arXiv subjects

Philip Whittington

Publications and source records attributed to Philip Whittington.

7 recordsLinked to original sources

arXiv2026

Pack, Remove, Reserve -- Online Knapsack with Second Thoughts

We study the online proportional knapsack problem with two paid forms of recourse. Items arrive one by one and must be handled immediately, without knowledge of the future: an algorithm may pack an item $x$, reject it, or reserve it for possible later use at proportional cost $\alpha x$; additionally, it may at any time remove previously packed items, at proportional cost $\beta y$ for each removed item $y$. Reservation and removal have each been analyzed in isolation, but their combination raises a natural question: is the better of the two mechanisms always optimal on its own, or is there a region in the parameter space spanned by $\alpha$ and $\beta$ in which they genuinely enter into a symbiosis? So far, this question has only been answered for the special case of free removal ($\beta = 0$), leaving the vast majority of the parameter space unexplored. We close this gap, determining matching upper and lower bounds on the competitive ratio for every pair of cost parameters $(\alpha, \beta)$ and revealing three qualitatively different regimes. In some regions, reservation alone already achieves the optimal ratio; in others, removal alone does. However, most interestingly, in the heart of the parameter space lies a symbiosis region in which combining both mechanisms is strictly better than either one on its own. The optimal algorithm in the symbiosis region is a natural blend of the two known single-mechanism strategies: postponing commitment by reserving until a threshold is reached, then packing greedily and revising via removal.

cs.DS↗

arXiv2026

Tokenisation via Convex Relaxations

Tokenisation is an integral part of the current NLP pipeline. Current tokenisation algorithms such as BPE and Unigram are greedy algorithms -- they make locally optimal decisions without considering the resulting vocabulary as a whole. We instead formulate tokeniser construction as a linear program and solve it using convex optimisation tools, yielding a new algorithm we call ConvexTok. We find ConvexTok consistently improves intrinsic tokenisation metrics and the bits-per-byte (BpB) achieved by language models; it also improves downstream task performance, but less consistently. Furthermore, ConvexTok allows the user to certify how far their tokeniser is from optimal, with respect to a certain objective, via a lower bound, and we empirically find it to be within 1\% of optimal at common vocabulary sizes.

cs.CL↗

arXiv2025

Tokenisation over Bounded Alphabets is Hard

Recent works have shown that tokenisation is NP-complete. However, these works assume tokenisation is applied to inputs with unboundedly large alphabets -- an unrealistic assumption, given that in practice tokenisers operate over fixed-size alphabets, such as bytes or Unicode-characters. We close this gap by analysing tokenisation over bounded alphabets, considering two natural variants: bottomup tokenisation and direct tokenisation, where we must, respectively, select a sequence of merge operations or a vocabulary whose application optimally compresses a dataset. We prove that even with binary alphabets, both variants are not only NP-complete, but also APX-hard and thus admit no polynomial-time approximation scheme (unless P=NP). We further show that direct tokenisation remains NP-complete even when applied to unary alphabets. These results establish that the computational intractability of tokenisation is not an artifact of large alphabets or complex constructions, but a fundamental barrier. Overall, our results explain why current practical algorithms such as BPE and UnigramLM are heuristic, and point toward approximation algorithms being an important path going forward for tokenisation research.

cs.CL↗

arXiv2025

Time-Optimal $k$-Server

The time-optimal $k$-server problem minimizes the time spent serving all requests instead of the distances traveled. We give a lower bound of $2k-1$ on the competitive ratio of any deterministic online algorithm for this problem, which coincides with the best known upper bound on the competitive ratio achieved by the work-function algorithm for the classical $k$-server problem. We provide further lower bounds of $k+1$ for all Euclidean spaces and $k$ for uniform metric spaces. For the latter, we give a matching $k$-competitive deterministic algorithm. Our most technical result, proven by applying Yao's principle to a suitable instance distribution on a specifically constructed metric space, is a lower bound of $k+\mathcal{O}(\log k)$ that holds even for randomized algorithms, which contrasts with the best known lower bound for the classical problem that remains polylogarithmic. With this paper, we hope to initiate a further study of this natural yet neglected problem.

cs.DS↗

arXiv2024

Tokenisation is NP-Complete

In this work, we prove the NP-completeness of two variants of tokenisation, defined as the problem of compressing a dataset to at most $δ$ symbols by either finding a vocabulary directly (direct tokenisation), or selecting a sequence of merge operations (bottom-up tokenisation).

cs.DS↗

arXiv2024

Online String Attractors

In today's data-centric world, fast and effective compression of data is paramount. To measure success towards the second goal, Kempa and Prezza [STOC2018] introduce the string attractor, a combinatorial object unifying dictionary-based compression. Given a string $T \in Σ^n$, a string attractor ($k$-attractor) is a set of positions $Γ\subseteq [1,n]$, such that every distinct substring (of length at most $k$) has at least one occurrence that contains one of the selected positions. String attractors are shown to be approximated by and thus measure the quality of many important dictionary compression algorithms such as Lempel-Ziv 77, the Burrows-Wheeler transform, straight line programs, and macro schemes. In order to handle massive amounts of data, compression often has to be achieved in a streaming fashion. Thus, practically applied compression algorithms, such as Lempel-Ziv 77, have been extensively studied in an online setting. To the best of our knowledge, there has been no such work, and therefore are no theoretical underpinnings, for the string attractor problem. We introduce a natural online variant of both the $k$-attractor and the string attractor problem. First, we show that the Lempel-Ziv factorization corresponds to the best online algorithm for this problem, resulting in an upper bound of $\mathcal{O}(\log(n))$ on the competitive ratio. On the other hand, there are families of sparse strings which have constant-size optimal attractors, e.g., prefixes of the infinite Sturmian words and Thue-Morse words, which are created by iterative application of a morphism. We consider the most famous of these Sturmian words, the Fibonacci word, and show that any online algorithm has a cost growing with the length of the word, for a matching lower bound of $Ω(\log(n))$. For the online $k$-attractor problem, we show tight (strict) $k$-competitiveness.

cs.DS↗

arXiv2024

The 2-Attractor Problem is NP-Complete

A $k$-attractor is a combinatorial object unifying dictionary-based compression. It allows to compare the repetitiveness measures of different dictionary compressors such as Lempel-Ziv 77, the Burrows-Wheeler transform, straight line programs and macro schemes. For a string $T \in Σ^n$, the $k$-attractor is defined as a set of positions $Γ\subseteq [1,n]$, such that every distinct substring of length at most $k$ is covered by at least one of the selected positions. Thus, if a substring occurs multiple times in $T$, one position suffices to cover it. A 1-attractor is easily computed in linear time, while Kempa and Prezza [STOC 2018] have shown that for $k \geq 3$, it is NP-complete to compute the smallest $k$-attractor by a reduction from $k$-set cover. The main result of this paper answers the open question for the complexity of the 2-attractor problem, showing that the problem remains NP-complete. Kempa and Prezza's proof for $k \geq 3$ also reduces the 2-attractor problem to the 2-set cover problem, which is equivalent to edge cover, but that does not fully capture the complexity of the 2-attractor problem. For this reason, we extend edge cover by a color function on the edges, yielding the colorful edge cover problem. Any edge cover must then satisfy the additional constraint that each color is represented. This extension raises the complexity such that colorful edge cover becomes NP-complete while also more precisely modeling the 2-attractor problem. We obtain a reduction showing $k$-attractor to be NP-complete and APX-hard for any $k \geq 2$.

cs.CC↗