Uncertainty quantification

1. Foundations of Uncertainty Quantification

1.1 Definitions and goals of UQ

Uncertainty quantification (UQ) is the discipline of characterizing uncertainty in model inputs and assumptions, propagating that uncertainty through computations, and reporting resulting uncertainty in model outputs. The central goal is not only to provide point predictions, but also to attach statistically meaningful measures—such as probability distributions, confidence intervals, or error bars—that reflect how uncertainty in the world and in knowledge translates to uncertainty in inferences or forecasts. In practice, UQ aims to improve decision quality by making uncertainty explicit and by testing whether reported uncertainty is consistent with observed discrepancies.

1.2 Sources of uncertainty (input, model, numerical)

Uncertainty in quantitative modeling typically arises from multiple layers. Input uncertainty includes measurement error, noise in observations, variability in environmental conditions, and uncertainty in externally specified quantities. Model uncertainty covers imperfect representation of reality, including limited expressiveness of the model class and uncertain relationships among variables. Numerical uncertainty arises from computational procedures, such as discretization choices, iterative solver tolerances, and floating-point effects. Distinguishing these components is important because they often require different mitigation strategies and affect how uncertainty should be interpreted.

1.3 Probability, statistics, and uncertainty representations

UQ relies on probabilistic and statistical concepts to express uncertainty. In many applications, uncertainty is represented through distributions over unknown quantities, through predictive distributions over outputs, and through summary statistics like means, quantiles, and credible intervals. Frequentist notions (e.g., confidence intervals, coverage) and Bayesian notions (e.g., posterior distributions, credible intervals) provide different interpretations but are often used together in applied workflows. Regardless of framework, a key requirement is internal consistency: uncertainty statements should correspond to the procedures used to compute them and to how the model is expected to behave.

1.4 Aleatoric vs. epistemic uncertainty

A widely used distinction separates aleatoric uncertainty from epistemic uncertainty. Aleatoric uncertainty reflects irreducible randomness in the data-generating process (for example, measurement noise or intrinsic variability). Epistemic uncertainty reflects incomplete knowledge—stemming from limited data, uncertain parameters, or uncertainty about which model is correct. This separation helps determine whether collecting more data can reduce the uncertainty: epistemic uncertainty is generally reducible, while aleatoric uncertainty typically persists even with additional measurements.

1.5 Calibration and interpretation of uncertainty outputs

Calibration concerns whether predicted uncertainty matches observed outcomes. For instance, if a model reports 90% intervals, then across repeated experiments the true quantity should fall inside those intervals about 90% of the time. Good calibration improves trust in uncertainty estimates and supports downstream decision-making. Interpretation also depends on the underlying assumptions: the same numeric interval may correspond to different meanings under different statistical interpretations, so UQ reporting should clarify what the interval represents and how it was computed.

2. Modeling Uncertainty

2.1 Modeling measurement noise and stochastic inputs

To incorporate input uncertainty, measurements are often modeled as noisy observations of latent quantities. A common approach is to assume additive noise (e.g., observation = truth + error) with a specified noise distribution. For time series or spatial data, noise may be correlated, requiring covariance models. Stochastic inputs—such as uncertain boundary conditions or random forcing terms—are represented by probability distributions or random fields, enabling the model to account for variability that cannot be eliminated by better parameter estimation alone.

2.2 Parameter uncertainty and priors/posteriors

Many models contain parameters that cannot be known exactly, due to finite data or limited experimental characterization. In Bayesian UQ, prior distributions encode initial beliefs, and observed data update them to form posterior distributions. In frequentist settings, parameter uncertainty may be expressed via sampling distributions or asymptotic approximations around an estimated parameter. In either case, the uncertainty in parameters propagates to predictions, and the choice of parameterization and prior (or regularization) can influence the resulting uncertainty measures.

2.3 Model-form uncertainty and structural assumptions

Even with perfect parameter estimation, a model may be wrong in structure. Model-form uncertainty accounts for discrepancies between simplified assumptions and true behavior. Approaches include adding discrepancy terms, using alternative model classes, or incorporating flexible components (such as Gaussian processes or kernel-based residual models) to capture systematic deviations. Structural uncertainty is often harder to quantify because it blends together misspecification, omitted physics, and unmodeled dynamics, requiring careful diagnostics and validation.

