Measurement uncertainty is described fully and quantitatively by a probability distribution on the set of the possible values of the measurand.[1] This concept, central to metrology, reflects the incomplete knowledge about the quantity being measured and is typically represented by a probability distribution, often quantified as a standard deviation or an interval with a specified coverage probability.[2]The framework for evaluating and expressing measurement uncertainty is provided by the Guide to the Expression of Uncertainty in Measurement (GUM), first published in 1993 by the International Organization for Standardization (ISO) in collaboration with other international bodies, updated in 2008 by the Joint Committee for Guides in Metrology (JCGM), and further revised with ISO/IEC Guide 98-1:2024 (GUM-1, Edition 2).[1] The GUM defines key terms such as standard uncertainty (u), which is the uncertainty expressed as a standard deviation; combined standard uncertainty (u_c), obtained by combining contributions from multiple input quantities via the law of propagation of uncertainty; and expanded uncertainty (U), which multiplies the combined standard uncertainty by a coverage factor (k) to achieve a desired level of confidence, typically k=2 for approximately 95% coverage assuming a normal distribution.[3] These components ensure that measurement results are reported with a quantified indication of their reliability, incorporating all relevant sources of variability.[2]Uncertainties are evaluated using two complementary methods: Type A evaluation, based on statistical analysis of repeated observations (e.g., standard deviation of the mean), and Type B evaluation, relying on other information such as prior knowledge, manufacturer specifications, or assumed distributions when statistical data are insufficient.[4] The combined standard uncertainty is then calculated as the positive square root of the sum of the variances (or covariances where applicable) of these components, following the measurement model Y = f(X_1, X_2, ..., X_n) where the measurand Y depends on input quantities X_i.[4] This approach, widely adopted by industries and national metrology institutes like the National Institute of Standards and Technology (NIST), underscores the importance of measurement uncertainty in ensuring traceability, comparability, and quality in scientific and technical applications.[2]
Fundamentals
Definition and Scope
Measurement uncertainty is a parameter associated with the result of a measurement that characterizes the dispersion of the values that could reasonably be attributed to the measurand.ISO/IEC Guide 98-3:2008 The measurand is the quantity intended to be measured, and uncertainty quantifies the range within which the true value is likely to lie, reflecting the incompleteness of knowledge about the measurement process.ISO/IEC Guide 98-3:2008Unlike measurement error, which is defined as the difference between the measured value and the true value of the measurand, uncertainty does not attempt to correct for error but instead provides a statistical description of the possible errors.ISO/IEC Guide 98-3:2008 Measurement error includes both random and systematic components, but since the true value is generally unknowable, the focus in metrology shifts to uncertainty as a quantifiable expression of doubt.ISO/IEC Guide 98-3:2008Uncertainty thus encompasses contributions from both random effects (which vary unpredictably) and systematic effects (which are consistent but often unresolvable), providing a comprehensive assessment rather than a point estimate of deviation.ISO/IEC Guide 98-3:2008The scope of measurement uncertainty extends to all domains of measurement science, including physical metrology (e.g., length, mass, time), chemical metrology (e.g., concentration, pH), and engineering metrology (e.g., dimensional tolerances in manufacturing).ISO/IEC Guide 98-3:2008 For instance, measuring the length of an object with a ruler might yield a value of 100 mm with an uncertainty of ±0.1 mm, indicating that the true length lies within 99.9 mm to 100.1 mm at a specified confidence level.ISO/IEC Guide 98-3:2008 This framework ensures reliable interpretation across diverse applications, from laboratory calibrations to industrial quality control.In standard notation, the measurand is denoted as Y, which depends on input quantities X_i through a measurement model Y = f(X_1, X_2, \dots, X_n).ISO/IEC Guide 98-3:2008 The standard uncertainty u(Y) represents the standard deviation of Y, while the expanded uncertainty U = k \, u(Y) incorporates a coverage factor k (typically 2 for approximately 95% coverage) to define an interval containing a specified proportion of the distribution of possible values.ISO/IEC Guide 98-3:2008 This notation underpins uncertainty evaluation in decision-making processes, such as compliance testing against specifications.ISO/IEC Guide 98-3:2008
