NEWS.md

distributional (development version)

Improvements

• Transformed distributions now work better for monotonic increasing functions when the domain of the untransformed distribution is R (#129).

Breaking changes

• The quantile() method for dist_multivariate_normal() now defaults to equicoordinate quantiles. To obtain marginal quantiles, you should use quantile(dist, p, type = "marginal").

distributional 0.5.0

New features

• support() now shows whether the interval of support is open or closed (@venpopov, #97).
• Added default method for cdf() which estimates the CDF using Monte Carlo integration (@robjhyndman, #122).

Probability distributions

• Added dist_gk() for g-and-k distributions.
• Added dist_gh() for g-and-h distributions.
• Added dist_gev() for the Generalised Extreme Value distribution and dist_gpd() for the Generalised Pareto distribution (@robjhyndman, #124).

Improvements

• dist_mixture() now displays the components of the mixture when the output width is sufficiently wide (@statasaurus, #112).
• generate() now respects dimnames() for multivariate distributions.
• dist_mixture() now supports multivariate distributions (@robjhyndman, #122).

Bug fixes

• Fixed error when using '-' as a unary operator on a distribution different from dist_normal() (@venpopov, #95)
• Density for transformed distributions now correctly gives 0 instead of NaNs for values outside the support of the distribution (@venpopov, #97)
• Fixed quantile() and cdf() for transformed distributions with monotonically decreasing transformations (#100).
• Fixed multivariate dist_sample() methods not structuring multivariate results correctly as matrices.
• The cdf() method for dist_multivariate_normal() now gives the P(X <= q) rather than P(X > q) for consistency with all other cdf() methods.
• The quantile() method for dist_multivariate_normal() now correctly gives the boundaries when p=0 or p=1 when type="equicoordinate".
• The quantile() method for dist_multivariate_normal() now only square roots the diagonal elements when type="marginal".

distributional 0.4.0

Breaking changes

• All graphics related functionality has been removed from the package in favour of the ggdist (https://cran.r-project.org/package=ggdist) package. This breaking change was done to substantially reduce the package's dependencies, focusing the functionality on representing vectors of distributions.

distributional 0.3.2

Small patch to resolve issues with CRAN checks.

Bug fixes

• Fixed object structure resulting from transforming sample distributions (#81).
• Improved reliability of quantile(<dist_mixture>).
• Defined cdf(<dist_sample>) as Pr(X <= x), not Pr(X < x).
• Fixed S3 generic argument name p for log_quantile().

distributional 0.3.1

New features

• Add Math and Ops methods for sample distribution, which applies the functions directly to the samples.
• Added mean and sd as aliases for mu and sigma respectively in dist_normal() and dist_student_t() to match arguments of the stats package interface (#76).
• Added scale argument for alternative specification for dist_burr() and dist_gamma().

Improvements

• Generics introduced by this package now allow na.rm and other parameters to be passed to distribution methods, even if these parameters aren't used. The package no longer checks the usage of ... with the ellipsis package, if you'd like to check that all ... are used, you can write your own wrapping functions.
• Lists of functions can now be used in dist_transformed(), allowing the transformation to differ for each distribution.
• covariance() and other matrix output functions of multivariate distributions now name the result using the distribution's dimension names.
• Improve handling of mixture distribution quantiles at boundaries {0,1}.

Bug fixes

• Fixed issue with computing multiple values from a univariate distribution with named dimensions (#79).

distributional 0.3.0

New features

Probability distributions

• Added dist_categorical() for the Categorical distribution.
• Added dist_lognormal() for the log-normal distribution. Mathematical conversion shortcuts have also been added, so exp(dist_normal()) produces dist_lognormal().

Generics

• Added parameters() generic for obtaining the distribution's parameters.
• Added family(<distribution>) for getting the distribution's family name.
• Added covariance() to return the covariance of a distribution.
• Added support() to identify the distribution's region of support (#8).
• Added log_likelihood() for computing the log-likelihood of observing a sample from a distribution.

Improvements

• variance() now always returns a variance. It will not default to providing a covariance matrix for matrices. This also applies to multivariate distributions such as dist_multivariate_normal(). The covariance can now be obtained using the covariance() function.
• dist_wrap() can now search for distribution functions in any environment, not just packages. If the package argument is NULL, it will search the calling environment for the functions. You can also provide a package name as before, and additionally an arbitrary environment to this argument.
• median() methods will now ignore the na.rm option when it does not apply to that distribution type (#72).
• dist_sample() now allows for missing values to be stored. Note that density(), quantile() and cdf() will remove these missing values by default. This behaviour can be changed with the na.rm argument.
• <hilo> objects now support non-numeric and multivariate distributions. <hilo> vectors that have different bound types cannot be mixed (#74).
• Improved performance of default methods of mean() and variance(), which no longer use sampling based means and variances for univariate continuous distributions (#71, @mjskay)
• dist_binomial() distributions now return integers for quantile() and generate() methods.
• Added conditional examples for distributions using functions from supported packages.

Bug fixes

• Fixed fallback format() function for distributions classes that have not defined this method (#67).

Breaking changes

• variance() on a dist_multivariate_normal() will now return the diagonal instead of the complete variance-covariance matrix.
• dist_bernoulli() will now return logical values for quantile() and generate().

distributional 0.2.2

New features

• Added is_distribution() to identify if an object is a distribution.

