The Good, the Bad, Algorithmic Noise Tolerance (Ant), the Ugly

Abstract

Computational units implemented on nanoscale physical substrates are susceptible to errors that can be catastrophic if not mitigated. Statistical error compensation techniques have become prevalent to safeguard computational units against such hardware-failures. Algorithmic Noise Tolerance (ANT) is one such technique that utilizes a low-fidelity replica unit to detect and bypass such failures occurring within the primary (main) computational unit. Connections between ANT and the binary hypothesis testing as well as the information theoretic CEO problem have been explored for sub-exponential error profiles, quadratic and logarithmic distortion functions. However, there exist fundamental performance limits of ANT approach even without such model-dependent restrictions. The purpose of this paper is to explore fidelity-dependent conditions that are universal over the statistical properties of the computational units under which, the overall performance of ANT is arbitrarily close to the fundamental limits.