Historical Development
The foundations of measurement uncertainty trace back to early 19th-century advancements in handling random errors in astronomical and scientific observations. In 1805, Adrien-Marie Legendre introduced the method of least squares in his work Nouvelles méthodes pour la détermination des orbites des comètes, providing a technique to minimize the impact of observational errors by finding the best fit to data through squared residuals.[5] Four years later, in 1809, Carl Friedrich Gauss published Theoria Motus Corporum Coelestium, where he derived the normal (Gaussian) distribution as the probability density for random errors, assuming they arise from numerous small, equally likely causes, and justified the least squares method probabilistically.[6] These developments shifted error analysis from ad hoc corrections to a systematic, probabilistic framework, laying the groundwork for modern statistical treatment of measurement variability.[7]In the mid-20th century, metrology began distinguishing between "measurement error"—an unknown signed deviation—and "measurement uncertainty"—a quantified expression of doubt in the result—to better address systematic effects and incomplete knowledge. This distinction gained traction in the 1970s amid growing international efforts to standardize practices, prompted by a 1977 letter from Ernest Ambler of the NationalBureau of Standards to the BIPM highlighting inconsistencies in error reporting.[8] By the late 1970s, the International Committee for Weights and Measures (CIPM) issued Recommendation INC-1 (1980), advocating for uncertainty as a positive parameter encompassing both random and systematic contributions, rather than focusing solely on errors.[9] Influential statisticianRonald A. Fisher advanced this evolution in the early 20th century through his work on statistical inference, particularly in Statistical Methods for Research Workers (1925), where he emphasized maximum likelihood estimation and fiducial inference to quantify uncertainty in experimental measurements, influencing metrology's adoption of rigorous probabilistic tools.[10]The 1980s marked a pivotal transition from traditional "error analysis" to "uncertainty evaluation," driven by BIPM initiatives to promote non-negative, comprehensive quantification that avoids implying directionality in unknowns. This reform, originating from BIPM consultations in 1980–1981, culminated in the first edition of the International Vocabulary of Basic and General Terms in Metrology (VIM, 1984), which defined uncertainty as a parameter characterizing the dispersion of values reasonably attributable to the measurand.[11] The BIPM, alongside ISO Technical Committee 12 (Quantities and Units), played a central role in this standardization by coordinating international consensus through working groups.[12] Key milestones followed with the publication of the Guide to the Expression of Uncertainty in Measurement (GUM, ISO/IEC Guide 98-3) in 1993, developed by BIPM, IEC, IFCC, ILAC, ISO, IUPAC, IUPAP, and OIML to provide unified rules for uncertainty evaluation.[13] The GUM was revised in 2008 (JCGM 100:2008) with supplements like JCGM 101:2008 on Monte Carlo methods for distribution propagation, and further updates in the 2010s, such as JCGM 102:2011 on extending Monte Carlo methods to any number of output quantities.[14] Subsequent developments, including JCGM 105:2023 (GUM Part 6) on the realisation of measurement models and ISO/IEC Guide 98-1:2024 providing an updated introduction to uncertainty evaluation, continue to refine and support the GUM framework as the global standard for expressing measurement uncertainty.[15][1][16]
Sources and Classification
Type A Evaluation
Type A evaluation of measurement uncertainty refers to the method of estimating standard uncertainty through the statistical analysis of a series of repeated observations of the measurand. This approach captures the random variability inherent in the measurement process by treating the observations as a sample from an underlying probability distribution. It is particularly applicable when multiple independent measurements can be performed under repeatable conditions, providing a data-driven estimate of the dispersion in the results.[14]The procedure for Type A evaluation involves first obtaining nindependent observations, denoted as x_1, x_2, \dots, x_n, of the quantity being measured. The sample mean \bar{x} is calculated as\bar{x} = \frac{1}{n} \sum_{i=1}^n x_i.The experimental standard deviation s of the observations is then determined from the sample variance:s = \sqrt{ \frac{1}{n-1} \sum_{i=1}^n (x_i - \bar{x})^2 }.Finally, the standard uncertainty associated with the mean, u(\bar{x}) or simply u(x), is obtained by dividing the standard deviation by the square root of the number of observations:u(x) = \frac{s}{\sqrt{n}}.This yields an estimate of the standard deviation of the mean, reflecting the precision of the measurement result derived from the series.