Elementary prior related functions

Density (pdf / lpdf), distribution function (cdf / ccdf), quantile function (quant), random generation (rng), mean, standard deviation (sd), and marginal variants of the functions (mpdf, mlpf, mcdf, mccdf, mquant) for prior distributions. Joint quantiles are unavailable for vector factor priors; use mquant() for their marginal quantiles. The marginal methods and var()/sd() of a bound orthonormal or mean-difference mnormal/mt factor prior describe its continuous levels: they take the scale of each level from the declared factor design, a level whose design row vanishes identically (the reference level of a treatment factor inside an interaction) is the atom at zero that the cell declares and does not enter the common scale, and levels whose non-zero rows differ in norm have no common marginal and refuse. Missing evaluation values NA_real_ and NaN propagate distinctly through density, distribution and quantile methods for moment, inverse-moment and inverse-gamma scalar priors, including their marginal methods. Generic scalar CDF and complementary CDF methods also retain this distinction under truncation. This statement does not cover truncated Normal quantiles. Continuous truncation uses logarithmic interval masses, including separate signed radial intervals for nonlocal priors. An unresolved positive normalizer stops with BayesTools_numerical_unavailable before a cached density evaluator or bridge calculation is constructed. Existing Normal-only rounding adjustments are retained; other truncated quantiles never clip an unavailable result into their bounds. Ordinary Gamma normalization is unavailable for a numeric subnormal shape when truncation cuts its positive support: finite base R log tails can be wrong in this floating-point regime. Constructors retain their supported parameter domain and whole-positive-support mass remains exactly one. Raw untruncated base R Gamma extreme calculations are not certified by this normalization guard.

Valid interior numerical failures give NaN with one BayesTools_numerical_unavailable warning per vector. Established inverse-gamma or nonlocal natural density or quantile range limits retain their raw zero or infinite result with BayesTools_numerical_range_limit. Ordinary family density range results retain their existing behavior when their log density is available. Both numerical classes inherit BayesTools_numerical_condition, warning, and condition. Conditions record operation, family, requested_scale, indices, and reason, with call = NULL. Exact support endpoints and rounded natural probabilities with available log tails do not warn. Sampling failures are errors additionally inheriting BayesTools_prior_rng_unavailable; all original uniforms are consumed before signaling. Finite-required consumers and JAGS refuse unavailable draws, density values, and normalization, without redraw or model repair.

# S3 method for class 'prior'
rng(x, n, ...)

# S3 method for class 'prior'
cdf(x, q, ...)

# S3 method for class 'prior'
ccdf(x, q, ...)

# S3 method for class 'prior'
lpdf(x, y, ...)

# S3 method for class 'prior'
pdf(x, y, ...)

# S3 method for class 'prior'
quant(x, p, ...)

# S3 method for class 'prior'
mcdf(x, q, ...)

# S3 method for class 'prior'
mccdf(x, q, ...)

# S3 method for class 'prior'
mlpdf(x, y, ...)

# S3 method for class 'prior'
mpdf(x, y, ...)

# S3 method for class 'prior'
mquant(x, p, ...)

Arguments

x

prior distribution

n

number of observations

...

additional arguments of rng(); the other functions ignore them. transform_factor_samples (default TRUE): whether the samples of orthonormal, mean-difference and ordered factor priors are transformed to the level scale (FALSE: the raw coefficients); sample_components (default FALSE): whether mixture and spike-and-slab priors return the sampled component (or inclusion indicator) instead of the values; quantity (ordered priors only, default "level"): the draws returned, one of "level" (the transformed level samples), "coefficient" and "increment" (the raw coefficients), "total" (the totals of the ordered sums) or "allocation" (the allocation shares).

q

vector or matrix of quantiles

y

vector of observations

p

vector of probabilities

Value

pdf (mpdf) and lpdf (mlpdf) give the (marginal) density and the log of (marginal) density, cdf (mcdf) and ccdf (mccdf) give the (marginal) distribution and the complement of (marginal) distribution function, quant (mquant) give the (marginal) quantile function, and rng generates random deviates for an object of class 'prior'.

Details

Numeric Dirichlet concentrations must be finite and at least 0.01. This supported-input minimum does not guarantee representable draws for every accepted specification. Dirichlet and cumulative weight-function rng() calls stop with class BayesTools_dirichlet_rng_unavailable, parent class BayesTools_prior_rng_unavailable, and classes error and condition if sampling produces non-finite entries, entries outside [0, 1], or a simplex row with zero total. The condition has call = NULL and message: "Dirichlet draws are unavailable because sampling produced non-finite or degenerate simplex draws. Use a less extreme 'alpha' specification." There is no retry or repair. Individual rounded zero or one entries are retained and do not represent new prior atoms. Ordinary expression-valued Dirichlet priors retain their JAGS-only support; R random generation refuses priors containing parameter expressions.

Examples

# create a standard normal prior distribution
p1 <- prior(distribution = "normal", parameters = list(mean = 1, sd = 1))

# generate a random sample from the prior
rng(p1, 10)
#>  [1] -1.5923277  2.3140022  0.3644570  0.5700212  0.8306817  1.6122182
#>  [7]  1.6783402  1.5679520  0.4274574 -0.3632913

# compute cumulative density function
cdf(p1, 0)
#> [1] 0.1586553

# obtain quantile
quant(p1, .5)
#> [1] 1

# compute probability density
pdf(p1, c(0, 1, 2))
#> [1] 0.2419707 0.3989423 0.2419707