Continuity of the Solution of a Non-Parametric Bayesian Statistical Calibration Procedure

1. Problem Statement

The paper addresses the Statistical Calibration Problem (SCP), a specific type of stochastic inverse problem. The goal is to infer the distribution of input parameters (the "Trial-Generating Distribution" or TGD, denoted $P^t_\Lambda$) for a computer model (simulator) $Q: \Lambda \to D$, given observed field data distributed according to a pushforward measure $P_D = Q P^t_\Lambda$.

• The Challenge: The map $Q$ is often many-to-one (non-injective), meaning multiple input configurations can produce the same output. Consequently, the inverse problem is ill-posed; many different input distributions could generate the same observed output distribution.
• The Existing Approach: Previous work (e.g., Prasadan et al., 2022) proposed a non-parametric Bayesian solution using the disintegration of measures. This approach updates a prior distribution $P_p$ using observed data to produce a posterior distribution $\hat{P}_\Lambda$ on the input space such that $Q \hat{P}_\Lambda = P_D$.
• The Gap: While the existence and density properties of this solution were established, the continuity and stability of the solution operator with respect to perturbations in the input distribution (TGD) had not been rigorously proven. In practical applications, experimental conditions change over time or space, and input distributions may evolve or converge to deterministic limits (Dirac measures). It is crucial to know if small changes in the input distribution lead to small changes in the inferred solution.

2. Methodology

The authors analyze the continuity of the SCP solution operator using tools from measure theory, specifically disintegration of measures and surface integrals.

• Framework:

  • Input Space ($\Lambda$): A compact subset of $\mathbb{R}^n$.
  • Output Space ($D$): A subset of $\mathbb{R}^m$.
  • Map ($Q$): A smooth, non-injective map representing the computer model.
  • Prior ($P_p$): A user-specified prior distribution on $\Lambda$ with density $\rho_p$.
  • Solution: The SCP solution $\hat{P}_\Lambda$ is constructed by disintegrating the prior $P_p$ over the contours (level sets) of $Q$. The density of the solution is given by:
    $$\hat{\rho}_\Lambda(\lambda) = \frac{\rho_D(Q(\lambda)) \rho_p(\lambda)}{\tilde{\rho}_{p,D}(Q(\lambda))}$$
    where $\rho_D$ is the density of the observed pushforward, and $\tilde{\rho}_{p,D}$ is the pushforward density of the prior (a surface integral over the contour $Q^{-1}(q)$).

• Analytical Tools:

  • Total Variation (TV) Metric: Used to measure the distance between probability measures.
  • Weak Convergence: Used to analyze sequences of measures converging to limits (including Dirac measures).
  • Co-area Formula: Used to represent conditional probabilities as surface integrals over the level sets of $Q$.
  • Assumptions: The analysis relies on regularity conditions (Assumptions 1–4) regarding the smoothness of $Q$, the rank of its Jacobian, and the properties of the prior. Crucially, the paper introduces Assumption 7 to handle limiting distributions that may not have densities (e.g., Dirac measures).

3. Key Contributions

The paper makes three primary theoretical contributions regarding the stability of the SCP:

1. Uniform Continuity in Total Variation (TV):
   The authors prove that the SCP solution operator is non-expansive and uniformly continuous with respect to the Total Variation metric.

   • Result: If the input TGDs $P^t_\Lambda$ and $R^t_\Lambda$ are close in TV distance, their corresponding SCP solutions $\hat{P}_\Lambda$ and $\hat{R}_\Lambda$ are at least as close (in fact, the distance is bounded by the distance of their pushforwards).
   • Implication: This guarantees that small perturbations in the experimental input distribution (e.g., due to noise or slight shifts in sampling) result in controlled, small changes in the inferred parameter distribution.

2. Local Limit Theorem (Pointwise to TV Convergence):
   The paper establishes that if a sequence of TGD densities converges pointwise almost everywhere to a limit density, the corresponding sequence of SCP solution densities converges in the Total Variation metric. This connects standard density convergence to the stability of the calibration procedure.

3. Weak Continuity for Broad Classes of Distributions:
   This is the most significant extension. The authors prove that the SCP operator is weakly continuous.

   • Scope: This holds for sequences of TGDs converging to limits that are not necessarily absolutely continuous with respect to the Lebesgue measure.
   • Specific Limits: The results explicitly cover limits that are Dirac measures (deterministic parameters) or finite mixtures of Dirac measures.
   • Constructive Results: The paper provides explicit constructions for the limiting SCP solution in these cases. For a Dirac limit $\delta_{\lambda^*}$, the solution converges to the regular conditional probability of the prior given the output $Q(\lambda^*)$, i.e., $P_p(\cdot | Q(\lambda^*))$.

4. Key Results and Theorems

• Theorem 3.1 (TV Stability): $d_{TV}(\hat{P}_\Lambda, \hat{R}_\Lambda) \leq d_{TV}(P^t_\Lambda, R^t_\Lambda)$. The solution operator does not amplify errors in the input distribution.
• Theorem 3.2 (Local Limit): Pointwise convergence of TGD densities implies TV convergence of SCP solutions.
• Theorem 4.1 (Weak Convergence): If a sequence of TGDs converges weakly to a limit satisfying specific regularity conditions (Assumption 7), the sequence of SCP solutions converges weakly to a unique limit measure.
• Theorem 4.3 (Dirac Limit): If the TGD converges weakly to a Dirac measure $\delta_{\lambda^*}$, the SCP solution converges weakly to the conditional prior distribution $P_p(\cdot | Q(\lambda^*))$. This mathematically unifies the stochastic calibration framework with deterministic inverse problems.

5. Significance and Applications

• Theoretical Unification: The results bridge the gap between Bayesian calibration for random parameters and deterministic inverse problems. They show that treating a fixed unknown parameter as a random variable with vanishing variance (a sequence of TGDs converging to a Dirac) yields a consistent solution.
• Robustness in Practice: The continuity results validate the use of SCP in dynamic environments where experimental conditions drift (e.g., changing soil properties in groundwater modeling or aging materials in structural engineering). It assures practitioners that the calibration will adapt smoothly to these changes rather than exhibiting erratic behavior.
• Handling Discrete/Deterministic Limits: By proving weak continuity for Dirac measures, the paper allows the SCP framework to be applied to problems where the "true" parameter is fixed but unknown, a common scenario in physical sciences.
• Numerical Validation: The authors demonstrate these properties through two examples:
  1. A synthetic Gaussian mixture converging to point masses, showing the solution concentrating on the correct level sets.
  2. A concrete compressive strength dataset, showing how the solution adapts when the data subset (concrete age) changes, demonstrating the continuity of the solution with respect to the data distribution.

In summary, this paper provides the rigorous mathematical foundation necessary to trust the non-parametric Bayesian calibration procedure in real-world scenarios where data distributions are imperfect, evolving, or converge to deterministic limits.