[14]The validity of Type A evaluation relies on several key assumptions. The observations must be independent, meaning that the value of one does not influence another, and the measurement process should be stationary, with constant mean and variance across the series. Additionally, the method typically assumes that the probability distribution of the measurement errors follows a Gaussian (normal) distribution, which justifies the use of the sample standard deviation as an unbiased estimator of the population standard deviation. Violations of these assumptions, such as dependence between observations or non-stationarity, can lead to biased uncertainty estimates.[14]To illustrate, consider repeated measurements of voltage using a digital voltmeter under controlled conditions. Suppose 10 independent readings are taken, yielding a sample mean \bar{V} = 10.0 V and an experimental standard deviation s = 0.2 V. The standard uncertainty is then u(V) = 0.2 / \sqrt{10} \approx 0.063 V, indicating the variability attributable to random effects in the measurement series. This example demonstrates how the sample variance directly informs the uncertainty, with the mean serving as the best estimate of the true voltage.[14]For constructing coverage intervals around the measurement result, the degrees of freedom \nu = n - 1 associated with the Type A standard uncertainty are used. This value enters into the Student's t-distribution, which provides a factor t for achieving a specified coverage probability (e.g., 95%), such that the interval is \bar{x} \pm t \cdot u(x). The finite \nu accounts for the additional uncertainty in estimating the standard deviation from a limited sample.[14]Type A evaluation has inherent limitations, as it requires multiple observations to compute a meaningful variance and is thus inapplicable to single measurements. It also focuses solely on random components and cannot address non-random effects, such as systematic biases, which are instead evaluated through Type B methods based on other information.[14]
Type B Evaluation
Type B evaluation of measurement uncertainty refers to the method of estimating standard uncertainty using means other than the statistical analysis of a series of repeated observations, often incorporating systematic effects or other components based on available knowledge and information.[14] This approach relies on scientific judgment, prior data, or documentation rather than experimental repetition, making it particularly suitable for situations where multiple measurements are impractical.[17]Common sources for Type B evaluations include calibration certificates providing uncertainty statements, manufacturer specifications for instruments, empirical experience from previous measurements, and assessments of physical limits such as instrument resolution or environmental influences like temperature variations.[14] For instance, a calibration report might specify an expanded uncertainty, which can be converted to standard uncertainty by dividing by the coverage factor (typically 2 for 95% confidence assuming a normal distribution).[17] Manufacturer data often provide bounds on performance, while experience might inform estimates of effects like drift or hysteresis not captured in repeated trials.[14]The procedure for Type B evaluation involves assuming a probability distribution for the input quantity based on the nature of the information and then deriving the standard uncertainty as the standard deviation of that distribution. For uniform (rectangular) distributions over bounds from a to b, the standard uncertainty is given byu = \frac{b - a}{\sqrt{12}},which assumes equal likelihood across the interval, often applied to resolution limits or conservative bounds.[14] For triangular distributions, suitable for gradual changes where extremes are less probable, the formula isu = \frac{b - a}{2 \sqrt{6}}.If prior data suggest a normal distribution, the standard uncertainty can be taken directly from the reported standard deviation or converted from a confidence interval (e.g., dividing a 95% interval by 1.96).[14] Asymmetric bounds can be handled by adapting the rectangular formula, such as u = \sqrt{(b_+ + b_-)^2 / 12}, where b_+ and b_- are the positive and negative deviations.[14]A representative example is the uncertainty from instrument resolution \delta, where the reading is equally likely to lie anywhere within \pm \delta/2, leading to a uniform distribution and u = \delta / \sqrt{12}.[17] Another is the contribution from temperature effects, where a coefficient c (e.g., thermal expansion) combined with a temperature variation \Delta T (full range) yields an uncertainty component u(T) = c \cdot \frac{\Delta T}{\sqrt{12}}, assuming uniform distribution over the range.[14]All Type B uncertainties are expressed as standard deviations to ensure consistency when combining with Type A evaluations in an overall uncertainty budget.[14] This method's advantages include its applicability to single measurements, where repeated observations are not feasible, and its ability to incorporate expert judgment and diverse data sources for a more complete assessment.[17]
