Low Frequency L\'evy Copula Estimation

Let $X$ be a $d$-dimensional L\'evy process with L\'evy triplet $(\Sigma,\nu,\alpha)$ and $d\geq 2$. Given the low frequency observations $(X_t)_{t=1,\ldots,n}$, the dependence structure of the jumps of $X$ is estimated. The L\'evy measure $\nu$ describes the average jump behavior in a time unit. Thus, the aim is to estimate the dependence structure of $\nu$ by estimating the L\'evy copula $\mathfrak{C}$ of $\nu$, cf. Kallsen and Tankov \cite{KalTan}. We use the low frequency techniques presented in a one dimensional setting in Neumann and Rei{\ss} \cite{NeuRei} and Nickl and Rei{\ss} \cite{NicRei} to construct a L\'evy copula estimator $\widehat{\mathfrak{C}}_n$ based on the above $n$ observations. In doing so we prove $$\widehat{\mathfrak{C}}_n\to \mathfrak{C},\quad n\to\infty$$ uniformly on compact sets bounded away from zero with the convergence rate $\sqrt{\log n}$. This convergence holds under quite general assumptions, which also include L\'evy triplets with $\Sigma\neq 0$ and $\nu$ of arbitrary Blumenthal-Getoor index $0\leq\beta\leq 2$. Note that in a low frequency observation scheme, it is statistically difficult to distinguish between infinitely many small jumps and a Brownian motion part. Hence, the rather slow convergence rate $\sqrt{\log n}$ is not surprising. In the complementary case of a compound Poisson process (CPP), an estimator $\widehat{C}_n$ for the copula $C$ of the jump distribution of the CPP is constructed under the same observation scheme. This copula $C$ is the analogue to the L\'evy copula $\mathfrak{C}$ in the finite jump activity case, i.e. the CPP case. Here we establish $$\widehat{C}_n \to C,\quad n\to\infty$$ with the convergence rate $\sqrt{n}$ uniformly on compact sets bounded away from zero. Both convergence rates are optimal in the sense of Neumann and Rei{\ss}.