1 Principles of uncertainty
Uncertainty propagation translates uncertainty from inputs—such as instrument readings, estimated parameters, or modeled variables—into an uncertainty for an output computed by a measurement or calculation model. The output is typically a function of several quantities, and the goal is to quantify how variability in the inputs contributes to variability in the result.
In practice, uncertainty propagation serves both measurement and modeling tasks: it supports decision-making, enables comparison of computed outputs, and clarifies which inputs dominate the overall uncertainty. The approach chosen depends on model complexity, availability of distributional information, and whether relationships among inputs are independent or correlated.
1.1 Types of uncertainty (random vs. systematic)
Uncertainty is often categorized by source. Random uncertainty represents variability that fluctuates from one observation to another, such as measurement noise or short-term fluctuations in experimental conditions. Systematic uncertainty reflects effects that shift results in a consistent direction, including calibration offsets, model bias, drift, or imperfect physical assumptions.
In uncertainty propagation, both kinds can be treated mathematically, but they may require different modeling choices. Random effects are frequently represented using statistical distributions, whereas systematic effects are sometimes represented via bias parameters or as structured uncertainties that may be correlated across measurements.
1.2 Probability distributions and uncertainty representation
A key principle is that uncertainty can be represented probabilistically. Instead of describing an input by a single value plus an interval, uncertainty propagation often uses an assumed distribution for each input quantity (for example, Gaussian, uniform, or lognormal). The distribution determines how likely different input values are.
For derived outputs, the goal is not just to compute a single uncertainty number but to characterize the output distribution implied by the model and the input distributions. Even when the output is summarized by a standard deviation or interval, the distributional assumptions influence the result.
1.3 Correlation and dependence between inputs
Inputs are not always independent. Correlation arises when two measurements share sources of error or when a common calibration influences multiple parameters. Dependence can also be created by constraints, shared underlying variables, or transformations that couple quantities.
Neglecting correlation can substantially misstate the propagated uncertainty, especially in models where sensitivities to different inputs have opposing signs. Therefore, uncertainty propagation frameworks often incorporate covariance information or more general dependence structures.
1.4 Confidence levels and coverage concepts
Uncertainty summaries are commonly tied to confidence or coverage. A “confidence level” typically refers to a probability statement under repeated sampling (e.g., a 95% interval intended to contain the true value in 95% of repeated experiments). Coverage concepts are related but depend on the construction method and assumptions.
When distributions are non-Gaussian or nonlinear transformations are strong, standard confidence interpretations may fail if a method assumes linearity or normality. Selecting an appropriate interval type requires aligning the method (analytical or simulation-based) with the intended coverage claim.
2 Mathematical foundations
Uncertainty propagation is grounded in probability theory and approximation methods. At its core, the output quantity is treated as a function of uncertain inputs, and the uncertainty in the output is derived from how input variability maps through that function.
2.1 Uncertainty in functions of variables
Let the output be \(y = f(x_1, x_2, \dots, x_n)\), where each input \(x_i\) is uncertain. If the inputs have means and spreads (and possibly covariances), the task is to determine the resulting spread of \(y\). Mathematical approaches range from local linearization—relying on derivatives—to sampling-based computation of the full implied output distribution.
2.2 First-order (linear) propagation
First-order propagation approximates the function near the nominal input values by a linear model. This approach is accurate when uncertainties are small relative to the scale on which the function curvature changes.
2.2.1 Jacobians and partial derivatives
In a multivariable setting, the sensitivity of the output to each input is given by partial derivatives. The Jacobian matrix collects these derivatives: each element represents how a small change in an input affects the output.
Under linearization, the variance of the output can be expressed using these sensitivities along with the input variances and covariances. This formulation highlights that uncertainty propagation is essentially a weighted accumulation of input uncertainties, where weights depend on local slopes.
2.2.2 Covariance-matrix formulation
When inputs are correlated, variances alone are insufficient. The covariance matrix \(\Sigma\) encodes both marginal variances and pairwise covariances. In first-order propagation, the output variance can be written compactly using the Jacobian vector (or matrix for vector outputs) and \(\Sigma\).
This covariance-matrix view is useful because it generalizes cleanly: once \(\Sigma\) and the sensitivities are known, correlated uncertainty contributions are handled systematically, without ad hoc adjustments.
2.3 Higher-order effects
If the function \(f\) is strongly nonlinear over the uncertainty range, linearization may not capture the true behavior. Higher-order expansions incorporate curvature and other effects that shift both the mean and variance of the output.
2.3.1 Second-order (Taylor) approximations
A second-order Taylor expansion includes terms involving second derivatives. These terms can contribute to the mean shift (the expected output can differ from the function evaluated at nominal means) and can change the uncertainty size compared with the first-order result.
