To learn how to calculate process capability, collect at least 50 measurements from a stable process, estimate its standard deviation, and compare the specification width to the process width. Cp equals (USL – LSL) / 6σ and shows whether the process spread fits inside the tolerance. Cpk equals the smaller of (USL – mean) / 3σ and (mean – LSL) / 3σ, which adds centering. Pp and Ppk use the same formulas with the overall standard deviation instead of the within-subgroup one. For most customers, a Cpk of 1.33 or higher is the pass mark.
The formulas take one line each. The trouble starts with the inputs. Spec limits get confused with control limits, unstable data goes into the calculation, and two analysts with the same measurements report different answers because one used within-subgroup sigma and the other used overall sigma. A Cpk that looks fine on paper can then hide a process that quietly ships defects.
This guide walks through the calculation in seven steps, using one worked example from start to finish: a shaft diameter with a tolerance of 10.00 mm ± 0.05 mm. You will set the limits, check stability, estimate sigma, calculate all four indices by hand and in Excel, and read the results. Lean Six Sigma Experts has worked with manufacturers and other organizations since 2011, and process capability is a standard topic in our Six Sigma training.
What the four indices mean and what you need first
Cp and Cpk describe what a stable process can do, using short-term variation inside subgroups. Pp and Ppk describe what the process actually did over the whole study, using all the variation in the data. Knowing which pair you need, and what must be true before you calculate, saves you from a result that looks precise and means nothing.
Capability versus performance
The four indices answer different questions. Two of them look only at spread, and two also look at how well the process is centered between the limits.
| Index | What it measures | Sigma it uses | Typical use |
|---|---|---|---|
| Cp | Spread versus tolerance, ignores centering | Within-subgroup (short-term) | Potential of a stable process |
| Cpk | Spread and centering | Within-subgroup (short-term) | Ongoing capability of a stable process |
| Pp | Spread versus tolerance, ignores centering | Overall (long-term) | Actual spread across the whole study |
| Ppk | Spread and centering | Overall (long-term) | New processes, pilot runs, short runs |
Use Cp and Cpk on a mature process that you have shown to be in statistical control. Use Pp and Ppk when the process is too new to judge stability, such as a pilot run or a short pre-production lot. On a stable process the two pairs land close together, and the gap between them is a diagnostic you will use in Step 5.
Cp and Cpk tell you what the process can do when it behaves, while Pp and Ppk tell you what it did.
What you need before you start
Every capability index rests on a few assumptions. If one fails, the number is still computable, but it stops predicting anything.
- Specification limits. An upper limit (USL), a lower limit (LSL), or one of them, taken from the drawing or the customer.
- Continuous data. Lengths, weights, torque, cycle times. Pass/fail counts need defect-rate methods such as DPMO instead.
- A trusted measurement system. Run a measurement system analysis or gauge R&R first. Gauge noise gets counted as process variation, so a poor gauge makes your process look worse than it is.
- A stable process. No special-cause signals on a control chart.
- Roughly normal data. The standard formulas assume a normal distribution. Step 2 covers what to do if yours is not.
- Enough data. About 50 independent values is the usual floor for a reasonable estimate, and 100 or more is better for a real capability study.
When all six are in place, the calculation itself takes minutes. When one is missing, fix it first.
Step 1. Set spec limits and collect data
Write down the USL and LSL from the drawing or customer requirement, then collect 100 or more measurements in time order. Limits come from outside the process, and the data has to cover enough time to show how the process really behaves. Everything else in the calculation depends on these two inputs.
Get the specification limits right
Specification limits are the voice of the customer. They come from the drawing, the contract or the product standard, never from your own data. In the example, the nominal diameter is 10.00 mm with a tolerance of ±0.05 mm, so LSL = 9.95 mm and USL = 10.05 mm. The tolerance width is 0.10 mm.
Control limits are a different thing. They are calculated from the process and say what the process does, not what is allowed. Mixing up spec limits and control limits is the most common error in capability studies, and it produces indices that look healthy for the wrong reason. If your characteristic has only one limit, such as a minimum fill weight or a maximum flatness, you will use a one-sided index, covered in Step 4.
Collect enough data in the right way
You can collect data in two ways, and the choice affects how you estimate sigma in Step 3.

- Subgrouped data. Take 4 or 5 consecutive parts at each sampling time. A common design is 25 subgroups, so 25 × 5 = 125 measurements. This is the example we will use.
- Individual data. Take one reading per sampling time, which suits slow or expensive measurements. You will chart it with an individuals and moving range (I-MR) chart.
