The top stalk of a reduced word is the graded coefficient module
Let \(\mathbf s=(s_1,\ldots,s_m)\) be a reduced word in a finite-generator Coxeter system, with product \(b\). Let \(A\) be the symmetric coefficient algebra, with its degree-one grading.
The only subexpression of \(\mathbf s\) with weight \(b\) is the full subexpression retaining every letter. Evaluation at that coordinate gives a graded \(A\)-linear isomorphism \[ M_{\mathbf s}^{\,b}\xrightarrow{\sim}A. \] Thus the top weight stalk is free of rank one with generator in degree zero.
Proof
Step 1. Force a top-weight subexpression to retain every letter
Let \(\varepsilon\) be a subexpression with weight \(b\), and let \(\mathbf t\) be its selected subword. If it retains \(r\) letters, then \(r\le m\).
The length of the product of any \(r\)-letter word is at most \(r\). Since the selected product is \(b\) and the original word is reduced, \[ m=\ell(b)=\ell(w(\mathbf t))\le r\le m. \] Hence \(r=m\). A subexpression retaining \(m\) of the \(m\) positions retains all of them, proving uniqueness of the full index.
Step 2. Realize every value of the surviving coordinate
Define a word element \(u_{\mathbf s}\) recursively by \[ u_{\varnothing}=1,\qquad u_{s\mathbf t}=1\otimes u_{\mathbf t}. \] Every evaluation sends this element to \(1\). For the empty word this is immediate. At each tensor step, the evaluation formula multiplies by the first scalar \(1\) and either keeps or reflects the remaining evaluation. The reflection fixes \(1\), so induction preserves that value.
For any \(a\in A\), \(A\)-linearity therefore gives \(E_\varepsilon(a u_{\mathbf s})=a\). In particular, the full coordinate takes every value in \(A\).
Step 3. Identify the top image stalk
The top stalk is an image inside the product indexed by top-weight subexpressions. Step 1 says that this index set has exactly one element. Two stalk vectors with the same full-coordinate value therefore agree in every coordinate, proving injectivity of evaluation on the stalk.
Step 2 gives surjectivity. The coordinate map is linear, so it defines the asserted module isomorphism.
Step 4. Check both grading directions
A stalk vector belongs to degree \(n\) exactly when all its coordinates belong to \(A_n\). There is only the full coordinate, so \[ v\in(M_{\mathbf s}^{\,b})_n \quad\Longleftrightarrow\quad E_{\mathrm{full}}(v)\in A_n. \] This proves that the isomorphism and its inverse preserve degree. The scalar \(1\in A_0\) thus corresponds to a degree-zero basis vector of the top stalk.
For the one-letter word \(s\), the full coordinate is the second coordinate of the congruence pair \((f,g)\). Its image stalk is \(A\), with \(g\) as the top coordinate.
References
The source declarations below give the constructions and proofs.
Lean formalization
OAI.KLInvariance.TitsSpace.bottFullIndex: BottSamelson.lean:237.
OAI.KLInvariance.TitsSpace.bottWeight_eq_top: BottSamelson.lean:269.
OAI.KLInvariance.TitsSpace.bottUnit: BottSamelson.lean:276.
OAI.KLInvariance.TitsSpace.bottEvaluation_unit: BottSamelson.lean:280.
OAI.KLInvariance.TitsSpace.bottTopCoordinate_bijective: BottSamelson.lean:300.
OAI.KLInvariance.TitsSpace.bottTopStalkEquiv: BottSamelson.lean:314.
OAI.KLInvariance.TitsSpace.bottTopStalkEquiv_homogeneous: BottSamelson.lean:502.
OAI.KLInvariance.TitsSpace.bottTopHom: BalancedPairings.lean:2066.
OAI.KLInvariance.TitsSpace.bottTopInvHom: BalancedPairings.lean:2075.