Process Capability Explained: Cp, Cpk, Pp and Ppk

Every manufacturing team eventually runs into the same argument. Someone looks at a batch of parts, sees that most of them pass, and declares the process fine. Someone else pulls the measurement data, runs the numbers, and says the process is one bad shift away from scrapping a whole run. Both people are looking at the same parts. The difference is that one of them is measuring capability and the other is measuring luck. Process capability indices exist to settle that argument with math instead of opinion.

The four indices Cp, Cpk, Pp and Ppk get grouped together because they answer closely related questions, but they are not interchangeable. Reporting the wrong one, or reporting all four without knowing what separates them, is one of the most common ways a capability study loses credibility. This guide walks through each index, the formulas behind them, a full worked example with real numbers, and the mistakes that quietly ruin capability reports. By the end you will know which number to trust and when.

What process capability actually measures

Process capability compares the natural variation of a process to the tolerance you are allowed. The tolerance is set by the customer, the drawing, or the specification. The variation is set by the process itself: the machine, the material, the operator, the environment. Capability is the ratio between what you are permitted and what you actually produce.

Two things can go wrong. The spread of your output can be too wide to fit inside the tolerance, or the spread can be narrow enough but sitting off to one side so that one edge crosses a limit. Those are two different failures with two different fixes. Widening a machine’s precision is an equipment problem. Re-centering an off-target process is usually an adjustment problem. Capability indices were designed so that a single report can tell you which of the two you are facing, and that is why you need more than one number.

The specification limits matter here. The upper specification limit, or USL, is the largest value a part can have and still pass. The lower specification limit, or LSL, is the smallest. The distance between them is the tolerance width. Capability studies only make sense when both the specification and a stable, measured process exist. Without a real specification you have nothing to compare against, and the indices become meaningless decimals.

The formulas for Cp and Cpk

Cp is the spread-only index. It compares the total tolerance width to six times the short-term standard deviation:

Cp = (USL – LSL) / (6 x sigma)

The six comes from the fact that a normal distribution places about 99.73 percent of its output inside plus or minus three standard deviations, so six sigma covers the practical full width of the process. If Cp equals 1.0, the process spread exactly fills the tolerance with no room to spare. If Cp is 2.0, the spread fits twice over. Notice that Cp says nothing about where the process is centered. A process can have an excellent Cp and still make scrap if it drifts to one side, because Cp only looks at width.

Cpk fixes that blind spot by measuring the distance from the process mean to the nearest specification limit, in units of three sigma:

Cpk = min[(USL – mean) / (3 x sigma), (mean – LSL) / (3 x sigma)]

Because Cpk takes the minimum of the two one-sided distances, it always reports the worse edge, the side that is closest to failing. If the process is perfectly centered, the two halves are equal and Cpk equals Cp. As the process drifts off target, Cpk falls below Cp. The gap between them is a direct measure of how far off-center you are. That single comparison, Cpk versus Cp, is one of the most useful readings in all of quality control.

The formulas for Pp and Ppk

Pp and Ppk use exactly the same arithmetic as Cp and Cpk. The only thing that changes is which standard deviation you feed into them. Cp and Cpk use the short-term, within-subgroup sigma, the variation you see inside a small batch collected close together in time. Pp and Ppk use the overall, long-term sigma calculated from the entire data set, which includes the slow drift, the tool wear, the shift changes, and the material lot differences that accumulate over hours or days.

Pp = (USL – LSL) / (6 x sigma_overall)

Ppk = min[(USL – mean) / (3 x sigma_overall), (mean – LSL) / (3 x sigma_overall)]

Because the overall sigma is almost always larger than the within-subgroup sigma, Pp and Ppk are usually a little lower than Cp and Cpk. That difference is not a rounding error. It is information. When Ppk is much smaller than Cpk, the process is capable in the short run but unstable over time, and the extra variation is coming from between the subgroups rather than within them. A large gap points you at drift and special causes rather than at the machine’s inherent precision.

How to calculate capability step by step

The mechanics are the same every time, and following them in order keeps you from the classic errors.

First, confirm the process is stable enough to study. Plot the data on a control chart and check that it is in statistical control. Capability numbers calculated on an out-of-control process describe nothing you can repeat, because the process has no single distribution to summarize.

Second, gather the specification. Write down the USL and the LSL from the drawing or the customer document, and compute the tolerance width as USL minus LSL. Do not guess these values, and do not confuse control limits with specification limits. Control limits come from the process. Specification limits come from the requirement. They are not the same thing.

Third, estimate both sigmas. Calculate the within-subgroup sigma for Cp and Cpk, usually from the average subgroup range or the pooled standard deviation. Calculate the overall sigma for Pp and Ppk from every individual measurement in the study.

Fourth, compute the process mean from all the data. Fifth, plug the mean, the two sigmas, and the two specification limits into the four formulas above. Sixth, and this is the step people skip, compare the indices to each other, not just to a target. The comparison between Cp and Cpk, and between Cpk and Ppk, is where the diagnosis lives.

