On the bottleneck stability of rank decompositions of multi-parameter persistence modules (Q6592197)

The authors first prove that the minimal rank decomposition is not stable with respect to the signed bottleneck dissimilarity. Then, they propose a new signed barcode that arises from a relative minimal projective resolution. The authors prove that their barcode has the desired stability properties. This is part of a bigger program that proposes using relative homological algebra in studying multiparameter persistence (the relative approach means that a subclass of the class of all exact sequences is fixed -- it is called an exact structure -- and all definitions, e.g., projectives, injectives, resolutions, etc., are with respect to only the exact sequences from this class).\N\NMain contributions:\N\begin{itemize}\N\item It is not possible to make the signed bottleneck dissimilarity between minimal rank decompositions by rectangles stable in any sense. I.e., there is no function \(f\) such that \(\hat{d}_B(\mathfrak{B}^\square(M), \mathfrak{B}^\square(N)) \leq f(d_I(M, N))\) for all finitely presented persistence modules \(M\), \(N\). Here \(d_I\) is the interleaving distance between the modules, and \(\hat{d}_B\) is the signed bottleneck dissimilarity between their minimal rank decomposition by rectangles \(\mathfrak{B}^\square(\cdot)\).\N\item Definition of a new exact structure called rank exact and of the rank exact decomposition. There are two differences: a) switching from rectangles to hook modules and b) requirement that the signed barcode comes from a minimal projective resolution relative to the rank exact structure.\N\item Bottleneck stability result for (ordinary, that is, not signed) decompositions by hook modules: if \(P\) and \(Q\) can be decomposed as a direct sum of at most countably many hook modules and \(P\) and \(Q\) are \(\varepsilon\)-interleaved, then there exists a \((2n-1)\varepsilon\)-matching between their barcodes \(\mathfrak{B}(P)\) and \(\mathfrak{B}(Q)\).\N\item Bottleneck dissimilarity for the rank exact decomposition of signed barcodes is stable: it is upper-bounded by \((2n-1)^2 d_I(M, N)\).\N\item Global dimension of the rank exact structure is \(2n-2\). In particular, for \(n=1\), the authors obtain the decomposition theorem for finitely presentable \(1\)-parameter persistence modules.\N\item A no-go theorem for distances between signed barcodes. Roughly speaking, there does not exist a non-trivial dissimilarity that satisfies some natural properties and the triangle inequality.\N\end{itemize}