Tolerance Stack-Ups in Mechanical Design: From Worst-Case Arithmetic to Functional Variation

A tolerance stack-up is best understood as a model of functional uncertainty. Dimensions are merely one way in which that uncertainty enters the assembly. Form error, orientation, location, datum establishment, thermal growth, elastic compliance and process drift can all move the functional output even when every individual part is “within tolerance”.

This is why simple arithmetic chains are useful but insufficient. The most important step is not choosing between worst-case and RSS; it is constructing the correct functional model. A beautifully calculated stack built on the wrong datum scheme can be more dangerous than a rough calculation that represents the actual assembly.

The discussion below uses the language of ISO GPS because it provides a coherent framework for size, geometry and datums, but the reasoning is equally applicable to ASME-based drawing systems when the applicable rules are defined consistently.

1. Start from the functional requirement, not the dimension chain

Every meaningful stack should begin with a measurable functional requirement. Examples include:

• minimum clearance between a connector and panel opening;
• maximum shaft-to-seal eccentricity;
• gear centre distance;
• minimum thread engagement;
• display-to-window offset;
• gap available for a thermal interface pad;
• distance between a sensor and target;
• minimum wall thickness after multiple surfaces and features are located.

Call that output y. The design contains input variables x₁ … xₙ. The engineering task is to understand how changes in each input move the output. In the simplest linear chain the sensitivities are ±1. In a real assembly, lever arms, angles, radial clearances and contact conditions can make the sensitivity very different.

This sensitivity view is valuable because it tells the designer which dimensions are worth controlling. A ±0.05 mm tolerance on a feature with near-zero sensitivity to the critical function may add cost with almost no benefit, while a seemingly generous tolerance on a high-sensitivity lever arm may dominate the failure probability.

NIST research on design tolerancing makes a similar point from a product-model perspective: tolerance decisions are most valuable when they are brought into the design process early and linked to assembly function, rather than applied only after detailed geometry is fixed.

2. Express the assembly as a sensitivity model

For a one-dimensional stack, assign a sign to each contributing dimension according to whether increasing it increases or decreases the closing dimension. If the functional gap is:

The signs matter for nominal value, but a classical worst-case tolerance uses the absolute magnitude of each contribution because each variable can move in the adverse direction.

For rotational or two-dimensional problems, the model should capture geometry rather than forcing the problem into an arbitrary linear chain. A small angular error θ over a lever arm L creates approximately L·θ lateral motion for small angles. Hole position within a circular tolerance zone does not behave like independent ±X and ±Y rectangular limits. A floating fastener joint changes the contact condition once clearance is consumed. These are model choices, not calculator choices.

3. Worst-case analysis: conservative for a reason

Worst-case analysis assumes every contributor reaches the most unfavourable permitted value simultaneously. For a linear stack with sensitivities of magnitude one:

This can look excessively conservative because the simultaneous occurrence of every extreme may be improbable. But probability is not always the relevant design criterion. Worst-case is appropriate when:

• the function must be guaranteed for any conforming part combination;
• production volumes are low and statistical assumptions cannot be supported;
• the consequence of assembly failure is high;
• suppliers or processes may change and the distribution is not controlled;
• limits are contractual but process centring data is not available;
• the variables are correlated or deliberately sorted in a way that invalidates simple statistical independence.

Worst-case is also an excellent architecture test. If the design needs implausibly tight component tolerances to pass a worst-case stack, it may be better to shorten the chain, introduce float, change the datum structure or make the interface self-locating rather than purchase precision everywhere.

4. Statistical RSS: powerful, but only when its assumptions are true

Root-sum-square analysis is often introduced as “the realistic method” because random variation is unlikely to place every feature at its limit simultaneously. For independent contributors with approximately normal distributions, the variance of a linear output can be propagated as:

The common mistake is to take drawing tolerances, square them, add them and take the square root without asking what the tolerances represent statistically. A specification limit is not a standard deviation. A supplier may run off-centre, use a non-normal process, sort product, change tooling cavities, or deliver a mixed population from several machines. The RSS result can then be precise-looking but physically unjustified.

Correlation matters as well. The more general form includes covariance terms:

Correlation is not an academic edge case. Features machined in one setup can share datum error; dimensions moulded in one cavity can move together with shrinkage; PCB features referenced to the same imaging/tooling system are not necessarily independent; temperature changes move several dimensions together according to material CTE.

