bartz 0.7.0__py3-none-any.whl → 0.8.0__py3-none-any.whl

bartz/mcmcstep/_moves.py

@@ -0,0 +1,904 @@
+# bartz/src/bartz/mcmcstep/_moves.py
+#
+# Copyright (c) 2024-2026, The Bartz Contributors
+#
+# This file is part of bartz.
+#
+# Permission is hereby granted, free of charge, to any person obtaining a copy
+# of this software and associated documentation files (the "Software"), to deal
+# in the Software without restriction, including without limitation the rights
+# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+# copies of the Software, and to permit persons to whom the Software is
+# furnished to do so, subject to the following conditions:
+#
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+#
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+# SOFTWARE.
+
+"""Implement `propose_moves` and associated dataclasses."""
+
+from functools import partial
+
+import jax
+from equinox import Module
+from jax import numpy as jnp
+from jax import random
+from jaxtyping import Array, Bool, Float32, Int32, Integer, Key, UInt
+
+from bartz import grove
+from bartz._profiler import jit_and_block_if_profiling
+from bartz.jaxext import minimal_unsigned_dtype, split, vmap_nodoc
+from bartz.mcmcstep._state import Forest, field, vmap_chains
+
+
+class Moves(Module):
+    """Moves proposed to modify each tree."""
+
+    allowed: Bool[Array, '*chains num_trees'] = field(chains=True)
+    """Whether there is a possible move. If `False`, the other values may not
+    make sense. The only case in which a move is marked as allowed but is
+    then vetoed is if it does not satisfy `min_points_per_leaf`, which for
+    efficiency is implemented post-hoc without changing the rest of the
+    MCMC logic."""
+
+    grow: Bool[Array, '*chains num_trees'] = field(chains=True)
+    """Whether the move is a grow move or a prune move."""
+
+    num_growable: UInt[Array, '*chains num_trees'] = field(chains=True)
+    """The number of growable leaves in the original tree."""
+
+    node: UInt[Array, '*chains num_trees'] = field(chains=True)
+    """The index of the leaf to grow or node to prune."""
+
+    left: UInt[Array, '*chains num_trees'] = field(chains=True)
+    """The index of the left child of 'node'."""
+
+    right: UInt[Array, '*chains num_trees'] = field(chains=True)
+    """The index of the right child of 'node'."""
+
+    partial_ratio: Float32[Array, '*chains num_trees'] | None = field(chains=True)
+    """A factor of the Metropolis-Hastings ratio of the move. It lacks the
+    likelihood ratio, the probability of proposing the prune move, and the
+    probability that the children of the modified node are terminal. If the
+    move is PRUNE, the ratio is inverted. `None` once
+    `log_trans_prior_ratio` has been computed."""
+
+    log_trans_prior_ratio: None | Float32[Array, '*chains num_trees'] = field(
+        chains=True
+    )
+    """The logarithm of the product of the transition and prior terms of the
+    Metropolis-Hastings ratio for the acceptance of the proposed move.
+    `None` if not yet computed. If PRUNE, the log-ratio is negated."""
+
+    grow_var: UInt[Array, '*chains num_trees'] = field(chains=True)
+    """The decision axes of the new rules."""
+
+    grow_split: UInt[Array, '*chains num_trees'] = field(chains=True)
+    """The decision boundaries of the new rules."""
+
+    var_tree: UInt[Array, '*chains num_trees 2**(d-1)'] = field(chains=True)
+    """The updated decision axes of the trees, valid whatever move."""
+
+    affluence_tree: Bool[Array, '*chains num_trees 2**(d-1)'] = field(chains=True)
+    """A partially updated `affluence_tree`, marking non-leaf nodes that would
+    become leaves if the move was accepted. This mark initially (out of
+    `propose_moves`) takes into account if there would be available decision
+    rules to grow the leaf, and whether there are enough datapoints in the
+    node is instead checked later in `accept_moves_parallel_stage`."""
+
+    logu: Float32[Array, '*chains num_trees'] = field(chains=True)
+    """The logarithm of a uniform (0, 1] random variable to be used to
+    accept the move. It's in (-oo, 0]."""
+
+    acc: None | Bool[Array, '*chains num_trees'] = field(chains=True)
+    """Whether the move was accepted. `None` if not yet computed."""
