Uncertainty analysis is the systematic study of how incomplete knowledge, imperfect observations, natural variability, and modeling assumptions affect the results of an investigation. Its objects may be measured quantities, estimated parameters, model predictions, or decisions based on those predictions. Rather than providing only a best estimate, it describes a range or distribution of possible outcomes, the assumptions supporting that description, and the sources responsible for its spread.
The subject connects statistics, probability, numerical computation, measurement science, and decision-making. It is indispensable wherever a numerical result must be interpreted beyond its reported digits: a laboratory measurement, an engineering safety calculation, an economic forecast, or a projection of environmental change. The closely related term uncertainty quantification often emphasizes mathematical representation and computation, while uncertainty analysis can also include identifying sources, evaluating assumptions, and communicating limitations. Usage varies among disciplines.
A complete analysis distinguishes uncertainty in inputs from uncertainty in model structure and numerical calculation, preserves important dependencies, and expresses results in a form appropriate to their intended use. Its methods range from local error-propagation formulas to simulation, statistical inference, interval calculations, and decision analysis.
Scope and kinds of uncertainty
Uncertainty is not identical to error. An error is a difference between a reported or calculated value and a relevant reference value; uncertainty describes incomplete knowledge about a quantity or result. A correction can remove an estimated systematic effect without eliminating uncertainty about that correction. For example, adjusting a thermometer using a calibration certificate changes its indicated value, but the calibration reference, interpolation procedure, and subsequent instrument drift remain imperfectly known.
A widely used distinction separates aleatory uncertainty, associated with variability represented as random, from epistemic uncertainty, associated with limited knowledge. Variation in the strength of nominally identical manufactured components may be treated as aleatory, whereas uncertainty about the parameters of their strength distribution is epistemic. The distinction depends on the question and the level of description. A quantity that is unknown but fixed for one component may vary across a population of components. Additional measurements may reduce uncertainty about a particular component without eliminating variability in the population.
Other classifications identify where uncertainty enters an investigation. Measurement uncertainty concerns observations and their interpretation. Sampling uncertainty arises because only part of a population is observed. Parameter uncertainty concerns coefficients estimated from data. Structural uncertainty concerns the equations, mechanisms, or functional forms used to describe a system. Scenario uncertainty concerns alternative future conditions, such as different emissions pathways or demand trajectories. Numerical uncertainty arises when an exact mathematical calculation is approximated on a computer. These categories can overlap, so an inventory of sources is generally more useful than a rigid taxonomy.
Uncertainty analysis also differs from sensitivity analysis. The former asks how uncertain a result is, given specified information and assumptions; the latter asks how changes or uncertainty in inputs influence that result. A model can be highly sensitive to a parameter that is known accurately, making its contribution to uncertainty small. Conversely, a moderately influential parameter may dominate output uncertainty if it is poorly constrained. Sensitivity calculations therefore become uncertainty attribution only when combined with information about plausible input variation.
The object of analysis must be defined carefully. Uncertainty about the average lifetime of a device is different from uncertainty about the lifetime of the next device produced. Uncertainty about a historical temperature average differs from uncertainty about tomorrow’s local temperature. Such distinctions determine whether variability among individual outcomes belongs in the final result, whether additional observations can reduce the relevant uncertainty, and which probability statement is meaningful.
Historical development and measurement practice
Modern uncertainty analysis grew partly from astronomy, geodesy, and the mathematical treatment of observational errors. Adrien-Marie Legendre published the method of least squares in 1805, and Carl Friedrich Gauss developed influential probabilistic treatments of observational error in the early nineteenth century. These methods connected repeated observations, assumptions about error distributions, and the precision of estimated quantities. Their descendants include least-squares estimation, covariance-based error propagation, and statistical adjustment of measurement networks.
During the twentieth century, industrial measurement and international comparability created a need for terminology that worked across laboratories and technical fields. The Guide to the Expression of Uncertainty in Measurement, commonly called the GUM, was first published in 1993 under the auspices of seven international organizations. Its framework treats a measurement result as an estimate accompanied by information about uncertainty. The Joint Committee for Guides in Metrology subsequently maintained the guide and developed supplementary documents, including guidance on distribution propagation using Monte Carlo methods in 2008.
A central GUM distinction concerns methods of evaluating measurement uncertainty. Type A evaluation uses statistical analysis of observed values, such as repeated measurements. Type B evaluation uses other information, including calibration certificates, instrument specifications, reference data, and knowledge of physical behavior. These labels describe evaluation methods, not an absolute separation between random and systematic effects. A correction for a systematic effect can have an uncertainty evaluated by either method, depending on the available evidence.
