Three Questions About a Diffusion Parameter
A diffusion parameter can be difficult to estimate for three distinct reasons. Different parameter values may generate exactly the same distribution of all possible observations; that is a structural identifiability problem. Alternatively, the model may distinguish the values in principle, while the available observation times and sample size provide too little information to separate them; that is a practical identifiability problem. Finally, even an identifiable parameter may not admit a precise estimate from a particular finite data set.
These distinctions matter in pharmacokinetic SDEs because a diffusion coefficient describes variability in an unobserved path, whereas data are collected at selected times and may include measurement error. Identifiability is therefore a property of the model together with its observation design, not a property of a parameter name alone.
The last definition is an asymptotic one: it asks whether increasingly informative data can support consistent estimation. It does not mean that every finite data set yields a narrow confidence interval. In other statistical settings, “estimable” can also mean that an unbiased estimator exists; that is a different, finite-sample usage and is not the sense used here.
Structural Identifiability Is About the Observation Law
For an SDE model, structural identifiability is assessed after specifying what is observed: for example, exact states at fixed times, noisy concentrations, or repeated trajectories. A symmetry of the parameterization is a direct warning sign. If changing parameters leaves the entire observation law unchanged, no inference method can distinguish the corresponding values from those observations.
Even the simplest diffusion shows why parameter conventions matter. For \(dX_t=\sigma\,dW_t\), the law of the path depends on \(\sigma^2\), not on the sign of \(\sigma\). If the parameter space allows both signs, \(\sigma\) and \(-\sigma\) are structurally indistinguishable. One convention is to parameterize by the variance rate \(q=\sigma^2\); another is to restrict \(\sigma\geq 0\). This removes the sign symmetry, but it does not guarantee that a particular observation design is informative.
Worked Example: An Exactly Matched Noisy OU Observation
Let \(X_0=0\) and consider \(dX_t=-X_t\,dt+\sigma\,dW_t\). At time \(T=1\), observe \(Y=X_1+\epsilon\), where \(\epsilon\sim N(0,\tau^2)\) is independent of the process. Suppose \(\sigma^2\) and \(\tau^2\) are both unknown and positive. Can one observation identify them separately, even if this experiment is repeated across independent subjects?
Multiplying the SDE by \(e^t\) and integrating gives \(X_1=\sigma\int_0^1 e^{-(1-s)}\,dW_s\). The stochastic integral is Gaussian with mean zero and variance \(\sigma^2\int_0^1e^{-2(1-s)}\,ds\). Write \(c=(1-e^{-2})/2\approx0.432332358\). Then \(Y\sim N(0,c\sigma^2+\tau^2)\).
Choose the target observation variance \(0.25\). For \(\sigma^2=0.25\), set \(\tau^2=0.25-0.25c\approx0.14191691\). For \(\sigma^2=0.40\), set \(\tau^2=0.25-0.40c\approx0.07706706\). Both error variances are positive, and both parameter pairs give exactly \(Y\sim N(0,0.25)\), since each has \(c\sigma^2+\tau^2=0.25\). Independent repetitions of this same observation design have the same joint distribution under either pair. Thus the two variance components are not structurally identifiable from these data, regardless of how many subjects are observed.
The identifiable quantity in this design is the sum \(c\sigma^2+\tau^2\), not its two components. More observations of the same type can estimate that sum more accurately, but cannot break the exact symmetry. A different observation design or additional information about one component would be needed.
Estimability and a Proof of Its Structural Limit
A useful consequence of structural nonidentifiability is that no estimator can be consistent for the full parameter at two distinct parameter values that produce the same observation law. This is not a limitation of a particular optimizer, filter, or Bayesian algorithm; it follows from probability alone.
Proof. Choose \(\varepsilon\) smaller than half the distance between \(\theta_1\) and \(\theta_2\). The sets of estimates within \(\varepsilon\) of \(\theta_1\) and within \(\varepsilon\) of \(\theta_2\) are disjoint. Consistency at \(\theta_1\) would make the probability of the first set tend to one. Consistency at \(\theta_2\) would make the probability of the second set tend to one. But for each \(n\), the distribution of the data—and hence of \(\hat\theta_n\)—is identical under the two parameter values. Under one common distribution, two disjoint events cannot both have probabilities tending to one. This is a contradiction. \(\square\)
The theorem also clarifies why an estimator of an identifiable function can still exist when the full parameter vector is not identifiable. In the OU example, the function \(c\sigma^2+\tau^2\) is the same across equivalent parameter pairs, so it is not ruled out by this argument; the separate components are.
