Time-Scale Invariance in Dynamical Models

Why the Clock Matters

A mechanistic dynamical model describes how a system changes over time. Its rate constants therefore have units tied to the time coordinate: a first-order rate might be reported per hour, for example. If observations reveal the shape of a trajectory but do not establish how the recorded time relates to physical time, a faster process on a slower clock can produce the same observations as a slower process on a faster clock. The issue is not necessarily a lack of precise measurements; it can be a structural ambiguity in the model and experiment.

The central question is whether the data establish an absolute time scale, rather than only the trajectory’s shape relative to an unknown clock. The distinction matters when a rate is to be interpreted in physical units or compared across experiments.

Deriving the Rate Transformation

Consider a state vector \(x(t)\) with dynamics

Here \(u(t)\) is an external input, \(k\) denotes rate constants, and \(\eta\) denotes other parameters. Suppose the measured output is \(y(t)=h(x(t);\eta)\), with no explicit dependence on time. Define \(s=ct\) and \(z(s)=x(s/c)\). Applying the chain rule gives

This equation is valid for any \(c>0\). To express it within the same model family, a parameter transformation must make the right-hand side representable by that family. In the common case where the dynamics are a sum of rate constants multiplying rate laws,

set \(k'_j=k_j/c\) for every \(j\), and define the transformed input \(u'(s)=u(s/c)\). Then the transformed-coordinate model is

The initial state is still \(z(0)=x_0\), and the output at corresponding times is unchanged: \(h(z(s);\eta)=h(x(s/c);\eta)\). If the right-hand side is locally Lipschitz in the state on the domain of interest, the initial-value problem has a unique solution there. Thus the transformed model’s solution is precisely the original trajectory expressed in the new time coordinate. Without uniqueness, the transformation still gives the same initial-value problem, but equality of unique trajectories cannot be concluded.

This rate transformation applies when the model family is closed under multiplying its entire dynamic right-hand side by \(1/c\). Parameters that do not scale the speed of the dynamics—for example, an initial concentration or a static observation-scale parameter—are not automatically divided by \(c\). A model with fixed time-dependent inputs must also transform those inputs to preserve the trajectory. That condition is important: a known input schedule can supply the time reference that observations of the states alone lack.

Worked Examples

Assumptions, Diagnostics, and Alternatives

The invariance argument is conditional, not a claim that every dynamical model has an unknown time scale. Check the model and experimental design. The state equations must permit the required rate transformation; the observation function must not introduce an independent calibrated clock that changes under the transformation; and any input must be transformed along with the time coordinate. For a unique trajectory, the initial-value problem also needs an appropriate uniqueness condition, such as local Lipschitz continuity of the right-hand side in the state over the relevant domain.

Useful diagnostic questions include:

• Are observation times known in physical units, or are they only labels from an uncalibrated clock?
• Does multiplying every dynamic rate by the same positive factor and inversely rescaling time leave the model’s output unchanged?
• Are input transition times, sampling intervals, or other external time references independently known?
• Do rate estimates change inversely when the time unit changes, as dimensional consistency requires?

A numerical fitting algorithm may return a particular rate even when the absolute rate is structurally ambiguous. A selected starting value, constraint, or optimizer convention can choose one point among equivalent parameter-and-clock descriptions; it does not make that rate identifiable from the observations. In contrast, a poorly estimated rate with calibrated times may be a practical estimation problem rather than this exact invariance. The distinction is whether the model admits an exact transformation preserving ideal observations, not simply whether a fitted confidence interval is wide.

Time-scale invariance differs from dose or exposure scaling: the latter concerns rescaling an input’s magnitude, while time-scale invariance changes the speed of the dynamics and the units attached to rates. Multiple outputs can help distinguish state mechanisms, but they do not by themselves calibrate a shared unknown clock if every output is observed on that same clock. A known event time, an independently measured rate, or a calibrated observation schedule can supply the missing reference. If no such reference exists, it may be more informative to report rates in instrument-time units or estimate a rate relative to a specified reference rate than to present an unsupported absolute rate.

Common Pitfall

A common error is to say, “The data are dense and the trajectory is smooth, so the absolute rate must be identifiable.” More observations can define the curve more precisely without telling us how its time axis maps to hours or seconds. The correction is to test whether a change of clock paired with the corresponding change of rates preserves every observed output. If it does, denser observations on that same uncalibrated clock do not alone resolve the absolute scale. Conversely, do not claim ambiguity when the experiment includes a known time reference: a fixed, independently calibrated input schedule may make the transformed model inconsistent with the design.