+
+    to_prune: None | Bool[Array, '*chains num_trees'] = field(chains=True)
+    """Whether the final operation to apply the move is pruning. This indicates
+    an accepted prune move or a rejected grow move. `None` if not yet
+    computed."""
+
+
+@jit_and_block_if_profiling
+@vmap_chains
+def propose_moves(key: Key[Array, ''], forest: Forest) -> Moves:
+    """
+    Propose moves for all the trees.
+
+    There are two types of moves: GROW (convert a leaf to a decision node and
+    add two leaves beneath it) and PRUNE (convert the parent of two leaves to a
+    leaf, deleting its children).
+
+    Parameters
+    ----------
+    key
+        A jax random key.
+    forest
+        The `forest` field of a BART MCMC state.
+
+    Returns
+    -------
+    The proposed move for each tree.
+    """
+    num_trees = forest.leaf_tree.shape[0]
+    keys = split(key, 2)
+    grow_keys, prune_keys = keys.pop((2, num_trees))
+
+    # compute moves
+    grow_moves = propose_grow_moves(
+        grow_keys,
+        forest.var_tree,
+        forest.split_tree,
+        forest.affluence_tree,
+        forest.max_split,
+        forest.blocked_vars,
+        forest.p_nonterminal,
+        forest.p_propose_grow,
+        forest.log_s,
+    )
+    prune_moves = propose_prune_moves(
+        prune_keys,
+        forest.split_tree,
+        grow_moves.affluence_tree,
+        forest.p_nonterminal,
+        forest.p_propose_grow,
+    )
+
+    u, exp1mlogu = random.uniform(keys.pop(), (2, num_trees))
+
+    # choose between grow or prune
+    p_grow = jnp.where(
+        grow_moves.allowed & prune_moves.allowed, 0.5, grow_moves.allowed
+    )
+    grow = u < p_grow  # use < instead of <= because u is in [0, 1)
+
+    # compute children indices
+    node = jnp.where(grow, grow_moves.node, prune_moves.node)
+    left, right = (node << 1) | jnp.arange(2)[:, None]
+
+    return Moves(
+        allowed=grow_moves.allowed | prune_moves.allowed,
+        grow=grow,
+        num_growable=grow_moves.num_growable,
+        node=node,
+        left=left,
+        right=right,
+        partial_ratio=jnp.where(
+            grow, grow_moves.partial_ratio, prune_moves.partial_ratio
+        ),
+        log_trans_prior_ratio=None,  # will be set in complete_ratio
+        grow_var=grow_moves.var,
+        grow_split=grow_moves.split,
+        # var_tree does not need to be updated if prune
+        var_tree=grow_moves.var_tree,
+        # affluence_tree is updated for both moves unconditionally, prune last
+        affluence_tree=prune_moves.affluence_tree,
+        logu=jnp.log1p(-exp1mlogu),
+        acc=None,  # will be set in accept_moves_sequential_stage
+        to_prune=None,  # will be set in accept_moves_sequential_stage
+    )
+
+
+class GrowMoves(Module):
+    """Represent a proposed grow move for each tree."""
+
+    allowed: Bool[Array, ' num_trees']
+    """Whether the move is allowed for proposal."""
+
+    num_growable: UInt[Array, ' num_trees']
+    """The number of leaves that can be proposed for grow."""
+
+    node: UInt[Array, ' num_trees']
+    """The index of the leaf to grow. ``2 ** d`` if there are no growable
+    leaves."""
+
+    var: UInt[Array, ' num_trees']
+    """The decision axis of the new rule."""
+
+    split: UInt[Array, ' num_trees']
+    """The decision boundary of the new rule."""
+
+    partial_ratio: Float32[Array, ' num_trees']
+    """A factor of the Metropolis-Hastings ratio of the move. It lacks
+    the likelihood ratio and the probability of proposing the prune
+    move."""
+
+    var_tree: UInt[Array, 'num_trees 2**(d-1)']
+    """The updated decision axes of the tree."""
+
+    affluence_tree: Bool[Array, 'num_trees 2**(d-1)']
+    """A partially updated `affluence_tree` that marks each new leaf that
+    would be produced as `True` if it would have available decision rules."""