Propagation and Combination
Uncertainty in Indirect Measurements
In indirect measurements, the quantity of interest, or measurand Y, is not observed directly but is instead calculated from several input quantities X_1, X_2, \dots, X_n through a functional relationship Y = f(X_1, X_2, \dots, X_n), where each input X_i carries its own standard uncertainty u(X_i).[14] This approach is essential in fields like metrology, where direct measurement of complex properties is impractical, and uncertainties from the inputs propagate to affect the reliability of Y.[14]Common examples include the determination of density \rho = m / V, where mass m and volume V are measured inputs, or speed v = s / t, derived from distance s and time t.[14] In the density case, uncertainties in the balance reading for m and the dimensional measurement for V (e.g., via calipers or pycnometry) directly influence the final \rho value. Similarly, for speed, errors in timing devices or distance markers contribute to u(v). These illustrate how indirect methods rely on accurate input evaluations to ensure the output reflects true variability.[14]For small uncertainties, the variation in Y, denoted \delta Y, can be approximated using the first-order Taylor expansion:\delta Y \approx \sum_{i=1}^n \left( \frac{\partial f}{\partial X_i} \right) \delta X_i,where \partial f / \partial X_i are partial derivatives evaluated at the best estimates of the inputs.[14] This linearization simplifies propagation by treating small deviations linearly, providing a foundational step before more detailed variance calculations.This approximation rests on key assumptions, including the independence of the input quantities (i.e., no significant correlations between X_i and X_j for i \neq j) and approximate linearity of f within the uncertainty ranges, which supports Gaussian-like propagation of uncertainties.[14] Violations, such as strong nonlinearities, may require higher-order terms or alternative methods like Monte Carlo simulation.[14]Within the Guide to the Expression of Uncertainty in Measurement (GUM) framework, developing the measurement model Y = f(X_1, \dots, X_n) precedes any uncertainty evaluation, ensuring all relevant inputs and influences are identified systematically.[14] This model forms the basis for constructing an uncertainty budget, which tabulates contributions from each source to transparently assess their impact on u_c(Y), the combined standard uncertainty.A representative uncertainty budget for density \rho = m / V (e.g., for a solid sample) might appear as follows, based on GUM principles:| Input Quantity | Best Estimate | Standard Uncertainty u(X_i) | Sensitivity Coefficient c_i = \partial f / \partial X_i | Contribution |c_i| \cdot u(X_i) |
|----------------|---------------|-----------------------------------|---------------------------------------------------------------|-----------------------------------------|
| Mass m | 100 g | 0.05 g | $1/V (≈ 0.02 cm^{-3}) | 0.001 g/cm³ |
| Volume V | 50 cm³ | 0.2 cm³ | -m/V^2 (≈ -0.04 g/cm⁶) | 0.008 g/cm³ |Here, contributions are computed using the linear approximation, with the total \delta \rho estimated from their sum in quadrature for independent inputs.[14] Such budgets highlight dominant sources, guiding improvements in measurement precision.[14]
Law of Propagation of Uncertainty
The law of propagation of uncertainty, also known as the propagation of variance, quantifies how uncertainties in input quantities contribute to the uncertainty in the output of a measurement model Y = f(X_1, X_2, \dots, X_N), where Y is the measurand and the X_i are the input estimates. This method assumes a functional relationship and uses partial derivatives to weight the contributions from each input's standard uncertainty u(x_i). It forms the basis for the standard approach in metrology for combining Type A and Type B uncertainties in indirect measurements.For uncorrelated input quantities, where covariances between them are zero, the combined standard uncertainty u_c(y) is calculated asu_c(y) = \sqrt{ \sum_{i=1}^N c_i^2 u^2(x_i) },with sensitivity coefficients c_i = \left. \frac{\partial f}{\partial x_i} \right|_{x_1, \dots, x_N} evaluated at the best estimates of the inputs. This expression arises from the propagation of variance under the assumption of independence.The derivation relies on a first-order multivariate Taylor series expansion of f around the input estimates:y \approx f(x_1, \dots, x_N) + \sum_{i=1}^N \frac{\partial f}{\partial x_i} (x_i - x_{i,\text{best}}),where higher-order