Tolerance Stack-Up — Worst-Case, RSS & Statistical Methods

A Tolerance Stack Is a Structural Model of an Assembly

A tolerance stack-up predicts how permitted variation in individual dimensions changes an assembly-level quantity such as gap, overlap, alignment, bearing engagement, seal compression or interface position. The arithmetic can look simple, but the engineering model behind it is not. The analyst must decide which dimensions actually contribute, which signs they take, which datums control them, whether contributors are independent, and whether the output varies linearly with each input. In structural work the output is often not merely a dimension: a gap may determine contact sequence; an offset may create bending; an interference may establish preload; a hole-location error may change fastener load sharing. The correct stack therefore starts from the physical load path and assembly sequence rather than from every tolerance printed on the drawing.

Linear Stack-Up and Sensitivity Coefficients

For a response y that can be written locally as a function of dimensional variables x_i, a first-order stack uses sensitivity coefficients a_i = ∂y/∂x_i. For a simple axial chain the coefficients are often +1 or -1. For angular, radial or kinematic geometry they may differ from unity and can change with configuration. Writing the stack in sensitivity form makes the assumptions visible and creates a direct bridge to FEA or geometric simulation when the response is not a simple dimension.

y ≈ y₀ + Σ aᵢ(xᵢ − xᵢ₀)

Worst-Case Arithmetic — What It Guarantees

Worst-case analysis places every contributing dimension at the permitted limit that drives the output in the unfavourable direction. For a linear independent chain this is the sum of absolute tolerance contributions. Its strength is conceptual clarity: if the drawing permits all of those limits simultaneously, worst-case establishes a hard geometric bound without requiring assumptions about process distributions. Its weakness is that it can be extremely conservative for long chains, especially when many contributors are statistically centred and simultaneous extremes are improbable. Worst-case is nevertheless appropriate where interchangeability, minimum clearance, hard interference, regulatory acceptance or safety requires every conforming build to work. The analyst should state whether the result is a guaranteed geometric limit or merely a screening bound.

RSS and Statistical Tolerancing

Root-sum-square methods assume that contributors behave as random variables rather than all sitting simultaneously at their limits. For independent zero-mean variables, output variance is the sum of each input variance weighted by the square of its sensitivity. This can produce a much narrower predicted assembly distribution than arithmetic worst case, but only if the assumed distributions and independence are credible. A common mistake is to divide every bilateral drawing tolerance by three and call the result a standard deviation. That may be convenient, but it is not evidence. Production data, process capability, supplier control plans or justified engineering distributions should define the statistical model.

σᵧ² ≈ Σ (aᵢ² σᵢ²)  for independent inputs

Correlation, Common Datums and Common Process Effects

Dimensions are frequently correlated. Two holes cut in one CNC setup may shift together because the datum is offset; two thicknesses may share a material-gauge bias; left and right features may be mirror-controlled by the same fixture. Positive correlation can increase stack variation while negative correlation can reduce it. Ignoring correlation is not automatically conservative. The covariance form of the stack makes this explicit and should be used when common manufacturing causes are important. Correlation should come from process knowledge or measured data, not from arbitrary coefficients selected to obtain a convenient result.

σᵧ² ≈ aᵀ Σₓ a

When the Stack Is Nonlinear

Clearance, angle, contact sequence and geometric fit can create nonlinear outputs. A positional tolerance may change radial eccentricity through a square-root relationship; a mechanism may switch contact faces; a flexible bracket may close one gap before another. In these cases a first-order stack can miss both the mean shift and the tail behaviour. The next level is parameterised geometric evaluation, deterministic corner cases or sampling through a kinematic or FE model. Before using Monte Carlo, however, map the response with a small designed set of points. If the surface is monotonic and nearly linear, simple methods may remain adequate. If it changes slope, mode or contact state, the nonlinearity must be retained.

From Dimensional Output to Structural Consequence

The stack result becomes useful only when connected to a structural decision. A 0.4 mm offset has no inherent severity; its significance depends on whether it changes fastener engagement, introduces bending, reduces seal compression, shifts bearing contact or consumes a clearance margin. Where the structural response is sensitive, use the stack to define physically compatible geometry states and evaluate those states in the structural model. Avoid combining the worst dimensional stack with independent worst material, friction and load values unless that combined state is actually possible or is intentionally conservative and labelled as such.

Verification and Reporting

A defensible tolerance stack records the datum path, sign convention, nominal chain, permitted ranges, assumed distributions, correlations and resulting output. Independently rebuild the chain from the drawing or CAD definition and test simple perturbations to confirm the sign and sensitivity of each contributor. If statistical methods are used, compare their predicted mean and spread against measured assemblies when data become available. Report both the geometric result and the structural implication, and distinguish hard drawing limits from probabilistic production expectations. This makes the stack auditable and prevents a statistical estimate from being mistaken for a guaranteed manufacturing bound.

• Trace every contributor to a controlled drawing, model or measured feature.
• State whether the assessment is worst-case, statistical or a hybrid.
• Check correlation and common-datum effects.
• Verify nonlinear or contact-driven outputs with explicit geometry states.
• Link the dimensional result to a structural acceptance quantity.

Key takeaways

• Use worst-case arithmetic when every permitted limit must assemble or remain safe; use statistical methods only when their production assumptions are justified.
• A drawing tolerance is a permitted range, not automatically a probability distribution.
• Correlations, datum structure and nonlinear geometry can make a simple RSS stack materially wrong.
• The structural quantity of interest should drive the tolerance model, not the convenience of the dimension chain.