Histograms with Specification Limits: 7 Patterns That Tell You What Cpk Cannot

A histogram with specification limits overlaid is the single fastest way to see whether a process is in trouble. In one glance you can tell whether the process is centered between the specs, whether the natural spread fits inside the tolerance band, and whether the data distribution looks anything like the normal curve that Cpk and Cp calculations assume. The numerical capability indices (Cpk, Cp, Pp, Ppk) are summaries of what the histogram already shows, but they can hide problems that a picture surfaces immediately.

This guide walks through how to interpret a histogram with specification limits like a practitioner—what the common shapes mean, what they imply about the process, and how they connect to the decisions downstream: is the process capable, is it stable, does Cpk mean what we think it means, and what needs to change to improve things.

What You’re Looking At

A histogram with specification limits has three critical reference marks:

• LSL (Lower Specification Limit) and USL (Upper Specification Limit)—vertical lines at the customer-dictated limits. Any data point outside these lines is, by definition, a nonconformity.
• Target value (if specified)—the ideal point, usually at the midpoint of LSL and USL, sometimes offset if the spec is one-sided or asymmetric.
• Process distribution—the bars themselves. Their center, spread, and shape carry the actual information.

The distance USL − LSL is the specification spread. The natural process variation, defined as 6σ (plus or minus 3 standard deviations covers 99.73% of a normal distribution), is the process spread. Cp measures how much space exists between these two spreads:

A Cp of 1.0 means the process spread equals the spec spread—barely fitting with no margin. Cp of 1.33 gives 25% margin; Cp of 2.0 gives a factor-of-2 cushion. But Cp says nothing about where the process is centered. For that you need Cpk, which penalizes off-center processes.

Pattern 1: Centered and Narrow (What You Want)

The process mean sits near the midpoint of LSL and USL; the distribution is narrow enough that the tails drop to near zero well inside both spec limits. Shape resembles a bell curve.

• What it means: The process is capable and centered. Cp and Cpk will both be high (> 1.33 or > 1.67 depending on industry).
• What to do: Maintain control. Monitor with X-bar and R charts to detect any drift before it reaches the specs.

Pattern 2: Centered but Wide

The mean is centered between LSL and USL, but the distribution bleeds over both spec limits or comes very close. Tails extend beyond the specs on both sides.

• What it means: Process variability is too large for the tolerance. Cp < 1.0. Even if you perfectly center the process, it would still produce nonconformities on both tails.
• What to do: Reduce variation. Tighter tolerances on input materials, better fixturing, operator training, machine capability upgrades—whatever lever reduces σ. Re-centering alone will not solve this.

Pattern 3: Narrow but Off-Center

The distribution is narrow enough to fit within the spec band, but the mean is shifted toward one limit. One tail gets close to (or past) the near spec while the other tail has huge margin to the far spec.

• What it means: Process has capability (Cp is high) but is poorly centered (Cpk is low). Cp − Cpk is large. Immediate risk of nonconformities on the near side.
• What to do: Adjust the process mean. This is usually a setup adjustment (tooling offset, temperature setpoint, flow rate) that moves the mean toward the target. Variation reduction is secondary—centering is the primary fix.

Pattern 4: Bimodal (Two Peaks)

The histogram shows two distinct peaks instead of one. It may still fit within the spec band, but the shape is not a single bell curve.

• What it means: Two processes are mixed in the data. Common causes: two operators with different biases, two machines, two shifts, two material lots, two tool-wear states. The process is not truly “one process”—it is two processes that happen to overlap in the same tolerance band.
• What to do: Stratify the data. Split by operator, machine, shift, or lot and look at each stratum’s histogram separately. One stratum may be centered and narrow while another is off-center. Treat them as separate processes until the root cause is identified and eliminated.

Pattern 5: Truncated or Cliff-Edge

The histogram has a sharp vertical edge right at or near a spec limit, with no gradual tapering. Everything past the edge is absent.

• What it means: Someone is 100% inspecting and discarding (or reworking) parts that fall outside the limit before they enter the data. This is screening, not process capability. The reported data is cleaned; the underlying process is producing nonconformities that are being caught downstream.
• What to do: Find the inspection gate that’s cleaning the data. Include screened parts in the capability study—capability is measured against what the process actually produces, not what passes inspection. Then address the upstream variation that makes the screening necessary.

Pattern 6: Skewed

The distribution has a long tail on one side (positive skew or negative skew). The peak is near one end and the tail stretches toward the other.

• What it means: Either (a) the process is genuinely non-normal (physical constraints like zero-bounded measurements—flatness, concentricity, surface roughness often skew), or (b) a special cause is pulling occasional samples far from the mean.
• What to do: If the skew is intrinsic (zero-bounded data), use a non-normal capability method (Johnson or Box-Cox transformation, or percentile-based Cpk). If the skew comes from occasional outliers, investigate those points as special causes via Western Electric or Nelson rules.

Pattern 7: Normal Curve Fits Poorly

The histogram bars deviate clearly from the overlaid normal curve (flat-topped, multi-modal, heavy-tailed, or gapped).

• What it means: The normality assumption behind Cpk calculations does not hold. The reported Cpk will be biased—sometimes optimistically, sometimes pessimistically—depending on the specific deviation.
• What to do: Test for normality (Anderson-Darling, Shapiro-Wilk). If non-normal, use a non-normal capability method or a distribution-free percentile approach. Flag the Cpk as approximate when reporting to customers.

Building a Defensible Histogram for Capability Analysis

1. Collect sufficient sample size. Minimum 30 data points for preliminary analysis; 100+ for a robust capability study. PPAP submissions usually require 30+ consecutive parts from a significant production run.
2. Verify the measurement system first. If Gage R&R is high, histogram spread includes measurement noise that you will mistake for process variation. Measurement system validation is the prerequisite to any capability study.
3. Use 6–20 bins. Too few bins obscures shape; too many produces a noisy histogram that doesn’t show patterns clearly. Sturges’ rule (bins = ceil(log2(n) + 1)) or the Freedman-Diaconis rule are standard defaults.
4. Overlay the spec limits clearly. Use high-contrast vertical lines. The histogram without spec limits is just a distribution picture; with spec limits, it becomes a capability picture.
5. Overlay a normal curve fitted to the data. Immediate visual check of whether the distribution is approximately normal—Cpk depends on it.
6. Mark target value if it exists. Customers who specify a target (not just spec limits) care about centering on target, not just fit between limits.

Connecting the Histogram to the Numerical Indices

The relationships between what you see and what the indices report:

A Cpk below 1.0 is always a problem. A Cpk above 1.33 with a Cp-minus-Cpk gap of zero is a capable and centered process. A Cpk above 1.33 with a Cp-minus-Cpk of 0.5 is a capable process that’s about to drift into trouble. The histogram shows the why; the indices show the magnitude. Using them together—not either alone—is what experienced quality engineers do.

For the downstream decisions (which control chart type to use, how to set control limits, when to escalate to corrective action), see choosing the right control chart and the SPC control chart tool that generates both histogram and time-series views from uploaded data.

The standard reference for process capability methodology remains the AIAG SPC Reference Manual, which documents histogram interpretation, capability index calculation, and the non-normal-distribution methods cited above. For purely technical references on distributional testing, the NIST Engineering Statistics Handbook covers Anderson-Darling and other normality tests in practitioner-accessible form.