+
+
+@partial(vmap_nodoc, in_axes=(0, 0, 0, 0, None, None, None, None, None))
+def propose_grow_moves(
+    key: Key[Array, ' num_trees'],
+    var_tree: UInt[Array, 'num_trees 2**(d-1)'],
+    split_tree: UInt[Array, 'num_trees 2**(d-1)'],
+    affluence_tree: Bool[Array, 'num_trees 2**(d-1)'],
+    max_split: UInt[Array, ' p'],
+    blocked_vars: Int32[Array, ' k'] | None,
+    p_nonterminal: Float32[Array, ' 2**d'],
+    p_propose_grow: Float32[Array, ' 2**(d-1)'],
+    log_s: Float32[Array, ' p'] | None,
+) -> GrowMoves:
+    """
+    Propose a GROW move for each tree.
+
+    A GROW move picks a leaf node and converts it to a non-terminal node with
+    two leaf children.
+
+    Parameters
+    ----------
+    key
+        A jax random key.
+    var_tree
+        The splitting axes of the tree.
+    split_tree
+        The splitting points of the tree.
+    affluence_tree
+        Whether each leaf has enough points to be grown.
+    max_split
+        The maximum split index for each variable.
+    blocked_vars
+        The indices of the variables that have no available cutpoints.
+    p_nonterminal
+        The a priori probability of a node to be nonterminal conditional on the
+        ancestors, including at the maximum depth where it should be zero.
+    p_propose_grow
+        The unnormalized probability of choosing a leaf to grow.
+    log_s
+        Unnormalized log-probability used to choose a variable to split on
+        amongst the available ones.
+
+    Returns
+    -------
+    An object representing the proposed move.
+
+    Notes
+    -----
+    The move is not proposed if each leaf is already at maximum depth, or has
+    less datapoints than the requested threshold `min_points_per_decision_node`,
+    or it does not have any available decision rules given its ancestors. This
+    is marked by setting `allowed` to `False` and `num_growable` to 0.
+    """
+    keys = split(key, 3)
+
+    leaf_to_grow, num_growable, prob_choose, num_prunable = choose_leaf(
+        keys.pop(), split_tree, affluence_tree, p_propose_grow
+    )
+
+    # sample a decision rule
+    var, num_available_var = choose_variable(
+        keys.pop(), var_tree, split_tree, max_split, leaf_to_grow, blocked_vars, log_s
+    )
+    split_idx, l, r = choose_split(
+        keys.pop(), var, var_tree, split_tree, max_split, leaf_to_grow
+    )
+
+    # determine if the new leaves would have available decision rules; if the
+    # move is blocked, these values may not make sense
+    leftright_growable = (num_available_var > 1) | jnp.stack(
+        [l < split_idx, split_idx + 1 < r]
+    )
+    leftright = (leaf_to_grow << 1) | jnp.arange(2)
+    affluence_tree = affluence_tree.at[leftright].set(leftright_growable)
+
+    ratio = compute_partial_ratio(
+        prob_choose, num_prunable, p_nonterminal, leaf_to_grow
+    )
+
+    return GrowMoves(
+        allowed=num_growable > 0,
+        num_growable=num_growable,
+        node=leaf_to_grow,
+        var=var,
+        split=split_idx,
+        partial_ratio=ratio,
+        var_tree=var_tree.at[leaf_to_grow].set(var.astype(var_tree.dtype)),
+        affluence_tree=affluence_tree,
+    )
+
+
+def choose_leaf(
+    key: Key[Array, ''],
+    split_tree: UInt[Array, ' 2**(d-1)'],
+    affluence_tree: Bool[Array, ' 2**(d-1)'],
+    p_propose_grow: Float32[Array, ' 2**(d-1)'],
+) -> tuple[Int32[Array, ''], Int32[Array, ''], Float32[Array, ''], Int32[Array, '']]:
+    """
+    Choose a leaf node to grow in a tree.
+
+    Parameters
+    ----------
+    key
+        A jax random key.
+    split_tree
+        The splitting points of the tree.
+    affluence_tree
+        Whether a leaf has enough points that it could be split into two leaves
+        satisfying the `min_points_per_decision_node` requirement.
