Kl.Sheaves
Moment-graph sheaves, support, and duality
This topic constructs modules on a labelled Bruhat interval, studies their compatible sections, and uses support and duality to relate stalks to outgoing kernels. Its inputs are the ordered interval, polynomial edge labels from the extended realization, and finite graded modules. Its outputs are boundary sheaves, stalk–costalk duality, and the actual-character reciprocity used in the invariance proof.
The two-vertex interval \(1<s\) and \(A=\mathbb R[\alpha]\) provide a running example. The same stalks, restrictions, and sections will carry us from the first definitions to an explicit duality map.
A graph sheaf is an explicit collection of modules and linear maps on a directed graph. Its section equations compare values at adjacent vertices after applying the two edge restrictions.
Modules at vertices and edges
Write the vertices as \(x<y\), and let \(e:x\to y\) be the single edge. Define \[ B^x=A,\qquad B^y=A,\qquad B^e=A/(\alpha). \] The symbols \(B^x,B^y\) are the vertex modules, also called stalks. The symbol \(B^e\) is the edge module. There are two restriction maps \[ \rho_{x,e}:B^x\to B^e,\qquad \rho_{y,e}:B^y\to B^e. \] In this example both are the quotient map \(f\mapsto f\bmod\alpha\).
The direction \(x\to y\) specifies which endpoint is lower and which is upper. Both restrictions go into the common edge module, where the two endpoint values can be compared.
On a general graph the definition supplies this data for every vertex and edge. A graded sheaf also requires compatible gradings and degree-zero restriction maps.
The coefficient action on an actual edge
For a root label \(r_a\in A\), the upper-quotient rule makes \(r_a\) annihilate the edge module. The action therefore factors through the edge coefficient ring \(E_a=A/(r_a)\): \[ [f]\cdot v=fv. \] If two representatives differ by \(r_ag\), their actions differ by \(g(r_av)=0\).
A basis \(m_i\) of the upper stalk descends to the vectors \(\rho^+_e(m_i)\) in the edge module. They form an \(E_a\)-basis by the scalar-kernel basis theorem: both the upper restriction and coordinate reduction have kernel \(r_aB^y\). This is the actual edge-module theorem, which fixes the coefficient action used when comparing submodules inside an edge.
Section equations and a homogeneous basis
An assignment chooses one stalk vector at every vertex. For the example it is a pair \((f,g)\in A^2\). It is a compatible section when the two values agree after restriction: \[ f\bmod\alpha=g\bmod\alpha, \] or equivalently \(f-g\in\alpha A\).
Thus the global section module is \[ \Gamma(B)=\{(f,g)\in A^2:f-g\in\alpha A\}. \] Write \(f-g=\alpha h\). Then \[ (f,g)=g(1,1)+h(\alpha,0). \] The expression is unique: the second coordinate determines \(g\), and injectivity of multiplication by \(\alpha\) determines \(h\). Therefore \(\Gamma(B)\) is free with homogeneous basis degrees zero and one.
Each stalk is free of rank one. The global section module has rank two, with basis degrees zero and one. The degree-one generator records the freedom to change the lower value by a multiple of the edge label.
Global generation and flabbiness
Global generation asks whether every stalk vector can be realized as the value of a global section. In the example, a prescribed polynomial \(f\) at either vertex is realized by \((f,f)\), so both stalks are generated by global sections.
An upper set is a vertex subset that contains every vertex above any of its members. The nontrivial proper upper set in this example is \(\{y\}\). Flabbiness asks whether every compatible section on an upper set extends to a global section.
A polynomial \(g\) prescribed at \(y\) extends to \((g,g)\), so the example is flabby. On a graph with several vertices, prescribing compatible values on an entire upper set imposes more conditions than prescribing a value at one stalk. The two extension statements become separate inputs to the boundary-image theorem.
Relative flabbiness is the extension statement for sets closed under the edges that remain relevant after localization. It will be used in localization and edge-image comparison when some edge modules become zero.
Word sheaves and their root quotients
The tensor word supplies a concrete sheaf by taking its image stalk at each vertex and the pair-projection quotient at each edge. The word-length induction proves that its stalks are finite free and that an upper edge restriction has kernel exactly the root label times its upper stalk. This identifies the edge module with that root quotient.
A basis of the extended realization changes the symmetric coefficient algebra into a polynomial ring. The transported scalar action is \(p\cdot m=e^{-1}(p)m\), so an old root multiple \(\alpha_E n\) becomes the new multiple \(e(\alpha_E)n\). The same underlying basis vectors remain a basis over the polynomial ring. Thus the polynomial word sheaf has the free stalks and upper quotients required by the later boundary-summand construction.
Word lifts, interval extensions and global sections
For a reduced word with product \(b\), the word stalks vanish outside \([1,b]\). Given an upper subset \(U\) of this interval, take its upper closure in the ambient group. It meets \([1,b]\) exactly in \(U\). Extend the prescribed assignment by zero outside the interval. An edge whose upper endpoint lies below \(b\) uses the original compatibility condition; an edge whose upper endpoint lies outside the support uses stalk surjectivity at the lower endpoint and the zero upper stalk.
The completed word lifting theorem supplies one word vector for this ambient upper tuple. Its interval projections form the global extension, proving word-sheaf flabbiness.
On the full interval, the section map is an isomorphism from the word module to all global sections. Its proof uses the free image of the full vertex projection, the same prime-class denominators and the evaluation determinant. This gives the actual module identification used to transport word pairings and adjoints.
