Robin Ryder

Evolutionary Tempo, Supertaxa, and Living Fossils Abstract. A relationship between the rate of molecular change and diversification has long been discussed, on both theoretical and empirical grounds. Howev

Do diversification rates and molecular evolution covary? Maybe, maybe not, but it’s surprising how much falls into place if we assume they do doi.org/10.1093/sysb...

The wonderful Niamh Cahill is recruiting a PhD student (co-supervised by me!) to work on estimating and projecting population exposure to extreme sea levels. Please contact me if interested!

I am pleased to announce that together with some friends, we are organising a workshop on Non-Reversible MCMC Sampling, taking place at Newcastle University from 8–10 September 2025.

Details on the programme and registration can be found at the workshop website (sites.google.com/view/probai-...).

Exciting new opening @nccrlanguage.bsky.social with the option of permanent extension. If you work in cutting-edge stats and are interested in joining a highly inter-disciplinary, vibrant community, don't hesitate to apply at tinyurl.com/3fxwjk8e

New on the blog: Using Bayesian tools to be a better frequentist

Turns out that for neg. bin. regression with small samples, standard frequentist tools fail to achieve their stated goals. Bayesian computation ends up providing better frequentist guarantees. www.martinmodrak.cz/2025/07/09/u...

Posterior distribution for two parameters. Left: parameter R0. The posterior using a model at the departmental level is much more peaked that the posteriror using a model at the national level. Right: paramer ν. There is only one curve at the national curve, but many individual posteriors at the departmental level, showing a lot of heterogeneity.

Map of France showing the paramer ν_k at each department. values are lower in the North-East (ν_k of the order of 0.5 and larger in the West (values varying between 1 and 3.5).

The methods scale well (by ABC standards). Here is the output on real data with about 300 parameters to infer. Notice that the local parameters do vary quite a lot, and allowing for that variation allows for better inference of the global parameters.

Performance of 7 algorithms. For most of the figure, the ordering is: worst ABC-SMC, the ABC-Vanilla or ABC-SMC, then permABC-Vanilla or permABC-SMC, then Under-Matching, then Over-Sampling

We have found that both Over Sampling and Under Matching give significant improvements with mis-specified models or outliers, as these schemes make it easier to catch the tails of the posterior predictive.

We then progressively increase M₀<M₁<… <Mₜ=K, giving another sequential scheme. We could combine over sampling and under matching but haven't considered that in this paper.

In Under Matching, we take the complementary approach: we simulate synthetic data of the appropriate size, but the distance between synthetic and observed data is only computed on the closest M₀ compartments. Again, matching is easier, since we start with the easiest parts of the data.

We then progressively decrease L₀>L₁>… >Lₜ=K. This defines a different sequential scheme, which is practical because we are allowing for permutations of the data.

In Over Sampling, we will instead simulate "too much" data. Say the observed data have K=94 compartments: we will initially simulate synthetic data with L₀=1000 compartments, and keep the best match for each compartment. Note that matching is easier, even for a low value of ε!

Say you are targeting a pseudo-posterior with some threshold ε. A standard ABC-SMC scheme consists in considering a (possibly adaptive) schedule ε₀>ε₁>…>εₜ=ε.
Each iteration of the SMC tightens the approximation, and eventually you reach the target pseudo-posterior with low value of ε.

We then implement two new sequential strategies:
➡️ over sampling and
➡️ under matching.

These differ from traditional ABC-SMC sequences of distributions.

Performance curve of various algorithms. For most of the figure, the ordering is: worst ABC-PMC, then ABC-Vanilla and ABC-SMC, best either permABC-Vanilla or permABC-SMC

Taken alone, this natural idea already gives us a gain of an order of magnitude.
(In this figure, lower is better.)

Vanilla ABC: d(x,z)<ε permABC: min_{σ∈S_K) d(y, z_σ)<ε

We exploit the model structure to accelerate the inference. We use a linear assignment algorithm to find the permutation of the synthetic data that best matches the observed data.

Standard ABC methods, including ABC-SMC, are slow, have huge computational cost, or simply fail to reach an acceptable approximation level.

You can easily imagine other comparable applications, with data split into exchangeable compartments. With an intractable likelihood, inference is difficult, because of parameter size and model structure.

