Re: DesNumNgb

From: Volker Springel <volker_at_MPA-Garching.MPG.DE>
Date: Wed, 01 Mar 2006 22:35:07 +0100

Yves Revaz wrote:
>
> Dear All,
>
> In the new implementation of sph, the smoothing length is not
> constraints by the number
> of neighbours but by imposing that the volume defined by the smoothing
> length contains
> a constant mass (eq. 6 of Springel 2005).
> However, when running gadget (with DesNumNgb=50), I get the typical
> following values
> out of the routine "density_evaluate" :
>
> numngb_inbox = 12 # numbers of neighbours in a box of
> dimension 2h x 2h x 2h
> numngb = 10 # numbers of neighbours in a sphere of
> radius h weighted_numngb = 49.936302 # weighted number of
> neighbours
>

Hi Yves,

I don't really think that these are typical numbers. Please note that
density_evaluate only looks for the neighbours on the local processor, i.e.
there can be additional contributions on other domains that are added in
with separate calls of this function on other processors.

Nevertheless, the number of actual neighbours inside the radius h can be
smaller than the value of DesNumNgb in certain cases (in general it scatters
around this value while weighted_numngb is always close to DesNumNgb). This
happens in particular for particles at the centre of density peaks.

The most extreme cases of this behaviour can happen if one has a very small
clump of a few particles (~10 or smaller) embedded in a sea of a particle
distribution of much lower density. Then the kernel radius h for the
particles of such a clump will be much larger than the radius of the clump,
letting several of the neighbours contribute with the full kernel weight at
zero lag (r/h<<1). This means they contribute strongly to the weighted
neighboor sum, making the absolute number of neighbours quite a bit smaller
than DesNumNgb.

However, I don't really see a severe deficit due to this. In this situation
one probably doesn't want that the clump particles couple strongly to
particles of the low density phase that are far away - after all, the
clumps have a fairly sharp "edge" in such a case. Also, low density
particles will still couple in the force equation to the clump particles
(their h will overlap with the clump particles from a larger volume than the
other way round) in a one-sided way. Finally, while in the more traditional
scheme to determine h one would set a slightly higher h in this case for the
clump particles, the additional neighbours that are seen would all lie at
the edge of the smoothing kernel, and would then contribute little to the
SPH sums. So I don't think that there is a significant difference in the
dynamics in this case between the two schemes.

Volker


> I have been first surprised by the fact that from 10 effective
> neighbours we finally obtain
> 49.936302 weighted number of neighbours ! However, this may be
> understood from the
> kernel definition and the weighted number of neighbours definition :
>
> wk ~ 8/(pi*h3) # for r/h<<1
>
> weighted_numngb ~= 4/3*pi*h3 *rho * / m ~= 4/3*pi*h3 *
> numngb*8/(pi*h3)*m / m ~= 10.66 numngb
>
>
> Thus, when setting DesNumNgb we do not really set the effective number
> of neighbours,
> but rather ~10 times (depending on the distribution around the particle)
> the effective number
> of neighbours. As a consequence, with DesNumNgb=50, only few particles
> are taking into account
> to compute the hydrodynamical forces (about 10 in the latter case),
> while in classical sph one expect
> to have at least 30 neighbours.
>
> Could someone comment on that ?
>
> Thanks.
>
>
>
Received on 2006-03-01 22:35:19

This archive was generated by hypermail 2.3.0 : 2022-09-01 14:03:41 CEST