Interpretation as a Modeling Choice

Why the Interpretation Is Part of the Model

An SDE written as $dC_t=a(C_t)\,dt+b(C_t)\,dW_t$ is not fully specified until the meaning of its stochastic integral is clear. With a state-dependent diffusion coefficient, interpreting the noise integral in the Itô or Stratonovich sense can define different stochastic models, even if the displayed coefficient functions look identical. That distinction matters in pharmacokinetics whenever random variation is linked to the current concentration or another evolving state.

The central modeling question is not just “How large is the noise?” It is also “What does the stated drift describe?” Under the Itô convention, the drift has a precise local conditional-mean interpretation. Under the Stratonovich convention, it is the drift in a symmetric-integral representation; it should not automatically be read as that same conditional mean rate. The diffusion coefficient describes the strength and direction of process noise, not assay error.

Two Integral Interpretations

Definition: Let $W_t$ be standard Brownian motion relative to a filtration $(\mathcal F_t)$, and let $X_t$ be an adapted process. An Itô SDE $dX_t=a(X_t)\,dt+b(X_t)\,dW_t$ means the integral equation $$ X_t=X_0+\int_0^t a(X_s)\,ds+\int_0^t b(X_s)\,dW_s, $$ where the stochastic integral is the Itô integral, defined using adapted left-endpoint sums. A Stratonovich SDE $dX_t=a(X_t)\,dt+b(X_t)\circ dW_t$ uses the Stratonovich integral, defined for continuous semimartingales as the limit in probability of symmetric (midpoint) sums. In both cases, the drift and diffusion coefficients must satisfy conditions that ensure the equation has a well-defined solution; for example, globally Lipschitz coefficients with at most linear growth give a unique strong solution.

The notation $\circ dW_t$ explicitly marks the Stratonovich interpretation. Without the circle, the standard default in much mathematical and statistical work is Itô, but a pharmacometric model should state its convention rather than rely on a reader’s default. The integral definitions are the key: left-endpoint sums use information available at the start of each interval, whereas symmetric sums evaluate the integrand across the interval.

With several Brownian drivers, a vector diffusion coefficient describes the state’s sensitivity to each driver. If $B$ is the diffusion matrix in an Itô model driven by independent Brownian motions, the instantaneous covariance matrix is $BB^\mathsf{T}$. Thus the diffusion coefficient gives local process-variation information, while the drift describes the systematic component under the chosen interpretation.

What the Itô Drift Means

For an Itô model, under regularity conditions, the drift is the local conditional mean rate. The following result makes that statement precise. Suppose $a$ and $b$ are bounded, continuous, globally Lipschitz functions, and $X_t$ is the resulting solution. Then, for each fixed $t$,

\lim_{h\downarrow 0}\frac{E[X_{t+h}-X_t\mid\mathcal F_t]}{h}=a(X_t) \quad\text{in }L^1, \qquad \lim_{h\downarrow 0}\frac{\operatorname{Var}(X_{t+h}-X_t\mid\mathcal F_t)}{h}=b(X_t)^2 \quad\text{in }L^1.

Justification. Write the increment as $A_h+M_h$, where $A_h=\int_t^{t+h}a(X_s)\,ds$ and $M_h=\int_t^{t+h}b(X_s)\,dW_s$. The Itô integral has conditional mean zero, so $E[M_h\mid\mathcal F_t]=0$. Continuity of $X$ and $a$, together with boundedness, gives $h^{-1}E[A_h\mid\mathcal F_t]\to a(X_t)$ in $L^1$.

For the variance, conditional Itô isometry gives $E[M_h^2\mid\mathcal F_t]=E[\int_t^{t+h}b(X_s)^2\,ds\mid\mathcal F_t]$; dividing by $h$ and using continuity and boundedness gives the limit $b(X_t)^2$. Also $|A_h|\leq \|a\|_\infty h$, so its conditional variance is $O(h^2)$. The conditional covariance between $A_h$ and $M_h$ is $O(h^{3/2})$ by conditional Cauchy–Schwarz and the $O(h)$ bound on $E[M_h^2\mid\mathcal F_t]$. These drift and cross terms vanish after division by $h$, proving the variance limit. $\square$

This result gives the Itô coefficients an operational meaning: $a(X_t)$ is the instantaneous conditional mean rate, and $b(X_t)^2$ is the instantaneous conditional variance rate. In a multivariate model, the corresponding variance-rate object is the diffusion matrix times its transpose. These are local statements, not claims that a finite-time conditional mean or variance equals the coefficient itself.

Worked Examples

Worked Example: Additive Noise in a Concentration Model

Consider a concentration state $C_t$ satisfying $dC_t=(u-kC_t)\,dt+\sigma\,dW_t$, where $u=2$ concentration units per hour, $k=0.2$ per hour, and $\sigma=1$ concentration unit per square-root hour. Let $C_0=10$. The diffusion is constant, so its integrand does not depend on the evolving state. Left-endpoint and midpoint sums for this constant integrand are identical; the Itô and Stratonovich readings therefore agree for this model.

The deterministic equilibrium is $u/k=2/0.2=10$. The Itô drift at $C_0=10$ is $u-kC_0=2-0.2(10)=0$, so the initial local conditional mean rate is zero. The local variance rate is $\sigma^2=1$ concentration-units squared per hour.

For a finite-time check, the solution is $C_t=10e^{-0.2t}+10(1-e^{-0.2t})+\int_0^t e^{-0.2(t-s)}\,dW_s$. The first two terms sum to 10. At $t=1$ hour, the mean is therefore 10, and Itô isometry gives $$ \operatorname{Var}(C_1)=\int_0^1 e^{-0.4(1-s)}\,ds =\frac{1-e^{-0.4}}{0.4} \approx 0.8242. $$ The model represents continuous process noise in concentration. It does not represent assay error in a measured concentration.