A full worked example

Numbers make this concrete. Suppose a machined feature has an upper specification limit of 10.5 and a lower specification limit of 9.5, so the tolerance width is exactly 1.0. The process has been running and you measure a mean of 10.1 with a sigma of 0.10.

Start with Cp, the spread-only index:

Cp = 1.0 / (6 x 0.10) = 1.0 / 0.60 = 1.667

A Cp of 1.667 is healthy. The tolerance is well wider than the process spread, so on width alone this process has margin to spare. If you stopped here you would call the process capable and move on. That would be a mistake, because the mean of 10.1 is not sitting in the middle of the tolerance, which runs from 9.5 to 10.5 and is centered on 10.0. The process is running high.

Now compute Cpk, which will expose the off-center condition. Work out both one-sided distances:

Upper side: (10.5 – 10.1) / (3 x 0.10) = 0.4 / 0.30 = 1.333

Lower side: (10.1 – 9.5) / (3 x 0.10) = 0.6 / 0.30 = 2.000

Cpk = min[1.333, 2.000] = 1.333

The upper side is tighter because the process is running toward the top of the tolerance, so the mean sits closer to the USL. Cpk reports that worse edge, 1.333, well below the Cp of 1.667. The gap between them, about 0.33, is the direct cost of being off-center. The process can physically fit the tolerance, but where it is actually centered has thrown away roughly a fifth of its margin.

Here is the encouraging part. The fix does not require a better machine. If you re-center the mean from 10.1 to 10.0, the two one-sided distances become equal, and Cpk climbs to 1.667, matching Cp exactly. You recover the entire lost margin with an adjustment, not a capital purchase. That is the practical payoff of reporting Cpk alongside Cp: it tells you the problem is centering, and centering is usually cheap to fix.

At a Cpk of 1.33 the expected defect rate is roughly 30 to 60 parts per million, depending on whether you account for the usual long-term sigma shift. That is why 1.33 became a widespread target: it corresponds to about four sigma of margin and keeps defects in the tens of parts per million rather than the thousands. The lesson from the whole example is compact. Cp tells you whether the tolerance can fit the spread. Cpk tells you whether it actually does, given where the process happens to be centered.

How to read and apply the result

Once you have the four numbers, interpret them as a set rather than one at a time. A Cpk at or above 1.33 is the usual line for a capable, centered process on a controlled line. Between 1.0 and 1.33 the process is producing acceptable parts but has thin margin, and any drift will start generating scrap. Below 1.0 the process is not capable, meaning the spread crosses at least one specification limit and defects are being produced as a matter of routine.

The comparisons carry the diagnosis. When Cpk is close to Cp, the process is well centered and your attention should go to reducing spread if you want more margin. When Cpk is meaningfully below Cp, re-center before you touch anything else, because centering is the cheaper and faster gain. When Ppk sits far below Cpk, the process is drifting between subgroups and you have a stability problem, so the fix is to hunt down special causes with a control chart rather than to chase the tolerance. Reading the indices this way turns a capability report from a pass or fail stamp into an action list.

Common mistakes that ruin capability studies

The first and most damaging mistake is calculating capability on an unstable process. If the control chart shows out-of-control points, the process has no single distribution, and any Cp or Cpk you compute is a fiction that will not repeat next week. Prove stability first, then measure capability.

The second is confusing specification limits with control limits. Control limits describe what the process does. Specification limits describe what the customer requires. Feeding control limits into a capability formula produces a number that looks reasonable and means nothing.

The third is ignoring the normality assumption. The standard formulas assume the data is roughly normal. Strongly skewed or bounded data, such as flatness, roundness, or anything that cannot go below zero, can make the indices badly misleading. For those cases you need a transformation or a non-normal capability method.

The fourth is reporting Cpk while quietly using the overall sigma, or reporting Ppk while using the within-subgroup sigma. The label has to match the sigma. Mixing them defeats the entire reason both pairs exist. The fifth is trusting capability from too little data. A handful of parts gives an unstable estimate of sigma, and small samples tend to make a process look better than it is.

When capability indices do not apply

Capability indices assume a continuous measurement, a two-sided or clearly one-sided specification, and a stable, roughly normal process. Several common situations break those assumptions, and forcing an index onto them gives a confident wrong answer.

Attribute data, where parts are simply good or bad rather than measured on a scale, does not have a sigma in the usual sense, so Cp and Cpk do not apply. There you work with defect rates, DPMO, and sigma level instead. One-sided characteristics, such as a minimum strength with no upper limit, only have a meaningful Cpk on the side that has a specification. Highly skewed processes need non-normal capability analysis. And any process that has not been shown to be in control should be stabilized before capability is even attempted, because capability is a property of a stable process and nothing else.

Three expert tips

Always report Cp and Cpk together, never Cpk alone

A lone Cpk hides whether the loss of margin comes from spread or from centering. Reporting both lets any reader see the gap between them and diagnose the process in one glance. If you can only publish one number, publish Cpk, but the pair is far more useful and costs nothing extra to compute.

