Statistical inference for the intensity in a partially observed jump diffusion (Q2414732)

Consider a latent signal described by the jump diffusion $dX_t = aX_t dt + bdW_t + \delta dN_t$, $X_0 = x_0 \in \mathbb{R}$, where $W_t$ is a standard Brownian motion and $N_t$ is an independent Poisson process with intensity $\lambda$ on the filtered probability space $(\Omega, \mathcal{F}, \mathcal{P}, (\mathcal{F}_t )_{t\in[0,T]})$ for $T > 0$ and constants $a, b, \delta \in \mathbb{R}$. The diffusion is not observed, but instead partial information is propagated via the observation process $Y_t = h \int_0^t X_s ds + B_t\), \(t \in [0, T]$, where the driving noise, $B_t$, is another standard Brownian motion, independent of $W_t$ and $N_t$, and $h\in \mathbb{R}$ is a constant that explains the relationship between the signal and its observation. The authors study the asymptotic behavior of the consistent Least Squares Estimator (LSE) of the intensity $\lambda$, establishing its large and moderate deviation principles on the basis of the Gartner-Ellis theorem. Together with the central limit theorem, these results compose the limit theorems of the LSE of $\lambda$ with different dominated scale $c_n$, where $c_n = n$ for large deviations and $c_n = n^a$, $a \in (0, 1/2)$, for moderate deviations. They also apply the moderate deviation principle to find the approximate power of the associate hypothesis testing for $\lambda$ and to construct an effective confidence interval for $\lambda$.