terms are neglected. Taking the variance of both sides yields\text{Var}(Y) \approx \sum_{i=1}^N \left( \frac{\partial f}{\partial x_i} \right)^2 \text{Var}(X_i)for the linear or weakly nonlinear case, with the standard deviation u_c(y) = \sqrt{\text{Var}(Y)}. This approximation extends naturally to the multivariate setting by considering the Jacobian matrix of partial derivatives.When input quantities are correlated, the full law incorporates covariance terms:u_c^2(y) = \sum_{i=1}^N c_i^2 u^2(x_i) + 2 \sum_{i=1}^{N-1} \sum_{j=i+1}^N c_i c_j u(x_i, x_j),where u(x_i, x_j) denotes the covariance between X_i and X_j, often expressed via the correlation coefficient r(x_i, x_j) = u(x_i, x_j) / [u(x_i) u(x_j)]. In matrix notation, this is u_c^2(y) = \mathbf{c}^T \mathbf{U} \mathbf{c}, with \mathbf{c} as the column vector of sensitivity coefficients and \mathbf{U} as the covariance matrix of the inputs.To express the uncertainty with a specified coverage probability, the expanded uncertainty is U = k \, u_c(y), where the coverage factor k is selected from the Student's t-distribution for a normal probability distribution of Y. For large degrees of freedom, k \approx 2 corresponds to approximately 95% coverage probability.A representative example is the calculation of the area A = \pi r^2 of a circle from a measured radius r. The sensitivitycoefficient is c_r = \frac{\partial A}{\partial r} = 2 \pi r, so the combined standard uncertainty is u_c(A) = |2 \pi r| \, u(r), or in relative terms, u_c(A)/A \approx 2 \, u(r)/r. For instance, if r = 2.8461 m with u(r) = 0.005 m, then A \approx 25.447 m² and u_c(A) \approx 0.089 m².[18]The first-order law is valid when the relative uncertainties are small (u(x_i)/x_i \ll 1) and the measurement function is approximately linear over the range of input variations; for nonlinear functions or larger uncertainties, higher-order Taylor terms or alternative methods may be required to account for asymmetry or non-normal distributions.
Sensitivity analysis in measurement uncertainty involves the determination of sensitivity coefficients, which quantify how variations in input quantities affect the output measurand. The sensitivity coefficient c_i for an input quantity X_i in a measurement model Y = f(X_1, X_2, \dots, X_n) is defined as the partial derivative c_i = \left. \frac{\partial Y}{\partial X_i} \right|_{X^*}, evaluated at the best estimates X^* of the input quantities.[14] This coefficient is a key element in the law of propagation of uncertainty, where it scales the standard uncertainty of each input to contribute to the combined uncertainty of the measurand.[14] Recent supplements to the GUM framework, such as Part 6 on developing measurement models (JCGM 106:2020), provide additional guidance on evaluating these coefficients in complex models.[19]Analytical computation of sensitivity coefficients is preferred when the measurement model f is differentiable, allowing direct calculation of the partial derivatives. For simple functional forms, these derivatives often simplify to constants or expressions involving the input estimates. For instance, in the measurement of electrical power P = V \cdot I, where V is voltage and I is current, the sensitivity coefficients are c_V = I and c_I = V, evaluated at the nominal values of current and voltage, respectively.[14] Such analytical methods ensure precision and are recommended in the Guide to the Expression of Uncertainty in Measurement (GUM) for models where explicit differentiation is feasible.[14]When analytical differentiation is impractical due to complex or non-analytic models, numerical methods provide approximations to the sensitivity coefficients. A common approach is the finite difference method, where c_i \approx \frac{f(X^* + \Delta X_i) - f(X^* - \Delta X_i)}{2 \Delta X_i}, with \Delta X_i chosen as a small perturbation relative to the uncertainty in X_i.[14] This empirical estimation is explicitly supported in the GUM for cases where partial derivatives cannot be obtained analytically, ensuring applicability to a wide range of measurement scenarios.[14]The sensitivity coefficient c_i interprets the relative influence of each input on the measurand, indicating the change in Y per unit change in X_i. To assess contributions to overall uncertainty, the percentage contribution of an input is often computed as \left( \frac{c_i u(X_i)}{u_c(Y)} \right) \times 100\%, where u(X_i) is the standard uncertainty of the input and u_c(Y) is the combined standard uncertainty of the output; this provides an approximate measure of each input's impact in linear approximations.[14] Higher values of |c_i| relative to other coefficients highlight inputs that disproportionately affect the measurand's uncertainty.Uncertainty in the sensitivity coefficients themselves arises if the input estimates X^* have associated uncertainties, which can propagate to c_i through the partial derivative evaluation. However, this effect is generally neglected in standard uncertainty evaluations because the uncertainties in c_i are typically small compared to those in the inputs, as noted in specialized analyses of propagation rules.