Tolerance stack-up: methods, formulas and sources
What the tolerance stack-up calculator does, the formulas it uses, and where they come from. New to the subject? Start with the beginner tutorial.
The model
A dimension chain is a list of lines. Each line has a nominal value N, a lower and an upper tolerance, and a direction: it either adds to the closing dimension (+) or subtracts from it (−). The closing dimension is the result of the loop:
CD = Σ si · Ni, with si = +1 or −1.
A positive CD is a gap between the neighbouring parts, a negative CD means interference. The lines are treated as independent of each other, and the chain is linear (every line is a length along one direction).
Worst case
Every line is placed at the limit that makes the closing dimension as large (or as small) as possible. For a line that adds, the upper limit goes into the maximum and the lower limit into the minimum; for a line that subtracts, it is the other way round. The result contains every assembly the tolerances allow, so it needs no assumption about distributions. Its drawback is that it is usually very cautious for long chains.
Statistical method (RSS)
Each line has a standard deviation σi and a mean. For independent lines the variance of the closing dimension is the sum of the variances, and the mean is the sum of the means:
σCD = √(Σ σi2), meanCD = Σ si · meani
The statistical range is mean ± kσCD, where you choose k (3, 4, 5.5 or any value). The share of assemblies inside the range is read from a normal distribution, so this method assumes the result is close to normal. With a few lines that have very different shapes, or with a mean shift, the normal assumption can be poor; that is why the calculator also shows the two methods below. The “RSS shifted mean” block uses the true means of the lines (for example after a mean shift or an uneven tolerance such as −0.05 / −0.01) instead of the middle of the limits.
Distribution of each line
The mean and σ of a line come from its distribution and its tolerance width T (upper limit minus lower limit):
- Normal: σ = T / (6·Cp). With Cp = 1 the tolerance limits are at ±3σ. An optional mean shift moves the mean away from the middle of the limits.
- Uniform: σ = T / √12, which is a/√3 for a half-width a = T/2.
- Trapezoidal: a flat top and sloping sides; you give where the flat top starts and ends as a percentage of the tolerance width. σ follows from the trapezoid shape; the uniform and the triangular distributions are the two extreme cases.
- Custom: you type the mean, σ, skewness and kurtosis. The calculator fits a generalised lambda distribution to the skewness and kurtosis. It accepts kurtosis between about 1.8 and 20, with kurtosis ≥ skewness2 + 1 and a one-humped shape, and refuses other combinations with a message.
Numerical convolution
The probability density of a sum of independent quantities is the convolution of their densities. The calculator computes this numerically for the closing dimension with the fast Fourier transform, on a grid of 16,384 cells that covers ±14σ. It is used to show the true shape of the result and to read limits for the chosen k. The result is trusted out to about 7σ; beyond that, rounding in the grid makes very small probabilities unreliable. The limits it shows are the values that leave the same share outside as a normal distribution does at ±kσ.
Monte Carlo
The Monte Carlo method draws random values for every line and adds them up, many times. The calculator draws 1,000,000 assemblies in the browser, each line from its own distribution, and builds a histogram of the closing dimension. Because it is random, a rerun gives slightly different numbers, mostly in the extreme tails, where only a few samples fall.
Skewness and kurtosis of the result
The cumulants of a sum of independent variables are the sums of the cumulants of the variables. The calculator uses this to give the skewness and kurtosis of the closing dimension from those of the lines (a line that subtracts changes the sign of its third cumulant). The kurtosis can never be smaller than skewness2 + 1.
Temperature
The change of a length with temperature is described by the coefficient of linear thermal expansion α: the fractional change in length per degree. The calculator uses a size factor
f = 1 + α · 10−6 · (T − Tref), with α in 10−6 per °C and T in °C.
A difference of one kelvin equals a difference of one degree Celsius, so the same α works in both. A value given per °F must be multiplied by 1.8. Two modes are offered, and they are a modelling choice, not a standard:
- Mean and tolerances (default): the nominal value, the limits, the mean and σ of the line are all multiplied by f.
- Mean only: the line moves by N·(f − 1); its tolerance width and σ stay as given. This fits cases where the tolerances are defined at the reference temperature and the part only changes its size.
The expansion is taken as linear in temperature, which is a good approximation over small temperature ranges.
Limits of the calculator
- Linear chains only: every line is a length in one direction. Angles, 2D and 3D chains and geometric tolerances are not covered.
- Lines are independent. Correlated dimensions (for example two features made in the same operation) are not modelled.
- The results depend on the distributions you choose. A Cp or a shape that does not match the real process gives results that do not match the real parts.
- This calculator is an aid for engineers. Check results that matter against your own methods, your company rules and the applicable standards.
Sources
Public sources used for the formulas above. Last checked: 7 November 2026.
- JCGM 100:2008, Evaluation of measurement data – Guide to the expression of uncertainty in measurement (GUM), BIPM. Rectangular distribution (variance a2/3, section 4.3.7), trapezoidal distribution (section 4.3.9) and the combined standard uncertainty as the square root of the sum of squares (sections 5.1.2 and 5.1.3). doi.org/10.59161/JCGM100-2008E
- JCGM 101:2008, Evaluation of measurement data – Supplement 1 to the “Guide to the expression of uncertainty in measurement” – Propagation of distributions using a Monte Carlo method, BIPM. doi.org/10.59161/JCGM101-2008
- Wikipedia, Tolerance analysis: worst-case analysis (variables placed at their tolerance limits, 100 percent of parts assemble) and the statistical and Monte Carlo approaches.
- Wikipedia, Cumulant: the n-th order cumulant of a sum of independent variables is the sum of their n-th order cumulants.
- Wikipedia, Kurtosis: the kurtosis is bounded below by the squared skewness plus 1 (Pearson, 1916).
- Wikipedia, Convolution of probability distributions: the distribution of a sum of independent random variables is the convolution of their distributions.
- P. Ruckdeschel and M. Kohl, General Purpose Convolution Algorithm in S4-Classes by means of FFT (arXiv:1006.0764): computing such convolutions numerically with the fast Fourier transform.
- Freimer, Mudholkar, Kollia and Lin (1988), A Study of the Generalized Tukey Lambda Family, Communications in Statistics – Theory and Methods, 17, 3547–3567, and the NIST Dataplot reference page GLDPDF for the quantile function Q(p) = λ1 + (1/λ2)·[(pλ3 − 1)/λ3 − ((1 − p)λ4 − 1)/λ4].
- Wikipedia, Monte Carlo method: repeated random sampling to obtain numerical results.
- Wikipedia, Thermal expansion: the coefficient of linear thermal expansion is the fractional change in length per degree of temperature change; its SI unit is the inverse kelvin.
The temperature factor and the two temperature modes, the fitting of the custom distribution and the grid used for the convolution are choices made in this calculator and are not taken from the sources above. Picrezo is an independent project and is not affiliated with any other software or company.