William Gilpin
banner
wgilpin.bsky.social
William Gilpin
@wgilpin.bsky.social
asst prof at UT Austin physics interested in chaos, fluids, & biophysics.

https://www.wgilpin.com/
hi david, thank you very much :)
June 17, 2025 at 2:06 AM
This work was inspired by amazing recent work on transients by the dynamical systems community: Analogue KSAT solvers, slowdowns in gradient descent during neural network training, and chimera states in coupled oscillators. (12/N)
June 16, 2025 at 5:30 PM
For the Lotka-Volterra case, optimal coordinates are the right singular vectors of the species interaction matrix. You can experimentally estimate these with O(N) operations using Krylov-style methods: perturb the ecosystem, and see how it reacts. (11/N)
June 16, 2025 at 5:30 PM
This variation influences how we reduce the dimensionality of biological time series. With non-reciprocal interactions (like predator prey), PCA won’t always separate timescales. The optimal dimensionality-reducing variables (“ecomodes”) should precondition the linear problem (10/N)
June 16, 2025 at 5:30 PM
As a consequence of ill-conditioning, large ecosystems become excitable: small changes cause huge differences in how they approach equilibrium. Using the FLI, a metric invented by astrophysicists to study planetary orbits, we see caustics indicating variation in solve path (9/N)
June 16, 2025 at 5:30 PM
How would hard optimization problems arise in nature? I used genetic algorithms to evolve ecosystems towards supporting more biodiversity, and they became more ill-conditioned—and thus more prone to supertransients. (8/N)
June 16, 2025 at 5:30 PM
So ill-conditioning isn’t just something numerical analysts care about. It’s a physical property that measures computational complexity, which translates to super long equilibration times in large biological networks with trophic overlap (7/N)
June 16, 2025 at 5:30 PM
More precisely: the expected equilibration time of a random Lotka-Volterra system scales with the condition number of the species interaction matrix. The scaling matches the expected scaling of the solvers that your computer uses to do linear regression (6/N)
June 16, 2025 at 5:30 PM
We can think of ecological dynamics as an analogue constraint satisfaction problem. As the problem becomes more ill-conditioned, the ODEs describing the system take longer to “solve” the problem of who survives and who goes extinct (5/N)
June 16, 2025 at 5:30 PM
But is equilibrium even relevant? In high dimensions, stable fixed points might not be reachable in finite time. Supertransients due to unstable solutions that trap dynamics for increasingly long durations. E.g, pipe turbulence is supertransient (laminar flow is globally stable) (4/N)
June 16, 2025 at 5:30 PM
Dynamical systems are linear near fixed points, so May used random matrix theory to show large random ecosystems are usually unstable. The biodiversity we see in the real world requires finer-tuned structure from selection, niches, et al. that recover stability (3/N)
June 16, 2025 at 5:30 PM
A celebrated result in mathematical biology is Robert May’s “stability vs complexity” tradeoff. In large biological networks, we can’t possibly measure all N^2 interactions among N species, genes, neurons, etc. What is our null hypothesis for their behavior? (2/N)
June 16, 2025 at 5:30 PM
Code here: github.com/abao1999/panda

Paper here: arxiv.org/abs/2505.13755

Study lead by Jeff Lai and Anthony Bao
GitHub - abao1999/panda: Patched Attention for Nonlinear Dynamics
Patched Attention for Nonlinear Dynamics. Contribute to abao1999/panda development by creating an account on GitHub.
github.com
May 22, 2025 at 4:49 AM
The attention architecture allows the model to handle much higher-dimensional inputs at testing than it ever saw during training, so we asked it to forecast two chaotic PDE (a fluid flow and KS equation). Not bad, given that the model has never seen a PDE before (6/7)
May 22, 2025 at 4:49 AM
We fed the model mixes of pure frequencies & measured its response. The activations lit up in complex patterns, indicating nonlinear resonance & mode-mixing, akin to triad interactions visible in turbulent bispectra. Compare these activations to Arnold webs in N-body chaos (5/7)
May 22, 2025 at 4:49 AM
We find a scaling law relating performance and the number of chaotic systems. Even if we control for total amount of training timepoints, more pretraining ODEs improves the model.
May 22, 2025 at 4:49 AM
What is the generalization signal for pretrained model to handle unseen chaotic systems? Post training, attention rollouts show recurrence maps and Toeplitz matrices, suggesting the model learns to implement complex numerical integration strategies to extend the context (5/8)
May 22, 2025 at 4:49 AM
Panda beats pure time-series foundation models at zero-shot forecasting unseen dynamical systems. That means that the model sees a snippet of an unseen chaotic system as context, and autonomously continues the dynamics (no weights are updated) (4/8)
May 22, 2025 at 4:49 AM
We made a novel chaotic systems dataset for pretaining by taking 135 hand-curated chaotic ODE (e.g. Lorenz, Rossler, etc) and mutating/recombining/selecting their ODE to select for chaoticity (3/8)
May 22, 2025 at 4:49 AM
Heroic effort co-lead by UT PhD students Jeff Lai & Anthony Bao, who implemented a new channel-attention architecture combining PatchTST, Takens embeddings, & Extended Dynamic Mode Decomp. They trained the whole thing on AMD GPUs! (2/8)
May 22, 2025 at 4:49 AM