A standard uncertainty is expressed as a standard deviation. If independent repeated observations share a common variance, the uncertainty associated with their mean is commonly estimated as , where is the sample standard deviation. This reduction does not apply to a common calibration offset shared by every observation. Averaging many readings can suppress independent fluctuations while leaving uncertainty from a shared reference essentially unchanged. Accordingly, a measurement model must describe the physical and procedural sources of uncertainty rather than treating every contribution as repeatable noise.
An expanded uncertainty is often written , where is the combined standard uncertainty and is a coverage factor. Under an approximately normal distribution and suitable conditions, gives an interval with coverage close to 95 percent. This is not a universal rule: asymmetry, limited degrees of freedom, and nonlinear measurement models can require other constructions. Measurement reports therefore distinguish the numerical uncertainty from the probabilistic interpretation attached to it, including the coverage probability or the method used to select the coverage factor.
Mathematical representations and dependence
A common starting point represents uncertain inputs by a vector and a quantity of interest by . Each component is a random variable, and the inputs are described by a joint probability distribution. The objective is then to determine properties of the induced distribution of : its center, spread, quantiles, tail probabilities, or dependence on other outputs. Specifying only a central estimate and an uncertainty percentage is generally insufficient to define this problem.
Distributional choices encode substantive assumptions. A uniform distribution can represent equal plausibility within stated bounds, but boundedness alone does not imply uniformity. Positive quantities such as durations or concentrations may require distributions that exclude negative values. A normal approximation may be appropriate for a sum of many small contributions, yet inappropriate for strongly skewed quantities or rare-event tails. Even the expected value and variance need not exist for every probability model; the Cauchy distribution is a standard example. Methods based on finite second moments therefore require more than a formally specified density.
Dependence is often as important as the marginal distributions. The covariance matrix records entries
It captures pairwise linear dependence but does not determine a general joint distribution. Outside special families, zero covariance does not imply statistical independence. Inputs may share a calibration source, environmental exposure, dataset, or conservation constraint. Sampling each independently can create combinations that are physically impossible or assign incorrect probabilities to simultaneous extremes. Copulas provide one way of separating marginal distributions from dependence structure, although their tail behavior remains a modeling assumption.
Not all uncertainty descriptions assign a single probability distribution. Interval analysis represents an input by bounds such as and seeks bounds on resulting outputs. Sets of probability distributions can represent uncertainty about distributional parameters or shapes. These approaches distinguish statements about what is possible from statements about what is probable. A guaranteed enclosure requires justified input bounds and mathematically valid propagation; an interval selected from limited observations does not automatically have that status.
Hierarchical representations can preserve the difference between variability and incomplete knowledge. For example,
separates outcome variability conditional on parameters from uncertainty about those parameters. In a Bayesian interpretation, the second distribution can represent current knowledge about ; in a population model, it may instead describe actual variation between groups. The same mathematical form can therefore support different interpretations. Documenting what each layer represents is essential to avoiding double counting or misleading claims about reducibility.
Analytical propagation of uncertainty
For a differentiable scalar model , local propagation begins with a first-order Taylor expansion about an input estimate :
If the input covariance matrix is , the resulting variance approximation is
The components of the gradient are local sensitivity coefficients. This formula is exact for affine models with finite input variances and otherwise depends on the adequacy of the linear approximation over the region containing appreciable probability.
For several outputs, , the corresponding expression is
where is the Jacobian matrix evaluated at the expansion point. The off-diagonal elements matter when outputs are subsequently combined. Two reported results may each have modest uncertainty but share a strong common error source. Ignoring their covariance can substantially misstate the uncertainty of their difference or ratio. Conversely, common-mode effects may cancel in a comparison even when the individual quantities remain uncertain.
For a product , first-order propagation gives
where and are nonzero input estimates. If their relative standard uncertainties are 2 percent and 3 percent and their covariance is zero, the relative output uncertainty is approximately percent. Under perfect positive correlation it becomes 5 percent; under perfect negative correlation, the first-order result becomes 1 percent. Thus adding uncertainties in quadrature is justified by the dependence assumptions, not simply by having multiple inputs.
Curvature affects both location and spread. With a Hessian matrix , a second-order approximation to the mean is
Consequently, evaluating a nonlinear function at mean inputs need not produce its mean output. For a zero-mean normal variable , the model illustrates a particularly clear failure of first-order propagation: the derivative at zero vanishes, but has positive mean and nonzero variance.
