Measurement Uncertainty
Write the measurement model, enter your readings, certificates and meter specs, and get the uncertainty budget with its expanded uncertainty and coverage factor. A Monte Carlo run then checks whether the GUM shortcut holds for your model.
1 · Measurement model
Operators + − * / ^ and parentheses; functions sqrt exp ln log10 sin cos tan abs asin acos atan sinh cosh tanh atan2 hypot; constant pi. Every other name is an input quantity.
2 · Input quantities
Type A takes repeated readings. Type B converts a certificate, a tolerance, a resolution or a meter spec to a standard uncertainty. Leave ν blank for ∞, or write a reliability like 25% (GUM G.4.2 gives ν = 8).
| Symbol | Type | Value xi | Parameters | νi | u(xi) |
|---|
3 · Correlations
Only for inputs that share an error source: one reference, one meter, one temperature. Most budgets have none.
4 · Coverage & Monte Carlo
Result (GUM)
Uncertainty budget
Sorted by contribution to uc². The top row is where to spend effort; halving a 5 % row changes almost nothing.
| Input | xi | u(xi) | νi | ci | |ci|·u(xi) | Share of uc² | Cum. |
|---|
Monte Carlo cross-check (JCGM 101)
From readings and specs to a defensible ± number
The GUM (JCGM 100:2008) method has four steps. Write the measurand as a function of its inputs. Give each input a best estimate and a standard uncertainty, meaning one standard deviation. Propagate them through the model. Then multiply by a coverage factor to get an interval that holds the true value with a stated probability.
Standard uncertainty of each input
A meter spec is a limit, not a standard deviation. "±(0.0035 % of reading + 0.0005 % of range)" on the 10 V range at 5 V is a = 0.000175 + 0.00005 = 0.000225 V, so u = 0.13 mV. Entering the limit as if it were u overstates the meter by √3. Entering it as a k = 2 normal understates it.
Propagation (law of propagation of uncertainty)
The tool computes t quantiles from the inverse regularized incomplete beta function, not by table lookup, so fractional ν works too. The test suite checks it against 26 entries of GUM Table G.2, including k(∞, 95.45 %) = 2.00, k(10, 95.45 %) = 2.28 and t₉₉(16) = 2.92. The "GUM H.1 end gauge" example reproduces the GUM's worked budget: uc = 32 nm, νeff = 16, k = 2.92, U₉₉ = 93 nm.
Why a Monte Carlo cross-check
The GUM formula linearizes the model at the estimate and assumes the result is Gaussian or a scaled t. That is fine for sums and gentle products. It fails for a quantity near a minimum (a magnitude, a power near zero), for strong curvature, and for budgets dominated by one rectangular input. JCGM 101 (GUM Supplement 1) propagates the full distributions instead:
Random numbers come from the enhanced Wichmann–Hill generator, Box–Muller normals and the ratio method for t draws, all as specified in JCGM 101 Annex C. A seed makes runs repeatable. The tool also splits the run into 10 blocks, and the spread of the block results shows whether M is large enough. S1 §8.2 asks for endpoints that are stable to δ/5 before you trust a validation.
The verdict has four grades. Validated means both endpoint gaps are within δ. Inconclusive means the gap is no bigger than the Monte Carlo noise, so raise M. Agrees to 1 digit means the GUM interval is right to one significant digit of u but not two. Not validated means quote the Monte Carlo interval. Each grade also says whether the GUM interval is wider (conservative) or narrower (optimistic). A budget dominated by rectangular inputs typically gets a slightly conservative GUM interval. A magnitude or a ratio near zero gets one that is simply wrong, as the "Magnitude near zero" example shows.
Limits
- This is an engineering estimate and teaching aid, not accredited laboratory software. For ISO/IEC 17025 work, validate it against your own reference budgets.
- The GUM column is first order: it omits the higher-order terms of GUM 5.1.2. Use the Monte Carlo result when the two disagree.
- Correlated inputs are sampled through a Gaussian copula. That is exact for normal inputs and approximate for rectangular, triangular and U-shaped ones. A correlated Type A or finite-ν input is sampled as a normal. With correlations, Welch–Satterthwaite is only approximate.
- Type A inputs with n ≤ 3 have a t-distribution with infinite variance (S1 §6.4.9.4). The Monte Carlo interval is still valid, but its standard deviation is not.
- The Monte Carlo run uses a fixed M, not the adaptive procedure of S1 §7.9. The block spread tells you when to raise M. Inexact limits (curvilinear trapezoid) and asymmetric limits are not modelled.
References
- JCGM 100:2008, Evaluation of measurement data — Guide to the expression of uncertainty in measurement (GUM), BIPM. §4.2–4.3, §5.1–5.2, Annex G (Table G.2, Welch–Satterthwaite), Annex H.1.
- JCGM 101:2008, Supplement 1 to the GUM — Propagation of distributions using a Monte Carlo method, BIPM. §6.4, §7.7, §7.9.2, §8, §9.2, Annex C.
- B. N. Taylor, C. E. Kuyatt, Guidelines for Evaluating and Expressing the Uncertainty of NIST Measurement Results, NIST Technical Note 1297, 1994.
- EA-4/02 M: 2022, Evaluation of the Uncertainty of Measurement in Calibration, European co-operation for Accreditation.
- B. A. Wichmann, I. D. Hill, "Generating good pseudo-random numbers," Computational Statistics & Data Analysis 51 (2006) 1614–1622.
- NIST/DLMF, Digital Library of Mathematical Functions, §8.17 (incomplete beta function); Abramowitz & Stegun 26.7.5 (t quantile expansion).
- T. Lafarge, A. Possolo, "The NIST Uncertainty Machine," NCSLI Measure 10(3), 2015 (Gaussian copula for correlated inputs).
- Scope references only, no code used: StephenDPerryJr/uncertainty-budget-calculator, kyb8801/measurement-uncertainty-mcp (MIT).
Building a test fixture that has to hold a tolerance?
Kyros designs verification rigs and production test stations, and the uncertainty budgets that show a pass/fail limit is real and not just noise.
