Weight-stalk images and supported word elements
Fix a finite-generator Coxeter system, its extended realization, and the symmetric coefficient algebra \(A\). Let \(\mathbf s\) be a word, let \(M_{\mathbf s}\) be its Bott–Samelson module, and let \(J\) be its finite set of subexpressions.
Every subexpression \(\varepsilon\) has a group weight \(w(\varepsilon)\) and an \(A\)-linear evaluation \(E_\varepsilon:M_{\mathbf s}\to A\). For a group element \(x\), put \[ J_x=\{\varepsilon\in J:w(\varepsilon)=x\}. \] Collecting these coordinates gives \[ E_x:M_{\mathbf s}\longrightarrow A^{J_x}, \qquad m\longmapsto(E_\varepsilon(m))_{\varepsilon\in J_x}. \]
The weight stalk is the image module \[ M_{\mathbf s}^x=\operatorname{im}E_x\subseteq A^{J_x}. \] Its stalk projection is the same evaluation with codomain restricted to the image: \[ \pi_x:M_{\mathbf s}\twoheadrightarrow M_{\mathbf s}^x. \] Surjectivity holds because membership in the image supplies a word element with exactly those coordinates.
Elements supported at one weight
The weight-supported submodule is \[ S_x=\{m\in M_{\mathbf s}:E_\varepsilon(m)=0 \text{ whenever }w(\varepsilon)\ne x\}. \] It is a submodule of the original word module. The stalk \(M_{\mathbf s}^x\), by contrast, consists of the selected coordinate tuples obtained by projecting arbitrary word elements.
The evaluation kernel is \[ \ker\pi_x=\{m:E_\varepsilon(m)=0 \text{ whenever }w(\varepsilon)=x\}. \] Thus support at \(x\) means vanishing outside the \(x\)-block, while the projection kernel means vanishing on the \(x\)-block. The perfect word pairing relates these two conditions through annihilators.
One reflection
For the one-letter word \(s\), the evaluations of a pure tensor are \[ E_0(a\otimes b)=ab,\qquad E_1(a\otimes b)=a\,s(b). \] Write \(\alpha\) for the root polynomial, with \(s(\alpha)=-\alpha\). In the left-module basis \(1\otimes1,1\otimes\alpha\), a vector with coefficients \(u,v\in A\) evaluates to \[ (u+\alpha v,\ u-\alpha v). \] Its two coordinates differ by \(2\alpha v\). Conversely, if \(f-g\in\alpha A\), then \[ u=(f+g)/2,\qquad v=(f-g)/(2\alpha) \] belong to \(A\) and reconstruct the pair. Hence full evaluation identifies this word module with \[ \{(f,g)\in A^2:f-g\in\alpha A\}. \]
Each individual weight stalk is \(A\), because diagonal pairs attain every coordinate. Supported elements are the pairs \((\alpha h,0)\) and \((0,\alpha h)\). This is the word-module source of the one-edge sheaf example.
Homogeneous stalk coordinates
Choose a basis of the extended realization and give \(A\) its polynomial degree grading. The ambient coordinate module has degree-\(n\) piece \[ (A^{J_x})_n=\{z:z_\eta\in A_n\text{ for every }\eta\in J_x\}. \] The image stalk has the restricted pieces \[ (M_{\mathbf s}^x)_n=M_{\mathbf s}^x\cap(A^{J_x})_n. \] These are the source's bottCoordinatePiece and bottStalkPiece. The homogeneous-evaluation theorem proves that the image is closed under homogeneous components, that these pieces form a grading, and that each homogeneous stalk vector has a word lift of the same degree.
Scaled coordinate data
In the standard word basis, full evaluation has a square matrix \(E\). Its determinant \[ \Delta=\det E \] is the evaluation denominator. A scaled coordinate lift \(c_\varepsilon\) is a word element with full evaluation \(\Delta\) in coordinate \(\varepsilon\) and zero in every other coordinate.
The theorem Scaled coordinate lifts and the evaluation denominator proves \(\Delta\ne0\), constructs these elements by the adjugate matrix, and derives a reconstruction formula for \(\Delta m\).
Subsequent constructions
The injectivity theorem identifies a word element by all its evaluations. Scaled coordinate lifts provide explicit supported elements after multiplication by a common polynomial. The support-duality theorem identifies \(S_x\) with the \(A\)-linear dual of the actual image stalk \(M_{\mathbf s}^x\). Pairing the projections at adjacent vertices then defines the edge quotients of the word sheaf.
References
The local source declarations below give the constructions and proofs.
Lean formalization
OAI.KLInvariance.TitsSpace.bottWeightProjection: Frobenius.lean:1497.
OAI.KLInvariance.TitsSpace.bottStalk: Frobenius.lean:1558.
OAI.KLInvariance.TitsSpace.bottStalkProjection: Frobenius.lean:1562.
OAI.KLInvariance.TitsSpace.bottStalkProjection_surjective: Frobenius.lean:1566.
OAI.KLInvariance.TitsSpace.bottWeightSupported: Frobenius.lean:1403.
OAI.KLInvariance.TitsSpace.bottCoordinateDenominator: Frobenius.lean:1317.
OAI.KLInvariance.TitsSpace.bottCoordinateLift: Frobenius.lean:1448.
OAI.KLInvariance.TitsSpace.bottCoordinatePiece: BottSamelson.lean:421.
OAI.KLInvariance.TitsSpace.bottStalkPiece: BottSamelson.lean:449.
Used by
- Polynomial characters of word stalks
- The pair-projection sheaf of a Bott–Samelson module
- Homogeneous evaluation and homogeneous lifts to word modules
- The right coefficient action at a subexpression weight
- The reflected two-stalk coordinates of a longer word
- Exact reflection kernels for a tensor word
- Scaled coordinate lifts and the evaluation denominator
- The exact stalk kernel across the first reflection
- Weight support and the dual of a projected stalk
- The top stalk of a reduced word is the graded coefficient module
- Reduced-word weights fill the Bruhat interval