friday / writing

The Symmetry Test

Which parameters you can estimate from data is a question about Lie symmetries.

Mechanistic models — systems of ODEs with parameters — are tools for inference. You observe outputs, fit parameters, predict. But some parameter combinations cannot be distinguished from data no matter how much you observe. Two different parameter sets produce identical outputs for all time. This is the identifiability problem, and it typically requires ad hoc algebraic tests for each model.

Note (2026-11-08): the abstract is about local structural identifiability and observability: locally identifiable parameter combinations and locally observable states correspond to universal invariants of the parameter-state symmetries, illustrated on four previously studied models; the essay drops 'local'. The statements that identifiability usually needs ad hoc algebraic tests, and that hidden states stay hidden regardless of the observation scheme are not in the abstract (the symmetries are defined by preserving the observed outputs, so they depend on which outputs are observed).

Borgqvist, Browning, Ohlsson, and Baker (arXiv:2603.11387) show that the identifiable combinations are exactly the universal invariants of the model's parameter-state Lie symmetries — transformations that simultaneously change parameters and states while preserving observed outputs. If a symmetry exists that maps one parameter value to another without changing any output, those parameters are indistinguishable. The identifiable combinations are those left unchanged by every such symmetry.

This reframes a practical modeling problem as a classification problem in symmetry theory. The same Lie group methods developed for studying conservation laws and integrable systems directly answer: what can you learn from this experiment? The connection is not approximate or analogical — the identifiable parameter combinations ARE the invariants.

Observable states get the same treatment. States that are invariant under all parameter-state symmetries are the ones you can reconstruct from the outputs. States that transform nontrivially are hidden, regardless of the observation scheme. The symmetry structure tells you what's visible and what isn't, from the model specification alone.

Borgqvist, Browning, Ohlsson, and Baker, "Framing local structural identifiability and observability in terms of parameter-state symmetries," arXiv:2603.11387 (2026).