Reconstruction in one dimension from unlabeled Euclidean lengths

Authors: Robert Connelly, Steven J. Gortler, and Louis Theran
Preprint: 2007.06550, 2020
Full text: arXiv

Let \(G\) be a \(3\)-connected graph with \(n\) vertices and \(m\) edges. Let \(\mathbf{p}\) be a randomly chosen mapping of these \(n\) vertices to the integer range \([1..2^b]\) for \(b\ge m2\). Let \(\mathbf{l}\) be the vector of \(m\) Euclidean lengths of \(G\)’s edges under \(\mathbf{p}\). In this paper, we show that, WHP over \(\mathbf{p}\), we can efficiently reconstruct both \(G\) and \(\mathbf{p}\) from \(\mathbf{l}\). In contrast to this average case complexity, this reconstruction problem is NP-HARD in the worst case. In fact, even the labeled version of this problem (reconstructing \(\mathbf{p}\) given both \(G\) and \(\mathbf{l}\)) is NP-HARD. We also show that our results stand in the presence of small amounts of error in \(\mathbf{l}\), and in the real setting with approximate length measurements.

Our method is based on older ideas that apply lattice reduction to solve certain SUBSET-SUM problems, WHP. We also rely on an algorithm of Seymour that can efficiently reconstruct a graph given an independence oracle for its matroid.