The root-sink expressibility lemma (A2) #
Statements split from the original Statements.lean skeleton (one file per proving task);
see AUDIT-NOTES A2 and papers/inflation-nontermination/paper/sections/12-pair-source.tex
(Lemma lem:rootsink) for the mathematics.
isAISet_iff_decomposition, isAISet_glue and gExpFeasible_iff_gAIFeasible are proved.
Two statements of the original skeleton were false as written and have been removed:
gInjectable_iff_raw needed 1 ≤ t (the corrected form is gInjectable_iff_raw_of_one_le,
and not_forall_gInjectable_iff_raw refutes the unrestricted one), and expressible_iff_ai
holds only for targets satisfying the ancestral-independence prescriptions (its intended
content is Expressible.isAISet together with gExpFeasible_iff_gAIFeasible). The two
consequences flip_gExpFeasible and compatible_gExpFeasible are proved at the end of this
file from the root-sink lemma. Everything in this file is proved.
Auxiliary material #
Everything in this section is new; the five statements of the task are unchanged and appear below in their original form.
Incidence, ancestors and the shared-parent relation #
Every copied observation has at least one copied latent ancestor.
Connectivity in the shared-parent graph, on the ambient type #
If everything connected to a inside T already lies in the smaller set S, then
connectivity inside T from a is connectivity inside S.
Components and decompositions #
A set with an AI decomposition is an AI set: the component of a member is contained in any block containing it, because a step of the shared-parent graph cannot leave a block.
A chain from X to Y inside X ∪ Y ∪ Z yields a trail that is active given Z:
cut the chain at the first place where it returns to X or reaches Y.
Injectability #
AUDIT-NOTES A1/A2: for 1 ≤ t the working definition of injectability agrees with the
primitive Wolfe–Spekkens–Fritz condition. The hypothesis 1 ≤ t is needed for the backward
direction: it supplies a copy index for the sources that the set leaves unconstrained.
Pushforward helpers #
Restricting further can only increase the mass of a fibre: the marginal of a nonnegative weight on a smaller set dominates the marginal on a larger one.
The marginal of an AI set #
Under a witness satisfying the ancestral-independence prescriptions, the marginal on a set with an AI decomposition is the product of the injectable marginals of its blocks.
Under d-separation, the component of a member of X ∪ Y ∪ Z in the shared-parent graph
lies inside X ∪ Z or inside Y ∪ Z: a shortest path joining X to Y inside a component
has all its interior in Z, so it is an active trail.
The unrestricted form of gInjectable_iff_raw_of_one_le is false: at t = 0 a scenario
with a vertex has no copied observations at all (every vertex is incident to a source, and
there is no map from a nonempty type to Fin 0), so the only set of copied observations is
∅. That set satisfies the primitive Wolfe–Spekkens–Fritz condition vacuously, while
GInjectable ∅ asks for a global index assignment Γ.Edge → Fin 0, which does not exist.
A2: the root-sink expressibility lemma #
An AI set is exactly a set presented as a union of pairwise ancestrally independent injectable blocks: the blocks may be taken to be the connected components of the shared-parent graph (AUDIT-NOTES A2).
AUDIT-NOTES A2, the geometric half of the root-sink lemma. If X ∪ Z and Y ∪ Z are AI
sets and X is d-separated from Y by Z, then every connected component of the
shared-parent graph on X ∪ Y ∪ Z lies inside X ∪ Z or inside Y ∪ Z, hence is
injectable, so X ∪ Y ∪ Z is again an AI set. (Take a shortest path inside a component from
X to Y: its internal vertices lie in Z, so it is an active trail.)
Consequences of the glue lemma #
Subsets of AI sets are AI sets, so every expressible set is an AI set.
From the AI prescriptions to the expressible prescriptions #
Intersecting the blocks of an AI decomposition of W with a subset A gives an AI
decomposition of A.
Equations
Instances For
AUDIT-NOTES A2: a witness satisfying the injectable and ancestral-independence prescriptions satisfies every expressible prescription.
From the expressible prescriptions to the AI prescriptions #
Expressibility transports along an equality of sets.
Ancestrally independent sets are d-separated by the empty set: a trail between them
would need an interior vertex, which the empty conditioning set cannot supply.
Two ancestrally independent expressible sets glue, with Z = ∅, to their union.
Under a witness satisfying the expressible prescriptions, ancestrally independent sets are independent.
Iterated gluing with Z = ∅: the marginal on a union of pairwise ancestrally
independent injectable sets is the product of their marginals.
AUDIT-NOTES A2: a witness satisfying the expressible prescriptions satisfies the ancestral-independence prescriptions.
AUDIT-NOTES A2, the consequence for the hierarchies: for pair-source scenarios, which are root-sink scenarios, the recursively expressible hierarchy coincides with the ancestral-independence hierarchy at every order.
Local flips preserve the recursively expressible test, via gExpFeasible_iff_gAIFeasible
and flip_gAIFeasible. (This does not follow from the flip structure alone: flipping the
conditioning coordinates of a glue prescription does not commute with the conditional law;
the identification of the glue prescriptions with products of injectable marginals is what
makes it true.)
A genuine model gives a witness satisfying every expressible prescription: run the
inflated model (compatible_gAIFeasible) and use the root-sink lemma.