Toric geometry of path signature varieties (Q2221770)

For a map \(X:[0,1]\rightarrow {\mathbb R}^d\), and its projection \(X_i\) on the \(i\)-th coordinate, for a fixed \(k\in{\mathbb N}\), \textit{K. T. Chen} defined in [Trans. Am. Math. Soc. 89, 395--407 (1958; Zbl 0097.25803)] the \(k\)-th signature of \(X\) as the \(k\)-tensor \(\sigma^{(k)}(X)\in ({\mathbb R}^d)^{\otimes k}\) whose \((i_1,\ldots,i_k)\)-th entry is the iterated integral \[\int_0^1\int_0^{t_k}\cdots \int_0^{t_3}\int_0^{t_2} \dot{X}_{i_1}(t_1)\cdots \dot{X}_{i_k}(t_k)\, dt_1\cdots dt_k\] and \(\sigma^{(0)}(X)=1\). The sequence \(\sigma(X)=(\sigma^{(k)}(X): k\geq 0)\) is the \textit{signature} of the path \(X\) and there is a truncated version of it \(\sigma^{\leq m}(X)=(\sigma^{(k)}(X): 0\leq k\leq m)\). For smooth \(X\), the iterated integrals do not give any extra information not derived from \(X\), however when \(X\) is not smooth, there can exist sequences of smooth paths \(X^n, Y^n\) both point-wise converging to \(X\) and with corresponding sequences of iterated integrals also converging but to different limits; these limits are no longer iterated integrals, but they satisfy the so-called Chen identities [loc. cit.] Fixing the class of paths and an integer \(k\), the \(k\)-th signature \(\sigma^{(k)}\) is an algebraic map into \(({\mathbb R}^d)^{\otimes k}\) and the (Zariski) closure of its image is called the \textit{\(k\)-th signature variety}. In the paper under review, the authors consider signature varieties for two classes of paths. First, they consider the signature variety of rough paths, which are first defined as the Zariski closure of the image of rough paths of order \(m\) in the tensor space \(({\mathbb R}^d)^{\otimes k}\), but since this is only a semi-algebraic subset, they complexify first and then take the projectivization in \(({\mathbb C}^d)^{\otimes k}\). The corresponding signature variety has some similarities with the classical Veronese variety and is known as the \textit{rough Veronese variety} \({\mathcal R}_{d,k,m}\). One such similarity considers the fact that the classical Veronese variety is the image of the map given by degree \(k\) monomials in the usual grading, and it is known that the rough Veronese variety \({\mathcal R}_{d,k,m}\) is the closure of the image of a weighted projective space by a map given by all monomials of weighted degree \(k\). Looking for more similarities, and since the classical Veronese variety is defined by quadrics, the corresponding property for the rough Veronese variety is not true by a counterexample in Proposition 28 of [\textit{F. Galuppi}, Linear Algebra Appl. 583, 282--299 (2019; Zbl 1432.14041)]. The first main result of the paper under review is that, in general, does not even exist a bound to the degree of the generators of the ideal defining the rough Veronese variety (Proposition 2.9). On the other hand, the second main result (Proposition 2.11) shows that \({\mathcal R}_{d,k,m}\) is defined by quadrics outside of a coordinate linear subspace of large codimension. Using toric geometry, the authors characterize the cases in which \({\mathcal R}_{d,k,m}\) is an embedding of the weighted projective space and conditions that make it (projectively) normal and examples when it is not. Lastly, they obtain formulas for the dimension and degree of \({\mathcal R}_{d,k,m}\) in Proposition 2.21. The second family of signature varieties that the authors study correspond to the class of axis-parallel paths, for which they obtain a combinatorial parametrization in Lemma 3.3 and using this description they show that the variety is toric in some cases.