Properties of the EMCEL scheme for approximating irregular diffusions (Q2069772)

The paper under review is concerned with a so-called EMCEL scheme (the full name is the scheme of `embeddable Markov chain with expected time lag \(h\)') for approximating a general diffusion on an interval. Roughly speaking, unlike the Euler scheme, this scheme obtains an approximating sequence by cutting time into (stochastic) pieces with the same expectation. More precisely, let \(Y\) be a one-dimensional continuous strong Markov process on an open, half-open or closed interval \(I\subset \mathbb{R}\). Assume that \(Y\) is regular in the sense that for every \(y\in I^\circ\), the interior of \(I\), and \(x\in I\), \(P_y\left[ H_x(Y)<\infty\right]>0\), where \(H_x(Y):=\inf\{t\geq 0: Y_t=x\}\), and that \(Y\) is in the natural scale. Set \(H_{a,b}(Y):=H_a(Y)\wedge H_b(Y)\) for \(a,b\in \bar{I}\), the closure of \(I\). Take an i.i.d sequence \((\xi_k)_{k\in \mathbb{N}}\) of random variables satisfying \(P(\xi_k=\pm 1)=1/2\). Given a discretization parameter \(h\in (0,\bar{h}]\), where \(\bar{h}>0\) is fixed, the crucial element to defining the approximating sequence of Markov chains is the so-called scale factor \(\hat{a}_h: I\rightarrow [0,\infty)\), which is chosen in such a way that the expected time it takes \(Y\) started in \(y\in I\) to leave the interval \((y-\hat{a}_h(y),y+\hat{a}_h(y))\) is equal to \(h\), i.e., \(\hat{a}_h\) satisfies for all \(y\in I\) that \(E_y\left[H_{y-\hat{a}_h(y), y+\hat{a}_h(y)}(Y) \right]=h\). Then for an initial value \(y_0\in I\), the Markov chain is recursively defined on the time grid \(\{kh: h\in \mathbb{N}_0\}\) as \[ \hat{X}^h_0=y_0,\quad \hat{X}^h_{(k+1)h}=\hat{X}^h_{kh}+\hat{a}_h(\hat{X}^h_{kh})\xi_{k+1}. \] This Markov chain can be extended to a continuous-time process, denoted by \(\hat{X}^h_t\), by linear interpolation. Then it can be shown that \(\hat{X}^h_t\) converges weakly to \(Y\) as \(h\downarrow 0\). The paper under review investigates the properties of the scale factor \(\hat{a}_h\). At first, it discusses the asymptotic properties of \(\hat{a}_h\) as \(h\downarrow 0\). It decreases to \(0\) in a certain meaning and the convergence ratio is also explored. Secondly, it is examined how \(\hat{a}_h(y)\) depends on \(y\). It turns out that for \(h\in (0,\bar{h}]\) and \(z\in \{-1,1\}\), the mapping \(y\mapsto y+\hat{a}_h(y)z\) is nondecreasing. This leads to certain results translating into some good stability features of the scheme. Finally, an ODE satisfied by \(\hat{a}_h\) is obtained.