2.4 Emulation and surrogate models for uncertain systems

When the underlying simulator is expensive, emulators or surrogate models approximate the input-output relationship. Typically trained on a set of simulator evaluations, these surrogates provide fast predictions and can produce uncertainty estimates themselves. Common emulator families include regression-based models, Gaussian process surrogates, and neural-network models with uncertainty estimates. Key concerns include surrogate bias, the reliability of emulator uncertainty, and how uncertainty should be combined with residual numerical or measurement uncertainties.

2.5 Handling missing data and incomplete information

Real datasets often include missing observations, censoring, or partially observed variables. UQ can address this by treating missingness mechanisms explicitly (when known) or by modeling latent variables. Strategies range from multiple imputation to Bayesian marginalization over unobserved quantities. When missingness is informative, assumptions about the missing data process can become a dominant contributor to epistemic uncertainty, so sensitivity checks are frequently necessary.

3. Uncertainty Propagation

3.1 Deterministic propagation concepts

Uncertainty propagation aims to map uncertainty in inputs to uncertainty in outputs. In the simplest setting, if the model is deterministic and inputs are random, the output becomes a random variable with a distribution induced by the input distributions. Deterministic propagation methods may include analytic transforms for simple systems or numerical evaluation over representative input samples. While conceptually straightforward, practical systems rarely yield closed-form results.

3.2 Monte Carlo methods (basic and advanced)

Monte Carlo (MC) propagation estimates output uncertainty by drawing many samples of uncertain inputs, running the model for each sample, and computing statistics of the resulting outputs. Basic MC converges slowly, but it is broadly applicable and easy to implement. Advanced variants include importance sampling (to focus on important regions), stratified sampling (to reduce variance), and adaptive schemes that concentrate samples where uncertainty impact is largest. For high computational cost models, MC is often paired with surrogates or reduced-sample approaches.

3.3 Variance reduction techniques

Variance reduction accelerates Monte Carlo convergence without changing the underlying expectation being estimated. Methods include common random numbers (reusing random seeds across comparisons), antithetic variates (using negatively correlated samples), control variates (subtracting known correlated quantities), and stratification. These techniques can significantly improve efficiency, especially when estimating tail probabilities or sensitivity indices.

3.4 Polynomial chaos expansions

Polynomial chaos expansions (PCE) represent the model output as a series expansion in orthogonal polynomials of the input random variables. When assumptions about input distributions and independence are satisfied, PCE can yield accurate approximations and allows analytic computation of moments and sometimes distributional information. Truncation order and polynomial selection affect accuracy, and dependence among inputs may require generalized or copula-based constructions.

3.5 Stochastic collocation methods

Stochastic collocation computes model outputs at strategically chosen points in the input space and then reconstructs output statistics using interpolation or quadrature rules. Compared with MC, collocation can be more efficient for problems with moderate effective dimension. Accuracy depends on smoothness of the model response and the selection of collocation points. For nonlinear or high-dimensional systems, collocation may still become expensive, motivating sparse or adaptive variants.

3.6 Linearization and first-order approximations

For models that are approximately smooth around nominal values, uncertainty propagation can use local linearization. By expanding outputs to first order in uncertain inputs, the method yields approximate mean and covariance of outputs. Linearization is computationally cheap and often provides insight into dominant uncertainty directions, but it can fail when uncertainty is large, nonlinear effects are strong, or outputs exhibit non-Gaussian behavior.

3.7 Bayesian propagation through probabilistic models

In Bayesian UQ, uncertainty propagation naturally arises from probabilistic modeling: the posterior over parameters and latent variables induces a predictive distribution over outputs. Bayesian methods integrate over unknown quantities, often using sampling or approximation methods. This yields principled uncertainty estimates that reflect both data noise and parameter/model uncertainty. Practical considerations include computational cost and the quality of approximations such as variational inference.

4. Sensitivity Analysis

4.1 Purpose and outputs of sensitivity analysis

Sensitivity analysis quantifies how variations in uncertain inputs influence variations in outputs. While UQ focuses on the uncertainty in predictions, sensitivity analysis explains *why* certain outputs are uncertain and which inputs contribute most. Outputs may include rankings of influential factors, measures of output variance attributable to each input, or qualitative assessments of robustness. This information supports prioritization of data collection and informs modeling decisions.

4.2 Local vs. global sensitivity methods

Local sensitivity examines behavior near a nominal point, typically via derivatives or small perturbations. It is useful for identifying immediate dependencies and for approximately linear regions. Global sensitivity evaluates the impact of inputs across their full uncertainty ranges, capturing nonlinearities and interactions. Global methods are more informative when uncertainties are large or when input-output relationships are highly nonlinear.

