Article — Relative Error Calculator
Relative error calculator: percent error and beyond
Relative error is the absolute difference between a measured and a true value, divided by the magnitude of the true value: |V_measured − V_true| ÷ |V_true|. Multiplying by 100% gives percent error. Because it's dimensionless, relative error lets you compare accuracy across measurements with completely different scales.
Two thermometers each off by 1°C are not equally bad. Off by 1°C when reading body temperature (37°C) is a 2.7% error — clinically meaningful. Off by 1°C when reading oven temperature (200°C) is a 0.5% error — completely irrelevant. Relative error captures this difference; absolute error does not.
What is relative error?
Relative error is the standard dimensionless measure of accuracy in science and engineering. Two measured values disagree with a reference value by some absolute amount; relative error divides that absolute amount by the reference itself, producing a fraction that does not depend on units. A 0.5% relative error is 0.5% whether you measured millimetres or megaparsecs.
The concept dates to the seventeenth century, when astronomers and surveyors realised that comparing measurement quality across different physical scales required normalisation. In 1993 the International Bureau of Weights and Measures published the Guide to the Expression of Uncertainty in Measurement (GUM), which formalised relative error and uncertainty for all of metrology.
Tycho Brahe's pre-telescope observations of planetary positions had relative errors around 0.5% — extraordinary for the late sixteenth century. Kepler later used those measurements to derive the three laws of planetary motion. If Tycho's errors had been even five percent, the elliptical orbit pattern would have hidden in the noise and the laws would have been delayed by decades.
The relative error formula
The full set of error formulas fits in a small box:
Absolute |V_measured − V_true|Relative |V_measured − V_true| ÷ |V_true|Percent Relative × 100%Signed (V_measured − V_true) ÷ V_trueThe denominator is always the true value, never the measured value. Putting the measured value in the denominator gives a different result for the same physical disagreement depending on which way the measurement drifted, which makes comparison across measurements impossible.
Absolute vs. relative error
Absolute error has units; relative error does not. Three quick examples make the distinction concrete.
- Building height: measured 45.3 m, true 45.0 m. Absolute 0.3 m. Relative 0.67%.
- Lab balance: measured 2.48 g, true 2.50 g. Absolute 0.02 g. Relative 0.8%.
- Water density: measured 0.9970 g/cm³, true 0.9970 g/cm³. Absolute ≈ 0. Relative ≈ 0%.
Use absolute error when the physical magnitude matters — manufacturing tolerances, drug doses, surveying. Use relative error when comparing methods or instruments across scales — calibration checks, inter-lab comparisons, scientific papers.
Interpreting percent error values
Different fields treat the same percent error very differently. A 5% error is unacceptable in metrology and routine in market research.
- Under 1% — laboratory calibration, astronomy, atomic physics.
- 1–5% — most school and university experiments.
- 5–10% — typical engineering tolerances.
- 10–20% — rough estimates and quick-look calculations.
- Over 20% — usually signals a systematic problem worth investigating.
Signed relative error and direction
The unsigned form |V_m − V_t| ÷ |V_t| answers "how far off". The signed form (V_m − V_t) ÷ V_t answers "how far off and in which direction". Positive means the measurement overestimates; negative means it underestimates.
Direction matters when systematic errors are at play. If repeated measurements consistently show a positive signed error, the instrument has a bias — it reads high. If signed errors fluctuate symmetrically around zero, the errors are random rather than systematic. Most laboratory calibration procedures use signed error to detect bias and unsigned error to characterise overall accuracy.
When publishing results, report unsigned percent error for accuracy comparisons and signed error for direction. A graph of signed errors against time exposes drift; a histogram of unsigned errors shows the precision distribution.
Relative error propagation
When a calculation combines multiple measured values, errors compound according to predictable rules. The combination differs between operations.
- Addition and subtraction — absolute errors combine in quadrature: Δz = √(Δx² + Δy²).
- Multiplication and division — relative errors combine in quadrature: (Δz/z)² = (Δx/x)² + (Δy/y)².
- Power z = xⁿ — relative error multiplies by |n|: Δz/z = |n| · Δx/x.
- Logarithm z = ln(x) — error becomes absolute: Δz = Δx/x.
- Exponential z = exp(x) — error becomes relative: Δz/z = Δx.
Quadrature combination assumes the individual errors are independent and random. If two measurements share a common bias — both calibrated against the same flawed reference — errors add linearly rather than in quadrature, and the combined error is larger than the rule predicts.
Common relative-error mistakes
When the true value is 0, relative error is undefined. Report the absolute error in physical units, or compare against a small nominal reference value if one is appropriate for the context.
Five errors come up again and again:
- Dividing by the measured value — should be the true (accepted) value.
- Ignoring units consistency — measured and true must be in the same units before computing the difference.
- Premature rounding — round only the final result; carry extra digits through intermediate steps.
- Confusing percentage points with percent error — 5/10 = 50% error, but 52% − 48% = 4 percentage points, not 4%.
- Forgetting absolute value — unsigned form requires it; signed form omits it.
NIST routinely measures atomic masses to relative errors of one part in 10⁹ — equivalent to weighing a 1,000 kg car and knowing the mass to within one milligram. Mass spectrometry combined with frequency standards is what makes such precision possible.
Three sources of measurement error deserve special attention because they show up everywhere. Systematic errors are biases — the instrument reads consistently high or low. Random errors fluctuate around the true value without preferred direction. Gross errors are mistakes — recording a digit wrong, mislabelling samples, miscounting trials. Systematic errors are caught by calibration against a reference. Random errors are reduced by averaging repeated measurements. Gross errors are caught only by independent review and replication.
The distinction matters because each error type responds to different mitigations. Averaging 10 measurements cuts random error by roughly √10 ≈ 3.2, but does nothing for a systematic bias. Recalibrating against a NIST-traceable standard removes systematic bias but leaves random scatter untouched. Gross errors hide until someone double-checks the workflow — which is why scientific papers carry version-controlled data files and peer review.