Improvements

• Improved NA structure of distributions, allowing it to work with is.na() and vctrs vector resizing / filling functionality.
• Added as.character(<hilo>) method, allowing datasets containing hilo() objects to be saved as a text file (#57).

Bug fixes

• Fixed issue with hdr() range size incorrectly being treated as 100-size, giving 5% ranges for 95% sizes and vice-versa (#61).

distributional 0.2.1

A small performance and methods release. Some issues with truncated distributions have been fixed, and some more distribution methods have been added which improve performance of common tasks.

New features

Probability distributions

• Added dist_missing() for representing unknown or missing (NA) distributions.

Improvements

• Documentation improvements.
• Added cdf() method for dist_sample() which uses the emperical cdf.
• dist_mixture() now preserves dimnames() if all distributions have the same dimnames().
• Added density() and generate() methods for sample distributions.
• Added skewness() method for dist_sample().
• Improved performance for truncated Normal and sample distributions (#49).
• Improved vectorisation of distribution methods.

Bug fixes

• Fixed issue with computing the median of dist_truncated() distributions.
• Fixed format method for dist_truncated() distributions with no upper or lower limit.
• Fixed issue with naming objects giving an invalid structure. It now gives an informative error (#23).
• Fixed documentation for Negative Binomial distribution (#46).

distributional 0.2.0

New features

Probability distributions

• Added dist_wrap() for wrapping distributions not yet added in the package.

Methods

• Added likelihood() for computing the likelihood of observing a sample from a distribution.
• Added skewness() for computing the skewness of a distribution.
• Added kurtosis() for computing the kurtosis of a distribution.
• The density(), cdf() and quantile() methods now accept a log argument which will use/return probabilities as log probabilities.

Improvements

• Improved documentation for most distributions to include equations for the region of support, summary statistics, density functions and moments. This is the work of @alexpghayes in the distributions3 package.
• Documentation improvements
• Added support for displaying distributions with View().
• hilo() intervals can no longer be added to other intervals, as this is a common mistake when aggregating forecasts.
• Incremented d for numDeriv::hessian() when computing mean and variance of transformed distributions.

Deprecated features

• Graphics functionality provided by autoplot.distribution() is now deprecated in favour of using the ggdist package. The ggdist package allows distributions produced by distributional to be used directly with ggplot2 as aesthetics.

distributional 0.1.0

First release.

New features

Object classes

• distribution: Distributions are represented in a vectorised format using the vctrs package. This makes distributions suitable for inclusion in model prediction output. A distribution is a container for distribution-specific S3 classes.
• hilo: Intervals are also stored in a vector. A hilo consists of a lower bound, upper bound, and confidence level. Each numerical element can be extracted using $, for example my_hilo$lower to obtain the lower bounds.
• hdr: Highest density regions are currently stored as lists of hilo values. This is an experimental feature, and is likely to be expanded upon in an upcoming release.

Generic functions

Values of interest can be computed from the distribution using generic functions. The first release provides 9 functions for interacting with distributions:

• density(): The probability density/mass function (equivalent to d...()).
• cdf(): The cumulative distribution function (equivalent to p...()).
• generate(): Random generation from the distribution (equivalent to r...()).
• quantile(): Compute quantiles of the distribution (equivalent to q...()).
• hilo(): Compute probability intervals of probability distribution(s).
• hdr(): Compute highest density regions of probability distribution(s).
• mean(): Obtain the mean(s) of probability distribution(s).
• median(): Obtain the median(s) of probability distribution(s).
• variance(): Obtain the variance(s) of probability distribution(s).

Graphics

• Added an autoplot() method for visualising the probability density function ([density()]) or cumulative distribution function ([cdf()]) of one or more distribution.
• Added geom_hilo_ribbon() and geom_hilo_linerange() geometries for ggplot2. These geoms allow uncertainty to be shown graphically with hilo() intervals.

Probability distributions

• Added 20 continuous probability distributions: dist_beta(), dist_burr(), dist_cauchy(), dist_chisq(), dist_exponential(), dist_f(), dist_gamma(), dist_gumbel(), dist_hypergeometric(), dist_inverse_exponential(), dist_inverse_gamma(), dist_inverse_gaussian(), dist_logistic(), dist_multivariate_normal(), dist_normal(), dist_pareto(), dist_student_t(), dist_studentized_range(), dist_uniform(), dist_weibull()
• Added 8 discrete probability distributions: dist_bernoulli(), dist_binomial(), dist_geometric(), dist_logarithmic(), dist_multinomial(), dist_negative_binomial(), dist_poisson(), dist_poisson_inverse_gaussian()
• Added 3 miscellaneous probability distributions: dist_degenerate(), dist_percentile(), dist_sample()

Distribution modifiers

• Added dist_inflated() which inflates a specific value of a distribution by a given probability. This can be used to produce zero-inflated distributions.
• Added dist_transformed() for transforming distributions. This can be used to produce log distributions such as logNormal: dist_transformed(dist_normal(), transform = exp, inverse = log)
• Added dist_mixture() for producing weighted mixtures of distributions.
• Added dist_truncated() to impose boundaries on a distribution's domain via truncation.