+    p_propose_grow
+        The unnormalized probability of choosing a leaf to grow.
+
+    Returns
+    -------
+    leaf_to_grow : Int32[Array, '']
+        The index of the leaf to grow. If ``num_growable == 0``, return
+        ``2 ** d``.
+    num_growable : Int32[Array, '']
+        The number of leaf nodes that can be grown, i.e., are nonterminal
+        and have at least twice `min_points_per_decision_node`.
+    prob_choose : Float32[Array, '']
+        The (normalized) probability that this function had to choose that
+        specific leaf, given the arguments.
+    num_prunable : Int32[Array, '']
+        The number of leaf parents that could be pruned, after converting the
+        selected leaf to a non-terminal node.
+    """
+    is_growable = growable_leaves(split_tree, affluence_tree)
+    num_growable = jnp.count_nonzero(is_growable)
+    distr = jnp.where(is_growable, p_propose_grow, 0)
+    leaf_to_grow, distr_norm = categorical(key, distr)
+    leaf_to_grow = jnp.where(num_growable, leaf_to_grow, 2 * split_tree.size)
+    prob_choose = distr[leaf_to_grow] / jnp.where(distr_norm, distr_norm, 1)
+    is_parent = grove.is_leaves_parent(split_tree.at[leaf_to_grow].set(1))
+    num_prunable = jnp.count_nonzero(is_parent)
+    return leaf_to_grow, num_growable, prob_choose, num_prunable
+
+
+def growable_leaves(
+    split_tree: UInt[Array, ' 2**(d-1)'], affluence_tree: Bool[Array, ' 2**(d-1)']
+) -> Bool[Array, ' 2**(d-1)']:
+    """
+    Return a mask indicating the leaf nodes that can be proposed for growth.
+
+    The condition is that a leaf is not at the bottom level, has available
+    decision rules given its ancestors, and has at least
+    `min_points_per_decision_node` points.
+
+    Parameters
+    ----------
+    split_tree
+        The splitting points of the tree.
+    affluence_tree
+        Marks leaves that can be grown.
+
+    Returns
+    -------
+    The mask indicating the leaf nodes that can be proposed to grow.
+
+    Notes
+    -----
+    This function needs `split_tree` and not just `affluence_tree` because
+    `affluence_tree` can be "dirty", i.e., mark unused nodes as `True`.
+    """
+    return grove.is_actual_leaf(split_tree) & affluence_tree
+
+
+def categorical(
+    key: Key[Array, ''], distr: Float32[Array, ' n']
+) -> tuple[Int32[Array, ''], Float32[Array, '']]:
+    """
+    Return a random integer from an arbitrary distribution.
+
+    Parameters
+    ----------
+    key
+        A jax random key.
+    distr
+        An unnormalized probability distribution.
+
+    Returns
+    -------
+    u : Int32[Array, '']
+        A random integer in the range ``[0, n)``. If all probabilities are zero,
+        return ``n``.
+    norm : Float32[Array, '']
+        The sum of `distr`.
+
+    Notes
+    -----
+    This function uses a cumsum instead of the Gumbel trick, so it's ok only
+    for small ranges with probabilities well greater than 0.
+    """
+    ecdf = jnp.cumsum(distr)
+    u = random.uniform(key, (), ecdf.dtype, 0, ecdf[-1])
+    return jnp.searchsorted(ecdf, u, 'right', method='compare_all'), ecdf[-1]
+
+
+def choose_variable(
+    key: Key[Array, ''],
+    var_tree: UInt[Array, ' 2**(d-1)'],
+    split_tree: UInt[Array, ' 2**(d-1)'],
+    max_split: UInt[Array, ' p'],
+    leaf_index: Int32[Array, ''],
+    blocked_vars: Int32[Array, ' k'] | None,
+    log_s: Float32[Array, ' p'] | None,
+) -> tuple[Int32[Array, ''], Int32[Array, '']]:
+    """
+    Choose a variable to split on for a new non-terminal node.
+
+    Parameters
+    ----------
+    key
+        A jax random key.
+    var_tree
+        The variable indices of the tree.
+    split_tree
+        The splitting points of the tree.
+    max_split
+        The maximum split index for each variable.
+    leaf_index
+        The index of the leaf to grow.