ISO 22514 provides a broader framework for process capability and performance. The 2026 revision of ISO 22514-2 explicitly addresses time-dependent process models rather than assuming every process remains in a simple state of statistical control. That is a useful reminder: the real production distribution deserves evidence.

5. Monte Carlo and non-linear variation

Monte Carlo analysis is valuable when the relationship between inputs and output is non-linear, when contacts switch, when angular and radial effects interact, or when the input distributions are not conveniently represented by a closed-form equation.

The method is conceptually simple:

1. Define each input distribution from credible process or supplier data.
2. Define correlations where they physically exist.
3. Sample the inputs repeatedly.
4. Resolve the assembly/function for each sample.
5. Inspect the resulting output distribution and predicted failure rate.

The difficulty lies in steps 1 and 2. Ten million simulations do not rescue poor assumptions. If every input is assigned a centred normal distribution simply because the software asks for one, the result can be less trustworthy than a transparent worst-case calculation.

Monte Carlo becomes especially useful for geometric problems involving contact, clearance and orientation, but the model should be verified against limit cases and physical builds. Simulation is a reasoning tool, not an authority.

6. GD&T changes the problem from size arithmetic to geometry

Traditional ± dimensions are weak at describing functional geometry. A hole can satisfy its size tolerance and still be badly located; a face can have an acceptable average location while being warped; a pin pattern can meet pairwise dimensions while rotating relative to the mating interface.

The ISO GPS framework separates these concepts. ISO 8015 defines fundamental GPS principles. ISO 5459:2024 defines datums and datum systems, while ISO 1101:2017 defines the symbol language for geometrical tolerancing.

For assembly analysis, the practical consequence is that a tolerance stack may need to include:

• size variation;
• position of a feature relative to a datum reference frame;
• orientation of a surface or axis;
• form variation that changes real contact;
• datum feature variation and the way the datum simulator establishes the reference frame;
• material modifiers and any available bonus tolerance.

ISO 2692:2021 formalises maximum material, least material and reciprocity requirements. Maximum material concepts are particularly useful for assembly functions such as pin/hole fit because size and geometric location can be coupled in a way that reflects the real virtual condition of the mating feature.

The important point is not to “add GD&T” decoratively. The specification should control the geometry that drives function and provide an inspection method that is economically sensible.

7. Datum strategy controls what variation means

A dimension only has engineering meaning relative to how the part is established in the assembly. If a panel is physically seated on three pads and located by two tabs, but its drawing dimensions critical holes from a cosmetic outer edge, inspection can certify parts that assemble poorly and reject parts that would assemble correctly.

Datum selection should normally follow functional contact and constraint:

• primary datum: the surface or set of points that establishes the dominant seating plane;
• secondary datum: the feature that controls the next most important orientation/translation;
• tertiary datum: the feature required to close the remaining necessary freedom.

That does not mean every assembly literally uses a three-two-one fixture, but the logic is powerful. It prevents a drawing from establishing a reference frame that the real product never sees.

Datum shift and clearance can also be beneficial. If the joint genuinely permits float, the tolerance model should allow it instead of pessimistically assuming every fastener hole sits concentrically at nominal position. Conversely, if friction or clamp sequence can freeze the joint unpredictably, the apparent float may not be reliable enough to claim as functional margin.

8. Thermal expansion and elastic deformation belong in the stack

Geometric tolerances are normally specified at a reference temperature. ISO 1:2022 establishes the standard reference temperature concept for dimensional and geometrical properties. Products, however, rarely operate permanently at the reference condition.

The first-order thermal change in length is:

For a 100 mm aluminium feature using α ≈ 23 µm/m·K over a 50 K temperature rise, the free expansion is approximately 0.115 mm. A polymer feature with α ≈ 70 µm/m·K over the same length and temperature range would move about 0.35 mm. The differential movement between them is roughly 0.235 mm before moisture absorption, moulding anisotropy or constraint are considered.

That can exceed the machining tolerance on many precision features. Thermal behaviour therefore belongs in the functional stack whenever temperature changes the interface.

Elastic deformation can be equally important. Screw preload can pull thin sheet metal flat, distort plastic bosses, compress gaskets and bend PCBs. The assembled geometry is the result of tolerance plus stiffness plus load. For high-sensitivity interfaces, a purely rigid-body stack is incomplete.

9. Process capability is evidence, not permission to ignore tolerance

Capability indices are useful when a stable, understood process exists. They describe the relationship between process variation and specification, not the intrinsic correctness of the specification.

