compiled-knowledge 4.0.0a20__cp313-cp313-macosx_11_0_arm64.whl

ck/pgm.py

@@ -0,0 +1,3475 @@
+"""
+For more documentation on this module, refer to the Jupyter notebook docs/4_PGM_advanced.ipynb.
+"""
+from __future__ import annotations
+
+import math
+from abc import ABC, abstractmethod
+from dataclasses import dataclass
+from itertools import repeat as _repeat
+from typing import Sequence, Tuple, Dict, Optional, overload, Set, Iterable, List, Union, Callable, \
+    Collection, Any, Iterator
+
+import numpy as np
+
+from ck.utils.iter_extras import (
+    combos_ranges as _combos_ranges, multiply as _multiply, combos as _combos
+)
+from ck.utils.np_extras import NDArrayFloat64, NDArrayUInt8
+
+# What types are permitted as random variable states
+State = Union[int, str, bool, float, None]
+
+# An instance (of a sequence of random variables) is a tuple of integers
+# that are state indexes, co-indexed with a known sequence of random variables.
+Instance = Sequence[int]
+
+# A key identifies an instance.
+# A single integer is treated as an instance with one dimension.
+Key = Union[Sequence[int], int]
+
+# The shape of a sequence of random variables (e.g., a PGM, Factor or PotentialFunction).
+Shape = Sequence[int]
+
+DEFAULT_TOLERANCE: float = 0.000001  # For checking CPT sums.
+
+
+class PGM:
+    """
+    A probabilistic graphical model (PGM) represents a joint probability distribution over
+    a set of random variables. Specifically, a PGM is a factor graph with discrete random variables.
+
+    Add a random variable to a PGM, pgm, using `rv = pgm.new_rv(...)`.
+
+    Add a factor to the PGM, pgm, using `factor = pgm.new_factor(...)`.
+
+    A PGM may be given a human-readable name.
+    """
+
+    def __init__(self, name: Optional[str] = None):
+        """
+        Create an empty PGM.
+
+        Args:
+            name: an optional name for the PGM. If not provided, a default name will be
+                created using `default_pgm_name`.
+        """
+        self._name: str = name if name is not None else default_pgm_name(self)
+        self._rvs: Tuple[RandomVariable, ...] = ()
+        self._shape: Shape = ()
+        self._indicators: Tuple[Indicator, ...] = ()
+        self._factors: Tuple[Factor, ...] = ()
+
+    @property
+    def name(self) -> str:
+        """
+        Returns:
+            The name of the PGM.
+        """
+        return self._name
+
+    @property
+    def number_of_rvs(self) -> int:
+        """
+        Returns:
+            How many random variables are defined in this PGM.
+        """
+        return len(self._rvs)
+
+    @property
+    def shape(self) -> Shape:
+        """
+        Returns:
+            a sequence of the lengths of `self.rvs`.
+        """
+        return self._shape
+
+    @property
+    def number_of_indicators(self) -> int:
+        """
+        Returns:
+            How many indicators are defined in this PGM, i.e., `sum(len(rv) for rv in self.rvs)`.
+        """
+        return len(self._indicators)
+
+    @property
+    def number_of_states(self) -> int:
+        """
+        Returns:
+            What is the size of the state space, i.e., `multiply(len(rv) for rv in self.rvs)`.
+        """
+        return number_of_states(*self._rvs)
+
+    @property
+    def number_of_factors(self) -> int:
+        """
+        Returns:
+            How many factors are defined in this PGM.
+        """
+        return len(self._factors)
+
+    @property
+    def number_of_functions(self) -> int:
+        """
+        Returns:
+            How many potential functions are defined in this PGM, including zero potential functions.
+        """
+        return sum(1 for _ in self.functions)
+
+    @property
+    def number_of_non_zero_functions(self) -> int:
+        """
+        Returns:
+            How many potential functions are defined in this PGM, excluding zero potential functions.
+        """
+        return sum(1 for _ in self.non_zero_functions)
+
+    @property
+    def rvs(self) -> Sequence[RandomVariable]:
+        """
+        Returns:
+            All the random variables, in `idx` order, which is the same as creation order.
+
+        Ensures:
+            `self.rvs[rv.idx] = rv`
+        """
+        return self._rvs
+
+    @property
+    def rv_log_sizes(self) -> Sequence[float]:
+        """
+        Returns:
+            [log2(len(rv)) for rv in self.rvs]
+        """
+        return [math.log2(len(rv)) for rv in self.rvs]
+
+    @property
+    def indicators(self) -> Sequence[Indicator]:
+        """
+        Returns:
+            All the random variable indicators.
+
+        Ensures:
+            the indicators of a random variable are adjacent,
+            the indicators of a random variable are in state index order,
+            the random variables are in the same order as `self.rvs`.
+        """
+        return self._indicators
+
+    @property
+    def factors(self) -> Sequence[Factor]:
+        """
+        Returns:
+            All the factors, in `idx` order, which is the same as creation order.
+
+        Ensures:
+            `self.factors[factor.idx] = factor`
+        """
+        return self._factors
+
+    @property
+    def functions(self) -> Iterable[PotentialFunction]:
+        """
+        Iterate over all in-use potential functions of this PGM, including
+        zero potential functions.
+
+        Returns:
+            An Iterable over all potential functions (including zero potential functions).
+        """
+        seen: Set[int] = set()
+        for factor in self._factors:
+            function = factor.function
+            if id(function) not in seen:
+                seen.add(id(function))
+                yield function
+
+    @property
+    def non_zero_functions(self) -> Iterable[PotentialFunction]:
+        """
+        Iterate over all in-use potential functions of this PGM, excluding
+        zero potential functions.
+
+        Returns:
+            An Iterable over all potential functions (excluding zero potential functions).
+        """
+        seen: Set[int] = set()
+        for factor in self._factors:
+            function = factor.function
+            if not (isinstance(function, ZeroPotentialFunction) or id(function) in seen):
+                seen.add(id(function))
+                yield function
+
+    def new_rv(self, name: str, states: Union[int, Sequence[State]]) -> RandomVariable:
+        """
+        Add a new random variable to this PGM.
+
+        The returned random variable will have an `idx` equal to the value of
+        `self.number_of_rvs` just prior to adding the new random variable.
+
+        Assumes:
+            Provided states contain no duplicates.
+
+        Args:
+            name: a name for the random variable.
+            states: either an integer number of states or a sequence of state values. If a
+                single integer, `n`, is provided then the states will be 0, 1, ..., n-1.
+                If a sequence of states are provided then the states must be unique.
+
+        Returns:
+            a RandomVariable object belonging to this PGM.
+        """
+        return RandomVariable(self, name, states)
+
+    def new_factor(self, *rvs: RandomVariable) -> Factor:
+        """
+        Add a new factor to this PGM where the factor connects
+        the given random variables.
+
+        The returned factor will have a ZeroPotentialFunction as its potential function.
+        The potential function may be changed by calling methods on the returned factor.
+
+        The returned factor will have an `idx` equal to the value of
+        `self.number_of_factors` just prior to adding the new factor.
+
+        Assumes:
+            The given random variables all belong to this PGM.
+            The random variables contain no duplicates.
+
+        Args:
+            *rvs: the random variables.
+
+        Returns:
+            a Factor object belonging to this PGM.
+        """
+        return Factor(self, *rvs)
+
+    def new_factor_implies(
+            self,
+            rv_1: RandomVariable,
+            state_idxs_1: int | Collection[int],
+            rv_2: RandomVariable,
+            state_idxs_2: int | Collection[int],
+    ) -> Factor:
+        """
+        Add a sparse 0/1 factor to this PGM representing:
+            rv_1 in state_idxs_1  ==>  rv_2 in states_2.
+        That is:
+            factor[s1, s2] = 1, if s1 not in state_idxs_1 or s2 in states_2;
+                           = 0, otherwise.
+
+        Args:
+            rv_1: The first random variable.
+            state_idxs_1: state idxs of the first random variable.
+            rv_2: The second random variable.
+            state_idxs_2: state idxs of the second random variable.
+
+        Returns:
+            a Factor object belonging to this PGM, with a configured sparse potential function.
+        """
+        if isinstance(state_idxs_1, int):
+            state_idxs_1 = (state_idxs_1,)
+        if isinstance(state_idxs_2, int):
+            state_idxs_2 = (state_idxs_2,)
+
+        factor = self.new_factor(rv_1, rv_2)
+        f = factor.set_sparse()
+        for i_1 in rv_1.state_range():
+            if i_1 not in state_idxs_1:
+                for i_2 in rv_2.state_range():
+                    f[i_1, i_2] = 1
+            else:
+                for i_2 in rv_2.state_range():
+                    if i_2 in state_idxs_2:
+                        f[i_1, i_2] = 1
+        return factor
+
+    def new_factor_equiv(
+            self,
+            rv_1: RandomVariable,
+            state_idxs_1: int | Collection[int],
+            rv_2: RandomVariable,
+            state_idxs_2: int | Collection[int],
+    ) -> Factor:
+        """
+        Add a sparse 0/1 factor to this PGM representing:
+            rv_1 in state_idxs_1  <==>  rv_2 in state_idxs_2.
+        That is:
+            factor[s1, s2] = 1, if s1 in state_idxs_1 == s2 in state_idxs_2;
+                           = 0, otherwise.
+
+        Args:
+            rv_1: The first random variable.
+            state_idxs_1: state idxs of the first random variable.
+            rv_2: The second random variable.
+            state_idxs_2: state idxs of the second random variable.
+
+        Returns:
+            a Factor object belonging to this PGM, with a configured sparse potential function.
+        """
+        if isinstance(state_idxs_1, int):
+            state_idxs_1 = (state_idxs_1,)
+        if isinstance(state_idxs_2, int):
+            state_idxs_2 = (state_idxs_2,)
+
+        factor = self.new_factor(rv_1, rv_2)
+        f = factor.set_sparse()
+        for i_1 in rv_1.state_range():
+            in_1 = i_1 in state_idxs_1
+            for i_2 in rv_2.state_range():
+                in_2 = i_2 in state_idxs_2
+                if in_1 == in_2:
+                    f[i_1, i_2] = 1
+        return factor
+
+    def new_factor_functional(
+            self,
+            function: Callable[[...], int],
+            result_rv: RandomVariable,
+            *input_rvs: RandomVariable
+    ) -> Factor:
+        """
+        Add a sparse 0/1 factor to this PGM representing:
+            result_rv ==  function(*rvs).
+        That is:
+            factor[result_s, *input_s] = 1, if result_s == function(*input_s);
+                                       = 0, otherwise.
+        Args:
+            function: a function from state indexes of the input random variables to a state index
+                of the result random variable. The function should take the same number of arguments
+                as `input_rvs` and return a state index for `result_rv`.
+            result_rv: the random variable defining result values.
+            *input_rvs: the random variables defining input values.
+
+        Returns:
+            a Factor object belonging to this PGM, with a configured sparse potential function.
+        """
+        factor = self.new_factor(result_rv, *input_rvs)
+        f = factor.set_sparse()
+        for input_s in _combos([list(rv.state_range()) for rv in input_rvs]):
+            result_s = function(*input_s)
+            f[(result_s,) + input_s] = 1
+        return factor
+
+    def indicator_pair(self, indicator: Indicator) -> Tuple[RandomVariable, State]:
+        """
+        Convert the given indicator to its RandomVariable and State value.
+
+        Args:
+            indicator: the indicator to convert.
+
+        Returns:
+            (rv, state) where
+            rv: is the random variable of the indicator.
+            state: is the random variable state of the indicator.
+        """
+        rv = self._rvs[indicator.rv_idx]
+        state = rv.states[indicator.state_idx]
+        return rv, state
+
+    def indicator_str(self, *indicators: Indicator, sep: str = '=', delim: str = ', ') -> str:
+        """
+        Render indicators as a string.
+
+        For example:
+            pgm = PGM()
+            a = pgm.new_rv('A', ('x', 'y', 'z'))
+            b = pgm.new_rv('B', (3, 5))
+            print(pgm.indicator_str(a[0], b[1], a[2]))
+        will print:
+            A=x, B=5, A=z
+
+        Args:
+            *indicators: the indicators to render.
+            sep: the separator to use between the random variable and its state.
+            delim: the delimiter to used when rendering multiple indicators.
+
+        Returns:
+            a string representation of the given indicators.
+        """
+        return delim.join(
+            f'{_clean_str(rv)}{sep}{_clean_str(state)}'
+            for rv, state in (
+                self.indicator_pair(indicator)
+                for indicator in indicators
+            )
+        )
+
+    def condition_str(self, *indicators: Indicator) -> str:
+        """
+        Render indicators as a string, grouping indicators by random variable.
+
+        For example:
+            pgm = PGM()
+            a = pgm.new_rv('A', ('x', 'y', 'z'))
+            b = pgm.new_rv('B', (3, 5))
+            print(pgm.condition_str(a[0], b[1], a[2]))
+        will print:
+            A in {x, z}, B=5
+
+        Args:
+            *indicators: the indicators to render.
+        Return:
+            a string representation of the given indicators, as a condition.
+        """
+        indicators: List[Indicator] = sorted(indicators, reverse=True)
+        cur_rv: Set[Indicator] = set()
+        cur_idx: int = -1  # rv_idx of the rv we are currently working on, -1 means not yet started.
