Adapted topologies and higher rank signatures (Q6104023)

The authors show in their main result that on the space of discrete time stochastic processes adapted to a fixed filtration \((\mathcal{F}_t)_t\), the rank \(r\) adapted topology \(\tau_r\) due to \textit{D. N. Hoover} and \textit{H. J. Keisler} [Trans. Am. Math. Soc. 286, 159--201 (1984; Zbl 0548.60019)] agrees with the initial topology of the feature map \(\mathbf{X}\mapsto\bar S_{r+1}(\operatorname{Law}(\hat{\mathbf{X}}^r))=\mathbb{E}[\bar{\mathbf{X}}_0^{r+1}]\). For all signature constructions in the article, \(S\) is a robust signature, i.e., a suitable normalization of the usual iterated-integrals signature, based on the work of \textit{I. Chevyrev} and \textit{H. Oberhauser} [``Signature moments to characterize laws of stochastic processes'', Preprint, \url{arXiv:1810.10971}]. The topology \(\tau_r\) is defined as convergence under expectation \(\mathbb{E}[f(X_n)-f(X)]\to 0\) for all adapted functionals \(f\) that involve up to \(r\) times taking conditional expectation. \(\mathbf{X}^r\) is the rank \(r\) prediction process obtained by iteratively taking conditional propability \(r\) times, extending the notion by \textit{D. J. Aldous}, Weak convergence and general theory of processes. Unpublished Draft of Monograph (1981)], and \(\hat{S}_{r+1}\) is the rank \(r+1\) expected signature. Furthermore, \(\bar{\mathbf{X}}_0^{r+1}=\mathbb{E}[S(\bar{\mathbf{X}}_0^{r})|\mathcal{F}_t]\) is the rank \(r+1\) conditional signature process. Algebraically, the target space of a higher rank (expected) signature is constructed by iterated application of the non-unital free associative algebra functor (naturally leading to a multi-grading), finally adding a unit and taking a suitable normed completion towards a space of formal series. The authors thus successfully demonstrate that while the (robust, rank \(r+1=1\)) expected signature is insufficient as a feature map when studying discrete time optimal stopping problems, it still can be effectively used to create a suitable refined feature map, namely \(\mathbf{X}\mapsto\bar S_{2}(\operatorname{Law}(\hat{\mathbf{X}}^1))\). Iterating this procedure, the higher rank robust signatures capture more and more information about the filtration of the adapted process. As pointed out by the authors themselves, the multi-grading of the higher rank tensor algebras has an immediate and powerful stochastic interpretation, describing ``the interplay of the law and the filtration of the process \(\mathbf{X}\) in a hierarchical manner'' (p. 2139)