Tolerance stack-up for beginners: worst case, RSS and convolution

What a tolerance stack-up is, and how the three main ways of calculating it work, explained step by step with numbers. You do not need to know anything about statistics.

1. What is a tolerance stack-up?

No two parts are exactly the same. A drawing might say a part is 15.00 mm long, but the real parts come out as 14.93, 15.04, 15.08 and so on. The allowed difference is the tolerance, written like 15.00 ± 0.10.

When several parts are put together, their small differences add up. A tolerance stack-up answers one question: with all these parts, how big or small can the gap (or overlap) between them be? This matters because a gap that is too small means the parts do not fit, and a gap that is too large means play or leaks.

2. The example we will use

Three parts are placed side by side inside a box. We want to know the gap that is left.

PartSize (mm)ToleranceDirection
Box (inside length)40.00± 0.20+
Part A15.00± 0.10−
Part B12.00± 0.10−
Part C12.50± 0.10−

The box gives space (+). Each part uses up space (−). With every part at its nominal size, the gap is

40.00 − 15.00 − 12.00 − 12.50 = 0.50 mm

This is called the closing dimension. But the real parts are not all exactly nominal, so the real gap will be a little different every time. Say we need the gap to be at least 0.20 mm.

3. Method 1: worst case

The safest idea: assume the worst possible combination. The gap is smallest when the box is as short as allowed and every part is as long as allowed.

smallest gap = 39.80 − 15.10 − 12.10 − 12.60 = 0.00 mm
largest gap = 40.20 − 14.90 − 11.90 − 12.40 = 1.00 mm

A shortcut: just add all the tolerances. 0.20 + 0.10 + 0.10 + 0.10 = 0.50, so the gap is 0.50 ± 0.50 mm.

Worst case guarantees the result, as long as every part is inside its tolerance. Here it says the gap could be as small as 0.00 mm, so we cannot promise our 0.20 mm requirement. To fix that with worst case, we would have to make the tolerances tighter, and tighter parts cost more.

4. Why worst case is often too careful

For the gap to be 0.00 mm, all four parts must be at their extreme in exactly the wrong direction at the same time. Think of throwing four dice and needing four sixes: possible, but very rare. Most real assemblies have some parts a bit too big and others a bit too small, and the differences partly cancel out. Statistical methods use this fact.

5. Method 2: RSS (root sum square)

RSS is a quick statistical shortcut. It needs one idea: how much do the real parts spread around the nominal size? The usual measure is the standard deviation, written σ (sigma). A common assumption is that the tolerance covers ±3σ, so almost all parts (99.73 %) are inside it. Then:

σ = tolerance ÷ 3

(In the calculator this is the setting Normal, Cp = 1.) The recipe has four steps:

  1. Find σ of each part.
  2. Square each σ.
  3. Add the squares.
  4. Take the square root. This is the σ of the gap.
PartToleranceσ = tol ÷ 3σ²
Box0.200.06670.004444
Part A0.100.03330.001111
Part B0.100.03330.001111
Part C0.100.03330.001111
Sum0.007778

σ of the gap = √0.007778 = 0.0882 mm. The gap is then 0.50 ± 3 × 0.0882 = 0.50 ± 0.265 mm, which is 0.235 to 0.765 mm.

0.00.20.40.60.81.0Worst case: 0.000 to 1.000RSS ±3σ: 0.235 to 0.765RSS ±4σ: 0.147 to 0.853requirement: gap ≥ 0.2Gap between the parts (mm)
Worst case (grey) against RSS at ±3σ and ±4σ (blue) for the example. RSS is much narrower than worst case. The ±3σ range stays above the 0.2 mm requirement; the wider ±4σ range reaches below it.

Why squares? If the parts were all wrong in the same direction, the errors would add up in a straight line (worst case). But parts are wrong in random directions, so errors partly cancel. Squaring and then taking the square root is the mathematically correct way to add independent random errors. It is the same reason a random walk of 100 steps gets about 10 steps away, not 100.

The same with ±4σ

±3σ leaves about 2,700 assemblies in a million outside the range (about 1 in 370). Some companies want a safer range and use ±4σ instead. The σ of the gap does not change, you only multiply it by 4 instead of 3:

0.50 ± 4 × 0.0882 = 0.50 ± 0.353 mm, which is 0.147 to 0.853 mm.

