Explained clearly 5 min read

The Impact of Rounding Negative Numbers on Uncertainty Propagation: A GUM Perspective

Rounding negative numbers can bias estimates and inflate uncertainty if rounded values are propagated. Under the GUM, propagate unrounded values and round only the final estimate and uncertainty.

Short Answer

Rounding negative numbers can bias estimates and inflate uncertainty if rounded values are propagated. Under the GUM, propagate unrounded values and round only the final estimate and uncertainty.

This reference article examines how rounding negative numbers interacts with uncertainty propagation under the Guide to the Expression of Uncertainty in Measurement (GUM). It is part of our precision-and-rounding reference library, not merely a calculator page. The site’s significant figures calculator implements these conventions, but the educational value lies in understanding why sign, rounding mode, and propagation order matter.

Rule Statement

Under the GUM, the measurand Y = f(X1, …, XN) is estimated by y = f(x1, …, xN). The combined standard uncertainty is obtained from the law of propagation of uncertainty, JCGM 100:2008 (GUM) Clause 5.1.2:

uc2(y) = Σ (∂f/∂xi)2 u2(xi) + 2 Σ Σ (∂f/∂xi)(∂f/∂xj) u(xi, xj).

Rounding is not a formal uncertainty component unless a rounded value is used as an input to a subsequent calculation. Negative numbers obey the same magnitude rules as positive numbers, but half-way conventions can be sign-dependent. For round-to-nearest, the rounding error is bounded by ±0.5q, where q is the quantum (the increment between adjacent representable values). If rounding is treated as a genuine rectangular-distribution error source, its standard uncertainty is uround = q/√12. For asymmetric methods such as floor, ceiling, or truncation toward zero, a sign-dependent bias appears.

Key rule: propagate unrounded values whenever possible; round only the final estimate and its uncertainty. GUM Clause 7.2.2 recommends that uncertainty be reported with at most two significant digits and that the estimate be rounded consistently with the uncertainty.

Standards Citation

  • JCGM 100:2008 (GUM), Clause 5.1.2 — law of propagation of uncertainty for correlated and uncorrelated inputs.
  • JCGM 100:2008 (GUM), Clause 5.2.2 — expanded uncertainty U = k uc(y).
  • JCGM 100:2008 (GUM), Clause 7.2.2 — numerical reporting of estimate and uncertainty; avoid excessive digits.
  • ISO 80000-1:2022, Clause 7.3 — general rules for rounding, including half-way conventions.
  • ASTM E29-22, Section 6 — using significant digits and rounding in test data conformance.
  • NIST TN 1297, Section 7 — guidelines for evaluating and expressing uncertainty, including rounding for reporting.

Convention Comparison Table

Rounding method −2.5 −3.5 −2.4 −2.6 Bias for negative inputs
Half away from zero −3 −4 −2 −3 Symmetric; zero mean
Half up (toward +∞) −2 −3 −2 −3 Positive for half cases
Half down (toward −∞) −3 −4 −2 −3 Negative for half cases
Half toward zero −2 −3 −2 −3 Positive for half cases
Half even (banker’s) −2 −4 −2 −3 Symmetric; zero mean
Floor −3 −4 −3 −3 Negative bias
Ceiling −2 −3 −2 −2 Positive bias
Truncate toward zero −2 −3 −2 −2 Positive bias for negatives

Worked Examples

Example 1: Rounding a negative correction before propagation

Suppose a calibration correction is c = −0.037 V with standard uncertainty u(c) = 0.004 V. If the correction is rounded to two decimal places before use, crounded = −0.04 V. The quantum is q = 0.01 V. Treating rounding as a rectangular error source:

  1. Half-width = q/2 = 0.005 V.
  2. uround = q/√12 = 0.01/3.464 = 0.00289 V.
  3. Combined standard uncertainty: uc = √(0.0042 + 0.002892) = 0.00493 V.
  4. Expanded uncertainty with k = 2: U = 2 × 0.00493 = 0.00986 V ≈ 0.010 V.

Report: c = −0.04 V ± 0.01 V. If the unrounded value had been propagated, uc = 0.004 V and U = 0.008 V, reported as −0.037 V ± 0.008 V. Intermediate rounding inflated U by about 23%.

Example 2: Difference of negative temperatures

Let T1 = −12.34 °C with u = 0.05 °C and T2 = −8.76 °C with u = 0.05 °C. The temperature difference is ΔT = T1 − T2 = −3.58 °C. Since ∂ΔT/∂T1 = 1 and ∂ΔT/∂T2 = −1, the signs vanish in the variance:

uc = √(0.052 + 0.052) = 0.071 °C. With k = 2, U = 0.14 °C, reported as −3.58 °C ± 0.14 °C.