Worked Example: Same State-Dependent Coefficients, Different Sums

Consider the noise-only state-dependent model $dC_t=0.4C_t\,dW_t$ or $dC_t=0.4C_t\circ dW_t$, with $C_0=5$. The same displayed coefficients do not by themselves settle the model: the first equation uses left-endpoint Itô sums and the second uses symmetric Stratonovich sums.

To see the distinction at the level of a numerical increment, take a step with $\Delta W=0.2$ (in square-root-hour units). The Itô left-endpoint update is $C_1=C_0+0.4C_0\Delta W=5+0.4(5)(0.2)=5.4$.

The corresponding implicit midpoint update for the Stratonovich integral is $C_1=C_0+0.4(C_0+C_1)\Delta W/2$. Substitution gives $C_1=5+0.04(5+C_1)$, hence $0.96C_1=5.2$, and $C_1=5.2/0.96\approx5.4167$. This one-step comparison illustrates the different sum conventions; it is not an exact finite-time transition calculation. Because the diffusion depends on $C_t$, a model statement must specify which convention defines the process.

At $C=5$, the diffusion coefficient is $0.4(5)=2$ concentration units per square-root hour, so its square gives a local variance rate of 4 in the Itô model. The state dependence means that the noise amplitude changes as the concentration changes; it does not mean that the assay becomes noisier.

Worked Example: Separating Process Noise From Assay Error

Suppose a latent concentration follows the additive-noise model in the first example, and an assay reports $Y_t=C_t+\epsilon_t$, where $\epsilon_t$ is independent measurement error with mean zero and variance $0.25$ concentration-units squared. At an instant when $C_t=10$, the process model’s instantaneous variance rate is $\sigma^2=1$ per hour. The assay-error variance, $0.25$, is instead a variance attached to a measurement.

These quantities have different units and roles: process variance accumulates over time, whereas the assay variance describes scatter around a latent state at a sampling time. Replicate assays can inform measurement error, but they do not by themselves establish a continuous-time diffusion. A model that intends both sources should represent both, rather than folding assay error into $b(C_t)$.

When the Convention Must Be Explicit

For a constant diffusion coefficient, the Itô and Stratonovich stochastic sums agree, so the convention does not change the model through the noise integral. With state-dependent diffusion, the sums generally differ because the coefficient changes along the Brownian-driven path. In that case, the equation needs an explicit convention, and a drift coefficient must be interpreted in relation to it. The later material on concentration-dependent diffusion conversion treats how to translate between conventions; the modeling decision here is to avoid silently treating their drift labels as interchangeable.

Model feature	What the diffusion says	Why the convention matters
Constant additive noise	Fixed local process-noise amplitude	Left and symmetric sums agree for a constant integrand
State-dependent noise	Local noise amplitude changes with the state	The Itô and Stratonovich integral definitions generally differ
Assay error only	Nothing about continuous process variation	Use an observation model; an SDE convention is not a substitute
Several Brownian drivers	Each driver contributes to state variation	State both the convention and how drivers enter the diffusion

A practical specification should state the state variable, time units, drift units, diffusion units, Brownian driver or drivers, and integral convention. The diffusion has units of state per square-root time; its square has units of state squared per time. Those checks can expose mistakes before fitting. The Itô local-moment result above is a useful interpretation check for an Itô model; it should not be applied by simply reading the drift of a Stratonovich equation as the conditional mean rate.

Common Mistakes and Limitations

Leaving the convention unstated. A state-dependent diffusion with no declared interpretation is ambiguous. State Itô or Stratonovich explicitly.
Calling the drift “the mean change” without qualification. For an Itô SDE, it is the local conditional mean rate under the stated conditions. A Stratonovich drift is the coefficient in a symmetric-integral representation, not automatically that rate.
Treating diffusion as assay error. Diffusion describes continuous process variation. Measurement error belongs in the observation model.
Assuming every noise term is state-dependent. Additive noise is constant in the state; in that case the two integral conventions agree.
Reading a single numerical step as an exact solution. Euler and midpoint updates approximate their respective integrals. Step-size behavior and model assumptions still need checking.

The interpretation does not settle whether a chosen stochastic model is biologically plausible. A diffusion can assign nonzero probability to negative concentrations, for example, and an SDE may poorly represent abrupt doses or discrete dose-to-dose variability. Model choice also depends on the intended mechanism, data resolution, and whether continuous process variation is supported by the observations. An explicit convention makes a model reproducible; it does not make an unsuitable model appropriate.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions.

For an Itô SDE with drift $a(X_t)$ and diffusion $b(X_t)$, what are the local conditional mean rate and variance rate?
Does a constant diffusion coefficient make the Itô and Stratonovich noise integrals agree? Why?
A model uses $dC_t=-kC_t\,dt+\gamma C_t\,dW_t$. What specification is needed before its meaning is unambiguous?
Why is replicate-assay scatter not, by itself, evidence for a continuous-time diffusion?

Solutions. (1) Under the stated regularity conditions, the rates are $a(X_t)$ and $b(X_t)^2$, respectively. (2) Yes. For a constant integrand, the left-endpoint and midpoint sums are the same sum, so the integral does not depend on that convention. (3) State whether the equation is Itô or Stratonovich; the coefficient depends on the state, so the integral interpretation matters. (4) Replicate scatter describes measurement variation at sampling times. A diffusion describes process variation accumulating continuously over time, so it requires a process model rather than an observation-error argument alone.