Why Unknown Initial Amounts Matter
A declining concentration can reflect a large starting amount followed by rapid decay, or a smaller starting amount followed by slower decay. Whether these explanations can be distinguished depends on what the experiment measures. In a simple first-order decay model, the full absolute output over time contains information about both the starting amount and the rate. A single summary, such as the initial decline rate, may contain only their product.
The distinction is one of structural identifiability: under ideal, noise-free observation and the stated model, can different parameter values produce exactly the same measured output? It is not a claim that finite, noisy data will estimate the parameters precisely.
Here, \(A\) has units of amount, \(k\) has units of inverse time, and \(x(t)\) has units of amount. The unknown initial condition is part of the unknown quantity set: it must be identified along with \(k\), rather than treated as known by default.
What the Full Trajectory Reveals
For a positive output, taking logarithms gives a straight-line relationship between log amount and time. The intercept is \(\log A\), and the slope is \(-k\). Thus, the trajectory contains two distinct features: its level determines \(A\), and its rate of decline determines \(k\).
The same separation can be seen without fitting a line. At time zero, \(x(0)=A\). The initial derivative is \(x'(0)=-kA\), so if both the initial amount and derivative are available, their ratio gives \(k\). More generally, for any two distinct times \(t_1<t_2\) in the observation interval, the ratio of outputs eliminates \(A\):
These formulas require positive outputs and distinct times. With \(k>0\), the later output is smaller, so \(x(t_1)/x(t_2)>1\) and the calculated rate is positive. Once \(k\) is recovered, either output gives the same \(A\) if the model is exact.
Worked Example: Recovering Both Parameters From a Decay Trajectory
Suppose a hypothetical assay measures an amount of \(120.0\) mg at \(t=0\) h and \(44.15\) mg at \(t=5\) h. Assume the first-order model and the absolute-output measurement assumptions above.
The two times are distinct, both outputs are positive, and the later output is smaller, consistent with decay. The rate estimate implied by these exact-form equations is
The displayed ratio is rounded; using the unrounded logarithm gives approximately \(0.2000\ \mathrm{h}^{-1}\). Because the first measurement is at time zero, the initial amount is directly \(A=x(0)=120.0\) mg. The resulting model is \(x(t)=120.0e^{-0.2000t}\) mg, with \(t\) measured in hours. The two observations distinguish the starting amount from the decline rate under the assumed model; they do not establish that the model is biologically correct.
The uniqueness claim has a direct proof. Suppose two parameter pairs, \((A,k)\) and \((A',k')\), generate the same positive outputs at \(t_1\) and \(t_2\), where \(t_2>t_1\). Their output ratios must also match. Taking logarithms of those ratios gives \(-k(t_2-t_1)=-k'(t_2-t_1)\). Since \(t_2-t_1\ne 0\), it follows that \(k=k'\). Substituting this equality into the output at \(t_1\) gives \(A=A'\). Hence no distinct parameter pair can produce those same two outputs.
When a Measurement Reveals Only a Product
The conclusion changes if the observation is not the trajectory but only its initial slope. Since \(x'(0)=-kA\), measuring that derivative reveals the product \(kA\), not the two factors separately. Every positive pair satisfying the same product gives the same slope. One output value at a single nonzero time can also leave a continuum of compatible pairs: it supplies one equation, \(x(t_*)=Ae^{-kt_*}\), for two unknowns.
Worked Example: An Initial Decline Rate Does Not Separate the Parameters
A hypothetical experiment reports an initial decline rate of \(-24.0\) mg/h, but does not separately measure the initial amount. Under the model, this result implies
For example, \(A=120.0\) mg and \(k=0.2000\ \mathrm{h}^{-1}\) give \(kA=24.0\) mg/h. So do \(A=80.0\) mg and \(k=0.3000\ \mathrm{h}^{-1}\), because \(80.0\times0.3000=24.0\). These parameter pairs predict different trajectories after time zero, but the same initial slope. The slope measurement alone therefore identifies their product, not \(A\) and \(k\) individually.
