[gmx-users] Error estimation with TI

[David Mobley]

Maik,

I agree with David on this one. It has been shown that using your method can
lead to pretty significant hysteresis, and it's difficult to be sure whether
you're going slow enough to equilibrate at any lambda value. In addition,
it's harder to parallelize (you are running one or several long trajectories
rather than a bunch of different trajectories) and hard to extend (if you
run at a bunch of different lambda values, and you decide you need more
sampling at one of them, you can just extend that simulation; this way you
have to start over from scratch). Also, it's not clear, in my opinion, how
to properly estimate errors in this approach (I don't think there's a
straightforward way to compute the autocorrelation time, since the fact that
you are changing lambda as a function of time will affect apparent
correlations).

With regards to using blocking to try and figure out the error from your
data, I don't know that this would work very well. Ideally, your blocks
should be of a length comparable to the autocorrelation time; if you make
them too short, you'll tend to underestimate the error, and if you make them
too long, you'll tend to overestimate it. But if you can't figure out the
autocorrelation time it's hard to know whether blocking is working right, I
think.

To reiterate: My suggestion would be to avoid using slow growth. Run
simulations at fixed lambda values, which (a) works better, with no
hysteresis (and don't use the simulation at, say, lambda=0 as a starting
point for the simulation at lambda=0.1!), (b) is more parallelizable, and
(c) makes error analysis easier.

David


On 1/17/06, Maik Goette <mgoette at mpi-bpc.mpg.de> wrote:
>
> Dear all, dear David
>
> Again a question concerning error estimation.
> This time it's due to thermodynamic integration.
> When I simulate my system, let's say 5 ns, from lambda 0 to 1, I get my
> dG/dl values.
> At the moment, I am simulating both directions (0->1 & 1->0), integrate
> over both, and the difference between my back and forth sims is the
> error estimate.
> But I think, one should use a method, similar to FEP, to "guess" the
> error from the dg/dl-data, maybe also with generating blocks with regard
> to the autocorrelation.
> But somehow, the delta lambda has to be included, not?
>
> Regards
>
> --
> Maik Goette, Dipl. Biol.
> Max Planck Institute for Biophysical Chemistry
> Theoretical & computational biophysics department
> Am Fassberg 11
> 37077 Goettingen
> Germany
> Tel.  : ++49 551 201 2310
> Fax   : ++49 551 201 2302
> Email : mgoette[at]mpi-bpc.mpg.de
>          mgoette2[at]gwdg.de
> WWW   : http://www.mpibpc.gwdg.de/groups/grubmueller/
> _______________________________________________
> gmx-users mailing list
> gmx-users at gromacs.org
> http://www.gromacs.org/mailman/listinfo/gmx-users
> Please don't post (un)subscribe requests to the list. Use the
> www interface or send it to gmx-users-request at gromacs.org.
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://maillist.sys.kth.se/pipermail/gromacs.org_gmx-users/attachments/20060117/85bf4425/attachment.html>