Modern ML predictions models are surprisingly accurate in practice and incorporating their power into algorithms has led to a new research direction. Algorithms with predictions have already been used to improve on worst-case optimal bounds for online problems and for static graph problems. With this work, we initiate the study of the complexity of {\em data structures with predictions}, with an emphasis on dynamic graph problems. Unlike the independent work of v.d.~Brand et al.~[arXiv:2307.09961] that aims at upper bounds, our investigation is focused on establishing conditional fine-grained lower bounds for various notions of predictions. Our lower bounds are conditioned on the Online Matrix Vector (OMv) hypothesis. First we show that a prediction-based algorithm for OMv provides a smooth transition between the known bounds, for the offline and the online setting, and then show that this algorithm is essentially optimal under the OMv hypothesis. Further, we introduce and study four different kinds of predictions. (1) For {\em $\varepsilon$-accurate predictions}, where $\varepsilon \in (0,1)$, we show that any lower bound from the non-prediction setting carries over, reduced by a factor of $1-\varepsilon$. (2) For {\em $L$-list accurate predictions}, we show that one can efficiently compute a $(1/L)$-accurate prediction from an $L$-list accurate prediction. (3) For {\em bounded delay predictions} and {\em bounded delay predictions with outliers}, we show that a lower bound from the non-prediction setting carries over, if the reduction fulfills a certain reordering condition (which is fulfilled by many reductions from OMv for dynamic graph problems). This is demonstrated by showing lower and almost tight upper bounds for a concrete, dynamic graph problem, called $\# s \textrm{-} \triangle$, where the number of triangles that contain a fixed vertex $s$ must be reported.
翻译:暂无翻译