Simple cases (assuming uncorrelated errors)
| Operation | Propagated error |
|---|---|
| Sum / difference z = x ± y | σ_z = √(σ_x² + σ_y²) |
| Product / quotient z = x·y or x/y | (σ_z/z)² = (σ_x/x)² + (σ_y/y)² |
| Power z = x^n | σ_z/z = |n| · σ_x/x |
| Log z = ln(x) | σ_z = σ_x / x |
| Exp z = e^x | σ_z / z = σ_x |
| Constant multiple z = c·x | σ_z = |c| · σ_x |
General formula
| Function of many vars | σ_z² ≈ Σ (∂f/∂xᵢ · σ_xᵢ)² + 2 Σ ∂f/∂xᵢ · ∂f/∂xⱼ · cov(xᵢ, xⱼ) |
|---|---|
| Uncorrelated | Covariance term = 0 → σ_z² = Σ (∂f/∂xᵢ · σ_xᵢ)² |
Correlation matters
- The formulas above assume uncorrelated errors — if variables are correlated, include the covariance term.
- Calibration drift, temperature, and voltage supply often introduce correlation.
- Monte Carlo with the joint distribution is the safest approach when correlations are messy.
Best practices
- Report uncertainty and unit with every measurement: "12.3 ± 0.2 V" not just "12.3 V".
- Use standard uncertainty (1σ) unless stated otherwise.
- Expanded uncertainty uses a coverage factor (k = 2 is common, ~95% confidence for normal).
- Always distinguish systematic (bias) from random (noise) errors.
Worked examples
A quotient. A resistor is found from V = 12.0 ± 0.2 V and I = 2.00 ± 0.05 A, so R = 6.00 Ω. The relative errors are 0.2 / 12.0 = 1.67 percent and 0.05 / 2.00 = 2.5 percent. Combined in quadrature: √(0.0167² + 0.025²) = √(0.000278 + 0.000625) = 0.0300, or 3.0 percent, so σ_R = 6.00 × 0.030 = 0.18 Ω. Report R = 6.00 ± 0.18 Ω. Adding the errors linearly would give 4.2 percent, an overestimate when the errors are independent.
A sum. Lengths of 10.0 ± 0.3 cm and 5.0 ± 0.4 cm total 15.0 cm with σ = √(0.3² + 0.4²) = √0.25 = 0.5 cm.
A power. The area of a circle with r = 5.0 ± 0.1 cm is πr² = 78.5 cm². The relative error doubles for a square: 2 × 0.1 / 5.0 = 4 percent, so σ_A = 78.5 × 0.04 = 3.1 cm².
Traps to avoid
- Subtracting nearly equal numbers. 100.0 − 99.5 with ± 0.3 on each gives 0.5 ± 0.42, an error of 85 percent. Redesign the measurement to avoid differencing large values.
- Treating a repeated variable as independent. For z = x × x, the error is 2 σ_x / x (power rule), not √2 times the relative error.
- Quoting too many digits. Round the uncertainty to one or two significant figures, then round the value to the same decimal place.
- Mixing systematic and random errors. Averaging reduces random noise by √n but does nothing for a calibration offset.