Known and Unknown Initial States in Turnover Models

Why the Initial State Matters

Turnover models describe a quantity that is continually produced and removed, such as a circulating biomarker with ongoing synthesis and first-order degradation. The initial amount can be known from a baseline measurement, or it can be unknown and estimated along with the kinetic parameters. These are different assumptions about the information available to identify the model, but an unknown initial state does not automatically make the kinetic parameters structurally unidentifiable.

The key is the shape of the observed trajectory. A nonconstant trajectory contains a transient whose exponential rate and long-run level can, under ideal observation, distinguish turnover from production. A constant trajectory has no transient: it reveals only a balance between production and removal. This tutorial compares those cases for a simple turnover ODE and makes explicit the assumptions behind the conclusions.

Model and assumptions: Let $R(t)$ be a directly observed state, measured in amount per volume, and suppose it satisfies $$ \frac{dR}{dt}=k_{\mathrm{in}}-k_{\mathrm{out}}R(t), \qquad R(0)=R_0. $$ Here $k_{\mathrm{in}}>0$ is a constant production rate with units of amount per volume per time, and $k_{\mathrm{out}}>0$ is a first-order removal rate with units of inverse time. The parameters are constant; there is no external input, observation scaling, measurement delay, or process noise. Structural conclusions below concern an exact output trajectory observed over an interval, not a finite noisy data set.

Output Expression and the Role of the Initial State

The equilibrium value is the state at which production and removal balance. Setting the derivative to zero gives $R_{\mathrm{eq}}=k_{\mathrm{in}}/k_{\mathrm{out}}$. Solving the ODE makes the transient from $R_0$ to that equilibrium explicit.

\frac{dR}{dt}+k_{\mathrm{out}}R=k_{\mathrm{in}}.

Multiplying by the integrating factor $e^{k_{\mathrm{out}}t}$ gives $\frac{d}{dt}\left(e^{k_{\mathrm{out}}t}R(t)\right)=k_{\mathrm{in}}e^{k_{\mathrm{out}}t}$. Integrating from 0 to $t$, then solving for $R(t)$, yields

R(t)=R_0e^{-k_{\mathrm{out}}t} +\frac{k_{\mathrm{in}}}{k_{\mathrm{out}}}\left(1-e^{-k_{\mathrm{out}}t}\right) =R_{\mathrm{eq}}+(R_0-R_{\mathrm{eq}})e^{-k_{\mathrm{out}}t}.

If $R_0$ is known, it is fixed by independent information or a baseline observation; it is not an additional parameter to infer from the trajectory. If $R_0$ is unknown, it is an additional unknown alongside $k_{\mathrm{in}}$ and $k_{\mathrm{out}}$. In either case, a nonconstant ideal trajectory has an exponential transient, and its rate and asymptotic level carry information about the kinetic parameters.

Structural identifiability: A parameter is structurally identifiable under a specified model and observation setup if distinct admissible parameter values cannot produce the same ideal output trajectory. This is a statement about uniqueness with exact, sufficiently rich observations. It does not guarantee that noisy or sparse measurements estimate the parameter precisely.

What a Nonconstant Trajectory Identifies

Differentiating the ODE once gives a useful input-output relation. Since $k_{\mathrm{in}}$ and $k_{\mathrm{out}}$ are constant, $R''(t)=-k_{\mathrm{out}}R'(t)$. At any time when the trajectory is changing, this gives the removal rate from the ratio of derivatives. The original ODE then gives the production rate.

k_{\mathrm{out}}=-\frac{R''(t)}{R'(t)}, \qquad k_{\mathrm{in}}=R'(t)+k_{\mathrm{out}}R(t), \qquad \text{when } R'(t)\ne 0.

This is also a short uniqueness argument. If two parameter pairs generate the same nonconstant exact output, they generate the same $R'$ and $R''$ at times when $R'\ne0$. The derivative ratio forces the same $k_{\mathrm{out}}$; substituting into the ODE forces the same $k_{\mathrm{in}}$. The initial state is $R(0)$, so it is also determined by the output if it was not supplied in advance.

The output expression gives the same conclusion in another form. A nonconstant trajectory can be written $A+Ce^{-\lambda t}$, where $\lambda=k_{\mathrm{out}}$, $A=k_{\mathrm{in}}/k_{\mathrm{out}}$, and $C=R_0-A$. Its exponential rate identifies $k_{\mathrm{out}}$; its asymptote identifies $A$, so $k_{\mathrm{in}}=k_{\mathrm{out}}A$; and its value at time zero identifies $R_0=A+C$. If $R_0$ was already known, the last quantity is a check on the model rather than an additional unknown.

