Rao’s quadratic entropy and maximum diversification indexation 3
this debate by formally establishing the principles at play be-
hind the MD approach through a new formulation based on
Rao’s Quadratic Entropy.
2.2. Rao’s quadratic entropy
Let us first recall the general definition of Rao’s Quadratic
Entropy (RQE). RQE† is a general approach to measuring di-
versity introduced by Rao (1982a, 1982b). Given a population
of individuals P, it is defined as the average difference between
two randomly drawn individuals from
P. More formally, sup-
pose that each individual in
P is characterized by a set of
measurement X and denote by P the probability distribution
function of X.RQEof
P is defined as:
H
D
(P) =
d(X
1
, X
2
)P(dX
1
)P(dX
2
), (4)
where the non-negative, symmetric ‘dissimilarity’ function
d(., .) expresses the difference between two individuals from
P. When X is a discrete random variable, H
D
(P) becomes
H
D
(P) = p
Dp, (5)
where p is column vector of probabilities with elements p
i
=
P(X = x
i
), ∀i = 1, ..., N and D = (d
ij
)
N
i, j=1
is the dis-
similarity matrix, where N is a number of individuals. The
interpretation of RQE is straightforward: the higher is H
D
(P),
the higher is the diversity of individuals among the population
P. RQE has been used extensively in fields such as statistics
(Rao 1982a, 1982b, Nayak 1986a, 1986b), ecology (Champely
and Chessel 2002, Pavoine et al. 2005, Ricotta and Szeidl
2006, Pavoine and Bonsall 2009, Pavoine 2012, Zhao and Naik
2012), energy policy (Stirling 2010) and income distribution
(Nayak and Gastwirth 1989).
Carmichael et al. (2015) recognize its potential for portfolio
theory and propose R QE as the theoretical underpinning for
a novel class of measures of portfolio diversification. To do
so, the magnitudes
P, X , P and D above are transferred to
a portfolio selection setting. Consider a universe of N assets
(risky or not) and denote by w = (w
1
, ..., w
N
)
a given long-
only portfolio associated to this universe of assets, with w
i
the weight of asset i in w. Next, define the random variable
X to take the finite values 1, ..., N (N assets) and its proba-
bility distribution P(X = i) = w
i
, ∀i = 1, ..., N , so that
it is associated to the random experiment whereby assets are
randomly selected (with replacement) from the portfolio w.
Then Carmichael et al. (2015) define RQE of a portfolio w
as half of the mean difference between two randomly drawn
(with replacement) assets from portfolio w:
H
D
(w) =
1
2
w
D w, (6)
where D = (d
ij
)
N
i, j=1
is a dissimilarity matrix between the
various assets of the portfolio. Carmichael et al. (2015) demon-
strate how H
D
(w) now defines a valid class of portfolio di-
all have the same volatility, with the correlation matrix defined as
follows: ρ
12
= 1, ρ
3,4
=−1andρ
ij
= 0,(i, j) = (1, 2) and (3, 4).
One can show that DR
2
w
MD
=+∞, a counter-intuitive result.
†RQE is also referred to as Diversity Coefficient (Rao 1982b)or
Quadratic Entropy (Rao and Nayak 1985).
versification measures, which meets ex-ante desirable proper-
ties of diversification, unifies several portfolio diversification
measures and utility functions that have been analyzed in the
literature and provides a flexible but formal approach for fund
managers to develop new, diversified portfolios.
The interpretation of H
D
(w) is straightforward. All things
equal, the higher H
D
(w) is, the more portfolio w is diversified,
because the more dissimilar are assets, the less is the probability
that they do poorly at the same time and in the same propor-
tion. A well-diversified portfolio can therefore be obtained by
maximizing (6).
In practice, one needs to specify the dissimilarity matrix
D. As argued in Carmichael et al. (2015), this added flexibil-
ity represents both an advantage and a challenge when using
RQE to quantify portfolio diversification degree. The matrix
D measures the difference, or distance, between assets.‡ One
possible example defines d
ij
as
d
ij
= (1 − ρ
ij
), (7)
where ρ
ij
measures the correlation between asset i and j ; d
ij
thus measures the difference in terms of correlation between
assets i and j, which is high when assets have low correlation.
Carmichael et al. (2015) review other suitable candidates.
2.3. Maximum diversification meets Rao’s quadratic entropy
We can now reconsider the maximum diversification indexa-
tion in the context of RQE. In that context, note that the square
of DR minus 1 gives
DR
2
(w) −1 =
w
σ
2
w
w
− 1. (8)
Since (8) preserves the preference ordering on the set of long-
only portfolios, the MDP can also be obtained by maximizing
DR
2
− 1. Notice further that (8) can also be written as§
DR
2
(w) −1 =
w
w
w
w
. (9)
The numerator w
wis RQE of portfolio w with the dissim-
ilarity matrix = (γ
ij
) having
γ
ij
= (1 − ρ
ij
)σ
i
σ
j
(10)
as typical element. In this formulation, γ
ij
measures the differ-
ence in terms of correlation and volatility between assets i and
j. This difference is high when assets have moderate-to-high
volatility and low correlation. RQE’s maximization based on
(RQE
hereafter) therefore gives more weight to assets with
moderate-to-high volatility and low correlation between them.
‡Taking account of that difference is the key distinguishing feature
of RQE relative to other entropy measures used in finance, notably
Shannon entropy and Tsallis Entropy (see Zhou et al. 2013,
Glasserman and Xu 2014).
§The development of
w
σ
2
and w
w are
w
σ
2
=
N
i=1
w
2
i
σ
2
i
+
N
i, j =1
w
i
w
j
σ
i
σ
j
and w
w =
N
i=1
w
2
i
σ
2
i
+
N
i, j =1
w
i
w
j
ρ
ij
σ
i
σ
j
respectively. Therefore, DR
2
(w) − 1 =
N
i , j=1
(1−ρ
ij
)σ
i
σ
j
w
i
w
j
w
w
=
w
w
w
w
,where = (γ
ij
)
N
i, j =1
with
γ
ij
= (1 − ρ
ij
)σ
i
σ
j
.