Minimum correlation for any bivariate Geometric distribution

Abstract

Consider a bivariate Geometric random variable where the first component has parameter $p_1$ and the second parameter $p_2$. It is not possible to make the correlation between the marginals equal to -1. Here the properties of this minimum correlation are studied both numerically and analytically. It is shown that the minimum correlation can be computed exactly in time $O(p_1^{-1} \ln(p_2^{-1}) + p_2^{-1} \ln(p_1^{-1}))$. The minimum correlation is shown to be nonmonotonic in $p_1$ and $p_2$, moreover, the partial derivatives are not continuous. For $p_1 = p_2$, these discontinuities are characterized completely and shown to lie near (1- roots of 1/2). In addition, we construct analytical bounds on the minimum correlation.