Invariant Image Reparameterisation: Bridging Symbolic and Numerical Methods for Identifiability Analysis, Model Reduction, and Prediction

summary

The gist

Invariant Image Reparameterisation (IIR) is an approach to parameter identifiability analysis that links symbolic reparameterisation ideas with local numerical calculations to determine a…

In short

Invariant Image Reparameterisation (IIR) links symbolic reparameterization ideas with numerical calculations to find a basis-independent representation of parameter combinations affecting model behavior. It replaces complex symbolic conditions with a single numerical Jacobian calculation, bridging symbolic and numerical methods for identifiability analysis and prediction.

Key concepts

Invariant Image
This is a reduced, basis-independent mathematical representation of the parameter combinations that control how an observable model behaves. It captures the essential information about parameter dependencies regardless of the specific coordinate system used to describe them.
Symbolic Reparameterisation Condition
This is a theoretical condition derived from differential equations on the row and null spaces of an auxiliary mapping Jacobian. It describes when a transformation can be written symbolically, which is then approximated numerically at a single reference point.
Sparse Monomial Bases
These are specific coordinate representations chosen for the invariant image. They are often preferred over standard methods because they yield parameter combinations that are more interpretable and mechanistically meaningful for understanding the model's underlying physics or biology.

Terminology used across episodes

This episode discusses

The paper

Invariant Image Reparameterisation: Bridging Symbolic and Numerical Methods for Identifiability Analysis, Model Reduction, and Prediction · Read on arXiv

Oliver J. Maclaren, Ruanui Nicholson, Joel A. Trent, Joshua Rottenberry, Matthew J. Simpson

University of Auckland · Queensland University of Technology

Structural and practical parameter non-identifiability issues are common when mathematical models are used to interpret data. Such issues motivate model reparameterisation and reduction methods. Here, we consider Invariant Image Reparameterisation (IIR), which asks when symbolic reparameterisation conditions can be replaced by numerical derivative calculations at a single reference point. The central object is the invariant image: a reduced, basis-independent representation of the parameter combinations controlling observable model behaviour. We show that when a one-to-one componentwise transformation makes observable behaviour depend only on fixed linear combinations of the transformed parameters, a single numerical Jacobian determines the associated lower-dimensional reparameterisation space. This includes models depending on monomial combinations of the original parameters. We also give a first-order invariance condition that can rule out a minimal reduction of this form and extracts candidate invariant null directions for an exact non-minimal reduction. In structurally identifiable but practically weakly informed settings, the same calculations separate strongly and weakly informed parameter combinations. The invariant image admits multiple coordinate representations: the SVD gives a default orthonormal basis ordered by local identifiability, while sparse monomial bases are often more interpretable. Treating these coordinates as interest parameters in Profile-Wise Analysis gives likelihood-based uncertainty quantification and prediction. We demonstrate the method on parameterised normal models with Poisson-limit, extended Poisson-limit, and non-limit cases, and on the repressilator, a nonlinear differential equation model of gene regulation. A Julia implementation of IIR, with these and further examples, is available at https://github.com/omaclaren/reparam.

DOI: 10.1016/j.mbs.2026.109844

Transcript

Introduction to the show: ident: Statistics. The day in statistics: methods, theory and the inference under them.

Tom: I'm Tom.

Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.

Tom: We're looking at the paper: "Invariant Image Reparameterisation".

Jane: Here's the gist of the paper. Invariant Image Reparameterisation (IIR) is an approach to parameter identifiability analysis that links symbolic reparameterisation ideas with local numerical calculations to determine a basis-independent representation of parameter combinations associated with observable model behaviour.

Tom: This method is significant because it allows the symbolic reparameterisation condition to be replaced by a single numerical Jacobian calculation at a single reference point, bridging symbolic and numerical methods for identifiability analysis, model reduction, and prediction.

Jane: Core Concept of Invariant Image Reparameterisation The central object in this approach is the invariant image, which is defined as a reduced, basis-independent representation of the parameter combinations controlling observable model behaviour.

Tom: The authors show that when a one-to-one componentwise transformation makes observable behaviour depend only on fixed linear combinations of the transformed parameters, a single numerical Jacobian determines the associated lower-dimensional reparameterisation space.

Paper summary: Tom: To unpack that a bit more, the paper claims that when you have this one-to-one componentwise transformation, and the observable behavior only depends on fixed linear combinations of those transformed parameters, then a single numerical Jacobian calculation at one reference point is enough to determine the associated lower-dimensional reparameterisation space.

Jane: That sounds like it simplifies the symbolic condition into something we can actually compute with numbers, which is huge because symbolic methods often get stuck in idealized settings.

Lu: They show that this process finds a reduced parameter space called the invariant image, and they even mention that this image has multiple ways to be described, like using Singular Value Decomposition or sparse monomial bases.

Meng: That's interesting because if you can choose between the SVD basis and these interpretable monomial bases, it suggests we might get representations that are both mathematically sound and actually make sense for us when we look at the data.

Lalam: And they introduce a first-order invariance condition, which they use to tell the difference between minimal reductions, like finding the invariant part of the local null space in non-minimal cases.

Tom: So it's not just finding *any* reduction, but using that first-order condition to distinguish between different types of reductions—minimal versus non-minimal—which is a key distinction in identifiability analysis.

Jane: It’s like they are giving us a tool to decide if we're looking at the most fundamental set of parameters or just some partial simplification.

Lu: The symbolic condition itself is represented by differential conditions on the row and null spaces of an auxiliary mapping Jacobian, and they show how that can be exactly represented by local derivative information at a single reference point under certain structural assumptions.

Meng: That assumption about the transformed auxiliary mapping being written in terms of reduced transformed coordinates, like phi star theta star equals phi bar A theta star equals phi bar eta, is what lets them get this clean numerical form.

Lalam: The specific conditions they derive are that the row of A must contain the row of D theta star phi star theta stars, and the null space of A must be contained in the null space of D theta star phi star theta stars.

Conclusion: Tom: So what does this paper actually mean when we look at it from a broader perspective? The title Invariant Image Reparameterisation really captures that they are bridging the gap between symbolic reparameterisation ideas and actual numerical calculations for identifiability analysis, model reduction, and prediction.

Jane: It suggests that we don't always need those heavy, global symbolic calculations to know what parameters matter; sometimes a local numerical calculation at one spot gives you the same information up to some numerical tolerances.

Lu: The implication is that we can use these single-point Jacobian calculations to find reduced parameter spaces, even when we are dealing with models that might be structurally identifiable but practically weakly informed.

Meng: For deployment purposes, this could mean faster model reduction because instead of a global symbolic search, you just run one numerical check at the reference point to see what the relevant combinations are.

Lalam: It also opens up ways to choose which coordinates you use for that reduced space—like choosing between the default SVD construction and these more interpretable sparse monomial bases—which is a practical win for how we interpret our results.

Tom: So, essentially, they provide a numerical way to get information about identifiable parameter combinations without needing the full machinery of symbolic methods across the entire model structure.

Jane: It moves us closer to having practical tools for model reduction and prediction where we are dealing with real data and not just perfectly idealized mathematical setups.

More episodes

← Home