Convergence of Markov processes near saddle fixed points (Q2370099)

The author studies sequences \((X^N_t)\) \((t\geq 0)\) of Markov processes in two dimensions whose fluid limit is a stable solution of an ordinary differential equation of the form \[ \dot x_t= Bx_t+ \tau(x_t)\;\text{ such that }\;\tau(x)= 0(|x|^2),\tag{\(*\)} \] where \(B\) is the matrix \(\left(\begin{smallmatrix} -\mu & 0\\ 0 &\lambda\end{smallmatrix}\right)\) with eigenvalues \(-\mu< 0\) and \(\lambda> 0\). Let denote the flow associated with \((*)\). There exists some \(x_0\neq 0\) such that \(\varphi_t(x_0)\to 0\) \((t\to\infty)\). The set of such \(x_0\) is a stable manifold \(M_s\). There exists also some \(x_\infty\) such that \(\varphi^{-1}_t(x_\infty)\to 0\) \((t\to\infty)\). The set of such \(x_\infty\) is an unstable manifold \(M_u\). Consider a sequence of Markov processes \((X^N_t)\) such that \(X^N_0= x_0\) for \(x_0\in M_s\) and assume that the variance of fluctuations of \(X^N_t\) is inversely proportional to \(N\). It is shown that there occur three different types of behaviour depending on the time scale: (A) On compact time intervals, \(X^N_t\) converges to the stable solution of \((*)\), the fluctuations around this limit being of order \(N^{-1/2}\). (B) There exists some \(\overline x_0\neq 0\) (only depending on \(x_0\)) and a Gaussian random variable \(Z_\infty\) such that if \(t\) lies in the interval \([R,(2\lambda)^{-1}\log N-R]\), then \[ X^N_t=\overline x_0 e^{-\mu t}(e_1+ \varepsilon_1)+ N^{-1/2} Z_\infty e^{\lambda t}(e_2+ \varepsilon_2) \] for some \(\varepsilon_i(t, N)\to 0\) uniformly in \(t\) in probability as \(R\), \(N\to\infty\), where \(\{e_1,e_2\}\) is the standard basis for \(\mathbb R^2\). (C) On time intervals of a fixed length around \((2\lambda)^{-1}\log N\), \(X^N_t\) converges to the unstable solution of \((*)\). The most interesting behaviour occurs on time intervals of fixed length around \((2(\lambda+\mu))^{-1}\log N\), as for these values of \(t\), \(\overline x_0 e^{-\mu t}\) and \(N^{-1/2}Z_\infty e^{\lambda t}\) are of the same order. It is shown that at these times \(X^N_t\) crosses all straight lines through \(0\), and also that \(|X^N_t|\) attains its minimum value when \(t\) is in this range. By using scaling limits, the author also rederives a result of \textit{J. F. C. Kingman} [Bull. Lond. Math. Soc. 31, No. 5, 601--606 (1999; Zbl 0933.60007)] for the OK Corral gunfight model, which says that \[ N^{-3/4} S^N\to 2^{3/4}|Z|^{1/2}\text{ (weakly) } (Z\sim N(0,1/3)), \] where \(S^N\) is the number of survivors when the process terminates.