Paper 2023/261

A Greedy Global Framework for Lattice Reduction Using Deep Insertions

Abstract

LLL-style lattice reduction algorithms iteratively employ size reduction and reordering on ordered basis vectors to find progressively shorter, more orthogonal vectors. DeepLLL reorders the basis through deep insertions, yielding much shorter vectors than LLL. DeepLLL was introduced alongside BKZ, however, the latter has received greater attention and has emerged as the state-of-the-art. We first show that LLL-style algorithms work with a designated measure of basis quality and iteratively improves it; specifically, DeepLLL improves a sublattice measure based on the generalised Lovász condition. We then introduce a new generic framework X-GG for lattice reduction algorithms that work with a measure X of basis quality. X-GG globally searches for deep insertions that minimise X in each iteration. We instantiate the framework with two quality measures — basis potential (Pot) and squared sum (SS) — both of which have corresponding DeepLLL algorithms. We prove polynomial runtimes for our X-GG algorithms and also prove their output to be X-DeepLLL reduced. Our experiments on non-preprocessed bases show that X-GG produces better quality outputs whilst being much faster than the corresponding DeepLLL algorithms. We also compare SS-GG and the FPLLL implementation of BKZ with LLL-preprocessed bases. In small dimensions (40 to 210), SS-GG is significantly faster than BKZ with block sizes 8 to 12, while simultaneously also providing better output quality in most cases. In higher dimensions (250 and beyond), by varying the threshold $\delta$ for deep insertion, SS-GG offers new trade-offs between the output quality and runtime. On the one hand, it provides significantly better runtime than BKZ-5 with worse output quality; on the other hand, it is significantly faster than BKZ-21 while providing increasingly better output quality after around dimension 350.