Finally, both a normalized boundary sheaf and the reduced word sheaf have top stalk \(A\) in degree zero. Their top identifications give inverse maps. The word sheaf has free stalks, root quotients, flabbiness and global generation, so the top-splitting theorem extends those maps to a graded boundary summand.
The boundary image as the target of the new stalk
Suppose only the portion of a sheaf strictly above \(x\) has been constructed. Restrict a section there to each edge leaving \(x\). The resulting tuple lies in the product of the outgoing edge modules. The set of all such tuples is the boundary image at \(x\).
For the one-edge example, the upper region has the stalk \(A\) at \(y\), and the boundary map is \[ A\longrightarrow A/(\alpha). \] Its image is all of \(A/(\alpha)\). To attach a stalk at \(x\), we need a free module surjecting onto this image. The minimal graded free cover is the quotient map \(A\to A/(\alpha)\), giving exactly the stalk and restriction we started with.
For several outgoing edges, the boundary image need not be the full product: values arriving from the upper region must already satisfy its internal section equations. Taking a minimal cover of this image retains those constraints. The theorem The stalk surjects onto the boundary image identifies it with the image of the lower stalk map: global generation gives one containment, and extension from the strict upper set gives the other.
Start at the top vertex \(b\), where the stalk is normalized to \(A\), and repeat the construction downward. Finiteness of the vertex set guarantees termination. The boundary-existence theorem verifies all required sheaf properties. The uniqueness theorem compares two such constructions by identifying their upper boundary images and extending the resulting isomorphism through the minimal covers.
Kernels remember which outgoing conditions have been imposed
For a set \(F\) of outgoing edges at \(x\), define \[ K_x(F)=\{m\in B^x:\rho_{x,e}(m)=0\text{ for every }e\in F\}. \] No conditions gives \(K_x(\varnothing)=B^x\). Imposing all outgoing conditions gives the full outgoing kernel, called the costalk in the later comparison.
In the one-edge example, \[ K_x(\varnothing)=A,\qquad K_x(\{e\})=\alpha A,\qquad K_y(\varnothing)=A. \] The last equality holds because the top vertex has no outgoing edges. Removing the condition \(e\) enlarges the lower kernel from \(\alpha A\) to \(A\).
On a larger graph there are intermediate kernels corresponding to different subsets of conditions. The simultaneous induction establishes freeness for the suffix kernels used in the comparison. At each step, an exact sequence measures a change of kernel by its image in one edge module.
Commuting restriction maps in a sheaf isomorphism
A sheaf isomorphism consists of an isomorphism at every vertex and edge, commuting with both restrictions. For example, if \(j_x\) and \(j_e\) are the lower stalk and edge maps, they must satisfy \[ j_e\rho^B_{x,e}=\rho^C_{x,e}j_x. \] This equation ensures that a kernel or image in \(B\) goes to the corresponding kernel or image in \(C\).
In the free module \(A\), the submodules \(A\) and \(\alpha A\) both have rank one, and their inclusions have different images. A commuting sheaf isomorphism tracks those inclusions. Later comparisons use it to place the relevant edge images in a common module.
Boundary images under a partial isomorphism
An isomorphism already constructed strictly above \(x\) gives an isomorphism \(J\) on the product of outgoing edge modules. Transporting a compatible upper assignment through its vertex maps commutes with the boundary map, coordinate by coordinate. Applying the inverse maps gives equality of the two boundary images.
For generated, flabby sheaves, the boundary-image identity replaces these with the outgoing stalk images. The boundary-image transport theorem therefore gives \[ J(\operatorname{im}d_x^B)=\operatorname{im}d_x^C. \] This identifies the actual target module used in both graded stalk comparison and the local-ring stalk splitting. The ungraded partial-isomorphism extension then defines each incoming edge map from the new upper-stalk quotient; its graded variant also checks preservation of degrees.
Extending a morphism through one stalk
A partial morphism on an upper set \(U\) includes an edge map whenever the edge's target lies in \(U\). For a maximal vertex \(x\) outside \(U\), all outgoing targets of \(x\) are already present. Their edge maps give a degree-zero map between the boundary products \[ j_{\partial x}:\prod_{e:x\to y}B^e\longrightarrow\prod_{e:x\to y}C^e. \]
Applying the known vertex maps to a section strictly above \(x\) preserves compatibility, by the two commuting squares. Its boundary values are transformed by \(j_{\partial x}\). The boundary-image theorem then gives \[ j_{\partial x}(\operatorname{im}r_B)\subseteq\operatorname{im}r_C, \] using generation and flabbiness of both sheaves.
Set \(T=\operatorname{im}r_C\). The required map \(B^x\to T\) lifts through the surjection \(C^x\to T\): choose a homogeneous free basis of \(B^x\), lift each image, and retain its homogeneous component of the same degree. The resulting map \(l_x:B^x\to C^x\) satisfies \(r_C l_x=j_{\partial x}r_B\), which verifies every outgoing lower square.
For a new incoming edge \(e:z\to x\), the upper-quotient rule descends \(l_x\) to an edge map. Indeed \(l_x(\alpha_e v)=\alpha_e l_x(v)\), so the old upper kernel maps into the new upper kernel. Homogeneous preimages show that this quotient map preserves degree. These constructions extend the partial morphism to \(U\cup\{x\}\). Removing \(x\) from the finite missing set supplies the strictly smaller induction instance.
The boundary-morphism lifting theorem, one-vertex morphism extension theorem and full morphism extension theorem develop these three steps separately.
