Many physical constants and scientific reference quantities are not measured as isolated values. They are calculated from several measured inputs through mathematical relationships. Each input may have uncertainty, and those uncertainties influence the final calculated result.
Uncertainty propagation provides a systematic way to evaluate this influence. The simplest approach uses a first-order approximation based on partial derivatives, also known as the law of propagation of uncertainty. It is appropriate when the measurement model is sufficiently smooth and the input uncertainties are small enough for linearization to be reasonable.
Suppose a calculated quantity y depends on independent input quantities x₁, x₂, ..., xₙ. The sensitivity coefficient for each input is the partial derivative of y with respect to that input. It describes how strongly a small change in that input affects the calculated result.
For independent inputs, the combined variance is the sum of the squared products of sensitivity coefficients and input standard uncertainties. For correlated inputs, covariance terms must be included. Ignoring important correlations can either overestimate or underestimate the final uncertainty.
For a simple product or quotient of independent quantities, relative uncertainties are often combined approximately in quadrature. For example, if y = ab, the relative variance is approximately the sum of the relative variances of a and b. This result follows from first-order propagation and assumes that the input uncertainties are sufficiently small and the inputs are independent.
For a power relationship y = xⁿ, the relative standard uncertainty is approximately |n| times the relative standard uncertainty of x. This explains why powers can amplify uncertainty.
When the model is strongly nonlinear, the input uncertainties are large, the output distribution is asymmetric or the inputs have complicated probability distributions, first-order propagation may not be sufficient. Monte Carlo methods can instead propagate probability distributions through the measurement model by repeatedly sampling input values and calculating corresponding outputs.
Uncertainty propagation is particularly important for derived physical constants. For example, the classical electron radius can be calculated from the elementary charge, vacuum permittivity, electron mass and speed of light. In modern SI, some of these inputs are exact by definition, while others have experimental uncertainty. The uncertainty of the derived result must reflect the actual uncertainty status and correlations of the inputs.
A complete uncertainty calculation therefore requires more than substituting numerical values into an equation. It requires a clearly defined measurement model, input uncertainty information, sensitivity coefficients, covariance data where relevant and an appropriate propagation method.
Uncertainty Propagation in Calculated Constants
Uncertainty propagation describes how uncertainties in measured input quantities influence a calculated result. It is essential when physical constants or reference quantities are derived from experimental measurements, mathematical relationships or other constants.
For y = f(x₁, x₂, ..., xₙ), the first-order law of propagation of uncertainty is:
uc²(y) = ΣᵢΣⱼ (∂f/∂xᵢ)(∂f/∂xⱼ)Cov(xᵢ,xⱼ)
For independent inputs:
uc²(y) = Σᵢ [ (∂f/∂xᵢ)u(xᵢ) ]²
For y = ab with independent inputs:
[u(y)/|y|]² ≈ [u(a)/|a|]² + [u(b)/|b|]²
For y = xⁿ:
u(y)/|y| ≈ |n|u(x)/|x|
Explanation
Variables
| Symbol | Meaning | Si Unit |
|---|---|---|
y | Calculated output quantity | Depends on the output |
xᵢ | Input quantity | Depends on the input |
u(xᵢ) | Standard uncertainty of input quantity xᵢ | Same as xᵢ |
uc(y) | Combined standard uncertainty of the output | Same as y |
∂f/∂xᵢ | Sensitivity coefficient describing the response of y to changes in xᵢ | Unit of y divided by unit of xᵢ |
Cov(xᵢ,xⱼ) | Covariance between input quantities | Product of the units of xᵢ and xⱼ |
Example or application
A worked example has not yet been added.
Definition
Uncertainty propagation is the mathematical process of determining how uncertainties and correlations in input quantities contribute to the uncertainty of a calculated output quantity.
Purpose
To determine the uncertainty of a calculated quantity from the uncertainties and statistical relationships of its input quantities.
Importance
Uncertainty propagation is indispensable when deriving physical constants, calculating experimental results, combining reference data and comparing theoretical predictions with measurements. It ensures that calculated values retain an honest representation of the reliability of their inputs.
Sensitivity Coefficients
The sensitivity coefficient cᵢ = ∂f/∂xᵢ measures the local response of the output to an input. A large absolute sensitivity coefficient means that a small change in that input can produce a comparatively large change in the output. However, the contribution to uncertainty also depends on the magnitude of the input uncertainty.
