Dynamic Time Warping Under Translation: Approximation Guided by Space-Filling Curves

Abstract: The Dynamic Time Warping (DTW) distance is a popular measure of similarity for a variety of sequence data. For comparing polygonal curves $\pi, \sigma$ in $\mathbb{R}<sup>d$, it provides a robust, outlier-insensitive alternative to the Fr\'echet distance. However, like the Fr\'echet distance, the DTW distance is not invariant under translations. Can we efficiently optimize the DTW distance of $\pi$ and $\sigma$ under arbitrary translations, to compare the curves' shape irrespective of their absolute location? There are surprisingly few works in this direction, which may be due to its computational intricacy: For the Euclidean norm, this problem contains as a special case the geometric median problem, which provably admits no exact algebraic algorithm (that is, no algorithm using only addition, multiplication, and $k$-th roots). We thus investigate exact algorithms for non-Euclidean norms as well as approximation algorithms for the Euclidean norm: - For the $L_1$ norm in $\mathbb{R}<sup>d$, we provide an $\mathcal{O}(n<sup>{2(d+1)})$-time algorithm, i.e., an exact polynomial-time algorithm for constant $d$. Here and below, $n$ bounds the curves' complexities. - For the Euclidean norm in $\mathbb{R}<sup>2$, we show that a simple problem-specific insight leads to a $(1+\varepsilon)$-approximation in time $\mathcal{O}(n<sup>3/\varepsilon<sup>2)$. We then show how to obtain a subcubic $$\widetilde{\mathcal{O}}(n<sup>{2.5}/\varepsilon<sup>2)$$ time algorithm with significant new ideas; this time comes close to the well-known quadratic time barrier for computing DTW for fixed translations. Technically, the algorithm is obtained by speeding up repeated DTW distance estimations using a dynamic data structure for maintaining shortest paths in weighted planar digraphs. Crucially, we show how to traverse a candidate set of translations using space-filling curves in a way that incurs only few updates to the data structure.