Mass-Action Parameter Products and Identifiability

Why a Reaction Rate Can Hide Its Components

Mass-action kinetics often make reaction rates proportional to products of reactant concentrations. If one reactant is maintained at a constant concentration, the observed dynamics may depend on the reaction constant and that concentration only through their product. A time course can then determine the product precisely, yet provide no way to distinguish the factors that produced it.

This is a structural identifiability issue, not merely a shortage of noisy observations. The key question is whether different admissible parameter values can produce exactly the same ideal output over time. The analysis below uses a deliberately explicit model: a reactant \(A\) is consumed by reaction with \(B\), while \(B\) is held at a constant but unknown concentration.

Because \(B(t)=B_0\) throughout the observation interval, the mass-action equation for \(A\) is

The composite coefficient \(\kappa\) has units of inverse time. The reduced equation is often called pseudo-first-order kinetics: the loss rate is first order in the observed concentration \(A\), even though the underlying reaction is second order in its reactants.

What the Ideal Time Course Determines

Solving the reduced differential equation with \(A(0)=A_0\) gives

When \(A_0\) is known and positive, an ideal observation of \(A(t)\) at any \(t>0\) determines the composite coefficient:

Equivalently, the log concentration has slope \(-\kappa\):

These expressions do not, however, provide separate estimates of \(k\) and \(B_0\). Both enter the output only through their product. The distinction between identifying a combination and identifying its components is essential: a perfectly determined \(\kappa\) does not imply that either factor is determined.

Why the Separate Constants Cannot Be Recovered

A proof of nonidentifiability must show that distinct allowed parameter values produce exactly the same output, not merely that a numerical fit is uncertain. Here, for any \(c>0\), define a new parameter pair by

The new pair remains positive and satisfies

It therefore gives the same reduced differential equation and the same initial condition. By the solution formula, its output is identical at every time:

For any original positive pair, choosing \(c\ne 1\) gives a distinct pair with exactly the same ideal trajectory. Thus \(k\) and \(B_0\) are not separately structurally identifiable in this observation model. The identifiable quantity is their product \(\kappa\). This conclusion assumes there are no other observations or constraints that provide independent information about either factor.

What Additional Information Can Resolve the Ambiguity?

Separate recovery requires information that does not depend solely on \(kB_0\). For instance, if \(B_0\) is measured independently and is nonzero, then the identified product gives \(k=\kappa/B_0\). Conversely, independent knowledge of \(k\) allows \(B_0=\kappa/k\). Such information changes the inference problem; collecting more observations of \(A(t)\) under the unchanged model only refines the same composite coefficient.

Model Checks, Alternatives, and Limitations

The conclusion depends on the assumptions that make \(B\) constant and unobserved. If \(B\) is depleted appreciably, then \(B(t)\) is not a fixed factor and the reduced equation \(dA/dt=-\kappa A\) is not the stated mass-action model. If \(B(t)\) is known as a time-varying input, the equation instead contains the known concentration profile, and the observed dynamics can provide information about \(k\) directly. These are different models and should not be treated as interchangeable.

Useful checks include whether the assumptions about buffering and other loss processes are scientifically justified, whether the observed trajectory is compatible with exponential decay, and whether the fitted rate has units of inverse time. A systematic departure from a straight line in a plot of \(\ln A(t)\) against time may indicate model misspecification, time-varying \(B\), or another process affecting \(A\). Such a departure does not by itself establish which alternative explanation is correct.

A common fitting error is to estimate \(k\) and \(B_0\) as separate free parameters, obtain one numerical pair, and report both as if the data uniquely determined them. In this model, many pairs give identical predictions, so an optimizer may return a value selected by its starting point or numerical constraints rather than by the output information. A small residual error cannot validate separate identifiability. A more transparent parameterization estimates \(\kappa\) directly and reports that \(k\) and \(B_0\) require independent information.

The main alternatives are to fit the reduced model in terms of \(\kappa\), to impose a scientifically justified known value for one factor, or to collect an independent measurement that constrains a factor. A more detailed model in which both reactants change can also be appropriate if its assumptions match the experiment. Simply adding more measurements of the same output under the same constant-\(B_0\) conditions does not break the exact product ambiguity.

Finally, this is a structural conclusion for the stated idealized model. In finite noisy data, even the composite coefficient may be estimated imprecisely; conversely, precise estimation of \(\kappa\) does not remove the structural ambiguity in its factors. The result also relies on \(A_0>0\) and positive \(k,B_0\): if \(A_0=0\), the predicted \(A\) output is identically zero and contains no information about \(\kappa\).