Noninverse correlation induction: Guidelines for algorithm development (Q1813816)

The authors propose a new interesting simple practical procedure for generating two correlated random variables \(X\), \(Y\) having \(F\), respectively \(G\), as distribution functions. A known method of inducing correlation between the \(X\) and \(Y\) uses \(X=F^{-1}(U)\) and \(y=G^{-1}(V)\) where \(U\), \(V\) are \((0,1)\) uniform correlated random variables. As particular cases consider \(V=U\) or the antithetic procedure with \(V=1-U\). Generally this technique is heavily to be applied because we must invert the cumulative distribution functions \(F\) and \(G\). The authors suggest the generation of the correlated random variables \(X\), \(Y\) by applying a general rejection procedure where the dependent uniform random variables \(U\) and \(V\) are used. Therefore, if the \((X,Y)\) pair is accepted then the dependent variables \(U\), \(V\) induce a correlation between the random variates \(X\), \(Y\). Otherwise, in the case of a rejection, they try to generate independent \(X\), \(Y\) random variables or to run again the first rejection step, with a new uniform random vector \((U,V)\). Two illustrations of this method are also given for exponential and beta distributions. The proposed technique is easily to be programmed on the computer. In addition the authors renounce to perform laborious calculations for determining the theoretical induced correlation because the experimental correlation can be obtained by a computer simulation technique.