Spread the sampling over a long enough window to include normal sources of variation, such as shift changes, operators, material lots and tool wear. A study taken in one calm hour flatters the process and inflates Ppk. Record every value in time order, because the stability check in Step 2 needs the sequence. A common rule is that the gauge should resolve to one tenth of the tolerance or finer, so for a 0.10 mm tolerance, measure to 0.001 mm if you can.
Your capability index can only be as honest as the spec limits and the sampling window behind it.
Step 2. Confirm the process is stable and normal
Before you calculate process capability, plot the data on a control chart and check the distribution shape. Capability indices only predict future output if the process is stable, and the standard formulas only convert to defect rates if the data is close to normal. Skipping this step is how people publish a Cpk of 1.5 for a process that is drifting.
Check stability with a control chart
Plot subgrouped data on an X-bar and R chart, or individual data on an I-MR chart. Look for points outside the control limits, runs of eight or more points on one side of the center line, and steady trends. If you find a special cause, such as a tool change or a bad material lot, investigate it and remove it before you calculate. Then collect fresh data.

Do not use the capability indices themselves to judge stability. A Cpl of 1 and a Cpu of 2 say the process sits closer to the lower limit. They say nothing about whether it is stable. Only the chart can tell you that. If your process cannot be made stable, report Pp and Ppk and state clearly that they describe past performance only.
Check normality and choose a method
Build a histogram and a normal probability plot. Then run a normality test such as Anderson-Darling. A p-value below 0.05 suggests the data is not normal. Real data is often skewed when a characteristic is bounded at zero, as with flatness, runout or leak rate.
If the data is not normal, you have three options:
- Transform the data with a Box-Cox transformation, and apply the spec limits through the same transformation. This is the preferred fix when it works.
- Use percentile-based indices. Replace 6σ with the spread between the 0.135th and 99.865th percentiles, and replace the mean with the median. This mimics the ±3σ coverage of a normal curve.
- Fit the right distribution, such as Weibull or lognormal, in statistical software.
Our shaft diameter data passes both checks, so the standard formulas apply.
A stable, roughly normal process is the entry ticket, and without it the index is just arithmetic.
Step 3. Estimate within and overall sigma
Capability indices need two different standard deviations. Use within-subgroup sigma for Cp and Cpk, and the overall standard deviation of all the data for Pp and Ppk. Using the wrong one is the main reason two analysts disagree on the same data.
Within-subgroup sigma for Cp and Cpk
Within-subgroup sigma estimates only the short-term, common-cause variation. For subgrouped data, divide the average subgroup range by the constant d2, which depends on subgroup size:
σ_within = R-bar / d2
| Subgroup size n | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 |
|---|---|---|---|---|---|---|---|---|---|
| d2 | 1.128 | 1.693 | 2.059 | 2.326 | 2.534 | 2.704 | 2.847 | 2.970 | 3.078 |
In the example, each subgroup has 5 parts and the average range is R-bar = 0.0279 mm. So σ_within = 0.0279 / 2.326 = 0.0120 mm. For individual data, use the average moving range of two consecutive values with d2 = 1.128, so σ_within = MR-bar / 1.128. If your subgroup sizes vary, the average standard deviation divided by the constant c4 (S-bar / c4) handles that case.
Overall sigma for Pp and Ppk
Overall sigma is the ordinary sample standard deviation of every measurement, ignoring subgroups. It also captures shifts and drift between subgroups, so it is normally the larger of the two. In the example, the 125 values give s = 0.0140 mm.
The two numbers tell a story before you calculate anything. Here the overall sigma is 17% larger than the within sigma (0.0140 / 0.0120 = 1.17), which means some variation is coming from between subgroups, not just within them. Step 5 shows what that does to the indices. Keep both numbers on your worksheet, labeled, so nobody mixes them up later.
Within sigma goes into Cp and Cpk, and overall sigma goes into Pp and Ppk, with no exceptions.
Step 4. Calculate Cp and Cpk
Cp is the spec width divided by six within-sigma, and Cpk is the distance from the mean to the nearest limit divided by three within-sigma. Cp tells you whether the process could fit, and Cpk tells you whether it does fit given where the mean sits. Calculate both, because the gap between them shows how much capability you lose to off-center running.
Cp formula and what it ignores
Cp = (USL - LSL) / (6 × σ_within)
For the shaft: Cp = (10.05 – 9.95) / (6 × 0.0120) = 0.10 / 0.072 = 1.39. The 6σ in the denominator is the width the process naturally uses, which covers about 99.73% of output for a normal process. Cp above 1 means the tolerance is wider than the process spread. Cp says nothing about centering, so a process can post a great Cp while producing scrap at one limit.
