Word objects and Bott–Samelson tensor induction
Let \(A\) be the symmetric coefficient algebra of the extended realization. For a simple reflection \(s\), write \(A^s\) for the subalgebra fixed by its coefficient action and \(\alpha_s\) for its degree-one root polynomial.
For a word \(\mathbf s\), the word module is defined recursively by \[ M_{\varnothing}=A,\qquad M_{s\mathbf t}=A\otimes_{A^s}M_{\mathbf t}. \] The tensor relation identifies \((ar)\otimes m\) with \(a\otimes(rm)\) for \(r\in A^s\).
The two coefficient actions
The empty-word module has the usual left and right multiplication by \(A\). At an induction step, the left action is \[ c(a\otimes m)=(ca)\otimes m, \] and the right action is \[ (a\otimes m)\cdot c=a\otimes(m\cdot c). \] The old right action commutes with the old left \(A\)-action. It therefore respects the tensor relation over \(A^s\), and it commutes with the new left action.
A recursively constructed finite basis
The reflection Frobenius extension has \(A^s\)-basis \(1,\alpha_s\). If \((b_\varepsilon)\) is an \(A\)-basis of \(M_{\mathbf t}\), then \((b_\varepsilon,\alpha_s b_\varepsilon)_\varepsilon\) is an \(A^s\)-basis. Extending scalars gives the \(A\)-basis \[ b_{(a,\varepsilon)}=1\otimes\alpha_s^a b_\varepsilon, \qquad a\in\{0,1\}. \] Starting with \(b_\varnothing=1\), a word of length \(m\) therefore has \(2^m\) basis vectors indexed by \(\{0,1\}^m\).
Their degrees are \[ d_\varepsilon=\varepsilon_1+\cdots+\varepsilon_m. \] The word grading is defined by the coefficient condition \[ \sum_\varepsilon c_\varepsilon b_\varepsilon\in(M_{\mathbf s})_n \quad\Longleftrightarrow\quad c_\varepsilon\in A_{n-d_\varepsilon}\text{ for every }\varepsilon. \] The coefficient decomposition gives a unique finite homogeneous decomposition of every vector.
Evaluation indices and group weights
The same finite index set labels the evaluation coordinates. Its group-weight function is \[ w(\varepsilon)=s_1^{\varepsilon_1}\cdots s_m^{\varepsilon_m}. \] These group labels are used to group evaluations into weight stalks.
For one letter \(s\), the basis vectors \(1\otimes1\) and \(1\otimes\alpha_s\) evaluate to \((1,1)\) and \((\alpha_s,-\alpha_s)\) in the coordinates labelled \(1,s\). Each basis vector therefore contributes to both weight coordinates. The basis degree and the coordinate's group label describe two parts of the construction.
The induced pairing
The empty-word pairing is multiplication in \(A\). At a prepending step, let \(\operatorname{tr}_s:A\to A^s\) be the reflection trace. The new pairing satisfies \[ \langle a\otimes m,b\otimes n\rangle =ab\,\operatorname{tr}_s(\langle m,n\rangle). \] The trace value is included in \(A\). The tensor-pairing theorem proves that this is a perfect pairing and preserves symmetry and balance of the right coefficient action.
The degree theorem The word pairing lowers degree by the word length proves \[ \langle(M_{\mathbf s})_i,(M_{\mathbf s})_j\rangle \subseteq A_{i+j-|\mathbf s|}. \] All these constructions apply to any word. For a reduced word, the later top-stalk theorem identifies its top coordinate with the degree-zero coefficient module.
References
The source declarations below give the constructions and proofs.
Lean formalization
OAI.KLInvariance.TitsSpace.bottSamelson: Frobenius.lean:394.
OAI.KLInvariance.TitsSpace.BottSamelsonModule: Frobenius.lean:399.
OAI.KLInvariance.TitsSpace.BottIndex: Frobenius.lean:947.
OAI.KLInvariance.TitsSpace.bottWeight: Frobenius.lean:1056.
OAI.KLInvariance.TitsSpace.bottBasis: Frobenius.lean:967.
OAI.KLInvariance.TitsSpace.bottDegree: Frobenius.lean:957.
OAI.KLInvariance.TitsSpace.bottPiece: Frobenius.lean:1014.
OAI.KLInvariance.TitsSpace.bottSamelsonRight: Frobenius.lean:405.
OAI.KLInvariance.TitsSpace.bottSamelsonDuality: Frobenius.lean:401.