Exact Epistemic Uncertainty

• Exact Epistemic Uncertainty is defined as the precise quantification of uncertainty arising from both variance-driven (data and procedural randomness) and bias-driven (model misspecification) components.
• The methodology emphasizes a rigorous bias–variance decomposition, separating systematic model error from irreducible aleatoric noise using simulation-based estimation and information-theoretic measures.
• Practices recommended include decomposing epistemic risk into bias and variance terms, ensuring accurate uncertainty calibration crucial for high-stakes applications such as scientific simulation and safety-critical systems.

Exact epistemic uncertainty quantifies an agent's lack of knowledge about the true data-generating process or model parameters, distinct from irreducible (aleatoric) noise. In contrast to ad hoc or approximate heuristics, "exact" epistemic uncertainty is characterized by metrics, decompositions, and evaluation protocols that are theoretically justified, permit bias-aware estimation, and, where possible, admit computation with full decomposition into bias and variance components, often supported by simulation or information-theoretic measures. Rigorous research has established that without precise accounting of model bias, variance due to finite data, and procedural randomness, prevailing machine learning methodologies systematically underestimate epistemic risk and misattribute systematic error to aleatoric uncertainty.

1. Formal Decompositions: Epistemic Uncertainty as Bias, Variance, and Model Misspecification

A foundational result is the bias–variance decomposition of the expected prediction error for supervised regression. Let $x \in \mathcal X$ be a test input and $y \sim p(y|x)$, with $y = f(x) + \epsilon$, $\epsilon \sim \mathcal N(0,\sigma^2(x))$. For a learning algorithm producing a prediction $\hat y = \hat y_{N,\gamma}(x)$ (random due to data $\mathcal D_N$ and procedural randomness $\gamma$), the total mean-squared error admits the following decomposition (Jiménez et al., 29 May 2025):

$$\mathbb{E}_{y|\ x}\mathbb{E}_{\mathcal D_N,\gamma}[(y - \hat y)^2] = \underbrace{\sigma^2(x)}_{\text{aleatoric}} + \underbrace{\operatorname{Var}_{\mathcal D_N,\gamma}[\hat y]}_{\text{variance-driven epistemic}} + \underbrace{(\mathbb{E}_{\mathcal D_N,\gamma}[\hat y] - f(x))^2}_{\text{bias-driven epistemic}}.$$

Epistemic uncertainty thus consists of two primary contributions:

• Variance-driven: Sensitivity to training data and sampling or procedural choices.
• Bias-driven: Systematic error due to model misspecification or inadequate hypothesis space.

The variance-driven epistemic uncertainty itself can be further decomposed via the law of total variance into:

• Procedural uncertainty: Expected variance from algorithmic randomness (e.g., initialization, stochastic training).
• Data-driven uncertainty: Variance from different dataset realizations.

It is essential to recognize that model bias (i.e., hypothesis-class misspecification) is formally part of epistemic uncertainty; this term is irreducible unless the hypothesis class is rich enough to contain the true conditional $p(y|x)$.

2. Failure Modes of Standard Second-Order Uncertainty Quantification

Prevailing second-order uncertainty quantification methods—such as Bayesian deep ensembles, variational inference, Laplace approximations, and evidential deep learning—attempt to quantify epistemic uncertainty by representing the predictive distribution as a second-order distribution over model predictions or parameters. These techniques allocate uncertainty by computing

$$\underbrace{\mathbb{E}_{\theta\sim q}[\mathrm{Var}(y|x,\theta)]}_{\text{aleatoric estimate}} +\underbrace{\mathrm{Var}_{\theta\sim q}(\mathbb{E}[y|x,\theta])}_{\text{epistemic estimate}}$$

where $q$ is a second-order (posterior or ensemble) distribution (Jiménez et al., 29 May 2025).

However, such variants systematically omit the squared bias term $(\mathbb{E}[\hat y]-f(x))^2$, and hence cannot provide exact epistemic uncertainty. In regions with high model bias (e.g., outside the training manifold), these methods misattribute systematic error to aleatoric noise, leading to underestimation of epistemic uncertainty and overestimation of aleatoric uncertainty—a phenomenon documented in empirical analyses and simulation studies (Jiménez et al., 29 May 2025).