4.3 Variance-based methods (e.g., Sobol indices)

Variance-based sensitivity analysis decomposes output variance into contributions from individual inputs and combinations of inputs. Sobol indices are a common tool: first-order indices measure the effect of each input alone, while total-effect indices include interactions with other variables. These measures connect directly to uncertainty propagation and provide a clear variance attribution framework. Interpretation requires attention to assumptions about input distributions and independence.

4.4 Derivative-based and screening approaches

Derivative-based methods use gradients (or approximations to gradients) to estimate sensitivities, which is effective for differentiable models and moderate dimension. Screening approaches aim to quickly identify unimportant inputs using relatively few model evaluations. They often rely on heuristic or coarse statistics and then refine the analysis on a smaller subset of variables. Such strategies are valuable when full global sensitivity is computationally prohibitive.

4.5 Surrogate-assisted sensitivity analysis

When the forward model is costly, sensitivity analysis can be performed using surrogates. By training an emulator on sampled runs, one can compute sensitivity measures with far fewer calls to the original simulator. However, the sensitivity results depend on emulator fidelity; uncertainty in the surrogate itself may propagate into the sensitivity estimates. Good practice includes checking surrogate accuracy and verifying that sensitivity conclusions remain stable under reasonable modeling choices.

4.6 Interpreting sensitivity results for decision-making

Sensitivity results should guide action. Inputs with high influence on output uncertainty are candidates for improved measurement, better modeling, or refined experimental design. Conversely, inputs with negligible influence may be treated with simpler assumptions, enabling computational savings. Interpretation should consider not only statistical influence but also practical constraints, such as measurement cost, time, and feasibility of changing certain assumptions.

5. Inference and Bayesian UQ

5.1 Bayesian modeling workflow

A typical Bayesian UQ workflow begins by specifying a probabilistic model: priors for unknown quantities, a likelihood for observed data given those quantities, and a forward model linking parameters to predictions. Next, inference computes the posterior distribution of unknowns conditioned on data. Finally, posterior predictive distributions propagate uncertainty to outputs of interest, often accompanied by model checks to assess whether the probabilistic assumptions capture observed behavior.

5.2 Likelihood functions and posterior inference

The likelihood function encodes how likely the observed data are under different parameter values. Choices for likelihood reflect noise assumptions (e.g., Gaussian, Poisson, or heavy-tailed errors) and can strongly affect posterior uncertainty. Posterior inference combines likelihood with priors to yield a normalized posterior distribution. In complex models, exact normalization may be infeasible, motivating numerical sampling or approximation algorithms.

5.3 Markov chain Monte Carlo (MCMC) overview

MCMC generates samples from the posterior by constructing a Markov chain whose stationary distribution is the posterior. After burn-in, samples approximate expectations under the posterior. MCMC methods include random-walk variants, Hamiltonian methods, and Gibbs sampling for conditionally conjugate structures. Key practical issues include convergence diagnostics, effective sample size, autocorrelation, and the computational burden of repeatedly evaluating the model and likelihood.

5.4 Variational inference for uncertainty estimation

Variational inference approximates the posterior with a tractable family of distributions by optimizing an objective that minimizes divergence from the true posterior. It is often faster than MCMC, enabling application to larger datasets or more complex models. The trade-off is that variational approximations can underestimate uncertainty if the chosen family is too restrictive. Careful assessment via predictive checks is therefore essential.

5.5 Ensemble and particle-based approaches

Ensemble methods represent uncertainty using multiple models or particles that evolve in time as new data arrive. Particle filters are common in sequential settings, where latent states change dynamically. Ensemble techniques can also support parameter estimation and state prediction under uncertainty, especially for nonlinear systems. These methods face challenges such as sample degeneracy and the need for resampling or regularization.

5.6 Posterior predictive checks and model validation

Posterior predictive checks evaluate whether simulated data drawn from the posterior resemble the actually observed data. This provides a diagnostic of model adequacy beyond parameter estimates. Common comparisons include distributions of residuals, summary statistics, and calibration of prediction intervals. Validation is not purely a formality: it helps detect misspecified likelihoods, inadequate discrepancy modeling, or systematic errors that can lead to misleading uncertainty quantification.