If each temperature is rounded to 0.1 °C first: −12.3 °C and −8.8 °C, giving ΔT = −3.5 °C. The rounding error is +0.08 °C. Each rounding contributes uround = 0.1/√12 = 0.0289 °C, so uc = √(0.052 + 0.052 + 0.02892 + 0.02892) = 0.081 °C and U = 0.16 °C. The sign of ΔT does not change the variance, but it does change the bias if asymmetric rounding is used.

Counter-Examples

  1. Using q/2 as the standard uncertainty. This overestimates uround by a factor of √3. The correct rectangular standard uncertainty is q/√12.
  2. Assuming all half-up methods round negatives away from zero. Half-up toward +∞ maps −2.5 to −2, while half-away-from-zero maps −2.5 to −3. The reported value and bias differ.
  3. Double rounding. Rounding −1.245 to two decimals gives −1.25 under half-away, then to one decimal gives −1.3. Direct rounding to one decimal gives −1.2 because −1.245 is closer to −1.2. Double rounding introduces avoidable error.
  4. Treating floor as unbiased for negative values. Floor always moves negative values more negative, producing a negative bias. Ceiling and truncation toward zero produce a positive bias for negative inputs.

Common Mistakes

  • Rounding intermediate negative values before uncertainty propagation.
  • Mixing half-up and half-even conventions in the same calculation chain.
  • Reporting U with more digits than the estimate, contrary to GUM Clause 7.2.2.
  • Forgetting that truncation toward zero biases negative values upward (less negative).
  • Applying significant-figure rules to uncertainty instead of GUM’s two-significant-digit guidance.
  • Using software defaults without documenting the negative half-way behavior.

Software Behavior Note

Tool Negative half-way behavior Example: round(−2.5) to integer
Python round() Half to even −2
Excel ROUND Half away from zero −3
Google Sheets ROUND Half away from zero −3
MATLAB round Half away from zero −3
R round() Half to even (IEC 60559) −2
JavaScript Math.round Half toward +∞ −2

The GUM does not prescribe a software rounding mode. For reproducibility, document the mode and avoid rounding before propagation. See our related Banker’s Rounding and Rounding Methods references.

Discipline Note

Chemistry often uses significant figures and half-up conventions, while physics and engineering metrology typically follow the GUM and ISO/IEC Guide 98-3. In analytical chemistry, ASTM E29 and ISO 80000-1 are common. Negative values appear in corrections, temperature differences, pH, electrode potentials, and celestial magnitudes. The sign should not alter the uncertainty propagation variance, but it does alter bias for asymmetric rounding. For further reading, see our Error Propagation and GUM Guide articles.

Quick Reference Table

Situation Rule Uncertainty impact
Round-to-nearest negative number Use a symmetric convention; half-even is preferred for reproducibility uround = q/√12 if treated as a source
Truncation toward zero (negative) Rounded value is greater than true value Positive bias, mean +q/2
Floor (negative) Rounded value is less than true value Negative bias, mean −q/2
Final reporting Round estimate and U together; report U to two significant digits Do not propagate rounded values

FAQ

Does the sign of a negative number affect GUM uncertainty propagation?

No for variance: squared sensitivity coefficients remove the sign. Yes for bias if asymmetric rounding is used, because floor, ceiling, and truncation treat negative values directionally.

How do I treat rounding uncertainty for negative values?

If rounding is a genuine input error, model it as a rectangular distribution with half-width q/2, so u = q/√12. For final reporting, the GUM treats rounding as a presentation step, not an uncertainty component.

What rounding mode should I use for negative numbers?

ISO 80000-1 and ASTM E29 allow half-up or half-even depending on context. For metrology, half-even reduces bias; document the choice.

Can rounding negative numbers change expanded uncertainty?

Yes, if rounded values are propagated. Intermediate rounding increases u_c and can shift the estimate. Use unrounded values until the final report, then round the estimate and U consistently.

Verified sources

References

  1. JCGM 100:2008, Evaluation of measurement data — Guide to the expression of uncertainty in measurement (GUM), Clauses 5.1.2, 5.2.2, 7.2.2.
  2. ISO 80000-1:2022, Quantities and units — Part 1: General, Clause 7.3.
  3. ASTM E29-22, Standard Practice for Using Significant Digits in Test Data to Determine Conformance with Specifications, Section 6.
  4. NIST Technical Note 1297, Guidelines for Evaluating and Expressing the Uncertainty of NIST Measurement Results, Section 7.
  5. JCGM 101:2008, Supplement 1 to the GUM — Propagation of distributions using a Monte Carlo method.

Leave a Reply

Your email address will not be published. Required fields are marked *