Cpk formula, step by step
Calculate the upper and lower halves separately, then take the smaller.
Cpu = (USL - mean) / (3 × σ_within)
Cpl = (mean - LSL) / (3 × σ_within)
Cpk = min(Cpu, Cpl)
The 125 shaft readings have a mean of 10.012 mm, so:
- Cpu = (10.05 – 10.012) / (3 × 0.0120) = 0.038 / 0.036 = 1.06
- Cpl = (10.012 – 9.95) / (3 × 0.0120) = 0.062 / 0.036 = 1.72
- Cpk = min(1.06, 1.72) = 1.06
You can cross-check with the offset form, Cpk = Cp × (1 – k), where k = |midpoint – mean| / (tolerance width / 2). Here k = |10.00 – 10.012| / 0.05 = 0.24, and 1.39 × 0.76 = 1.06. Cpk can never exceed Cp. If yours does, you have a formula error. If the mean falls outside the spec limits, Cpk turns negative.
One-sided specifications
With only one limit, there is no Cp. Report the matching half and call it Cpk. A minimum fill weight with only an LSL gives Cpk = Cpl. A maximum flatness with only a USL gives Cpk = Cpu. For two-sided specs, the identity Cp = (Cpu + Cpl) / 2 also holds: (1.06 + 1.72) / 2 = 1.39.
Cp tells you what centering could buy you, and Cpk tells you where you actually stand.
Step 5. Calculate Pp and Ppk
Pp and Ppk use the same formulas as Cp and Cpk, with the overall standard deviation s in place of σ_within. Calculate them from the same data, then compare each pair. The comparison is where the real insight is.
Pp and Ppk formulas
Pp = (USL - LSL) / (6 × s)
Ppu = (USL - mean) / (3 × s)
Ppl = (mean - LSL) / (3 × s)
Ppk = min(Ppu, Ppl)
With s = 0.0140 mm and a mean of 10.012 mm:
- Pp = 0.10 / (6 × 0.0140) = 0.10 / 0.084 = 1.19
- Ppu = 0.038 / 0.042 = 0.90
- Ppl = 0.062 / 0.042 = 1.48
- Ppk = 0.90
Use Pp and Ppk as your headline numbers when the process is new or you cannot show stability, since they make no promise about the future. Report them as a description of what the study measured.
Compare the pairs to find the problem
Put the four results side by side.
| Pair | Short-term (within) | Long-term (overall) | Ratio |
|---|---|---|---|
| Spread only | Cp = 1.39 | Pp = 1.19 | 1.17 |
| Spread and centering | Cpk = 1.06 | Ppk = 0.90 | 1.17 |
Each gap points to a different fix. A Cp well above Pp means variation is entering between subgroups, from drift, setup changes, tool wear or lot changes. A Cp well above Cpk means the process is off-center, and moving the mean toward 10.00 mm is the cheapest improvement you can make. In a truly stable process, Cpk and Ppk come out nearly equal.
Ppk is usually less than or equal to Cpk. If Cpk comes out lower than Ppk, check your subgrouping and sample size. A small sample over a short time can make the R-bar estimate overstate the standard deviation.
The gap between Cpk and Ppk shows how much of your variation is drift, not noise.
Step 6. Interpret the index values
A Cpk of 1.33 or higher is the common minimum for an established process, and 1.67 or higher is often required for new or critical characteristics. Below 1.00, the process cannot hold the tolerance. Each value of Cpk maps to a distance from the mean to the nearest limit, measured in standard deviations, and that distance sets the defect rate.
Read the number against a threshold
Multiply Cpk by 3 to get the short-term sigma level. The table shows the expected output beyond the nearest limit for a normal, stable process.
| Cpk | Z to nearest limit | Expected out of spec beyond that limit | How to read it |
|---|---|---|---|
| 0.67 | 2 | about 2.3% | Not capable |
| 1.00 | 3 | about 1,350 ppm | Marginal, no room for drift |
| 1.33 | 4 | about 32 ppm | Capable, common minimum |
| 1.67 | 5 | about 0.3 ppm | Strong, often required for critical features |
| 2.00 | 6 | about 1 ppb | Six sigma, short-term |
For a perfectly centered process, both tails count. Cp of 1.00 then corresponds to 0.27% out of spec, Cp of 1.33 to about 64 ppm, Cp of 1.67 to about 0.6 ppm and Cp of 2.00 to about 2 ppb. Many automotive customers ask for a Ppk of at least 1.67 on an initial study and a Cpk of at least 1.33 on ongoing production, so check your own customer’s requirement. The 3.4 defects per million quoted for Six Sigma assumes a 1.5 sigma long-term shift on top of a short-term sigma level of 6.