6. Experimental Design and UQ

6.1 Designing experiments to reduce uncertainty

Experimental design aims to choose measurements that reduce uncertainty in quantities of interest. In UQ terms, the design affects which aspects of the model are better constrained and which uncertainty components dominate after observing new data. Good designs align measurements with parameters that influence outputs strongly or with regions where the model’s predictions are most sensitive. The approach often requires integrating the cost of experiments with the expected information gain.

6.2 Active learning for efficient data collection

Active learning selects new data points iteratively, using current uncertainty estimates to decide where additional observations are most valuable. The goal is to achieve desired accuracy with fewer simulations or experiments. Typical acquisition functions prioritize regions with high predictive uncertainty, high expected improvement, or high expected reduction in posterior variance. For surrogate-based workflows, active learning also helps keep the emulator accurate where it matters.

6.3 Optimal design criteria (overview)

Optimal design criteria provide objective functions for selecting experimental settings. Examples include maximizing expected information gain, minimizing expected posterior variance, or optimizing expected utility under specific loss functions. The choice of criterion depends on the task: whether the priority is parameter estimation, accurate prediction in a particular region, or uncertainty reduction for decision-making. In many cases, the criteria are approximated via sampling or surrogate models to remain computationally feasible.

6.4 Adaptive sampling strategies

Adaptive sampling modifies the sampling scheme based on intermediate results. Rather than fixing a design upfront, the method updates where to sample next using current uncertainty or error indicators. This is common in numerical UQ, where refinement depends on estimated local error, and in data-driven modeling, where new observations guide the next iteration. Adaptive strategies can improve efficiency but require robust stopping rules and safeguards against feedback-driven bias.

6.5 Balancing cost, accuracy, and uncertainty reduction

Real experiments and simulations have nontrivial cost, so UQ-related design must trade off statistical gains against expense. A design that heavily reduces uncertainty in one part of the output space may be inefficient if the decision task is sensitive to other regions. Multi-objective perspectives help balance overall accuracy, reliability, and practical constraints. In many workflows, approximate optimality is sufficient when supported by validation and sensitivity analysis.

7. Numerical Methods and Implementation

7.1 Discretization error and numerical uncertainty

Many forward models are based on continuous equations, which are approximated on discrete grids or meshes. The discretization error reflects the difference between the continuous solution and its numerical approximation. This error may be systematic and can correlate with input uncertainty. Proper UQ often treats discretization uncertainty as an additional component, especially when comparing results across resolutions or when tight error tolerances are required.

Mesh refinement improves accuracy by increasing resolution, but it also increases computational burden. Uncertainty-aware discretization uses error estimates to decide where to refine or how to choose resolution parameters so that numerical errors remain below a target level relative to other uncertainties. Adaptive mesh refinement can allocate resources to regions with high gradients or large sensitivity. The aim is to avoid confusing discretization artifacts with genuine uncertainty in inputs or model structure.

7.3 Solver tolerances and propagation of computational error

Iterative solvers and stopping tolerances introduce additional error. If solver tolerances are too loose, numerical results may vary due to computation rather than due to modeled randomness. UQ workflows may incorporate the effect of solver error by tightening tolerances, estimating residual error, or modeling computational uncertainty explicitly. Ensuring that computational errors are controlled is a prerequisite for meaningful uncertainty statements.

7.4 High-dimensional problems and computational scaling

UQ methods face the curse of dimensionality: as the number of uncertain inputs grows, the cost of sampling, quadrature, or polynomial expansions can increase rapidly. Practical solutions include dimension reduction (e.g., identifying influential directions), sparse approximations, exploiting structure like independence, or using surrogates to reduce evaluation time. Computational scaling is frequently the limiting factor in industrial applications.

7.5 Practical workflows and software ecosystems

Implementation typically involves managing sampling plans, surrogate training, numerical solvers, and statistical post-processing. Reproducible workflows include fixed random seeds, version-controlled dependencies, and documentation of model assumptions and uncertainty settings. Many toolchains support UQ via libraries for probability distributions, sampling, emulation, and sensitivity analysis. Practical success depends on integrating these components into consistent pipelines, along with validation steps that verify uncertainty quality.

8. Validation, Verification, and Reporting

8.1 Verification vs. validation terminology

Verification checks whether a computational implementation correctly solves the intended equations (e.g., grid convergence and solver correctness). Validation assesses whether the model, including its uncertainty assumptions, adequately represents the real process for the intended use. Confusing these terms can lead to misplaced confidence: a code may be verified while the model remains invalid, and the associated uncertainty estimates may be inaccurate for the application.

