Choose your reading route
eduKateSG · Why Mathematics? · Process Capability Analysis: Cp, Cpk, Pp, Ppk and Specification Limits
Start with the question, audit the worked example, or jump to the assumptions and practice. Every chapter stays open for uninterrupted reading.
Understand the question
Define the quantity and follow the mechanism.Route 2 · Chapters 9–16
Reproduce the example
Audit the numbers one step at a time.Route 3 · Chapters 17–24
Check assumptions
See where design and calibration matter.Route 4 · Chapters 25–32
Explore applications
Connect the method to a concrete investigation.Route 5 · Chapters 33–42
Study and verify
Practise, inspect failure cases and read further.
Full chapter index · Practice and checking · Mathematics Learning Hub
Why is mathematics important when a process must fit real requirements? Process capability analysis compares a stable process with specification limits. Cp, Cpk, Pp and Ppk separate spread, centring, short-term variation and longer-term performance so that improvement targets the mechanism rather than a headline number.
Did You Know? Two processes can have the same Cp but different Cpk because Cp ignores where the mean sits. A process centred at 10.0 and one centred at 10.1 can have the same spread, yet the second is closer to the upper limit.
Full chapter index
Chapters 1–8 · Understand the question
Chapters 9–16 · Reproduce the example
Chapters 17–24 · Check assumptions
Chapters 25–32 · Explore applications
Chapters 33–42 · Study and verify
Process capability asks how the distribution of a stable process compares with specification limits that define acceptable performance. It is not a score for one item and not a substitute for inspecting time order. A capable process should be predictable enough for its estimated location and spread to have a meaningful relationship with requirements. Capability analysis therefore begins with the process, measurement system, sampling window and specification source. The index is the last step of a chain of evidence, not the first number to request.
The lower specification limit, LSL, and upper specification limit, USL, should come from functional, engineering, customer or regulatory needs. They are not estimated from process data. Widening specifications to improve an index changes the requirement rather than the process. Some characteristics have only one meaningful limit, such as a maximum contaminant level or minimum strength. Record the source, version, unit, rounding rule and whether limits apply to individual output, a lot average or another defined quantity.
For a roughly normal stable process, the interval from the mean minus three standard deviations to the mean plus three standard deviations contains most individual output under the model. Its width is six standard deviations, which motivates Cp. “Natural” does not mean guaranteed, and sample estimates have uncertainty. If data are skewed, multimodal, rounded or dependent, a normal six-sigma description may be misleading. Plot the data and time order before compressing spread into a denominator.
Cp equals (USL minus LSL) divided by six times a within-process standard deviation. It compares specification width with short-term process spread and ignores where the process mean sits. In the worked example, LSL=9.5, USL=10.5 and within sigma=0.10, so Cp=1.0/0.60=1.67. That says the tolerance is 1.67 times the modelled six-sigma width. It does not say the process is centred, stable or guaranteed to meet every future requirement.
Cpk uses the smaller distance from the mean to either specification, divided by three within standard deviations. With mean 10.1, USL distance is 0.4 and LSL distance is 0.6. CPU=0.4/0.30=1.33 and CPL=0.6/0.30=2.00, so Cpk=1.33. The upper side limits capability. The difference between Cp 1.67 and Cpk 1.33 is a centring penalty. If the mean were exactly 10.0 under the same spread, Cpk would equal Cp.
| Quantity | Formula | Value |
|---|---|---|
| Cp | 1.0/(6×0.10) | 1.67 |
| Cpk | min(0.4/0.30, 0.6/0.30) | 1.33 |
| Pp | 1.0/(6×0.15) | 1.11 |
| Ppk | min(0.4/0.45, 0.6/0.45) | 0.89 |
Pp and Ppk use an overall standard deviation calculated across the sampled window rather than a within-subgroup estimate. With overall sigma=0.15, Pp=1.0/(6×0.15)=1.11. Ppk is the smaller of 0.4/0.45=0.89 and 0.6/0.45=1.33, so Ppk=0.89. These indices describe observed longer-term performance in the chosen window. Different software may define within variation differently, which is why every report must name the estimator rather than showing four unexplained labels.