Second-order approximations are computationally feasible for modest-dimensional problems, but their accuracy depends on the smoothness of the model and the magnitude of uncertainty.
2.3.2 When linearization fails
Linear propagation may fail when uncertainties are large, the function has sharp features (such as thresholds or saturation), or the output distribution is truncated by constraints (e.g., physical non-negativity). It also struggles when the output depends on ratios or products of uncertain quantities, which can yield skewed or heavy-tailed distributions.
In such situations, simulation-based methods or alternative analytic treatments may provide more reliable uncertainty estimates.
2.4 Combining independent uncertainties
If inputs are independent and the output uncertainty is computed from a linear approximation, the combination often reduces to a sum of contributions.
2.4.1 Root-sum-of-squares (RSS)
For independent inputs, the variance contributions add. A common expression is the root-sum-of-squares (RSS), where each uncertainty component is weighted by its sensitivity (often a derivative magnitude). RSS yields a total standard uncertainty as the square root of the sum of squared contributions.
RSS is widely used due to its simplicity, but it implicitly depends on independence and local linearity (or on conditions where higher-order terms are negligible).
2.4.2 Treatment of mixed units and scaling
Uncertainty propagation naturally accommodates mixed units because sensitivities convert input units into the output unit. Proper derivative-based weighting ensures that components are combined consistently in the output’s scale.
Scaling matters because derivative values change with units and with the chosen parametrization. A reparameterization that reduces sensitivity magnitudes near nominal values can improve numerical stability, even if the physical uncertainty remains unchanged.
3 Analytical propagation workflows
Analytical workflows aim to compute propagated uncertainty with formulas, derivatives, and covariance algebra. They are efficient when the model is differentiable and when assumptions about distributions and linearity are reasonable.
3.1 Identify inputs and define the model
The first step is to list the uncertain inputs and define the functional relationship that maps inputs to the output. The model should reflect the computation used in the measurement or derived observable.
Equally important is establishing what “nominal values” represent—means, medians, or best estimates—and whether those values were produced by a fit that already accounts for measurement variability.
3.2 Specify assumptions about distributions
Analytical methods require distributional assumptions, even if only implicitly. For first-order variance propagation, a distribution-free approach can be used if uncertainty is characterized by moments (mean, variance, covariance). When higher-order moments or confidence intervals are needed, distributional form becomes important.
If inputs are known to be non-Gaussian (for example, strictly positive parameters), using a mismatched distribution can distort intervals or underestimate tail probabilities.
3.3 Compute sensitivities to inputs
Sensitivities are computed via partial derivatives or by constructing a Jacobian. This step can be carried out symbolically, through automatic differentiation, or via numerical differentiation.
Care must be taken with numerical derivatives: step size selection affects accuracy, and discontinuities or noisy evaluation can introduce spurious sensitivities that distort the final uncertainty.
3.4 Assemble the propagated uncertainty
With sensitivities and input covariance information, the output variance is computed using the appropriate propagation formula. For independent cases, RSS may suffice; for correlated inputs, the covariance-matrix formulation is used.
If higher-order terms are included, the workflow must also incorporate how second-order contributions affect the mean and variance, which may require additional moment calculations.
3.5 Report results (units, units consistency, and confidence)
A complete report states the nominal output, the propagated standard uncertainty (or equivalent), and any expanded uncertainty with an associated coverage or confidence concept. Units consistency should be checked: uncertainty components must correspond to the same output quantity and scale.
The report should clarify the assumed confidence level or coverage basis, especially when the method yields intervals based on normality or on asymptotic approximations.
4 Numerical propagation methods
Numerical propagation methods estimate output uncertainty by evaluating the model many times under sampled input realizations. These techniques are especially valuable when nonlinearities are strong or output distributions are difficult to characterize analytically.
4.1 Monte Carlo simulation
Monte Carlo simulation approximates the output distribution by repeated sampling from the input distributions, propagating each sample through the model.
4.1.1 Sampling strategies for input distributions
A sampling strategy specifies how to generate plausible input draws. For each input, draws are taken from its assumed distribution using the corresponding mean and standard uncertainty (or other parameterization).
If uncertainty is represented empirically, samples may be drawn from observed data distributions rather than from a parametric model. In all cases, the sample generation method should match the uncertainty representation used in the analysis.
4.1.2 Convergence and number of trials
Monte Carlo results converge as the number of trials increases, but the rate depends on the variance of the estimator and the tail behavior of the output. Practitioners typically monitor stability of summary statistics (such as output mean and standard deviation) as trial count increases.
