Lightnews — Scholar-powered news

LightNews

Sam Power

@spmontecarlo.bsky.social

2.3K followers 1.9K following 1.1K posts

Lecturer in Maths & Stats at Bristol. Interested in probabilistic + numerical computation, statistical modelling + inference. (he / him). Homepage: https://sites.google.com/view/sp-monte-carlo Seminar: https://sites.google.com/view/monte-carlo-semina

sites.google.com

Posts Media Videos Starter Packs

Sam Power @spmontecarlo.bsky.social · 1d

Quite a funny MO comment / discussion, for me:

Sam Power @spmontecarlo.bsky.social · 2d

absolutely fascinating (standard Steinerberger fare); be sure to open this!

arxiv.org/abs/2510.11571
'Robust Online Sampling from Possibly Moving Target Distributions'
- François Clément, Stefan Steinerberger

Sam Power @spmontecarlo.bsky.social · 2d

from source (not sure how informative):

Sam Power @spmontecarlo.bsky.social · 2d

This one evaded my lists; quite cool stuff!

arxiv.org/abs/2507.00272
'Iteratively Saturated Kalman Filtering'
- Alan Yang, Stephen Boyd

Sam Power @spmontecarlo.bsky.social · 3d

www.arxiv.org/abs/2510.04460
'Perspectives on Stochastic Localization'
- Bobby Shi, Kevin Tian, Matthew S. Zhang

Sam Power @spmontecarlo.bsky.social · 3d

arxiv.org/abs/2510.04088
'Offline Reinforcement Learning in Large State Spaces: Algorithms and Guarantees'
- Nan Jiang, Tengyang Xie

arxiv.org/abs/2510.04473
'Introduction to Interpolation-Based Optimization'
- Lindon Roberts

Sam Power @spmontecarlo.bsky.social · 3d

lots of nice review-type works up last week:

arxiv.org/abs/2510.04932
'MCMC for State Space models'
- Paul Fearnhead, Chris Sherlock

arxiv.org/abs/2510.03951
'Dimension dependence of critical phenomena in long-range percolation'
- Tom Hutchcroft

Sam Power @spmontecarlo.bsky.social · 4d

very convenient that { function, gradient, Hessian } come in such neat alphabetical order

Sam Power @spmontecarlo.bsky.social · 4d

... since there is plenty of value in identifying subsets of the problem space for which better solutions are available, rather than complaining too much about the fact that these solutions aren't universal (which was anyways a big ask!).

Sam Power @spmontecarlo.bsky.social · 4d

I used to pull away from this a bit, on the basis that some of the historical motivation for ABC and co. involved models which are far worse-behaved than this, and so these new proposals wouldn't help much with these original problems. With the benefit of hindsight, this was a bit misguided ...

Sam Power @spmontecarlo.bsky.social · 4d

Hopefully it's clear that "culturally wrong" is not really a criticism here. Anyways, I've had similar reactions in the context of simulation-based inference (still ABC to me, but what can one do?), where you often see new methodology which assumes that the forward model is differentiable.

Sam Power @spmontecarlo.bsky.social · 4d

... and they are often quite loud about how they don't want to have to deal with derivatives through their simulation pipeline, and so on. So in this regard, the AD-compatibility of the EnKF is clearly true in an objective sense, but almost feels "culturally wrong" if this is your main context.

Sam Power @spmontecarlo.bsky.social · 4d

Not on this particular occasion, but certainly in the past!

Sam Power @spmontecarlo.bsky.social · 4d

Yeah, I'm used to a similar story (though I wonder if the story is the same for x-derivatives and θ-derivatives, which is not completely obvious). I roughly believe the story, but equally, there is such strong soft information as to just how useful gradients are, so it's hard to ignore indefinitely.

Sam Power @spmontecarlo.bsky.social · 4d

I had some chats lately with a couple of cosmologists, and it was really the same story at most links of the chain (gradients not quite available, supercomputers involved, emulator-curious), with the quasi-distinction that they were pretty sure that gradients would help if they were available.

Sam Power @spmontecarlo.bsky.social · 4d

... after how many years of "re-doing the numerics would be annoying, but clearly useful" do you (collectively) cave in, and make the leap? Anyways, as you say, there are other reasons why it's not quite so simple.

Sam Power @spmontecarlo.bsky.social · 4d

I very much agree with the spirit of this comment; most of my other thoughts relate to some other points which have come up here (e.g. memory as the real bottleneck in this and other applications). What I found myself reflecting on was that if the only bottleneck had been effort (roughly), then ...

Sam Power @spmontecarlo.bsky.social · 4d

At dinner, but in the mean time, I could maybe advise to look up Marc Bocquet (who maybe has some nice slides IIRC).

Sam Power @spmontecarlo.bsky.social · 4d

... and they are often quite loud about how they don't want to have to deal with derivatives through their simulation pipeline, and so on. So in this regard, the AD-compatibility of the EnKF is clearly true in an objective sense, but almost feels "culturally wrong" if this is your main context.

Sam Power @spmontecarlo.bsky.social · 4d

A side point, which came up through some of this reading: some works will make the remark that the Ensemble Kalman Filter is nicely compatible with automatic differentiation. This is true on its face, but always takes a minute for me, given that the EnKF is huge in numerical weather prediction ...

Sam Power @spmontecarlo.bsky.social · 4d

I got back onto this topic via arxiv.org/abs/2312.05910; an early paper is proceedings.mlr.press/v9/turner10a..., and a couple of intermediate works which I found useful are proceedings.neurips.cc/paper_files/... and proceedings.mlr.press/v97/ialongo1....

Sam Power @spmontecarlo.bsky.social · 4d

A glimpse behind the curtain. Yet more care is needed when looking at things like variational inference with inducing variable approaches (which also occupy an interesting space vis-à-vis conditioning, randomness, etc.).

Sam Power @spmontecarlo.bsky.social · 4d

I got back onto this topic via arxiv.org/abs/2312.05910; an early paper is proceedings.mlr.press/v9/turner10a..., and a couple of intermediate works which I found useful are proceedings.neurips.cc/paper_files/... and proceedings.mlr.press/v97/ialongo1....

Sam Power @spmontecarlo.bsky.social · 4d

There are also some nice subtleties around the difference between conditioning on a quantity and fixing a quantity, which seems to often be a slightly messy point when talking about random functions. Anyways, I find it all quite satisfying to resolve.

Sam Power @spmontecarlo.bsky.social · 4d

This weekend, I have been doing a bit of reading about Gaussian Process State Space Models (GP-SSMs). Aside from being interesting on the modelling and inference sides, they are a remarkably good exercise for what it really means to write down a joint distribution.

Sam Power @spmontecarlo.bsky.social · 4d

if the list of speakers looks unfamiliar to you, then I can recommend { Tessera, Salez, van Handel, Magee, Sahasrabudhe } as some who I can vouch for having given good talks, one way or another

Sam Power @spmontecarlo.bsky.social · 4d

very cool workshop last week - www.newton.ac.uk/event/oggw03/ - lots of talks from good speakers on interesting works. keen to watch my way through them when I get a moment!