+    blocked_vars
+        The indices of the variables that have no available cutpoints. If
+        `None`, all variables are assumed unblocked.
+    log_s
+        The logarithm of the prior probability for choosing a variable. If
+        `None`, use a uniform distribution.
+
+    Returns
+    -------
+    var : Int32[Array, '']
+        The index of the variable to split on.
+    num_available_var : Int32[Array, '']
+        The number of variables with available decision rules `var` was chosen
+        from.
+    """
+    var_to_ignore = fully_used_variables(var_tree, split_tree, max_split, leaf_index)
+    if blocked_vars is not None:
+        var_to_ignore = jnp.concatenate([var_to_ignore, blocked_vars])
+
+    if log_s is None:
+        return randint_exclude(key, max_split.size, var_to_ignore)
+    else:
+        return categorical_exclude(key, log_s, var_to_ignore)
+
+
+def fully_used_variables(
+    var_tree: UInt[Array, ' 2**(d-1)'],
+    split_tree: UInt[Array, ' 2**(d-1)'],
+    max_split: UInt[Array, ' p'],
+    leaf_index: Int32[Array, ''],
+) -> UInt[Array, ' d-2']:
+    """
+    Find variables in the ancestors of a node that have an empty split range.
+
+    Parameters
+    ----------
+    var_tree
+        The variable indices of the tree.
+    split_tree
+        The splitting points of the tree.
+    max_split
+        The maximum split index for each variable.
+    leaf_index
+        The index of the node, assumed to be valid for `var_tree`.
+
+    Returns
+    -------
+    The indices of the variables that have an empty split range.
+
+    Notes
+    -----
+    The number of unused variables is not known in advance. Unused values in the
+    array are filled with `p`. The fill values are not guaranteed to be placed
+    in any particular order, and variables may appear more than once.
+    """
+    var_to_ignore = ancestor_variables(var_tree, max_split, leaf_index)
+    split_range_vec = jax.vmap(split_range, in_axes=(None, None, None, None, 0))
+    l, r = split_range_vec(var_tree, split_tree, max_split, leaf_index, var_to_ignore)
+    num_split = r - l
+    return jnp.where(num_split == 0, var_to_ignore, max_split.size)
+    # the type of var_to_ignore is already sufficient to hold max_split.size,
+    # see ancestor_variables()
+
+
+def ancestor_variables(
+    var_tree: UInt[Array, ' 2**(d-1)'],
+    max_split: UInt[Array, ' p'],
+    node_index: Int32[Array, ''],
+) -> UInt[Array, ' d-2']:
+    """
+    Return the list of variables in the ancestors of a node.
+
+    Parameters
+    ----------
+    var_tree
+        The variable indices of the tree.
+    max_split
+        The maximum split index for each variable. Used only to get `p`.
+    node_index
+        The index of the node, assumed to be valid for `var_tree`.
+
+    Returns
+    -------
+    The variable indices of the ancestors of the node.
+
+    Notes
+    -----
+    The ancestors are the nodes going from the root to the parent of the node.
+    The number of ancestors is not known at tracing time; unused spots in the
+    output array are filled with `p`.
+    """
+    max_num_ancestors = grove.tree_depth(var_tree) - 1
+    index = node_index >> jnp.arange(max_num_ancestors, 0, -1)
+    var = var_tree[index]
+    var_type = minimal_unsigned_dtype(max_split.size)
+    p = jnp.array(max_split.size, var_type)
+    return jnp.where(index, var, p)
+
+
+def split_range(
+    var_tree: UInt[Array, ' 2**(d-1)'],
+    split_tree: UInt[Array, ' 2**(d-1)'],
+    max_split: UInt[Array, ' p'],
+    node_index: Int32[Array, ''],
+    ref_var: Int32[Array, ''],
+) -> tuple[Int32[Array, ''], Int32[Array, '']]:
+    """
+    Return the range of allowed splits for a variable at a given node.
+
+    Parameters
+    ----------
+    var_tree
+        The variable indices of the tree.
+    split_tree
+        The splitting points of the tree.
+    max_split
+        The maximum split index for each variable.
+    node_index
+        The index of the node, assumed to be valid for `var_tree`.
+    ref_var
+        The variable for which to measure the split range.