Worked Example: Known Initial State and a Nonconstant Response

Hypothetical data. A biomarker has a known baseline $R_0=10$ mg/L. An ideal continuous response is nonconstant, with initial derivatives $R'(0)=8$ mg/L/day and $R''(0)=-3.2$ mg/L/day$^2$. Recover the two kinetic parameters.

The derivative relation gives

k_{\mathrm{out}}=-\frac{R''(0)}{R'(0)} =-\frac{-3.2\ \mathrm{mg/L/day^2}}{8\ \mathrm{mg/L/day}} =0.400\ \mathrm{day^{-1}}.

Then use the ODE at baseline:

k_{\mathrm{in}}=R'(0)+k_{\mathrm{out}}R_0 =8\ \mathrm{mg/L/day} +(0.400\ \mathrm{day^{-1}})(10\ \mathrm{mg/L}) =12.0\ \mathrm{mg/L/day}.

The implied equilibrium is $12.0/0.400=30.0$ mg/L, and the fitted trajectory is $R(t)=30.0-20.0e^{-0.400t}$ mg/L, with $t$ in days. It begins at 10 mg/L and increases toward 30 mg/L. The known baseline is consistent with the trajectory's intercept.

Worked Example: Unknown Initial State and a Nonconstant Response

Hypothetical data. Suppose an ideal response is $R(t)=20.0-15.0e^{-0.300t}$ mg/L, where time is in days, but the baseline was not supplied separately. Show what can be recovered from the output.

The exponential rate is $0.300\ \mathrm{day^{-1}}$, so $k_{\mathrm{out}}=0.300\ \mathrm{day^{-1}}$. The asymptote is $20.0$ mg/L, giving $k_{\mathrm{in}}=(0.300\ \mathrm{day^{-1}})(20.0\ \mathrm{mg/L})=6.00\ \mathrm{mg/L/day}$. The initial state follows from evaluating the trajectory at zero:

R_0=R(0)=20.0-15.0=5.00\ \mathrm{mg/L}.

Thus the three unknowns $R_0$, $k_{\mathrm{in}}$, and $k_{\mathrm{out}}$ are all determined by this ideal nonconstant curve. Treating $R_0$ as unknown adds a parameter, but the observed intercept and transient amplitude provide the corresponding information in this case.

The Constant-Output Exception

The derivative argument requires a changing trajectory. If $R(t)=c$ throughout the observation interval, then $R'=0$, and the ODE reduces to $k_{\mathrm{in}}=k_{\mathrm{out}}c$. The output identifies the balance, not the two rates separately. This is a genuine structural ambiguity: many distinct positive parameter pairs produce exactly the same output.

R(t)=c \quad\Longrightarrow\quad k_{\mathrm{in}}=k_{\mathrm{out}}c,\qquad R_0=c.

Knowing the initial state does not remove this ambiguity if it equals the constant observed state. It confirms $R_0=c$, but it does not reveal how quickly production and removal are each occurring. Conversely, if a known $R_0$ differs from a constant observed value, the exact model assumptions cannot all hold: a solution with positive $k_{\mathrm{in}}$ and $k_{\mathrm{out}}$ cannot jump from $R_0$ to $c$ at time zero.

Worked Example: Two Parameter Pairs With the Same Constant Output

Hypothetical data. A directly observed biomarker remains at $10.0$ mg/L for the entire ideal observation interval, and its initial state is known to be $10.0$ mg/L. Consider two candidate parameter pairs.

For $k_{\mathrm{out}}=0.200\ \mathrm{day^{-1}}$, the balance condition requires $k_{\mathrm{in}}=(0.200)(10.0)=2.00\ \mathrm{mg/L/day}$. For $k_{\mathrm{out}}=0.500\ \mathrm{day^{-1}}$, it requires $k_{\mathrm{in}}=(0.500)(10.0)=5.00\ \mathrm{mg/L/day}$.

R(t)=10.0+(10.0-10.0)e^{-k_{\mathrm{out}}t}=10.0\ \mathrm{mg/L}

Both pairs produce precisely the same output for every $t$. In fact, any $k_{\mathrm{out}}>0$ paired with $k_{\mathrm{in}}=10.0k_{\mathrm{out}}$ in compatible units does so. The initial state is known, yet the kinetic rates are not separately identifiable from this constant trajectory.