Within variation aims to estimate short-term background variation from rational subgroups or a specified moving-range method. Overall variation includes shifts, drift and other between-time changes present in the sample. A large gap between Cpk and Ppk can indicate that the process does not maintain its short-term behaviour over the full period, but it can also reflect estimator choices or sampling. Plot the components and time order. Do not treat the gap as a diagnosis without investigating production and measurement records.
Capability predicts performance only if the process represented by the estimates is sufficiently stable for that purpose. A time series with a mid-period mean shift has no single persistent mean, even if a pooled histogram looks smooth. A control chart or another time-aware analysis should identify regimes. Stabilise the process or report separate periods with clear justification. NIST describes process stability as stationarity assessed through plots, control charts and time-series ideas. A single index cannot repair a mixed history.
Cp and Pp compare spread with the two-sided tolerance and ignore centring. Cpk and Ppk include centring by using the nearer limit. Cp/Cpk use a defined within variation estimate; Pp/Ppk use overall variation. Read them together: Cp much greater than Cpk suggests off-centring, while Cpk much greater than Ppk suggests longer-term instability or drift. These patterns are prompts for investigation, not universal causal rules. Always pair them with plots, counts and the sampling window.
When only an upper limit matters, use CPU=(USL−mean)/(3 sigma) with the appropriate within or overall variation. When only a lower limit matters, use CPL=(mean−LSL)/(3 sigma). Do not create a fictional opposite limit or report the minimum of two sides when only one requirement exists. One-sided indices still rely on distribution, stability and measurement assumptions. If values below zero are physically impossible or the tail shape is skewed, a normal-distance index may not describe risk well.
The midpoint of specifications is not always the engineering target. A process may have an asymmetric loss around a preferred value even when limits are symmetric. Cpm-type indices include deviation from a target in the denominator, but the target must be documented and the formula stated. Reporting mean, target, midpoint and distances directly is often clearer. A high Cpk can coexist with a mean that is economically or clinically suboptimal if the limits are broad.
Cycle times, particle counts and concentrations can be skewed or bounded. Options include a justified transformation, fitting an appropriate distribution, using nonparametric percentile methods or reporting direct conformance estimates with uncertainty. Each method answers a slightly different question. Check whether the tail near the specification is represented. Do not select a distribution only because it produces a favourable index. Preserve the raw scale in communication even if modelling occurs on a transformed scale.
| Index | Variation estimate | Centring included |
|---|---|---|
| Cp | Within | No |
| Cpk | Within | Yes |
| Pp | Overall | No |
| Ppk | Overall | Yes |
A histogram with two peaks may combine machines, suppliers, recipes or shifts. A single mean and standard deviation can place the centre in a region where almost no item occurs. Stratify known sources and decide whether one overall capability statement has operational meaning. If customers receive a random mix, the mixture may matter; if each machine is controlled separately, machine-specific capability may guide improvement. Pooling is a business and process decision, not a default statistical convenience.
Serial dependence changes the effective information in a sample and can affect within and overall variation estimates. A slowly drifting process can produce many similar consecutive values that make a naive short-term estimate look small. Inspect lag plots and control charts, consider the sampling interval and use a time-series model when appropriate. Capability should describe the distribution relevant to delivered output, so modelling dependence must not erase genuine long-term variation customers experience.
Observed variation combines product variation and measurement variation. Bias shifts the apparent mean toward a specification, while poor repeatability widens the observed spread. If measurement variation is material, capability of the recorded values is not the same as capability of the product. Conduct measurement system analysis, maintain calibration and consider decision rules near limits. Report units and resolution. An index with three decimals is not credible when the gauge records only coarse increments.
Capability data should represent the process, products and time horizon named in the claim. A convenient hour after machine setup cannot support a year-long performance claim. Sample across relevant shifts, materials, tools and environmental conditions or state the narrower scope. Avoid only sampling accepted output after screening, which truncates the distribution. Record exclusions, rework and missing values. If destructive testing selects special units, explain how selection relates to delivered production.
