[HN Gopher] Monte Carlo Long-Range Interacting System Simulations
       ___________________________________________________________________
        
       Monte Carlo Long-Range Interacting System Simulations
        
       Author : sandoze
       Score  : 39 points
       Date   : 2023-07-30 16:34 UTC (6 hours ago)
        
 (HTM) web link (www.uni-leipzig.de)
 (TXT) w3m dump (www.uni-leipzig.de)
        
       | conformist wrote:
       | Arxiv: https://arxiv.org/abs/2207.14670
       | 
       | If I understand correctly after skimming, one of the fundamental
       | ideas behind this appears to be similar to the well-known Fast
       | Multiple Method [1]. It's also a tree-based approach where far
       | away points are aggregated into larger chunks?
       | 
       | [1] https://en.m.wikipedia.org/wiki/Fast_multipole_method
        
         | physicsguy wrote:
         | It's exactly the same as the Barnes-Hut method from a quick
         | glance, there's nothing particularly new in this paper as far
         | as I can see. There are various papers which have merit on
         | releasing specific codes that support this, but people have
         | been using it in spin systems and even in micromagnetics for
         | years. I did my PhD in this area years ago and implemented it
         | in Monte Carlo and dynamical simulations at the time... and I
         | cited papers going back to the 80s and 90s when I wrote up my
         | thesis!
        
           | conformist wrote:
           | Ah, yes, and they do indeed cite Barnes-Hut, too. So the
           | (claimed) novelty seems to be how they connect it to Monte
           | Carlo steps.
        
           | radioactivist wrote:
           | While the spatial decomposition matches what is done in
           | Barnes-Hut, the details of the underlying algorithm are
           | somewhat different (they outline this in the introduction).
           | 
           | In particular, their scheme using exact bounds on energy
           | differences (evaluated using a hierarchical tree as in BH)
           | but in such a way that no approximation is being made. The
           | tree is evaluated to whatever depth is needed to decide
           | whether to accept/reject the MC move (which in worse case
           | could be a brute-force sum over the whole lattice/system) --
           | this is different I think than BH or other multipole inspired
           | methods (which have a kind of "truncation" or "tolerance"
           | parameter).
           | 
           | [This also works well with systems where update are local and
           | not global, which I think is a difference from some other
           | spatial partitioning schemes -- but I'm less conversant with
           | that aspect].
        
             | physicsguy wrote:
             | I read the full paper after posting my initial comment. Not
             | really - BH does use a "opening angle" parameter, but for
             | FMM you use an order of expansion which provides a strict
             | error bound in terms of accuracy which was derived in
             | Greengard and Rohklin's original paper on but have been
             | expanded to other potentials.
             | 
             | To me it's just a minor (but nonetheless interesting)
             | variation of an existing method; there have been many of
             | these. They don't compare it to other tree based methods in
             | performance in the paper which says to me that is merits
             | aren't really that clear... with long range potentials FP
             | inaccuracies mean none of this is exact anyway.
        
       | Fede_V wrote:
       | Isn't this a stochastic variant of Barnes-Hut:
       | https://en.wikipedia.org/wiki/Barnes%E2%80%93Hut_simulation
        
       | ckrapu wrote:
       | Any commentary on what this means for the wider world of MCMC?
        
         | tnecniv wrote:
         | I imagine a similar scheme can be used for general inference if
         | you have a way to cluster components of your model in a similar
         | fashion even if the parameters are not spatial. Perhaps your
         | model has some kind of natural tree structure.
        
       | punnerud wrote:
       | Duplicate? https://news.ycombinator.com/item?id=36930857
        
       | phyalow wrote:
       | Welcome to reinforcement learning
        
         | whatever1 wrote:
         | Really, communities should start talking to each other.
        
         | angus-mackaiver wrote:
         | Could you elaborate on that or point to any literature/websites
         | because I'm curious to learn about what these have in common.
         | Thank you!
        
       ___________________________________________________________________
       (page generated 2023-07-30 23:01 UTC)