Short Answer
Most engineering results are not measured directly. A pipe’s cross-sectional area comes from a measured diameter, a heat duty comes from measured temperatures and a flow rate, and a density comes from separate mass and volume readings. Every input carries uncertainty, and combining the numbers does not make that uncertainty disappear — it travels into the result. Error propagation, also called propagation of uncertainty, is the set of standard formulas for estimating how the uncertainty in measured inputs becomes uncertainty in a calculated quantity.
Reporting those uncertainty bounds matters: a result whose error is too large can be useless for its intended purpose, so understanding how error propagates or compounds in computations is a core part of experimental engineering work. This article explains how propagation works, gives the working rules for common operations, walks through worked examples, and closes with a practical procedure, the assumptions behind the formulas, and the mistakes that most often distort results.
What Error Propagation Is
A measurement is reported as a best estimate plus an uncertainty interval, written in a form such as 4.0 ± 0.2 cm. In engineering data analysis, measurement outcomes are modeled as random variables — quantities whose repeated values show variation — and any calculation built on them becomes a function of those random inputs. Because engineers use measurement values to perform calculations, the uncertainty in those measurements must be carried through to the result; this step is a standard topic in engineering data-analysis courses.
Two forms of uncertainty appear throughout:
- Absolute uncertainty has the same units as the measurement itself, for example ±0.2 cm.
- Relative (fractional) uncertainty is the uncertainty divided by the value — 0.2 ÷ 4.0 = 5% in that example — and is dimensionless. It is the natural quantity for products, quotients, and percentage statements.
The rules below are stated in terms of these two quantities, and converting between them is a routine step in any uncertainty calculation.
Bias and Precision: Two Kinds of Uncertainty
Not all uncertainty behaves the same way. Precision error — the random scatter seen when a measurement is repeated — can be quantified with statistical methods applied to repeat-sampled data. Bias error, a systematic offset shared by every reading of an instrument, cannot be uncovered using statistical methods: repeating the same measurement with the same device shows scatter but never reveals a constant offset. Bias is instead estimated from experience and knowledge of the apparatus, or by comparison with a more accurate instrument — a direct check that is usually impractical because of its cost and time.
This distinction matters for propagation. The formulas in this article move precision-type, statistical uncertainty through a calculation; the same mathematics does not expose bias. Computing a cylinder’s volume from separately measured dimensions illustrates the difference: statistical propagation tells you how small random errors in the radius and height combine, while an unnoticed scale or calibration offset in either instrument shifts every computed volume the same way and must be assessed and reported separately as part of the total uncertainty.
How Uncertainty Moves Through a Formula
Think of a calculation as a function f(x₁, x₂, …) that maps measured inputs to a result. The mechanism of propagation is sensitivity. Near the measured values, a small change in an input produces a small change in the result, and the ratio between them is the slope of the function with respect to that input — its partial derivative, written ∂f/∂x.
Replacing the function with this local linear, tangent-line approximation is the Taylor-series step that underlies the standard formulas. These formulas can be derived rigorously: the linear rules follow from the definitions of expected value and variance, and the general nonlinear rule from a Taylor-series approximation. For independent inputs — errors that vary independently of one another — the variances add:
sresult² ≈ (∂f/∂x₁ · s₁)² + (∂f/∂x₂ · s₂)² + …
Each input contributes its uncertainty scaled by its sensitivity. Because input errors vary in sign, some push the result high while others push it low, so they partially cancel: independent uncertainties therefore combine as a root-sum-square — the square root of the sum of squares, called combining in quadrature — rather than by straight addition. A straight sum is instead the conservative worst-case rule: it assumes every input errs at once in the direction that hurts. Both conventions appear in practice, as quick rules of thumb for rough estimates, and as variance-based formulas for statistical estimates, and the difference is visible in the examples below.
Rules for Common Operations
The propagation rules depend on whether quantities were added, subtracted, multiplied, or divided. The table summarizes the two conventions: the conservative rule of thumb, which bounds the largest possible error, and the statistical quadrature estimate for independent random errors.
| Operation on the inputs | Conservative rule of thumb | Statistical estimate (independent errors) |
|---|---|---|
| Addition or subtraction | Add the absolute uncertainties | Combine absolute uncertainties in quadrature |
| Multiplication or division | Add the relative (fractional) uncertainties | Combine relative uncertainties in quadrature |
| Multiply by an exact constant C | Absolute uncertainty multiplied by |C|; relative uncertainty unchanged | Same |
| Raise an input to a power n | Relative uncertainty multiplied by |n| | Same |
| Any other function f(x) of one input | uresult ≈ |df/dx| · ux | Same linear approximation |
Two notes on using the table. First, exact constants — π, factors such as 4 or 2, and defined unit conversions — carry no uncertainty and contribute nothing. Second, a complete formula is handled by decomposing it into these operations in order; for example, the area of a circle, A = πD²/4, is an exact constant times a squared measurement, so only the power rule and the constant rule apply.
Worked Examples
The following calculations use representative values; they are illustrative rather than specific experimental data.
Adding two measured lengths
Take A = 4.0 ± 0.2 cm and B = 2.0 ± 0.1 cm. For addition or subtraction, the absolute uncertainties are added, so A + B = 6.0 ± 0.3 cm. The same rule applies to differences, which produces the cancellation effect shown in the third example.