Cp, Cpk, Pp and Ppk are estimates. Small samples can make them unstable, especially when the true process lies near a decision threshold. Confidence intervals or bootstrap intervals convey uncertainty under stated assumptions. The distribution of Cpk estimators is more complicated than a standard-error formula for the mean. Software can help, but the method and assumptions should be documented. A lower confidence bound may be more relevant for qualification than a point estimate, though policy determines the decision rule.
Rules such as 1.00, 1.33 or 1.67 are conventions used in some industries and contracts, not universal laws of mathematics. Required capability depends on consequence, inspection, process control, supplier agreement and uncertainty. Do not label a process “world class” from one threshold. State who set the criterion and why. High-stakes characteristics may require direct tail-risk evidence, engineered controls and stricter requirements beyond a capability index.
Under a stable normal model with known parameters, specification tail areas can estimate nonconforming probability. This is model-based, not a guaranteed defect count. Extrapolating tiny tail probabilities far beyond the observed data is highly sensitive to mean, spread and distribution assumptions. Compare predicted and observed nonconformance where sample size permits, and show confidence limits. If inspection removes nonconforming output, distinguish process yield before screening from outgoing quality after screening.
Capability language can illuminate a repeatable service process, such as whether a scanning or printing workflow keeps page alignment within a functional tolerance. It should not be used to classify students as capable or incapable. Human learning has changing goals, context and development, and a score distribution is not a production process with engineering specifications. Use low-stakes technical processes to teach the mathematics, then discuss why ethical domain boundaries matter.
Suppose a shaft diameter must lie between 9.5 and 10.5 mm. A stable centred process with sigma 0.10 has Cp=Cpk=1.67. If its mean moves to 10.1, Cp stays 1.67 while Cpk falls to 1.33. Reducing sigma improves both; recentring improves Cpk but not Cp. This decomposition helps engineers choose between adjustment and variation reduction. Confirmation data after the change must show that the improvement persists.
A service may require completion below a maximum time rather than between symmetric limits. Use a one-sided model and recognise the right-skewed distribution. Reporting a normal Cpk from raw times can misrepresent the long tail. A lognormal or empirical percentile analysis may be more interpretable. Also distinguish a service target from a hard maximum and account for censored cases that were still open at the reporting cut-off.
A system can have a latency service-level objective such as 95 percent below a threshold. A Cp-style normal index is often less natural than direct percentile monitoring because latency is skewed and traffic mix changes. Still, the capability framework teaches valuable questions: is the system stable, which distribution describes the tail, what is the requirement, and how much measurement or sampling error exists? Use the metric the service promise actually defines.
If Cp is low, reduce spread through process redesign, better controls, materials, maintenance or robust settings. If Cp is adequate but Cpk is low, investigate centring and drift. If Cpk exceeds Ppk, examine between-time changes. Improvements should target mechanisms, not the index formula. Recalculate only after collecting a new representative period and document changed conditions. Never delete legitimate tail observations simply to make the index rise.
NIST defines Cp as tolerance width divided by six times a standard-deviation estimate and Cpk as the smaller standardised distance from the mean to the two specification limits. It also emphasises that capability reflects how observed variation fits specifications. Read the NIST process-capability section and capability assessment guidance. These formulas support the worked example; the sampling and specification evidence determine whether using them is defensible.
Reproduce Cp 1.67, Cpk 1.33, Pp 1.11 and Ppk 0.89 from the stated example. Check units and both distances to limits. Plot the sequence, histogram, normal probability plot and subgroup variation. Recalculate after centring the mean at 10.0 and after changing sigma. Test a one-sided case and a skewed dataset. Compare results across two software tools using the same estimator definitions. Every discrepancy should be traced to a formula, denominator or data-filter choice.