Only about 63 ppm (1 in 16,000) fall outside ±4σ. The price is a wider range: its lower end, 0.147 mm, is now below our 0.20 mm requirement. So with ±4σ the statistical range no longer proves that we meet the requirement, while with ±3σ it did. This is not a contradiction: the real share of assemblies below 0.20 mm is the same in both cases (about 335 in a million, see below). The ±3σ or ±4σ choice only decides how wide a range you want to report.

RangeSmallest gapLargest gapOutside the range
RSS ±3σ0.2350.7652,700 ppm
RSS ±4σ0.1470.85363 ppm
Worst case0.0001.0000 (if all parts are in tolerance)

How many assemblies would fail?

With the σ of the gap we can also estimate the share of assemblies below the requirement. The mean gap is 0.50 and the requirement is 0.20, which is (0.50 − 0.20) ÷ 0.0882 = 3.4 sigmas away. A normal distribution has about 335 ppm beyond 3.4σ. “ppm” means parts per million: about 335 assemblies in a million (roughly 1 in 3,000) would have a gap under 0.20 mm. Worst case could not promise this at all; RSS says it is rare. Whether 1 in 3,000 is acceptable is an engineering decision.

The catch. RSS assumes the gap follows a bell curve (a normal distribution). That is true when the parts are normal, as above, and it is nearly true with many parts. It can be wrong with few parts that are not bell-shaped. The next section shows why.

6. Adding distributions: the dice idea

To go beyond RSS we need to add not just numbers but whole distributions: for each part, how likely each size is. Start with something familiar: two dice. Each die is flat (every face is equally likely). What is the chance of each total?

There is only one way to get a 2 (1+1), two ways to get a 3 (1+2 and 2+1), and six ways to get a 7. The number of ways, out of 36 combinations, forms a triangle:

1223344556675849310211112Total of the two dice (number above each bar = ways to get it)
Two flat distributions (two dice) added together give a triangle, not a flat shape.

To find the chance of a total, you add up the chances of all the pairs that make it. Doing this for every total is called a convolution. The same works for parts that can take any size, not only 1 to 6: for every possible gap, add up the chances of all the size combinations that give it. A computer does it with a fast method called the FFT.

The advantage: convolution uses the real shape of each part, whether bell-shaped, flat, or something else, and gives the real shape of the gap. It needs no bell-curve assumption.

Convolution with ±3σ and ±4σ

For the box and three parts of section 2, every part is bell-shaped (normal), and the sum of bell curves is again a bell curve. So the convolution gives exactly the RSS limits: 0.235 to 0.765 mm at ±3σ and 0.147 to 0.853 mm at ±4σ. Convolution only starts to give different numbers when the parts are not bell-shaped, as in sections 8 and 9.

To get limits from a convolution, we use the same “risk” as a bell curve has at ±3σ (2,700 ppm outside) or at ±4σ (63 ppm outside), and find where the real shape of the gap has that share outside.

7. Convolution step by step

What is the “shape” of a part?

For every part, the shape says how likely each size is. The shape is a curve (or a staircase of bars) whose total area is 1, meaning 100 %. Higher means more likely. The four shapes you will meet most often:

Box: Normal−tol+tolPart A: Uniform−tol+tolPart B: Triangular−tol+tolPart C: Trapezoidal−tol+tolHeight = how likely each size is. Each shape is centered on its nominal size.
A normal part is most likely near the nominal and rare at the tolerance limits. A uniform part is equally likely anywhere inside its tolerance. A triangular and a trapezoidal part lie in between.

Adding two shapes on a grid

A computer cannot work with a smooth curve, so it cuts the size axis into small equal steps (a grid) and stores the chance of each step. Here is a tiny example you can follow by hand. Imagine two parts whose deviation from nominal can only be −0.10, −0.05, 0, +0.05 or +0.10 mm:

We add the two deviations. For every possible total, we add up the chances of all the pairs that give it. There are 5 × 9 = 45 equal combinations in total:

Total deviation (mm)−0.20−0.15−0.10−0.050+0.05+0.10+0.15+0.20
Combinations (of 45)136898631
Chance2.2 %6.7 %13.3 %17.8 %20 %17.8 %13.3 %6.7 %2.2 %

For example, a total of 0 can happen in 5 ways: A = −0.10 with B = +0.10 (1 × 1), A = −0.05 with B = +0.05 (1 × 2), A = 0 with B = 0 (1 × 3), A = +0.05 with B = −0.05 (1 × 2) and A = +0.10 with B = −0.10 (1 × 1). That is 1 + 2 + 3 + 2 + 1 = 9 combinations out of 45. This “multiply the chances of each pair, then add all the pairs that give the same total” is exactly what a convolution is. The result (1, 3, 6, 8, 9, 8, 6, 3, 1) is already starting to look like a bell.