[20]A practical example illustrates the role of sensitivity coefficients in revealing relationships between inputs. For the resistance measurement R = V / I, the coefficients are c_V = 1 / I and c_I = -V / I^2, showing a direct proportionality to voltage and an inverse proportionality to current; the ratio c_V / c_I = -I / V underscores the opposing influences, with the negative sign indicating that uncertainty in current inversely affects resistance uncertainty.[14]Sensitivity analysis is crucial for identifying dominant uncertainty sources, enabling targeted efforts to reduce overall measurement uncertainty by focusing on inputs with the largest |c_i| u(X_i) contributions. This prioritization is integral to uncertainty budgeting and supports efficient improvement in measurement reliability across fields like metrology and engineering.[14]
Models for Multiple Outputs
In measurement models that produce multiple output quantities, the relationship is expressed as a vector-valued function \mathbf{Y} = f(\mathbf{X}), where \mathbf{Y} is the vector of output quantities and \mathbf{X} is the vector of input quantities. The uncertainty associated with \mathbf{Y} is characterized by a covariance matrix U_y, which provides a complete description of the joint uncertainties among the outputs.The propagation of uncertainty in this multivariate framework approximates the covariance matrix as U_y \approx J U_x J^T, where U_x is the covariance matrix of the inputs, and J is the Jacobianmatrix containing the partial derivatives of the outputs with respect to the inputs. The diagonal elements of U_y represent the variances (or standard uncertainties squared) of the individual output quantities u^2(Y_k), while the off-diagonal elements capture the covariances \text{cov}(Y_j, Y_k) between pairs of outputs, accounting for potential correlations arising from shared inputs or model dependencies.The sensitivity matrix J is defined by its elements J_{kj} = \frac{\partial Y_k}{\partial X_j}, which quantify how variations in each input affect each output. For instance, in coordinate metrology, the position coordinates (x, y, z) of a point may be derived from measured distances to reference points, resulting in a full covariance matrix that describes both the uncertainties in each coordinate and their interdependencies due to geometric correlations in the measurements.[21]These models find application in multidimensional measurements, such as spectroscopic analysis where multiple spectral intensities or concentrations are estimated simultaneously, or in image processing where pixel values or feature coordinates form correlated outputs; such approaches are detailed in the Guide to the Expression of Uncertainty in Measurement (GUM) Supplement 2 from 2011.[22] Computation of U_y can be performed analytically for linear or simple nonlinear models, but numerical methods, such as finite differences to approximate the Jacobian, are often employed for complex functions.
Alternative Approaches
Uncertainty as an Interval
In measurement uncertainty analysis, a non-probabilistic approach represents the possible values of a measurand Y as a bounded interval Y \in [Y_{\min}, Y_{\max}], derived from intervals of input quantities X_i \in [X_{i,\min}, X_{i,\max}]. This method, known as interval arithmetic, treats uncertainty as a range of potential values without assuming probability distributions, providing guaranteed enclosures for the output when inputs are bounded.[23][24]The natural interval extension computes the output interval by substituting input intervals into the functional form f and evaluating at the endpoints, such as for the sum y = x_1 + x_2, yielding [y_{\min}, y_{\max}] = [x_{1,\min} + x_{2,\min}, x_{1,\max} + x_{2,\max}]. For more complex functions, arithmetic operations on intervals follow rules like addition and multiplication of bounds, ensuring the result encloses all possible values from combinations within the input intervals. This extension maintains an inclusion property, where the computed interval always contains the true range of Y, though it may overestimate due to the dependency problem, which arises when the same variable appears multiple times, leading to loss of correlationinformation.[23][25]To address overestimation, Hansen's method applies to monotonic functions by selecting appropriate endpoint combinations based on the function's increasing or decreasing nature, yielding tighter bounds. Moore's interval analysis further refines this by incorporating rigorous techniques, such as mean-value extensions or subdivision, to compute verified enclosures that minimize width while preserving inclusion. These approaches ensure computational tractability for bounding solutions in nonlinear systems.