Will your assembly fit when every part is somewhere inside its tolerance? List the dimensions in the loop between two faces, mark whether each one opens or closes the gap, and enter its deviations. The calculator gives the mean gap and its range by worst case, by RSS (root sum of squares) and by modified RSS, shows which part drives the variation, estimates how many assemblies would fall outside your gap limits, and tells you how much to tighten to make the worst case pass.
Engineering estimate, not a quote. A one-dimensional stack assumes every dimension acts along one line; angles, form errors and part deflection need a 2D/3D analysis.
Every gap or fit in an assembly is the result of a chain of part dimensions. If you start at one face of the gap and walk through the parts to the opposite face, each dimension you cross either moves you toward the gap's positive direction or back. The gap is the signed sum of those dimensions, and because each dimension can be anywhere inside its tolerance, the gap has a range. A stack-up calculates that range so you can compare it with what the design needs: enough clearance to assemble and move, not so much that the parts rattle or a seal leaks.
Step 1: draw the loop. Pick the gap, choose a positive direction across it, and trace from one side of the gap through every contacting part back to the other side. Each step through a part is one dimension: the distance between the two faces where that part touches its neighbours. Mark it + if it points in the positive direction and − if it points back. Only dimensions in the loop count, and each appears once.
Step 2: convert to mean and equal tolerance. Drawings often carry unequal tolerances such as +0/−0.12. Before adding, convert each to its mid-point and a symmetric tolerance: mean = nominal + (upper + lower) ÷ 2 and t = (upper − lower) ÷ 2. This is the value the parts will scatter around if the process is centred in its tolerance.
Step 3: mean gap. Add the means with their signs. If a dimension acts on the gap through a ratio, such as a diameter whose radius sits in the loop, multiply its mean and tolerance by that sensitivity factor.
Step 4: worst case. Add all the tolerances. The gap is the mean ± Σt. This range is guaranteed: if every part is in tolerance, every assembly is inside it.
Step 5: RSS. If the dimensions are independent, centred and roughly normal, their variances add, not their ranges. Treating each tolerance as the same number of standard deviations, the gap tolerance is √Σt². Because it is very unlikely that every part sits at the bad end of its tolerance at once, RSS is narrower than worst case, and the more parts in the stack, the bigger the difference.
Step 6: modified RSS and defect rate. Real processes drift and are rarely perfectly centred, so many teams multiply the RSS result by a correction factor, usually 1.5 (Bender's factor), and never let it exceed the worst case. From the RSS standard deviation the calculator also estimates the fraction of assemblies outside your gap limits, and the gap's Cpk, under the stated assumptions.
A gear shaft assembly: a molded housing pocket 39.80 ± 0.15 mm deep holds a ball bearing 12.00 +0/−0.12 mm wide, a spacer 8.00 ± 0.05, a gear hub 18.00 ± 0.08 and a thrust washer 1.50 ± 0.04. The end gap between the washer and the housing wall must stay between 0.05 mm (so nothing binds) and 0.80 mm (so the gear does not walk enough to mis-mesh).
The two methods disagree in a way that matters. By RSS the gap is comfortably inside its limits and a defect is a one-in-a-million event, if the parts really are centred and normally distributed. By worst case, a housing at its shortest with every other part at its longest leaves a 0.02 mm interference and the gear binds. Notice also that the bearing's one-sided tolerance moved the mean gap from 0.30 to 0.36: anyone who added nominals would have been 0.06 mm off before considering tolerance at all. Finally, the housing pocket alone carries 61.5% of the RSS variance. Tightening it from ±0.15 to ±0.08 makes the worst case exactly 0.05 to 0.67 mm and cuts the RSS tolerance to √0.0205 = 0.143 mm; tightening the washer from ±0.04 to ±0.02 instead reduces the RSS tolerance by under 2%. Press Load the worked example above to reproduce every number.
The contribution columns answer the most useful question a stack-up can ask: where should the tolerance budget be spent? Under worst case each part contributes in proportion to its tolerance, but under RSS it contributes in proportion to the square of its tolerance, so the largest tolerance dominates far more than its share of the sum suggests. In the example the housing has 39.5% of the worst-case sum but 61.5% of the RSS variance. Tightening a small contributor buys almost nothing, and loosening it costs almost nothing. That is the real value of the RSS view even for teams that sign off on worst case: it shows which tolerances can be opened up to make parts cheaper and easier to make, and which one must be controlled.