8.2 Comparing predicted uncertainty with observed errors

A core validation check compares predicted uncertainty with empirical discrepancies. For example, one can evaluate whether the width of prediction intervals matches the frequency of errors observed on held-out data. If intervals are systematically too narrow, uncertainty is underreported; if too wide, uncertainty may be overly conservative or based on miscalibrated assumptions. Such comparisons help diagnose whether uncertainty originates from model misspecification, incorrect noise assumptions, or inadequate propagation.

8.3 Coverage metrics and reliability diagrams overview

Coverage metrics quantify how often true values fall within predicted intervals or exceed thresholds. Reliability diagrams visualize calibration by comparing empirical frequencies to nominal levels across multiple quantiles or confidence thresholds. These tools are especially useful for assessing the probabilistic quality of predictive distributions. Coverage alone may not capture all aspects of distributional shape, so it is often complemented by additional diagnostic measures.

8.4 Proper scoring rules and uncertainty quality measures

Proper scoring rules evaluate predictive distributions using a single numerical score that rewards well-calibrated, sharp predictions. Common choices include log score and continuous ranked probability score (CRPS). These measures help compare competing uncertainty estimators and guide model improvement. Because scoring rules depend on the form of the forecast distribution, they encourage accurate representation of both central tendency and uncertainty spread.

8.5 Reporting standards and reproducibility

Transparent reporting enables others to understand and reproduce UQ results. Essential elements include describing uncertainty sources and assumptions, specifying distributions for uncertain inputs, documenting sampling methodology or approximation settings, and clarifying which uncertainty measures were computed. Reproducibility also requires reporting randomization practices, data preprocessing steps, and software versions. Good reporting supports independent verification and trustworthy use in decision contexts.

9. Advanced Topics

9.1 Multi-fidelity UQ and model bridging

Multi-fidelity UQ combines models of varying resolution or accuracy, such as coarse simulations with expensive high-fidelity simulations. Bridging methods align outputs across fidelities, often using calibration, correction models, or probabilistic linking. This can reduce computational cost while preserving uncertainty quality. The central challenge is to model the correlation between fidelities so that uncertainty estimates appropriately reflect discrepancies between approximations.

9.2 UQ for inverse problems

Inverse problems infer unknown quantities from observed outputs, making uncertainty central because the mapping from observations to parameters may be ill-posed or non-identifiable. UQ in this context often uses regularization, probabilistic priors, and posterior computations that reflect instability. Methods such as Bayesian inverse modeling can yield posterior distributions over parameters and propagates them to derived quantities. Validation is particularly important because posterior uncertainty can be sensitive to prior choices and likelihood assumptions.

9.3 Robust optimization under uncertainty

Robust optimization seeks decisions that perform well across plausible realizations of uncertain inputs rather than optimizing for a single predicted scenario. UQ provides the distributions or uncertainty sets needed to define robust objectives or constraints. Approaches include chance-constrained formulations and min-max criteria. The result is often a decision policy or design that trades nominal performance for guaranteed or probabilistically bounded risk.

9.4 UQ in time-dependent and dynamical systems

Time-dependent models introduce uncertainty that evolves through dynamics, often amplifying due to nonlinearities or chaotic behavior. UQ may require sequential methods, such as filtering and smoothing, or propagators that account for state uncertainty over time. Key issues include computational cost over long horizons and maintaining calibration as uncertainty accumulates. Diagnostics must consider both transient and asymptotic behavior of predictive intervals.

9.5 Causal considerations in uncertainty high-level overview

Causal considerations distinguish predictive uncertainty from causal inference about interventions. While UQ focuses on uncertainty in predictions of outcomes given a model, causal uncertainty addresses what would happen under changes to variables. In high-level terms, integrating causal structure with UQ requires careful identification assumptions and appropriate modeling of counterfactual outcomes. Without valid causal structure, uncertainty estimates may quantify prediction error rather than uncertainty about causal effects.

9.6 When UQ assumptions break: limitations and diagnostics

UQ depends on assumptions about distributions, independence, model adequacy, and numerical error behavior. Assumptions can fail when data are nonstationary, when noise is heteroscedastic and mis-modeled, when outputs exhibit strong non-Gaussian tails, or when the chosen model class cannot represent key patterns. Diagnostics include calibration checks, posterior predictive tests, residual analysis, sensitivity to prior choices, and investigations of out-of-distribution behavior. Identifying broken assumptions is crucial because it determines whether uncertainty estimates are trustworthy.