iiinova — where your digital self keeps evolving

Adrien Corenflos @adriencorenflos.bsky.social · 26/09/2026

I have lived here for almost a year I'll have you know.

130

Adrien Corenflos @adriencorenflos.bsky.social · 23/01/2026

According to AI, the exam I've written for my students "does not spark any joy".

150

Adrien Corenflos @adriencorenflos.bsky.social · 20/10/2025

We believe there are many nice and important consequences to the construction. One we find particularly nice is that it gives diagnostics of convergence for MCMC algorithms in *any* f-divergence, e.g. the total variation. This is because we essentially have f-div ≤ Σₙ f(2N Wⁿ) for "correct" weights.

200

Adrien Corenflos @adriencorenflos.bsky.social · 20/10/2025

We prove that this sequence of weights is consistent (in the Monte Carlo sense: it gives corrected weighted averages) and that it converges to the constant weights 1/(2N): this is what you'd have if your chain was at stationarity.

100

Adrien Corenflos @adriencorenflos.bsky.social · 20/10/2025

This is what the original illustration showed. You make pairs, wait until they meet, at which point you equalise their weights: both particles are one and the same, so it doesn't matter! Once that's done, you just need to swap them (shuffle) with the other pairs that met, and the cycle continues.

100

Adrien Corenflos @adriencorenflos.bsky.social · 20/10/2025

The idea is to make several Markov chains interact and exchange information via their weights. To do so we rely on *couplings* of chains X₀, X₁, ..., Xₙ, which are computable joint distributions the chains satifying P(Xₙ = Yₙ | Xₙ₋₁, Yₙ₋₁) > 0. sites.google.com/site/pierrej...

100

Adrien Corenflos @adriencorenflos.bsky.social · 20/10/2025

A little self promotion: Hai-Dang Dau (NUS) and I recently released this pre-print, which I'm not half proud of. arxiv.org/abs/2510.07559 The main problem we solve in it is to construct importance weights for Markov chain Monte Carlo. We achieve it via a method we call harmonization by coupling.

1163

Adrien Corenflos @adriencorenflos.bsky.social · 16/10/2025

I was complaining about the same thing so my friends got me a katana!

010

Adrien Corenflos @adriencorenflos.bsky.social · 13/07/2025

I'm not that much into painful Gaussian algebra, but to each their own kink I suppose.

110

Adrien Corenflos @adriencorenflos.bsky.social · 05/07/2025

I'm a big fan of the fuse in the plug, it's lacking a critical ergonomic feature.

Diagram showing the slight angle formed by the europlug so that it stays put.

000

Adrien Corenflos @adriencorenflos.bsky.social · 02/07/2025

What's short for "Decision maker"? Didi?

000

Adrien Corenflos @adriencorenflos.bsky.social · 27/06/2025

HMC explained in a comic strip. Note the final velocity flip.

Gemini generated image with the following prompt:

Make a cartoon strip illustration of MCMC. The strip shows a kid climbing a ladder to a slide, going down it enthusiastically, but at the bottom the kids mother is waiting and tells him to stop and go back sternly. The last panel shows the kid going back up the slide, but via the slide itself, not the ladder.

No text. Make sure physics is correct.

050

Adrien Corenflos @adriencorenflos.bsky.social · 03/06/2025

The gist of it is that, even though PMMH may eventually work better than PGibbs (because it is based on a better acceptance function in the MCMC sense), it's brittle in the low particle regime. On the other hand, PGibbs works out of the box (after some minor modification of the original algorithm).

010

Adrien Corenflos @adriencorenflos.bsky.social · 03/06/2025

I recently pushed a note on ArXiv, arxiv.org/abs/2505.04611, the title of which I am particularly proud of. I argue there against a misconception that parameter estimation in state-space models is usually better done with particle marginal Metropolis Hastings (PMMH) than particle Gibbs (PGibbs).

190

Adrien Corenflos @adriencorenflos.bsky.social · 25/04/2025

Canary wharf ruined star wars for me.

010

Adrien Corenflos @adriencorenflos.bsky.social · 18/04/2025

Today we remember the passion of Gauss who suffered so we didn't have to. en.wikipedia.org/wiki/Date_of...

