Reversal of the full path product in an induced dihedral interval
Let \(W\) be a finite-generator Coxeter system, \(P\) an actual root plane, and \(B\) a plane-basis datum with induced two-generator system \(D\). Choose \(\mathrm{lo}\le_D\mathrm{hi}\), and let \(V=[\mathrm{lo},\mathrm{hi}]_D\), \(E\) be its directed reflection edges.
Let \(\iota:V\to V\) and \(g:E\to E\) be the interval and edge flips, and let \(\lambda(e)=\lambda_B(e)\) be the actual geometric root slope from the interval-flip construction. For composable edges \(e,h\), \[ \lambda(e)<\lambda(h) \Longleftrightarrow \lambda(g(h))<\lambda(g(e)). \tag{1} \]
Let \(p\) be a list containing every edge of \(E\) exactly once, weakly decreasing for \(\lambda\). For any semiring \(R\) and coefficient \(T\in R\),
\[
\bigl(\mathcal D_p(T)\delta_{\mathrm{hi}}\bigr)_{\mathrm{lo}}
=
\bigl(\mathcal D_{p^{\mathrm{rev}}}(T)\delta_{\mathrm{hi}}\bigr)_{\mathrm{lo}}.
\tag{2}
\]
This is PlaneBasis.dispatch_reverse_endpoints.
Relation to the OpenAI notes
A complete intrinsic interval supplies the root-plane identity.
OpenAI notes: The notes state the increasing and decreasing path formulas and develop the two-vertex internal ranks and angular comparisons in the dihedral section. bg:paths, dih:shape, dih:midpoint-order.
Lean (primary): The local identity fixes the two endpoints of a full intrinsic interval and switches its internal-rank vertices. The middle-descent test proves reversal of adjacent actual-root slope comparisons. A length-preserving involution then identifies the two endpoint entries for a full sorted edge list and its reverse. DyerPaths:1563, DyerPaths:1594, DyerPaths:1620, DyerPaths:1048, DyerPaths:1722, DyerPaths:1752, DyerPaths:1889.
Role in the proof: The Blueprint gives the interval and edge maps, verifies both directions of sorted path transport and states the precise endpoint-entry equality. The later ambient plane-block application uses this intrinsic identity through the coset graph.
Proof
Step 1. The internal middle vertex
Take composable edges \(e:x\to y\) and \(h:y\to z\) in the full interval. Let \(L=\ell_D(\mathrm{lo})\), \(H=\ell_D(\mathrm{hi})\). Since \(x,z\) belong to the interval and the two edges strictly increase length, \[ L\le\ell_D(x)<\ell_D(y)<\ell_D(z)\le H. \] Thus \(L<\ell_D(y)<H\), and the interval flip sends \(y\) to \(\theta y\). The interval-flip construction sends the two edges to a composable pair \(\iota x\to\theta y\to\iota z\).
Step 2. The change of simple descent
The internal-rank theorem applies to \(y\): its length is positive, and the element \(\mathrm{hi}\) has greater length. Hence \(y\) has exactly one of the two simple left descents.
For either simple generator \(s_i\), the identities
\[
s_i\theta(y)=\theta(s_{1-i}y),\qquad
\ell_D(\theta w)=\ell_D(w)
\]
give
\[
\ell_D(s_i\theta y)<\ell_D(\theta y)
\Longleftrightarrow
\ell_D(s_{1-i}y)<\ell_D(y).
\]
For \(i=0\), uniqueness of the descent therefore gives
\[
s_0\text{ descends at }\theta y
\Longleftrightarrow
s_0\text{ does not descend at }y.
\tag{3}
\]
These are the source's descent_swap and descent_swap_false_iff_not.
Step 3. The middle-vertex slope test
The middle-vertex slope test for the plane basis gives \[ \lambda(e)<\lambda(h) \Longleftrightarrow s_0\text{ descends at }y. \tag{4} \] Apply the same test to the flipped composable pair: \[ \lambda(g(e))<\lambda(g(h)) \Longleftrightarrow s_0\text{ descends at }\theta y. \tag{5} \] The two roots labelling a composable upward pair are distinct. Indeed, equal roots would give the same reflection \(\tau\), so the two successive reflections would return the last vertex to the first, contradicting the two strict length increases. Geometric slope is injective on positive roots, so the two slopes in each pair are also distinct.
By (3)–(5), the comparison in (5) holds exactly when the comparison in (4) fails. For two distinct real numbers, failure of one strict comparison gives the opposite strict comparison. Hence (4) is equivalent to \(\lambda(g(h))<\lambda(g(e))\), proving (1).
