Core Concepts
By combining a fast list labeling algorithm F and a reliable list labeling algorithm R, the embedding F⊳R achieves the best of both worlds: worst-case cost O(WR), amortized expected cost O(GF(x)) on any input sequence x, and lightly-amortized expected cost O(ER) overall.
Abstract
The list labeling problem is a fundamental data structure problem that involves storing a dynamic set of n elements in sorted order in an array of size (1+Θ(1))n, supporting both insertions and deletions. Over the past four decades, there has been extensive research on list labeling, leading to improvements in three key directions: low-latency (worst-case) bounds, high-throughput (expected) bounds, and adaptive bounds for important workloads.
However, these three directions of research have remained largely disjoint, as the techniques that enable progress in one direction often worsen the bounds in the others. The authors show that this tension is not fundamental by developing a new data structural technique called the embedding F⊳R, which combines any three list labeling solutions to cherry-pick the best worst-case, adaptive, and expected bounds from each.
The key idea is to hierarchically embed a fast algorithm F into a reliable algorithm R, with the F-emulator maintaining a simulated copy of F and gradually transforming the actual state to match it, while the R-shell handles buffering and consolidation of work. This approach overcomes the three major challenges: the deadweight problem, the input-interference problem, and the imbalance problem.
The authors prove that F⊳R simultaneously achieves the best of the three performance criteria: worst-case cost O(WR), amortized expected cost O(GF(x)) on any input sequence x, and lightly-amortized expected cost O(ER) overall. This result is shown to be composable, allowing the embedding to be applied recursively to combine three list labeling algorithms with different guarantees.