Tensor extension and compatible sections
For a commutative scalar extension \(R\to S\), define \(B_S^x=S\otimes_R B^x\) and \(B_S^e=S\otimes_R B^e\), with restrictions \(1\otimes\rho^\pm_e\). A compatible assignment \(f\) gives \((s\otimes f_x)_x\), producing \[ \Theta_\Omega:S\otimes_R\Gamma_\Omega(B)\longrightarrow\Gamma_\Omega(B_S). \]
The section constraint map collects the differences of the two restrictions on edges internal to \(\Omega\). Its kernel is exactly \(\Gamma_\Omega(B)\). The finite assignment and edge-product maps commute with tensoring, and on a pure tensor the constraint identity is \[ s\otimes\rho^-_e(f_{s(e)})-s\otimes\rho^+_e(f_{t(e)}) =s\otimes\bigl(\rho^-_e(f_{s(e)})-\rho^+_e(f_{t(e)})\bigr). \] Flatness preserves the kernel sequence. Thus the flat section base-change theorem makes \(\Theta_\Omega\) an isomorphism for a finite graph.
A tensor stalk value is a finite sum of pure tensors, so global sections generating its second factors generate the value after extension. Under the finite-flat hypotheses, a section on an upper set likewise comes from a finite sum of tensor sections; extending each original section proves flabbiness of \(B_S\). The scalar-extension property theorem also transports free stalks, finite generation and the upper-quotient rule.
The selected outgoing constraint map applies the same construction at one stalk. It identifies \(K_x(F)\) with a kernel, giving the flat kernel-and-image theorem used in localization. That theorem preserves the embeddings of both edge images into the same extended edge module.
Pullback through the surviving edges
Let \(f:G\to H\) be a graph map. Pullback copies the modules at \(f_V(x)\) and \(f_E(a)\), with restrictions identified through the two endpoint equations. A global section pulls back coordinatewise, so generation passes to the pullback. Upper quotients pass through the same endpoint identifications.
For flabbiness, take a set \(L\) of target edges whose complement has zero edge modules. Assume \(f_V\) is injective and every edge in \(L\) starting at an image vertex \(f_V(x)\) has an outgoing lift starting at \(x\). If \(\Omega\) is upper in \(G\), then \(f_V(\Omega)\) is closed under \(L\): lift a surviving edge, use the upper-set property at its source, and map its target back to \(H\).
A section on \(\Omega\) pushes to an assignment on \(f_V(\Omega)\). On a surviving edge, its compatibility equation is the equation on the lifted edge; on any other edge it takes values in the zero module. When \(H\) has finitely many vertices and its sheaf is flabby, the relative extension theorem extends this assignment globally. Pullback then gives the desired extension on \(G\).
The selected-kernel pullback theorem uses the same outgoing lifts at one fixed vertex. Every selected ambient equation either comes from a selected source edge or vanishes in a zero module. Consequently it gives equality of the two kernels inside the identical stalk. This kernel statement applies using just those local lift and zero-module hypotheses.
The localized coset construction supplies this setup with \(L\) equal to the ambient edges whose reflection labels lie in the plane subgroup.
Minimal covers detect a boundary-sheaf isomorphism
Consider a graded morphism between two normalized boundary sheaves, bijective at the distinguished stalk. Suppose its stalk maps are already bijective strictly above \(x\). Upper quotient maps then make the relevant edge maps bijective.
These component maps identify compatible upper tuples and their boundary images. Global generation and flabbiness identify each boundary image with the image of its outgoing stalk map. At \(x\), the two stalk maps are therefore minimal free covers of a common graded module.
The specified morphism between those covers is bijective by the minimal-cover comparison theorem. Finite downward induction gives bijectivity at every stalk, and upper quotients give it at every edge. This is the top-detection theorem.
It is also the correction step in top splitting. Extend an inclusion and retraction from the top to sheaf morphisms \(i,p\). Their composite \(f=pi\) is invertible by top detection. Replacing \(i\) by \(if^{-1}\) gives a split pair on the whole sheaf.
Boundary models and extension from a lower ideal
For each vertex \(c\), construct a normalized boundary sheaf on the lower ideal \(\{x:x\le c\}\), then extend it by zero. The result \(O_c\) has top stalk \(A\), vanishes at vertices outside that ideal, and has zero edge module whenever the edge ends outside it. Inside the ideal, its restrictions are the original ones. At \(x\ne c\), the full outgoing kernel is contained in \(JO_c^x\).
For the single edge \(x\to y\), the model \(O_y\) is the congruence sheaf already constructed. The other model \(O_x\) has stalk \(A\) at \(x\), zero at \(y\), and zero edge module. Thus the two possible centers give different local data.
A graded model sum starts with the zero sheaf and repeatedly adds finite blocks of shifts \(O_c(d)\), with \(d\ge0\). A basis generator of degree \(r\) in \(O_c\) has degree \(r+d\) in the shifted sheaf. A finite sum uses componentwise restrictions, so its compatible sections and outgoing kernels split componentwise.
Decomposition through a stalk cover and its remainder
Suppose a free-stalk, globally generated, flabby sheaf \(B\), with the prescribed upper quotients, has already been identified with a model sum \(C\) on an upper set \(U\). Take a maximal vertex \(x\) outside \(U\).
