Gluing semi-orthogonal decompositions (Q2186020)

In the first part of this interesting paper the authors consider the question of gluing semi-orthogonal decompisotions of dg categories. In the second, several applications to algebraic geometry are discussed. The main application is a construction of semi-orthogonal decompositions on the root stack for a (not necessarily simple) normal crossing divisor. The gluing of linear categories, as formalised by homotopy limits of dg categories, plays in important role in algebraic geometry. For example it allows one to express perfect complexes on a variety in terms of perfect complexes on a cover. It is natural to ask if one can recover structures on the limit from structures on the input. A particularly fruitful structure on a dg category \(\mathcal C\) (or on a triangulated category, or stable \(\infty\)-category etc.)\ is a semi-orthogonal decomposition, which describes how \(\mathcal C\) is assembled from a collection of admissible subcategories \(\{\mathcal C_x\}\). If \(\mathcal C = \varprojlim \mathcal C_i\) and the \(\mathcal C_i\), have semi-orthogonal decompositions, the authors ask if one can find a semi-orthogonal decomposition of \(\mathcal C\). To construct such a decomposition they first rigidify the notion of semi-orthogonal decompositions and define a \emph{pre-ordered semi-orthogonal decomposition} (psod) by equipping the indexing set with the structure of a pre-order, such that the semi-orthogonality condition \(\mathcal C_y \subset \mathcal C_x^\perp\) holds if \(y < x\). They also define an \emph{ordered structure} on a functor between dg categories with psod's, which is an order-reflecting contravariant map of the indexing preorders. Then the authors prove the following gluing result: Let \(\mathcal C_i\) be a diagram of triangulated dg categories, such that every \(\mathcal C_i\) has a psod with indexing preorder \(P_i\), and all functors have ordered structures, inducing a diagram of preorders. Then \(\varprojlim \mathcal C_i\) has a psod indexed by \(\varinjlim P_i\) as long as \(\varinjlim P_i\) is finite and directed. The proof constructs psod's on the limits explicitly. The authors describe an application to conservative descent that illustrates how dg categorical methods simplify a more complicated proof in triangulated categories. But the main interest of this paper is in root stacks for (not necessarily simple) normal crossing divisors. These were defined in [\textit{N. Borne} and \textit{A. Vistoli}, Adv. Math. 231, No. 3--4, 1327--1363 (2012; Zbl 1256.14002)], this paper unfortunately does not recall the somewhat subtle definition. The main application of the gluing result is then as follows: Given a general normal crossing divisor \(D \subset X\) the dg category of perfect complexes on the \(r\)-th root stack has a semi-orthogonal decomposition \(\mathrm{Perf}(\sqrt[r]{(X,D)}) = \langle \mathcal A^k_\chi \mid (k,\chi) \in (\mathbb Z_{D, r}, \leq)\rangle\), where \(\mathbb Z_{D,r}\) is a preorder depending only on \(r\) and the maximal codimension of strata in \(D\). The admissible subcategories \(\mathcal A^k_\chi\) are given by perfect complexes on the disjoint union of normalizations of the strata of \(D\) in codimension \(k\). (In the simple normal crossing case the strata are already normal.) The case that \(D\) is simple normal crossing was proven in [\textit{D. Bergh} and \textit{O. M. Schnürer}, Adv. Math. 360, Article ID 106882, 39 p. (2020; Zbl 1453.14048)], where the assumption of simple normal crossing is not easy to find, cf.\ Remark 4.1 of the paper under review. The proof of the theorem proceeds from the simple normal crossing case. The idea is to choose an étale cover \(U\) of \(X\) such that the \((U, D|_U)\) is simple normal crossing. One then expresses \(\mathrm{Perf}(\sqrt[r]{(X,D)})\) as the limit of perfect copmlexes over the nerve of the étale cover \(\sqrt[r]{(U,D|_U)} \to \sqrt[r]{(X,D)}\). Using their own previous work [``Parabolic semi-orthogonal decmpositions and Kummer flat invariants of log schemes'', Preprint, \url{arXiv:1803:06398}] this allows the authors to construct an infinite semi-orthogonal decomposition on the infinite root stack \(\sqrt[\infty]{(X,D)}\). As an application of this the authors use the semi-orthogonal decomposition of \(\mathrm{Perf}(\sqrt[\infty]{(X,D)})\) to compute its additive invariants. In particular the Kummer flat \(K\)-theory of \((X,D)\) is known to be the algebraic \(K\)-theory of \(\mathrm{Perf}(\sqrt[\infty]{(X,D)})\), thus it decomposes as a direct sum over the indexing set of the semi-orthogonal decomposition.