The measurable kernel evaluator over pattern-local latents (R4 evaluator layer, #107) #
The second R4 layer, on the settled #137 interface: a kernel family assigns to every relation symbol and every abstract equality pattern a measurable Boolean function of the pattern-local latents; the evaluator reads each coordinate of the infinite structure off the global latent source through its local window. No latent source measure, sampler pushforward, dissociation, or representation theorem here — those are the subsequent layers; this file settles the evaluator and its equivariance.
RelKernelFamily S— the label-free representing kernelsf_{r,π}, measurable on the finite productPatternLatentIndex π → ℝ;RelCoord.localLatents— the local latent window of a coordinate: the global source restricted alongpatternLatentIndexEquivCoordand the canonical projection;RelKernelFamily.evalStructure— the evaluated infinite structure, measurable in the global source;RelKernelFamily.evalStructure_relabel— equivariance: relabeling the evaluated structure is evaluating at the relabeled latents — the statement that will make exchangeability of the evaluated law a coordinatewise check.
Kernel families #
A representing kernel family: for every relation symbol and every abstract equality pattern, a measurable Boolean function of the pattern-local latents. Label-free by construction — the kernel sees only the pattern and its blocks, never vertex labels.
The kernel at a relation symbol and pattern.
Each kernel is measurable on the finite latent product.
Instances For
The local latent window #
The local latent window of a coordinate: the global latent source read through the coordinate's pattern-local indices (via the canonical equivalence and projection).
Equations
- c.localLatents ω A = ω ((RelSignature.patternLatentIndexEquivCoord c) A).toLatentIndex
Instances For
The evaluator #
The evaluated infinite structure: each coordinate is its kernel applied to its local latent window.
Equations
- F.evalStructure ω c = F.toFun c.fst c.equalityPattern (c.localLatents ω)
Instances For
The evaluator is measurable in the global latent source.
Equivariance #
The value of a support block (definitional): patternQuotientEquivSupport sends the
block of position i to the sort-tagged value at i.
Membership characterization of the labeled avatar of a block-set: a tagged value lies
in patternLatentIndexEquivCoord c A iff it is the support value of one of the blocks
in A.
Membership characterization of the relabeling equivalence: the forward direction of
CoordLatentIndex.congrMap is the image under Sigma.map id.
Transport of pattern-local indices along rfl is the identity.
Cast taming: membership of a block in a transported pattern-local index reduces to
membership in the original — both sides typecheck because the two patterns share the
positions Fin (S.arity r).
The transport square: the labeled avatar of a block-set of the relabeled coordinate
is the relabeling (CoordLatentIndex.congrMap) of the labeled avatar of the same block-set
read through the pattern identification RelCoord.equalityPattern_map.
The local latent window is equivariant: the window of the relabeled coordinate at a block-set is the window of the original coordinate, read off the source precomposed with the global latent-index action, at the transported block-set.
Kernel applications transport along pattern equalities: if the patterns agree and the
latent arguments correspond under PatternLatentIndex.congr, the kernel values agree.
Equivariance of the evaluator: relabeling the evaluated structure by a sortwise
family of permutations is evaluating at the source precomposed with the latent-index action
of the permutations. Orientation: relabel σ reads coordinate c at RelCoord.map σ c,
so the source is precomposed with LatentIndex.map σ (the forward action, not σ⁻¹) —
relabeling the structure = precomposing the latent source with the latent-index action.