Two processes can have exactly the same variation and therefore the same Cp. Move one process mean toward the upper specification limit, however, and its Cpk falls. That one comparison explains why Cp is called potential capability and Cpk reflects the currently limiting side.
The indices compress a dataset into useful ratios. They do not prove that the process will remain stable, that the data are normally distributed, that the measurement system is adequate, or that a universal threshold applies.
Capability belongs at the end of a decision sequence
Capability compares estimated process behavior with engineering specifications. That comparison becomes credible only when the measurements are trustworthy and the process behavior is sufficiently stable for one mean and one variation estimate to represent the period being claimed.
Reversing the order creates confident-looking but fragile conclusions. A high Cpk calculated from unstable data is a description of a mixed historical dataset, not reliable evidence of predictable future capability.
The formulas separate spread from centering
Cp: potential capability
Cp compares the specification width with the estimated natural process spread. The mean does not appear in the formula, so Cp does not change when the distribution shifts but its standard deviation stays the same.
Cpk: capability from the current mean
Cpk uses the weaker side. As the mean approaches one limit, that side’s distance shrinks and Cpk falls—even when the process spread is unchanged.
For a perfectly centered two-sided process, CPU and CPL are equal, so Cp and Cpk are equal. In real data they may differ slightly because the sample mean is not exactly at the specification midpoint. A large gap indicates a meaningful centering loss relative to the available tolerance.
Worked example: same spread, different location
Both prepared datasets contain 50 values with sample standard deviation 0.03990 mm. The specification is 9.800 to 10.200 mm. Dataset A is centered at 10.000 mm. Dataset B adds 0.035 mm to every value, preserving the spread while moving the mean toward the USL.
A · Centered process
Cp = CpkThe mean is at the specification midpoint, so the upper and lower one-sided indices are equal.
B · Shifted process
Cp > CpkThe spread is identical, but the upper side now limits capability. CPU is 1.38 while CPL is 1.96.
The shift did not make the process more variable. That is why Cp remains 1.67. It consumed upper-side margin, so CPU and Cpk fell. This is the practical value of reading both indices: Cp shows the opportunity available through centering, while Cpk shows the performance from the current mean.
Dataset B · Cp substitution
The result is identical to Dataset A because the specification width and sample standard deviation are identical.
Dataset B · Cpk substitution
CPL = (10.035 − 9.800) ÷ (3 × 0.03990) = 1.96
Cpk = min(1.38, 1.96) = 1.38
The upper distance is smaller, so CPU becomes the limiting one-sided index.
The midpoint of the specifications is not automatically the preferred process target. Functional risk may be asymmetric, tool wear may move the process predictably, and a control plan may intentionally aim away from the geometric midpoint. The Cp–Cpk gap identifies lost margin; engineering determines whether recentering is appropriate and where the safe target should be.
Read the capability graphs before accepting the summary
The panel below uses Dataset B and the same graph family implemented in the Mechatrovich Process Capability Analysis Toolkit: a fitted histogram against the specification limits, a measurement-order plot, and a boxplot. Hover over chart marks for values.
Process Capability Report — Diameter
50 measurements · LSL 9.800 mm · USL 10.200 mm · sample standard deviation
Distribution / Histogram + Fitted Normal Curve
Measurement Order
Distribution / Boxplot
The histogram explains the Cp–Cpk gap, but the measurement-order plot is essential: a distribution can look smooth after time order is removed while still containing a shift, trend, cycle, or mixture.
The histogram shows a compact distribution positioned above the specification midpoint. The trend plot shows whether that position is consistent or created by a time-dependent change. The boxplot highlights median, quartiles, and possible outliers, but does not test normality or statistical control by itself.
Read the graphs as complementary evidence. If the trend reveals a late shift, the combined histogram and its fitted curve no longer represent one stable operating state. If the boxplot flags an extreme point, investigate the source before choosing whether it belongs to the process population. If the histogram shape is strongly skewed, the normal model—not the data—may be the problem.
What common Cp and Cpk patterns mean
| Pattern | Likely interpretation | First engineering question |
|---|---|---|
| Cp ≈ Cpk and both high | The spread is small relative to tolerance and the mean is near the midpoint. | Is the process stable and is the sigma method appropriate? |
| Cp high, Cpk lower | Potential spread fits, but centering consumes margin on one side. | Is the shift intentional, safe, and stable—or caused by drift, setup, or wear? |
| Cp and Cpk both low | Variation is too large for the tolerance; centering may also contribute. | Which sources dominate the variation after stratification? |
| Cpk negative | The sample mean is beyond at least one specification limit. | Is the specification, unit, data column, and process condition correct? |
| Cpk greater than Cp | For a standard two-sided calculation using the same mean and sigma, this should not occur. | Were different sigma estimates, limits, formulas, or rounding used? |
A good histogram cannot prove stability
Capability calculations ignore sequence unless the analyst preserves and reviews it. If the first half of a dataset runs at one mean and the second half at another, the combined histogram may look broad and approximately bell-shaped. A single Cpk then mixes two operating states.
