Package index • ripr

Spaces

A space is a set a measure can live on. Every space says how many coordinates one of its elements carries and whether a given point is one of them; a finite one can also list them all.

space(): Spaces
space_dim(): The dimension of one element of a space
contains(): Does a point belong to a space?
enumerate_space(): Every element of a finite sample space
is_finite_space(): Is this sample space finite?
count_space(): The space of k-category count vectors summing to n_trials

Observation models

A parametric_family owns the model p_theta(x): a compiled log-likelihood, score and sampler over a sample space.

parametric_family(): Parametric families
multinomial_family(): Multinomial sampling family
gaussian_family(): Gaussian sampling family with known covariance
compile_loglik(): Compile the log-likelihood function for a fixed set of outcomes
score(): Score d log P_theta(x) / d theta
kernel_draw(): Draw one observation from P_theta per parameter

Regions

Subsets of the real numbers that carry convex geometry. A convex_region can be charted, projected onto, and searched by an optimiser; a union_region is a finite union of them and need not be convex, and its parts may overlap. Every region answers parts() for what it was declared as and cells() for a triangulation of its parts.

region(): Regions of a parameter space
convex_region(): Convex regions
union_region(): A finite union of convex regions
empty_region(): The region with nothing in it
polyhedron_region(): A convex polyhedron given by its generators
h_region(): A convex polyhedron given by its half-space description
polytope_region(): Convex hull of a set of vertices
simplex_region(): A simplex: an affinely independent vertex set
halfspace_region(): Halfspace given by a normal and an offset
point_region(): A single parameter point
real_region(): The whole of R^d as a region
chart(): Coordinate chart for a parameter space
project(): Euclidean projection onto a parameter space
parts(): The convex regions a region was declared as
cells(): The convex regions a region decomposes into for optimisation
is_empty(): Is a region empty?
is_bounded(): Is a region bounded?

Set algebra

Operations that build new regions from old ones, and decide questions about them, composed in exact rational arithmetic and converted to doubles once per chain.

region_algebra |,ripr::region,ripr::region-method &,ripr::region,ripr::region-method -,ripr::region,ripr::region-method ==,ripr::region,ripr::region-method !=,ripr::region,ripr::region-method: Set algebra on regions
disjoin(): Transform a region's parts into a disjoint cover of the union

Null hypotheses

A null is a family together with a finite union of convex parameter regions.

null_model(): A null hypothesis: a family together with its parameter region

Distributions

A distribution is a law over a space. Pushing one over a parameter space through a family’s kernel induces a mixture distribution over the family’s sample space.

distribution(): Distributions over a space
log_density(): Log density of a distribution
draw(): Draw outcomes from a distribution
family.distribution: The family a distribution was induced from, if any
mixture(): The distribution P_W induced by a mixing measure W and a family
discrete_dist(): Distributions with finite support
dirac(): A point mass at a single value
finite_dist(): A distribution on finitely many atoms, sum_c w_c delta_{theta_c}
continuous_dist(): Continuous distributions
gaussian_dist(): Multivariate Gaussian distribution
dirichlet() truncated_dirichlet() dirichlet_dist(): Dirichlet priors over the simplex
n_atoms(): Number of atoms in a distribution
atoms(): Atoms of a distribution
weights.distribution weights,ripr::continuous_dist-method: Weights of a distribution
discretise(): Replace a distribution with the empirical distribution of draws from it
prune(): Drop atoms below a weight threshold and renormalise

Random variables

Callable variables on a sample space, with arithmetic for building likelihood ratios.

random_variable(): A random variable on a sample space
likelihood(): The likelihood of a distribution, as a random variable
random_variable_arithmetic: Arithmetic on random variables and scalars

Quadrature engines

Expectations under the alternative are approximated by a quadrature rule: exact enumeration, Monte Carlo, or Gauss-Hermite.

exact_engine(): Exact enumeration over a finite sample space
mc_engine(): Monte Carlo integration against draws from the alternative
gh_engine(): Gauss-Hermite quadrature against a Gaussian alternative

Fitting the RIPr

ripr_init() builds a starting state; the step verbs advance it by Frank–Wolfe, Li–Barron and/or EM; ripr_finish() turns it into a distribution.

ripr_control(): Tuning for the RIPr optimiser
ripr_state(): State of a RIPr fit
ripr_init(): Begin a RIPr fit
oracles: KL Minimisation Oracles
fw_step(): Frank–Wolfe step
lb_step(): Li–Barron greedy step
em_step(): EM sweep
weight_step(): Fully-corrective weight step
ct_step(): Csiszar–Tusnady step
sfw_step(): Sliding Frank–Wolfe step
kl_flat() gap_below() support_gap_below(): Predicates for early stopping
ripr_finish(): Turn a fitted state into a mixture
ripr_fit(): A finished RIPr fit

Certification

Proving (or searching for) an upper bound on the null expectation of a random variable, to correct a fitted ratio into a genuine e-variable.

certify(): Certify an upper bound on the largest null expectation
ripr_certificate(): A certified upper bound on the largest null expectation
e_variable(): An e-variable from a certified bound
certify_trace(): Record how a certification ran, for inspection and plotting
sup_lb(): Estimate the largest null expectation of a random variable
ripr_search(): A searched lower bound on the largest null expectation