How to run a comparison 01 Pick a preset or type a model. The presets include y = x², x₁·x₂ and the pendulum, so the interesting failures are one click away. 02 Give every input a distribution. Normal, rectangular, triangular, scaled-and-shifted t, log-normal or arcsine, with ν where it is known. 03 Set M, seed and coverage. The default is 100 000 trials, seed 20250101 and 95%. 04 Read both rows. y, u(y), the symmetric interval and the §8 differences; the verdict text says how to interpret them. 05 Decide what to report. When the endpoint differences stay inside the tolerance, quote the first-order result; when they exceed it, quote Monte Carlo.
Worked example Seed 20250101, M = 100 000, 95% coverage; δ is the §8 numerical tolerance on the standard uncertainty.
Model
GUM u(y)
MC u(y)
MC 95% interval
Endpoint differences vs δ = 5.00e-2
Verdict
x^2, x ~ N(0, 1)
0
1.41508161
[0.00092727, 5.05221926]
9.27e-4 / 5.05 — agree: no
nonlinear: use MC (skewness 2.7978 )
x1 * x2, N(10, 1), N(5, 0.5)
7.07106781
7.0995449
[36.9247005, 64.60907171]
7.84e-1 / 7.50e-1 — not resolved at this M
raise M
x1 + x2 + x3 + x4, four N(0, 1)
2
2.00529595
[−3.90620878, 3.96027218]
1.37e-2 / 4.03e-2 — inside δ
first-order validated
The square preset is the cleanest failure in the whole science section: at x = 0 the derivative 2x vanishes, so the GUM row reads u = 0 and a zero-width interval, while the true distribution is χ²-like with mean 1.00083028 , standard uncertainty 1.41508161 and a 95% interval from 0.00092727 up to 5.05221926 . Monte Carlo also recognises the shape: skewness 2.7978 , and a shortest interval [0 , 3.84543713 ] that is much tighter than the symmetric one.
The §8 test, and what it needs JCGM 101 §8 fixes a numerical tolerance δ = ½·10^l, where u(y) = c·10^l with c an integer of the chosen digits — 5.00e-2 for these examples. The first-order result is validated only if both interval endpoints differ from the Monte Carlo endpoints by no more than δ. Two conditions must hold for that comparison to mean anything: the Monte Carlo noise must be small against δ, and the intervals must be the same kind. The tab also reports whether the comparison is resolved — requiring δ ≥ 3× the sampling noise u(y)/√M — because at small M the verdict flips randomly rather than measuring nonlinearity. The square preset stays resolved from M = 10⁴ to 10⁶ while failing agreement every time; the product preset's u differs from the GUM value by only 0.4011% of the Monte Carlo value (the panel measures the gap against u_MC) yet its verdict at M = 100 000 is not resolved , because the endpoint differences (0.784 and 0.750 ) dwarf δ and the noise (2.25e-2 ) is too large for a clean decision.
Limits and what it does not do
Monte Carlo is only as good as the input distributions. A rectangular distribution is not a normal one with the same σ, and the coverage interval changes shape accordingly — see the four uniform inputs, whose interval [−3.86488045, 3.90530028] is narrower than the normal case's ±3.92 at the same u.
The model must be a typed expression. Correlations enter as an explicitly entered matrix; a measurement whose inputs share a drift without a recorded correlation is silently optimistic.
The trap: a tiny u difference is not a pass The product preset's standard uncertainties differ by only 0.4011% of the Monte Carlo value — small enough to look like agreement — but the §8 check compares intervals, not u, and fails. Conversely, agreement can be a coincidence of a favourable M: before quoting a validated verdict, read the resolved line. The third trap lives in the square preset's zero: u = 0 is not a strong statement, it is the linearisation reporting that the slope is flat at the evaluation point, which says nothing about the spread 1.4 units away.
Where it is used Calibration and test reports A report that quotes a linearised uncertainty for a curved model can be wrong in the second digit. Running both methods and attaching the §8 verdict makes the choice of result auditable.
Teaching why derivatives are not enough The square and product presets are small enough to compute by hand: one shows a vanishing derivative, the other the missing u₁²u₂² term. They turn "use Monte Carlo" from a slogan into a shown failure.
Privacy Parsing, sampling and the comparison run in your browser with a seeded generator; the model and inputs are never uploaded.
References
BIPM JCGM, JCGM 100:2008 — the GUM first-order method. bipm.org (访问日期:2026-10-07)
BIPM JCGM, JCGM 101:2008 — Monte Carlo propagation, §7.9 tolerance and §8 validation. bipm.org (访问日期:2026-10-07)
NIST, Technical Note 1297 — evaluating and expressing uncertainty. nist.gov (访问日期:2026-10-07)
Wikipedia, Propagation of uncertainty — linearisation and its failure. en.wikipedia.org (访问日期:2026-10-07)
Wikipedia, Monte Carlo method — sampling through a model. en.wikipedia.org (访问日期:2026-10-07)
Calculators in this hub
Hand-picked tools, one click away. The mini versions compute live and carry your values into the full calculator.
Sources & review
Reviewed by CalcX Editorial Team
Updated 2026-10-07