Initial State Distributions and Diffuse Initialization

Why the Initial State Matters

A Kalman filter begins with a distribution for the state before it uses the first observation. In pharmacokinetics, that state might contain compartment amounts, concentrations, or latent effects. If its initial mean or uncertainty is poorly specified, the first concentration updates can reflect the chosen starting assumption as much as the data. This matters especially when observations begin soon after a dose, when the model has not yet had time to reduce uncertainty through state evolution.

An informative initialization assigns finite uncertainty to every state direction. A diffuse initialization represents a direction whose initial level is not assigned a proper finite variance. Diffuse initialization is not a claim that the state has literally infinite variance; it is a limiting device that lets the data determine an initially unknown level without choosing an arbitrary, very large variance.

This tutorial uses a linear Gaussian state transition \(x_{k+1}=A_kx_k+b_k+w_k\), with \(w_k\sim N(0,Q_k)\), and a linear concentration observation \(y_k=H_kx_k+\varepsilon_k\), with \(\varepsilon_k\sim N(0,R_k)\). The observation errors are independent of the state and process errors, and \(R_k>0\) for the scalar observations considered below. The known input \(b_k\), including a known dose contribution, changes the predicted mean but not the covariance. For a concentration obtained by scaling a central amount \(x_k\), for example, \(H_k\) can include the reciprocal of the relevant volume.

Propagating Initial Uncertainty

The state prediction carries both parts of the initial covariance forward. If the current mean is \(a_k\), the predicted mean is \(a_{k+1|k}=A_ka_k+b_k\). For an informative covariance \(P_k\), the predicted covariance is \(P_{k+1|k}=A_kP_kA_k^{\mathsf T}+Q_k\). For a diffuse representation, apply the same linear transformation separately to its two components:

The distinction is important: process noise adds to the finite component, not to the diffuse component. A state transition may also change the rank of the diffuse component. For example, if a transition maps an initially unknown direction to zero, that uncertainty no longer appears in the next state; if the transition preserves the direction, the uncertainty remains diffuse.

Exact Diffuse Measurement Updates

Consider one scalar observation \(y=Hx+\varepsilon\), where \(H\) is a row vector and \(\varepsilon\sim N(0,R)\). Suppose the predicted state has mean \(a\) and covariance \(\kappa P_\infty+P_*\). Define the innovation \(v=y-Ha\), the diffuse innovation variance \(F_\infty=HP_\infty H^{\mathsf T}\), and the finite innovation variance \(F_*=HP_*H^{\mathsf T}+R\). The following update applies when \(F_\infty>0\), so this observation contains information about at least one diffuse direction.

Here \(a^+\), \(P_\infty^+\), and \(P_*^+\) denote the updated mean and covariance components. The finite covariance formula accounts for the first finite-order contribution to the ordinary Gaussian update; it is not obtained by simply applying an ordinary Kalman update to \(P_\infty\).

For completeness, set \(U=P_\infty H^{\mathsf T}\) and \(W=P_*H^{\mathsf T}\). The gain expansion is \(K(\kappa)=U/F_\infty+\kappa^{-1}(W/F_\infty-UF_*/F_\infty^2)+O(\kappa^{-2})\). Substituting this and \(P=\kappa P_\infty+P_*\) into \(P^+=P-K(\kappa)HP\), the coefficient of \(\kappa\) is \(P_\infty-UU^{\mathsf T}/F_\infty\). The finite term simplifies to \(P_*-K_\infty HP_*-P_*H^{\mathsf T}K_\infty^{\mathsf T}+K_\infty F_*K_\infty^{\mathsf T}\), as stated. This proves the scalar-observation update formulas. \(\square\)

What the First Concentration Data Can—and Cannot—Resolve

A concentration observation only reduces diffuse uncertainty in directions that affect the observed concentration. For a scalar observation, \(F_\infty=HP_\infty H^{\mathsf T}\) measures the diffuse contribution to its uncertainty. If \(F_\infty=0\), this particular observation has no diffuse component in its prediction. The update above is then inapplicable: it cannot be used with division by zero, and it cannot claim to resolve a diffuse direction. The measurement may still update finite uncertainty through \(F_*\).

In a pharmacokinetic application, whether an initial amount is identifiable from an early concentration depends on the observation mapping and state dynamics, not just on how many measurements are collected. Repeated observations can resolve additional directions if their observation mappings or intervening transitions provide new information. Repeating an observation that is insensitive to a remaining diffuse direction does not, by itself, make that direction finite.

Informative, Exact Diffuse, and Large-Variance Choices

An exact diffuse update is a principled alternative to setting an extremely large initial variance. For a finite but large \(\kappa\), an ordinary update remains mathematically valid, but computations can lose precision when large covariance terms are combined or when the filter is implemented with limited numerical accuracy. Conversely, exact diffuse initialization is not a substitute for a genuinely informative prior: if prior knowledge supports a finite uncertainty, encoding that information in \(P_*\) is more appropriate than declaring the direction diffuse.

A useful diagnostic is to inspect the rank and scale of \(P_\infty\) after each early observation. A decrease in rank indicates that some diffuse uncertainty has been resolved; directions left in the null space of the observation mapping may remain diffuse. Also check that the finite covariance remains symmetric and positive semidefinite up to numerical tolerance. A negative variance, an unexplained increase in diffuse rank, or a division by a near-zero \(F_\infty\) can indicate a coding or conditioning problem. When \(F_\infty\) is zero or numerically tiny, do not apply the diffuse scalar formula as if the observation resolved diffuse uncertainty.