Worked Diffusion Estimation with Exact Brownian Observations
Worked Example: Estimating a Brownian Variance Rate
Suppose \(X_t\) follows \(dX_t=\sqrt{q}\,dW_t\), with known \(X_0\), unknown \(q>0\), and exact observations at fixed times \(0=t_0<t_1<\cdots<t_n\). Let \(\Delta_i=t_i-t_{i-1}\) and \(Z_i=X_{t_i}-X_{t_{i-1}}\). Find the maximum-likelihood estimator and its sampling variance.
The increments are independent, with \(Z_i\sim N(0,q\Delta_i)\). Their joint log-likelihood, up to constants not depending on \(q\), is \(\ell(q)=-\frac{n}{2}\log q-\frac{1}{2q}\sum_{i=1}^n Z_i^2/\Delta_i\). Differentiating gives \(\ell'(q)=-n/(2q)+\sum_i Z_i^2/(2q^2\Delta_i)\). Setting this to zero yields \(\hat q=n^{-1}\sum_i Z_i^2/\Delta_i\). The second derivative at this solution is negative, so it is the maximum.
For each increment, \(Z_i/\sqrt{q\Delta_i}\sim N(0,1)\), and hence \(Z_i^2/(q\Delta_i)\) has a chi-squared distribution with one degree of freedom. Independence gives \(n\hat q/q=\sum_i Z_i^2/(q\Delta_i)\sim\chi_n^2\). Therefore \(E[\hat q]=q\) and, using the chi-squared variance \(2n\), \(\operatorname{Var}(\hat q)=2q^2/n\). This model and design identify \(q\): the increment distributions determine their variances \(q\Delta_i\), and each \(\Delta_i\) is positive. Precision still depends on \(n\); a structurally identifiable diffusion rate can be noisy when few increments are observed.
This calculation relies on exact states, zero drift, and independent Brownian increments over known intervals. Changing those assumptions changes the likelihood and may change what can be identified. The example is a benchmark for distinguishing a clear structural result from the more difficult practical question of how much information a real design supplies.
Practical Identifiability: Information in a Finite Design
Practical identifiability is not a second form of exact equality. It describes weak but nonzero separation between observation distributions over the parameter values of interest. A likelihood may have a unique maximum and still be broad, asymmetric, or nearly flat in one parameter direction. The observation schedule, measurement precision, number of subjects, and parameter scale all affect that uncertainty.
Worked Example: Information About Diffusion with Measurement Error
Let \(dX_t=-X_t\,dt+\sqrt{\lambda}\,dW_t\), with \(X_0=0\). For each of \(m\) independent subjects, observe \(Y=X_{0.05}+\epsilon\), where \(\epsilon\sim N(0,0.01)\) independently. Calculate the Fisher information per observation about \(\lambda\) at \(\lambda=0.04\), and the information-based standard error for \(m=10\).
The OU solution gives \(X_{0.05}\sim N(0,a\lambda)\), where \(a=\int_0^{0.05}e^{-2(0.05-s)}\,ds=(1-e^{-0.1})/2\approx0.04758129\). Consequently \(Y\sim N(0,V)\), with \(V=0.01+a\lambda\). At \(\lambda=0.04\), \(V\approx0.01+(0.04758129)(0.04)=0.01190325\).
For one observation, the log-density, apart from constants, is \(\ell(\lambda)=-\frac12\log V-\frac{Y^2}{2V}\), where \(dV/d\lambda=a\). Its score is \(\ell'(\lambda)=a(Y^2-V)/(2V^2)\). Since a centered normal variable of variance \(V\) has \(\operatorname{Var}(Y^2)=2V^2\), the expected squared score is \(I(\lambda)=a^2/(2V^2)\). Substituting the values gives \(I(0.04)\approx(0.04758129)^2/[2(0.01190325)^2]\approx7.989\) per observation. For ten independent observations, the information is approximately \(79.89\), giving the Fisher-information-based standard error \(1/\sqrt{79.89}\approx0.112\).
