Reference / Defaults

Defaults

This page lists the thresholds and run-time defaults that the code sets as named constants; each function’s parameter defaults are in its signature on the API pages. The tables are generated from the code. Every engine takes rng=None, which draws fresh entropy from the operating system: pass a seed for reproducible draws. Names that start with an underscore are internal: they are listed to explain behaviour, not as settings to change.

Automatic engine choice

Name	Value	Module	What it controls
EXTEND_TARGET	1.2	binference.inference	An extension of a Gibbs run aims for this many times target_ess.
EXTEND_MAX_FACTOR	3	binference.inference	One extension of a Gibbs run adds at most this many times the draws so far.
EXTEND_MIN_FACTOR	0.5	binference.inference	One extension of a Gibbs run adds at least this many times the draws so far.
PROBE_PROPOSALS	1000	binference.inference	Proposals of the first pass that can stop a Laplace trial early.
PROBE_DRAWS	500	binference.inference	Draws of the first pass that can stop a Laplace trial early.
PROBE_ESS_SHARE	0.5	binference.inference	A Laplace trial stops when its first pass projects an importance ESS below this share of target_ess.

Results and diagnostics

Name	Value	Module	What it controls
TAIL_QUANTILES	(0.05, 0.95)	binference.diagnostics	Quantiles whose indicators define the tail ESS.
K_HAT_MAX	0.7	binference.diagnostics	Pareto k-hat at or above which importance sampling is unreliable, reached from about 2,000 weights; with n weights the limit is min(1 - 1/log10(n), K_HAT_MAX).
PSIS_MIN_WEIGHTS	5	binference.diagnostics	Pareto smoothing needs at least this many importance weights.
PSIS_TAIL	(0.2, 3)	binference.diagnostics	The generalised Pareto is fitted to the ceil(min(a n, b sqrt(n))) largest of n weights, for (a, b) these.
N_DRAWS	500	binference.results	Population draws the per-object methods use by default.
RHAT_MAX	1.01	binference.results	Largest split R-hat that PopulationResult.check accepts by default.
ESS_MIN_PER_CHAIN	100	binference.results	Smallest bulk and tail ESS per chain that PopulationResult.check accepts by default.
MIXING_MIN_DRAWS	8	binference.results	Chains shorter than this many draws get no R-hat or ESS.
EBFMI_MIN	0.3	binference.results	PopulationResult.check warns when a NUTS chain’s E-BFMI is below this.
MAX_DEPTH_SHARE	0.05	binference.results	PopulationResult.check warns when a larger share of NUTS iterations hit max_depth.
SAMPLE_VARIANCE_MAX	1	binference.results	PopulationResult.check warns when the summed Monte Carlo variance of likelihoods from samples exceeds this.
EVAL_BUDGET	16777216	binference.results	Most values a density function returns per call when values are drawn (2**24 floats are 128 MB).