Watch the Ppk to Cpk gap as your stability signal

Short-term capability can look excellent while the process quietly drifts across shifts. Track Ppk against Cpk over time. When the gap widens, the process is becoming unstable between subgroups, and that is your early warning to investigate before the drift turns into scrap. A steady, small gap is a sign of a well-controlled line.

Fix centering before you spend money on precision

When Cpk trails Cp, the worked example shows the gain hiding in plain sight. Re-centering the mean is usually a setup adjustment, not a capital project, and it can recover the entire lost margin in a single change. Reach for tighter tolerancing machines only after you have confirmed the process is already centered and still short.

Free quality control calculators

You do not have to run these formulas by hand. The Process Capability Calculator (Cp, Cpk, Pp, Ppk) computes all four indices from your specification limits and data, including both the short-term and overall sigma so you can see the stability gap directly. For the surrounding work, the rest of the quality control suite covers each neighboring task:

• Control Chart Calculator (X-bar R, X-bar S, I-MR) to confirm the process is stable before you run a capability study.
• Acceptance Sampling Calculator (AQL, ANSI Z1.4) for lot-by-lot inspection plans.
• Gage R&R Calculator (MSA, AIAG Method) to check that your measurement system is trustworthy enough to judge capability.
• Process Yield Calculator (FPY, RTY, DPU) for first pass yield and rolled throughput yield.
• Sigma Level & DPMO Calculator to translate defect rates into sigma level for attribute data.

All of these live in the Quality Control hub, which collects the full set of tools with short guides for each.

Frequently asked questions

What is the difference between Cp and Cpk?

Cp measures only the process spread against the tolerance width and ignores where the process is centered. Cpk measures the distance from the mean to the nearest specification limit, so it also captures centering. When the process is perfectly centered, Cp and Cpk are equal. When it drifts off target, Cpk falls below Cp, and the gap tells you how far off-center you are.

What is the difference between Cpk and Ppk?

They use identical formulas but different standard deviations. Cpk uses the short-term, within-subgroup sigma and describes the process at its best. Ppk uses the overall, long-term sigma from the whole data set, so it includes drift and shift. Cpk is short-term capability; Ppk is long-term performance.

What is a good Cpk value?

A widely used target is Cpk of 1.33 or higher, which corresponds to about four sigma of margin and a defect rate in the tens of parts per million. Some industries require 1.67 or 2.0 for critical characteristics. Below 1.0 the process is not capable and produces defects as a matter of routine.

Why is my Cpk lower than my Cp?

Because the process is off-center. Cp only sees the spread, so it stays high as long as the spread fits. Cpk measures the nearest edge, so it drops as the mean moves toward one specification limit. A Cpk below Cp is the signal to re-center the process, which usually recovers the lost margin without any equipment change.

Can Cpk be higher than Cp?

No. Cpk is always less than or equal to Cp. The two are equal only when the process is perfectly centered between the specification limits. Any off-center condition can only pull Cpk down, never push it above Cp. If a report shows Cpk above Cp, there is a calculation error.

Which sigma do I use for capability?

Use the within-subgroup, short-term sigma for Cp and Cpk, usually derived from the average range or pooled standard deviation. Use the overall, long-term sigma calculated from every individual measurement for Pp and Ppk. The label you report must match the sigma you used, or the numbers mislead.

Do I need the process to be stable first?

Yes. Capability describes a stable process with a single, repeatable distribution. If the control chart shows out-of-control points, the process has no fixed distribution to summarize, and any Cp or Cpk you compute will not repeat. Stabilize the process on a control chart before running a capability study.

What does a large gap between Cpk and Ppk mean?

It means the process is capable in the short run but unstable over time. The extra variation is coming from between subgroups, from drift, tool wear, shift changes, or material lots, rather than from the machine’s inherent precision. A large gap points you toward finding and removing special causes rather than tightening the tolerance.

How much data do I need for a reliable study?

Enough to estimate sigma stably and cover the sources of long-term variation. Many practitioners collect at least 25 subgroups, or on the order of 100 or more individual measurements, spread across shifts and lots. Too few parts gives an unstable sigma estimate and usually makes the process look better than it truly is.

What if my data is not normal?

The standard formulas assume roughly normal data. For strongly skewed or bounded characteristics, such as flatness or anything that cannot go below zero, the ordinary indices mislead. Use a data transformation or a non-normal capability method that models the actual distribution, and state which method you used in the report.

Can I use capability indices for pass or fail attribute data?

No. Cp, Cpk, Pp and Ppk need a continuous measurement and a standard deviation. Attribute data, where parts are simply good or bad, has no sigma in that sense. For attribute processes, work with defect rates, DPMO, and sigma level instead, which the Sigma Level and DPMO calculator handles.

Process capability is not a single verdict but a short conversation between four related numbers. Cp asks whether the tolerance can hold the spread. Cpk asks whether it does given the centering. Pp and Ppk ask both questions again over the long run, where drift shows up. Read them together, match each index to the right sigma, and prove stability before you start, and the four numbers will tell you exactly what to fix and roughly what it will cost. That is the whole point of a capability study: to trade an argument about luck for a decision backed by data.