Pair merge compression hard to approximate on binary and unary strings
Pair-Merge Tokenisation Is Hard to Approximate on Binary Strings and Strongly NP-Hard on Unary Datasets
Data Structures and AlgorithmsComputational ComplexityComputation and Language
Summary
The paper studies the difficulty of compressing strings by merging pairs of symbols, a common approach in text tokenisation. The authors show that finding near-optimal merges efficiently is very hard—even for simple binary strings—unless widely believed complexity assumptions fail. They also prove that on single-letter (unary) datasets, the problem of finding optimal merges is strongly NP-complete, meaning it remains hard even when numeric values are small. This clarifies the computational limits of some popular tokenisation strategies.
What this means in practice
- •For compression algorithm designers: Understand the inherent limits on efficiently approximating pair-merge tokenisation methods when compressing binary strings to guide algorithm development.
- •For software engineers working on text tokenisation: Recognize that achieving near-optimal token compression with fixed merges is computationally infeasible for certain inputs, influencing design of heuristic tokenisers.
A theory result. No direct application yet.
Authors
Kundan Baliga, Aiden Xie, Aditya Byju
Abstract
Tokenisers such as Byte-Pair Encoding (BPE) shorten text by repeatedly merging an adjacent pair of symbols into a new one; fewer tokens mean less computation. We ask whether an efficient algorithm can choose $k$ merges that leave a string of length $n$ almost as short as possible; doing nothing is already within a factor $n$. We prove that, unless P = NP, no polynomial-time algorithm comes within a factor $c\,n/\log n$, for some constant $c>0$, already for one binary string and for BPE-style merges that replace every occurrence. The key idea is that $k$ merges form a grammar with $k$ rules: a vertex-cover gap becomes a grammar-size gap, which we amplify with relabelled copies, encode in binary and realise with explicit merges. For merges that may skip occurrences, a simple algorithm achieves $O(n/\log\log n)$, and no Karp reduction can prove hardness beyond order $n/\log n$ unless coNP $\subseteq$ NP/poly. Over a one-letter alphabet, both tokenisation problems left open by Kastreva et al. are strongly NP-complete on datasets with repeated strings.