Mixture
Fits a mixture of continuous distributions to univariate data by partitioning the sorted sample, fitting a component to each part, and ranking the assembled mixtures.
By default the facade reproduces the configuration the designed experiment measured as best: a Jenks initial partition, refinement by local search over the cut positions, and separable family selection. Every one of those is a parameter of fit, so the unrefined baseline and the exhaustive family search remain available for comparison — but a user who asks for nothing in particular gets the configuration the evidence supports rather than the simplest one.
The number of components to consider is intersected with what the data can structurally support before anything is fitted, so an impossible request is reported rather than discovered as a failure part-way through.
Which entry point to use. fitSpecified is the documented path and fit over a range is the sensitivity tool, which is the opposite of what a distribution-fitting facade usually implies. The reason is measured rather than stylistic: across the designed experiment the criterion recovers the true number of components on about 36% of attempts, under-selects eight times more often than it over-selects, and is confidently wrong when it is wrong — the median criterion gap to the truth is around 19, so the size of the gap is not a signal that the choice is reliable. An analyst who has a mechanism in mind, or who has read a count off a plot, holds better information than the search does, and this class should not silently overrule them.
So: name the count when you have one, and use the range form to ask how sensitive the answer is to that choice. Neither is more supported than the other computationally; the difference is which question is being asked.
Parameters
the observations, which need not be sorted; a sorted copy is taken
the component fitter, wrapped in a cache by this class
the criterion used to rank assembled mixtures
the least number of observations permitted in a group
the least number of distinct values permitted in a group
Constructors
Properties
The structural facts of the sample, computed once.
A histogram of the sample.
The largest number of components this sample can structurally support.
The observations in sorted order.
The sample statistics.
Functions
Whether the fitted density reproduces the sample it came from.
How many modes the data appears to have, and whether that appearance survives smoothing.
Everything worth knowing about the sample before fitting anything to it.
Fits a mixture with the number of components the caller has chosen.
Fits the numbers of components the caller is choosing between.
How far the reported density moves when the cuts between the components are nudged.
The component counts this sample does not rule out.
Tests a component count the analyst proposed against one more than itself.