Does PDM require a sine-wave pattern?

No; that flexibility is a central reason to use phase-based dispersion.

Can theta exceed one?

Yes, depending on the grouping and normalization, as the example shows.

Does uneven sampling make every result invalid?

No, but it affects coverage, aliases, and appropriate inference.

AI Glossary / Statistics and probability / Intermediate

Phase dispersion minimization

Definition

Phase dispersion minimization, usually abbreviated PDM, is a method for finding candidate periods in time-series data. It folds observations at trial periods, groups nearby phases, and measures scatter within those groups. A period that aligns repeated features tends to produce smaller within-phase dispersion. PDM can be useful for repeating patterns that are not well described by a sinusoid.

What it means

Folding means mapping observations from different cycles onto a single cycle. If the trial period matches a repeating pattern, observations from similar points in that pattern should tend to agree.
A low dispersion value identifies a candidate worth examining, not automatic proof of a real physical cycle. Sampling gaps, noise, trends, harmonics, and choices about phase bins can create or alter minima in the search result.

How it works

Begin with measurement times, observed values, and information about uncertainty. Inspect the raw series for data errors, changes in measurement conditions, and trends. Any detrending or exclusions should be documented because they can change the apparent periodic structure.
Choose a range of trial periods or frequencies and a suitable search resolution. Frequency is the reciprocal of period. A grid that is uniform in frequency is not uniform in period, and a coarse grid can miss a narrow minimum.
For each trial period P and reference time t0, compute phase as the fractional part of (t minus t0)/P. This places each observation between phase zero and one, irrespective of which cycle it came from.
Divide the cycle into phase bins and compute dispersion within those bins. In a simple nonoverlapping-bin formulation, pool the within-bin sums of squared deviations with their degrees of freedom, then divide that pooled variance by the variance of all observations. The resulting theta statistic is small when phase grouping substantially reduces scatter.
Implementations differ in binning, overlapping covers, weighting, and treatment of sparsely populated bins. State the particular version rather than assuming all PDM programs calculate identical numbers. A bin with too few observations cannot support an ordinary sample-variance estimate.
Repeat across the search grid and examine local minima. Plot the data folded at promising periods, including plausible multiples and fractions of them. A repeated pattern with similar subfeatures can produce competing harmonic candidates.
Evaluate significance and period uncertainty using methods appropriate to the sampling and noise. Searching many trial frequencies changes the interpretation of an unusually low value. A randomization procedure must preserve relevant dependence; arbitrary shuffling is inappropriate when correlated noise is central.
Confirm the pattern with additional observations or a separate time interval where possible. The analysis should distinguish a stable repeating process from a temporary pattern or an observing schedule that itself repeats.

Three practical examples

Astronomy: researchers search for periods in variable-star brightness curves with nonsinusoidal shapes.
Sensor analysis: engineers investigate recurring features in an unevenly sampled monitoring record.
AI-assisted science: a system proposes candidate periods for expert review while retaining the folded plots and search settings.

Think of it this way

Think of wrapping a long strip of measurements around a cylinder. If the circumference matches the true repeating length, similar features line up. PDM scores how tightly the aligned observations cluster within sections of that circumference.

Worked example

Consider eight observations at times 0 through 7, with values 0, 10, 1, 11, -1, 9, 0, and 10. Use two nonoverlapping phase bins, [0, 0.5) and [0.5, 1), with reference time zero.
The overall mean is 5, and the total squared deviation is 204. The overall sample variance is therefore 204/7.
At trial period 2, the bins contain 0, 1, -1, 0 and 10, 11, 9, 10. Their means are 0 and 10; each has squared deviation sum 2. The pooled within-bin variance is (2 + 2)/(8 minus 2) = 4/6. Theta is (4/6)/(204/7) = 7/306, approximately 0.02288.
At trial period 4 with the same two-bin rule, the bin means are 4.5 and 5.5, and the combined within-bin squared deviation is 202. Theta becomes (202/6)/(204/7) = 707/612, approximately 1.15523.
Period 2 gives much tighter phase grouping in this illustration. The result depends on the specified coarse binning and is not a significance test or a uniquely established period.

Illustrative case study

Fictional case: Rowan Observatory uses PDM to investigate a small irregularly sampled brightness record. The analyst reviews several low-theta candidates, checks the observing schedule for aliases, and plans additional observations where the competing folded curves disagree. The team records a candidate period with uncertainty instead of publishing the minimum grid point as a definitive discovery.

Common misconceptions and mistakes

Do not assume the deepest minimum is always the fundamental period. Harmonics, aliases, and multiple physical processes can produce competing minima.
PDM does not eliminate the need to consider noise or sampling. A small theta is not itself a probability.
Avoid selecting bins or deleting observations solely to strengthen a preferred result. Different bin counts trade phase resolution against the number of observations available for estimating scatter.
An apparently periodic signal may drift in phase or amplitude, limiting a fixed-period description.

Business use

PDM is most relevant to businesses handling scientific or instrument time series with potentially repeating, non-sinusoidal signals. It can provide candidate cycles for inspection, but operational forecasts or control decisions should use validated models and explicitly account for uncertainty and changing conditions.

Check your understanding

Why is period 2 favored in the example?

Answer: It groups low values together and high values together, leaving little within-bin scatter.

What would be needed before treating that candidate as a real cycle?

Answer: Appropriate uncertainty assessment, sampling checks, and preferably independent confirmation.

Related terms

Time series
Aliasing (factorial experiments)
Variance
phase folding
periodogram
aliasing
harmonic
periodicity
Lomb–Scargle periodogram

Learn next

Learn next how phase-bin choices influence sensitivity, how observing schedules create aliases, and how search-wide significance differs from evaluating one predetermined period.

Sources

Sources: PyAstronomy, Phase Dispersion Minimization documentation, https://pyastronomy.readthedocs.io/en/latest/pyTimingDoc/pyPDMDoc/pdm.html ; IRAF, pdm task documentation, https://iraf.readthedocs.io/en/latest/tasks/noao/astutil/pdm.html . These primary developer references identify the Stellingwerf method, phase-bin search, and implementation choices. The explicitly simplified calculation and fictional case are original.

Previous: Phase correlation Next: Phase (Fourier)