Capability indices cannot establish stability, prove normality, verify specifications, identify a cause, guarantee future conformance or compensate for measurement error. They compress complex behaviour and can hide mixtures and time changes. Very high estimated values from small samples are especially fragile. Tail predictions may rely on extrapolation. Keep the raw plots and process evidence beside the summary. If assumptions fail, use an alternative analysis or make a narrower claim.
Lead with the requirement, population, time window and stability evidence. Show mean, within and overall standard deviations, sample size, indices and intervals. State estimator definitions, distribution method and any transformation. Include observed nonconforming counts and relevant plots. Explain whether the limiting side is upper or lower and which improvement is proposed. A reader should be able to reproduce each index and understand why it matters.
Start with the four worked calculations. Draw specifications on a number line and mark mean plus or minus three within sigmas. Move the mean while holding spread fixed, then widen spread while holding the mean fixed. Compare Cp with Cpk and Pp with Ppk. Next inspect a shifted time series whose pooled histogram looks acceptable. Finish with a skewed one-sided example. This sequence builds arithmetic, graph sense, assumptions and decision judgement.
Capability analysis is valuable because it separates requirement, centring, short-term variation and long-term performance. It gives students a practical reason to care about standard deviation, distributions and time order. Used carefully, it directs improvement toward the process rather than toward cosmetic reporting. The optimistic goal is not a larger number by any means; it is a process whose evidence, measurement and requirements agree.
Use one row per time-ordered observation with columns for subgroup, source and acceptance result. Calculate subgroup statistics, within variation with a clearly labelled method, overall standard deviation, mean and distances to limits. Put LSL, USL and target in protected cells linked to a requirement record. Build time, histogram and probability plots before calculating indices. Add confidence intervals through validated formulas or a documented add-in. Test the sheet with the worked values and ensure blank cells are not converted to zero.
Keep raw data immutable and store specifications, target, estimator choice and analysis window in a configuration file. Write tested functions for Cp, CPU, CPL, Cpk, Pp and Ppk. Assert that USL exceeds LSL and that standard deviations are positive. Produce time-order diagnostics before indices. For non-normal analysis, compare fitted distributions and empirical quantiles without overwriting the original scale. Bootstrap by resampling at an appropriate unit when independence permits, and preserve seeds and package versions.
Capability methods appear in manufacturing, metrology, semiconductor processing, pharmaceuticals, food production, laboratories, supplier quality and some service operations. Practitioners combine statistics with engineering requirements, measurement science, safety, regulation and improvement methods. A capability index alone does not qualify someone or certify a product. Students can develop relevant skills through algebra, probability, data visualisation, careful units and transparent reporting, then explore engineering, statistics, operations and quality roles.
Measure a safe repeated process such as the width of paper strips cut with a fixed guide. Set a functional tolerance before collecting data. Take small rational subgroups over several sessions, recording tool, operator and paper lot. Chart stability, examine the measurement resolution, then calculate within and overall indices. Introduce a justified centring adjustment and collect new data. Compare the mechanism and indices without mixing the pre- and post-change regimes. Discuss whether the tolerance itself was evidence-based.
Cp is not the predicted percentage conforming. Cpk does not replace a control chart. Ppk is not automatically “worse” than Cpk; it uses a different variation estimate. A specification is not a confidence interval. A high index does not prove normality or guarantee zero defects. One decimal change in the mean can matter when it moves toward a limit. Widening specifications is not process improvement. Thresholds such as 1.33 are contextual conventions, not universal mathematical truths.
The statistical process control article establishes stability before capability. The robust parameter design article shows how to reduce sensitivity to noise. The response-surface methodology article explores settings and curvature. Capability then checks whether the improved process fits real requirements over a declared period.
Students should calculate the example by hand and then verify it in a spreadsheet. Parents can ask where specifications came from, whether the process was stable and whether the measurement system is adequate. Encourage learners to explain the four indices in sentences before comparing thresholds. Use technical examples rather than applying factory language to children. The most transferable habit is tracing each number to a source: requirement, mean, within variation, overall variation or model assumption.
