How Stable Is Parameter Identifiability in Data-Driven Biological Modeling?

Why the Fisher approximation misreads weak directions

Practical identifiability asks whether the data pin down a model's parameters to a finite range. The usual gold standard, profile likelihood, answers this by repeatedly re-optimizing the model while holding one parameter fixed, which becomes prohibitively expensive for high-dimensional biological systems. A cheaper alternative replaces the loss Hessian with the Fisher information matrix built from the model's Jacobian, but that step discards the second-order curvature of the model output.

The paper's argument is that this discarded term matters most exactly where identifiability is in doubt. Along weakly constrained parameter directions the Fisher information is small, so the neglected Hessian contribution can be comparable to or larger than it. A parameter may then be labeled non-identifiable even though second-order curvature constrains it, or its uncertainty may be misjudged by more than an order of magnitude. The authors keep the full loss Hessian K = F + r*⊗H, where F is the Fisher matrix, H is the Hessian tensor of the stacked observations, and r* is the residual at the fitted parameters.

Stability rates depend on residual and perturbation size

The paper defines identifiability as stability of the estimated parameters under perturbations of the data: every perturbed minimizer must stay within a controlled distance of the original estimate. With a zero residual, the loss Hessian reduces to the Fisher matrix. If that matrix is singular, displacements split into a range component controlled at a Lipschitz rate by the Jacobian and a null-space component that the Fisher matrix leaves free but that the Hessian tensor controls only at a Hölder-1/2 rate, provided its coercivity on the Fisher null space is positive. When the Fisher matrix is full rank, the Lipschitz rate returns.

A nonzero residual changes the picture. The term r*⊗H lifts the weak eigenvalues of the Fisher matrix, and the loss Hessian becomes positive definite, giving Lipschitz stability. Between these limits, the residual norm and the perturbation size together decide the order: Lipschitz below a crossover scale proportional to the residual norm, Hölder-1/2 above it. This is a statement about local behavior under the paper's smoothness assumptions, not a claim that every biological model attains either rate.

A coordinate test that keeps other parameters free

For a single parameter, the paper works within a local quadratic approximation of the least-squares loss and eliminates the remaining parameters through a generalized Schur complement. This yields a profiled curvature κ and a Lipschitz coefficient L for each coordinate, and the coordinate is identifiable if and only if κ is positive. The construction tolerates singular Jacobians and nuisance Hessian blocks, so it does not require the full parameter vector to be well behaved.

The distinction from subset-based screening methods is central. The Quaiser and Yao procedures assess whether a subset of parameters is estimable conditional on fixing or excluding the rest, so retaining a parameter does not establish its identifiability when all parameters are free. The new criterion evaluates each coordinate with all others free, which is why the authors argue it can recover the profile-likelihood classification without repeated optimization.

JAK–STAT shows where Fisher overstates sensitivity

The JAK–STAT signaling benchmark fits time courses of two observables with seven unknown quantities: four kinetic parameters, two scaling factors, and one initial condition. Profile likelihood identified p1, p3, and p4 as practically identifiable, with profiles crossing the confidence thresholds on both sides, while the profiles of p2, s1, s2, and x1(0) were flat. The proposed criterion gave the same split from the local Jacobian and Hessian alone: the three identifiable coordinates had finite Lipschitz coefficients, and the other four had zero curvature and unbounded coefficients.

For the identifiable coordinates, the Hessian-aware coefficient was smaller than its Fisher counterpart, most markedly for p3, meaning the Fisher approximation overestimated how sensitive those estimates are to data perturbations. The subset methods diverged: Quaiser removed p2 and x1(0) and kept s1 and s2, while Yao with a threshold of 0.04 excluded only s2. Both retained parameters whose full-profile behavior was flat.

The second-order analysis separated two mechanisms behind the four non-identifiable coordinates. Along one eigendirection, first-order curvature, coercivity, and residual-induced curvature all vanished to numerical precision; this is a scaling symmetry that leaves both observables unchanged, so no stability rate applies. Along the other weak direction, first-order curvature was about 10^-6 while the residual and coercivity terms were of order 10^-3, so the direction is weakly informed but Hessian-constrained. The noise level lay above the crossover, placing that direction in the Hölder regime at the resolution of the data.

HIV patients differ in how much the correction matters

The HIV host–virus model has six parameters describing target-cell turnover, infection, viral production, and clearance, fitted to log-transformed plasma viral load for two patients with contrasting post-peak trajectories. For patient 3, the four largest eigenvalues of the corrected and Fisher Hessians were closely aligned, but the smallest eigenvalue rose by nearly three orders of magnitude after correction. The change was concentrated in the least constrained direction, while eigenvector orientations shifted little, with the largest pairwise perturbation about 1.1×10^-2.

For patient 8, the two eigenvalue spectra nearly overlapped. The authors summarize the contrast with a diagnostic that compared the corrected and Fisher uncertainty estimates against MCMC results: the Fisher approximation was adequate when that diagnostic was below 1 and overestimated interval widths when it was well above 1. The practical lesson is that residual curvature must be judged against the local information structure rather than eigenvector changes alone, and that the size of the correction is patient-specific rather than a fixed property of the model.

Amyloid-β prediction intervals shift in time

In the amyloid-β accumulation application, the paper propagates the quadratic parameter posterior through the model to obtain prediction intervals. The corrected Hessian spectrum showed a clear spectral gap, which the authors treat as support for using the subthreshold eigenspace as an effective null subspace. Differences in the regularized contributions from that subspace shifted the age of the narrowest prediction interval from 84.36 to 80.73 years, moving it into the period of rapid amyloid-β accumulation.

This result matters because it links parameter-level corrections to a clinically legible quantity, the timing of minimum predictive uncertainty. It also marks a boundary: the paper distinguishes an exact null space from a subspace defined by spectral truncation, and notes that sloppiness, a hierarchy of sensitivities, is not by itself evidence of practical non-identifiability. The threshold-based partition is defensible here because a clear gap was present, not because truncation is generally safe.

When the Hessian correction changes the verdict

The paper's priority is a residual-aware reading of identifiability. The Fisher approximation is adequate when the corrected and uncorrected information structures nearly coincide, as in patient 8 and the well-constrained JAK–STAT coordinates. It changes the verdict when the smallest eigenvalue is lifted substantially, as in patient 3, or when a weak direction is constrained by second-order curvature rather than first-order sensitivity. In those cases, the relevant question is not only whether a parameter is identifiable but at what stability rate and at what perturbation scale.

The framework requires only local derivative evaluations, which is what makes it tractable for large models where profile likelihood is infeasible. Its conclusions remain local and rest on a quadratic approximation of the loss, so the authors present higher-order analysis as necessary where that approximation breaks down. The most useful takeaway for a modeler is to compare the corrected and Fisher spectra and the residual norm before trusting a Fisher-based identifiability label, because the direction that decides the label is often the one the Fisher matrix constrains least.

Sources used in this answer