Gini coefficient (aka Gini Index) is a quantity used in economics to describe income inequality. It is 0 for uniformly distributed income, and approaches 1 when all income is in hands of one individual.

For probabilities we have:

$$ G(\{p_i\}_{i\in\{1...n\}}) = \sum_{i=1}^n \frac{2i-n-1}{n}p_i, $$

where $p_i$ is a non-decreasing sequence, or in the continuous case:

$$ G(p) = \int_0^1 \left(1-2\int_0^xp(y)dy\right)dx = \int_0^1 (2x-1)p(x)dx $$ where probability density function $p(x)$ is non-decreasing.

As it is a measure of uniformness of a probability distribution, I was curious if it is possible to relate Gini coefficient to some Rényi entropy $H_q$ (for example, collision entropy $H_2$)? (Or Tsallis entropy $T_q$ as for the continuous case formulas are shorter.)

I got an upper bound

$$ G(p) \leq \sqrt{\tfrac{1}{3}(e^{-H_2(p)}-1)} = \sqrt{-\tfrac{1}{3}T_2(p)} $$ with Schwarz inequality for $(2x-1)$ and $(p(x)-1)$.

It works similarly for discrete probabilities:

$$ G(\{p_i\})\leq\sqrt{\tfrac{n}{3}}\sqrt{1-\tfrac{1}{n^2}}\sqrt{e^{-H_2(\{p_i\})}-\tfrac{1}{n}}. $$

Is there any non-trivial lower bound? That is, except for

$$G(p)\geq 1 - e^{H_0(p)} = -T_0(p),$$ $$G(\{p_i\})\geq 1 - \tfrac{1}{n}e^{H_0(\{p_i\})}.$$

Is there an upper bound that is saturated for $p = \{0,0,\cdots,0,1\}$?

  • Am I tired or should the RHS of the second display actually be negated? – cardinal Nov 27 '13 at 2:06
  • @cardinal You are right, fixed. – Piotr Migdal Nov 27 '13 at 11:24
  • 1
    Something is off with your lower bound on G({p_i}). G({0,...,0,1}) is equal to 1-1/n, but H_0({0,...,0,1})=0, so your bound would give G \geq 1. Is the factor "-1" in the definition of G correct? – Marcin Kotowski Nov 27 '13 at 17:25
  • p_i is the amount of money possessed by (n-i)'th richest person? – Marcin Kotowski Nov 28 '13 at 5:24
  • @MarcinKotowski Here $G$ is biases, maximally $1-1/n$ (you can "unbias" it by multiplying by $n/(n-1)$, exactly the same reason as for variance). $H_0(\{0,0,...,1\})=0$ so the lower bound is $1-1/n\geq 1-e^0/n$, right? (Or I am missing something.) – Piotr Migdal Nov 28 '13 at 10:46
up vote 4 down vote accepted

As of now, the best I got is:

$$ 1 - e^{H_{1/2}(p)} \leq G(p) \leq \sqrt{1-e^{2 H_{1/2}(p)}}, $$ where $e^{H_{1/2}(p)} = 2 \ln \int_0^1p^{1/2}(x)dx$.

Both inequalities are saturated for both $G(p)=0$ and $G(p)=1$.

The lower bound is better than the previous one ($1 - e^{H_{0}(p)}$), as Rényi entropy $H_q(p)$ is non-decreasing in $q$.

The higher bound is different from $\sqrt{\tfrac{1}{3}e^{-H_2(p)}-1}$ (there is no inequality between them).

And for practical purposes their quadratic mean seems to be a good approximation, $$ G(p) \approx \sqrt{1 - e^{H_{1/2}(p)}}. $$

To get some taste of it, here is a plot for the lower, the upper bound and the approximation:

enter image description here

(Sampled for distributions of the form $p(x) \propto (1-x+r)^{-\alpha} + c$, where $r$, $\alpha$ and $c$ are some parameters.)

Proof

We have: $$ G(p) = \int_0^1 (2x-1) p(x) dx = \tfrac{1}{2} \int_0^1 \int_0^1 |p(x) - p(y)|dx dy. $$ (Actually, the right hand side is more general, as it does not require any ordering of p(x).)

For the lower bound we use $$ |p(x) - p(y)| = \left|\sqrt{p(x)} + \sqrt{p(y)}\right| \left| \sqrt{p(x)} - \sqrt{p(y)} \right| \geq \left( \sqrt{p(x)} - \sqrt{p(y)} \right)^2 $$ and perform the integration.

For the upper bound we use Schwartz inequality for $\left|\sqrt{p(x)} + \sqrt{p(y)}\right|$ and $\left| \sqrt{p(x)} - \sqrt{p(y)} \right|$.

Your Answer

 
discard

By clicking "Post Your Answer", you acknowledge that you have read our updated terms of service, privacy policy and cookie policy, and that your continued use of the website is subject to these policies.

Not the answer you're looking for? Browse other questions tagged or ask your own question.