API reference

Everything in hyperiax lives behind a small public surface; the sections below list it module by module. Each entry links to a per-symbol page generated from its docstring.

hyperiax.core

Tree topology, immutable Tree pytree, schema, views, sweep decorators.

hyperiax.Topology(parents, size, depth, ...)	Immutable tree topology with precomputed dispatch layouts.
hyperiax.Tree(topology, schema, data)	A topology + a typed dict of (N, *trailing) arrays.
hyperiax.Schema(fields)	Ordered, hashable collection of named FieldSpecs.
hyperiax.FieldSpec(shape[, dtype])	Per-node trailing shape and dtype of a single Tree field.
hyperiax.SweepFn(direction, fn, reads, ...)	A direction-tagged, pure Tree -> Tree transform.
hyperiax.Node(fields)	Sliced per-node fields for one level (or one selected subset).
hyperiax.Children(fields)	View of the children of all parents at the current up-sweep level.
hyperiax.ChildrenAxis(flat, segments, ...)	Virtual children-axis proxy for unequal-degree trees.

hyperiax.up(*[, reads, reads_children, ...])	Decorator: mark a function as an up-sweep.
hyperiax.down(*[, reads, reads_parent])	Decorator: mark a function as a down-sweep.
hyperiax.symmetric_topology(depth, degree)	A regular tree where every internal node has exactly degree children.
hyperiax.from_parents(parents, *[, names])	Thin re-export of Topology.from_parents().
hyperiax.from_newick(source, *[, schema])	Read a Newick tree (literal or file path) into a hyperiax Tree.
hyperiax.to_newick(tree)	Serialize a Tree back to a Newick string.
hyperiax.HyperiaxError	Base class for all hyperiax-raised exceptions.
hyperiax.MissingField	A sweep or tree access references a name not in the tree's schema.
hyperiax.SchemaMismatch	A value's shape or dtype does not match the field's spec.

hyperiax.prebuilt

Ready-to-use sweeps and helpers for common message-passing tasks.

Phylogenetic weighted mean

hyperiax.prebuilt.phylo_mean()

Return a SweepFn that fills each non-leaf node's estimated_value with the edge-length-weighted average of its children's estimated_value.

BFFG — discrete-edge (linear-Gaussian) sweeps

hyperiax.prebuilt.bffg.discrete_schema(d)	Field layout for a discrete-edge BFFG tree.
hyperiax.prebuilt.bffg.init_discrete_tree(...)	Seed a discrete-edge BFFG tree with leaf observations and initial anchors.
hyperiax.prebuilt.bffg.discrete_bf_sweep(...)	Build the discrete-edge backward-filtering up-sweep (Theorem 14 §6.1).
hyperiax.prebuilt.bffg.discrete_forward_sweep(...)	Build the unconditional forward-sampling down-sweep.
hyperiax.prebuilt.bffg.discrete_fg_sweep(...)	Build the discrete-edge forward-guided down-sweep (Theorem 14 §6.1).
hyperiax.prebuilt.bffg.discrete_refine_anchor()	Build the discrete anchor-refinement down-sweep (Algorithm 3 §7.1).

BFFG — continuous-edge (SDE) sweeps

hyperiax.prebuilt.bffg.continuous_schema(d, ...)	Field layout for a continuous-edge (SDE) BFFG tree.
hyperiax.prebuilt.bffg.init_continuous_tree(...)	Seed a continuous-edge BFFG tree with leaf observations and initial anchors.
hyperiax.prebuilt.bffg.continuous_bf_sweep(...)	Build the continuous-edge backward-filtering up-sweep (Theorem 23 §7.1).
hyperiax.prebuilt.bffg.continuous_forward_sweep(...)	Build the unconditional SDE forward-sampling down-sweep.
hyperiax.prebuilt.bffg.continuous_fg_sweep(...)	Build the continuous-edge forward-guided down-sweep (Theorem 23 §7.1).
hyperiax.prebuilt.bffg.continuous_refine_anchor()	Build the continuous anchor-refinement down-sweep (Algorithm 3 §7.1).
hyperiax.prebuilt.bffg.propagate_linearization(tree)	Propagate continuous BFFG linearisation anchors top-down.

For MCMC over BFFG-guided latents and hyperparameters, hyperiax composes with NumPyro — see Gaussian MCMC — BFFG with a hand-rolled kernel and Gaussian NUTS — and what its existence proves for worked examples.