To add a third part, we do the same again: convolve the result with the shape of the third part. And so on, one part after another. The order does not matter.

Parts that subtract

A part with a − sign takes space away. Its shape must first be mirrored left to right, and then it is added like the others. For symmetric shapes (normal, uniform, triangular, symmetric trapezoid) the mirror image looks the same, so nothing visible changes. For a lopsided shape it matters, and the calculator does it for you.

The full recipe

  1. Shape of each part: its distribution, as a deviation from its nominal size.
  2. Mirror the parts with a − sign.
  3. Convolve the shapes one after another. The result is the shape of the gap deviation.
  4. Shift the result by the closing dimension at nominal sizes (here 0.50 mm), so it becomes the shape of the gap itself.
  5. Accumulate the chances from the left. This gives, for every gap, the share of assemblies that are smaller than it.
  6. Read the limits where the share is the risk you accept: for ±3σ that is 0.135 % below the lower limit and 0.135 % above the upper limit; for ±4σ, 0.0032 % on each side. To check a requirement, read the share below the required minimum (or above the maximum) directly.

How the computer does it quickly

Convolving two shapes of 16,384 steps the long way means about 268 million multiplications for every pair. Instead, the calculator uses the FFT (fast Fourier transform): it converts each shape into another form where convolution becomes a simple multiplication, multiplies them, and converts back. The result is the same, but much faster. The calculator uses a grid of 16,384 steps that covers ±14σ of the result, which is reliable for risks down to about 7σ (beyond that, tiny rounding errors in the grid become larger than the probabilities themselves).

8. A complete example with different distributions

We use the same box, the same sizes and the same tolerances as in section 2, but now every part has its own distribution, as in a real assembly where the parts come from different processes:

PartSize (mm)ToleranceDir.Distribution
Box40.00± 0.20+Normal (Cp = 1)
Part A15.00± 0.10−Uniform
Part B12.00± 0.10−Triangular
Part C12.50± 0.10−Trapezoidal (flat top from 25 % to 75 % of the width)

Step 1: the σ of each part and the RSS result

The σ of a part depends on its shape. For a tolerance width T (upper limit minus lower limit, here 0.40 for the box and 0.20 for the parts):

PartFormulaσ (mm)σ²
Box (normal)T ÷ 60.06670.004444
Part A (uniform)T ÷ √120.05770.003333
Part B (triangular)T ÷ (2√6)0.04080.001667
Part C (trapezoid 25/75)T × 0.22820.04560.002083
Sum0.011528

σ of the gap = √0.011528 = 0.1074 mm. This number is exact for any shapes, because variances always add up for independent parts. So RSS gives: ±3σ = 0.50 ± 0.322 = 0.178 to 0.822, and ±4σ = 0.50 ± 0.430 = 0.070 to 0.930. (With all parts normal, section 5, we had σ = 0.0882: the flatter shapes of A, B and C spread more.)

Step 2: the convolution

Now we follow the recipe: the four shapes are mirrored where needed (they are symmetric, so no visible change), convolved one after another, and shifted by 0.50. The result is the true shape of the gap:

0.00.20.40.60.81.0requirement ≥ 0.2RSS ±3σconvolutionnormal curve (same σ)Gap between the parts (mm)
The true distribution of the gap (blue area) against the normal curve that RSS assumes (red dashed). In the middle the two are almost the same. The difference is in the tails, shown enlarged below.
0.00.10.20.3requirement ≥ 0.2normal curve (RSS)true shape (convolution)Gap between the parts (mm). Vertical scale magnified 6 times.
Zoom on the lower tail, with the vertical scale magnified 6 times. Below a gap of about 0.2 mm the bell curve (red dashed) lies above the true shape (blue). So a bell curve overstates how many assemblies have a small gap.

Its center is at 0.50 and its σ is 0.1074, the same as RSS. But its shape is not exactly a bell: its kurtosis is 2.86 (a bell curve has 3), so the tails are lighter. The tails are what decides the limits. We read them as in step 6 of the recipe:

RangeSmallest gapLargest gapOutside the range
Worst case0.0001.0000
RSS ±3σ0.1780.822really 1,770 ppm, not 2,700
Convolution, risk of 3σ (2,700 ppm)0.1900.8102,700 ppm
RSS ±4σ0.0700.930really 15 ppm, not 63
Convolution, risk of 4σ (63 ppm)0.0990.90163 ppm

Read the last column of the RSS rows: the RSS range is wider than it needs to be. The ±3σ RSS range is crossed by only 1,770 assemblies in a million, not by 2,700 as a bell curve would suggest. At ±4σ the difference grows: 15 against 63. So here RSS is on the safe side, but it asks you to accept a lower minimum gap (0.070) than is really needed (0.099).