Engines

Name	Value	Module	What it controls
_INIT_BUFFER	75	binference.samplers._nuts_core	Iterations of the first fast phase of Stan-style warm-up.
_BASE_WINDOW	25	binference.samplers._nuts_core	Iterations of the first slow window of Stan-style warm-up; each next one is twice as long.
_TERM_BUFFER	50	binference.samplers._nuts_core	Iterations of the last fast phase of Stan-style warm-up.
INTERIOR_START_ITER	50	binference.samplers.base	Most regularised EM steps of the start of Gibbs and of MAP’s EM and L-BFGS paths when init is None.
PROPOSAL_FACTOR	2	binference.samplers.laplace	Importance proposals per draw when n_proposals is None.
EIGEN_FLOOR	1e-08	binference.samplers.laplace	Eigenvalues of a Hessian that is not positive definite are raised to this times the largest (at least one).
_TARGET_ACCEPT	0.8	binference.samplers.nuts	target_accept when it is None, except for logistic priors.
_TARGET_ACCEPT_LOGISTIC	0.9	binference.samplers.nuts	target_accept for logistic priors when it is None.
_DENSE_MAX_DIM	200	binference.samplers.nuts	metric="auto" is dense for logistic priors up to this many dimensions, and up to this many the warm-up curvature is averaged over every chain’s point, not taken at one.
_HESSIAN_MAX_DIM	500	binference.samplers.nuts	Up to this many dimensions, warm-up starts the metric from the curvature at the mode.
_CENTRING_MAX_SD	0.28	binference.samplers.nuts	Logistic priors stay non-centred when the median spread of the modes’ log-scales given the field exceeds this.
_GEOMETRY_ITER	25	binference.samplers.nuts	Warm-up iteration of the partial centring of logistic priors, or the first slow window’s start if that is earlier.
_DUAL_AVERAGING	(10, 0.05, 0.75)	binference.samplers.nuts	t0, gamma and kappa of the dual-averaging step-size adaptation.
_ROUNDOFF	1000	binference.samplers.optimize	A MAP run that can no longer improve its objective in floating point is converged when its optimality gap is below this many times tol.
_MIXSQP_START_ITER	10	binference.samplers.optimize	Regularised EM steps of mix-SQP’s start when init is None.

Priors

Name	Value	Module	What it controls
_TINY	1e-300	binference.priors.base	Floor of drawn fractions and rates, so that their logarithms stay finite.
_KERNELS	{'rbf': (None, 1.3), 'matern52': (2.5, 1.8), 'matern32': (1.5, 2.2), 'matern12': (0.5, 8.5)}	binference.priors.logistic	Smoothness nu (None for the RBF) and resolution factor f of each kernel of LogisticGP: m sines resolve length scales down to f times the box half-width over m.
_MAX_BASIS	1024	binference.priors.logistic	Most basis functions of LogisticGP in all when n_basis is None.

Inputs

Name	Value	Module	What it controls
DEFAULT_CHUNK	4096	binference._kernels	Objects per block of the chunked likelihood kernels.
MIN_EVIDENCE	1e-250	binference._kernels	An object whose normalised evidence is below this is impossible.
SPARSE_FILL	0.1	binference._kernels	Likelihoods are stored sparse below this share of non-zero entries.
MAX_ELEMENTS	4194304	binference._quadrature	Most function values a quadrature evaluates in one call.
DEFAULT_OVERSAMPLE	{1: 8, 2: 4}	binference.basis	Quadrature cells per order-1 component and axis when oversample is None, by dimension (2 above two dimensions; order 0 takes 1).
OUTSIDE_THRESHOLD	0.9	binference.likelihoods	An object with more than this share of its likelihood outside the bins is mostly outside.
OUTSIDE_SHARE	0.1	binference.likelihoods	A warning is raised when more than this share of the objects are mostly outside the bins.
FEW_SAMPLES_SHARE	0.1	binference.likelihoods	from_samples warns when more than this share of the objects have fewer than min_samples effective samples.
REPRESENTATION_SHARE	0.5	binference.likelihoods	A warning is raised when more than this share of posterior rows sum to one only under the other representation.
REPRESENTATION_RTOL	0.01	binference.likelihoods	Relative tolerance within which a posterior row counts as summing to one.
MIGRATION_ATOL	0.001	binference.response	With efficiency, a warning is raised unless the matrix’s columns sum to one within this.

Selection

Name	Value	Module	What it controls
STEEP_RATIO	2	binference.selection	A selection that changes by more than this factor between neighbouring bins raises a warning.
S_FLOOR	1e-12	binference.selection	Smallest detection probability a draw of an inferred selection takes.

Validation

Name	Value	Module	What it controls
SBC_MAX_BINS	20	binference.validation	Most bins of an SBC rank histogram when n_bins is None.
SBC_SIMULATIONS_PER_BIN	5	binference.validation	Fewest simulations per bin of an SBC rank histogram when n_bins is None.