Area of a circular pipe cross-section
Suppose a pipe diameter is measured as D = 20.0 ± 0.2 mm, a relative uncertainty of 0.2/20.0 = 1%. The area A = πD²/4 squares the diameter, so the power rule doubles the relative uncertainty to 2%. Numerically, A = 314 mm², and 2% of that is about 6 mm², giving A = 314 ± 6 mm². Because π and 4 are exact, the diameter is the sole source of uncertainty.
A temperature difference
Let a hot stream be at 85.0 ± 0.5 °C and a cold stream at 80.0 ± 0.5 °C. The difference is 5.0 °C in magnitude, but the absolute uncertainties add to ±1.0 °C under the worst-case rule — a 20% relative uncertainty in a result built from inputs known to about 0.6%. Combining the two independent uncertainties in quadrature instead gives √(0.5² + 0.5²) ≈ 0.7 °C, still about 14%. Either way, subtracting two similar numbers magnifies relative uncertainty dramatically. This is why a heat duty computed from Q = ṁcpΔT can be far more uncertain than its inputs: cancellation shrinks the result while the input uncertainties remain.
A Practical Procedure for Propagating Uncertainty
The following sequence turns propagation into a routine part of any calculation:
- Write the computation explicitly. Express the result as an equation in the measured inputs, separating exact constants. You cannot propagate what you have not formalized.
- Establish each input’s uncertainty and its type. Use statistical analysis of repeated measurements for precision error; use instrument specifications, resolution, and judgment for bias estimates.
- Decompose the formula. Break it into sums, differences, products, quotients, powers, and constant factors, and apply the matching rule — or, if the formula does not decompose neatly, evaluate the partial derivatives directly.
- Choose a combination method and state it. Use quadrature for independent statistical uncertainties; use straight addition when a guaranteed worst-case bound is required. Do not mix them silently.
- Convert between absolute and relative forms as needed and keep units consistent: uncertainties carry units, and only like quantities may be combined.
- Report the result as value ± uncertainty, rounded sensibly. The uncertainty is normally kept to one or two significant figures, and the result is rounded to the same decimal place, as in 4.0 ± 0.2 cm.
- Identify the dominant contributor. The input whose sensitivity-weighted uncertainty dominates sets the precision of the whole result; improving that one measurement yields the most benefit, while refinements elsewhere buy little.
Assumptions and Limitations
- Small-uncertainty (linear) approximation. The formulas come from a first-order Taylor expansion, so they are exact for linear operations and approximate for nonlinear functions. They are reliable when uncertainties are small relative to the curvature of the function; strongly nonlinear results with large input uncertainty deserve more careful treatment.
- Independence. The quadrature rule assumes input errors are independent. Reusing the same instrument, or exposing two inputs to the same drift, correlates them; the full multivariate propagation formula includes covariance terms to handle this, and ignoring a real correlation understates the result’s uncertainty.
- What the number means. A quadrature result is a statistical estimate at roughly a one-standard-deviation level, not a guarantee; a straight-sum result is a bound but usually overestimates for independent inputs. State which convention a reported figure uses.
- Bias is untouched. Propagation addresses precision-type uncertainty only; total uncertainty requires combining the propagated precision estimate with separately assessed bias.
Common Mistakes to Avoid
- Using the wrong rule for the operation — adding relative uncertainties across a sum, or absolute uncertainties across a product. The rules key to the operation performed on the quantities.
- Overlooking cancellation. Differences of similar quantities — temperature differences, pressure drops — can inherit large relative uncertainty even when the inputs are precisely known.
- Assigning uncertainty to exact constants, including π and defined conversion factors, which inflates results for no reason.
- Counting correlated errors as independent, for example treating repeated uses of one instrument as separate, independent measurements.
- Reporting spurious precision, such as 314.159 ± 6.288 mm², where the propagation supports 314 ± 6 mm².
- Mistaking a propagated precision estimate for total accuracy and omitting bias from the reported uncertainty.
Practical Takeaways
Propagation of uncertainty is the bridge between measuring inputs and trusting outputs. Write the calculation explicitly, classify each input’s uncertainty, apply the rule that matches each operation, combine the contributions by a stated convention, and report value ± uncertainty to commensurate precision. The most useful output is often diagnostic: the dominant contributor shows where better instrumentation or more repeats will actually pay off — and where they will not.
FAQ
Should I add uncertainties or combine them in quadrature?
Straight addition gives a conservative worst-case bound that cannot understate the error, while quadrature (root-sum-square) gives a more realistic statistical estimate when input errors are independent and random in sign. Choose based on whether you need a guaranteed bound or a realistic estimate, and state which convention you used.
Do exact constants like π or unit-conversion factors contribute uncertainty?
No. Defined constants and conversion factors are treated as exact and contribute nothing. Only measured quantities — and empirical correlations that carry their own stated uncertainty — contribute to the propagated uncertainty.
Why can a temperature difference be much more uncertain, relatively, than the temperatures themselves?
Subtracting two similar values cancels most of the magnitude while the absolute uncertainties remain, so a difference built from inputs known to a fraction of a percent can carry tens of percent relative uncertainty. Combining independent errors in quadrature softens the effect but does not remove it.
Does propagated uncertainty represent the total error of a result?
No. Propagation addresses precision-type, statistical uncertainty. Bias must be estimated separately — from knowledge of the apparatus or by comparison with a more accurate instrument — and then combined with the precision estimate to give the total uncertainty.
Leave a Reply