As recently shown by Shirts and Chodera[107], all the
configurations produced by a REM simulation of
replicas, each
characterized by a distribution function
, can be effectively
used to obtain equilibrium averages for any target distribution
, using the so-called Multistate Bennett Acceptance Ratio
(MBAR) estimator, which is illustrated in the following.
In the ORAC REM implementation, the most general distribution
function for replica
is given by Eq. 5.15. Given that
for each replica
one has saved
configurations of the kind
, it can be easily shown that
 |
(5.24) |
Eq. 5.24 holds for any arbitrary bridge function
,
In particular, choosing[107]
 |
(5.25) |
Eq. 5.24 transforms as
 |
(5.26) |
where we have collapsed the two indices
, for the configurations,
and
, for the replicas in one single index
running on all
the configurations
produced by the REMD. Except for an arbitrary
multiplicative factor, Eq. 5.26 can be solved iteratively for
the partition function
.
At the beginning of the process, one sets
for all
. .
At the iteration
we have that
 |
(5.27) |
where the weights depend on the
calculated at the previous iteration
 |
(5.28) |
and we have used the definition of 5.15 for the replica
distribution in ORAC .
Once the
have been determined,
for
an arbitrary distribution
can be calculated using the configurations sampled
in the REMD simulation:
 |
(5.29) |
Setting for example
, where
is the
target distribution
(i.e. that of the replica 1) and
is an
arbitrary configurational property, we obtain
 |
(5.30) |
From Eq. 5.29, we have that
and that
. Substituting these
results into Eq. 5.30, we obtain
 |
(5.31) |
where the weights
for all sampled points in the REMD simulation are given by
 |
(5.32) |
In summary, using the configurational energies from all
replicas,
we first solve iteratively the system 5.26 for all
(except for a multiplicative factor), with
. In doing this, the weights
(including
) are also
determined. Finally configurational averages at, e.g., the target
distribution can be determined using all the REMD configurations
by means of Eq. 5.31.
procacci
2021-12-29