we have
mixing behavior of solid substrate particles in a rotating drum
established using asymptotic arguments based on either
fermenter. Biotechnol. Bioeng. 75 (6), 666675.
the probability distributions or the values of the mixing Sen, A., 1973. On Economic Inequality. Clarendon Press, Oxford,
indices. pp. 2446.
We have illustrated with specic examples as to how the Shannon, C.E., 1948. A mathematical theory of communication.
formulation of mixing indices may be inuenced by the choice Bell Syst. Tech. J. 27, 379423, 623656.

Please cite this article in press as: Gu, Z., Chen, J.J.J., A probabilistic analysis of some selected mixing indices. Chem. Eng. Res. Des. (2014),
http://dx.doi.org/10.1016/j.cherd.2014.04.014
CHERD-1558; No. of Pages 11
ARTICLE IN PRESS
chemical engineering research and design x x x ( 2 0 1 4 ) xxxxxx 11

Theil, H., 1972. Studies in Mathematical and Managerial Too, J.R., Rubison, R.M., Fan, L.T., 1979. Studies on
Economics: Statistical Decomposition Analysis: With multicomponent solids mixing and mixtures: II. Estimation of
Applications in the Social and Administrative Sciences, vol. mixing index and contact number by spot sampling of a
14. North-Holland Publishing Company, Amsterdam. multicomponent mixture in an incompletely mixed state.
Theil, H., Finizza, A.J., 1971. A note on the measurement of racial Powder Technol. 23 (1), 99113.
integration of schools by means of informational concepts. J. Trivellato, F., 2011. On the efciency of turbulent mixing in
Math. Sociol. 1 (2), 187193. rotating stirrers. Chem. Eng. Process.: Process Intensif. 45 (6),
Thiel, W.J., Lai, F., Hersey, J.A., 1981. Comments on suggestions 500506.
on the nomenclature of powder mixtures. Powder Technol.
28 (1), 117118.

Please cite this article in press as: Gu, Z., Chen, J.J.J., A probabilistic analysis of some selected mixing indices. Chem. Eng. Res. Des. (2014),
http://dx.doi.org/10.1016/j.cherd.2014.04.014

You might also like