Understanding Robust Catalytic Computing

2026-05-10

AI Summary

This work investigates how to tolerate more flexible reset errors in catalytic computation to enable its robust application in space-bounded algorithms. We introduce three novel fault-tolerant models—stochastic total corruption, stochastic expectation-bounded error, and deterministic initial-state error with bounded expectation—and provide the first systematic characterization of their computational power. Through complexity-theoretic analysis, randomization and derandomization techniques, and space–time reductions, we nearly completely characterize these models in general settings and within logarithmic space and polynomial time. Under standard derandomization assumptions, we further show that almost all catalytic classes collapse within logarithmic space, revealing deep connections between fault-tolerant catalytic computation and classical complexity classes.

Abstract

Catalytic computing concerns space bounded computation which starts with memory full of data that have to be restored by the end of the computation. Lossy catalytic computing, defined by Gupta et al. (2024) and fully characterized by Folkertsma et al. (ITCS 2025), is the study of allowing a small number of errors when resetting the catalytic tape at the end of a computation. Such a notion is useful when considering the robust use of catalytic techniques in the study of ordinary space-bounded algorithms. To that end however, defining and characterizing less strict notions of error was left open by Folkertsma et al. (ITCS 2025) and other works such as Mertz (B. EATCS, 2023). We expand the definition of possible resetting error in three natural ways: 1. randomized catalytic computation which can completely destroy the catalytic tape with some probability over the randomness 2. randomized catalytic computation which makes a bounded number of errors in expectation over the randomness 3. deterministic catalytic computation which makes a bounded number of errors in expectation over the initial catalytic tape itself We show a near complete characterization of the above models, both in the general case and in the logspace polynomial-time regime, by showing equivalences either between one another, to errorless catalytic space models, or to standard time or space complexity classes. Under a derandomization assumption, we show a near full collapse of all existing catalytic classes in the logspace regime.

Problem

• catalytic computing
• space-bounded computation
• resetting error
• randomized computation
• derandomization

Innovation

• catalytic computing
• space-bounded computation
• randomized algorithms
• derandomization
• computational complexity