+        cur_str: str = ''  # accumulated result string
+        while len(indicators) > 0:
+            this_ind = indicators.pop()
+            if this_ind.rv_idx != cur_idx:
+                if cur_idx >= 0:
+                    cur_str = self._condition_str_rv(cur_str, cur_rv)
+                cur_rv = set()
+                cur_idx = this_ind.rv_idx
+            cur_rv.add(this_ind)
+        if cur_idx >= 0:
+            cur_str = self._condition_str_rv(cur_str, cur_rv)
+        return cur_str
+
+    def instance_str(
+            self,
+            instance: Instance,
+            rvs: Optional[Sequence[RandomVariable]] = None,
+            sep: str = '=',
+            delim: str = ', ',
+    ) -> str:
+        """
+        Render an instance as a string.
+
+        The result looks something like 'X=x, Y=y, Z=z' where X, Y, and X are
+        random variables and x, y, and z are the states represented by the
+        given instance.
+
+        Args:
+            instance: the instance to render.
+            rvs: the random variables that the instance refers to. If rvs is None, then `self.rvs` is used.
+            sep: the separator to use between the random variable and its state.
+            delim: the delimiter to used when rendering multiple indicators.
+
+        Returns:
+            a string representation of the indicators implied by the given instance.
+        """
+        if rvs is None:
+            rvs = self.rvs
+        assert len(instance) == len(rvs)
+        return self.indicator_str(
+            *[rv[state] for rv, state in zip(rvs, instance)],
+            sep=sep,
+            delim=delim
+        )
+
+    def state_str(
+            self,
+            instance: Instance,
+            rvs: Optional[Sequence[RandomVariable]] = None,
+            delim: str = ', ',
+    ) -> str:
+        """
+        Render the states of an instance.
+
+        The result looks something like 'x, y, z' where x, y, and z are
+        the states of the random variables represented by the given instance.
+
+        Args:
+            instance: the instance to render.
+            rvs: the random variables that the instance refers to. If rvs is None, then `self.rvs` is used.
+            delim: the delimiter to used when rendering multiple indicators.
+
+        Returns:
+            a string representation of the states implied by the given instance.
+        """
+        if rvs is None:
+            rvs = self.rvs
+        assert len(instance) == len(rvs)
+        return delim.join(str(rv.states[i]) for rv, i in zip(rvs, instance))
+
+    def instances(self, flip: bool = False) -> Iterable[Instance]:
+        """
+        Iterate over all possible instances of this PGM, in natural index
+        order (i.e., last random variable changing most quickly).
+
+        Args:
+            flip: if true, then first random variable changes most quickly.
+
+        Returns:
+            an iteration over tuples, each tuple holds random variable state indexes
+            co-indexed with this PGM's random variables, `self.rvs`.
+        """
+        return _combos_ranges(tuple(len(rv) for rv in self._rvs), flip=not flip)
+
+    def instances_as_indicators(self, flip: bool = False) -> Iterable[Sequence[Indicator]]:
+        """
+        Iterate over all possible instances of this PGM, in natural index
+        order (i.e., last random variable changing most quickly).
+
+        Args:
+            flip: if true, then first random variable changes most quickly.
+
+        Returns:
+            an iteration over tuples, each tuples holds random variable indicators
+            co-indexed with this PGM's random variables, `self.rvs`.
+        """
+        for inst in self.instances(flip=flip):
+            yield self.state_idxs_to_indicators(inst)
+
+    def state_idxs_to_indicators(self, instance: Sequence[int]) -> Sequence[Indicator]:
+        """
+        Given an instance (list of random variable state indexes), co-indexed with the PGM's
+        random variables, `self.rvs`, return the corresponding indicators.
+
+        Assumes:
+            The instance has the same length as `self.rvs`.
+            The instance is co-indexed with `self.rvs`.
+
+        Args:
+            instance: the instance to convert to indicators.
+
+        Returns:
+            a tuple of indicators, co-indexed with `self.rvs`.
+        """
+        return tuple(rv[state] for rv, state in zip(self._rvs, instance))
+
+    def factor_values(self, key: Key) -> Iterable[float]:
+        """
+        For a given instance key, each factor defines a single value. This method
+        returns those values.
+
+        Args:
+            key: the key defining an instance of this PGM.
+
+        Returns:
+            an iterator over factor values, co-indexed with the factors of this PGM.
+        """
+        instance: Instance = check_key(self._shape, key)
+        assert len(instance) == len(self._rvs)
+        for factor in self._factors:
+            states: Sequence[int] = tuple(instance[rv.idx] for rv in factor.rvs)
+            value: float = factor.function[states]
+            yield value
+
+    @property
+    def is_structure_bayesian(self) -> bool:
+        """
+        Does the PGM structure correspond to a Bayesian network, where
+        each factor is taken to be a CPT and the first random variable of factor
+        is taken to be the child.
+
+        This method does not check the factor parameters to confirm they correspond
+        to valid CPTs.
+
+        Return:
+            True only if:
+                the number of factors equals the number of random variables,
+                each random variable appears exactly once as the first random variable of a factor,
+                there are no directed loops created by the factors.
+        """
+
+        # One factor per random variable.
+        if self.number_of_factors != self.number_of_rvs:
+            return False
+
+        # Each random variable is a child.
+        # Map each random variable to the factor it is a child of
+        child_to_factor: Dict[int, Factor] = {
+            factor.rvs[0].idx: factor
+            for factor in self._factors
+        }
+        if len(child_to_factor) != self.number_of_rvs:
+            return False
+
+        # Factors form a DAG
+        states: NDArrayUInt8 = np.zeros(self.number_of_factors, dtype=np.uint8)
+        for factor in self._factors:
+            if self._has_cycle(factor, child_to_factor, states):
+                return False
+
+        # All tests passed
+        return True
+
+    def factors_are_cpts(self, tolerance: float = DEFAULT_TOLERANCE) -> bool:
+        """
+        Are all factor potential functions set with parameters values
+        conforming to Conditional Probability Tables.
+
+        Assumes:
+            tolerance is non-negative.
+
+        Args:
+            tolerance: a tolerance when testing if values are equal to zero or one.
+
+        Returns:
+            True only if every potential function conforms to being a valid CPT.
+        """
+        return all(function.is_cpt(tolerance) for function in self.functions)
+
+    def check_is_bayesian_network(self, tolerance: float = DEFAULT_TOLERANCE) -> bool:
+        """
+        Is this PGM a Bayesian network.
+
+        Assumes:
+            tolerance is non-negative.
+
+        Args:
+            tolerance: a tolerance when testing if values are equal to zero or one.
+
+        Returns:
+            `is_structure_bayesian and check_factors_are_cpts(tolerance)`.
+        """
+        return self.is_structure_bayesian and self.factors_are_cpts(tolerance)
+
+    def value_product(self, key: Key) -> float:
+        """
+        For a given instance key, each factor defines a single value. This method
+        returns the product of those values.
+
+        Args:
+            key: the key defining an instance of this PGM.
+
+        Returns:
+            the product of factor values.
+        """
+        return _multiply(self.factor_values(key))
+
+    def value_product_indicators(self, *indicators: Indicator) -> float:
+        """
+        Return the product of factors, conditioned on the given indicators.
+
+        For random variables not mentioned in the indicators, then the result is the sum
+        of the value product for each possible combination of states of the unmentioned
+        random variables.
+
+        If no indicators are provided, then the value of the partition function (z)
+        is returned.
+
+        If multiple indicators are provided for the same random variable, then all matching
+        instances are summed.
+
+        This method has the same semantics as `ProbabilitySpace.wmc` without conditioning.
+
+        Warning:
+            this is potentially computationally expensive as it marginalised random
+            variables not mentioned in the given indicators.
+
+        Args:
+            *indicators: are indicators from random variables of this PGM.
+
+        Returns:
+            the product of factors, conditioned on the given instance. This is the
+            computed value of the PGM, conditioned on the given instance.
+        """
+        # # Create a filter from the indicators
+        # inst_filter: List[Set[int]] = [set() for _ in range(self.number_of_rvs)]
+        # for indicator in indicators:
+        #     rv_idx: int = indicator.rv_idx
+        #     inst_filter[rv_idx].add(indicator.state_idx)
+        # # Collect rvs not mentioned - to marginalise
+        # for rv, rv_filter in zip(self.rvs, inst_filter):
+        #     if len(rv_filter) == 0:
+        #         rv_filter.update(rv.state_range())
+        #
+        # def _sum_inst(_instance: Instance) -> bool:
+        #     return all(
+        #         (_state in _rv_filter)
+        #         for _state, _rv_filter in zip(_instance, inst_filter)
+        #     )
+        #
+        # # Accumulate the result
+        # sum_value = 0
+        # for instance in self.instances():
+        #     if _sum_inst(instance):
+        #         sum_value += self.value_product(instance)
+        #
+        # return sum_value
+
+        # Work out the space to sum over
+        sum_space_set: List[Optional[Set[int]]] = [None] * self.number_of_rvs
+        for indicator in indicators:
+            rv_idx: int = indicator.rv_idx
+            cur_set = sum_space_set[rv_idx]
+            if cur_set is None:
+                sum_space_set[rv_idx] = cur_set = set()
+            cur_set.add(indicator.state_idx)
+
+        # Convert to a list of states that we need to sum over.
+        sum_space_list: List[List[int]] = [
+            list(cur_set if cur_set is not None else rv.state_range())
+            for cur_set, rv in zip(sum_space_set, self.rvs)
+        ]
+
+        # Accumulate the result
+        return sum(
+            self.value_product(instance)
+            for instance in _combos(sum_space_list)
+        )
+
+    def dump_synopsis(
+            self,
+            *,
+            prefix: str = '',
+            precision: int = 3,
+            max_state_digits: int = 21,
+    ):
+        """
+        Print a synopsis of the PGM.
+        This is intended for demonstration and debugging purposes.
+
+        Args:
+            prefix: optional prefix for indenting all lines.
+            precision: a limit on the render precision of floating point numbers.
+            max_state_digits: a limit on the number of digits when showing number of states as an integer.
+        """
+        # limit the precision when displaying number of states
+        num_states: int = self.number_of_states
+        number_of_parameters = sum(function.number_of_parameters for function in self.functions)
+        number_of_nz_parameters = sum(function.number_of_parameters for function in self.non_zero_functions)
+
+        if math.log10(num_states) > max_state_digits:
+            log_states = math.log10(num_states)
+            exp = int(log_states)
+            man = math.pow(10, log_states - exp)
+            num_states_str = f'{man:,.{precision}f}e+{exp}'
+        else:
+            num_states_str = f'{num_states:,}'
+
+        log_2_num_states = math.log2(num_states)
+        if (
+                log_2_num_states == 0
+                or (
+                    log_2_num_states == int(log_2_num_states)
+                    and math.log10(log_2_num_states) <= max_state_digits
+                )
+        ):
+            log_2_num_states_str = f'{int(log_2_num_states):,}'
+        else:
+            log_2_num_states_str = f'{math.log2(num_states):,.{precision}f}'
+
+        print(f'{prefix}name: {self.name}')
+        print(f'{prefix}number of random variables: {self.number_of_rvs:,}')
+        print(f'{prefix}number of indicators: {self.number_of_indicators:,}')
+        print(f'{prefix}number of states: {num_states_str}')
+        print(f'{prefix}log 2 of states: {log_2_num_states_str}')
+        print(f'{prefix}number of factors: {self.number_of_factors:,}')
+        print(f'{prefix}number of functions: {self.number_of_functions:,}')
+        print(f'{prefix}number of non-zero functions: {self.number_of_non_zero_functions:,}')
+        print(f'{prefix}number of parameters: {number_of_parameters:,}')
+        print(f'{prefix}number of functions (excluding ZeroPotentialFunction): {self.number_of_non_zero_functions:,}')
+        print(f'{prefix}number of parameters (excluding ZeroPotentialFunction): {number_of_nz_parameters:,}')
+        print(f'{prefix}Bayesian structure: {self.is_structure_bayesian}')
+        print(f'{prefix}CPT factors: {self.factors_are_cpts()}')
+
+    def dump(
+            self,
+            *,
+            prefix: str = '',
+            indent: str = '  ',
+            show_function_values: bool = False,
+            precision: int = 3,
+            max_state_digits: int = 21,
+    ) -> None:
+        """
+        Print a dump of the PGM.
+        This is intended for demonstration and debugging purposes.
+
+        Args:
+            prefix: optional prefix for indenting all lines.
+            show_function_values: if true, then the function values will be dumped.
+            indent: additional prefix to use for extra indentation.
+            precision: a limit on the render precision of floating point numbers.
+            max_state_digits: a limit on the number of digits when showing number of states as an integer.