[24][23]Interval methods find application in scenarios lacking probabilistic data, such as worst-case tolerance analysis in engineering, where component dimensions are specified by tolerance intervals to guarantee assembly functionality. For instance, in mechanical design, input tolerances on lengths propagate to output clearances via interval arithmetic, providing conservative bounds for reliability assessment.[26]Compared to the Guide to the Expression of Uncertainty in Measurement (GUM), which employs probabilistic variance propagation assuming known distributions, interval arithmetic offers a deterministic alternative that yields wider bounds as worst-case guarantees but can be narrower than probabilistic intervals when distributions are unknown or non-normal.[14][24] This contrasts with probabilistic propagation by avoiding statistical assumptions, focusing instead on algebraic enclosures.[23]Limitations include pessimism for correlated inputs or non-monotonic functions, where dependency ignorance inflates intervals, potentially requiring advanced forms like centered or Taylor models for mitigation.[25][24]
Monte Carlo Simulation Methods
Monte Carlo simulation methods offer a numerical technique for propagating probability distributions of input quantities through a measurement model to approximate the distribution of the output quantity. This approach addresses limitations of analytical methods by directly simulating the propagation without relying on approximations such as linearization or Gaussian assumptions, as detailed in Supplement 1 to the Guide to the Expression of Uncertainty in Measurement (GUM Supplement 1, JCGM 101:2008).[27] It is particularly suited for complex measurement models where exact analytical solutions are infeasible.[28]The standard procedure begins by specifying the joint probability density function p(\mathbf{x}) for the input vector \mathbf{X} = (X_1, \dots, X_N), which may include marginal distributions and correlations. A large number M of independent random samples \mathbf{x}^{(m)}, for m = 1, \dots, M, are then drawn from p(\mathbf{x}). For each sample, the output value y^{(m)} = f(\mathbf{x}^{(m)}) is computed, where f represents the measurement model. The resulting set \{ y^{(m)} \} approximates the distribution of the output Y, with its probability density function estimated via methods such as histograms or kernel density estimation.[27][29]From the simulated outputs, the standard uncertainty u(Y) is obtained as the sample standard deviation of the y^{(m)} values. Coverage intervals, such as the shortest interval containing a specified probability (e.g., 95%), are determined using order statistics on the sorted samples; for instance, the central 68% interval approximates one standard uncertainty.[27] This empirical approach ensures reliable characterization even for non-standard cases.These methods excel in handling non-linear models and non-Gaussian input distributions, capturing features like skewness and multimodality that analytical propagation under the law of propagation of uncertainty cannot address.[28] For example, in propagating uncertainty through the Modification of Diet in Renal Disease (MDRD) equation for estimated glomerular filtration rate, \text{eGFR} = 175 \times (\text{SCr})^{-1.154} \times (\text{age})^{-0.203} \times (0.742 \text{ if female}) \times (1.212 \text{ if African American}), where inputs like serum creatinine follow normal distributions, Monte Carlo simulation with 10,000 trials yields an output mean of 41.5 mL/min/1.73 m² and standard uncertainty of 1.70 mL/min/1.73 m², revealing a slight positive skew and asymmetric coverage interval due to the reciprocal and power terms.[29]Convergence of the simulation requires sufficient trials, typically M \geq 10^5 to $10^6, to achieve stable estimates of the output mean, standard uncertainty, and coverage probabilities within acceptable precision (e.g., 0.1% relative error). Validation can involve comparing results to analytical propagation for linear Gaussian cases, confirming agreement within simulation variability.[28][27]Key advantages include independence from specific model assumptions, enabling application to arbitrary functional forms, and straightforward incorporation of input correlations through joint sampling from the multivariate distribution.[28] This makes the method versatile for metrological applications ranging from simple calibrations to complex systems.[29]Practical implementations are supported in software such as MATLAB's Statistics and Machine Learning Toolbox and Python's NumPy and SciPy libraries, which provide functions for random sampling from distributions like normal, lognormal, and uniform, along with tools for density estimation and order statistics, ensuring compliance with JCGM 101:2008 guidelines.[28]