+
+    Returns
+    -------
+    The range of allowed splits as [l, r). If `ref_var` is out of bounds, l=r=1.
+    """
+    max_num_ancestors = grove.tree_depth(var_tree) - 1
+    index = node_index >> jnp.arange(max_num_ancestors)
+    right_child = (index & 1).astype(bool)
+    index >>= 1
+    split = split_tree[index].astype(jnp.int32)
+    cond = (var_tree[index] == ref_var) & index.astype(bool)
+    l = jnp.max(split, initial=0, where=cond & right_child)
+    initial_r = 1 + max_split.at[ref_var].get(mode='fill', fill_value=0).astype(
+        jnp.int32
+    )
+    r = jnp.min(split, initial=initial_r, where=cond & ~right_child)
+
+    return l + 1, r
+
+
+def randint_exclude(
+    key: Key[Array, ''], sup: int | Integer[Array, ''], exclude: Integer[Array, ' n']
+) -> tuple[Int32[Array, ''], Int32[Array, '']]:
+    """
+    Return a random integer in a range, excluding some values.
+
+    Parameters
+    ----------
+    key
+        A jax random key.
+    sup
+        The exclusive upper bound of the range, must be >= 1.
+    exclude
+        The values to exclude from the range. Values greater than or equal to
+        `sup` are ignored. Values can appear more than once.
+
+    Returns
+    -------
+    u : Int32[Array, '']
+        A random integer `u` in the range ``[0, sup)`` such that ``u not in
+        exclude``.
+    num_allowed : Int32[Array, '']
+        The number of integers in the range that were not excluded.
+
+    Notes
+    -----
+    If all values in the range are excluded, return `sup`.
+    """
+    exclude, num_allowed = _process_exclude(sup, exclude)
+    u = random.randint(key, (), 0, num_allowed)
+    u_shifted = u + jnp.arange(exclude.size)
+    u_shifted = jnp.minimum(u_shifted, sup - 1)
+    u += jnp.sum(u_shifted >= exclude)
+    return u, num_allowed
+
+
+def _process_exclude(sup, exclude):
+    exclude = jnp.unique(exclude, size=exclude.size, fill_value=sup)
+    num_allowed = sup - jnp.sum(exclude < sup)
+    return exclude, num_allowed
+
+
+def categorical_exclude(
+    key: Key[Array, ''], logits: Float32[Array, ' k'], exclude: Integer[Array, ' n']
+) -> tuple[Int32[Array, ''], Int32[Array, '']]:
+    """
+    Draw from a categorical distribution, excluding a set of values.
+
+    Parameters
+    ----------
+    key
+        A jax random key.
+    logits
+        The unnormalized log-probabilities of each category.
+    exclude
+        The values to exclude from the range [0, k). Values greater than or
+        equal to `logits.size` are ignored. Values can appear more than once.
+
+    Returns
+    -------
+    u : Int32[Array, '']
+        A random integer in the range ``[0, k)`` such that ``u not in exclude``.
+    num_allowed : Int32[Array, '']
+        The number of integers in the range that were not excluded.
+
+    Notes
+    -----
+    If all values in the range are excluded, the result is unspecified.
+    """
+    exclude, num_allowed = _process_exclude(logits.size, exclude)
+    kinda_neg_inf = jnp.finfo(logits.dtype).min
+    logits = logits.at[exclude].set(kinda_neg_inf)
+    u = random.categorical(key, logits)
+    return u, num_allowed
+
+
+def choose_split(
+    key: Key[Array, ''],
+    var: Int32[Array, ''],
+    var_tree: UInt[Array, ' 2**(d-1)'],
+    split_tree: UInt[Array, ' 2**(d-1)'],
+    max_split: UInt[Array, ' p'],
+    leaf_index: Int32[Array, ''],
+) -> tuple[Int32[Array, ''], Int32[Array, ''], Int32[Array, '']]:
+    """
+    Choose a split point for a new non-terminal node.
+
+    Parameters
+    ----------
+    key
+        A jax random key.
+    var
+        The variable to split on.
+    var_tree
+        The splitting axes of the tree. Does not need to already contain `var`
+        at `leaf_index`.
+    split_tree
+        The splitting points of the tree.
+    max_split
+        The maximum split index for each variable.