+        """
+
+        next_prefix: str = prefix + indent
+        next_next_prefix: str = next_prefix + indent
+
+        print(f'{prefix}PGM id={id(self)} name={self.name!r}')
+        self.dump_synopsis(prefix=next_prefix, precision=precision, max_state_digits=max_state_digits)
+
+        print(f'{prefix}random variables ({self.number_of_rvs})')
+        for rv in self.rvs:
+            print(f'{next_prefix}{rv.idx:>3} {rv.name!r} ({len(rv)})', end='')
+            if not rv.is_default_states():
+                print(' [', end='')
+                print(', '.join(repr(s) for s in rv.states), end='')
+                print(']', end='')
+            print()
+
+        print(f'{prefix}factors ({self.number_of_factors})')
+        for factor in self.factors:
+            rv_idxs = [rv.idx for rv in factor.rvs]
+            if factor.is_zero:
+                function_ref = '<zero>'
+            else:
+                function = factor.function
+                function_ref = f'{id(function)}: {function.__class__.__name__}'
+
+            print(f'{next_prefix}{factor.idx:>3} rvs={rv_idxs} function={function_ref}')
+
+        print(f'{prefix}functions ({self.number_of_functions})')
+        for function in sorted(self.non_zero_functions, key=lambda f: id(f)):
+            print(f'{next_prefix}{id(function):>13}: {function.__class__.__name__}')
+            function.dump(prefix=next_next_prefix, show_function_values=show_function_values, show_id_class=False)
+
+        print(f'{prefix}end PGM id={id(self)}')
+
+    def _has_cycle(self, factor: Factor, child_to_factor: Dict[int, Factor], states: NDArrayUInt8) -> bool:
+        """
+        Support function for `is_structure_bayesian`.
+
+        A recursive depth-first-search to see if the factors form a DAG.
+
+        For a factor `f` the value of states[f.idx] is the search state.
+        Specifically:
+        state 0 => the factor has not been seen yet,
+        state 1 => the factor is seen but not fully processed,
+        state 2 => the factor is fully processed.
+
+        Args:
+            factor: the current Factor being checked.
+            child_to_factor: a dictionary from `RandomVariable.idx` to Factor
+                with that random variable as the child.
+            states: depth-first-search states, i.e., `states[i]` is the state of a factor with `Factor.idx == i`.
+        Returns:
+            True if a directed cycle is detected.
+        """
+        f_idx: int = factor.idx
+        match states.item(f_idx):
+            case 1:
+                return True
+            case 0:
+                states[f_idx] = 1
+                for parent in factor.rvs[1:]:
+                    parent_factor = child_to_factor[parent.idx]
+                    if self._has_cycle(parent_factor, child_to_factor, states):
+                        return True
+                states[f_idx] = 2
+                return False
+        return False
+
+    def _register_rv(self, rv: RandomVariable) -> None:
+        """
+        Called by the constructor of RandomVariable to record a newly created Random variable
+        of this PGM.
+
+        Args:
+            rv: the newly constructed random variable.
+        """
+        assert rv.pgm is self
+        self._rvs += (rv,)
+        self._shape += (len(rv),)
+        self._indicators += rv.indicators
+
+    def _condition_str_rv(
+            self,
+            cur_str: str,
+            cur_rv: Set[Indicator],
+            sep: str = ', ',
+            equal: str = '=',
+            elem: str = ' in ',
+    ) -> str:
+        """
+        Support method for `self.condition_str`.
+
+        This is a method renders a condition defined by a set of indicators, of the same random variable.
+
+        Args:
+              cur_str: the string to append to.
+              cur_rv: a set of indicators, all from the same random variable.
+              sep: the separator string to use between condition components.
+              equal: the string to use for _rv_ = _state_.
+              elem: the string to use for _rv_ in _set_.
+
+        Returns:
+            `cur_str` appended with the new condition, `cur_rv`.
+        """
+        if cur_str != '':
+            cur_str += sep
+        if len(cur_rv) == 1:
+            cur_str += self.indicator_str(*cur_rv, sep=equal)
+        else:
+            _cur_rv = sorted(cur_rv)
+            rv = self._rvs[_cur_rv[0].rv_idx]
+            states_str: str = sep.join(_clean_str(rv.states[ind.state_idx]) for ind in _cur_rv)
+            cur_str += f'{_clean_str(rv)}{elem}{{{states_str}}}'
+        return cur_str
+
+
+@dataclass(frozen=True, eq=True, slots=True)
+class Indicator:
+    """
+    An indicator identifies a random variable being in a particular state.
+
+    Indicators are immutable and hashable.
+
+    Note that an Indicator does not know which PGM it came from, therefore indicators from one PGM
+    are interchangeable with indicators of another PGM so long as corresponding random variables of the
+    PGMs are co-indexed (created in the same order) and corresponding random variables have the same
+    states.
+
+    Fields:
+        rv_idx: `rv.idx` where `rv` is the random variable referenced by this indicator.
+        state_idx: the state index of the state referenced by this indicator.
+    """
+    rv_idx: int
+    state_idx: int
+
+    def __lt__(self, other) -> bool:
+        """
+        Define a sort order over indicators.
+        When sorted, indicators are ordered by random variable index, then by state index.
+        """
+        if isinstance(other, Indicator):
+            if self.rv_idx < other.rv_idx:
+                return True
+            if self.rv_idx > other.rv_idx:
+                return False
+            return self.state_idx < other.state_idx
+        return False
+
+
+class RandomVariable(Sequence[Indicator]):
+    """
+    A random variable in a probabilistic graphical model.
+
+    Random variables are immutable and hashable.
+
+    Each RandomVariable has a fixed finite number of states.
+    Its states are indexed by integers, counting from zero.
+
+    Every RandomVariable object belongs to exactly one PGM object.
+
+    Every random variable has an index (counting from zero) which is its position
+    in the random variable's PGM list of random variables.
+
+    A random variable behaves like a sequence of Indicators, where each indicator represents a random
+    variable being in a particular state. Specifically for a random variable rv, len(rv) is the
+    number of states of the random variable and rv[i] is the Indicators representing that
+    rv is in the ith state. When sliced, the result is a tuple, i.e. rv[1:3] = (rv[1], rv[2]).
+
+    A RandomVariable has a name. This is for human convenience and has no functional purpose
+    within a PGM.
+    """
+
+    def __init__(self, pgm: PGM, name: str, states: Union[int, Sequence[State]]):
+        """
+        Create a new random variable, in the given PGM.
+
+        Assumes:
+            Provided states contain no duplicates.
+
+        Args:
+            pgm: the PGM that the random variable will belong to.
+            name: a name for the random variable.
+            states: either an integer number of states or a sequence of state values. If a
+                single integer, `n`, is provided then the states will be 0, 1, ..., n-1.
+                If a sequence of states are provided then the states must be unique.
+        """
+        self._pgm: PGM = pgm
+        self._name: str = name
+
+        if isinstance(states, int):
+            states = tuple(range(states))
+
+        self._states: Sequence[State] = tuple(states)
+        self._inv_states: Dict[State, int] = {state: idx for idx, state in enumerate(self._states)}
+
+        if len(self._inv_states) != len(self._states):
+            raise ValueError('random variable states are not unique')
+
+        self._offset: int = pgm.number_of_indicators
+        self._idx: int = pgm.number_of_rvs
+        self._indicators: Sequence[Indicator] = tuple(Indicator(self._idx, i) for i in range(len(self._states)))
+
+        # Register self with our PGM
+        # noinspection PyProtectedMember
+        pgm._register_rv(self)
+
+    @property
+    def pgm(self) -> PGM:
+        """
+        Returns:
+            The PGM that this random variable belongs to.
+        """
+        return self._pgm
+
+    @property
+    def name(self) -> str:
+        """
+        Returns:
+            The name of this random variable.
+        """
+        return self._name
+
+    @property
+    def idx(self) -> int:
+        """
+        Returns:
+            The index of this random variable into the PGM.
+
+        Ensures:
+            `self.pgm.rvs[self.idx] is self`.
+        """
+        return self._idx
+
+    @property
+    def offset(self) -> int:
+        """
+        Returns:
+            The index into the PGM's indicators for the start of this random variable's indicators.
+
+        Ensures:
+            `self.pgm.indicators[self.offset + i] is self[i] for i in range(len(self))`.
+        """
+        return self._offset
+
+    @property
+    def states(self) -> Sequence[State]:
+        """
+        Returns:
+            the states of this random variable, in state index order.
+        """
+        return self._states
+
+    @property
+    def indicators(self) -> Sequence[Indicator]:
+        """
+        Returns:
+            the indicators of this random variable, in state index order.
+        """
+        return self._indicators
+
+    def state_range(self) -> Iterable[int]:
+        """
+        Iterate over the state indexes of this random variable, in order.
+
+        Returns:
+            range(len(self))
+        """
+        return range(len(self._states))
+
+    def factors(self) -> Iterable[Factor]:
+        """
+        Iterate over factors that this random variable participates in.
+        This method performs a search through all `self.pgm.factors`.
+
+        Returns:
+            an iterator over factors.
+        """
+        for factor in self._pgm.factors:
+            if self in factor.rvs:
+                yield factor
+
+    def markov_blanket(self) -> Set[RandomVariable]:
+        """
+        Return the set of random variable that are connected
+        to this random variable by a factor.
+        This method performs a search through all `self.pgm.factors`.
+
+        Returns:
+            a set of random variables connected to this random variable by any factor, excluding self.
+        """
+        result = set()
+        for factor in self.factors():
+            result.update(factor.rvs)
+        result.discard(self)
+        return result
+
+    def state_idx(self, state: State) -> int:
+        """
+        Returns:
+            the state index of the given state of this random variable.
+
+        Assumes:
+            the given state is a state of this random variable.
+        """
+        return self._inv_states[state]
+
+    def is_default_states(self) -> bool:
+        """
+        Are the states of this random variable the default states.
+        I.e., `self.states[i] == i, for all 0 <= i < len(self)`.
+
+        Returns:
+            True only if the states are the same as the state indexes.
+        """
+        return all(i == s for i, s in enumerate(self._states))
+
+    def __str__(self) -> str:
+        """
+        Returns:
+             the name of this random variable.
+        """
+        return self._name
+
+    def __call__(self, state: State) -> Indicator:
+        """
+        Get the indicator for the given state.
+        This is equivalent to self[self.state_idx(state)].
+
+        Returns:
+            an indicator of this random variable.
+
+        Assumes:
+            the given state is a state of this random variable.
+        """
+        return self._indicators[self._inv_states[state]]
+
+    def __hash__(self) -> int:
+        """
+        A random variable is hashable.
+        """
+        return self._idx
+
+    def __eq__(self, other) -> bool:
+        """
+        Two random variable are equal if they are the same object.
+        """
+        return self is other
+
+    def equivalent(self, other: RandomVariable | Sequence[Indicator]) -> bool:
+        """
+        Two random variable are equivalent if their indicators are equal. Only
+        random variable indexes and state indexes are checked.
+
+        This ignores the names of the random variable and the names of their states.
+        This means their indicators will work correctly in slot maps, even
+        if from different PGMs.
+
+        Args:
+            other: either a random variable or a sequence of Indicators.
+
+        Returns:
+            True only if they represent the same sequence of indicators.
+        """
+        indicators = self._indicators
+        if isinstance(other, RandomVariable):
+            return self.idx == other.idx and len(self) == len(other)
+        else:
+            return (
+                    len(indicators) == len(other) and
+                    all(indicators[i] == other[i] for i in range(len(indicators)))
+            )
+
+    def __len__(self) -> int:
+        """
+        Returns:
+            Number of states (or equivalently, the number of indicators) of this random variable.
+        """
+        return len(self._states)
+
+    def __iter__(self) -> Iterator[Indicator]:
+        """
+        Iterate over the indicators of this random variable.
+        """
+        return iter(self._indicators)
+
+    @overload
+    def __getitem__(self, index: int) -> Indicator:
+        ...
+
+    @overload
+    def __getitem__(self, index: slice) -> Sequence[Indicator]:
+        ...
+
+    def __getitem__(self, index):
+        """
+        Get the indexed (or sliced) indicators.
+        """
+        return self._indicators[index]
+
+    def index(self, value: Any, start: int = 0, stop: int = -1) -> int:
+        """
+        Returns the first index of `value`.
+        Raises ValueError if the value is not present.
+        Contracted by Sequence[Indicator].
+
+        Warning:
+            This method is different to `self.idx`.
+        """
+        if isinstance(value, Indicator):
+            if value.rv_idx == self._idx:
+                idx: int = value.state_idx
+                if stop < 0:
+                    stop = len(self) + stop + 1
+                if 0 <= idx < len(self) and start <= idx < stop:
+                    return value.state_idx
+        raise ValueError(f'{value!r} is not an indicator of the random variable')
+
+    def count(self, value: Any) -> int:
+        """
+        Returns the number of occurrences of `value`.
+        Contracted by Sequence[Indicator].
+        """
+        if isinstance(value, Indicator):
+            if value.rv_idx == self._idx and 0 <= value.state_idx < len(self):
+                return 1
+        return 0
+
+
+class RVMap(Sequence[RandomVariable]):
+    """
+    Wrap a PGM to provide convenient access to PGM random variables.
+
+    An RVMap of a PGM behaves exactly like the PGM `rvs` property. That it, it
+    behaves like a sequence of RandomVariable objects.
+
+    If the underlying PGM is updated, then the RVMap will automatically update.
+
+    Additionally, an RVMap enables access to the PGM random variable via the name
+    of each random variable.
+
+    for example, if `pgm.rvs[1]` is a random variable named `xray`, then
+    ```
+    rvs = RVMap(pgm)
+
+    # These all retrieve the same random variable object.
+    xray = rvs[1]
+    xray = rvs('xray')
+    xray = rvs.xray
+    ```
+
+    To use an RVMap on a PGM, the variable names must be unique across the PGM.
