[gmx-users] Long range electrostatic potential

[Mark Abraham]

> As far as I know it can't be done for individual groups. My best guess
> is that it has something to do with the 'M' in 'PME'. :-)

Well... sort of...

The real-space charges have to get gridded (there's that M), Fourier
transformed, then the long-range term in PME is formed from truncated
summation in reciprocal space, then back-transformed and interpolated off
the M. I'd guess you could do a group-wise decomposition of this
algorithm, but that would require P^2
transform-sum-backtransform-interpolation stages for P groups. This would
make the scaling of group-wise-LR PME something like P^2*N*log(N), which
you'd never implement as your default algorithm, since N*log(N) is
available. Then maybe you'd stop and ask yourself whether these group-wise
interaction energies mean much anyway...

Mark

PS Afterthought, maybe it would be less than P^2*N*log(N), since N might
be smaller. If you care, read the PME papers and go figure :-)