Every assembly is a stack-up. A shaft goes into a bore, a PCB slides into an enclosure, a wing rib mates to a spar. Each dimension carries a tolerance, and those tolerances combine — sometimes additively, sometimes not. How you model that combination determines whether the parts fit on the shop floor and whether the design is manufacturable at all.
Two methods dominate. Worst-case tolerancing assumes every component lands at its extreme simultaneously. Statistical tolerancing assumes dimensions follow a distribution — usually normal — and combines them using root-sum-square (RSS) mathematics. Worst-case guarantees fit at the extremes but over-constrains the design. Statistical tolerancing opens up the tolerance band, reduces cost, and aligns the drawing with what the process actually produces.
The choice is not academic. At a major aerospace manufacturer, I have seen machined fittings held to tolerances that added no functional value but drove scrap rates into double digits because the worst-case stack-up was modelled conservatively. Switching to statistical tolerancing, backed by capability data, recovered the margin without touching the process.
The Worst-Case Trap
Worst-case (arithmetic) tolerancing is linear. Take a five-part stack-up, each component held to ±0.05 mm. The worst-case assembly tolerance is 5 × 0.05 = ±0.25 mm. The design must accommodate that full range, or the parts will not assemble at the extremes.
The logic is simple and defensible: if every part arrives at its maximum material condition, the assembly still fits. But the probability of that event is vanishingly small. If each dimension is centred with a Cpk of 1.33, the chance that all five land at the extreme simultaneously is roughly one in several million.
The cost of that safety is real. Tighter component tolerances mean slower machining, more inspection, more scrap, and higher tooling wear. Suppliers charge for the precision, and they are right to. The designer asked for it.
The RSS Alternative
Statistical tolerancing replaces linear addition with the root-sum-square method. The same five-part stack-up, each at ±0.05 mm, yields an assembly tolerance of √(5 × 0.05²) = ±0.112 mm. That is less than half the worst-case result. The design envelope opens, component tolerances can be relaxed, and cost drops — provided the assumptions hold.
The core assumption is that the component dimensions are independent and normally distributed, centred on the nominal, with variation driven by common causes. If those conditions are met, the RSS result predicts assembly performance accurately. If they are not — if a process is drifted, a fixture is worn, or a supplier is running at the edge of the tolerance to maximise material usage — the RSS calculation understates the real risk.
Stack-Up Comparison: Five Components at ±0.05 mm
When Statistical Tolerancing Fails
RSS is a model, and models break when their inputs are wrong. The three failure modes are predictable: mean shift, non-normal distributions, and correlated dimensions.
Mean shift is the most common killer. A supplier holds ±0.05 mm but runs the process at +0.04 mm to save material or compensate for tool wear. The range looks compliant, the Cpk looks acceptable, but the distribution is no longer centred. The RSS assumption is violated, and the assembly tolerance is no longer ±0.112 mm — it is biased, and the bias compounds across multiple components.
Non-normality is the second failure mode. A process with a dominant tool-wear mechanism produces a uniform or trapezoidal distribution, not a Gaussian curve. RSS assumes the tails are thin; a uniform distribution has heavy tails relative to its range. The statistical calculation is now optimistic.
Correlated dimensions are the third. If two parts in the stack-up are machined on the same fixture, their variation is not independent. RSS treats them as uncorrelated; reality does not. The combined variance is underestimated.
I have audited plants where statistical tolerancing was applied to a deep drawing process with a known springback bias. The RSS calculation said the assembly was fine. The scrap report said otherwise. The gap between model and reality was closed only when the engineering team walked the floor and saw that the press was running at one end of its capability window.
Conditions for Valid Application
Before an organisation adopts statistical tolerancing, it must establish the conditions that make the mathematics valid. These are not optional prerequisites — they are the foundation. Skip them, and RSS becomes a number that looks rigorous but is not.
First, the processes producing the components must be demonstrably stable. That means SPC charts, capability studies, and evidence that the process is in statistical control. Without that, you have no basis for assuming a normal distribution.
Second, the capability must be sufficient. A Cpk of 1.33 is the floor. Below that, the tails are too heavy and the RSS result is unreliable. For critical or safety-related stack-ups, 1.67 is the appropriate target.
Third, there must be a mechanism to detect and respond to drift. If a process shifts after launch, the RSS assumption breaks. This requires ongoing SPC and a defined reaction plan — not a one-time capability study filed away at PPAP.
Maturity Ladder for Statistical Tolerancing
- OptimisationTolerances are actively managed using live capability data; cost reductions are pursued through tolerance relaxation
- RSS applicationStatistical stack-up analysis is applied to assemblies where processes are stable, centred, and capable
- Capability evidenceEvery tolerance used in a statistical stack-up is backed by current Cpk data and SPC records
- Process stabilitySPC is in place on critical characteristics; processes are in statistical control
- Process definitionPFMEA identifies critical-to-quality dimensions; control plans specify how they are monitored
Practical Rules for the Drawing
Statistical tolerancing must be explicit on the drawing. ISO 22081 and ASME Y14.5 both provide mechanisms for declaring that a tolerance is statistical, not arithmetic. If the drawing is silent, the supplier assumes worst-case — and prices accordingly.
The most expensive tolerance is the one the process can hold but the drawing does not allow.
State the method. Identify which dimensions participate in the statistical stack-up. Provide the capability assumption — the Cpk target the supplier must demonstrate. Require PPAP-level capability data for those dimensions at initial submission and at defined intervals thereafter.
Maintain a dynamic link between the tolerance calculation and the SPC data. If a supplier's process drifts and Cpk falls below the threshold, the statistical basis is broken. The tolerance must either be tightened — returning to worst-case for that component — or the process must be corrected. This is not a periodic review task; it is a continuous monitoring obligation.
Implementation Checklist
- Identify the stack-ups in your product where worst-case tolerancing is over-constraining the design or driving cost.
- For each, verify that the contributing processes are under statistical control and have demonstrated capability (Cpk ≥ 1.33).
- Confirm that the processes are centred on nominal. Mean shift is the most common reason RSS fails in practice.
- Check for independence. If two or more dimensions share a process, fixture, or setup, treat them as correlated and adjust the RSS calculation.
- Annotate the drawing with the statistical method, the required Cpk, and the SPC submission requirements.
- Establish a monitoring cadence. Capability at PPAP is necessary but not sufficient. The RSS assumption must hold for the life of the programme.
The Cost of Getting It Wrong
Used correctly, statistical tolerancing reduces cost, improves manufacturability, and aligns the drawing with process reality. Used incorrectly, it produces assemblies that do not fit — and the failure appears at the worst possible moment, on the line, at rate.
The decision to apply RSS is an engineering decision, not a drafting convention. It requires evidence, discipline, and a quality system that can detect drift before it reaches the customer. If that system exists, the mathematics is a powerful tool. If it does not, the mathematics is a liability.