+    """
+
+    def __init__(self, pgm: PGM, ignore_case: bool = False):
+        """
+        Construct an RVMap for the given PGM.
+
+        Args:
+            pgm: the PGM to wrap.
+            ignore_case: if true, the variable name are not case-sensitive.
+        """
+        self._pgm: PGM = pgm
+        self._ignore_case: bool = ignore_case
+        self.__rv_map: Dict[str, RandomVariable] = {}
+        self._reserved_names: Set[str] = {self._clean_name(name) for name in dir(self)}
+
+        # Force the rv map cache to be updated.
+        # This may raise an exception.
+        _ = self._rv_map
+
+    def _clean_name(self, name: str) -> str:
+        """
+        Adjust the case of the given name as needed.
+        """
+        return name.lower() if self._ignore_case else name
+
+    @property
+    def _rv_map(self) -> Dict[str, RandomVariable]:
+        """
+        Get the cached rv map, updating as needed if the PGM changed.
+        Returns:
+            a mapping from random variable name to random variable
+        """
+        if len(self.__rv_map) != len(self._pgm.rvs):
+            # There is a difference between the map and the PGM - create a new map.
+            self.__rv_map = {self._clean_name(rv.name): rv for rv in self._pgm.rvs}
+            if len(self.__rv_map) != len(self._pgm.rvs):
+                raise RuntimeError(f'random variable names are not unique')
+            if not self._reserved_names.isdisjoint(self.__rv_map.keys()):
+                raise RuntimeError(f'random variable names clash with reserved names.')
+        return self.__rv_map
+
+    def new_rv(self, name: str, states: Union[int, Sequence[State]]) -> RandomVariable:
+        """
+        As per `PGM.new_rv`.
+        Delegate creating a new random variable to the PGM.
+
+        Returns:
+            a RandomVariable object belonging to the PGM.
+        """
+        return self._pgm.new_rv(name, states)
+
+    def __len__(self) -> int:
+        return len(self._pgm.rvs)
+
+    def __getitem__(self, index: int) -> RandomVariable:
+        return self._pgm.rvs[index]
+
+    def items(self) -> Iterable[Tuple[str, RandomVariable]]:
+        return self._rv_map.items()
+
+    def keys(self) -> Iterable[str]:
+        return self._rv_map.keys()
+
+    def values(self) -> Iterable[RandomVariable]:
+        return self._rv_map.values()
+
+    def get(self, rv_name: str, default=None):
+        return self._rv_map.get(self._clean_name(rv_name), default)
+
+    def __call__(self, rv_name: str) -> RandomVariable:
+        return self._rv_map[self._clean_name(rv_name)]
+
+    def __getattr__(self, rv_name: str) -> RandomVariable:
+        return self(rv_name)
+
+
+class Factor:
+    """
+    A PGM factor over one or more random variables declares a relationship between
+    those variables. A Factor also has a potential function associated with
+    it which defines a real-number value with each combination of states of
+    the random variables.
+
+    The default potential function for a factor is a unique ZeroPotentialFunction.
+
+    The order of a Factors random variables is important as many things will be
+    co-indexed with the random variables. For example, the shape of a Factor is
+    the tuple of random variable lengths.
+
+    Note that multiple factors may share a potential function, so long as they all
+    belong to the same PGM object and have the same shape.
+    """
+
+    def __init__(self, pgm: PGM, *rvs: RandomVariable):
+        """
+        Add a new factor to the given PGM.
+
+        Assumes:
+            The given random variables all belong to this PGM.
+            The random variables contain no duplicates.
+
+        Args:
+            pgm: the PGM that the factor will belong to.
+            *rvs: the random variables.
+
+        Returns:
+            a Factor object belonging to this PGM.
+        """
+        if len(set(rvs)) != len(rvs):
+            raise ValueError('duplicated random variable in factor')
+        if len(rvs) == 0:
+            raise ValueError('must be at least one random variable')
+        if any(rv.pgm is not pgm for rv in rvs):
+            raise ValueError('random variable not from the same PGM')
+
+        self._pgm: PGM = pgm
+        self._idx: int = pgm.number_of_factors
+        self._rvs: Sequence[RandomVariable] = tuple(rvs)
+        self._shape: Shape = tuple(len(rv) for rv in rvs)
+
+        self._zero_potential_function: ZeroPotentialFunction = ZeroPotentialFunction(self)
+        self._potential_function: PotentialFunction = self._zero_potential_function
+
+        # Register self with our PGM
+        # noinspection PyProtectedMember
+        pgm._factors += (self,)
+
+    @property
+    def rvs(self) -> Sequence[RandomVariable]:
+        """
+        Returns:
+            The random variables of this factor.
+        """
+        return self._rvs
+
+    @property
+    def pgm(self) -> PGM:
+        """
+        Returns:
+            The PGM that this factor belongs to.
+        """
+        return self._pgm
+
+    @property
+    def idx(self) -> int:
+        """
+        Returns:
+            The index of this factor into the PGM.
+
+        Ensures:
+            `self.pgm.factors[self.idx] is self`.
+        """
+        return self._idx
+
+    @property
+    def shape(self) -> Shape:
+        return self._shape
+
+    @property
+    def number_of_states(self) -> int:
+        """
+        How many distinct states are covered by this Factor.
+        """
+        return self._potential_function.number_of_states
+
+    def __str__(self) -> str:
+        """
+        Return a human-readable string to represent this factor.
+        This is intended mainly for debugging purposes.
+        """
+        return '(' + ', '.join([repr(str(rv)) for rv in self._rvs]) + ')'
+
+    def __len__(self) -> int:
+        """
+        Returns:
+            the number of random variables.
+        """
+        return len(self._rvs)
+
+    @overload
+    def __getitem__(self, index: int) -> RandomVariable:
+        ...
+
+    @overload
+    def __getitem__(self, index: slice) -> Sequence[RandomVariable]:
+        ...
+
+    def __getitem__(self, index):
+        return self._rvs[index]
+
+    def instances(self, flip: bool = False) -> Iterable[Instance]:
+        """
+        Iterate over all possible instances, in natural index order (i.e.,
+        last random variable changing most quickly).
+
+        Args:
+            flip: if true, then first random variable changes most quickly
+
+        Returns:
+            an iterator over tuples, each tuple holds random variable
+            state indexes, co-indexed with this object's shape, i.e., self.shape.
+        """
+        return self.function.instances(flip)
+
+    def parent_instances(self, flip: bool = False) -> Iterable[Instance]:
+        """
+        Iterate over all possible instances of parent random variable, in
+        natural index order (i.e., last random variable changing most quickly).
+
+        Args:
+            flip: if true, then first random variable changes most quickly
+
+        Returns:
+            an iteration over tuples, each tuple holds random variable states
+            co-indexed with this object's 'parent' shape, i.e., `self.shape[1:]`.
+        """
+        return self.function.parent_instances(flip)
+
+    @property
+    def is_zero(self) -> bool:
+        """
+        Is the potential function of this factor set to the special 'zero' potential function.
+        """
+        return self._potential_function is self._zero_potential_function
+
+    @property
+    def function(self) -> PotentialFunction:
+        return self._potential_function
+
+    @function.setter
+    def function(self, function: PotentialFunction | Factor) -> None:
+        """
+        Set the potential function for this PGM factor to the given potential function
+        or factor.
+
+        Assumes:
+            The given potential function belongs to the same PGM as this Factor.
+            The potential function has the correct shape.
+        """
+        if isinstance(function, Factor):
+            function = function.function
+        assert isinstance(function, PotentialFunction)
+
+        if self._potential_function is function:
+            # nothing to do
+            return
+
+        if function.pgm is not self._pgm:
+            raise ValueError(f'the given function is not of the same PGM as the factor')
+
+        if function.shape != self._shape:
+            raise ValueError(f'incorrect function shape: expected {self._shape}, got {function.shape}')
+
+        if isinstance(function, ZeroPotentialFunction):
+            self.set_zero()
+        else:
+            self._potential_function = function
+
+    def set_zero(self) -> ZeroPotentialFunction:
+        """
+        Set the factor's potential function to its original ZeroPotentialFunction.
+
+        Returns:
+            the potential function.
+        """
+        self._potential_function = self._zero_potential_function
+        return self._potential_function
+
+    def set_dense(self) -> DensePotentialFunction:
+        """
+        Set to the potential function to a new `DensePotentialFunction` object.
+
+        Returns:
+            the potential function.
+        """
+        self._potential_function = DensePotentialFunction(self)
+        return self._potential_function
+
+    def set_sparse(self) -> SparsePotentialFunction:
+        """
+        Set to the potential function to a new `SparsePotentialFunction` object.
+
+        Returns:
+            the potential function.
+        """
+        self._potential_function = SparsePotentialFunction(self)
+        return self._potential_function
+
+    def set_compact(self) -> CompactPotentialFunction:
+        """
+        Set to the potential function to a new `CompactPotentialFunction` object.
+
+        Returns:
+            the potential function.
+        """
+        self._potential_function = CompactPotentialFunction(self)
+        return self._potential_function
+
+    def set_clause(self, *key: int) -> ClausePotentialFunction:
+        """
+        Set to the potential function to a new `ClausePotentialFunction` object.
+
+        Args:
+            *key: defines the random variable states of the clause. The key is a sequence of
+                random variable state indexes, co-indexed with `Factor.rvs`.
+
+        Returns:
+            the potential function.
+
+        Raises:
+             KeyError: if the key is not valid for the shape of the factor.
+        """
+        self._potential_function = ClausePotentialFunction(self, key)
+        return self._potential_function
+
+    def set_cpt(self, tolerance: float = DEFAULT_TOLERANCE) -> CPTPotentialFunction:
+        """
+        Set to the potential function to a new `CPTPotentialFunction` object.
+
+        Args:
+            tolerance: a tolerance when testing if values are equal to zero or one.
+
+        Returns:
+            the potential function.
+
+        Raises:
+            ValueError: if tolerance is negative.
+        """
+        self._potential_function = CPTPotentialFunction(self, tolerance)
+        return self._potential_function
+
+
+@dataclass(frozen=True, eq=True)
+class ParamId:
+    """
+    A ParamId identifies a parameter of a potential function.
+
+    Parameter identifiers uniquely identify every parameter within a PGM.
+
+    A ParamId is immutable and hashable.
+    """
+    function_id: int
+    param_idx: int
+
+
+class PotentialFunction(ABC):
+    """
+    A potential function defines the potential values for a Factor, where
+    a factor joins one or more variables of a PGM.
+
+    A potential function may be shared by several Factors of a PGM,
+    i.e., can be applied to multiple variables.
+
+    The `shape` of a potential function is a tuple of integers which defines
+    the number of variables, len(shape), and the number of states of each
+    variable, shape[i].
+
+    The potential function value for variable states (x = i, y = j, ...) is given by
+    self[i, j, ...], i.e., self.__getitem__((i, j, ...)). The tuple, (i, j, ...), is
+    known as a Key.
+
+    The values of a potential function are defined by potential function parameters.
+    The number of potential function parameters is given by number_of_parameters.
+    The value of each parameter is given by get_param(i), where i is the parameter index.
+
+    Every valid key of the potential function is mapped either mapped to a parameter or is
+    "guaranteed zero" which means that the value is zero and cannot be changed by changing
+    the values of the potential function's parameters.
+    """
+
+    def __init__(self, factor: Factor):
+        """
+        Create a potential function compatible with the given factor.
+
+        Ensures:
+            Does not hold a reference to the given factor.
+            Does not register the potential function with the PGM.
+
+        Args:
+            factor: which factor is this potential function is compatible with.
+        """
+        self._pgm: PGM = factor.pgm
+        self._shape: Shape = factor.shape
+        self._number_of_states = _multiply(self._shape)
+
+    @property
+    def pgm(self) -> PGM:
+        """
+        Returns:
+            The PGM that this potential function belong to.
+        """
+        return self._pgm
+
+    @property
+    def shape(self) -> Shape:
+        """
+        Returns:
+            The shape of this potential function.
+        """
+        return self._shape
+
+    @property
+    def number_of_rvs(self) -> int:
+        """
+        Returns:
+            The number of random variables in this potential function.
+        """
+        return len(self._shape)
+
+    @property
+    def number_of_states(self) -> int:
+        """
+        How many distinct states are covered by this potential function.
+
+        Returns:
+            The size of the state space of this potential function.
+        """
+        return self._number_of_states
+
+    @property
+    def number_of_parent_states(self) -> int:
+        """
+        How many distinct states are covered by this potential function parents,
+        i.e., excluding the first random variable.
+
+        Returns:
+            The size of the state space of this potential function parent random variables.
+        """
+        return _multiply(self._shape[1:])
+
+    def count_usage(self) -> int:
+        """
+        Check all PGM factors to count the number of times that this potential function
+        is used.
+
+        Returns:
+            the number of factors that use this potential function.
+        """
+        return sum(1 for factor in self._pgm.factors if factor.function is self)
+
+    def check_key(self, key: Key) -> Instance:
+        """
+        Convert the key into an instance.
+
+        Arg:
+            key: defines an instance in the state space of the potential function.
+
+        Returns:
+            an instance, which is a tuple of state indexes, co-indexed with `self.rvs`.
+
+        Raises:
+             KeyError: if the key is not valid for the shape of the factor.