Our shaft process has Cpk = 1.06, so about 760 ppm should exceed the upper limit in the short term. Ppk = 0.90 predicts roughly 3,300 ppm (0.33%) over the long run. That fails a 1.33 requirement, and the fix points to centering first.
Remember the uncertainty
Every index is an estimate from a sample. For normal data, an approximate 95% interval is Cpk ± 1.96 × √(1/(9n) + Cpk² / (2(n – 1))). With n = 125 and Cpk = 1.06, that is 1.06 ± 0.14, or about 0.91 to 1.20. The true Cpk could sit below 1.00. A Cpk of 1.33 from 30 parts proves much less than the same value from 300. Quote the interval when a decision depends on a threshold, and use statistical software for non-normal data.
Act on a low value
Pick the fix from the pattern, not from the number alone. If Cp is healthy but Cpk is low, recenter the process, usually with an offset or setup adjustment. If Cp itself is low, you must reduce variation, which is DMAIC Improve work: root cause analysis, designed experiments, mistake-proofing and control charts. Our Green Belt online training course runs about 28 hours and covers these statistical techniques with an emphasis on measuring. After any change, repeat the study with fresh data.

Read Cpk against 1.33, check the uncertainty, and let the Cp-to-Cpk gap tell you where to act.
Step 7. Calculate process capability in Excel
Excel has no built-in Cpk function, but five cells of formulas replace it. Put the subgroup data in a grid, compute each subgroup range, and let the summary cells do the rest. You can paste the layout below and adjust the cell references to your own sheet.
Set up the worksheet
Place 25 subgroups of 5 readings in A2:E26. In G2, enter the range formula and fill it down to G26. Then build the summary block in columns I and J.
G2 =MAX(A2:E2)-MIN(A2:E2) (fill down to G26)
I1 USL J1 10.05
I2 LSL J2 9.95
I3 Mean J3 =AVERAGE(A2:E26)
I4 Sigma within J4 =AVERAGE(G2:G26)/2.326
I5 Sigma overall J5 =STDEV.S(A2:E26)
I6 Cp J6 =(J1-J2)/(6*J4)
I7 Cpk J7 =MIN((J1-J3)/(3*J4),(J3-J2)/(3*J4))
I8 Pp J8 =(J1-J2)/(6*J5)
I9 Ppk J9 =MIN((J1-J3)/(3*J5),(J3-J2)/(3*J5))
The constant 2.326 is d2 for subgroups of 5. Swap in the value from the Step 3 table for your subgroup size. For individual data in column A, add a moving range column with =ABS(A3-A2) and set J4 to =AVERAGE(B3:B101)/1.128.
Estimate the defect rate and sanity-check
To estimate the total fraction out of spec from the overall spread, use the normal distribution function:
=NORM.DIST(J2,J3,J5,TRUE)+(1-NORM.DIST(J1,J3,J5,TRUE))
Swap J5 for J4 to get the short-term estimate. Run three quick checks before you trust the sheet. Cpk must be less than or equal to Cp. Ppk should rarely exceed Cpk by much. And the values should match your hand calculation from Steps 4 and 5, which for this example are Cp 1.39, Cpk 1.06, Pp 1.19 and Ppk 0.90. Excel does not test normality or draw confidence intervals for you, so for formal reports on non-normal data, use dedicated statistical software.
Five formulas in Excel reproduce the whole hand calculation, so check them against a worked example once.
Next steps after your capability study
Calculating process capability comes down to good inputs and a clear reading of four numbers. Set real spec limits, collect data that covers normal variation, confirm stability and normality, estimate within and overall sigma separately, then calculate Cp, Cpk, Pp and Ppk. Read Cpk against 1.33, let the gaps between the indices point to the fix, and keep the uncertainty in view when a decision hangs on a threshold.
If your study turns up a weak Cpk and you are not sure whether the cause is centering, drift or the gauge, a second pair of eyes shortens the search. Lean Six Sigma Experts offers a free 30-minute assessment with no obligation, and our consultants and trainers can help you run the study or build the skills in-house. Contact Lean Six Sigma Experts to talk through your data.