Analytical propagation is efficient when uncertainty is sufficiently localized and the model is smooth. It becomes unreliable near discontinuities, thresholds, singularities, or denominators that can approach zero. Derivatives may be obtained symbolically, by finite differences, or through automatic differentiation, but accurate derivatives do not establish that linearization is adequate. Comparing higher-order terms, changing the expansion point, or comparing with direct distribution propagation can reveal when the local approximation misses important features.
Simulation and distribution propagation
The Monte Carlo method propagates uncertainty by repeatedly drawing input vectors from their joint distribution and evaluating the model:
The resulting sample approximates the output distribution. Means, standard deviations, quantiles, and exceedance probabilities can be estimated without assuming that the output is normal. This makes simulation especially useful for nonlinear transformations, asymmetric distributions, mixtures, and models with multiple operating regimes.
Dependence must be preserved during sampling. For multivariate normal inputs, a factorization allows samples to be formed as , where the components of are independent standard normal variables. A Cholesky decomposition supplies such a factor for a positive-definite covariance matrix; singular cases require an appropriate alternative factorization. More general dependencies can be generated through conditional models or other joint-distribution constructions. Independently sampling fitted marginals is not equivalent to sampling a fitted joint distribution.
Monte Carlo introduces its own numerical uncertainty. For independent simulations with finite output variance , the standard error of the estimated mean is . Halving this error therefore requires approximately four times as many model evaluations. Quantile accuracy depends on the probability density near the target quantile, and tail probabilities can require far more samples than central summaries. If an event has probability , one million independent draws produce only one occurrence on average, which is insufficient for a precise relative estimate. Importance sampling changes the sampling distribution and applies weights to concentrate effort on relevant rare outcomes.
Other sampling designs improve efficiency under suitable conditions. Stratification and Latin hypercube sampling distribute evaluations across input ranges, while quasi-Monte Carlo methods use low-discrepancy point sets rather than independent random draws. Their benefits depend on smoothness, dimensional structure, and the quantity being estimated. They do not make a poor input model accurate. When inputs or parameters are sampled using Markov chain Monte Carlo, successive draws are dependent, so uncertainty estimates must account for effective rather than merely nominal sample size.
Simulation error must also be distinguished from uncertainty in the system being modeled. Increasing can make the estimated output distribution more precise while leaving that distribution broad. It does not reduce physical variability, resolve an unidentified parameter, or validate a structural assumption. A technically complete simulation analysis therefore reports both properties of the modeled output and the numerical accuracy with which those properties have been computed. For deterministic or bounded uncertainty descriptions, a Monte Carlo sample also cannot establish worst-case bounds: unobserved input combinations may produce more extreme outputs.
Statistical estimation and predictive uncertainty
When model parameters are estimated from observations, uncertainty analysis includes the consequences of using finite and imperfect data. In linear regression, for example, uncertainty about fitted coefficients affects the estimated mean response. Predicting a new observation introduces an additional contribution from outcome variation around that mean. The distinction between an interval for a mean response and a prediction interval is therefore substantive, not merely terminological: the latter is generally wider because it addresses a different unknown quantity.
A frequentist confidence interval is defined through the repeated-sampling behavior of a procedure. A nominal 95 percent procedure covers the fixed true parameter in 95 percent of hypothetical repetitions under its assumptions. Once a particular interval has been computed, the parameter is not assigned a 95 percent probability of lying inside it within that framework. Coverage can fail if sampling assumptions, error distributions, dependence structures, or model specifications are incorrect. Nominal levels describe a mathematical property under stated conditions, not an unconditional guarantee of scientific correctness.
Bayesian inference represents parameter uncertainty with a posterior distribution obtained by combining a likelihood and a prior. A credible interval contains a specified posterior probability, conditional on the model and available information. Prediction integrates over parameter uncertainty:
This posterior predictive distribution incorporates both uncertainty about and variation in future outcomes conditional on . It does not automatically include uncertainty from omitted mechanisms, incorrect likelihoods, or unconsidered models.
The bootstrap offers another approach by approximating the sampling distribution of an estimator through resampling. In the ordinary nonparametric bootstrap, observations are sampled with replacement from the observed dataset and the estimator is recomputed. The method can estimate standard errors and construct intervals without a closed-form sampling distribution. Its validity depends on the sampling structure and estimator. Time-dependent observations may require block methods, grouped observations may require cluster resampling, and extreme-value or boundary problems can defeat simple bootstrap approximations.