+        """
+        return check_key(self._shape, key)
+
+    def valid_key(self, key: Key) -> bool:
+        """
+        Is the given key valid.
+
+        Arg:
+            key: defines an instance in the state space of the potential function.
+
+        Returns:
+            True only if the given key is valid.
+        """
+        return valid_key(self._shape, key)
+
+    def valid_parameter(self, param_idx: int) -> bool:
+        """
+        Is the given parameter index valid.
+
+        Arg:
+            param_idx: a parameter index.
+
+        Returns:
+            True only if `0 <= param_idx < self.number_of_parameters`.
+        """
+        return 0 <= param_idx < self.number_of_parameters
+
+    @property
+    def is_sparse(self) -> bool:
+        """
+        Are there any 'guaranteed zero' parameters values.
+
+        Returns:
+            True only if `self.number_of_not_guaranteed_zero < self._number_of_states`.
+        """
+        return self.number_of_not_guaranteed_zero < self._number_of_states
+
+    @property
+    @abstractmethod
+    def number_of_not_guaranteed_zero(self) -> int:
+        """
+        How many of the states of this potential function are not 'guaranteed zero'.
+        That is, how many keys are associated with a parameter.
+
+        Returns:
+            The number of valid keys that are associated with a parameter.
+
+        Ensures:
+            0 <= self.number_of_not_guaranteed_zero <= self.number_of_states.
+        """
+        ...
+
+    @property
+    @abstractmethod
+    def number_of_parameters(self) -> int:
+        """
+        Get the number of parameters defining the potential function values.
+        Each valid key of the function maps either to a parameter
+        is 'guaranteed zero'.
+
+        Returns:
+            The number of parameters.
+
+        Ensures:
+            0 <= self.number_of_parameters <= self.number_of_not_guaranteed_zero.
+        """
+        ...
+
+    @property
+    @abstractmethod
+    def params(self) -> Iterable[Tuple[int, float]]:
+        """
+        Iterate the parameters and their associated values.
+
+        Returns:
+            An iterable over (param_idx, value) tuples, for every possible parameter.
+
+        Assumes:
+            The potential function is not mutated while iterating.
+        """
+        ...
+
+    @property
+    @abstractmethod
+    def keys_with_param(self) -> Iterable[Tuple[Instance, int, float]]:
+        """
+        Iterate the keys that have a parameter associated with them.
+
+        Returns:
+            An iterable over (key, param_idx, value) tuples, for every key with an associated parameter.
+
+        Assumes:
+            The potential function is not mutated while iterating.
+        """
+        ...
+
+    @abstractmethod
+    def __getitem__(self, key: Key) -> float:
+        """
+        Get the potential function value for the given instance key.
+
+        Arg:
+            key: defines an instance in the state space of the potential function.
+
+        Returns:
+            The value of the potential function for the given key.
+
+        Assumes:
+            self.valid_key(key).
+        """
+        ...
+
+    @abstractmethod
+    def param_value(self, param_idx: int) -> float:
+        """
+        Get the potential function value by parameter index.
+
+        Arg:
+            param_idx: a parameter index.
+
+        Assumes:
+            `self.valid_parameter(param_idx)`.
+        """
+        ...
+
+    @abstractmethod
+    def param_idx(self, key: Key) -> Optional[int]:
+        """
+        Get the parameter index for the given potential function random variables states (key).
+
+        Arg:
+            key: defines an instance in the state space of the potential function.
+
+        Returns:
+            either `None` indicating a "guaranteed zero" value, or the parameter index holding
+            the potential function value for the key.
+        """
+        ...
+
+    def is_cpt(self, tolerance=DEFAULT_TOLERANCE) -> bool:
+        """
+        Is the potential function set with parameters values conforming to a
+        Conditional Probability Table.
+
+        Every parameter value must be non-negative.
+        For every state of the parent (non-first slots)
+        the sum of the parameters over the child states (first slots)
+        must be either 1 or 0.
+
+        Assumes:
+            tolerance is non-negative.
+
+        Args:
+            tolerance: a tolerance when testing if values are equal to zero or one.
+
+        Returns:
+            True only if the potential function is compatible with being a CPT.
+        """
+        # This default implementation calculates the result the long way, by checking
+        # every valid key of the potential function.
+        # Subclasses may override this implementation.
+        low: float = 1.0 - tolerance
+        high: float = 1.0 + tolerance
+        for parent_state in self.parent_instances():
+            total: float = sum(
+                self[(state,) + tuple(parent_state)]
+                for state in range(self.shape[0])
+            )
+            if not ((low <= total <= high) or (0 <= total <= tolerance)):
+                return False
+        return True
+
+    def natural_param_idx(self, key: Key) -> int:
+        """
+        Get the natural parameter index for the given key. This is the same index as used
+        by a DensePotentialFunction with the same shape.
+
+        Args:
+            key: is a valid key of the potential function, referring to an instance in the factor's state space.
+
+        Assumes:
+            `self.valid_key(key)` is true.
+
+        Returns:
+            a hypothetical parameter index assuming that every valid key has a unique parameter
+            as per DensePotentialFunction.
+        """
+        return _natural_key_idx(self._shape, key)
+
+    def param_id(self, param_idx: int) -> ParamId:
+        """
+        Get a hashable object to represent the parameter with the given parameter index.
+
+        Arg:
+            param_idx: a parameter index.
+
+        Returns:
+             a hashable ParamId object for the parameter of this potential function.
+
+        Raises:
+            ValueError: if the parameter index is not valid.
+        """
+        if not (0 <= param_idx < self.number_of_parameters):
+            raise ValueError(f'invalid parameter index: {param_idx}')
+        return ParamId(id(self), param_idx)
+
+    def items(self) -> Iterable[Tuple[Instance, float]]:
+        """
+        Iterate over all keys and values of this potential function.
+
+        Returns:
+            An iterator over all (key, value) pairs, where key is an Instance and value
+            is the value of the potential function for the key.
+        """
+        for key in _combos_ranges(self._shape, flip=True):
+            yield key, self[key]
+
+    def instances(self, flip: bool = False) -> Iterable[Instance]:
+        """
+        Iterate over all possible instances, in natural index order (i.e.,
+        last random variable changing most quickly).
+
+        Args:
+            flip: if true, then first random variable changes most quickly
+
+        Returns:
+            an iterator over tuples, each tuple holds random variable
+            state indexes, co-indexed with this object's shape, i.e., self.shape.
+        """
+        return _combos_ranges(self._shape, flip=not flip)
+
+    def parent_instances(self, flip: bool = False) -> Iterable[Instance]:
+        """
+        Iterate over all possible instances of parent random variable, in
+        natural index order (i.e., last random variable changing most quickly).
+
+        Args:
+            flip: if true, then first random variable changes most quickly
+
+        Returns:
+            an iteration over tuples, each tuple holds random variable states
+            co-indexed with this object's 'parent' shape, i.e., `self.shape[1:]`.
+        """
+        return _combos_ranges(self._shape[1:], flip=not flip)
+
+    def __str__(self) -> str:
+        """
+        Provide a human-readable representation of this potential function.
+        This is intended mainly for debugging purposes.
+        """
+        shape_str: str = ', '.join(str(x) for x in self._shape)
+        return f'{self.__class__.__name__}({shape_str})'
+
+    def dump(
+            self,
+            *,
+            prefix: str = '',
+            indent: str = '    ',
+            show_function_values: bool = False,
+            show_id_class: bool = True,
+    ) -> None:
+        """
+        Print a dump of the function.
+        This is intended for debugging purposes.
+
+        Args:
+            prefix: optional prefix for indenting all lines.
+            indent: additional prefix to use for extra indentation.
+            show_function_values: if true, then the function values will be dumped.
+            show_id_class: if true, then the function id and class will be dumped.
+        """
+
+        shape_str: str = ', '.join(str(x) for x in self._shape)
+
+        if show_id_class:
+            print(f'{prefix}id: {id(self)}')
+            print(f'{prefix}class: {self.__class__.__name__}')
+        print(f'{prefix}usage: {self.count_usage()}')
+        print(f'{prefix}rvs: {self.number_of_rvs}')
+        print(f'{prefix}shape: ({shape_str})')
+        print(f'{prefix}states: {self._number_of_states}')
+        print(f'{prefix}guaranteed zero: {self._number_of_states - self.number_of_not_guaranteed_zero}')
+        print(f'{prefix}not guaranteed zero: {self.number_of_not_guaranteed_zero}')
+        print(f'{prefix}parameters: {self.number_of_parameters}')
+        if show_function_values:
+            next_prefix = prefix + indent
+            for key, param_idx, value in self.keys_with_param:
+                print(f'{next_prefix}{param_idx} {key} = {value}')
+
+
+class ZeroPotentialFunction(PotentialFunction):
+    """
+    A ZeroPotentialFunction behaves like a DensePotentialFunction
+    in that there is a parameter for each possible key.
+    However, a PGM user has no way to change parameter values.
+    Parameter values are always zero.
+    Despite the inability to change the value of the parameters,
+    no key is considered 'guaranteed zero'.
+
+    The primary use of a ZeroPotentialFunction is as a placeholder
+    within a factor, prior to parameter learning.
+    """
+    __slots__ = ()
+
+    def __init__(self, factor: Factor):
+        """
+        Create a potential function for the given factor.
+
+        Ensures:
+            Does not hold a reference to the given factor.
+            Does not register the potential function with the PGM.
+
+        Args:
+            factor: which factor is this potential function is compatible with.
+        """
+        super().__init__(factor)
+
+    @property
+    def number_of_not_guaranteed_zero(self) -> int:
+        return self.number_of_states
+
+    @property
+    def number_of_parameters(self) -> int:
+        return self.number_of_states
+
+    @property
+    def params(self) -> Iterable[Tuple[int, float]]:
+        for param_idx in range(self.number_of_parameters):
+            yield param_idx, 0
+
+    @property
+    def keys_with_param(self) -> Iterable[Tuple[Instance, int, float]]:
+        for param_idx, instance in enumerate(self.instances()):
+            yield instance, param_idx, 0
+
+    def __getitem__(self, key: Key) -> float:
+        self.check_key(key)
+        return 0
+
+    def param_value(self, param_idx: int) -> float:
+        if not self.valid_parameter(param_idx):
+            raise ValueError(f'invalid parameter index: {param_idx}')
+        return 0
+
+    def param_idx(self, key: Key) -> int:
+        return _natural_key_idx(self._shape, key)
+
+    def is_cpt(self, tolerance=DEFAULT_TOLERANCE) -> bool:
+        return True
+
+
+class DensePotentialFunction(PotentialFunction):
+    """
+    A dense (tabular) potential function.
+    There is one parameter for each valid key of the potential function.
+    The initial value for each parameter is zero.
+    It is possible independently change any value corresponding to any key.
+    """
+
+    def __init__(self, factor: Factor):
+        """
+        Create a potential function for the given factor.
+
+        Ensures:
+            Does not hold a reference to the given factor.
+            Does not register the potential function with the PGM.
+
+        Args:
+            factor: which factor is this potential function is compatible with.
+        """
+        super().__init__(factor)
+        self._values: NDArrayFloat64 = np.zeros(self.number_of_states, dtype=np.float64)
+
+    @property
+    def number_of_not_guaranteed_zero(self) -> int:
+        return self.number_of_states
+
+    @property
+    def number_of_parameters(self) -> int:
+        return self.number_of_states
+
+    def __getitem__(self, key: Key) -> float:
+        return self._values.item(self.param_idx(key))
+
+    def param_value(self, param_idx: int) -> float:
+        return self._values.item(param_idx)
+
+    def param_idx(self, key: Key) -> Optional[int]:
+        return self.natural_param_idx(key)
+
+    @property
+    def params(self) -> Iterable[Tuple[int, float]]:
+        # Type warning due to numpy type erasure
+        # noinspection PyTypeChecker
+        return enumerate(self._values)
+
+    @property
+    def keys_with_param(self) -> Iterable[Tuple[Instance, int, float]]:
+        for param_idx, key in enumerate(self.instances()):
+            value: float = self.param_value(param_idx)
+            yield key, param_idx, value
+
+    # Mutators
+
+    def __setitem__(self, key: Key, value: float) -> None:
+        """
+        Set the potential function value, for a given key.
+
+        Arg:
+            key: defines an instance in the state space of the potential function.
+            value: the new value of the potential function for the given key.
+
+        Assumes:
+            self.valid_key(key).
+        """
+        self._values[self.param_idx(key)] = value
+
+    def set_param_value(self, param_idx: int, value: float) -> None:
+        """
+        Set the parameter value.
+
+        Arg:
+            param_idx: is the index of the parameter.
+            value: the new value of the potential function for the given key.
+
+        Assumes:
+            self.valid_param(param_idx).
+        """
+        self._values[param_idx] = value
+
+    def clear(self) -> DensePotentialFunction:
+        """
+        Set all values of the potential function to zero.
+        
+        Returns:
+            self
+        """
+        return self.set_all(0)
+
+    def normalise_cpt(self) -> DensePotentialFunction:
+        """
+        Normalise the parameter values as if this was a CPT.
+        That is, treat the first random variable as the child and the others as parents;
+        for each combination of parent states, ensure the parameters over the child
+        states sum to 1 (or 0).
+        
+        Assumes:
+            There are no negative parameter values.