1. Recalculate all four worked indices. 2. Centre the mean and show which indices change. 3. Increase within sigma and explain the effect. 4. Give a one-sided example. 5. Explain why a mean shift breaks one pooled capability statement. 6. Describe two non-normal methods. 7. Trace a specification to its owner. 8. Show how measurement bias can inflate or reduce Cpk. 9. Write a conclusion with uncertainty and no guarantee language.
### Which index includes centring? Cpk and Ppk use the nearer specification limit.
### Why can Cpk exceed Ppk? Within variation can be smaller than overall variation when the process drifts or shifts.
### Must capability data be normal? No, but the method must fit the distribution and be stated.
### Can a stable process have low capability? Yes. Stability and fit to specifications are separate questions.
Report characteristic and unit; specification source, version and target; population and dates; sampling and subgroup plan; measurement-system evidence; stability analysis; missing values, rework and exclusions; distribution and dependence checks; mean; within estimator; overall standard deviation; Cp, Cpk, CPU/CPL where relevant, Pp and Ppk; intervals; observed nonconformance; software; plots; limiting side; proposed mechanism-based improvement; and scope of the conclusion.
Trace every specification to a controlled document and confirm units. Recalculate the worked values independently. Compare the raw sequence with subgroup labels and verify no post-screened or reworked values disappeared. Reproduce within and overall standard deviations from raw data. Check that the mean and limiting side match the plot. Repeat the analysis by source and period to expose mixtures. Challenge the normal model near each specification. Finally, test whether the conclusion changes under a justified alternative estimator or distribution; report that sensitivity rather than selecting the most flattering result.
Continue through the Mathematics Learning Hub. Read the NIST discussion of process stability, then study measurement system analysis so that observed spread and bias are understood. For improvement design, revisit robust parameter design and response-surface methods. The next useful action is to reproduce one small capability study from time-ordered raw data, with the specification source attached.
Extended practice laboratory
Practice case 1: Calculate Cp
Setup: Use LSL 9.5, USL 10.5 and within sigma 0.10.
Task: Compute Cp=(1.0)/(0.60)=1.67.
Checking target: State that Cp describes potential spread fit and ignores centring.
Practice case 2: Calculate Cpk
Setup: Use mean 10.1 with the same specifications and sigma.
Task: Compute CPU=1.33, CPL=2.00 and Cpk=1.33.
Checking target: Identify the upper specification as the limiting side.
Practice case 3: Calculate Pp and Ppk
Setup: Replace within sigma with overall sigma 0.15.
Task: Compute Pp=1.11 and Ppk=0.89.
Checking target: Explain why long-term drift can widen overall variation.
Practice case 4: Shift the mean
Setup: Move the mean from 10.1 to 10.0 without changing sigma.
Task: Recalculate Cpk and compare it with Cp.
Checking target: Centring improves Cpk but does not reduce spread.
Practice case 5: Widen the spread
Setup: Keep mean 10.0 and change sigma from 0.10 to 0.14.
Task: Recalculate Cp and Cpk.
Checking target: Name a process improvement that targets variation rather than arithmetic.
Practice case 6: Check stability first
Setup: Draw a time series with a late mean shift.
Task: Explain why one pooled capability number hides two regimes.
Checking target: Stratify or stabilise before using the index for prediction.
Practice case 7: Use a one-sided limit
Setup: Suppose only a maximum contaminant level matters.
Task: Calculate the appropriate upper capability index.
Checking target: Do not invent a lower specification for symmetry.
Practice case 8: Check non-normal data
Setup: Use a right-skewed cycle-time distribution.
Task: Compare transformation, distributional modelling and empirical quantiles.
Checking target: Report the method rather than applying normal six-sigma formulas silently.
Practice case 9: Audit specification ownership
Setup: Trace USL and LSL to an engineering or customer requirement.
Task: Separate specifications from observed process limits.
Checking target: Never widen specifications merely to improve an index.
Practice case 10: Report the analysis
Setup: State mean, within and overall variation, indices and intervals.
Task: Include stability evidence, distribution checks and sample window.
Checking target: Avoid presenting a point estimate as a guarantee of conforming output.