010

Adrien Corenflos @adriencorenflos.bsky.social · 17/04/2025

Also the mug 'cause I don't like people using mine (also YOLO)

010

Adrien Corenflos @adriencorenflos.bsky.social · 17/04/2025

If someone's looking for me at the next conference, I got myself the t-shirt version.

2100

Adrien Corenflos @adriencorenflos.bsky.social · 14/04/2025

e.g., today: I am computing some weird SDE bridge, with an unstable behaviour at the final time step; but eh, there's just one value of the bridge end point that can help.

020

Adrien Corenflos @adriencorenflos.bsky.social · 31/03/2025

Use something you did during your PhD. For mine I made a "firework" using MCMC. Not sure about "pretty" but I like it which is what matters.

230

Adrien Corenflos @adriencorenflos.bsky.social · 28/03/2025

My little particles are en route for the weekend!

070

Adrien Corenflos @adriencorenflos.bsky.social · 04/02/2025

Remind me, why am I doing this job again?

150

Adrien Corenflos @adriencorenflos.bsky.social · 03/01/2025

The exponential integrator doesn't do *too* bad, but the exponential drift exp(x) make it super unstable

000

Adrien Corenflos @adriencorenflos.bsky.social · 30/12/2024

This has already made it onto the dedicated Wikipedia page en.m.wikipedia.org/wiki/Email_s... 🤣🤣🤣

110

Adrien Corenflos @adriencorenflos.bsky.social · 30/12/2024

There seems to be some flexibility (ambiguity?) about what metric to choose even in the "simple" case of polytopes though. I'm wondering if there is a way to identify an optimal choice for these somehow.

010

Adrien Corenflos @adriencorenflos.bsky.social · 30/12/2024

The method is similar to Riemannian manifold MALA but uses metrics that are adapted to the domain at hand.

120

Adrien Corenflos @adriencorenflos.bsky.social · 30/12/2024

Interesting drop on arXiv over Christmas. These authors study the problem of sampling over constrained domains, which is usually much harder than unconstrained ones. arxiv.org/abs/2412.18701

160

Adrien Corenflos @adriencorenflos.bsky.social · 17/12/2024

Not exactly what I was expecting as a list for "Rogue-like books".

A list of books tagged "roguelike" on goodread, but they are all about programming languages/software development.

010

Adrien Corenflos @adriencorenflos.bsky.social · 09/12/2024

I'm trying to get a general sense of if the inequality worsens as I increase the number of RV's (identically distributed but not independent for example) or improves. I just tried empirically and it seems like it's worsening exponentially fast, but honestly I don't "feel" why.

120

Adrien Corenflos @adriencorenflos.bsky.social · 28/11/2024

Arguably it has too many. I had seen your response earlier this year discourse.julialang.org/t/state-of-a... when working on a paper where we used Julia over python (better support of variable-size lists essentially) and the sentence in the screenshot is really a deal breaker honestly.

230

Adrien Corenflos @adriencorenflos.bsky.social · 24/11/2024

Well, that was my first time uploading a GIF here, and *of course* it was not going to work, here, sequential case

120

Adrien Corenflos @adriencorenflos.bsky.social · 24/11/2024

This is salient on modern parallel hardware, as these were *designed* for low-precision arithmetic. In arxiv.org/abs/2207.00426 (just accepted at SIAM's JoSC 🥳) we allow for parallel decomposition in the Cholesky-space rather than the natural covariance-space of Gaussian-based Kalman filters.

110

Adrien Corenflos @adriencorenflos.bsky.social · 24/11/2024

Parallel scans accumulate sequences on GPUs (or other parallel hardware) at logarithmic cost in the size of the input. The canonical example is cumulative sums (a, a+b, a+b+c, ...) from an input (a, b, c, ...), but this is hardly the only use, and, e.g., Kalman filtering can be handled in parallel.

2102

Adrien Corenflos @adriencorenflos.bsky.social · 24/11/2024

I'm pretty happy about this specific addition which is the kind of thing I know you like to see in papers too -> isolate the simplest examples on which you know your method would fail and explain why. This was prompted by a reviewer highlighting (correctly) that it wasn't clear why SMC was needed.

010