Documentation

Graphon.RelAustinEnriched

The Austin base, its action, and the base-extended bundle: unit 2 (R4 converse, #107, #197) #

Route A (Austin) only. No Kallenberg machinery, and nothing here asserts that the two routes' outputs agree.

The auxiliary base carried alongside the representation is AustinBaseSpace S n = PooledRankLatentSpace S n × ClusterSpace S n: the pooled lower-rank latents together with the mixed clusters. It is the equivariant base over which the enriched kernel will later be built.

Its two components differ in kind, and conflating them would be a mistake. The latent component carries no rank-n latent, every pooled index having cardinality < n. The cluster component, by contrast, deliberately does carry rank-n blocks — mixed ones, at supports that are not wholly original. That is the point of the poll: the clusters are correlated structural data, which is exactly why conditioning on them is informative.

The action is fixed and proved strict before the bundle is built, because equivariance cannot be repaired downstream. No conditional distribution or kernel identity appears here; those belong to the kernel unit, and every such identity there is almost-everywhere under a named law.

Contents #

The auxiliary base #

@[reducible, inline]

The Austin base: pooled lower-rank latents together with the mixed clusters. The latent component carries no rank-n latent — every pooled index has cardinality < n — while the cluster component deliberately carries rank-n blocks at supports that are not wholly original. The base is therefore lower-rank in its latent part only; its cluster part is structural polling data.

Equations
Instances For

    The split lift #

    The split lift of a finitely supported relabeling: act on the original half, fix the spare half. Restricting to split permutations is what makes the original-latent component natural; a boundary-crossing permutation has no such law.

    Equations
    Instances For
      @[simp]

      The split lift preserves the spare half, which is what keeps mixed clusters mixed.

      The induced action on cluster indices #

      The split lift permutes the mixed cluster indices: it preserves cardinality, and preserves the presence of a spare vertex, so a mixed index stays mixed — including the all-spare ones.

      Equations
      • One or more equations did not get rendered due to their size.
      Instances For

        The action on the base #

        The action on the Austin base: the pooled latents move by the split lift, and each cluster coordinate is re-read at the moved index and transported by blockSpaceCongr. A pinned definition rather than an opaque field, so its laws are checkable.

        Equations
        • One or more equations did not get rendered due to their size.
        Instances For

          The laws of the action #

          Note the orientation: pooledRankLatentRelabel_comp is contravariant (Equiv.trans f g applies f first), so the action composes as austinBaseRelabel (σ * τ) = austinBaseRelabel τ ∘ austinBaseRelabel σ.

          Pointwise cluster naturality: relabeling the pooled structure and reading a cluster coordinate is reading the moved cluster at the moved coordinate. Stated at the block coordinate so that no fibre transport appears in the statement, and generically in the cluster index, so genuinely mixed and all-spare indices are covered alike.

          Pointwise evaluation of the base action on a cluster coordinate: a Bool-valued identity, so no dependent fibre appears. The bridge through which the function-level laws are proved.

          Composition law, pointwise, in the contravariant orientation forced by pooledRankLatentRelabel_comp.

          The function-level laws #

          The global naturality square, and the invariance it yields #

          The action on the pooled objects by the split lift.

          Equations
          • One or more equations did not get rendered due to their size.
          Instances For

            The action on the enriched objects: the representation coordinates move by σ, the auxiliary base by austinBaseRelabel.

            Equations
            • One or more equations did not get rendered due to their size.
            Instances For

              One square for all four components. Enriching after acting on the pooled objects is acting on the enriched objects after enriching. The structure and pooled-latent components are definitional, the clusters are pollingClusters_relabel, and the original latents are the split corollary restrictOriginalLatents_sumCongr. An exact function equality.

              Exact invariance of the enriched law. Nothing but Measure.map_map, the naturality square, and the extension's own invariance — no almost-everywhere reasoning and no component rewriting.

              The bundle #

              Compressing the enriched law drops the original rank-n latents, which are redundant: on the image of enrichedPollingMap they are restrictOriginalLatents of the pooled array, so nothing is lost and map_original recovers C.P from the pooled component alone.

              Forget the redundant original-latent coordinate.

              Equations
              Instances For

                The base-extended law: the representation's structure together with the Austin base, over the existing signature. No synthetic relational signature is introduced, and the pooled real latents remain real — they are never encoded as Boolean relation coordinates.

                Instances For

                  The bundle exists. The law is the exact pushforward of the enriched polling law; the four laws are transports of results already established, and the mutual conditional independence is the unit-1 witness moved forward by the route-local transport rather than reproved.

                  Equations
                  • One or more equations did not get rendered due to their size.
                  Instances For

                    Adapter spaces #

                    Introduced here only so that their standard-Borel structure is available and inferred rather than assumed; their maps and commuting laws belong to the bundle that follows. Stated for an arbitrary coherent basis — selecting one via nonempty_coherentBasis is what would introduce [Fintype S.Srt], and that is deliberately not done here.

                    @[reducible, inline]

                    The boundary space at A extended by the Austin base.

                    Equations
                    Instances For

                      The adapter into the coherent-basis factor API #

                      For an arbitrary coherent basis: selecting one via nonempty_coherentBasis is what would introduce [Fintype S.Srt], and that is deliberately not done here. The four commuting laws are exact, and the last is a prerequisite for eventual exact truncation rather than that statement itself.

                      The boundary at A, extended by the base.

                      Equations
                      Instances For

                        The rank-m layer. The base is not attached: the layer is where the fresh rank-m information lives, and the base is the lower-rank data it is read against.

                        Equations
                        Instances For

                          The exact-anchor layer at A.

                          Equations
                          Instances For

                            First commuting law: projecting the extended lower factor to the boundary, leaving the base untouched, is the extended boundary map.

                            Second commuting law: the exact-anchor layer factors through the rank-m layer.

                            Third commuting law: forgetting the auxiliary base from the joint lower/layer observation returns the corresponding pushforward of M.law.

                            This asserts nothing about independence, and the base is not free of structural information — the clusters are correlated structural polling data. What the identity says is only that the structural factor law is recovered once the base is discarded, which is what the adapter needs.

                            Factor-side recombination: recombining the lower and layer coordinates through lowerFactorSpaceSuccEquiv.symm returns the rank-(n+1) lower-factor law, on the nose — an equality of pushed measures with the auxiliary base forgotten, with no almost-everywhere weakening.

                            It is a prerequisite for exact truncation, not that statement itself: it mentions neither C.P nor rankLatentProjection, so it does not pin RankSuccessor.truncation. Exact truncation will come from map_original together with the source-splitting identity of the assembly unit.