Space-Efficient Quantum Error Reduction without log Factors

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 $\varepsilon$ -- e.g., if one wants to call the algorithm many times as a subroutine. The usual method, for both quantum and randomized algorithms, is majority voting, which incurs a multiplicative overhead of $O(\log\frac{1}{\varepsilon})$ from calling the algorithm this many times. Transducers are a recently introduced model of quantum computation, and it is possible to reduce the ``error'' of a transducer arbitrarily with only constant overhead, using a construction analogous to majority voting called purification. 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 purifier, that can be understood as a weighted walk on a line similar to a random walk interpretation of majority voting. Our purifier has much smaller space and time complexity than the previous one. Indeed, it only uses one additional counter, and only performs two increment and two decrement operations on each iteration. It also has quadratically better dependence on the soundness-completeness gap of the algorithm being purified. We prove that our purifier has optimal query complexity, even down to the constant, which matters when one composes quantum algorithms to super-constant depth. Purifiers can be seen as a way of turning a ``Monte Carlo'' quantum algorithm into a ``Las Vegas'' quantum algorithm -- a process for which there is no classical analogue. Our simplified construction sheds light on this strange quantum phenomenon, and could have implications for the complexity of composed quantum algorithms.