+            
+        Returns:
+            self
+        """
+        child = self._shape[0]
+        parents = self._shape[1:]
+        for parent_states in _combos_ranges(parents):
+            keys = [(c,) + parent_states for c in range(child)]
+            total = sum(self[key] for key in keys)
+            if total != 0 and total != 1:
+                for key in keys:
+                    self[key] /= total
+        return self
+
+    def normalise(self, grouping_positions: Sequence[int] = ()) -> DensePotentialFunction:
+        """
+        Convert the potential function to a CPT with 'grouping_positions' nominating
+        the parent random variables.
+        
+        I.e., for each possible key of the function with the same value at each
+        grouping position, the sum of values for matching keys in the factor is scaled
+        to be 1 (or 0).
+
+        Parameter 'grouping_positions' are indices into `self.shape`. For example, the
+        grouping positions of a factor with parent rvs 'conditioning_rvs', then
+        grouping_positions = [i for i, rv in enumerate(factor.rvs) if rv in conditioning_rvs].
+
+        Args:
+            grouping_positions: indices into `self.shape`.
+            
+        Returns:
+            self
+        """
+        _normalise_potential_function(self, grouping_positions)
+        return self
+
+    def set_iter(self, values: Iterable[float]) -> DensePotentialFunction:
+        """
+        Set the values of the potential function using the given iterator.
+
+        Mapping instances to *values is as follows:
+            Given Factor(rv1, rv2) where rv1 has 2 states, and rv2 has 3 states:
+            values[0] represents instance (0,0)
+            values[1] represents instance (0,1)
+            values[2] represents instance (0,2)
+            values[3] represents instance (1,0)
+            values[4] represents instance (1,1)
+            values[5] represents instance (1,2).
+
+        For example: to set to counts, starting from 1, use `self.set_iter(itertools.count(1))`.
+
+        Args:
+            values: an iterable providing values to use.
+            
+        Returns:
+            self
+        """
+        self._values = np.fromiter(
+            values,
+            dtype=np.float64,
+            count=self.number_of_parameters
+        )
+        return self
+
+    def set_stream(self, stream: Callable[[], float]) -> DensePotentialFunction:
+        """
+        Set the values of the potential function by repeatedly calling the stream function.
+        The order of values is the same as set_iter.
+
+        For example, to set to random numbers, use `self.set_stream(random.random)`.
+
+        Args:
+            stream: a callable taking no arguments, returning the values to use.
+            
+        Returns:
+            self
+        """
+        return self.set_iter(iter(stream, None))
+
+    def set_flat(self, *value: float) -> DensePotentialFunction:
+        """
+        Set the values of the potential function to the given values.
+        The order of values is the same as set_iter.
+        
+        Args:
+            *value: the values to use.
+            
+        Returns:
+            self
+            
+        Raises:
+            ValueError: if `len(value) != self.number_of_states`.
+        """
+        if len(value) != self.number_of_states:
+            raise ValueError(f'wrong number of values: expected {self.number_of_states}, got {len(value)}')
+        return self.set_iter(value)
+
+    def set_all(self, value: float) -> DensePotentialFunction:
+        """
+        Set all values of the potential function to the given value.
+
+        Args:
+            value: the value to use.
+            
+        Returns:
+            self
+        """
+        return self.set_iter(_repeat(value))
+
+    def set_uniform(self) -> DensePotentialFunction:
+        """
+        Set all values of the potential function 1/number_of_states.
+        
+        Returns:
+            self
+        """
+        return self.set_all(1.0 / self.number_of_states)
+
+
+class SparsePotentialFunction(PotentialFunction):
+    """
+    A sparse potential function.
+
+    There is one parameter for each non-zero key value.
+    The user may set the value for any key and parameters will
+    be automatically reconfigured as needed. Setting the value for
+    a key to zero disassociates the key from its parameter and
+    thus makes that key "guaranteed zero".
+    """
+
+    def __init__(self, factor: Factor):
+        """
+        Create a potential function for the given factor.
+
+        Ensures:
+            Does not hold a reference to the given factor.
+            Does not register the potential function with the PGM.
+
+        Args:
+            factor: which factor is this potential function is compatible with.
+        """
+        super().__init__(factor)
+        self._values: List[float] = []
+        self._params: Dict[Instance, int] = {}
+
+    @property
+    def number_of_not_guaranteed_zero(self) -> int:
+        return len(self._params)
+
+    @property
+    def number_of_parameters(self) -> int:
+        return len(self._params)
+
+    def __getitem__(self, key: Key) -> float:
+        param_idx: Optional[int] = self.param_idx(key)
+        if param_idx is None:
+            return 0
+        else:
+            return self._values[param_idx]
+
+    def param_value(self, param_idx: int) -> float:
+        return self._values[param_idx]
+
+    def param_idx(self, key: Key) -> Optional[int]:
+        return self._params.get(_key_to_instance(key))
+
+    @property
+    def params(self) -> Iterable[Tuple[int, float]]:
+        return enumerate(self._values)
+
+    @property
+    def keys_with_param(self) -> Iterable[Tuple[Instance, int, float]]:
+        for key, param_idx in self._params.items():
+            value: float = self._values[param_idx]
+            yield key, param_idx, value
+
+    # Mutators
+
+    def __setitem__(self, key: Key, value: float) -> None:
+        """
+        Set the potential function value, for a given key.
+        
+        If value is zero, then the key will become "guaranteed zero".
+
+        Arg:
+            key: defines an instance in the state space of the potential function.
+            value: the new value of the potential function for the given key.
+
+        Assumes:
+            self.valid_key(key).
+        """
+        instance: Instance = _key_to_instance(key)
+        param_idx: Optional[int] = self._params.get(instance)
+
+        if param_idx is None:
+            if value == 0:
+                # Nothing to do
+                return
+            param_idx = len(self._values)
+            self._values.append(value)
+            self._params[instance] = param_idx
+            return
+
+        if value != 0:
+            # Simple case
+            self._values[param_idx] = value
+            return
+
+        # This is the case where the key was associated with a parameter
+        # but the value is being set to zero, so we
+        # need to clear an existing non-zero parameter.
+        # This code operates by first ensuring the parameter is the last one,
+        # then popping the last parameter.
+
+        end: int = len(self._values) - 1
+        if param_idx != end:
+            # need to swap the parameter with the end.
+            self._values[param_idx] = self._values[end]
+
+            for test_instance, test_param_idx in self._params.items():
+                if test_param_idx == end:
+                    self._params[test_instance] = param_idx
+                    # There will only be one, so we can break now
+                    break
+
+        # Remove the parameter
+        self._values.pop()
+        self._params.pop(instance)
+
+    def set_param_value(self, param_idx: int, value: float) -> None:
+        """
+        Set the parameter value.
+
+        Arg:
+            param_idx: is the index of the parameter.
+            value: the new value of the potential function for the given key.
+
+        Assumes:
+            self.valid_param(param_idx).
+        """
+        self._values[param_idx] = value
+
+    def clear(self) -> SparsePotentialFunction:
+        """
+        Set all values of the potential function to zero.
+        
+        Returns:
+            self
+        """
+        self._values = []
+        self._params = {}
+        return self
+
+    def normalise_cpt(self) -> SparsePotentialFunction:
+        """
+        Normalise the parameter values as if this was a CPT.
+        That is, treat the first random variable as the child and the others as parents;
+        for each combination of parent states, ensure the parameters over
+        the child states sum to 1 (or 0).
+
+        Returns:
+            self
+        """
+        grouping_positions = list(range(1, self.number_of_rvs))
+        _normalise_potential_function(self, grouping_positions)
+        return self
+
+    def normalise(self, grouping_positions=()) -> SparsePotentialFunction:
+        """
+        Convert the potential function to a CPT with 'grouping_positions' nominating
+        the parent random variables.
+        
+        I.e., for each possible key of the function with the same value at each
+        grouping position, the sum of values for matching keys in the factor is scaled
+        to be 1 (or 0).
+
+        Parameter 'grouping_positions' are indices into function.shape. For example, the
+        grouping positions of a factor with parent rvs 'conditioning_rvs', then
+        grouping_positions = [i for i, rv in enumerate(factor.rvs) if rv in conditioning_rvs].
+
+        Returns:
+            self
+        """
+        _normalise_potential_function(self, grouping_positions)
+        return self
+
+    def set_iter(self, values: Iterable[float]) -> SparsePotentialFunction:
+        """
+        Set the values of the potential function using the given iterator.
+
+        Mapping instances to *values is as follows:
+            Given Factor(rv1, rv2) where rv1 has 2 states, and rv2 has 3 states:
+            values[0] represents instance (0,0)
+            values[1] represents instance (0,1)
+            values[2] represents instance (0,2)
+            values[3] represents instance (1,0)
+            values[4] represents instance (1,1)
+            values[5] represents instance (1,2).
+
+        For example: to set to counts, starting from 1, use `self.set_iter(itertools.count(1))`.
+
+        Args:
+            values: an iterable providing values to use.
+
+        Returns:
+            self
+        """
+        self.clear()
+        for instance, value in zip(self.instances(), values):
+            if value != 0:
+                self._params[instance] = len(self._values)
+                self._values.append(value)
+        return self
+
+    def set_stream(self, stream: Callable[[], float]) -> SparsePotentialFunction:
+        """
+        Set the values of the potential function by repeatedly calling the stream function.
+        The order of values is the same as set_iter.
+
+        For example, to set to random numbers, use `self.set_stream(random.random)`.
+
+        Args:
+            stream: a callable taking no arguments, returning the values to use.
+
+        Returns:
+            self
+        """
+        return self.set_iter(iter(stream, None))
+
+    def set_flat(self, *value: float) -> SparsePotentialFunction:
+        """
+        Set the values of the potential function to the given values.
+        The order of values is the same as set_iter.
+
+        Args:
+            *value: the values to use.
+
+        Returns:
+            self
+
+        Raises:
+            ValueError: if `len(value) != self.number_of_states`.
+        """
+        if len(value) != self.number_of_states:
+            raise ValueError(f'wrong number of values: expected {self.number_of_states}, got {len(value)}')
+        return self.set_iter(value)
+
+    def set_all(self, value: float) -> SparsePotentialFunction:
+        """
+        Set all values of the potential function to the given value.
+
+        Args:
+            value: the value to use.
+
+        Returns:
+            self
+        """
+        if value == 0:
+            return self.clear()
+        else:
+            return self.set_iter(_repeat(value))
+
+    def set_uniform(self) -> SparsePotentialFunction:
+        """
+        Set all values of the potential function 1/number_of_states.
+        
+        Returns:
+            self
+        """
+        return self.set_all(1.0 / self.number_of_states)
+
+
+class CompactPotentialFunction(PotentialFunction):
+    """
+    A compact potential function is sparse, where values for keys of
+    the same value are represented by a single parameter.
+
+    There is one parameter for each unique, non-zero key value.
+    The user may set the value for any key and parameters will
+    be automatically reconfigured as needed. Setting the value for
+    a key to zero disassociates the key from its parameter and
+    thus makes that key "guaranteed zero".
+    """
+
+    def __init__(self, factor: Factor):
+        """
+        Create a potential function for the given factor.
+
+        Ensures:
+            Does not hold a reference to the given factor.
+            Does not register the potential function with the PGM.
+
+        Args:
+            factor: which factor is this potential function is compatible with.
+        """
+        super().__init__(factor)
+        self._values: List[float] = []
+        self._counts: List[int] = []
+        self._map: Dict[Instance, int] = {}
+        self._inv_map: Dict[float, int] = {}
+
+    @property
+    def number_of_not_guaranteed_zero(self) -> int:
+        return len(self._map)
+
+    @property
+    def number_of_parameters(self) -> int:
+        return len(self._values)
+
+    @property
+    def params(self) -> Iterable[Tuple[int, float]]:
+        return enumerate(self._values)
+
+    @property
+    def keys_with_param(self) -> Iterable[Tuple[Instance, int, float]]:
+        for key, param_idx in self._map.items():
+            value: float = self._values[param_idx]
+            yield key, param_idx, value
+
+    def __getitem__(self, key: Key) -> float:
+        param_idx: Optional[int] = self.param_idx(key)
+        if param_idx is None:
+            return 0
+        else:
+            return self._values[param_idx]
+
+    def param_value(self, param_idx: int) -> float:
+        return self._values[param_idx]
+
+    def param_idx(self, key: Key) -> Optional[int]:
+        return self._map.get(_key_to_instance(key))
+
+    # Mutators
+
+    def __setitem__(self, key: Key, value: float) -> None:
+        """
+        Set the potential function value, for a given key.
+
+        If value is zero, then the key will become "guaranteed zero".
+        If the value is the same as an existing parameter value, then
+        that parameter will be reused.
+
+        Arg:
+            key: defines an instance in the state space of the potential function.
+            value: the new value of the potential function for the given key.
+
+        Assumes:
+            self.valid_key(key).