With n dimensions that all carry the same tolerance t, worst case gives n·t and RSS gives √n·t. For two dimensions RSS is 71% of worst case, for four it is 50%, for nine 33%, and for sixteen 25%. That is why RSS is attractive for long stacks and nearly pointless for short ones: on a two-part stack it saves little, and the assumption of independent, centred parts is hardest to defend when a single bad part drives the result. In the example, five unequal tolerances give an RSS of 0.191 against a worst case of 0.38, almost exactly half.
The catch with RSS is that it is a statement about populations. It says nothing about the one assembly built from a short housing and three long parts, which still exists and still binds. RSS is the right choice when that rare assembly is acceptable, for example because it is caught at end-of-line test, can be sorted, or simply has a mild consequence. It is the wrong choice for a safety function, a low-volume build where every unit matters, or parts from a process you have never measured.
The RSS assumptions were written with machined parts in mind, and formed parts break them in predictable ways. A multi-cavity mold produces one population per cavity, each with its own mean, because the cavities are never identical and fill differently; a stack that uses parts from cavity 1 and cavity 4 is not one centred distribution. Shrinkage moves the whole population: a molded dimension often sits off-centre for the life of a tool if the cavity was cut on a shrinkage estimate that turned out slightly wrong. Die-cast and stamped parts drift as the tool wears. All of these push the real result toward worst case, which is the reason for the modified RSS factor.
Two practices make stack-ups on formed parts far more reliable. First, cut critical cavity dimensions steel-safe: put the first-shot dimension on the side where removing steel corrects it, so the mean can be moved after first parts are measured. Second, once tools run, replace the drawing tolerances in the stack with measured means and spreads from first-article and capability studies, per cavity. That turns the stack from a prediction into a check, and it often shows that a worst-case fail is a real risk on one cavity and no risk at all on the others.
| Method | Use it when | Limit |
|---|---|---|
| Worst case (this page) | Short stacks, safety or functional-critical gaps, low volume, unmeasured processes | Pessimistic for long stacks; drives tight, expensive tolerances |
| RSS (this page) | Four or more independent dimensions, volume production, capable centred processes | Assumes centred normal distributions; a statement about populations, not each unit |
| Modified RSS, Bender factor (this page) | Volume parts without capability data, or formed parts with likely mean shifts | The factor is a rule of thumb, not derived from your process |
| Monte Carlo simulation | Non-normal or skewed distributions, measured data per cavity, non-linear stacks | Needs distributions for every input; results depend on them |
| 2D/3D tolerance analysis software | Angles, geometric tolerances interacting in several directions, hole patterns | Needs a full model and correct datum scheme |
The vector-loop stack-up, the conversion to mean and equal bilateral tolerance, and the worst-case and root-sum-square methods follow the practice described in B. R. Fischer, Mechanical Tolerance Stackup and Analysis, in M. F. Spotts, Dimensioning and Tolerancing for Quantity Production, applied to parts dimensioned and toleranced per ASME Y14.5 (which defines how the drawing is dimensioned; it does not itself prescribe a stack-up method). The modified RSS factor of 1.5 is from A. Bender, "Benderizing tolerances: a simple practical probability method of handling tolerances for limit-stack-ups" (1962). The out-of-limit estimate uses the standard normal distribution with each part tolerance taken as ±kσ, and the gap Cpk uses the usual process-capability definition. The normal tail is computed in your browser with a published complementary error function approximation accurate to better than 1.2 × 10⁻⁷ relative error. The flag thresholds the calculator uses (a single contributor above 50% of the RSS variance, a gap Cpk below 1.33) are page-set engineering practice, not standards.
A tolerance stack-up adds up the dimensions and tolerances of every part feature that lies between two surfaces of an assembly, to find the range of the gap or fit between them. You trace a closed loop from one side of the gap, through each part and its mating faces, to the other side, and give each dimension a plus sign if making it bigger opens the gap and a minus sign if it closes the gap. The mean gap is the signed sum of the means, and the tolerance of the gap is found either by adding all the tolerances (worst case) or by taking the square root of the sum of their squares (RSS).
Use worst case when every single assembly must fit, when the stack has only two or three dimensions, when the volume is low, or when a failure is a safety issue: the gap tolerance is the sum of all tolerances, so it is guaranteed if every part is in tolerance. Use RSS when the stack has four or more independent dimensions made in volume by capable processes: it assumes the parts are centred and normally distributed, so the extremes rarely combine, and the gap tolerance is the square root of the sum of squares. With n equal tolerances, RSS is smaller than worst case by a factor of √n: half as large for four dimensions, a third for nine. Many teams design to RSS and then check worst case or modified RSS for the critical stacks.