Data collection and identification constrain all these methods. A large dataset does not eliminate selection bias, unmeasured confounding, or ambiguity among observationally equivalent parameter combinations. In an inverse problem, several parameter settings may reproduce the observations while yielding different extrapolations. Regularization or prior information can stabilize estimation, but the resulting precision partly reflects those added assumptions. A narrow interval within one fitted model can coexist with substantial uncertainty across plausible models, populations, or future conditions.
Numerical models, emulators, and structural discrepancy
Computer-based uncertainty analysis often operates on a mathematical model that is itself only an approximation to a physical or social system. A useful conceptual decomposition is
where is the model, its parameters, a structural discrepancy, and an observation error. Model discrepancy represents systematic differences between the model and the target system that cannot be removed simply by selecting the correct parameter values. The decomposition is useful, but its components may be difficult to identify separately from limited observations.
Numerical approximation introduces another layer. A simulation of a partial differential equation uses a finite mesh, time step, convergence tolerance, and finite-precision arithmetic. The computed output , where denotes a resolution parameter, need not equal the exact solution of the mathematical model. Grid-refinement studies and error estimators assess this difference. Under an established convergence regime, a relation such as supports extrapolation. Discontinuities, unresolved scales, and unstable algorithms can invalidate the assumed rate.
Verification and validation address different questions. Verification concerns whether equations and algorithms are implemented and solved correctly; validation concerns how adequately the model represents the intended real-world system. Code verification may use analytical solutions or manufactured solutions. Solution verification examines numerical error for a particular calculation. Validation compares model behavior with relevant observations while accounting for uncertainty in those observations and in model inputs. Successful validation for one range of conditions does not establish accuracy in an unrelated regime.
High-cost simulations often require an emulator or surrogate model. A Gaussian process emulator predicts outputs between evaluated input settings and, under its probabilistic assumptions, supplies uncertainty about that interpolation. A polynomial chaos expansion approximates the response using polynomial functions orthogonal with respect to an input probability measure. Such expansions can make moments and sensitivity indices inexpensive to calculate. Their convergence depends on response regularity and dimensional structure; sharp transitions and many influential inputs can make simple representations ineffective.
Surrogate uncertainty must be distinguished from uncertainty in the original model and inputs. An emulator can reproduce a simulator accurately while both disagree with reality. Likewise, adjusting parameters to fit observations can conceal structural defects rather than resolve them. Calibration and discrepancy terms may compensate for each other, producing apparent fit without identifiable physical parameters. Independent evaluation data, informative experimental conditions, and explicit examination of extrapolation are therefore central to interpreting model-based uncertainty, especially when predictions concern conditions not represented in the calibration dataset.
Sensitivity analysis and attribution of uncertainty
Uncertainty attribution identifies which inputs, assumptions, or model components are responsible for variation in a specified output. Local methods use derivatives at a selected point, sometimes scaled by input and output magnitudes. They are inexpensive and useful for diagnosing nearby behavior, but they do not necessarily describe global effects over broad input ranges. A derivative can vanish at one point even when the corresponding input is important elsewhere, and a parameter can matter mainly through interactions with other parameters.
For independent inputs and a square-integrable response, variance-based methods use a functional decomposition,
whose components satisfy suitable zero-mean and orthogonality conditions. Output variance then separates into contributions from individual variables and their interactions. A first-order Sobol index is
It measures the fraction of variance explained by knowing alone under the specified input distribution.
A total-effect index includes interactions involving that input:
where denotes all inputs except . Consider , with independent zero-mean inputs having positive finite variances. Both first-order indices are zero, because conditioning on either input alone leaves a zero conditional mean. Yet both total-effect indices equal one: all output variance arises through the interaction. This example shows why one-at-a-time parameter changes can miss the mechanism generating uncertainty.
Dependence complicates attribution. When inputs are correlated, variation associated with one input can also carry information about another. The usual independent-input variance decomposition no longer provides the same unique interpretation. Conditional sensitivity measures and Shapley effects offer alternative allocations, but the chosen definition determines how shared contributions are assigned. A large sensitivity measure also does not establish causation; it describes relationships within the specified model and input distribution.
Variance is not always the relevant target. For structural reliability, the probability of crossing a failure threshold may matter more than total spread. For an asymmetric cost problem, sensitivity of an expected loss or an upper quantile may be more informative. Screening methods can identify potentially influential inputs before more expensive analysis, while derivative-based or distribution-based measures can expose different aspects of model behavior. Rankings are therefore conditional on the output, probability model, and metric selected. An input that is negligible for average behavior may dominate a rare but consequential tail event.
