Mixture
Everything produced by one modeling run.
Parameters
the evaluated candidates, best first
the structural facts of the sample
the fit results per group, keyed by number of components
the requested numbers of components that were not attempted, with the reason
the fraction of group fits answered from the cache
the numbers of components whose family search was cut short, with what was cut. These were attempted and produced an answer; that answer may simply not be the best available, which is a different fact from the count being unusable and is reported separately for that reason.
what the refiner did at each number of components
the observations the mixtures were fitted to, in order. Carried so that the diagnostics can be asked for without handing the data back in, which is how the library's single-distribution results behave.
the criterion that ranked the candidates. Needed to orient a comparison: whether a smaller value is an improvement is a property of the criterion, not of the number.
the numbers of components the caller asked for, when they asked rather than letting the criterion choose. Null after an ordinary search. It changes what the results mean: the recommendation comes from the caller's counts, and the criterion values are evidence about that choice rather than a selection over a searched range.
Constructors
Properties
The recommended candidate, or null when nothing could be fitted.
Whether a fitted mixture shows the signature of having been over-fitted.
How many families the fitter offered, inferred from what a group actually attempted: the candidates that succeeded plus the ones that were rejected. Needed by the extended criterion, which charges for the size of the model space being searched.
Whether the recommendation has more than one component.
How much better the recommended mixture is than the best single distribution, in the criterion's own units, oriented so that a positive number always means the mixture is better. Null when either is unavailable.
Whether the criterion prefers the recommended mixture to the best single distribution.
Whether the number of components was supplied by the caller rather than selected.
The observations the mixtures were fitted to, in order.
The recommendation as an ordinary continuous distribution, ready to be sampled from, plotted, or handed to a simulation model. Null when nothing could be fitted.
The recommended number of components, or null when nothing could be fitted.
The best single distribution, which is simply the best candidate at one component.
The single count the caller asked for, or null when they asked for several or none.
Functions
The standard report as a self-contained HTML page.
The standard report as Markdown, for pasting into a document that is already text.
The best candidate for each number of components attempted, which is the profile a reader needs in order to judge how decisively the recommendation was made.
How far the reported criteria agree about the number of components.
Why the criterion chose what it chose, assembled from the per-count profile.
Whether the recommended mixture's components are ones the data can actually tell apart, or null when the comparison could not be made.
The recommended mixture's components, one row per component.
The recommended mixture as a table, one row per component.
Whether this sample is large enough for the recommended component count to be answerable at all, or null when the question does not arise.
How much of a fresh sample the recommended fit cannot account for at all.
Every reported criterion's value at every number of components that was fitted.
Every reported criterion's value at every number of components, in long form.
What each criterion would choose, and whether it actually chose.
The four diagnostic plots for the recommended mixture, or null when nothing was fitted.
Goodness-of-fit tests for the recommended mixture, or null when nothing was fitted.
Every family that failed to fit some group, with the estimator's explanation and how often it happened. Estimator failure rates are an experimental result, so they are surfaced rather than logged and forgotten.
Writes the summary and opens the diagnostic plots.
Writes the HTML report to a file and opens it.
Everything a reader needs in order to judge the recommendation, in one place.
The standard mixture modeling report.
A direct answer to "was a mixture worth it?", comparing the recommendation against the best single distribution on the criterion and on the usual goodness-of-fit tests.