Uncertainty Propagation

Introduction

Chemists often calculate a final quantity from measured masses, volumes, concentrations and signals. The arithmetic gives a central value, but the uncertainty of each input also travels through the calculation. A calculated concentration written to six decimal places can still be poorly known if the sample mass, calibration or recovery is uncertain. Propagation connects the measurement model to an honest reported range.

Core explanation

Start with a measurement equation, such as c = n/V for molar concentration. Identify each input and its standard uncertainty, including calibration and preparation contributions where relevant. For independent inputs with small relative uncertainties, a sum or difference y = a ± b has u(y) = sqrt[u(a)^2 + u(b)^2] . For a product or quotient y = ab or a/b , the approximate relative standard uncertainty is the square root of the sum of squared relative input uncertainties. Constants considered exact do not add uncertainty; measured conversion factors do.

These familiar rules are special cases of the sensitivity-coefficient method. If y = f(x1, x2, ...) , then for small changes each input contributes approximately (∂f/∂xi)u(xi) in output units. Square and sum the contributions when inputs are independent. If two inputs are correlated, covariance terms must be included. The NIST guidance on combined standard uncertainty explicitly includes covariance when appropriate. The independence assumption should not be hidden behind a memorized formula.

Correlation matters in chemistry. If two concentrations are calculated from the same calibration slope, their errors may move together. Subtracting them can cancel some common calibration error, while adding them may reinforce it. Conversely, two aliquots measured with the same biased pipette share a systematic component that averaging does not remove. Treating shared errors as independent can overstate or understate final uncertainty depending on the calculation.

Uncertainty also has different sources. Type A components are evaluated from repeated observations; Type B components may come from calibration certificates, instrument specifications or prior knowledge. The labels describe evaluation method, not whether an error is random or systematic in ordinary language. Correct identified biases where possible, then include uncertainty in the correction. An uncertainty budget should name contributions such as balance calibration, volumetric glassware, blank variation, fit coefficients and sample heterogeneity.

For large relative uncertainties or strongly nonlinear operations, first-order formulas may fail. A concentration near zero, a ratio with an uncertain denominator, or an exponential rate constant can have asymmetric output uncertainty. Propagating simulated input distributions through the full equation can reveal that shape. NIST's overview of measurement uncertainty notes both linearized and distribution-propagation approaches. Whatever method is used, state assumptions and coverage: a standard uncertainty is not automatically a 95% interval.

Step-by-step reasoning

Write the equation and units. List each measured input, its estimate, standard uncertainty and any shared source of error. Decide whether small-error linear propagation is adequate. Calculate each input's sensitivity contribution and combine independent terms in quadrature, adding covariance if needed. Check the result's units and whether one component dominates. Report the central value and uncertainty with a stated interpretation, and revisit the measurement plan if the dominant uncertainty is too large for the decision.

Visual explanation

Draw a flow chart with measured mass and volume entering a concentration box. Each arrow carries a value and uncertainty. The output box shows both c and u(c) . A bar chart beneath it displays the squared contribution from mass, volume and calibration. The tallest bar tells the researcher which measurement improvement would most reduce final uncertainty; polishing a tiny contribution yields little benefit.

Real-world analogy

Estimating travel speed from distance divided by time requires uncertainty in both the map distance and the clock. A precise stopwatch cannot compensate for a very uncertain route length. If the same map scale is used for two routes, their errors can be correlated. Chemical calculations likewise inherit the limitations and shared errors of their inputs.

Real-world example

A laboratory reports a dissolved-metal mass by multiplying measured concentration by sample volume. The instrument gives a low concentration uncertainty, but the sample volume was estimated from a roughly filled container. Propagation shows the volume dominates uncertainty in total mass. Repeating the instrument reading will barely improve the result; using a calibrated volume or weighing the sample will. This is why an uncertainty budget can guide experimental effort.

Why?

Why are independent uncertainties combined by a square root of squares rather than simply added? Positive and negative random deviations can partly offset, and variances add for independent quantities. Adding absolute uncertainty bounds can be a conservative worst-case calculation, but it is a different interpretation. The root-sum-square rule requires standard uncertainties and suitable independence or covariance treatment.

Common misconception

“The answer has the precision of the calculator display.” Digits do not remove input uncertainty. “Percentage uncertainties always add” is a rough worst-case rule, not the usual combined standard uncertainty for independent small errors. “Systematic errors are excluded from uncertainty” is false: after correction, uncertainty in calibration and corrections remains. Do not state a 95% confidence claim from ±1u without justification.

Worked example

A sample contains n = 0.1000 ± 0.0010 mol and has volume V = 0.500 ± 0.005 L , where the plus-minus values are independent standard uncertainties. Concentration is c = n/V = 0.200 mol/L . Each input has 1% relative uncertainty. Thus u(c)/c ≈ sqrt[(0.0010/0.1000)^2 + (0.005/0.500)^2] = sqrt(0.01²+0.01²) ≈ 0.0141 . The standard uncertainty is 0.200 × 0.0141 ≈ 0.0028 mol/L , so report about 0.200 ± 0.003 mol/L as a standard-uncertainty result. If mole amount and volume share a calibration factor, revisit the independence assumption.

Quick check

1. Why can repeated measurements fail to reduce uncertainty from a shared miscalibrated pipette? Answer: All repeats inherit the same calibration component. Averaging reduces independent random variation, but the shared systematic component remains until independently checked or corrected.

Exam focus

Write the measurement equation first. Use absolute uncertainties for independent sums and relative uncertainties for products and quotients under small-error assumptions. Distinguish standard from expanded uncertainty and state the coverage factor if used. Mention covariance when inputs share a calibration or preparation step. Round the final result consistently with the uncertainty, not with all calculator digits.

Advanced insight

Uncertainty propagation can guide optimal experimental design. Sensitivity coefficients identify which input most influences the output; its uncertainty times sensitivity gives an approximate contribution. If one component dominates, spending effort on smaller ones barely improves the final result. For nonlinear models, Monte Carlo propagation can preserve asymmetric distributions, boundaries such as nonnegative concentration, and correlations that simple derivative formulas may obscure.

Summary

Derived chemical quantities inherit uncertainty from all relevant measured inputs. A measurement equation and uncertainty budget make those contributions explicit. Root-sum-square rules apply to independent small standard uncertainties, while shared errors require covariance and nonlinear cases may require distribution propagation. Report the resulting value with units, assumptions and a clear uncertainty interpretation.

Practice questions

1. Two independent masses are 10.0 ± 0.2 g and 4.0 ± 0.1 g , with standard uncertainties. Find uncertainty of their difference. Answer: The difference is 6.0 g and standard uncertainty is sqrt(0.2²+0.1²) ≈ 0.22 g , so approximately 6.0 ± 0.2 g at an appropriate reporting precision.

2. A product depends on two independent measurements each with 2% relative standard uncertainty. Approximate the product's relative uncertainty. Answer: sqrt(0.02²+0.02²) ≈ 0.028 , or about 2.8%, under the small-error approximation.

3. Why might a ratio of two values from the same calibration require a covariance term? Answer: Both values may shift together when the shared calibration slope changes. Treating them as independent ignores possible cancellation or reinforcement of that common error in the ratio.