The partial isomorphism includes every edge map whose target lies in \(U\). In particular it identifies all outgoing edge modules at \(x\). Generation and flabbiness identify the two boundary images. Both stalk maps can therefore be written as surjections to one graded module: \[ f:B^x\twoheadrightarrow T,\qquad \pi:C^x\twoheadrightarrow T. \] The existing centers lie in \(U\), so \(C\) is kernel-minimal at \(x\). Hence \(\pi\) is a minimal cover.
The free-cover remainder theorem constructs degree-zero maps \[ g:B^x\to C^x,\qquad h:C^x\to B^x,\qquad \pi g=f,\quad gh=1. \] It gives the explicit graded decomposition \[ B^x\simeq C^x\oplus K,\qquad v\longmapsto(gv,v-hgv),\qquad K=\ker g. \] The remainder \(K\) is a finite graded direct summand of a free module, so it has a homogeneous basis. If its basis degrees are \(d_1,\ldots,d_m\), add the block \(\bigoplus_jO_x(d_j)\).
This block has zero stalks at every old vertex in \(U\) and zero modules on every edge ending there. The old partial isomorphism thus extends by zero on the new components. At \(x\), the displayed decomposition supplies the stalk map; the equation \(\pi g=f\) verifies every outgoing square. Incoming edge maps are induced by the new stalk isomorphism through the upper quotient rule.
The missing-vertex set has decreased by one. Finite induction constructs an isomorphism of whole sheaves \[ B\simeq\bigoplus_{c,j}O_c(d_{c,j}). \] The local cover argument, the one-vertex extension, and the finite induction are developed separately in the linked proofs below.
One decomposition and two characters
Over a polynomial coefficient algebra, a finite free graded module with basis degrees \(d_i\) has character \(\sum_iq^{d_i}\). Its Hilbert series is that character divided by \((1-q)^N\), where \(N\) is the number of polynomial variables.
Record the model shifts at center \(c\) by \[ \mu_c(q)=\sum_jq^{d_{c,j}}. \] These are the multiplicity polynomials. The sheaf decomposition induces both the stalk decomposition and the outgoing-kernel decomposition. When the model kernels are free, both character identities therefore use the same \(\mu_c\).
For the example \(B=O_y\oplus O_x(2)\), the restrictions at \(x\) ignore the second summand: \[ B^x=A\oplus A(2)\longrightarrow A/(\alpha),\qquad (f,g)\longmapsto f\bmod\alpha. \]
| Vertex | Stalk | Full outgoing kernel | Stalk character | Kernel character |
|---|---|---|---|---|
| \(x\) | \(A\oplus A(2)\) | \(\alpha A\oplus A(2)\) | \(1+q^2\) | \(q+q^2\) |
| \(y\) | \(A\) | \(A\) | \(1\) | \(1\) |
Here \(\mu_y=1\) and \(\mu_x=q^2\). The lower model contributes \(q^2\) to both lower-vertex characters; the upper model contributes \(1\) to the stalk and \(q\) to the kernel.
Top splitting and the boundary pairing
A split inclusion consists of sheaf maps \(i:B\to W\) and \(p:W\to B\) with \(pi=1_B\). Their reverse composite \(ip\) is a projector on \(W\). The top-splitting theorem constructs such maps from a split pair at the normalized top; top detection corrects their composite on the whole boundary sheaf.
This splitting supplies the boundary summand of a reduced-word sheaf. Its projector and the word pairing produce the adjoint image and boundary pairing developed below. The full model decomposition supplies the simultaneous character formulas used later in cancellation. These constructions meet when boundary duality evaluates the kernels of all the lower models.
Upper restrictions with a fixed top
For a fixed top \(b\), let \(B(b)\) be the actual boundary sheaf on \([1,b]\). Its restriction to an upper interval is the family \[ \mathcal I_{u,b}=B(b)|_{[u,b]}. \] Every edge beginning in \([u,b]\) also ends there, so the inclusion retains all outgoing restrictions at each retained stalk. Consequently the restriction is a normalized boundary sheaf, and every selected outgoing kernel is the same submodule of the same stalk after identifying its retained edges.
This construction keeps \(B(b)\) and its top fixed while varying \(u\). The lower model \(O_c\) instead starts with \(B(c)\) on \([1,c]\) and extends by zero. The former is used to transfer an induction hypothesis to a smaller upper interval; the latter supplies the summands and multiplicity polynomials of decomposition.
From coordinate images to a sheaf
The pair-projection sheaf uses these stalk images. On an edge it remembers whether the two endpoint values come from one common word element. The simultaneous free-stalk/edge-kernel theorem proves that the stalks are free and that the upper restriction kernel is the edge label times the upper stalk.
A global word element gives compatible values on every edge. To recover a common word element from compatible coordinates, the flabbiness proof establishes a lifting theorem on upper sets through projective-range and denominator arguments.
The word module supplies the coordinate values, the word sheaf records their pairwise congruences, and the stalk and flabbiness theorems supply the hypotheses for decomposition and duality.
Supported sections and incoming label factors
A section supported at \(x\) has every other coordinate zero. Its value at \(x\) must therefore vanish under both the incoming upper restrictions and the outgoing lower restrictions. The full outgoing kernel already supplies the second group of equations.
For each incoming edge, the upper-quotient condition identifies its kernel with the multiples of that edge label. Let \[ \beta_x=\prod_{e:\,t(e)=x}\alpha_e. \] The theorem Incoming labels identify the outgoing kernel with vertex support proves that multiplication by \(\beta_x\), followed by zero extension, gives \[ K_x\xrightarrow{\sim}\Gamma(B)_{\{x\}}, \qquad k\longmapsto\delta_x(\beta_x k). \] The proof uses the prime linear labels twice. Distinct incoming primes force a supported stalk value to be divisible by their product. That product remains nonzero modulo every outgoing label, so it can be cancelled in each outgoing edge module.