Step 4. The reverse implication in the other orientation
The edge flip preserves composability and is an involution. Apply (1) to \(g(e),g(h)\), and use \(g^2=\operatorname{id}\). This gives \[ \lambda(h)<\lambda(e) \Longleftrightarrow \lambda(g(e))<\lambda(g(h)). \tag{6} \] Equations (1) and (6) are the local comparison statements needed in the two path-map directions.
Step 5. The two sorted schedules
The schedule \(p\) is weakly decreasing for \(\lambda\). Its reverse is weakly decreasing for the label \(-\lambda\): if \(e\) precedes \(h\) in the reverse list, then \(h\) precedes \(e\) in \(p\), so \[ \lambda(e)\le\lambda(h) \quad\Longrightarrow\quad -\lambda(h)\le-\lambda(e). \] Both lists contain every edge exactly once.
Equal \(\lambda\)-labels cannot occur on composable edges by the distinctness argument in Step 3. Equality of \(-\lambda\)-labels is equivalent to equality of \(\lambda\)-labels, so the same property holds for \(-\lambda\).
Step 6. Forward transport to the reverse schedule
Use sorted path transport with source labels \(\lambda\), target labels \(-\lambda\), schedules \(p,p^{\mathrm{rev}}\), vertex map \(\iota\), and edge map \(g\). The endpoint-map equations hold by the interval-flip construction, and \(\iota\) fixes \(\mathrm{lo},\mathrm{hi}\). For a composable pair with \(\lambda(h)<\lambda(e)\), (6) gives \(\lambda(g(e))<\lambda(g(h))\), or equivalently \[ -\lambda(g(h))<-\lambda(g(e)). \] This is the required local comparison for the target labels. All sorting, fullness and tie hypotheses were verified in Step 5. Therefore \[ q\in\mathcal P_p(\mathrm{lo},\mathrm{hi}) \Longrightarrow g(q)\in\mathcal P_{p^{\mathrm{rev}}}(\mathrm{lo},\mathrm{hi}). \tag{7} \]
Step 7. Backward transport
Now exchange the two schedules and labels. A composable pair decreasing for \(-\lambda\) satisfies \(\lambda(e)<\lambda(h)\). Equation (1) gives \(\lambda(g(h))<\lambda(g(e))\), the decreasing comparison for the target label \(\lambda\). The same endpoint equations, fullness and tie properties apply. Sorted path transport gives \[ q\in\mathcal P_{p^{\mathrm{rev}}}(\mathrm{lo},\mathrm{hi}) \Longrightarrow g(q)\in\mathcal P_p(\mathrm{lo},\mathrm{hi}). \tag{8} \]
Step 8. Equality of the endpoint entries
The map \(g\) is an involution, so its edgewise action has the same property and preserves path length. The two schedules have no repeated edge, and (7)–(8) give the two path-membership directions. The path-involution theorem now applies and proves (2).
If \(\mathrm{lo}=\mathrm{hi}\), there are no upward interval edges and both endpoint entries are the empty-path contribution one. For positive-rank intervals, the same argument pairs the selected paths with paths of the same number of edges in the opposite schedule. This is the intrinsic root-plane block identity used in the later geometric order-independence proof.
References
The local source declarations below give the construction and proof.
Lean formalization
OAI.KLInvariance.Dihedral.descent_swap: DyerPaths.lean:1011.
OAI.KLInvariance.Dihedral.unique_leftDescent: DyerPaths.lean:1019.
OAI.KLInvariance.Dihedral.descent_swap_false_iff_not: DyerPaths.lean:1048.
OAI.KLInvariance.TitsSpace.PlaneBasis.edgeRoot_ne_of_path: DyerPaths.lean:1715.
OAI.KLInvariance.TitsSpace.PlaneBasis.edgeRoot_slopes_rankFlip: DyerPaths.lean:1722.
OAI.KLInvariance.TitsSpace.PlaneBasis.intervalEdgeSlope_flip: DyerPaths.lean:1757.
OAI.KLInvariance.TitsSpace.PlaneBasis.intervalEdgeSlope_ties: DyerPaths.lean:1865.
OAI.KLInvariance.TitsSpace.PlaneBasis.intervalEdgeSlope_flip_reverse: DyerPaths.lean:1875.
OAI.KLInvariance.TitsSpace.PlaneBasis.dispatch_reverse_endpoints: DyerPaths.lean:1889.