+        """
+        instance: Instance = _key_to_instance(key)
+
+        param_idx: Optional[int] = self._map.get(instance)
+
+        if param_idx is None:
+            # previous value for the key was zero
+            if value == 0:
+                # nothing to do
+                return
+            param_idx: Optional[int] = self._inv_map.get(value)
+            if param_idx is not None:
+                # the value already exists in the function, so reuse it
+                self._map[instance] = param_idx
+                self._counts[param_idx] += 1
+            else:
+                # need to allocate a new value
+                new_param_idx: int = len(self._values)
+                self._values.append(value)
+                self._counts.append(1)
+                self._inv_map[value] = new_param_idx
+                self._map[instance] = new_param_idx
+            return
+
+        # the key previously had a non-zero value
+        prev_value: float = self._values[param_idx]
+
+        if value == prev_value:
+            # nothing to do
+            return
+
+        reference_count: int = self._counts[param_idx]
+        if reference_count == 1:
+            if value != 0:
+                # simple case
+                self._values[param_idx] = value
+            else:
+                # need to remove the parameter
+                self._remove_param(param_idx)
+                self._map.pop(instance)
+                self._inv_map.pop(prev_value)
+            return
+
+        # decrement the reference count of the previous parameter
+        self._counts[param_idx] = reference_count - 1
+
+        # allocate the key to a different parameter
+        param_idx: Optional[int] = self._inv_map.get(value)
+        if param_idx is not None:
+            # the value already exists in the function, so reuse it
+            self._map[instance] = param_idx
+            self._counts[param_idx] += 1
+        else:
+            # need to allocate a new value
+            new_param_idx: int = len(self._values)
+            self._values.append(value)
+            self._counts.append(1)
+            self._inv_map[value] = new_param_idx
+            self._map[instance] = new_param_idx
+
+    def set_iter(self, values: Iterable[float]) -> CompactPotentialFunction:
+        """
+        Set the values of the potential function using the given iterator.
+
+        Mapping instances to *values is as follows:
+            Given Factor(rv1, rv2) where rv1 has 2 states, and rv2 has 3 states:
+            values[0] represents instance (0,0)
+            values[1] represents instance (0,1)
+            values[2] represents instance (0,2)
+            values[3] represents instance (1,0)
+            values[4] represents instance (1,1)
+            values[5] represents instance (1,2).
+
+        For example: to set to counts, starting from 1, use `self.set_iter(itertools.count(1))`.
+
+        Args:
+            values: an iterable providing values to use.
+
+        Returns:
+            self
+        """
+        self.clear()
+        for instance, value in zip(self.instances(), values):
+            self[instance] = value
+        return self
+
+    def set_stream(self, stream: Callable[[], float]) -> CompactPotentialFunction:
+        """
+        Set the values of the potential function by repeatedly calling the stream function.
+        The order of values is the same as set_iter.
+
+        For example, to set to random numbers, use `self.set_stream(random.random)`.
+
+        Args:
+            stream: a callable taking no arguments, returning the values to use.
+
+        Returns:
+            self
+        """
+        return self.set_iter(iter(stream, None))
+
+    def set_flat(self, *value: float) -> CompactPotentialFunction:
+        """
+        Set the values of the potential function to the given values.
+        The order of values is the same as set_iter.
+
+        Args:
+            *value: the values to use.
+
+        Returns:
+            self
+
+        Raises:
+            ValueError: if `len(value) != self.number_of_states`.
+        """
+        if len(value) != self.number_of_states:
+            raise ValueError(f'wrong number of values: expected {self.number_of_states}, got {len(value)}')
+        return self.set_iter(value)
+
+    def set_all(self, value: float) -> CompactPotentialFunction:
+        """
+        Set all values of the potential function to the given value.
+
+        Args:
+            value: the value to use.
+
+        Returns:
+            self
+        """
+        self.clear()
+        if value != 0:
+            self._values = [value]
+            self._counts = [self.number_of_states]
+            self._inv_map = {value: 0}
+            self._map = {instance: 0 for instance in self.instances()}
+        return self
+
+    def set_uniform(self) -> CompactPotentialFunction:
+        """
+        Set all values of the potential function 1/number_of_states.
+
+        Returns:
+            self
+        """
+        return self.set_all(1.0 / self.number_of_states)
+
+    def clear(self) -> CompactPotentialFunction:
+        """
+        Set all values of the potential function to zero.
+        
+        Returns:
+            self
+        """
+        self._values = []
+        self._counts = []
+        self._map = {}
+        self._inv_map = {}
+        return self
+
+    def _remove_param(self, param_idx: int) -> None:
+        """
+        Remove the indexed parameter from self._params and self._counts.
+        If the parameter is not at the end of the list of parameters
+        then it will be swapped with the last parameter in the list.
+        """
+
+        # ensure the parameter is at the end of the list
+        end: int = len(self._values) - 1
+        if param_idx != end:
+            # swap `param_idx` with `end`
+            end_value: float = self._values[end]
+            self._values[param_idx] = end_value
+            self._counts[param_idx] = self._counts[end]
+            self._inv_map[end_value] = param_idx
+            for instance, instance_param_idx in self._map.items():
+                if instance_param_idx == end:
+                    self._map[instance] = param_idx
+
+        # remove the end parameter
+        self._values.pop()
+        self._counts.pop()
+
+
+class ClausePotentialFunction(PotentialFunction):
+    """
+    A clause potential function represents a clause From a CNF formula.
+    I.e. a clause over variables X, Y, Z, is a disjunction of the form: 'X=x or Y=y or Z=z'.
+
+    A clause potential function is guaranteed zero for a key where the clause is false,
+    i.e., when 'X != x and Y != y and Z != z'.
+
+    For keys where the clause is true, the value of the potential function
+    is given by the only parameter of the potential function. That parameter
+    is called the clause 'weight' and is notionally 1.
+
+    The weight of a clause is permitted to be zero, but that is _not_ equivalent to
+    guaranteed-zero.
+    """
+
+    def __init__(self, factor: Factor, key: Key, weight: float = 1):
+        """
+        Create a clause potential function for the given factor.
+
+        Ensures:
+            Does not hold a reference to the given factor.
+            Does not register the potential function with the PGM.
+
+        Raises:
+             KeyError: if the key is not valid for the shape of the factor.
+
+        Args:
+            factor: which factor is this potential function is compatible with.
+            key: defines the random variable states of the clause.
+        """
+        super().__init__(factor)
+        self._weight: float = weight
+        self._clause: Instance = self.check_key(key)
+        self._num_not_guaranteed_zero: int = _zero_space(self.shape)
+
+    @property
+    def number_of_not_guaranteed_zero(self) -> int:
+        return self._num_not_guaranteed_zero
+
+    @property
+    def number_of_parameters(self) -> int:
+        return 1
+
+    @property
+    def params(self) -> Iterable[Tuple[int, float]]:
+        return ((0, self._weight),)
+
+    @property
+    def keys_with_param(self) -> Iterable[Tuple[Instance, int, float]]:
+        value = self._weight
+        for i in range(self.number_of_rvs):
+            key = list(self._clause)
+            for j in range(self.shape[i]):
+                key[i] = j
+            yield tuple(key), 0, value
+
+    def __getitem__(self, key: Key) -> float:
+        instance: Instance = self.check_key(key)
+        for key_state_idx, clause_state_idx in zip(instance, self._clause):
+            if key_state_idx == clause_state_idx:
+                return self._weight
+        return 0
+
+    def param_value(self, param_idx: int) -> float:
+        if param_idx != 0:
+            raise IndexError(param_idx)
+        return self._weight
+
+    def param_idx(self, key: Key) -> Optional[int]:
+        instance: Instance = _key_to_instance(key)
+        if instance == self._clause:
+            return 0
+        else:
+            return None
+
+    def is_cpt(self, tolerance=DEFAULT_TOLERANCE) -> bool:
+        """
+        A ClausePotentialFunction can only be a CTP when all entries are zero.
+        """
+        return -tolerance <= self._weight <= tolerance
+
+    def is_sparse(self) -> bool:
+        return True
+
+    @property
+    def weight(self) -> float:
+        """
+        Returns:
+            the "weight" parameter defining the potential function.
+        """
+        return self._weight
+
+    @property
+    def clause(self) -> Instance:
+        """
+        Returns:
+            the clause defining the potential function.
+        """
+        return self._clause
+
+    # Mutators
+
+    @weight.setter
+    def weight(self, value: float) -> None:
+        """
+        Set the weight parameter to the given value.
+        """
+        self._weight = value
+
+    @clause.setter
+    def clause(self, key: Key) -> None:
+        """
+        Set the clause to the given key.
+
+        Raises:
+             KeyError: if the key is not valid for the shape of the factor.
+        """
+        self._clause = self.check_key(key)
+
+
+class CPTPotentialFunction(PotentialFunction):
+    """
+    A potential function implementing a sparse Conditional Probability Table (CPT).
+
+    The first random variable in the signature is the child, and the remaining random
+    variables are parents.
+
+    For each instantiation of the parent random variables there is a Conditioned Probability
+    Distribution (CPD) over the states of the child random variable.
+
+    If a CPD is not provided for a parent instantiation, then that parent instantiation
+    is taken to have probability zero (i.e., all values of the CPD are guaranteed zero).
+    """
+
+    def __init__(self, factor: Factor, tolerance: float):
+        """
+        Create a CPT potential function for the given factor.
+
+        Ensures:
+            Does not hold a reference to the given factor.
+            Does not register the potential function with the PGM.
+
+        Args:
+            factor: which factor is this potential function is compatible with.
+            tolerance: a tolerance when testing if values are equal to zero or one.
+
+        Raises:
+            ValueError: if tolerance is negative.
+        """
+        super().__init__(factor)
+
+        if tolerance < 0:
+            raise ValueError('tolerance cannot be negative')
+
+        self._child_size: int = self.shape[0]
+        self._parent_shape: Shape = self.shape[1:]
+        self._map: Dict[Instance, int] = {}
+        self._values: List[float] = []
+        self._inv_map: List[Instance] = []
+        self._tolerance = tolerance
+
+    @property
+    def number_of_not_guaranteed_zero(self) -> int:
+        return len(self._values)
+
+    @property
+    def number_of_parameters(self) -> int:
+        return len(self._values)
+
+    def is_cpt(self, tolerance=DEFAULT_TOLERANCE) -> bool:
+        if tolerance >= self._tolerance:
+            return True
+        else:
+            # The requested tolerance is tighter than ensured.
+            # Need to use the default method
+            return super().is_cpt(tolerance)
+
+    @property
+    def params(self) -> Iterable[Tuple[int, float]]:
+        return enumerate(self._values)
+
+    @property
+    def keys_with_param(self) -> Iterable[Tuple[Instance, int, float]]:
+        child_size: int = self._child_size
+        for param_idx, value in enumerate(self._values):
+            parent: Instance = self._inv_map[param_idx // child_size]
+            key: Instance = (param_idx % child_size,) + tuple(parent)
+            yield key, param_idx, value
+
+    def __getitem__(self, key: Key) -> float:
+        param_idx: Optional[int] = self.param_idx(key)
+        if param_idx is None:
+            return 0
+        else:
+            return self._values[param_idx]
+
+    def param_value(self, param_idx: int) -> float:
+        return self._values[param_idx]
+
+    def param_idx(self, key: Key) -> Optional[int]:
+        instance: Instance = self.check_key(key)
+        offset: Optional[int] = self._map.get(instance[1:])
+        if offset is None:
+            return None
+        else:
+            return offset + instance[0]
+
+    @property
+    def parent_shape(self) -> Shape:
+        """
+        What is the shape of the parents.
+        """
+        return self._parent_shape
+
+    @property
+    def number_of_parent_states(self) -> int:
+        """
+        How many combinations of parent states.
+        """
+        return _multiply(self._parent_shape)
+
+    @property
+    def number_of_child_states(self) -> int:
+        """
+        Number of child random variable states.
+
+        This is the same as the number of values in each conditional
+        probability distribution. This is equivalent to `self.shape[0]`.
+
+        Returns:
+            the number of child states.
+        """
+        return self._child_size
+
+    def get_cpd(self, parent_states: Key) -> List[float]:
+        """
+        Get the CPD conditioned on parent states indicated by `parent_states`.
+        
+        Args:
+            parent_states: indicates the parent states.
+
+        Returns:
+            The conditioned probability distribution.
+        """
+        parent_instance: Instance = check_key(self._parent_shape, parent_states)
+        offset: Optional[int] = self._map.get(parent_instance)
+        child_size: int = self._child_size
+        if offset is None:
+            return [0] * child_size
+        else:
+            return self._values[offset:offset + child_size]
+
+    def cpds(self) -> Iterable[Tuple[Instance, Sequence[float]]]:
+        """
+        Iterate over (parent_states, cpd) tuples.
+        This will exclude zero CPDs.
+        Do not change CPDs to (or from) zero while iterating over them.
+        
+        Get the CPD conditioned on parent states indicated by `parent_states`.
+        
+        Returns:
+            an iterator over pairs (instance, cpd) where,
+            instance: is indicates the state of the parent random variables.
+            cpd: is the conditioned probability distribution, for the parent instance.
+        """
+        for parent_instance, offset in self._map.items():
+            cpd = self._values[offset:offset + self._child_size]
+            yield parent_instance, cpd
+
+    # Mutators
+
+    def clear(self) -> CPTPotentialFunction:
+        """
+        Set all values of the potential function to zero.
+        
+        Returns:
+            self
+        """
+        self._map = {}
+        self._values = []
+        self._inv_map = []
+        return self
+
+    def set_uniform(self) -> CPTPotentialFunction:
+        """
+        Set each CPD to a uniform distribution.
