[gmx-users] dg/dl unit

[Berk Hess]

>From: Michael Shirts <mrshirts at gmail.com>
>Reply-To: Discussion list for GROMACS users <gmx-users at gromacs.org>
>To: gmx-users at gromacs.org
>Subject: Re: [gmx-users] dg/dl unit
>Date: Thu, 3 Nov 2005 09:22:59 -0500
>
> > Can I ask a naive question? Since when is lambda a unit? lambda is a
> > unitless quantity and quantities usually do not appear in a unit. If you
> > want to do it completely correct, I think that dG/dl has the unit 
>kJmol-1.
> >
> > If you would do a slow growth simulation in which you modify lambda as a
> > function of time, then you probably write out dG/dl * dl/dtime and the
> > unit would be kJmol-1ps-1.
>
>Here's my naive question- is there a legitimate reason to do a slow
>growth simulation, other than in the context of averaging a number of
>slow growth simulations, a la Jarzynski?  It's a completely
>uncontrolled approximation, because the ensemble always lags behind
>the Hamiltonian in a way that cannot be corrected for.  Shouldn't we
>all be doing fixed-lambda sampling, or (when I get my fixes in),
>WHAM/Bennett acceptance ratio methods?
>
>I suppose one could do slower and slower growth until it converges,
>but for anything but the smallest changes (changing just a couple of
>atom types, no disappearance/appearance of atoms), its would be very
>inefficient.

I missed the last paragraph in the previous mail.
Time should never enter into dG/dlambda, the growth should
always be slower than the slowest relaxation time in the system.
In practice with fixed lambda simulations one always has a better
control of the error.

I have tried the Bennett acceptance ratio method and did not
find significant differences with plain integration of dG/dlambda.
Although my experience is that people often take far too little
lambda points (the new soft-core lambda power of 1 in 3.3 helps a bit 
tough).
Michael, what is your opinion on this?

Berk.