3. Taxonomy of Epistemic Uncertainty Sources

A rigorous taxonomy clarifies the nuanced sources of epistemic uncertainty (Jiménez et al., 29 May 2025):

• Model (hypothesis-space) uncertainty: $p(y|x)\notin\mathcal H$, irreducible except by extending the hypothesis class.
• Estimation uncertainty: Stochasticity from finite data and procedural randomness, subdivided into:
  • Data-driven: Fluctuations due to dataset sampling.
  • Procedural: Fluctuations due to algorithmic choices and stochastic optimization.
• Distributional uncertainty: Arises under dataset or covariate shift $p(x,y)\neq\tilde p(x,y)$; technically a distinct but related object.

Within this taxonomy, exact epistemic uncertainty aggregates all but the irreducible aleatoric component.

4. Simulation-Based Protocols for Calibration and Measurement

For accurate empirical quantification, simulation-based protocols involve generating multiple independent resamples of training sets ($\mathcal D_N^{(i)}$) and procedural seeds ($\gamma_j$), training models on all composites, and then computing the empirical reference distribution

$$q_{N,\gamma}(\theta|x) = \frac{1}{n_dn_\gamma}\sum_{i,j}\delta_{\hat\theta_{N,\gamma_j}^{(i)}(x)}(\theta)$$

to measure:

• Total epistemic variance (variance across all $\hat y$)
• Split via law of total variance into procedural and data-driven components
• Bias via mean prediction versus the known ground-truth $f(x)$ (Jiménez et al., 29 May 2025)

This controlled protocol exposes the precise bias–variance structure and corrects for systematic underestimation of uncertainty by prevailing approaches.

5. Criteria and Metrics for Exact Epistemic Uncertainty

Exact epistemic uncertainty is defined as the sum of the variance-driven and bias-driven components:

$$U_{\mathrm{epi}}^{\mathrm{exact}}(x) = \underbrace{\mathrm{Var}_{\mathcal D,\gamma}[\hat y|x]}_{\text{variance-driven}} + \underbrace{(\mathbb{E}_{\mathcal D,\gamma}[\hat y|x] - f(x))^2}_{\text{bias}^2}$$

The corrective principle is to measure or estimate both terms, either in closed-form (when feasible) or empirically via resampling, and always report the sum as the exact epistemic uncertainty. If the bias term is large, conventional estimates of aleatoric uncertainty are not meaningful: part of the error attributed to irreducible noise is actually systematic epistemic error (Jiménez et al., 29 May 2025).

A summary of the decomposition and their properties is provided below:

Component	Notation	Description
Aleatoric	$\sigma^2(x)$	Irreducible data noise
Data-driven epi.	$$\mathrm{Var}_{\mathcal D}\big[\mathbb E_\gamma[\hat y]\big]$$	Finite-sample variance
Procedural epi.	$$\mathbb E_{\mathcal D}\big[\mathrm{Var}_\gamma(\hat y)\big]$$	Stochastic procedure variance
Bias-driven epi.	$(\mathbb E_{\mathcal D,\gamma}[\hat y]-f(x))^2$	Model/hypothesis bias

This decomposition is exact: the total expected error equals the sum of these components (Jiménez et al., 29 May 2025).

6. Impossibility of Empirical Risk Minimization for Exact Epistemic Uncertainty

No strictly proper second-order scoring rule exists that can compel a learner to honestly represent epistemic uncertainty using empirical risk minimization alone (Bengs et al., 2023). The reason is structural: a single sample $y$ provides no information about the uncertainty over the first-order distribution $p^*(\cdot|x)$, so frequentist loss-minimization is ill-posed for epistemic uncertainty estimation. Only hypothesis-driven (Bayesian) formulations, which encode epistemic uncertainty through explicit model priors or ensemble diversity, are theoretically capable of supporting exact epistemic quantification.

7. Recommendations for Practice and Research Impact

• Always separate variance-driven and bias-driven epistemic uncertainty. Simply reporting the (ensemble) variance of predictions is insufficient unless the model is perfectly specified and bias is negligible.
• Estimate bias terms via simulation, reference models, or held-out data. When $f(x)$ is not available, domain knowledge or improved modeling of the hypothesis space may be necessary.
• Recalibrate aleatoric uncertainty only when epistemic uncertainty is small. Misallocation of systematic risk into aleatoric estimates leads to misleading risk assessment.
• These principles underpin robust UQ in high-stakes applications (e.g., scientific simulation, safety-critical systems), where underestimating epistemic uncertainty can have catastrophic implications (Jiménez et al., 29 May 2025).

The unifying insight is that exact epistemic uncertainty is not a matter of elaborate UQ machinery, but of rigorous bias–variance accounting, model class assessment, and principled estimation protocols. This perspective establishes a baseline for all subsequent empirical and theoretical developments in uncertainty quantification.