Too few trials can yield noisy estimates of confidence intervals, particularly for quantiles in the tails. Adaptive strategies or variance reduction methods may be used, depending on computational budget.
4.1.3 Interpreting output distribution shape
Monte Carlo produces an empirical distribution for the output. The shape—symmetry, skewness, multimodality—can reveal issues with linearity assumptions or invalid extrapolations.
This distributional view supports robust summary choices. For example, quantile-based intervals may be more appropriate than symmetric intervals when the output is skewed.
4.2 Resampling approaches
Resampling methods generate new pseudo-datasets from existing data to reflect uncertainty in empirically estimated inputs.
4.2.1 Bootstrap for empirical uncertainties
Bootstrap involves resampling with replacement from observed data to approximate the sampling distribution of an estimator. In uncertainty propagation, the bootstrap can be used when input uncertainties come from finite datasets and when analytic uncertainty formulas are unavailable or unreliable.
The bootstrap’s validity depends on assumptions about how data represent the underlying process. It can be sensitive to small sample sizes or strong dependence structures.
4.2.2 Bayesian sampling as uncertainty propagation
Bayesian approaches treat uncertain quantities as random variables with posterior distributions. Propagating uncertainty then means sampling from the joint posterior and pushing those samples through the model to get an implied posterior for the output.
This framework naturally incorporates parameter uncertainty learned from data and provides a coherent basis for credible intervals. Computation often relies on Markov chain Monte Carlo or related algorithms.
4.3 Handling correlations in simulation
Correlations must be respected in numerical propagation to avoid misrepresenting joint variability.
4.3.1 Cholesky decomposition and covariance sampling
When dependence is reasonably described by a covariance matrix and a multivariate normal assumption (or when a transformation yields an approximately normal structure), one can sample correlated inputs using covariance-based methods such as Cholesky decomposition.
This procedure generates correlated draws that match the specified covariance structure, enabling straightforward Monte Carlo propagation.
4.3.2 Copulas and non-linear dependence overview
More general dependence can be modeled using copulas, which separate marginal distributions from the dependence structure. Copulas allow non-linear dependence patterns while retaining chosen marginals.
In uncertainty propagation, copulas can improve realism when correlation alone does not capture joint behavior, though they add modeling choices and require care in parameter estimation and validation.
5 Practical considerations and pitfalls
Uncertainty propagation is sensitive to modeling choices. Practical errors often come from incorrect assumptions, mis-specified distributions, unit mistakes, or overlooking discontinuities.
5.1 Nonlinearity and model mismatch
If the model used for propagation differs from the model used to compute nominal outputs, the propagated uncertainty may reflect the wrong sensitivities. Nonlinearity is another key risk: models with nonlinear transformations, large relative uncertainties, or discontinuities can invalidate linear approximations.
A mitigation strategy is to compare predicted uncertainty using linearization against simulation results for representative uncertainty magnitudes.
5.2 Truncation, bounds, and non-Gaussian outputs
Physical constraints can truncate input domains. For instance, certain quantities cannot be negative or exceed known limits. Truncated distributions can produce output distributions that are skewed or bounded.
Treating constrained inputs as unbounded normals can underestimate tail risks or distort confidence/coverage properties. Using distributions that respect bounds (or sampling that enforces them) improves fidelity.
5.3 Propagation with categorical or discrete inputs overview
When inputs are categorical (e.g., model selection) or discrete (e.g., counts), standard derivative-based methods are not directly applicable. Uncertainty propagation can still be performed by marginalizing over discrete possibilities, weighting outputs by their probabilities.
This scenario often results in mixture distributions, where the output can show multiple modes corresponding to different discrete input states.
5.4 Unit conversions and scaling errors
Unit conversion mistakes are a frequent source of error in uncertainty calculations. A consistent approach should include verifying that nominal values and uncertainties are expressed in the same unit system before propagation.
Scaling issues also arise when uncertainties are computed in different parameterizations (e.g., degrees versus radians), which changes derivatives and sensitivity coefficients if not handled correctly.
5.5 Sensitivity analysis vs. uncertainty propagation
Sensitivity analysis identifies how changes in inputs affect the output, but it does not necessarily quantify uncertainty in a statistically rigorous way. Uncertainty propagation incorporates both sensitivity and the magnitude and distribution of input variability.
While the two are related, confusing them can lead to overinterpretation. A sensitivity ranking alone does not provide an uncertainty interval unless uncertainty magnitudes are also incorporated.
6 Special topics
Some applications involve structured models, multi-stage computations, or parameter estimation pipelines that introduce additional dependence and complexity.