+    leaf_index
+        The index of the leaf to grow.
+
+    Returns
+    -------
+    split : Int32[Array, '']
+        The cutpoint.
+    l : Int32[Array, '']
+    r : Int32[Array, '']
+        The integer range `split` was drawn from is [l, r).
+
+    Notes
+    -----
+    If `var` is out of bounds, or if the available split range on that variable
+    is empty, return 0.
+    """
+    l, r = split_range(var_tree, split_tree, max_split, leaf_index, var)
+    return jnp.where(l < r, random.randint(key, (), l, r), 0), l, r
+
+
+def compute_partial_ratio(
+    prob_choose: Float32[Array, ''],
+    num_prunable: Int32[Array, ''],
+    p_nonterminal: Float32[Array, ' 2**d'],
+    leaf_to_grow: Int32[Array, ''],
+) -> Float32[Array, '']:
+    """
+    Compute the product of the transition and prior ratios of a grow move.
+
+    Parameters
+    ----------
+    prob_choose
+        The probability that the leaf had to be chosen amongst the growable
+        leaves.
+    num_prunable
+        The number of leaf parents that could be pruned, after converting the
+        leaf to be grown to a non-terminal node.
+    p_nonterminal
+        The a priori probability of each node being nonterminal conditional on
+        its ancestors.
+    leaf_to_grow
+        The index of the leaf to grow.
+
+    Returns
+    -------
+    The partial transition ratio times the prior ratio.
+
+    Notes
+    -----
+    The transition ratio is P(new tree => old tree) / P(old tree => new tree).
+    The "partial" transition ratio returned is missing the factor P(propose
+    prune) in the numerator. The prior ratio is P(new tree) / P(old tree). The
+    "partial" prior ratio is missing the factor P(children are leaves).
+    """
+    # the two ratios also contain factors num_available_split *
+    # num_available_var * s[var], but they cancel out
+
+    # p_prune and 1 - p_nonterminal[child] * I(is the child growable) can't be
+    # computed here because they need the count trees, which are computed in the
+    # acceptance phase
+
+    prune_allowed = leaf_to_grow != 1
+    # prune allowed  <--->  the initial tree is not a root
+    # leaf to grow is root  -->  the tree can only be a root
+    # tree is a root  -->  the only leaf I can grow is root
+    p_grow = jnp.where(prune_allowed, 0.5, 1)
+    inv_trans_ratio = p_grow * prob_choose * num_prunable
+
+    # .at.get because if leaf_to_grow is out of bounds (move not allowed), this
+    # would produce a 0 and then an inf when `complete_ratio` takes the log
+    pnt = p_nonterminal.at[leaf_to_grow].get(mode='fill', fill_value=0.5)
+    tree_ratio = pnt / (1 - pnt)
+
+    return tree_ratio / jnp.where(inv_trans_ratio, inv_trans_ratio, 1)
+
+
+class PruneMoves(Module):
+    """Represent a proposed prune move for each tree."""
+
+    allowed: Bool[Array, ' num_trees']
+    """Whether the move is possible."""
+
+    node: UInt[Array, ' num_trees']
+    """The index of the node to prune. ``2 ** d`` if no node can be pruned."""
+
+    partial_ratio: Float32[Array, ' num_trees']
+    """A factor of the Metropolis-Hastings ratio of the move. It lacks the
+    likelihood ratio, the probability of proposing the prune move, and the
+    prior probability that the children of the node to prune are leaves.
+    This ratio is inverted, and is meant to be inverted back in
+    `accept_move_and_sample_leaves`."""
+
+    affluence_tree: Bool[Array, 'num_trees 2**(d-1)']
+    """A partially updated `affluence_tree`, marking the node to prune as
+    growable."""
+
+
+@partial(vmap_nodoc, in_axes=(0, 0, 0, None, None))
+def propose_prune_moves(
+    key: Key[Array, ''],
+    split_tree: UInt[Array, ' 2**(d-1)'],
+    affluence_tree: Bool[Array, ' 2**(d-1)'],
+    p_nonterminal: Float32[Array, ' 2**d'],
+    p_propose_grow: Float32[Array, ' 2**(d-1)'],
+) -> PruneMoves:
+    """
+    Tree structure prune move proposal of BART MCMC.
