[gmx-users] What is the autocorrelation time

Christopher Neale chris.neale at mail.utoronto.ca
Sat May 19 03:38:55 CEST 2012

Thank you very much Patrick. This is exactly what I was looking for.  Based on what you sent, I'm going to go with the following text in my work:

The value of  tau_int is approximately half of the amount of time required, on average, to obtain a statistically independent (decorrelated) sample [ref Janke].

I really appreciate it!

Thank you for your help.

-- original message --

Hi Chris,I understand your question, this autocorrelation time puzzled me for a 
long time as well. Not far from the interpretation you give, Scott 
Feller defines it (http://dx.doi.org/10.1007/978-1-59745-519-0_7) as the 
time a given observable takes to lose the memory of its previous state, 
or in other words the time it takes to relax (that's why it's sometimes 
called relaxation time). He also discusses it as a tool to choose the 
block size for calculating an error estimate of an observable (one 
single simulation can be used as independant samples if each block size 
is >> autocorrelation time).
We also had a nice discussion some years ago on the mailing list on free 
energy calculation and error estimate: 
http://lists.gromacs.org/pipermail/gmx-users/2007-May/027281.html. John 
Chodera pointed me to a useful article from Wolfhard Janke (the link in 
the discussion is broken, here's the new one: 
http://www2.fz-juelich.de/nic-series/volume10/janke2.pdf). There you'll 
find a rigorous mathematical definition of autocorrelation time. Quoting 
this paper "This shows more clearly that only every 2
iterations the measurements are approximately uncorrelated and gives a 
better idea of the relevant effective size of the statistical sample" 
(tau_int is the integrated autocorrelation time; as you said the 
autocorrelation function is usually a single exponential, but sometimes 
it's more complex and one needs to evaluate it by integration of the 
autocorrelation function).
After all these considerations, the autocorrelation time can be seen as 
a tool to assess the time that is needed to have a good estimate of an 
observable: the simulation must be many many times longer than the 
autocorrelation time. And sometimes it's directly related to 
experimental observables (i.e. NMR relaxation experiments).
Hope it's useful,


Le 16/05/2012 23:39, Christopher Neale a écrit :
> Thank you Stephane.
> Unfortunately, neither of those links contains the information that I am
> seeking. Those links contain some example plots of autocorrelation
> functions including a discussion of time-spans over which the example
> time-series is autocorrelated and when it is not, but neither link
> defines the (exponential or integral) autocorrelation time except to
> show a plot and indicate when it is non-zero and when it fluctuates
> about zero.
> For example, I already know that the autocorrelation time describes the
> exponential decay of the correlation and that two values drawn from the
> same simulation are statistically independent if they are separated by a
> sufficient number of (accurate) autocorrelation times, but this
> information is not exactly a definition of the autocorrelation time.
> I am hoping to find a definition of the autocorrelation time in terms of
> the probability of drawing uncorrelated samples, although any complete
> definition will do.
> If anybody else has the time, I would appreciate it.
> Thank you,
> Chris.
> -- original message --
> Probably these links give you simple and clear response for your
> question
> http://idlastro.gsfc.nasa.gov/idl_html_help/Time-Series_Analysis.html
> and http://www.statsoft.com/textbook/time-series-analysis/ HTH Stephane


More information about the gromacs.org_gmx-users mailing list