Removing Online Exponential Net Search from Solovay-Kitaev

Abstract

The Solovay-Kitaev algorithm describes how to approximate, to arbitrary precision, a matrix in the special unitary group SU(d) using any fixed universal gate set. Although the algorithm scales as O(poly(log(1/$\epsilon$))), where $\epsilon$ is the maximum targeted approximation error, its running time depends exponentially on the qudit dimension d. This bad dependence can be traced to its explicit use of an $\epsilon$_0-net of size 2 $\Omega$(d^2) , which is queried O(poly(log(1/$\epsilon$))) times throughout the execution. For this reason, the standard Solovay-Kitaev theorem is usually stated for fixed d, with the base net and its lookup cost absorbed into the constants. We study the algorithmic problem in the variabledimension regime and show how to avoid searching an exponentially large precomputed net for each target unitary. In particular, we introduce the notion of a good exponential basis and show that such a basis can replace the usual depth-zero net-search routine. This yields a modification of the algorithm in which the use of an explicit net is fully moved to a preprocessing step. For instruction sets that already contain, or allow the efficient construction of, a good exponential basis, the resulting online synthesis algorithm is polynomial in d and polylogarithmic in 1/$\epsilon$. For arbitrary universal instruction sets, the exponential dependence on d^2 is not removed, but is isolated into a one-time additive preprocessing cost. Our technique uses differential-geometric methods to devise an integerized version of trotterization that replaces the depth-zero net query by a constructive local synthesis routine. The same framework also suggests possible extensions based on other discretized numerical integration schemes.