6.1 Propagation in multi-stage models
Multi-stage models compute intermediate quantities and then feed them into subsequent stages. Uncertainty propagation must account for how intermediate uncertainties correlate with later computations, especially when intermediates share inputs.
A common approach is to propagate uncertainties through the full computational graph rather than treating stages as independent. When stagewise propagation is done, careful handling of covariance between intermediate results is required.
6.2 Uncertainty propagation for derived observables
Derived observables—functions computed from fitted parameters or from multiple measurements—often involve transformations such as ratios, differences, exponentials, or nonlinear calibrations. These transformations can change distribution shapes, making linear approximations less reliable.
Uncertainty propagation for derived observables benefits from comparing analytical estimates with simulation, particularly when outputs are used to form decision thresholds or to compare to reference values.
6.3 Correlated calibration parameters
Calibration procedures frequently estimate parameters simultaneously, creating correlations among those parameters. Propagating uncertainty to calibrated results requires using the full covariance matrix from the calibration fit or an equivalent representation.
Ignoring parameter correlation can over- or under-estimate uncertainty, and it may also affect whether uncertainties in different calibrated channels appear spuriously independent.
6.4 Uncertainty budgets in experiments
An uncertainty budget organizes contributions from multiple sources—measurement noise, calibration uncertainty, environmental effects, and model assumptions—into a structured summary of total uncertainty.
Uncertainty propagation provides the mathematical backbone for budgets by mapping each contribution through the sensitivity structure of the measurement model. A budget should specify whether contributions are independent and how correlations are treated.
7 Validation and verification
Validation and verification ensure that propagated uncertainty estimates are consistent with data and robust against reasonable changes in assumptions.
7.1 Consistency checks with repeated measurements
Repeated measurements provide empirical evidence about output variability. Comparing observed output scatter with propagated uncertainty tests whether the uncertainty model captures the real variability.
Discrepancies can indicate missing uncertainty sources, incorrect distributional assumptions, or overlooked correlations.
7.2 Comparing analytical vs. Monte Carlo results
For differentiable models with moderate nonlinearities, analytical and Monte Carlo methods can agree closely. When results diverge, the difference often signals that linearization is insufficient or that distributional assumptions are inadequate.
Such comparisons guide method selection: if Monte Carlo yields more stable uncertainty intervals under varied conditions, it may be preferable.
7.3 Checking sensitivity to assumptions
Propagation relies on choices such as input distribution forms, parameter correlations, and confidence interval construction. Sensitivity-to-assumptions checks involve re-running propagation under alternative but plausible models to see whether conclusions change substantially.
If results are highly sensitive, uncertainty estimates should be presented with that sensitivity acknowledged, and more robust modeling or additional data may be warranted.
7.4 Stress-testing extreme parameter values
Stress testing evaluates model behavior near extremes of plausible inputs. This helps identify failure modes like truncation effects, nonlinear blow-up, or numerical instabilities in derivative calculations.
In simulation, stress tests can also detect situations where sampling does not adequately cover tail regions that dominate uncertainty in the output.
8 Reporting and standards of practice
Clear reporting supports interpretability and reuse. Standard practice aims to make uncertainty claims auditable by specifying the uncertainty model, assumptions, and computational method.
8.1 Choosing uncertainty levels (e.g., standard vs. expanded)
Results may be reported as standard uncertainty (a standard deviation equivalent) or as expanded uncertainty that scales a standard uncertainty by a coverage factor. The choice affects how directly the result maps to a confidence or coverage statement.
The reporting should state which uncertainty level is provided and how any scaling factor is justified (for example, based on normality assumptions or a specified coverage framework).
8.2 Communicating correlation information
Because correlations can dominate uncertainty in combined calculations, reporting should include covariance details when available. If correlations are unknown, the report should explain whether independence was assumed and what impact that assumption has.
For simulation-based methods, dependence handling should be described in terms of the method used to generate joint samples.
8.3 Documentation of assumptions and methods
A reproducible report includes the model equation, nominal input values, distribution assumptions, and the propagation approach. It should also list how sensitivities were computed (analytic derivatives, numerical differentiation, or automatic differentiation) or how sampling was performed.
Assumption documentation matters because uncertainty propagation is not uniquely determined by the output model; it depends on the uncertainty representation and methodological choices.
8.4 Reproducible uncertainty reporting templates
Templates help standardize reporting across projects and improve comparability. Useful templates include fields for: input values and uncertainties, correlation or covariance specifications, method type (analytical or simulation), number of Monte Carlo trials, interval construction details, and software or code versioning.
Reproducibility improves verification and reduces the chance that subtle modeling differences lead to conflicting uncertainty claims.