Asymptotic approximations for the distributions of multinomial goodness- of-fit statistics (Q802233)

Let \(\pi =(\pi_ 1,\pi_ 2,...,\pi_ k)\), \(\pi_ j>0\), \(\Sigma \pi_ j=1\), be a probability distribution. Let \(Y=(Y_ 1\), \(Y_ 2,...,Y_ k)\) be a random vector with the multinomial distribution \[ P(Y_ j=n_ j,\quad 1\leq j\leq k)=n!\prod^{k}_{j=1}(\pi_ j^{n_ j}/n_ j!),\quad n=n_ 1+n_ 2+...+n_ k. \] For testing the simple hypothesis \(H: \pi =p\), where \(p=(p_ 1,p_ 2,...,p_ k)\) is a fixed probability distribution, the following three statistics are commonly \(used:\) T\({}_ 1=\sum^{k}_{j=1}(Y_ j\)-np\({}_ j)^ 2/np_ j\); \(T_ 2=2\sum^{k}_{j=1}Y_ j\) \(\log (Y_ j/np_ j);\) and \(T_ 3=4\sum^{k}_{j=1}(\sqrt{Y_ j}\)-\(\sqrt{np_ j})^ 2.\) It is well known that, under H, each \(T_ j\) is asymptotically chi squared with k-1 degrees of freedom. There is, however, very little information on the magnitude of \(np_ j\) for which this asymptotic distribution can safely be applied. The authors develop expansions of the distribution of \(T_ 2\) and \(T_ 3\), whose error terms give information on the accuracy of the approximations mentioned. The paper is a valuable contribution to its field.