In the running example, \(\beta_x=1\) and \(\beta_y=\alpha\). Thus the lower kernel \(\alpha A\) gives the sections \((\alpha h,0)\), while the upper kernel \(A\) gives \((0,\alpha h)\) after multiplication by \(\alpha\).
Integral cuts and the stalk dual
Let \(p\) be the product of all edge labels. It annihilates every edge module. Consequently \[ C_xm=p\,\delta_x(m_x) \] is a global section for every global section \(m\). This is the integral vertex cut.
A perfect section pairing identifies \(\Gamma(B)\) with its \(A\)-linear dual. Its two support-orthogonality equations imply that the cuts are balanced: \[ \langle C_xm,n\rangle=\langle m,C_xn\rangle. \] The proof compares both expressions, multiplied by \(p\), with \(\langle C_xm,C_xn\rangle\), and cancels \(p\).
This balance equation identifies supported sections with the functionals that vanish on the evaluation kernel \(\ker\pi_x\). Global generation makes \(\pi_x:\Gamma(B)\to B^x\) surjective. A supported section therefore determines a stalk functional by \[ \Phi_x(m)(v)=\langle m,n\rangle\quad\text{for any section }n\text{ with }n_x=v. \] Two choices of \(n\) differ by an element of \(\ker\pi_x\), on which the pairing is zero. Conversely, a stalk functional \(\lambda\) is represented by \(D^{-1}(\lambda\circ\pi_x)\), which the support-annihilator theorem proves is supported at \(x\).
The two isomorphisms compose: \[ K_x\xrightarrow{\ k\mapsto\delta_x(\beta_x k)\ } \Gamma(B)_{\{x\}} \xrightarrow{\ \Phi_x\ }(B^x)^*. \] This is the costalk-duality map. It transfers a finite dual basis to the full outgoing kernel, proving that kernel free. The homogeneity equation of the pairing then determines the basis degrees.
Degree reversal under support duality
For a homogeneous basis vector \(m\) of degree \(d\), its coordinate functional has degree \(-d\): on a degree-\(j\) vector, it extracts a coefficient of degree \(j-d\).
The homogeneous-representative theorem proves that a degree-\(r\) functional on sections is represented through a shift-\(L\) perfect pairing by a section of degree \(r+L\). The proof takes homogeneous components and uses injectivity of the pairing isomorphism. Removing a nonzero incoming product of degree \(n_x\) then produces a kernel vector of degree \(r+L-n_x\), by injectivity of scalar multiplication. Thus every degree in the following calculation is justified by a component argument.
The constructed perfect pairing on a reduced word of length \(m\) sends input degrees \(i,j\) to polynomial degree \(i+j-m\). Support duality then identifies the full outgoing kernel at a vertex \(x\) with a shifted stalk dual. The product of incoming labels accounts for \(\ell(x)\) degrees. Thus a stalk basis degree \(d_i\) becomes a kernel basis degree \[ m-\ell(x)-d_i. \] Since \(m=\ell(b)\) for a reduced word with top \(b\), its character becomes \[ K_x^{\mathbf s}(q)=q^{\ell(b)-\ell(x)}C_x^{\mathbf s}(q^{-1}). \]
In our one-edge example, the lower stalk is \(A\), with basis degree zero, and its outgoing kernel is \(\alpha A\), with basis degree one. The formula predicts \(q\cdot1=q\), exactly as calculated directly.
We can also see the duality map itself in this example. Tensor induction gives \[ \langle a\otimes b,c\otimes d\rangle=ac\,\operatorname{tr}(bd). \] In the basis \(1\otimes1,1\otimes\alpha\), the pairing matrix is again \(\left(\begin{smallmatrix}0&1\\1&0\end{smallmatrix}\right)\), now over \(A\). Write two compatible section pairs as \((f,g)=(u+\alpha v,u-\alpha v)\) and \((f',g')=(u'+\alpha v',u'-\alpha v')\). Their pairing is \[ uv'+vu'=\frac{ff'-gg'}{2\alpha}. \] The displayed quotient belongs to \(A\): both pairs agree modulo \(\alpha\), so the numerator is divisible by \(\alpha\).
A section supported only at the lower vertex has the form \((\alpha h,0)\). The stalk functional \(f\mapsto cf\) is represented by the supported section \((2\alpha c,0)\), because \[ \langle(2\alpha c,0),(f,g)\rangle=cf. \] Thus the degree-zero dual generator of the stalk corresponds to a supported generator of degree one. The factor \(2\) is an invertible scalar. This calculation shows concretely how the outgoing kernel is related to a shifted stalk dual; the general support-duality theorem constructs this relation with all incoming labels accounted for.
Polynomial reflection of the kernel character
The kernel basis degrees above are nonnegative. Thus, for a duality shift \(L\) on \([1,b]\), every stalk basis degree \(d_i\) satisfies \(d_i\le L-\ell(x)\). The full outgoing-kernel module has a finite homogeneous basis with degrees \(L-\ell(x)-d_i\), so its character is polynomial reflection of the stalk character in degree \(L-\ell(x)\).
For the word sheaf of a reduced word and for the actual boundary sheaf, \(L=\ell(b)\). Both have the formula \[ K_x(q)=\operatorname{reflect}_{\ell(b)-\ell(x)}C_x(q). \] The degree bound identifies this polynomial with \(q^{\ell(b)-\ell(x)}C_x(q^{-1})\). The reflected-basis theorem spells out the nonzero-vector argument and the natural-number degrees used by Lean.
