Prime denominators for compatible word-stalk tuples
Fix a finite-generator Coxeter system with symmetric coefficient algebra \(A\), a word \(\mathbf t\), and any weight set \(U\subseteq W\). Let \(z=(z_x)_{x\in U}\), with \(z_x\in M_{\mathbf t}^x\).
The nonzero evaluation denominator \(\Delta_{\mathbf t}\) satisfies \[ \Delta_{\mathbf t}z\in L_{\mathbf t}(U) \] for every such tuple.
If \(z\) is compatible on all upward reflection edges within \(U\), then for every prime \(p\in A\) there is a scalar \(d\) with \[ p\nmid d,\qquad dz\in L_{\mathbf t}(U). \]
Relation to the OpenAI notes
Explicit word-image construction of the extension property.
OpenAI notes: The moment-graph sheaf theorem gives extension of compatible sections on upper sets for the normalized boundary sheaf through the cited Braden–MacPherson and Soergel bridges. bg:bmp.
Lean (primary): The word module gives coordinate images and stable coefficient characters. Reflection geometry makes each prime congruence class a single weight or a reflection pair. Explicit diagonal operators combine class lifts, and prime-denominator removal proves compatible-tuple lifting from image projectivity. The word induction supplies upper-image projectivity using its two kernel sequences. IntegralPairings:470, IntegralPairings:467, IntegralPairings:1299, IntegralPairings:1408, IntegralPairings:2057, IntegralPairings:1620, IntegralPairings:1833, BalancedPairings:292, Support:293.
Role in the proof: Upper-image lifting gives interval flabbiness after extension to the ambient upper closure. Full-interval section realization also has a direct free-image proof. Free stalks, root quotients, flabbiness, generation and the normalized word top then supply the graded boundary summand.
Proof
Step 1. Construct a generic scalar lift
For each subexpression \(\varepsilon\), define a scalar \[ c_\varepsilon= \begin{cases} (z_{w(\varepsilon)})_\varepsilon,&w(\varepsilon)\in U,\\ 0,&w(\varepsilon)\notin U. \end{cases} \] The scaled-coordinate theorem supplies \(m\in M_{\mathbf t}\) with \(E_\varepsilon(m)=\Delta_{\mathbf t}c_\varepsilon\) for every \(\varepsilon\).
Fix \(x\in U\). For every subexpression of weight \(x\), this identity gives the corresponding coordinate of \(\pi_x(m)=\Delta_{\mathbf t}z_x\). Thus the stalk vectors agree, proving \(p_U(m)=\Delta_{\mathbf t}z\). The same theorem gives \(\Delta_{\mathbf t}\ne0\).
Step 2. Lift a reflection pair
Suppose \(x,y\in U\) and \(y=tx\) for a reflection \(t\). Multiplication by a reflection changes the parity of Coxeter length. Indeed, assigning \(-1\) to every simple generator respects the Coxeter relations, which have even length, and defines a sign character. A conjugate of a simple generator has sign \(-1\), while an element of length \(n\) has sign \((-1)^n\). Thus \(\ell(tx)\) and \(\ell(x)\) have opposite parity and are unequal. If \(\ell(x)<\ell(y)\), compatibility gives a word vector with the prescribed pair \((z_x,z_y)\).
If the length inequality is reversed, then \(x=ty\), since \(t^2=1\). Apply compatibility to the upward edge \(y\to x\), and read the two values in the opposite order. This again gives a common lift of \(z_x,z_y\).
Step 3. Lift an entire prime-character class
Fix \(x\in U\). If its class within \(U\) consists only of \(x\), stalk surjectivity supplies a vector attaining \(z_x\), which realizes that class.
Otherwise choose \(y\ne x\) in its class. The prime-character theorem gives a reflection \(t\) with \(y=tx\), and says that every member of this class is \(x\) or \(y\). Step 2 therefore gives one word vector realizing all prescribed values in the class.
Step 4. Pass to the finite stalk presentation
Use \(J_U\), the finite set of subexpressions with weight in \(U\). Set \[ \widehat z_\varepsilon=z_{w(\varepsilon)},\qquad \widehat\chi_a(\varepsilon)=\chi_{w(\varepsilon)}(a). \] The image of \(\widehat p_U\) is stable under these weights.
For a fixed index \(\varepsilon\), Step 3 gives a word vector matching \(z\) at every group weight congruent to \(w(\varepsilon)\) modulo \(p\). Its finite projection therefore matches \(\widehat z\) at every index in the residue class of \(\varepsilon\), including all repeated copies of those weights.
Step 5. Assemble the class lifts and recover the original tuple
The class-lift assembly theorem applies to the finite image \(\operatorname{im}\widehat p_U\). It produces \(d\) with \(p\nmid d\) and a word vector \(m\) such that \[ \widehat p_U(m)=d\widehat z. \] The finite-presentation recovery argument gives \(\pi_x(m)=dz_x\) for every \(x\in U\). Hence \(dz\in L_{\mathbf t}(U)\), proving the second assertion.
References
The local source declarations below give the construction and proof.
Lean formalization
OAI.KLInvariance.TitsSpace.bottSet_scaled: IntegralPairings.lean:1950.
OAI.KLInvariance.TitsSpace.BottCompatible.reflection_lift: IntegralPairings.lean:1968.
OAI.KLInvariance.TitsSpace.BottCompatible.prime_lift: IntegralPairings.lean:1982.
OAI.KLInvariance.TitsSpace.BottCompatible.local_denominator: IntegralPairings.lean:2029.