Assumptions, Diagnostics, and Practical Limits

The conclusions depend on observing $R(t)$ itself on a known scale and on having the stated constant-rate model. If the measurement is instead $y(t)=sR(t)$ with an unknown scale $s$, or if the observed series is too short to show its transient, the recovery argument changes. Such complications are not part of the model used here and must not be silently treated as if the state were directly observed.

In finite experimental data, derivatives are not known exactly. Estimating a second derivative from noisy measurements can be unstable, so the derivative formulas establish structural uniqueness but are not necessarily the best numerical fitting procedure. Fitting the closed-form trajectory directly is a common alternative. For a well-resolved nonconstant curve, a diagnostic plot should show a coherent transient toward a stable level; systematic deviations from a single exponential can signal that constant production or first-order removal is inadequate.

Sparse observations, a short follow-up period, or a small transient can make the asymptote and decay rate difficult to estimate even when they are structurally identifiable. This is practical uncertainty, not the exact nonuniqueness of the constant-output case. Repeated observations over a sufficiently informative time interval can help with practical precision, but they cannot separate $k_{\mathrm{in}}$ and $k_{\mathrm{out}}$ from a truly constant output under this model.

Check the observation setup.
Confirm that the measured quantity is the state $R(t)$, time is known, and the model assumes constant $k_{\mathrm{in}}$ and $k_{\mathrm{out}}$.

Check for a transient.
If the ideal output is nonconstant, its exponential rate and equilibrium level identify the two kinetic parameters; $R(0)$ gives the initial state if needed.

Handle a constant response separately.
Use $k_{\mathrm{in}}=k_{\mathrm{out}}R$ and report that the two rates are not separately identified by this output.

Separate uniqueness from precision.
A theoretically identifiable transient may still be poorly estimated from noisy or sparse measurements.

Common Mistakes

Assuming an unknown initial state always confounds the rates. For a nonconstant exact trajectory, the transient rate and asymptote identify $k_{\mathrm{out}}$ and $k_{\mathrm{in}}$, while the intercept identifies $R_0$.
Assuming a known baseline guarantees identifiability. If the state remains at that baseline, the data give only $k_{\mathrm{in}}=k_{\mathrm{out}}R_0$.
Calling poor precision structural nonidentifiability. A short, noisy transient may yield uncertain estimates, even though distinct parameters cannot produce the same complete ideal nonconstant curve.
Interpreting a flat measured series as proof of exact equilibrium. Real data may look flat because the transient is small relative to measurement noise. The structural conclusion applies to an exactly constant ideal output under the stated model.

Key takeaway: In the first-order turnover model, a nonconstant ideal state trajectory identifies the turnover rate, production rate, and—if not already known—the initial state. A constant trajectory identifies only the production-to-removal balance, whether or not the initial state is known.

Check Your Understanding

Answer each question using the stated first-order turnover model. Solutions follow the questions.

For $R(t)=18-8e^{-0.25t}$ mg/L, with time in days, identify $R_0$, $k_{\mathrm{out}}$, and $k_{\mathrm{in}}$, including units.
If $R(t)=7.0$ mg/L exactly throughout observation and $R_0=7.0$ mg/L, what relation between the kinetic parameters is identified?
Why does an unknown $R_0$ not by itself prevent identifying both rates from a nonconstant ideal trajectory?
A known baseline is 12 mg/L, but an exact trajectory is constant at 9 mg/L from time zero onward. Can this be explained by the model with positive constant rates? Explain.

Solutions. (1) The exponential rate gives $k_{\mathrm{out}}=0.25\ \mathrm{day^{-1}}$, and the asymptote gives $k_{\mathrm{in}}=(0.25)(18)=4.50\ \mathrm{mg/L/day}$. At zero, $R_0=18-8=10$ mg/L. (2) The relation is $k_{\mathrm{in}}=7.0k_{\mathrm{out}}$, with compatible units; infinitely many positive rate pairs satisfy it. (3) The transient rate identifies $k_{\mathrm{out}}$, the asymptote identifies $k_{\mathrm{in}}/k_{\mathrm{out}}$, and the value at zero identifies $R_0$. (4) No. The solution is continuous and must start at $R(0)=12$ mg/L, so it cannot be exactly 9 mg/L from time zero under the stated assumptions.