The full kernel here consists of the stalk vectors killed by every outgoing lower restriction. The bundled kernelData and the kernel of the combined outgoing map have the same underlying vectors, which is the identification used by the character construction.
Adjoint maps, image sheaves, and the boundary pairing
The splitting theorem supplies maps \(i:B\to W\) and \(p:W\to B\), with \(pi=1\), from the normalized boundary sheaf into a reduced-word sheaf. Their composite \(e=ip\) is a projector. The section equivalence identifies its action on global sections with an endomorphism \(F\) of the word module.
The perfect word pairing defines \[ F^*=D^{-1}F^\vee D,\qquad \langle F^*m,n\rangle=\langle m,Fn\rangle. \] Composition reverses under the adjoint, so \(F^*\) is also a projector. Homogeneous representatives show that it has degree zero.
To descend this operator to stalks, fix a projection \(\pi_x\). Its kernel consists exactly of the word elements annihilated by every supported element at \(x\). The original sheaf map preserves that supported submodule. Hence, if \(\pi_x(m)=0\), \[ \langle s,F^*m\rangle=\langle Fs,m\rangle=0 \] for every supported \(s\), proving \(\pi_x(F^*m)=0\). Quotient descent defines the adjoint stalk map. The pair-projection quotient supplies the edge map and both commuting squares. This produces the graded sheaf projector \(e^*\).
Its image sheaf \(C\) has the component images as stalks and edge modules. A section of this image sheaf is fixed by the section projector, and every value of the section projector is an image-sheaf section. Thus \[ \Gamma(C)\simeq\operatorname{im}F^*. \] The algebraic adjoint-summand theorem now gives \[ \Gamma(C)\simeq\Gamma(B)^*,\qquad c\longmapsto(w\mapsto\langle c,i(w)\rangle). \] Its inverse represents the extended functional \(\lambda\circ p\) by the perfect word form.
The image \(C\) inherits the quotient rule, flabbiness, and global generation by inclusion, extension in the word sheaf, and projection. Its stalks are graded direct summands of free stalks, hence projective and then graded free.
The identity of \(B\) is primitive, so its ambient projector \(e\) is primitive. The adjoint preserves primitivity, and the image-splitting correspondence makes \(1_C\) primitive. At the top, both word and boundary stalks are \(A\). A split pair there consists of multiplication by two scalars whose product is one, so its reverse composite is also one. The fixed-stalk theorem makes \(e^*\) the identity at that same top stalk.
These facts supply a normalized top splitting into \(C\). Boundary recognition gives an isomorphism \(B\simeq C\). With \(j:C\to W\) and \(q:W\to C\), the specified canonical comparison is \[ h=qi:B\to C,\qquad jh=e^*i. \] At the top its inverse is \(pj\), since both ambient projectors are the identity. The recognized-target form of top detection then proves that this particular \(h\) is an isomorphism everywhere.
Pull the displayed duality back through \(h\). For global sections \(v,w\) of \(B\), \[ \langle h(v),i(w)\rangle =\langle F^*i(v),i(w)\rangle =\langle i(v),Fi(w)\rangle =\langle i(v),i(w)\rangle. \] The last equality uses \(Fi=i\). This proves that the induced perfect boundary pairing is the word form on the included sections. Its symmetry, degree shift, and support conditions pass through the inclusion maps.
Finally, a basis identifies the symmetric coefficient algebra with a polynomial ring by a degree-preserving algebra isomorphism \(\varphi\). Changing the scalar action transports each pairing value \(a\) to \(\varphi(a)\). The transported sheaf retains the quotient, generation, flabbiness, free-stalk, and primitive properties, so boundary recognition identifies it with the chosen polynomial-coordinate boundary sheaf. Pullback along that isomorphism supplies the pairing used in the polynomial character argument.
The word kernel and the R-convolution
The section isomorphism of a reduced word transports its perfect word pairing to all global sections. Its inverse preserves degree because all stalk projections detect each homogeneous component. A section supported at \(x\) pairs to zero with a section vanishing at \(x\), by the word support-annihilator theorem. These facts supply the actual polynomial word perfect pairing and its kernel duality.
Two results now meet at the word character. Geometry gives the kernel character as polynomial reflection. The positive Hecke-word bar identity gives the R-convolution of its standard coefficients. The word-character theorem identifies those coefficients with the actual word-stalk polynomials. In the Laurent polynomial ring, the unit powers simplify to \(q^{\ell(b)-\ell(x)}\); the proved degree bound and injective polynomial embedding then give \[ K_x^{\mathbf t}(q) =\operatorname{reflect}_{\ell(b)-\ell(x)}C_x^{\mathbf t}(q) =\sum_{z\le b}R_{x,z}(q)C_z^{\mathbf t}(q). \]
Cancellation of lower boundary models
Let the word sheaf decompose into shifted boundary models on \([1,b]\). For each center \(c\le b\), let \(\mu_c(q)\) count its shifts, and let \(L_z^c=C_{z,c}\) for \(z\le c\), with value zero otherwise.
Put \(\delta(x,c)=\max\{\ell(c)-\ell(x),0\}\), the natural-number rank difference. The same graded sheaf decomposition gives \[ C_z^{\mathbf t}=\sum_c\mu_cL_z^c,\qquad K_x^{\mathbf t}=\sum_c\mu_c\operatorname{reflect}_{\delta(x,c)}L_x^c. \] The second identity comes from the full kernels commuting with sums and shifts. Its coefficient \(\mu_c\) is therefore exactly the coefficient in the first identity.