Convert each dimension to a mean with an equal plus and minus tolerance t, then add the means with their signs and take the square root of the sum of the squared tolerances. For a housing pocket of 39.80 ± 0.15 mm containing a bearing 12.00 +0/−0.12 (mean 11.94 ± 0.06), a spacer 8.00 ± 0.05, a gear hub 18.00 ± 0.08 and a thrust washer 1.50 ± 0.04, the mean end gap is 39.80 − 11.94 − 8.00 − 18.00 − 1.50 = 0.36 mm. The squares sum to 0.0225 + 0.0036 + 0.0025 + 0.0064 + 0.0016 = 0.0366, so the RSS tolerance is √0.0366 = 0.191 mm and the gap is 0.169 to 0.551 mm. Worst case, the tolerances simply add to 0.38 mm and the gap is −0.02 to 0.74 mm, which means an interference is possible.
Modified RSS multiplies the RSS tolerance by a correction factor, usually 1.5, to allow for real parts that are not perfectly centred or normally distributed. The factor comes from A. Bender's 1962 paper on statistical tolerancing and is often called Benderizing. In the bearing stack example, the RSS tolerance of 0.191 mm becomes 1.5 × 0.191 = 0.287 mm, so the gap is 0.073 to 0.647 mm instead of 0.169 to 0.551 mm. The result should not exceed the worst-case sum, so the calculator caps it there. Modified RSS is a practical middle ground when you cannot accept worst-case tolerances but do not have process data to justify plain RSS.
Convert it to a mean with an equal bilateral tolerance before adding: mean = nominal + (upper deviation + lower deviation) ÷ 2 and t = (upper deviation − lower deviation) ÷ 2. A bearing width of 12.00 +0/−0.12 mm becomes 11.94 ± 0.06 mm. Skipping this step is a common error: using the nominal 12.00 in the sum puts the predicted gap 0.06 mm away from where the parts will actually centre, and using 0.12 as a plus and minus tolerance doubles that part's contribution. In the example the nominal-only gap is 0.30 mm but the true mean gap is 0.36 mm.
If each part tolerance equals ±3 standard deviations of a centred normal distribution, the gap's standard deviation is the RSS tolerance divided by three, and the fraction out of limits comes from the normal distribution. In the bearing stack example the gap has a mean of 0.36 mm and a standard deviation of 0.191 ÷ 3 = 0.0638 mm. The minimum allowed gap of 0.05 mm is 0.31 ÷ 0.0638 = 4.86 standard deviations below the mean, which corresponds to about 0.6 assemblies per million, and the gap's Cpk is 0.31 ÷ 0.191 = 1.62. These numbers are only as good as the assumptions: off-centre processes, multi-cavity molds and non-normal distributions can raise the real rate by orders of magnitude.
First re-centre the nominals so the mean gap sits in the middle of the allowed range, then tighten the dimension with the largest contribution. Under RSS a contributor's share is its tolerance squared over the sum of squares, so the biggest tolerance dominates. In the bearing stack the housing pocket at ±0.15 mm contributes 61.5% of the RSS variance; tightening it to ±0.08 mm cuts the RSS tolerance from 0.191 to 0.143 mm and makes the worst case exactly 0.05 to 0.67 mm, which passes. Tightening the washer from 0.04 to 0.02 instead reduces the RSS tolerance by under 2%. Other fixes are removing parts from the loop, locating parts against a common datum, or adding a shim or adjustment.
The direction sets the sign of each dimension's mean in the gap, so it decides the mean gap. It does not change the tolerance of the gap, because worst case adds the absolute tolerances and RSS adds their squares. To get the directions right, draw the loop: start at one face of the gap, step through each part along the measured direction, return to the other face, and mark each step plus if it moves you toward the positive gap direction and minus if it moves you back. A dimension that appears twice in the loop, or a dimension that is not in the loop at all, is the usual cause of a wrong mean gap.
Not sure which dimensions are in the loop, or which tolerance a molded or cast part can really hold? Ask our AI assistant. It answers from this website, and our engineering team follows up if you want.
AI assistant. Answers can be wrong; nothing it says is a quote or a promise.
This calculator gives an engineering estimate. Our engineering team reviews your CAD and drawings for the loop, the datum scheme and which tolerances belong on which part, while each change is still a drawing edit and not a tool modification.
Book a 30-min call Send a drawing or sample