Statistical Tolerance Analysis (RSS)
Statistical tolerance analysis uses probability to predict how tolerances combine. The most common method—Root Sum of Squares (RSS)—assumes tolerances follow normal distributions and won't all hit their limits simultaneously. This guide explains when RSS is valid, when it's not, and how to use it correctly.
The Statistical Argument
Worst-case analysis assumes every part hits its tolerance limit at the same time, in the worst direction. Statistically, this is extremely unlikely. Most parts land closer to nominal, and their variations partially cancel out.
RSS captures this reality mathematically. Instead of adding tolerances directly, it combines them as independent random variables:
RSS Formula
Where t₁, t₂, etc. are individual tolerance values. The result is always smaller than the sum of tolerances.
RSS vs. Worst-Case: A Comparison
Consider a stack with 5 dimensions, each with ±0.10mm tolerance:
| Method | Calculation | Result |
|---|---|---|
| Worst-Case | 5 × 0.10 | ±0.50mm |
| RSS | √(5 × 0.10²) | ±0.22mm |
RSS predicts 56% less variation. This difference matters—it can be the difference between parts that require expensive precision machining and parts that can be made with standard processes.
The Assumptions Behind RSS
RSS isn't magic—it's based on statistical assumptions. If these assumptions don't hold, RSS predictions are wrong:
1. Normal Distribution
RSS assumes each dimension follows a normal (Gaussian) distribution centered on the nominal. This is generally true for machined features but may not hold for cast, molded, or purchased parts with skewed distributions.
2. Statistical Independence
Each dimension must vary independently. If two features come from the same machining setup and their errors correlate, RSS underestimates the combined variation.
3. Adequate Process Capability
Standard RSS assumes a 3σ process (Cpk ≈ 1.0). If your supplier can't hold that capability, the actual distribution is wider than assumed and RSS is optimistic.
4. Sufficient Sample Size
Statistical averaging only works with large populations. For a one-off prototype or small batch (<100 units), RSS predictions are meaningless—you might get the unlucky worst-case combination.
When NOT to Use RSS
Do NOT use RSS when:
- • Safety is critical — Aircraft, medical devices, automotive safety systems
- • Production volume is low — Prototypes, custom builds, small batches (<100)
- • 100% interchangeability required — Field-replaceable parts with no sorting
- • Process capability is unknown — New suppliers or processes without Cpk data
- • Failure cannot be inspected — Sealed units or inaccessible interfaces
- • Distributions are non-normal — Castings, injection molding, or skewed processes
In these cases, use worst-case analysis. Yes, it requires tighter tolerances and costs more. But the alternative is accepting risk you can't quantify or manage.
When RSS Is Appropriate
RSS is valid when:
- High production volume — Thousands of units where statistics apply
- Established processes — Demonstrated Cpk ≥ 1.33 on contributing dimensions
- Cost pressure — WC tolerances would require expensive precision
- Fallout is manageable — You can inspect and scrap the rare failures
- Multiple independent sources — Parts from different suppliers, machines, lots
Understanding the Fallout Rate
RSS predicts a range that contains most assemblies, not all. Standard 3σ RSS covers 99.73% of the population—meaning 0.27% (about 3 in 1000) fall outside.
| Coverage | Sigma Level | Fallout Rate |
|---|---|---|
| 68.27% | 1σ | 317 per 1000 |
| 95.45% | 2σ | 45 per 1000 |
| 99.73% | 3σ (standard) | 2.7 per 1000 |
| 99.9937% | 4σ | 0.06 per 1000 |
If 2.7 failures per 1000 is unacceptable, either use worst-case or multiply your RSS result by a safety factor (e.g., 1.5× for ~4σ coverage).
Beyond Basic RSS: Advanced Methods
RSS is the simplest statistical method. More sophisticated approaches exist:
Monte Carlo Simulation
Randomly samples each dimension's distribution thousands of times and builds a histogram of results. Handles non-normal distributions and correlations but requires more data and computation.
Modified RSS (Bender Method)
Adds a correction factor for small numbers of contributors. Standard RSS assumes many contributors; with <5, Bender's method gives more conservative results.
Six Sigma Tolerancing
Assumes 6σ process capability and calculates tolerances to achieve target Cpk. Common in automotive and high-volume manufacturing.
For most engineering work, standard RSS is sufficient. Advanced methods add complexity without proportional benefit unless you have the process data to support them.
Best Practices
- Always run WC first — If WC passes, you don't need RSS
- Document your justification — State why RSS is acceptable (volume, Cpk, fallout plan)
- Know your process capability — Don't assume Cpk ≥ 1.33 without data
- Plan for fallout — How will you catch and handle the statistical outliers?
- Don't mix methods — Use WC or RSS for the entire stack, not both in one analysis
- Show both results to reviewers — Let them see the WC range even if you recommend RSS
Summary
Statistical tolerance analysis (RSS) predicts tighter assembly variation than worst-case by accounting for the statistical improbability of all tolerances hitting their limits simultaneously. It's valid for high-volume production with controlled processes and acceptable fallout risk.
RSS is not appropriate for safety-critical applications, low volumes, or situations where 100% interchangeability is required. When in doubt, use worst-case.