Propagation For Common Operations
| Operation | Model | Independent Input Result | Interpretation |
|---|---|---|---|
Addition or subtraction | y = a ± b | u²(y) = u²(a) + u²(b) | Absolute uncertainties combine in quadrature for independent inputs. |
Multiplication | y = ab | [u(y)/y]² ≈ [u(a)/a]² + [u(b)/b]² | Relative uncertainties combine in quadrature. |
Division | y = a/b | [u(y)/y]² ≈ [u(a)/a]² + [u(b)/b]² | Relative uncertainties combine in quadrature when the denominator is sufficiently far from zero. |
Power | y = aⁿ | u(y)/|y| ≈ |n|u(a)/|a| | The exponent scales the relative uncertainty. |
General function | y = f(a,b,...) | Use partial derivatives and the covariance matrix. | This method handles more complicated mathematical relationships within its approximation limits. |
Correlation Effects
Two input quantities are correlated when their variations are statistically related. Positive correlation can increase the uncertainty of a sum-like relationship, while negative correlation can reduce it. The actual effect depends on the mathematical structure of the measurement model. Correlations may arise because quantities share a calibration reference, are measured using the same instrument or are jointly estimated in a physical adjustment.
Exact And Measured Inputs
An input fixed exactly by an SI definition contributes no uncertainty through its assigned numerical value. An experimentally measured input contributes its evaluated uncertainty. In a derived expression, the uncertainty must be calculated using the actual status of each input, rather than assigning an uncertainty to every symbol automatically.
Monte Carlo Propagation
The Monte Carlo method evaluates a measurement model by repeatedly sampling input quantities from their assigned probability distributions and calculating the output for each sampled set. The resulting output distribution provides an estimate of the output value and uncertainty. This approach is useful when the model is nonlinear, the output distribution is asymmetric or analytical propagation is difficult. The number of trials, sampling assumptions and convergence should be documented.
Worked Example
Problem
A rectangular area is calculated from length L = 2.00 ± 0.01 m and width W = 3.00 ± 0.02 m. Assume the measurements are independent and the uncertainties are standard uncertainties. Calculate the area and its combined standard uncertainty.
Solution
Area: A = LW = 2.00 × 3.00 = 6.00 m².
For independent inputs, [u(A)/A]² = [u(L)/L]² + [u(W)/W]².
Relative uncertainty from length = 0.01/2.00 = 0.005.
Relative uncertainty from width = 0.02/3.00 ≈ 0.006667.
Combined relative uncertainty = √(0.005² + 0.006667²) ≈ 0.00833.
Absolute standard uncertainty = 6.00 × 0.00833 ≈ 0.050 m².
Result
A = 6.00 m² with combined standard uncertainty approximately 0.05 m², assuming independent input measurements and first-order propagation.
Worked Example With Power
Problem
A quantity is calculated using y = r², where r = 5.00 ± 0.02 cm. Find the approximate relative standard uncertainty in y.
Solution
For y = r², the relative uncertainty is approximately 2u(r)/r.
Substitute: 2 × 0.02/5.00 = 0.008.
Convert to percentage: 0.008 × 100 = 0.8%.
Result
Relative standard uncertainty in y ≈ 0.8%.
Practical Procedure
Write the complete mathematical model for the calculated quantity.
Identify every input quantity and its unit.
Obtain the standard uncertainty of each input.
Determine whether the inputs are independent or correlated.
Calculate the partial derivatives or sensitivity coefficients.
Evaluate the combined uncertainty using the covariance information.
Check whether first-order propagation is appropriate.
Use Monte Carlo propagation when the model or probability distributions require it.
Report the calculated result with its uncertainty and assumptions.
Common Misconceptions
Uncertainties should not always be added directly; the correct combination depends on the measurement model and correlations.
Relative uncertainty is useful for products and quotients, but absolute uncertainty is more natural for sums and differences.
A small uncertainty in an input does not guarantee a small output uncertainty if the model is highly sensitive to that input.
Correlated quantities cannot always be treated as independent.
Exact defining constants should not be assigned artificial measurement uncertainty.
First-order propagation is an approximation and is not universally valid.
Limitations And Considerations
First-order propagation may be inaccurate for strongly nonlinear models.
Input probability distributions may not be normal.
Correlations can be difficult to estimate but may be important.
Model uncertainty and unrecognized systematic effects may not be captured by input uncertainty calculations.
Monte Carlo results depend on appropriate input distributions and sufficient sampling.
Applications
Calculating derived physical constants.
Determining uncertainty in experimental equations.
Precision spectroscopy and metrology.
Engineering calculations involving measured dimensions and material properties.
Combining values from scientific databases.
Evaluating uncertainty in physical and chemical models.
Related Topics
SCI-REF-002-088
SCI-REF-002-089
SCI-REF-002-090
SCI-REF-002-091
SCI-REF-002-095
Sources and verification
Sources are shown as provided in the project data. Not every source has yet been independently verified. Please report errors or missing citations.
- JCGM 100 — Evaluation of Measurement Data: Guide to the Expression of Uncertainty in Measurement (GUM).
- JCGM 101 — Propagation of Distributions Using a Monte Carlo Method.
- Taylor, J. R. — An Introduction to Error Analysis: The Study of Uncertainties in Physical Measurements.