This standard error is a local information benchmark, not an exact finite-sample confidence interval. It is large relative to the value \(0.04\), indicating that ten observations may provide weak practical precision even though the model’s variance \(0.01+a\lambda\) changes with \(\lambda\) and therefore identifies it in principle.
The information calculation has a general interpretation. When the likelihood is smooth and regular, its curvature near a maximum measures how quickly the fit worsens as a parameter changes. A small curvature corresponds to a broad range of plausible values. For multiple parameters, the information matrix has directions, not just a single magnitude: a very small eigenvalue signals a combination of parameters that the design distinguishes poorly.
Diagnostics, Alternatives, and Limitations
A useful analysis starts by checking the observation law for exact symmetries, before examining numerical standard errors. Then assess finite-data uncertainty using complementary summaries rather than treating one optimizer output as proof of identification.
- Check parameter transformations. Ask whether distinct allowed parameter values leave the transition or observation distributions unchanged. Pay particular attention to parameters that enter only through a product, sum, or square.
- Inspect likelihood profiles. Fix one parameter at a sequence of values and re-optimize the others. A flat or nearly flat profile, a long ridge, or a boundary maximum indicates weak separation or model tension.
- Inspect information-matrix conditioning. A near-zero eigenvalue or a very large condition number signals a poorly determined direction. Scale parameters sensibly before comparing eigenvalues, since numerical conditioning can depend on units.
- Check sensitivity to the design. Compare which observation times, replicate counts, or measurement precisions would change the relevant distributions. A parameter that is weakly informed under one schedule may be better informed under another.
These diagnostics answer different questions. A symbolic symmetry or an exact equality of distributions establishes structural nonidentifiability. A broad profile or nearly singular information matrix suggests practical weakness but does not by itself prove exact nonidentifiability. Conversely, optimizer convergence only shows that an algorithm found a numerical solution; it does not establish that the likelihood has a unique, informative maximum.
| Approach | Question answered | Main limitation |
|---|---|---|
| Observation-law or symmetry analysis | Can distinct parameters produce exactly the same data distribution? | Requires a correct model and observation specification. |
| Likelihood profiles and information | How sharply does this finite design distinguish nearby values? | Local curvature can miss distant modes or boundary effects. |
| Bayesian posterior summaries | How does the data update a specified prior? | A proper, narrow posterior can reflect prior information even when the likelihood is weak or nonidentifying. |
No one diagnostic replaces the others. In particular, a prior can regularize an estimation problem and produce a usable posterior, but it does not change an exact symmetry in the likelihood. Practical conclusions also depend on the assumed SDE, the measurement model, the observation schedule, and the parameter region being considered. The criteria developed here do not determine how to handle drift nuisance parameters or state-dependent diffusion functions; those require additional model-specific analysis.
Common Pitfall
A common mistake is to call a parameter “identified” because an estimation algorithm returns one numerical value, or because two rounded variances look equal. A single optimizer result does not rule out a ridge, another mode, or an exact symmetry. And rounded values that coincide do not establish equality of distributions: exact structural claims require exact equality or a valid proof. State the observation law, check the parameter mapping, and keep approximation separate from equality.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- For \(dX_t=\sigma\,dW_t\), with \(\sigma\) allowed to be positive or negative, which parameter values generate the same path law?
- In the noisy OU example, what identifiable function of \(\sigma^2\) and \(\tau^2\) is determined by the one-time observation distribution?
- For the Brownian model with \(n\) exact increments, what is the variance of \(\hat q\), and what happens to it as \(n\) increases with \(q\) fixed?
- Does a narrow Bayesian posterior by itself prove that the likelihood structurally identifies every parameter? Explain.
Solutions. (1) The values \(\sigma\) and \(-\sigma\) produce the same law because the centered Gaussian increments depend on \(\sigma\) only through \(\sigma^2\). (2) The distribution determines \(c\sigma^2+\tau^2\), where \(c=(1-e^{-2})/2\); it does not determine the two terms separately. (3) \(\operatorname{Var}(\hat q)=2q^2/n\), which tends to zero as \(n\) increases, showing improving precision in this design. (4) No. Posterior concentration can be influenced by the prior; exact structural identifiability concerns whether the likelihood’s observation distributions differ across parameter values.