+
+    Parameters
+    ----------
+    key
+        A jax random key.
+    split_tree
+        The splitting points of the tree.
+    affluence_tree
+        Whether each leaf can be grown.
+    p_nonterminal
+        The a priori probability of a node to be nonterminal conditional on
+        the ancestors, including at the maximum depth where it should be zero.
+    p_propose_grow
+        The unnormalized probability of choosing a leaf to grow.
+
+    Returns
+    -------
+    An object representing the proposed moves.
+    """
+    node_to_prune, num_prunable, prob_choose, affluence_tree = choose_leaf_parent(
+        key, split_tree, affluence_tree, p_propose_grow
+    )
+
+    ratio = compute_partial_ratio(
+        prob_choose, num_prunable, p_nonterminal, node_to_prune
+    )
+
+    return PruneMoves(
+        allowed=split_tree[1].astype(bool),  # allowed iff the tree is not a root
+        node=node_to_prune,
+        partial_ratio=ratio,
+        affluence_tree=affluence_tree,
+    )
+
+
+def choose_leaf_parent(
+    key: Key[Array, ''],
+    split_tree: UInt[Array, ' 2**(d-1)'],
+    affluence_tree: Bool[Array, ' 2**(d-1)'],
+    p_propose_grow: Float32[Array, ' 2**(d-1)'],
+) -> tuple[
+    Int32[Array, ''],
+    Int32[Array, ''],
+    Float32[Array, ''],
+    Bool[Array, 'num_trees 2**(d-1)'],
+]:
+    """
+    Pick a non-terminal node with leaf children to prune in a tree.
+
+    Parameters
+    ----------
+    key
+        A jax random key.
+    split_tree
+        The splitting points of the tree.
+    affluence_tree
+        Whether a leaf has enough points to be grown.
+    p_propose_grow
+        The unnormalized probability of choosing a leaf to grow.
+
+    Returns
+    -------
+    node_to_prune : Int32[Array, '']
+        The index of the node to prune. If ``num_prunable == 0``, return
+        ``2 ** d``.
+    num_prunable : Int32[Array, '']
+        The number of leaf parents that could be pruned.
+    prob_choose : Float32[Array, '']
+        The (normalized) probability that `choose_leaf` would chose
+        `node_to_prune` as leaf to grow, if passed the tree where
+        `node_to_prune` had been pruned.
+    affluence_tree : Bool[Array, 'num_trees 2**(d-1)']
+        A partially updated `affluence_tree`, marking the node to prune as
+        growable.
+    """
+    # sample a node to prune
+    is_prunable = grove.is_leaves_parent(split_tree)
+    num_prunable = jnp.count_nonzero(is_prunable)
+    node_to_prune = randint_masked(key, is_prunable)
+    node_to_prune = jnp.where(num_prunable, node_to_prune, 2 * split_tree.size)
+
+    # compute stuff for reverse move
+    split_tree = split_tree.at[node_to_prune].set(0)
+    affluence_tree = affluence_tree.at[node_to_prune].set(True)
+    is_growable_leaf = growable_leaves(split_tree, affluence_tree)
+    distr_norm = jnp.sum(p_propose_grow, where=is_growable_leaf)
+    prob_choose = p_propose_grow.at[node_to_prune].get(mode='fill', fill_value=0)
+    prob_choose = prob_choose / jnp.where(distr_norm, distr_norm, 1)
+
+    return node_to_prune, num_prunable, prob_choose, affluence_tree
+
+
+def randint_masked(key: Key[Array, ''], mask: Bool[Array, ' n']) -> Int32[Array, '']:
+    """
+    Return a random integer in a range, including only some values.
+
+    Parameters
+    ----------
+    key
+        A jax random key.
+    mask
+        The mask indicating the allowed values.
+
+    Returns
+    -------
+    A random integer in the range ``[0, n)`` such that ``mask[u] == True``.
+
+    Notes
+    -----
+    If all values in the mask are `False`, return `n`. This function is
+    optimized for small `n`.
+    """
+    ecdf = jnp.cumsum(mask)
+    u = random.randint(key, (), 0, ecdf[-1])
+    return jnp.searchsorted(ecdf, u, 'right', method='compare_all')