Decisions, robustness, and the value of information
Uncertainty analysis supports decisions by connecting possible states of the world with consequences of alternative actions. In decision theory, an action , uncertain state , and loss function define a formal problem. Under a specified probability distribution, a common criterion chooses
The optimal action depends on more than the most likely outcome. Asymmetric losses, thresholds, and low-probability severe consequences can make decisions sensitive to parts of the distribution that contribute little to its mean.
For example, consider an uncertain engineering load and a resistance that can be increased at a cost. Failure probability depends on the joint distribution of load and resistance, including their tails and dependence. A design selected from mean values alone can differ substantially from one selected using the probability that load exceeds resistance. Reliability constraints sometimes take the form
where denotes acceptable performance. The resulting probability is conditional on the assumed model; it is not independent evidence that all relevant failure mechanisms have been represented.
When a single probability distribution is poorly supported, robust optimization considers performance across a specified uncertainty set. A minimax-loss formulation is
Distributionally robust formulations instead optimize against a set of probability distributions. Such methods do not eliminate assumptions: conclusions depend on the bounds or distribution set chosen. Excessively broad sets can produce highly conservative actions, while narrow sets can exclude important possibilities. Robustness must therefore be interpreted relative to an explicit class of departures from a reference model.
The value of information measures whether reducing uncertainty could improve a decision. For loss minimization, the expected value of perfect information can be written
The first term is the best expected loss before learning the state; the second assumes the action can be selected after the state is known. Under suitable conditions, the difference is nonnegative because information can be ignored if it is unhelpful. It bounds the expected benefit of a feasible study when compared on the same basis.
Information value is not the same as variance reduction. An experiment may sharply reduce uncertainty about an influential parameter without changing the preferred action. Conversely, a modest reduction near a decision threshold may be valuable. Experimental design can therefore target parameter precision, prediction accuracy, or expected decision improvement, with different resulting designs. Scenario analysis serves a related but distinct role: it tests decisions under specified alternative conditions without necessarily assigning probabilities to those conditions.
Reporting, evaluation, and methodological limits
An uncertainty statement is interpretable only when its target and conditions are explicit. A report should distinguish an individual outcome from a population mean, a fixed unknown parameter from future variability, and an unconditional prediction from one conditional on a scenario. Units, reference conditions, transformations, and the time horizon can all change the meaning of the result. An uncertainty attached to an annual average cannot automatically be used for daily extremes, and a relative uncertainty is ambiguous when its denominator is near zero.
Numerical summaries should identify their construction. A standard deviation, a standard error, a central 95 percent interval, and an upper 95 percent bound answer different questions. For skewed or multimodal distributions, a mean and symmetric error bar can conceal important structure. Quantiles, density plots, or separate descriptions of distinct regimes may be more informative. An uncertainty budget records input estimates, their uncertainty evaluations, dependence assumptions, and propagation methods; it also makes clear which sources have been omitted or treated separately.
Predictive uncertainty can be evaluated against repeated outcomes. Calibration concerns agreement between stated probabilities and observed frequencies: events assigned a probability near 0.7 should occur about 70 percent of the time across an appropriate collection of comparable forecasts. Coverage assesses whether prediction intervals include observations at their advertised rate. Sharpness concerns concentration of the predictive distribution. Wide intervals can achieve high coverage while being uninformative, so performance assessment generally considers both concentration and calibration, often through scoring rules that reward honest probability distributions.
Evaluation itself has uncertainty. Rare events provide little information about tail calibration, dependent outcomes reduce the effective number of tests, and changing environments can make past performance an imperfect guide to future behavior. Cross-validation helps assess predictive behavior within a data-generating setting, but does not automatically test transfer to a different population or operating regime. Good average performance can also conceal poor calibration in particular subgroups or ranges of the input space.
Fundamental limitations remain when relevant mechanisms are unrecognized, data do not identify competing explanations, or extrapolation extends far beyond observations. Assigning a narrow probability distribution cannot compensate for omitted possibilities. Conversely, simply listing many hypothetical models does not establish that their range is comprehensive. Model ensembles can share structural assumptions and common datasets, making agreement less informative than the number of members suggests.
Reproducibility requires enough detail to reconstruct the analysis: input distributions and bounds, dependencies, data-processing choices, model versions, sampling algorithms, numerical tolerances, and convergence diagnostics. Separate accounting of input uncertainty, model limitations, and computational error helps prevent a precise numerical calculation from being mistaken for a precise description of reality. Where no defensible probability model exists, explicitly conditional results or bounds preserve the distinction between quantified uncertainty and assumptions that remain unresolved.