A design with a needlessly tight tolerance can still have a capable process if enough money is spent on tooling, inspection or sorting. Conversely, an excellent functional specification can be produced by an unstable process with poor capability. Design and process questions must therefore be separated:

• Design question: what geometric range can the product tolerate and still function?
• Process question: can the chosen process repeatedly deliver inside that range with suitable margin?

When reliable capability data exists, statistical assembly analysis can use the measured distributions rather than default assumptions. When it does not, drawing limits remain the defensible basis for guarantee.

10. Tolerance synthesis: allocate variation where it costs least

Tolerance analysis asks what happens with a given set of tolerances. Tolerance synthesis asks how the available functional budget should be distributed across the contributors.

The cheapest solution is rarely to divide the budget equally. Different features have different:

• sensitivities to the function;
• manufacturing processes and natural capability;
• inspection cost;
• supplier risk;
• opportunities for adjustment or float;
• impact on other functions.

A rational allocation gives looser tolerances to low-sensitivity or expensive-to-control features and tighter control to high-sensitivity features that are economically capable. Better still, change the architecture so that high-variation dimensions do not sit in the critical chain.

This is where tolerance work becomes design work. NIST's research on tolerance synthesis frames the problem as allocating geometric variation to satisfy functional constraints rather than merely documenting an already-fixed geometry.

11. Tolerance Stack-Up Analysis Example: PCB Connector to Panel Opening

Consider a simplified horizontal alignment requirement between a PCB-mounted connector and an enclosure opening. Assume the nominal design provides ±1.00 mm lateral clearance from connector body to opening edge on each side.

Worst-case: ±(0.20 + 0.15 + 0.10 + 0.15) = ±0.60 mm. The minimum remaining lateral clearance is therefore 1.00 − 0.60 = 0.40 mm.

If those four limits were instead defensibly interpreted as equivalent independent statistical bounds with common coverage, an RSS combination of the numerical values would be about 0.31 mm. That predicts much more margin, but it is not a guaranteed limit and it is only meaningful if the statistical assumptions behind those numbers are supported.

Now add a 0.3° possible rotational error of the board about a mounting feature and suppose the connector mating face is 45 mm away. The additional lateral motion is approximately:

This is why the functional model matters. A purely translational chain could have produced a reassuring but incomplete answer. For more on this particular interface, see PCB-to-enclosure integration.

12. Verification and measurement uncertainty

The last step is to ensure that the specified characteristic can actually be verified. Measurement uncertainty, fixturing, datum simulation and sampling strategy all affect the confidence of the result.

A tolerance narrower than the practical measurement capability creates ambiguity at acceptance. Likewise, a datum scheme that cannot be reproduced on the CMM or production gauge with the same physical logic used in the assembly can produce technically precise but functionally irrelevant inspection data.

For high-value stacks, the verification plan should state:

• which characteristics are key or critical;
• how each datum is established during inspection;
• the measurement method and expected uncertainty;
• whether capability data will be collected;
• how changes in supplier, tooling or cavity will be handled;
• how the stack will be revisited if production data contradicts assumptions.

Common failure modes

A practical tolerance-analysis workflow

1. Write the functional requirement and acceptance limit in measurable terms.
2. Identify the physical contacts and datum structure that establish the assembly.
3. Draw the dimensional/geometric loop from one side of the function to the other.
4. Assign sensitivity and sign to each contributor.
5. Include orientation, contact, thermal and elastic effects where they can be material.
6. Run worst-case first to understand guaranteed margin.
7. Use statistical analysis only where input distributions and correlations are credible.
8. Change architecture before buying precision if the stack is fragile.
9. Allocate tolerances according to sensitivity, process capability and cost.
10. Define how the critical characteristics will be inspected and controlled.

Tolerance stack-up questions

Standards and research references

• ISO 8015:2011 — GPS fundamentals, concepts, principles and rules.
• ISO 5459:2024 — datums and datum systems.
• ISO 1101:2017 — geometrical tolerancing.
• ISO 2692:2021 — maximum material, least material and reciprocity requirements.
• ISO 14405-1:2025 — dimensional tolerancing of linear sizes.
• ISO 22514-1:2014 and ISO 22514-2:2026 — process capability and performance principles.
• NIST — Design for Tolerance of Electro-Mechanical Assemblies.
• NISTIR 6524 — Information Models for Design Tolerancing.