Substitute both formulas into the word R-convolution and interchange the finite sums: \[ 0=\sum_c\mu_c\left( \operatorname{reflect}_{\delta(x,c)}L_x^c -\sum_zR_{x,z}L_z^c\right). \] For \(c\ne b\), the center has smaller length. Strong induction gives reciprocity on \([1,c]\), and the lower-character convolution theorem transports it to the ambient sum. If \(x\not\le c\), both sides vanish by the two support conditions.
At the top, the word character is \(1\), and the only model contributing there is the model centered at \(b\), also with character \(1\). Thus \(\mu_b=1\). Every other defect vanishes, so the remaining equation is \[ \operatorname{reflect}_{\ell(b)-\ell(x)}C_{x,b} =\sum_{z\le b}R_{x,z}C_{z,b}. \] The boundary kernel degree bound gives the reciprocal expression. For \([1,s]\), the lower-vertex identity reads \(q=1+(q-1)\).
Scaled actual characters and normalized kernel endpoints
For \(t\ne0\), define the scaled actual-stalk character \[ c_x^b(t)=t^{\ell(x)-\ell(b)}C_{x,b}(t^2). \] Evaluating reciprocity at \(q=t^2\) and using the ordinary-to-normalized R relation gives \[ c_x^b(t)=\sum_{z\le b}\widetilde R_{x,z}(t^{-1}-t)c_z^b(t^{-1}). \] For \(x\in[a,b]\), terms with \(z\not\ge a\) vanish, so this formula uses the restricted vector on the upper interval.
In the real comparison range \(0<t<1\), the normalized kernel value multiplies the Hilbert sum at \(t^2\) by \((1-t^2)^Nt^{\ell(x)-\ell(b)}\). The empty-condition module is the stalk, giving \[ D_x(t,\varnothing)=c_x^b(t). \] The full outgoing kernel has reflected character, giving \[ D_x(t,\mathrm{all})=c_x^b(t^{-1}). \] The polynomial coefficient factors cancel, and the full-kernel power is \(t^{-d}(t^2)^d=t^d\). All outgoing edges at a vertex of \([a,b]\) remain in that upper restriction, so the same endpoint formulas apply there.
These are the actual-character vectors used in the schedule comparison. The scaled reciprocity identity identifies their relation to the normalized R matrix.
Local intersection maps and positive-ideal coordinates
For the actual boundary sheaf, write \(M=B(b)^x\), \(K_x=\ker\partial_x\), and \(\kappa_x:K_x\to M^*\) for the stalk–costalk isomorphism. The local intersection map is \[ I_x:M^*\longrightarrow M,\qquad I_x=\iota_x\kappa_x^{-1}. \] Its image is exactly \(K_x\): one inclusion comes from \(\iota_x\), and any \(k\in K_x\) is the image of \(\kappa_x(k)\).
The symmetry statement is an identity of coefficient-valued pairings, \(\lambda(I_x\mu)=\mu(I_x\lambda)\). It follows by multiplying the kernel pairing by the incoming label product \(\beta_x\), expressing that product as the pairing of two supported global sections, using global symmetry, and cancelling the nonzero \(\beta_x\).
Let \(m_i\) be a homogeneous stalk basis of degree \(d_i\), with dual basis \(\lambda_i\). If \(d=\ell(b)-\ell(x)\), inverse duality gives \(I_x\lambda_j\in M_{d-d_j}\), so the matrix entry \[ a_{ij}=\lambda_i(I_x\lambda_j) \] lies in \(A_{d-d_i-d_j}\). At \(x<b\), boundary minimality puts the entire image in \(A_+M\), hence every \(a_{ij}\) belongs to \(A_+\).
A specified nonzero diagonal entry \(a_{ii}\) therefore has strictly positive degree: negative graded pieces vanish, and augmentation kills \(A_+\) while fixing degree-zero elements. Thus \[ d-2d_i>0. \] The positive-ideal diagonal-degree theorem develops this argument for a general graded algebra, and the local diagonal-degree theorem applies it to the actual boundary stalk.
Boundary conditions and KL endpoint substitution
The section-duality structure packages a perfect pairing balanced under the cut operators. For the actual boundary family, the constructed symmetric pairing supplies that structure. Its stalk–costalk isomorphisms prove the full outgoing-kernel freeness condition, because the dual of a finite free stalk is finite free.
The stalk-character identification condition is the polynomial equality \[ C_x^{[u,b]}(q)=P_{x,b}(q)\qquad(x\in[u,b]). \] Here \(C\) counts homogeneous basis degrees of the chosen boundary stalks, while \(P\) is the canonical family defined by its R-recurrence and strict degree normalization. Boundary uniqueness makes \(C\) independent of the chosen normalized model, and upper restriction identifies it with the corresponding character on \([1,b]\).
Under the equality at \(x\), the KL endpoint substitution theorem gives \[ D_x(t,\varnothing)=t^{\ell(x)-\ell(b)}P_{x,b}(t^2) \] for \(0<t<1\). On \([1,b]\), a section duality with shift \(\ell(b)\) also gives \[ D_x(t,U_x)=t^{\ell(b)-\ell(x)}P_{x,b}(t^{-2}). \] The first identity substitutes the character into the free-stalk Hilbert formula; the second uses the full-kernel Hilbert formula and the same substitution.
