🚀 go-pugleaf

Here is the intro part of a paper by some other physicists.

\centerline{\bf Hidden Errors in Simulations and the Quality 
                of Pseudorandom Numbers}

by I. Vattulainen$^1$, K. Kankaala$^{1,2}$, J. Saarinen$^1$, and
T. Ala-Nissila$^3$.

In a recent letter Ferrenberg {\it et al.} (FLW) [1] present 
intriguing results arising from combinations of some random 
number generators and Monte Carlo acceleration algorithms. 
In particular, they observe systematic errors when the 
Wolff algorithm [2] is used at $T_c$ in the $2-d$ Ising model.
As an explanation of this, they propose that subtle correlations
arise in the random number sequences, in the sense that the
higher order bits of the random numbers are correlated.

We have recently carried out extensive statistical,
bit level, and visual tests for several commonly used
pseudorandom number generators in physics applications [3].
Two of these were used in Ref. 1, namely
R250 [4] and CONG [5].
Using the Wolff algorithm FLW discovered problems
with R250, but not with CONG. Thus, our test results bear
direct relevance to the existence of possible problems, and
differences between these two algorithms.

-- 
B. Narasimhan                             naras@cda.mrs.umn.edu
Division of Science and Math.
The University of Minnesota at Morris
Morris, MN 56267