On the lack of semimartingale property (Q2668503)

The authors extend the characterization of semimartingale functions due to \textit{E. Cinlar} et al. [Z. Wahrscheinlichkeitstheor. Verw. Geb. 54, 161--219 (1980; Zbl 0443.60074)] to the non-Markovian setting. They show that if a function of a semimartingale remains a semimartingale, then under certain conditions the function must have intervals where it is a difference of two convex functions. Under suitable conditions this property also holds for random functions. As an application, they prove that the median process defined by \textit{V. Prokaj} et al. [Ann. Inst. Henri Poincaré, Probab. Stat. 47, No. 2, 498--514 (2011; Zbl 1216.60048)] is not a semimartingale. This is defined as follows. Let \((D_t(x))_{t\geq 0}\) be the solution of \(d D_t(x) = \sigma(D_t(x))dB_t\), \(D_0(x)=x\in[0,1]\), where \(\sigma(x)=x\wedge(1-x)\). The median process is \((D^{-1}_t(1/2))_{t\geq 0}\). More generally, they show that \((D^{-1}_t (\alpha))_{t\geq 0}\) is not a semimartingale for \(\alpha\in(0, 1)\).