A non-uniform bound on Poisson approximation in somatic cell hybrid model (Q2486550)

Somatic cell hybrids are routinely used to assign particular human genes to particular human chromosomes. To do this we use the information of a panel of somatic cell hybrids. Let \(n\) be the number of hybrid clones in a panel. Each row of this panel corresponds to a particular clone. The columns of the panel are unique, and the panel has the unusual property that every pair of columns differs in entries. Minimum Hamming distance is a natural measure of the redundancy of a panel. The Hamming distance \(\rho (c_s, c_t)\) between two columns \(c_s\) and \(c_t\) from a random panel is just the number of entries in which they are different. Let \(C^{n}_{s}\) denote the colums \(s\) of a random panel of n clones. The authors give a non-uniform bound for approximating the point probabilities of the number of pairs of chromosomes for which the Hamming distance \(\rho (C^n_s, C^{n}_t)\) is less than some fixed Hamming distance \(d\) by using the Stein-Chen method. Let \(\Gamma = \{ \alpha = \{ s,t \} : 1 \leq s \neq t \leq 23 \}. \) For any two distinct columns \(C^{n}_{s}\) and \(C^{n}_{t}\), define \(X_{\{ s, t \}}\) to be the indicator of the event \(\rho ( C^{n}_{s}, C^{n}_{t})< d,\) where \(d\) is some fixed Hamming distance. For each \(\alpha \in \Gamma, \) the probability \(P(X_{\alpha} =1)\) is \[ p= \sum^{d-1}_{i=0} { {n}\choose{i}}q^{i}(1-q)^{n-i}, \] where \(q=2p_{0}(1-p_{0})\) is the probability that \(C^{n}_{s}\) and \(C^{n}_{t}\) differ in any entry. Let \(W=\sum_{\alpha\in \Gamma}X_{\alpha}\) be the number of pairs of chromosomes for which the Hamming distance \(\rho (C^{n}_{s}, C^{n}_{t})\) is less than \(d\). It seems reasonable to approximate \(W\) as a Poisson with mean \(\lambda=E[W]= {23\choose 2}p.\) For each \(\alpha\in\Gamma\), let \(B_{\alpha}=\{ \beta \in \Gamma : \alpha \cap \beta \neq \empty \}\). Let \[ b = \sum_{\alpha\in \Gamma}\sum_{\beta\in B_{\alpha}-\{\alpha\}} E[X_{\alpha}X_{\beta}]. \] The main purpose is to approximate the point probabilities of \(W\) by a Poisson distribution. The main result is as follows: Let \(W\), \(\lambda\) and \(b\) be defined as above. Then 1. For \(w_{0}\in \{ 1, \cdots, 253 \}\), \[ | P(W=w_0)-e^{-\lambda}{{\lambda^{w_0}}/{w_0\!}}|\leq \min \left\{\omega_0^{-1}/{w_0}, \lambda^{-1 }\right\} [\lambda\min\{ 1, 43p\}+\min\{ \lambda, b\}]. \] 2. For \(w_{0}=0\), \[ | P (W=0)-e^{-\lambda}|\leq \lambda^{-2}(\lambda + e^{-\lambda}-1)\max\{43\lambda p, b\}. \]