Aleksandrs Belovs, Stacey Jeffery (Feb 14 2025).

Abstract: Given an algorithm that outputs the correct answer with bounded error, say $1/3$, it is sometimes desirable to reduce this error to some arbitrarily small $\varepsilon$ -- for example, if one wants to call the algorithm many times as a subroutine. The usual method, for both quantum and randomized algorithms, is a procedure called majority voting, which incurs a multiplicative overhead of $O(\log\frac{1}{\varepsilon})$ from calling the algorithm this many times. A recent paper introduced a model of quantum computation called \emphtransducers, and showed how to reduce the ``error'' of a transducer arbitrarily with only constant overhead, using a construction analogous to majority voting called \emphpurification. Even error-free transducers map to bounded-error quantum algorithms, so this does not let you reduce algorithmic error for free, but it does allow bounded-error quantum algorithms to be composed without incurring log factors. In this paper, we present a new highly simplified construction of a purifier, that can be understood as a weighted walk on a line similar to a random walk interpretation of majority voting. In addition to providing a new perspective that is easier to contrast with majority voting, our purifier has exponentially better space complexity than the previous one, and quadratically better dependence on the soundness-completeness gap of the algorithm being purified. Our new purifier has nearly optimal query complexity, even down to the constant, which matters when one composes quantum algorithms to super-constant depth.

Arxiv: https://arxiv.org/abs/2502.09249