Posterior Geometry in Nonlinear ODE Models

Why Posterior Geometry Matters

In a nonlinear ODE model, several parameters can compensate for one another: changing one rate may be offset by changing an initial condition, capacity, or another rate. The resulting posterior can concentrate near a narrow, curved region rather than around a roughly circular cloud. This geometry is not merely a plotting feature. It affects which parameter combinations are well determined, how efficiently posterior sampling can explore the distribution, and whether separate summaries of each parameter represent plausible joint values.

A posterior is the distribution of unknown parameters conditional on the observed data and the specified model. Let \(\boldsymbol{\theta}\) denote the parameter vector, \(y\) the observed data, \(L(y\mid\boldsymbol{\theta})\) the likelihood, and \(\pi(\boldsymbol{\theta})\) the prior density. Bayes' rule gives

when the denominator is finite and nonzero. The posterior geometry depends on the ODE predictions, the observation model, and the prior, not on the ODE alone. For interpretation, assume the ODE solution is uniquely defined for parameter values under consideration, the numerical solver is accurate enough for the likelihood calculation, and the observation-error model is appropriate. If any of these conditions fails, apparent geometry may reflect a modeling or computational artifact.

A Simple ODE Model and a Curved Ridge

Consider a hypothetical response following a one-compartment-type rise toward a plateau:

where \(x(t)\) is the response in concentration units, \(A>0\) is the plateau concentration, \(k>0\) is a rate constant in inverse time, and \(t\) is time. This is a simplified illustration, not a claim that a particular biological system follows this mechanism. With only one exact observation \(x(4)=8\), every pair satisfying

has the same predicted observation. Solving for \(k\) gives a curve in the \((A,k)\) plane:

As the plateau \(A\) increases, the rate \(k\) needed to reach the observed value by time 4 decreases. The resulting dependence is negative and nonlinear. With noisy measurements and a prior, the posterior generally has a finite-width band around such a curve rather than an infinitely thin line.

Local Curvature and ODE Sensitivities

The local orientation of a posterior ridge can be understood through sensitivities: derivatives of model predictions with respect to parameters. For the illustrative response,

Near a parameter value, a small change in prediction is approximately the sum of each sensitivity multiplied by its parameter change. If one observation is to remain nearly unchanged, parameter changes must approximately satisfy

This is a local, linear approximation to the curved equal-prediction contour. With multiple observations, the sensitivities at each observed time contribute information about different directions in parameter space. If their effects are similar across all observation times, some combinations remain difficult to separate.

For independent Gaussian observation errors with known standard deviations \(\sigma_i\), the log likelihood, apart from constants, is

Near a well-fitting parameter value, the curvature of this log likelihood is often approximated by the sensitivity matrix weighted by the observation precisions. In a locally Gaussian approximation, this curvature describes the orientation and relative narrowness of the posterior cloud; prior curvature also contributes. This approximation is most informative when the posterior is concentrated in one smooth region. It can be misleading for strongly curved, bounded, or multimodal posteriors, and it does not replace inspection of the full joint distribution.

Reading Joint and Marginal Uncertainty

A joint posterior describes combinations of parameter values. A marginal posterior for one parameter is obtained by integrating the joint distribution over the other parameters. Marginal summaries are useful, but they discard information about which values occur together. For example, separate credible intervals for \(A\) and \(k\) do not imply that every combination inside the two intervals is plausible. On a narrow ridge, many such cross-combinations may have very low posterior density.

Diagnostics, Alternatives, and Limitations

Useful diagnostics begin with the model predictions and the joint posterior, not with a correlation coefficient alone. Examine pairwise plots on meaningful parameter scales, overlay model predictions from representative joint draws, and check whether apparent ridges persist under tighter ODE solver tolerances. Compare posterior regions with parameter bounds and prior support. A ridge terminating at a boundary may reflect a constraint or prior as much as the data. If predictions change irregularly when solver tolerances are tightened, numerical error may be distorting the apparent likelihood surface.

A linear correlation summarizes only linear association on the chosen scale. A curved ridge can have weak overall correlation even though the parameters are strongly dependent. Conversely, a high correlation does not by itself establish a biological compensation mechanism. It describes the posterior under the stated model, data, and prior. Re-expressing parameters, for example using a log scale for a positive rate, can change the visual shape and numerical correlation; it does not by itself create new information in the observations.

A local covariance or Hessian approximation is compact and useful when the posterior is approximately elliptical near its center. A joint posterior plot can reveal curvature that a single ellipse misses. Profile likelihoods can describe how fit changes as one parameter is fixed and the others are optimized, but they are not posterior distributions and do not include prior weighting in the same way. Marginal intervals provide compact parameter-specific summaries, while joint plots preserve dependence. These approaches answer related but different questions; none removes uncertainty caused by weakly informative data.

Ridges can make sampling difficult because a sampler must move along a narrow, possibly bending region while avoiding directions that lead to poor fit. This can result in slow exploration or a chain that appears to remain in a small part of the posterior. Geometry alone does not diagnose convergence: sampling diagnostics and adequate exploration are still necessary. Nor does a ridge prove that the model is structurally unidentifiable. It may reflect limited observations, noise, parameter bounds, or prior information; structural identifiability concerns what can be distinguished in principle under a specified model and idealized data.