Propagate independent input uncertainty with GUM and Monte Carlo. Compare distributions, histograms, paired samples and estimated 95% output coverage intervals.
Choose a formula, enter each mean and width, then select the distribution and trial count. Zero widths are valid, and custom expressions accept zero or negative means; blanks and negative widths are rejected. Edits restart the calculation and cancel the old run. Custom expressions use x1–x4, arithmetic and powers, comparisons and conditional expressions, and supported functions such as Math.pow and Math.sqrt (up to 2000 characters). Names are labels; expressions still use x1–x4. Shared URLs include the expression, labels and numeric settings. Old links without a formula require you to enter it.
For normal inputs, width means standard deviation σ. For uniform and symmetric triangular inputs, it means interval half-width a. All input variables are sampled independently. Reuse the same x number for repeated occurrences of one variable. There is no correlation matrix, CSV export or automatic unit conversion.
The default stress F/(b·h), with F=10000 N, b=50 mm and h=20 mm, has nominal value 10 MPa. Normal input standard deviations 200 N, 0.5 mm and 0.3 mm give first-order contributions 0.20, 0.10 and 0.15 MPa. Their root-sum-square is about 0.26926 MPa. MC statistics vary between finite random samples and also reflect nonlinearity.
For custom y=x1, mean 0 and width 0 give output mean 0, standard deviation 0 and coverage interval [0,0]. With mean 0 and normal σ=2, the population standard deviation is 2. Uniform half-width 2 gives 2/√3≈1.15470; symmetric triangular half-width 2 gives 2/√6≈0.81650.
Physical preset domains: stress F/(b·h) requires b,h>0, with signed F allowed; Reynolds requires density, diameter and viscosity>0 and speed U≥0; the power model requires positive stress amplitude σ and reference stress S. Every MC trial is checked against this domain. Any out-of-domain trial stops aggregation rather than being discarded. Normal inputs have unbounded tails, which may extend beyond the physical domain.
First-order GUM uses u²(y)≈Σ(∂f/∂xi)²u²(xi), with distribution widths converted to standard uncertainties and derivatives evaluated numerically at the input means. This local approximation can fail for nonlinear functions or domain boundaries. For y=x1² at mean x1=0, its first-order contribution is zero even when the output varies.
MC evaluates the formula on joint input draws and reports the output sample mean, sample standard deviation and 2.5th/97.5th sample percentiles. This estimates a 95% output coverage interval; it is not a confidence interval for the mean. Histogram bars count trials. Scatter points retain at most 500 actual paired trials. Undefined outputs stop aggregation and report their count; they are never silently discarded.
More trials do not correct unsuitable input distributions, missing correlations or an inadequate model. Relative uncertainty is undefined at mean 0 and unstable near zero. The power model(S/σ)^m has no reference life or calibrated S–N curve and does not predict absolute fatigue life.
For normal inputs it is standard deviation σ. For uniform and symmetric triangular inputs it is half-width a, converted to standard uncertainty a/√3 and a/√6 respectively.
GUM is a first-order approximation at the input means; MC estimates the output from a finite sample of the selected distributions. Check nonlinearity, trial count and input assumptions. MC is not automatically exact.
Relative uncertainty divides standard uncertainty by the absolute mean and is undefined at zero. As an explicit display rule, a ratio is also withheld when |MC mean| ≤ 3×sample SD/√trials because it is unstable near zero. This is not an ISO acceptance criterion. Standard deviation and coverage interval remain visible.
No. Keep units consistent in your formula. The default stress example uses newtons and millimetres, giving N/mm², equivalent to MPa.
BIPM — JCGM 101:2008 · NIST — Type B uncertainty · NIST — Coverage factors