Inversion cuts in a root plane
Let \(W\) be a finite-generator real Coxeter system, \(P\) a root plane, and \(B=(a_L,a_R)\) a plane basis for the injective geometric slope \(\sigma\). For \(w\in W\) and a positive root \(a\), write \[ \operatorname{Inv}_W(w,a)\quad\Longleftrightarrow\quad w^{-1}a\text{ is negative}. \] Choose positive roots \(a,b\in P\) such that \(\operatorname{Inv}_W(w,a)\) holds and \(\operatorname{Inv}_W(w,b)\) does not. Then: \[ \begin{aligned} \operatorname{Inv}_W(w,a_L)&\ \Longrightarrow\ \sigma(a)<\sigma(b),\\ \operatorname{Inv}_W(w,a_R)&\ \Longrightarrow\ \sigma(b)<\sigma(a),\\ \sigma(a)<\sigma(b)&\ \Longleftrightarrow\ \operatorname{Inv}_W(w,a_L). \end{aligned} \] Here \(w\) is any ambient group element. The assertions describe how the signs at \(w\) divide the slope-ordered roots of this plane.
Relation to the OpenAI notes
Ambient inversion cuts in place of a root-coordinate recurrence.
OpenAI notes: The notes set \(\kappa=-(a_L,a_R)/2\), where the parentheses denote the ambient Tits form. The midpoint-order lemma writes prefix roots using \(p_0=0\), \(p_1=1\) and \(p_{n+1}=2\kappa p_n-p_{n-1}\). It obtains \(\kappa=\cos(\pi/m)\) in finite type and \(\kappa\ge1\) in infinite type, then compares the resulting coordinate ratios. The functional ratio gives that angular order or its reverse. dih:midpoint-order.
Lean (primary): PlaneBasis chooses its endpoint roots in increasing geometric slope. Convexity of inversion signs gives a middle-descent test, which orders the two actual ambient prefix-root sequences. A separating element \(w_{k+1}\) compares the two rays. Unbounded lengths give the signed Before order; a maximum length gives the odd residue order modulo 2m. DyerPaths:923, DyerPaths:1106, DihedralIntervals:832, DihedralIntervals:844, DihedralIntervals:1101, DihedralIntervals:1143, DihedralIntervals:1251.
Role in the proof: The ordering conclusion is developed along Lean’s inversion and length proof. The additional classification of \(\kappa\) in the notes is not a conclusion of these Lean angular-order declarations.
Proof
Step 1. Exclude equal slopes for roots with different signs
If \(\sigma(a)=\sigma(b)\), injectivity gives \(a=b\). The inversion and noninversion assumptions would then contradict each other. Thus their slopes are distinct.
The same reasoning applies to any inverted root and any noninverted root used below.
Step 2. Use an inverted left endpoint
Assume \(a_L\) is inverted. Suppose the desired inequality fails. Step 1 gives \(\sigma(b)<\sigma(a)\). The endpoint bound \(\sigma(a_L)\le\sigma(b)\) is strict, since \(a_L\) is inverted and \(b\) is not. Consequently \[ \sigma(a_L)<\sigma(b)<\sigma(a). \]
In particular \(a_L\ne a\). Distinct positive roots in the root plane are linearly independent, so \(a_L,a\) span \(P\). The slope-cone theorem expresses the intermediate root as \[ b=r a_L+t a,\qquad r,t>0. \] Both \(w^{-1}a_L\) and \(w^{-1}a\) have nonpositive coordinates. Linearity gives \[ w^{-1}b=r\,w^{-1}a_L+t\,w^{-1}a, \] also with nonpositive coordinates. This says that \(b\) is inverted, contradicting its chosen sign. Hence \(\sigma(a)<\sigma(b)\).
Step 3. Use an inverted right endpoint
Assume \(a_R\) is inverted and suppose \(\sigma(b)<\sigma(a)\) fails. Distinctness of slopes gives \(\sigma(a)<\sigma(b)\). The upper endpoint inequality is strict at \(b\), because \(b\) and \(a_R\) have different inversion signs. Thus \[ \sigma(a)<\sigma(b)<\sigma(a_R). \] The roots \(a,a_R\) span \(P\). The slope-cone theorem gives \(b=r a+t a_R\) with \(r,t>0\). Applying \(w^{-1}\) again expresses a noninverted root as a nonnegative combination of two inverted roots, a contradiction. This proves \(\sigma(b)<\sigma(a)\).
Step 4. Recover the endpoint sign from a mixed comparison
Suppose \(\sigma(a)<\sigma(b)\) and \(a_L\) is noninverted. Every positive root in \(P\) belongs to the nonnegative cone of \(a_L,a_R\); in particular \[ a=r a_L+t a_R,\qquad r,t\ge0. \] If \(a_R\) were also noninverted, the noninverted part of the slope-cone theorem would make \(a\) noninverted. Therefore \(a_R\) is inverted.
Step 3 would give \(\sigma(b)<\sigma(a)\), contradicting the assumption. Hence \(a_L\) is inverted. The reverse implication is Step 2, proving the equivalence.
References
The local source declarations below give the construction and proof.
Lean formalization
OAI.KLInvariance.TitsSpace.PlaneBasis.slope_lt_of_inverted_left: DyerPaths.lean:881.
OAI.KLInvariance.TitsSpace.PlaneBasis.slope_lt_of_inverted_right: DyerPaths.lean:899.
OAI.KLInvariance.TitsSpace.PlaneBasis.slope_lt_iff_inverted_left: DyerPaths.lean:923.