+        
+        Returns:
+            self
+        """
+        self.clear()
+        for parent_states in self.parent_instances():
+            self.set_cpd_uniform(parent_states)
+        return self
+
+    def set_random(self, random: Callable[[], float], sparsity: float = 0) -> CPTPotentialFunction:
+        """
+        Set the values of the potential function to random CPDs.
+        
+        Args:
+            random: is a stream of random numbers, assumed uniformly distributed in the interval [0, 1].
+            sparsity: sets the expected proportion of probability values that are zero.
+            
+        Returns:
+            self
+        """
+        self.clear()
+        for parent_states in self.parent_instances():
+            self.set_cpd_random(parent_states, random, sparsity)
+        return self
+
+    def set(self, *rows: Tuple[Key, Sequence[float]]) -> CPTPotentialFunction:
+        """
+        Calls self.set_cpd(parent_states, cpd) for each row (parent_states, cpd)
+        in rows. Any unmentioned parent states will have zero probabilities.
+
+        Example usage, assuming three Boolean random variables:
+            pgm.Factor(x, y, z).set_cpt().set(
+                # y  z    x[0] x[1]
+                ((0, 0), (0.1, 0.9)),
+                ((0, 1), (0.1, 0.9)),
+                ((1, 0), (0.1, 0.9)),
+                ((1, 1), (0.1, 0.9))
+            )
+        
+        Args:
+            *rows: are tuples (key, cpd) used to set the potential function values.
+            
+        Raises:
+            ValueError: if a CPD is not valid.
+            
+        Returns:
+            self
+        """
+        self.clear()
+        for parent_states, cpd in rows:
+            self.set_cpd(parent_states, cpd)
+        return self
+
+    def set_all(self, *cpds: Optional[Sequence[float]]) -> CPTPotentialFunction:
+        """
+        Set all CPDs using the given `cpds` which are taken to be in order of the parent states
+        with the last variable of the parent changing state most rapidly, as per parent_states().
+
+        If insufficient CPDs are provided then the remaining parent instantiations are taken to be
+        impossible (i.e. not set and guaranteed zero).
+        If too many CPDs are provided then the extras are ignored.
+        Any list entry may be None, indicating 'guaranteed zero' for the associated parent states.
+
+        Args:
+            *cpds: are the CPDs used to set the potential function values.
+            
+        Raises:
+            ValueError: if a CPD is not valid.
+            
+        Returns:
+            self
+        """
+        self.clear()
+        for parent_states, cpd in zip(self.parent_instances(), cpds):
+            self.set_cpd(parent_states, cpd)
+        return self
+
+    def set_cpd(self, parent_states: Key, cpd: Optional[Sequence[float]]) -> CPTPotentialFunction:
+        """
+        Set the CPD of the given parent states to the given cpd.
+        If cpd is None or all zeros, then this is equivalent to clear_cpd(parent_states).
+        
+        Args:
+            parent_states: indicates the CPD to set, based on the parent states.
+            cpd: is a conditioned probability distribution, or None indicating `guaranteed zero`.
+            
+        Raises:
+            ValueError: if the CPD is not valid.
+            KeyError if the key is not valid.
+
+        Returns:
+            self
+        """
+        parent_instance: Instance = check_key(self._parent_shape, parent_states)
+
+        if cpd is None:
+            self._clear_cpd(parent_instance)
+            return self
+
+        if len(cpd) != self._child_size:
+            raise ValueError(f'CPD incorrect size: expected {self._child_size}, got {len(cpd)}')
+        if not all(0 <= value <= 1 for value in cpd):
+            raise ValueError(f'not a valid CPD: {cpd!r}')
+
+        total_value = sum(cpd)
+        if total_value < self._tolerance:
+            self._clear_cpd(parent_instance)
+            return self
+
+        if total_value < 1 - self._tolerance or total_value > 1 + self._tolerance:
+            raise ValueError(f'not a valid CPD: sum of values = {total_value}')
+
+        offset: Optional[int] = self._map.get(parent_instance)
+        child_size: int = self._child_size
+        if offset is None:
+            offset = len(self._values)
+            self._values.extend(cpd)
+            self._map[parent_instance] = offset
+            self._inv_map.append(parent_instance)
+        else:
+            self._values[offset:offset + child_size] = cpd
+
+        return self
+
+    def clear_cpd(self, parent_states: Key) -> CPTPotentialFunction:
+        """
+        Set the CPD of the given parent_states to all 'guaranteed zero'.
+        
+        Args:
+            parent_states: indicates the CPD to clear, based on the parent states.
+
+        Raises:
+            KeyError if the key is not valid.
+
+        Returns:
+            self
+        """
+        parent_instance: Instance = check_key(self._parent_shape, parent_states)
+        self._clear_cpd(parent_instance)
+        return self
+
+    def set_cpd_uniform(self, parent_states: Key) -> CPTPotentialFunction:
+        """
+        Set the CPD of the given parent_states to a uniform CPD.
+        
+        Args:
+            parent_states: indicates the CPD to clear, based on the parent states.
+
+        Raises:
+            KeyError if the key is not valid.
+
+        Returns:
+            self
+        """
+        num_states = self.number_of_child_states
+        cpd = [1.0 / num_states] * num_states
+        return self.set_cpd(parent_states, cpd)
+
+    def set_cpd_random(
+            self,
+            parent_states: Key,
+            random: Callable[[], float],
+            sparsity: float = 0,
+    ) -> CPTPotentialFunction:
+        """
+        Set the CPD of the given parent_states to a random CPD.
+
+        Args:
+            parent_states: identifies the CPD being set.
+            random: is a stream of random numbers, assumed uniformly distributed in the interval [0, 1].
+            sparsity: sets the expected proportion of probability values that are zero.
+
+        Returns:
+            self
+        """
+        cpd = np.zeros(self.number_of_child_states, dtype=np.float64)
+        if sparsity <= 0:
+            for i in range(len(cpd)):
+                cpd[i] = 0.0000001 + random()
+        else:
+            for i in range(len(cpd)):
+                if random() > sparsity:
+                    cpd[i] = 0.0000001 + random()
+        sum_value = np.sum(cpd)
+        if sum_value > 0:
+            cpd /= sum_value
+            return self.set_cpd(parent_states, cpd)
+        else:
+            return self.clear_cpd(parent_states)
+
+    def _clear_cpd(self, parent_instance: Instance) -> None:
+        """
+        Remove the parent instance from the parameters
+        """
+        offset: Optional[int] = self._map.get(parent_instance)
+        if offset is None:
+            # nothing to do
+            return
+
+        child_size: int = self._child_size
+        end_offset: int = len(self._values) - child_size
+        if offset != end_offset:
+            # need to swap parameters
+            end_cpd = self._values[end_offset:]
+            end_parent_instance = self._inv_map[-1]
+
+            self._values[offset:offset + child_size] = end_cpd
+            self._map[end_parent_instance] = offset
+            self._inv_map[offset // child_size] = end_parent_instance
+
+        self._map.pop(parent_instance)
+        self._inv_map.pop()
+        for _ in range(child_size):
+            self._values.pop()
+
+
+def default_pgm_name(pgm: PGM) -> str:
+    """
+    If no name is provided to a PGM constructor, then this will be the default name for the PGM.
+
+    Args:
+        pgm: a PGM object.
+
+    Returns:
+        a name for the PGM if none is given at construction time.
+    """
+    return 'PGM_' + str(id(pgm))
+
+
+def check_key(shape: Shape, key: Key) -> Instance:
+    """
+    Convert the key into an instance.
+
+    Args:
+        shape: the shape defining the state space.
+        key: a key into the state space.
+
+    Returns:
+        A instance from the state space, as a tuple of state indexes, co-indexed with the given shape.
+
+    Raises:
+        KeyError if the key is not valid.
+    """
+    _key: Instance = _key_to_instance(key)
+    if len(_key) != len(shape):
+        raise KeyError(f'not a valid key for shape {shape}: {key!r}')
+    if all((0 <= i <= m) for i, m in zip(_key, shape)):
+        return tuple(_key)
+    raise KeyError(f'not a valid key for shape {shape}: {key!r}')
+
+
+def valid_key(shape: Shape, key: Key) -> bool:
+    """
+    Is the given key valid.
+
+    Args:
+        shape: the shape defining the state space.
+        key: a key into the state space.
+
+    Returns:
+        True only if tke key is valid for the given shape.
+    """
+    try:
+        check_key(shape, key)
+        return True
+    except KeyError:
+        return False
+
+
+def number_of_states(*rvs: RandomVariable) -> int:
+    """
+    Returns:
+        What is the size of the state space, i.e., `multiply(len(rv) for rv in self.rvs)`.
+    """
+    return _multiply(len(rv) for rv in rvs)
+
+
+def rv_instances(*rvs: RandomVariable, flip: bool = False) -> Iterable[Instance]:
+    """
+    Enumerate instances of the given random variables.
+
+    Each instance is a tuples of state indexes, co-indexed with the given random variables.
+
+    The order is the natural index order (i.e., last random variable changing most quickly).
+
+    Args:
+        flip: if true, then first random variable changes most quickly.
+
+    Returns:
+        an iteration over tuples, each tuple holds state indexes
+        co-indexed with the given random variables.
+    """
+    shape = [len(rv) for rv in rvs]
+    return _combos_ranges(shape, flip=not flip)
+
+
+def rv_instances_as_indicators(*rvs: RandomVariable, flip: bool = False) -> Iterable[Sequence[Indicator]]:
+    """
+    Enumerate instances of the given random variables.
+
+    Each instance is a tuples of indicators, co-indexed with the given random variables.
+
+    The order is the natural index order (i.e., last random variable changing most quickly).
+
+    Args:
+        flip: if true, then first random variable changes most quickly.
+
+    Returns:
+        an iteration over tuples, each tuples holds random variable indicators
+        co-indexed with the given random variables.
+    """
+    return _combos(rvs, flip=not flip)
+
+
+def _key_to_instance(key: Key) -> Instance:
+    """
+    Convert a key to an instance.
+
+    Args:
+        key: a key into a state space.
+
+    Returns:
+        A instance from the state space, as a tuple of state indexes, co-indexed with the given shape.
+
+    Assumes:
+        The key is valid for the implied state space.
+    """
+    if isinstance(key, int):
+        return (key,)
+    else:
+        return tuple(key)
+
+
+def _natural_key_idx(shape: Shape, key: Key) -> int:
+    """
+    What is the natural index of the given key, assuming the given shape.
+
+    Args:
+        shape: the shape defining the state space.
+        key: a key into the state space.
+
+    Returns:
+        an index as per enumerated instances in their natural order, i.e.
+        last random variable changing most quickly.
+
+    Assumes:
+        The key is valid for the shape.
+    """
+    instance: Instance = _key_to_instance(key)
+    result: int = instance[0]
+    for s, i in zip(shape[1:], instance[1:]):
+        result = result * s + i
+    return result
+
+
+def _zero_space(shape: Shape) -> int:
+    """
+    Return the size of the zero space of the given shape. This is the number
+    of possible instances in the state space that do not have a zero in the instance.
+
+    The zero space is the same as the shape but with one less state
+    for each random variable.
+
+    Args:
+        shape: the shape defining the state space.
+
+    Returns:
+        the size of the zero space.
+    """
+    return _multiply(x - 1 for x in shape)
+
+
+def _normalise_potential_function(
+        function: Union[DensePotentialFunction, SparsePotentialFunction],
+        grouping_positions: Sequence[int],
+) -> None:
+    """
+    Convert the potential function to a CPT with 'grouping_positions' nominating
+    the parent random variables.
+    
+    I.e., for each possible key of the function with the same value at each
+    grouping position, the sum of values for matching keys in the factor is scaled
+    to be 1 (or 0).
+
+    Parameter 'grouping_positions' are indices into `function.shape`. For example, the
+    grouping positions of a factor with parent rvs 'conditioning_rvs', then
+    grouping_positions = [i for i, rv in enumerate(factor.rvs) if rv in conditioning_rvs].
+
+    Args:
+        function: the potential function to normalise.
+        grouping_positions: indices into `function.shape`.
+    """
+    if len(grouping_positions) == 0:
+        total = sum(
+            function.param_value(param_idx)
+            for param_idx in range(function.number_of_parameters)
+        )
+        if total != 0 and total != 1:
+            for param_key, param_idx, param_value in function.keys_with_param:
+                function.set_param_value(param_idx, param_value / total)
+    else:
+        group_sum = {}
+        for param_key, param_idx, param_value in function.keys_with_param:
+            group = tuple(param_key[i] for i in grouping_positions)
+            group_sum[group] = group_sum.get(group, 0) + param_value
+
+        for param_key, param_idx, param_value in function.keys_with_param:
+            group = tuple(param_key[i] for i in grouping_positions)
+            total = group_sum[group]
+            if total > 0:
+                function.set_param_value(param_idx, param_value / total)
+
+
+_CLEAN_CHARS: Set[str] = set('abcdefghijklmnopqrstuvwxyzABCDEFGHIJKLMNOPQRSTUVWXYZ0123456789_-+~?.')
+
+
+def _clean_str(s) -> str:
+    """
+    Quote a string if empty or not all characters are in _CLEAN_CHARS.
+    This is used when rendering indicators.
+    """
+    s = str(s)
+    if len(s) == 0 or not all(c in _CLEAN_CHARS for c in s):
+        return repr(s)
+    else:
+        return s