Using acor to compute error bars and autocorrelation

time
Jonathan Goodman
Courant Institute of Mathematical Sciences, NYU
[email protected]
May 7, 2009
Funded by DOE

The program acor estimates the errors and autocorrelation time of a time
series. It could be applied to real data or to data from Markov chain Monte
Carlo. It is intended to be used on fairly long time series. See the README
file for more details of the use.
The algorithm combines a direct calculation of the autocovariance function
C(t) with an averaging procedure. If the time series is X(t), the code first
computes and subtracts out the sample mean:
n
1X
X = X(t) ,
n t=1

and
X(t) ← X(t) − X .
Next it computes the first few values of the empirical auotcovariance function
n
1 X
C(t)
b = X(t + s)X(s) .
n − t s=1

(Warning, the actual code differs from this in a way that will be unimportant
for large n. I did this to make it run faster by having better cache performance.)
It estimates the diffusion coefficient
∞
X
D = C(t) ,
−∞

and the autocorrelation time

Y (t) = X(2t) + X(2t + 1) .

An estimate of D from the Y series can be used to estimate τ and D for the X
series. The difference is τ for the Y series should be half as long.
This process is applied recursively until the autocorrelation time is short
enough that the direct calculation works. This is τ ≤ 2 in the parameters that
the code has “out of the box”.

1
 The code comes with a tester acorTest that generates a time series with a
simple geometric autocovariance function. It creates a file of the format that
acor can read. The included makefile can create both: type make acorTest
or make acor. When I ran it on my system, I typed acorTest 4000000 ao to
make a file named ao with a time series of length four million. At this time,
the parameter a in acorTest was set to .9. Then I ran the program and got
output (edited for clarity, the output is all on one line):
jg> ./acor ao
sample mean = 5.0011, standard deviation = 0.00144837,
autocorrelation time = 19.1301, series length = 4000000

Warning: the program acorTest sets the seed for the random number generator
to a fixed value. This makes it so that if you generate two time series with the
same a you get the same time series. The exact answer is τ = 19 for a = .9.
There also is a C++ procedure called acor that receives the time series as
an argument.
Please let me know if this routine is helpful, and especially if it is not helpful.