AI Glossary / Statistics and probability / Intermediate

Kuiper's test

Definition

Kuiper's test is a goodness-of-fit or two-sample distribution-comparison test based on the sum of the largest upward and downward differences between cumulative distribution functions. Its statistic is commonly written V = D+ + D−. This differs from the Kolmogorov - Smirnov statistic, which takes the larger of the two deviations. Kuiper's statistic is particularly useful for circular or cyclic data because an appropriately defined comparison is invariant to the chosen origin around the circle. The statistic requires suitable null calibration before it can support a significance conclusion.

What it means

Angles, times of day and phases wrap around: a value just before the end of a cycle is close to one just after the beginning. Choosing midnight rather than noon as the starting point should not alter a test of uniformity around the same cycle. Kuiper's combination of positive and negative cumulative deviations addresses that origin issue. It can also be used for ordinary continuous measurements. The test detects distributional discrepancies, but a significant result does not identify their cause or establish a particular alternative distribution.

How it works

For a one-sample continuous null CDF F₀, sort the observations and calculate D+ = max[i/n − F₀(x₍ᵢ₎)] and D− = max[F₀(x₍ᵢ₎) − (i − 1)/n]. Add the two maxima to obtain V. For a uniform circular null, express observations as phases between zero and one using a consistent cycle length. A two-sample version compares two empirical CDFs. Obtain a p-value using calibration appropriate to the sample sizes and design. Estimated null parameters, discrete values, substantial ties or dependent observations can require different calibration rather than ordinary continuous-null formulas.

Three practical examples

• Cyclic timing: investigate whether event phases are uniformly distributed over a day when observation exposure supports that null.
• Directional measurements: compare angular observations with a specified circular reference distribution.
• Sample comparison: compare phase distributions from two independently sampled conditions while preserving the circular nature of the measurements. Different observation schedules or measurement processes can create apparent differences unrelated to the phenomenon of interest.

Think of it this way

Think of cutting a circular ribbon at any point and laying it out as a line. The cumulative discrepancy may move above or below the reference depending on where the cut is made. Kuiper's statistic adds the largest upward and downward departures, preserving the relevant overall discrepancy when the circle's origin changes. The ribbon analogy explains the origin property; it does not remove the need for independent observations or a defensible reference distribution.

Worked example

Take four phases 0.10, 0.20, 0.60 and 0.90 under a specified uniform-circle null. Since F₀(x) = x, the maximum upward empirical-CDF deviation is D+ = 0.30, attained after the second observation. The maximum downward deviation is D− = 0.15, attained just before 0.90. Thus V = 0.45. Now rotate every phase by 0.15 modulo one. The sorted phases become 0.05, 0.25, 0.35 and 0.75. Their upward maximum is 0.40 and downward maximum is 0.05, so V is still 0.45. The KS maximum would change from 0.30 to 0.40 in this example, while the Kuiper sum remains unchanged. The value 0.45 is a statistic, not a p-value or a 45% probability that the null is wrong. The four-observation example illustrates the calculation and rotation property only.

Illustrative case study

Fictional case study: Cedar Monitoring examines the daily phase of equipment alerts. A uniformity test finds a pattern, but analysts discover that some sensors are switched off during part of the day. They revise the null model to reflect observation exposure and investigate whether alerts are clustered within machines. The team also changes the plotted clock origin to make the peak easier to see, confirming that the uniform-circle statistic is unchanged when all phases are rotated consistently. Statistical significance alone would not have distinguished operational timing from the recording schedule.

Common misconceptions and mistakes

Common mistakes include using the KS p-value for a Kuiper statistic, interpreting the statistic as a probability, and assuming circular data are automatically uniform under the null. Origin invariance is not permission to rotate only one component of a comparison inconsistently. Ordinary continuous-null calibration does not automatically apply to rounded or discrete phases. If a period or reference parameters are selected by searching the same data, that selection must be reflected in inference. Scanning many periods or many groups also creates a multiple-testing problem.

Business use

Businesses can use Kuiper's test in cyclic monitoring, directional measurements and distribution checks when those questions fit the data. Record the phase convention, cycle length, reference distribution, sample size and calibration method. Examine plots and exposure patterns to understand the discrepancy's practical meaning. Account for repeated observations from the same units and for dependence over time. Software may use approximations for false-positive probabilities, especially in small samples or particular ranges, so verify that the reported calibration suits the intended decision rather than treating every returned decimal as exact.

People also ask

Is the test only for circular data?

No. It can compare continuous distributions more generally, but circular invariance is a notable advantage.

Is V the same as the KS statistic?

No. Kuiper adds D+ and D−; KS takes their maximum.

Does a significant result prove a single preferred phase?

No. Many forms of nonuniformity, including multiple peaks, can produce a discrepancy.

Check your understanding

Why does the rotated example retain V = 0.45 even though D+ changes?

Answer: The upward and downward deviations redistribute when the circular origin changes. Their sum remains the same for this consistently rotated uniform-circle comparison.

Related terms

• Phase (Fourier)
• Goodness-of-fit
• circular statistics
• Kolmogorov–Smirnov test
• empirical CDF
• uniformity test
• phase
• multiple testing

Learn next

reproduce the phase rotation calculation and compare one-sample with two-sample formulations. Then examine how unequal exposure, estimated periods and rounded measurements affect the null model and the calibration needed for a valid test.

More in Statistics and probability

• Kolmogorov-Smirnov test - The Kolmogorov - Smirnov test is a distribution-comparison test based on the largest vertical difference between cumulative distribution functions.
• Kriging - Kriging is a family of prediction methods that use a modeled covariance or variogram to combine observed values when estimating an unobserved value, often...
• Kruskal­Wallis one-way analysis of variance - The Kruskal - Wallis one-way analysis of variance is a rank-based test for comparing independent groups on an ordered or numerical outcome.
• Kurtosis - Kurtosis is a standardized fourth-moment measure of a distribution.
• Lack of fit - Lack of fit is systematic disagreement between a model's assumed relationship and the pattern supported by the data.
• Lack-of-fit sum of squares - The lack-of-fit sum of squares is the part of regression residual variation attributable to discrepancies between observed means at replicated predictor...

Sources

Sources: Astropy, kuiper documentation, https://docs.astropy.org/en/stable/api/astropy.stats.kuiper.html. PyCircStat2, hypothesis-testing documentation, https://circstat.github.io/pycircstat2/reference/hypothesis/ (two-sample circular comparison). M. A. Stephens, Use of the Kolmogorov–Smirnov, Cramér–Von Mises and Related Statistics Without Extensive Tables, https://academic.oup.com/jrsssb/article/32/1/115/7027180. Calculation and fictional case are original illustrations.

Previous: Kruskal­Wallis one-way analysis of variance Next: Kullback-Leibler divergence