NEWS.md
support() now shows whether the interval of support is open or closed (@venpopov, #97).cdf() which estimates the CDF using Monte Carlo integration (@robjhyndman, #122).dist_gk() for g-and-k distributions.dist_gh() for g-and-h distributions.dist_gev() for the Generalised Extreme Value distribution and dist_gpd() for the Generalised Pareto distribution (@robjhyndman, #124).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).dist_normal() (@venpopov, #95)quantile() and cdf() for transformed distributions with monotonically decreasing transformations (#100).dist_sample() methods not structuring multivariate results correctly as matrices.cdf() method for dist_multivariate_normal() now gives the P(X <= q) rather than P(X > q) for consistency with all other cdf() methods.quantile() method for dist_multivariate_normal() now correctly gives the boundaries when p=0 or p=1 when type="equicoordinate".quantile() method for dist_multivariate_normal() now only square roots the diagonal elements when type="marginal".Small patch to resolve issues with CRAN checks.
quantile(<dist_mixture>).cdf(<dist_sample>) as Pr(X <= x), not Pr(X < x).p for log_quantile().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).scale argument for alternative specification for dist_burr() and dist_gamma().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.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.dist_categorical() for the Categorical distribution.dist_lognormal() for the log-normal distribution. Mathematical conversion shortcuts have also been added, so exp(dist_normal()) produces dist_lognormal().parameters() generic for obtaining the distribution’s parameters.family(<distribution>) for getting the distribution’s family name.covariance() to return the covariance of a distribution.support() to identify the distribution’s region of support (#8).log_likelihood() for computing the log-likelihood of observing a sample from a distribution.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).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.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().is_distribution() to identify if an object is a distribution.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.
dist_missing() for representing unknown or missing (NA) distributions.cdf() method for dist_sample() which uses the emperical cdf.dist_mixture() now preserves dimnames() if all distributions have the same dimnames().density() and generate() methods for sample distributions.skewness() method for dist_sample().dist_truncated() distributions.dist_truncated() distributions with no upper or lower limit.dist_wrap() for wrapping distributions not yet added in the package.likelihood() for computing the likelihood of observing a sample from a distribution.skewness() for computing the skewness of a distribution.kurtosis() for computing the kurtosis of a distribution.density(), cdf() and quantile() methods now accept a log argument which will use/return probabilities as log probabilities.distributions3 package.View().hilo() intervals can no longer be added to other intervals, as this is a common mistake when aggregating forecasts.d for numDeriv::hessian() when computing mean and variance of transformed distributions.First release.
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.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).autoplot() method for visualising the probability density function ([density()]) or cumulative distribution function ([cdf()]) of one or more distribution.geom_hilo_ribbon() and geom_hilo_linerange() geometries for ggplot2. These geoms allow uncertainty to be shown graphically with hilo() intervals.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()
dist_bernoulli(), dist_binomial(), dist_geometric(), dist_logarithmic(), dist_multinomial(), dist_negative_binomial(), dist_poisson(), dist_poisson_inverse_gaussian()
dist_degenerate(), dist_percentile(), dist_sample()
dist_inflated() which inflates a specific value of a distribution by a given probability. This can be used to produce zero-inflated distributions.dist_transformed() for transforming distributions. This can be used to produce log distributions such as logNormal: dist_transformed(dist_normal(), transform = exp, inverse = log)
dist_mixture() for producing weighted mixtures of distributions.dist_truncated() to impose boundaries on a distribution’s domain via truncation.