Start / Is binference right for my data?

Is binference right for my data?

binference fits when three things hold. You have many objects. Each object carries its own noisy information about one quantity, which you can divide into bins (intervals, grid cells or categories). And you want the population of that quantity, or each object’s value informed by it. The tables below match what you have and what you want to binference’s inputs and outputs.

What you have

Each row names the constructor that turns your data into a Likelihoods, then the guide section that explains it and the tutorials that use it.

You have, for each object	binference reads it with	Read more
a measured value and its Gaussian error	Likelihoods.from_gaussian	guide, quick start, Empirical Bayes deconvolution
likelihoods or log-likelihoods on the bins	Likelihoods.from_table	guide, Mixed-stock analysis
a likelihood function or a SciPy distribution	Likelihoods.from_density	guide
a posterior on the bins, and the prior it was computed under	Likelihoods.from_table	guide, Posteriors on a grid
posterior samples, and the prior they were drawn under	Likelihoods.from_samples	guide, Posteriors on a grid
the bin it was measured in, and the detector’s response matrix	Likelihoods.from_binned	guide, Binned counts and a response: unfolding a spectrum
a classifier’s probabilities, and the class prevalence it was trained on	Likelihoods.from_table	guide, Classifier outputs: class fractions under label shift
a classifier’s scores, and a labelled validation set	Likelihoods.from_binned with Response.from_labelled	guide, Classifier outputs: class fractions under label shift, Mixed-stock analysis

Posteriors and posterior samples need their interim prior, which binference divides out (pitfalls).

What you want

You want	binference gives	Read more
the population, with its uncertainty	result.samples, result.summary, result.plot	guide, Gaussian measurements: a redshift distribution
a number computed from the population, such as its mean	result.functional	guide
each object’s value, informed by the population: its hierarchical posterior	result.assignment_probabilities, result.sample_values	guide, Empirical Bayes deconvolution
the histogram of these objects, the realised sample	result.realized_counts	guide
totals and rates from an incomplete catalogue, given its selection function	infer’s selection, then result.total, result.rates	guide, Selection effects and rates, Selection measured by injections
a quick single answer, a point estimate	method="map"	guide

What binference does not do

• Compute the per-object inputs. binference starts where each object already has a likelihood, a posterior, samples or a measured bin. Fitting photometry, training a classifier or simulating a detector happens before it.

• Use posteriors whose prior is unknown. Without its interim prior, binference cannot turn a posterior back into a likelihood.

• Fit a population without bins. The population is a histogram on your bins, optionally with a finer B-spline Basis. binference has no parametric population models, such as a power law, and no neural density models (Neural and simulation-based deconvolution).

• Let the population depend on other variables. All objects share one population; binference has no covariate-dependent populations.

• Model correlated objects. The model multiplies the objects’ likelihoods, so it treats their data as independent given the population (theory).