What Is Propagation of Uncertainty? (Lab Physics Specs)

Propagation of uncertainty estimates how uncertainty in measured inputs affects a calculated laboratory result. Start with a measurement model, assign each input a standard uncertainty, then combine contributions using the function’s sensitivities and any correlations. Report the result with units, uncertainty method, and coverage convention so readers can judge what the number means.

If you are unsure whether to add uncertainties, divide them, or use a different rule, start with the equation for the quantity you measured. The right method depends on that equation and on how the input measurements relate to one another. I use the steps below to make those choices clear and keep the final result honest.

Propagation of Uncertainty: Purpose, Measurement Model & Lab Overview

Propagation of uncertainty is the method used to estimate how input measurement uncertainties affect a calculated result. It does not remove uncertainty or guarantee that a result is correct. It provides a consistent way to carry uncertainty through a measurement model and report how reliable the calculated value is.

A measurement is rarely just a reading from one instrument. You may calculate density from mass and volume, or gravity from a measured pendulum length and period. In each case, uncertainty in the inputs affects uncertainty in the result.

The goal is to identify the quantity you want to measure, list its inputs, and show how those inputs determine the result. JCGM 100, known as the Guide to the Expression of Uncertainty in Measurement (GUM), and NIST Technical Note 1297 describe widely used approaches for doing this.

What is the measurement model?

A measurement model is the equation that connects the measured inputs to the quantity you want to report. It is written as (y=f(x_1,\ldots,x_n)), where (y) is the result and each (x_i) is an input value. List each input’s value, unit, and standard uncertainty before calculating.

For example, a pendulum experiment may use (g=4\pi^2L/T^2), where (L) is pendulum length, (T) is its period, and (g) is the calculated acceleration due to gravity. The model makes clear that period affects the result differently from length: period is squared and appears in the denominator.

Define the measurand, or quantity you intend to measure, precisely. “Pendulum length” might mean the distance from the pivot to the bob’s center, not the string length alone. A clear definition helps prevent an uncertain or inconsistent input from being mistaken for a calculation problem.

Key takeaway: Write the equation first, then identify each input and its unit.

Equations, Input Uncertainties & Calculation Requirements

The standard propagation equation combines input uncertainties according to how strongly the model responds to each input. It includes covariance terms for inputs that are related. For independent inputs, those terms are zero. This first-order method is usually suitable when uncertainties are small and the model behaves smoothly near the measured values.

For (y=f(x_1,\ldots,x_n)), the combined standard uncertainty is:

[ u_c^2(y)=\sum_i\left(\frac{\partial f}{\partial x_i}\right)^2u^2(x_i) +2\sum_{i<j}\frac{\partial f}{\partial x_i}\frac{\partial f}{\partial x_j}\operatorname{cov}(x_i,x_j). ]

Here, (u(x_i)) is the standard uncertainty of an input. The partial derivative tells you how much the result changes when that input changes. The covariance term accounts for inputs that vary together. Do not assume all inputs are independent without considering how they were measured.

For independent inputs, calculate each derivative, multiply its magnitude by that input’s standard uncertainty, square each contribution, add them, then take the square root. The result, (u_c(y)), is the combined standard uncertainty.

For products, quotients, and powers, a relative-uncertainty form can be convenient if the inputs are independent and the model has that form. For example, for (y=AB), the relative combined uncertainty is (\sqrt{[u(A)/A]^2+[u(B)/B]^2}). This shortcut does not apply to every equation. Use derivatives when the function or assumptions do not support it.

Which uncertainty belongs to each input?

Standard uncertainty expresses an uncertainty as a standard deviation. Type A uncertainty is evaluated by statistical analysis of data; Type B uncertainty is evaluated from other information, such as calibration records or instrument specifications. These labels describe how you evaluate uncertainty, not whether one source is more important.

For (n) independent repeated observations, calculate the sample mean (\bar{x}) and sample standard deviation (s). The standard uncertainty in the mean is (s/\sqrt n), not (s). This estimate describes the mean, while (s) describes the spread of individual readings.

Resolution, repeatability, calibration, and a manufacturer’s stated limit are different things. A display’s smallest step does not capture every source of uncertainty. If a digital instrument rounds to increments of size (q), and rounding error is reasonably treated as rectangular from (-q/2) to (+q/2), its standard uncertainty is (q/\sqrt{12}). This represents resolution alone, not the whole instrument uncertainty.

Convert stated limits into standard uncertainties only when you can justify a probability distribution. A rectangular distribution from (-a) to (+a) gives (u=a/\sqrt3). Calibration data may give a different distribution or a stated standard uncertainty. Record the basis for each conversion.

Key takeaway: Build an uncertainty budget from relevant sources, rather than treating one instrument specification as the complete answer.

How do you calculate a laboratory result?

