Locality and Statistical Error Reduction on Correlation Functions

We propose a multilevel Monte-Carlo scheme, applicable to local actions, which is expected to reduce statistical errors on correlation functions. We give general arguments to show how the efficiency and parameters of the algorithm are determined by the low-energy spectrum. As an application, we measure the euclidean-time correlation of pairs of Wilson loops in SU(3) pure gauge theory with constant relative errors. In this case the ratio of the new method's efficiency to the standard one increases as exp{m_0t/2}, where m_0 is the mass gap and t the time separation.