Martingale limit theorems of divisible statistics in a multinomial scheme with mixed frequencies (Q553023)

For each \(n\) let \(\{\nu_{ni}\}_{1\leq i\leq N}\) be a random vector which follows the multinomial distribution with sample size \(n\) and probabilities \(\{p_{ni}\}_{1\leq i\leq N}\). We define divisible statistics as the sums of some functions \(g_{ni}\) of the frequencies, i.e., \(\sum^N_{i=1} g_{ni}(\nu_{ni}, np_{ni})\). Divisible stasistics include many well-known and widely used statistics. Let \(Y_{ni}= (\nu_{ni}- np_{ni})/\sqrt{np_{ni}}\), and choose functions \(h_{ni}\) such that, for all \(n\), \[ \sum^N_{i=1} h_{ni}(Y_{ni}, np_{ni})= \sum^N_{i=1} g_{ni}(\nu_{ni}, np_{ni}) \] holds. In this paper, the authors define the partial-sum process as \[ X_n(t)= {1\over\sqrt{N}}\, \sum^{Nt}_{i=1} (h_{ni}(Y_{ni}, np_{ni})- \operatorname{E}h_{ni}(Y_{ni}, np_{ni}))\quad (0\leq t\leq T), \] and, to study some asymptotic properties of \(X_n(t)\), decompose \(X_n(t)\) into the martingale part \[ W_n(t)= {1\over\sqrt{N}}\, \sum^{Nt}_{i=1} (h_{ni}(Y_{ni}, np_{ni})- \operatorname{E}[h_{ni}(Y_{ni}, np_{ni})|{\mathcal H}^{(n)}_{i-1}]) \] and its compensator \[ K_n(t)= {1\over\sqrt{N}}\, \sum^{Nt}_{i=1} (\operatorname{E}[h_{ni}(Y_{ni}, np_{ni})|{\mathcal H}^{(n)}_{i-1}])- \operatorname{E}h_{ni}(Y_{ni}, np_{ni}), \] where \({\mathcal H}^{(n)}_i= \sigma\{\nu_{ni}: k\leq i\}\) \((0\leq i\leq N)\). Then, under some conditions, they obtain the weak convergence of both \(W_n\) and \(K_n\).