Check an appropriate control chart before the final capability statement. For rational subgroups, that may be an X̄–R or X̄–S chart. For individual observations, a different chart is required. The chart choice should follow the data structure, not the chart available by default.
Control limits are not specification limits. Control limits are estimated from process behavior and address statistical signals. Specification limits come from design or customer requirements and define acceptable output. A process can be stable but incapable, or capable-looking in the sample but unstable.
Normality, sigma method, and sample size change the claim
Normal capability
The familiar Cp and Cpk interpretation assumes an approximately normal population and a defensible process standard-deviation estimate. Review the histogram, probability behavior, process physics, and any approved test—not only a normality-test p-value.
Non-normal or mixed data
Strong skew, bounds, multiple streams, censoring, or mixtures can make normal-tail interpretation misleading. Stratify real sources first, then use an approved transformation, distribution model, or percentile method when appropriate.
Within versus overall standard deviation
Software may estimate short-term within-subgroup variation for Cp/Cpk and overall variation for Pp/Ppk. The labels and methods vary across procedures, but the core warning is universal: values produced with different sigma estimates answer different questions. Record the method beside the index and do not compare two unlabeled “Cpk” numbers from different tools.
One-sided specifications
When only one specification limit is meaningful, Cp is not a complete two-sided comparison. Report the relevant CPU or CPL—or the approved one-sided capability measure—rather than inventing a second limit. The process target and risk direction still matter, especially for characteristics such as maximum force, minimum strength, leakage, or contamination.
Sample size and uncertainty
An index estimated from a small sample can move substantially when a few observations change. More data help only when they represent the period and conditions in the claim. Fifty consecutive measurements from one short setup are not evidence for machines, cavities, lots, operators, or material batches that were never sampled.
A practical capability review
Verify the characteristic, unit, target, LSL, USL, drawing revision, and whether the specification is one- or two-sided.
Review resolution, calibration, relevant MSA evidence, rounding, data collection, and any transformations.
Keep time, machine, cavity, tool, operator, lot, material, and other known stratification fields.
Use an appropriate chart and investigate special causes before treating one mean and sigma as representative.
Look for skew, mixtures, outliers, bounds, and process mechanisms that affect the selected model.
Name the sigma estimate, index family, sample period, exclusions, subgrouping, and software or worksheet method.
State what period and process stream the evidence covers, the limiting side, remaining risks, and the next monitoring action.
Common mistakes that make Cpk look more certain than it is
- Calculating capability before checking the measurement system and process stability.
- Combining machines, cavities, tools, product families, or shifts without stratification.
- Sorting data before reviewing the measurement sequence.
- Removing outliers only because they reduce the index.
- Using specification limits as control limits or vice versa.
- Assuming that “no rejected sample” means the process is capable.
- Reporting Cp/Cpk without naming the sigma method or normality assumption.
- Creating a missing specification limit for a one-sided characteristic.
- Treating 1.33, 1.67, or another threshold as universal rather than requirement-specific.
What Cp and Cpk do not tell you
Cp and Cpk do not identify the physical source of variation. A low value may reflect tool wear, material variation, temperature, setup, machine-to-machine differences, cavity effects, measurement noise, or a mixture of several streams. Stratification and process knowledge are needed to turn the ratio into a corrective action.
They also do not state how many nonconforming pieces were actually observed. Capability indices describe an estimated distribution relative to limits; observed fallout is a separate count. Converting Cpk to an expected defect rate adds assumptions about distribution, stability, independence, and the sigma estimate. Those assumptions must be explicit before a projected tail rate is presented.
Finally, capability is not the same as product acceptance. Inspection determines whether measured pieces meet the specification. Capability describes the behavior of a process population under stated assumptions and conditions. Both can be useful, but neither should silently replace the other.
Continue from capability to process evidence
Use the free calculator for a quick normal-capability review, or continue to the X̄–R guide when the immediate question is process stability. The complete toolkit adds CSV import, cleanup, multiple dimensions, grouping, product graphs, and reporting.
For a full offline workflow with histogram, trend, boxplot, specification review, multiple dimensions, and report export, continue to the Mechatrovich Process Capability Analysis Toolkit.
References and method sources
- NIST/SEMATECH e-Handbook, What is Process Capability? — stable-process comparison, normality assumption, sample-size context, and capability-index definitions.
- NIST, Assessing Process Capability — Cp calculation and graphical interpretation.
- NIST Dataplot Reference, CPK — formal Cpk definition.
- AIAG & VDA SPC Manual, 1st Edition (2026) — current automotive-industry framework for SPC, process performance, and capability work.