For a pendulum, suppose (L=0.800\ \text{m}) with standard uncertainty (u(L)=0.001\ \text{m}), and (T=1.795\ \text{s}) with (u(T)=0.010\ \text{s}). Assume the inputs are independent standard uncertainties and use (g=4\pi^2L/T^2).

For this model, the relative combined standard uncertainty is:

[ \frac{u_c(g)}{g}= \sqrt{\left(\frac{u(L)}{L}\right)^2+ \left(2\frac{u(T)}{T}\right)^2}. ]

The factor of 2 comes from the squared period. Substitution gives (g\approx9.80\ \text{m/s}^2) and (u_c(g)\approx0.11\ \text{m/s}^2). The period uncertainty contributes more than the length uncertainty in this example. That identifies where better measurement may have the greatest effect.

This is an illustrative calculation, not a claim about a particular lab instrument. If the length and period share a source of error, such as a common timing or setup effect, assess whether they are correlated before setting covariance to zero.

Lab Execution, Data Quality & Common Student Errors

A dependable result comes from a documented sequence: define the measurand, characterize the inputs, propagate their uncertainties, then check the result and its assumptions. Keep raw data and instrument information with the calculation so another person can follow the same steps and see where each uncertainty came from.

In a teaching lab, I would check the measurement model before entering numbers into a spreadsheet. This simple habit catches errors such as using the string length instead of the pivot-to-bob distance, or entering a period for one swing when the data were collected over several swings.

What should you check before combining uncertainties?

An uncertainty budget is a list of input uncertainty sources and their contributions to the final result. It helps reveal missing sources, unsupported assumptions, and dominant contributions. The budget should note whether each value is Type A or Type B, how it was estimated, and whether correlations were considered.

Use this checklist:

• Define the measurand and write the measurement model.
• Record input values, units, repeated readings, and instrument details.
• Estimate Type A components from data and Type B components from documented information.
• Convert components to standard uncertainties using a justified method.
• Check whether any inputs share a source of variation.
• Apply the derivative and covariance formula, or use a validated numerical method.
• Confirm units and identify the largest contributions.

A common error is to treat resolution as the full uncertainty. Another is to add all absolute uncertainties directly. Neither rule works for every calculation. Apply the propagation method that fits the measurement model.

First-order propagation can be inadequate when uncertainty is large relative to the input, the model is strongly nonlinear over the input range, or the output distribution is not well described by a standard uncertainty. In those cases, consider a validated numerical method, such as Monte Carlo propagation, and describe the method used. Do not present a first-order result as exact.

Key takeaway: Document each source, check correlation and model behavior, and investigate the inputs that dominate the result.

If a lab reports (k\approx2), it may be used for roughly 95% coverage under suitable distribution and degrees-of-freedom assumptions. It is not a universal guarantee. The coverage depends on the uncertainty model and available information, so avoid calling every (k=2) interval a 95% confidence interval without support.

Round the uncertainty to a sensible number of significant digits, then round the measured result to the same decimal place. Avoid unexplained “±” values and excessive digits. A result such as (9.80214\pm0.11367) suggests more precision than the uncertainty supports.

When comparing with a reference value, state the comparison rule. For example, you can compare the difference between the measured and reference values with their combined uncertainty if the uncertainty definitions and independence assumptions support that method. A small difference alone does not prove agreement, and a reference value may have uncertainty of its own.

Key takeaway: Report what the uncertainty means, how it was obtained, and the rule used for any comparison.

Frequently asked questions

These short answers address common questions about uncertainty propagation in introductory physics labs. Apply the rules to the measurement model and uncertainty information for your own experiment; the same shortcut does not fit every calculation.

Is propagation of uncertainty the same as measurement error?
No. Propagation estimates how uncertainty in inputs affects a calculated result. Error is the difference between a measured value and a reference or true value, when known.

Do I add uncertainties or combine them in quadrature?
For independent inputs in the first-order method, combine derivative-weighted contributions in quadrature. Correlated inputs require covariance terms.

When do I use (s/\sqrt n)?
Use it to estimate the standard uncertainty of the mean from (n) independent repeated observations. It is not the standard deviation of individual readings.

Is instrument resolution the instrument’s total uncertainty?
No. Resolution is one possible contribution. Calibration, repeatability, and other relevant effects may also matter.

What does Type A uncertainty mean?
It is an uncertainty evaluated by statistical analysis of observations. Type B uncertainty is evaluated from other information, such as calibration or specifications.

What if two measured inputs are correlated?
Include their covariance in the propagation equation, or use a suitable numerical method that preserves their relationship. Do not set covariance to zero without reason.

How many digits should I report?
Round the uncertainty sensibly, then round the result to the same decimal place. Keep enough digits during intermediate calculations to avoid rounding effects.

(This article was written by one of our staff writers, Nathan Sterling. Visit our Meet the Team page.)