A rank calculation makes the difference precise. If the measured quantity is the initial slope \(g(A,k)=-Ak\), its sensitivity row is \((-k,-A)\). One scalar observation has at most rank one for two unknowns. By contrast, using the initial amount and the initial slope together gives the map \((A,k)\mapsto(A,-Ak)\), whose Jacobian determinant is \(-A\), nonzero for \(A>0\). Locally, those two quantities can therefore distinguish the parameters. This calculation supports the algebraic result; it does not replace the need to state what was actually measured.
Worked Example: Two Nonzero-Time Measurements
A hypothetical decay experiment reports \(x(2)=80.44\) mg and \(x(7)=29.59\) mg, with time in hours. Both observations are positive and occur at distinct times, so the two-point formulas apply.
Using the observation at \(t=2\) h to recover the initial amount gives
The rounded outputs are consistent with \(A=120.0\) mg and \(k=0.2000\ \mathrm{h}^{-1}\). The time-zero amount is inferred by extrapolating the assumed model; it was not directly observed. In an exact model, using the \(t=7\) h observation would give the same \(A\). Small discrepancies between two such calculations with measured data can arise from rounding, measurement noise, or model mismatch.
Assumptions, Diagnostics, and Limitations
The identifiability result is conditional on the model and the measurement definition. It can fail if the output is not absolute amount, if time is not calibrated, or if the data do not include information about change over time. For example, if the instrument reports an unknown constant multiple \(c x(t)\), the observed level is \(cA e^{-kt}\); the trajectory still displays a decay rate, but its level combines the unknown scale and \(A\). An unmodelled offset or a different decay process likewise changes the equations used for recovery.
Useful checks for this simple model include plotting log output against time and checking whether it is approximately linear, and checking whether pairwise rate calculations agree across distinct time pairs. Curvature or systematically changing pairwise rates can indicate that the first-order model is inadequate, though noise can also cause departures. A zero or negative observed amount is incompatible with taking a real logarithm as written and calls for examining the measurement model and data, not mechanically applying the formulas.
Structural identifiability and practical estimability should not be confused. If \(k\) is very small over a short observation interval, the output changes little; small measurement errors can then greatly affect the estimated rate. That is weak practical information, not necessarily structural nonidentifiability: an ideal exact trajectory that changes over time still separates the parameters. Conversely, repeated measurements of the same product-only quantity can reduce uncertainty about that product without resolving the individual factors.
| Observation available | What it determines under the model | Consequence |
|---|---|---|
| Absolute output at two distinct times | \(k\) from the output ratio, then \(A\) | Both parameters are structurally identifiable |
| Initial amount and initial slope | \(A\) directly, then \(k=-x'(0)/A\) | Both parameters are structurally identifiable for \(A>0\) |
| Initial slope alone | The product \(kA\) | Infinitely many positive pairs are compatible |
| One output at one time | The combination \(Ae^{-kt}\) | Infinitely many positive pairs are compatible |
A common mistake is to say, “The initial slope is measured precisely, so both the initial amount and rate are known.” Precision does not supply the missing information: the slope gives only \(kA\). The correction is to identify the observation map first, then ask whether distinct parameter pairs produce the same observed quantities. For a full absolute trajectory, the changing output separates level from rate; for a product-only summary, it does not.
Check Your Understanding
Answer each question using the model assumptions stated above. Solutions are included after the questions.
- What two features of the log-output trajectory correspond to \(A\) and \(k\)?
- A measurement gives \(x'(0)=-15.0\) mg/h and nothing else. What parameter combination is determined?
- Why are two exact positive outputs at distinct times sufficient to recover both parameters?
- Does a nearly flat, noisy trajectory prove that \(A\) and \(k\) are structurally nonidentifiable? Explain.
Solutions. (1) The intercept is \(\log A\), and the slope is \(-k\). (2) It determines \(kA=15.0\) mg/h, not the individual values. (3) Their ratio eliminates \(A\) and determines \(k\); substituting \(k\) into either output then determines \(A\). (4) No. A nearly flat trajectory may provide weak practical information, especially over a short interval, while an ideal changing trajectory under the stated model still identifies both parameters.