Step 3: the requirement (gap at least 0.20 mm)

With the full shape we can read the share below 0.20 directly. The mean is (0.50 − 0.20) ÷ 0.1074 = 2.8 sigmas above the requirement.

MethodAssemblies below 0.20 mm
RSS (bell curve)about 2,600 ppm
Convolutionabout 1,920 ppm
Monte Carlo (1,000,000 assemblies)about 1,920 ppm
Worst casecannot promise it (the gap can reach 0.00)

Convolution and Monte Carlo agree, and they say that about 1 assembly in 520 will be below 0.20 mm, instead of the 1 in 380 that RSS suggests. The difference is not huge here, but it grows quickly the further you go into the tails, which is exactly where requirements for rare failures are decided.

Remember. RSS gives you the right σ, but it assumes a bell-shaped gap when it turns σ into limits and failure rates. Convolution uses the real shape of every part, so it gives the right limits and rates, whatever the shapes are. When all parts are normal, the two agree.

9. When RSS and convolution disagree

A second example: a hole and a pin, whose sizes are equally likely anywhere inside their tolerance (a uniform distribution, like a die). The hole is 10.00 ± 0.30, the pin is 9.50 ± 0.30. The nominal gap is 0.50 mm.

MethodSmallest gapLargest gap
Worst case−0.1001.100
RSS, ±3σ−0.2351.235
Convolution, same risk as 3σ (2,700 ppm)−0.0691.069
RSS, ±4σ−0.4801.480
Convolution, same risk as 4σ (63 ppm)−0.0951.095

Look at the RSS rows: they claim the gap could be wider than worst case (and at ±4σ much wider). That is impossible, since worst case already uses the extreme sizes. RSS goes wrong because it assumes a bell curve, which has long tails, while two flat distributions added together make a triangle with hard ends (just like the dice).

-0.200.511.2possible gaps (triangle)worst caseRSS ±3σconvolutionGap between the parts (mm)
The possible gaps form a triangle. The RSS limits (red, dashed) fall outside it, where no assembly can be. The convolution limits (green) stay inside.

Here convolution is the better answer. (Real parts are usually somewhere between flat and bell-shaped. That is why the calculator lets you choose the distribution for each part.)

10. Monte Carlo: the brute-force check

A third way is to let the computer build the assemblies. For each one it picks a random size for every part, according to that part’s distribution, and records the gap. After a million assemblies, the results show how the gap really behaves. Monte Carlo agrees with convolution (in the hole and pin example at ±3σ: 1.069 and −0.068 against 1.0688 and −0.0688), but it is random, so each run is slightly different and the far tails are noisy. Convolution is exact up to the grid it uses, so the two methods are a good check on each other.

11. Which method should I use?

MethodGood forWatch out for
Worst caseSafety-critical fits, no failures allowed, few partsCan force very tight (expensive) tolerances
RSSA quick estimate with many parts of roughly bell-shaped sizesCan be wrong with few or non-bell-shaped parts
ConvolutionAny distributions, accurate limits and failure ratesAssumes independent parts
Monte CarloA check of the other results; very odd casesRandom: small differences between runs

A practical habit: compute worst case and a statistical method side by side. If worst case already meets the requirement, you are done. If only the statistical result does, the decision is about how much risk is acceptable.

12. Try the examples yourself

Open the tolerance stack-up calculator and enter the lines of the first example: choose + or − for each line, type the nominal and the lower and upper tolerance (for example −0.10 and 0.10), set the distribution to Normal with Cp = 1, and type 0.20 as the required minimum. You should see the numbers from this page. Then change the box to a hole and a pin with a Uniform distribution to see the second example. For the example of section 8, give the box Normal with Cp = 1, part A Uniform, part B Trapezoidal with b = 50 and c = 50 (which is a triangle), and part C Trapezoidal with b = 25 and c = 75. In the results, read the Numerical Convolution block and choose ±3σ or ±4σ in the sigma list.

13. Common mistakes

14. Learn more

The formulas behind the calculator, its limits and the public sources are on the methods and sources page. This tutorial is a simplified introduction, not a replacement for your own standards and company rules.

Open the calculator →