A hierarchical model with observations y_1, y_2,... y_K. Each y_i depends on a local parameter μ_i and on a global, shared parameter β.

We consider ABC for hierarchical models, with both local and global parameters. Our motivating example is an SIR model of early Covid data across 94 French departements: some parameters are shared across all departements (global), others are specific to each departement (local).

New preprint on arXiv, led by PhD student Antoine Luciano: 
Permutations accelerate Approximate Bayesian Computation.

arxiv.org/abs/2507.06037

There is strong phylogenetic signal in this trait, so we can try to reconstruct ancestral values at internal nodes.

Look where the arrow is pointing: we find a 99% probability that FD calls existed at that node, meaning that non-trivial compositionality already existed 11 million years ago(!).

We consider whether the communication systems of various species of tits, chickadees, and related birds, exhibit FD calls, a non-trivial composition of two base calls.

These are plotted on the known phylogeny: red nodes exhibit FD calls, grey nodes (or don't pass our rather stringent condition).

Birds combined calls more than 11 million years ago - Scientific Reports Scientific Reports - Birds combined calls more than 11&nbsp;million years ago

New paper, led by @ambresalis.bsky.social and just out in Scientific reports:

Birds combined calls more than 11 million years ago

www.nature.com/articles/s41...

This. ⤵️

Some of us are manifesting for the algorithm to be called Metropolis-Rosenbluth-Teller-Hastings, or MRTH. It better recognizes the scientists who had an instrumental role in the original paper. It's what I use in my papers and lectures. I hope it will eventually become the standard name!

Macroevolutionary analysis of polysynthesis shows that language complexity is more likely to evolve in small, isolated populations | PNAS Evolution of complexity in human languages has been vigorously debated, including the proposal that complexity can build in small, isolated populat...

Is language complexity more likely to evolve in small, isolated populations? We test this idea using global occurrence of one particular form of morphological complexity - polysynthesis (complex word forms that embody whole phrases) www.pnas.org/doi/10.1073/...

Enjoying discovering Singapore, and ready for BayesComp 2025!

I'll be giving a talk tomorrow (Wed) on Saddlepoint Monte Carlo and its application to exact Ecological Inference. Come say hi at 16:10 in room LT51 - it's a cool paper, I swear!

arxiv.org/abs/2410.18243

Screenshot of an e-mail which reads "Dear Dr.Ryder, Robin J. Hello! We understand that you have made outstanding achievements in the academic field and published many high-level papers. For example, your recent publication "Ryder, Robin J." has left a deep impression on us. We sincerely invite you to submit to our open access journals. Our journals have been included in the Scopus database and cover multiple disciplines such as environmental and earth sciences, ecological research, linguistics, building materials, sustainable oceans, medicine, etc., and are committed to publishing high-"

Sloppier-than-usual academic spam. 😬

I'd never thought of publishing a paper whose title would be my name. Sounds like a highly efficient plan, it would eliminate the whole headache of trying to pick a good title! 🙄

ISBA 2026: Call for Invited Sessions & Panel Discussions The Scientific Committee invites the submission of proposals for Invited Sessions & Panel Discussions for the 2026 ISBA World Meeting, to be held in Nagoya, Japan from 28 June to 3 July, 2026. Invited...

The Scientific Committee welcomes proposals for Invited Sessions for the 2026 #ISBA World Meeting. Please submit proposals using this form: forms.gle/5G3xJWJaaE6Q...

Proposals will be accepted until 11:59pm (anywhere on Earth) on 6 July 2025.

Also see the conf. site: isba2026.github.io

An Introduction to Bayesian Methods for the Social Sciences Lecturers: Antonietta Mira & Francesco Denti Modality: In-presence Week 1: 18-22 August 2025   Workshop contents and objectives Bayesian statistics has experienced a surge in popularity over the last ...

This looks like a great workshop:
➡️ An Introduction to Bayesian Methods for the Social Sciences, Lugano (Switzerland), 18-22 August 2025

Bayesian regression, network models, Bayesian clustering…

www.usi.ch/en/education...

Bienheureux ceux qui ne comprennent pas ce message.