The OpenAI notes obtain the equality of \(C\) and \(P\) through their cited character theorem and grading conversion. The Lean comparison route uses the proved reciprocity of \(C\) at its endpoints, then treats canonical \(P\) through polynomial rigidity. The character-identification condition and the KL-labelled endpoint lemmas record the additional identification explicitly.
Definitions and proofs by construction
Ordered graphs and labels
Sheaves, sections, and boundary data
- Module-valued sheaves on an ordered graph
- Free stalks, upper quotients, and global generation
- Graded sheaves and homogeneous restriction maps
- Compatible assignments and sections on a vertex subset
- Flabbiness and closure under directed edges
- Extension from subsets closed under surviving edges
- Boundary values supplied by upper sections
- The stalk surjects onto the boundary image
- The upper-end edge quotient rule
- The actual edge module over its coefficient ring
- Kernels of upward edge conditions
Operations and morphisms
- Morphisms and partial morphisms of graded graph sheaves
- Sheaf isomorphisms and their partial forms
- Lifting a partial morphism through the boundary image
- Adding one vertex to a partial sheaf morphism
- Extending a sheaf morphism from an upper set
- Tensor base change of a graph sheaf
- Section and outgoing-kernel constraint maps
- Sections commute with finite flat scalar extension
- Sheaf properties under scalar extension
- Sheaf properties under pullback along surviving edges
- Selected kernels under outgoing-complete graph pullback
- Zero sheaves, finite sums, and grading shifts
- Structural properties and kernels of finite sheaf sums
- Extension by zero from a lower ideal
- Sheaf properties preserved by lower-ideal extension
- Extending a partial sheaf isomorphism across one vertex
- Boundary and stalk images under an isomorphism above a vertex
Boundary sheaves and minimal extension
- Minimal-boundary sheaves and their normalized stalk
- Boundary-sheaf isomorphisms are detected at the normalized stalk
- Boundary recognition after a change of coefficient coordinates
- Existence of normalized boundary sheaves on finite ordered graphs
- Attaching a new bottom stalk by a minimal cover
- Uniqueness of the minimal-boundary sheaf up to graded isomorphism
- Boundary-sheaf pullback with complete outgoing edge lifts
- The boundary-sheaf family obtained by upper-interval restriction
- Stalk characters and selected kernels under upper-interval restriction
Summands, primitive objects, and decomposition
- Split sheaf summands and their projectors
- Image sheaves and the splitting of a projector
- Sections of an image sheaf are the image of the section projector
- Sheaf properties inherited by a graded split summand
- Primitivity of a split summand and its image projector
- Recognition of a boundary sheaf by a primitive top splitting
- Primitive sheaf projectors
- A nonzero split summand of a primitive sheaf is the whole sheaf
- Extending a split top stalk to a sheaf summand
- Indecomposability of the normalized boundary sheaf
- Boundary models supported on principal lower ideals
- Finite constructions from shifted boundary models
- Decomposing a free flabby sheaf into shifted boundary models
- Construction of normalized models at every vertex
- Structural properties of a model sum and minimality away from its centers
- Decomposing a stalk over an identified upper boundary
- Adding the next center in a boundary-model decomposition
Integral support and cut operators
- Sections supported at a vertex and their stalk values
- Integral vertex cuts and products of edge labels
- Vertex support equals the annihilator of an evaluation kernel
- Incoming labels identify the outgoing kernel with vertex support
Sheaves associated to words
- The pair-projection sheaf of a Bott–Samelson module
- Free word stalks and root quotients in polynomial coordinates
- Extending an interval assignment to its ambient upper closure
- Flabbiness of the actual word sheaf
- A reduced tensor word realizes all global sections
- A normalized boundary sheaf is a graded summand of a word sheaf
Perfect pairings and adjoints
- Balanced perfect duality on global sections
- Perfect graded pairings on reduced-word sections
- Symmetry of the stalk–costalk identification
- Perfect graded pairings and support orthogonality
- Support orthogonality makes integral cuts balanced
- Supported sections are dual to a generated stalk
- Integral adjoints for a perfect pairing
- Composition, balance, and grading of adjoints
- The adjoint image is dual to a split summand
- The integral word adjoint descends to a graded sheaf map
- Fixed stalks and the top normalization of adjoint projectors
- Recognition of the adjoint boundary image
- The canonical comparison with the adjoint image is an isomorphism
- Transport of perfect section pairings through coefficient and sheaf isomorphisms
- Primitivity is preserved by the adjoint operation
- Symmetric perfect pairings on the actual boundary sheaf
- The full outgoing kernel is dual to the stalk
- The degree of the inverse stalk–costalk duality map
Stalk characters and actual-character reciprocity
- Stalk and outgoing-kernel characters of graded sheaves
- The grading shift in stalk–costalk reciprocity
- The outgoing-kernel character of a reduced word
- Boundary kernel characters and their degree bound
- Polynomial R-convolution for word-stalk and kernel characters
- Outgoing kernels under sheaf isomorphisms, sums, and shifts
- One multiplicity family controls stalk and full-kernel characters
- Stalks and kernels of the lower-supported Bruhat models
- Simultaneous decomposition of Bott–Samelson stalk and kernel characters
- Lower-character normalization and convolution on an ambient interval
- Cancelling lower summands in character identities
- R-reciprocity for the actual boundary-sheaf character
- The scaled actual-stalk character vector
- Scaled reciprocity and its restriction to an upper interval
- Normalized empty and full kernels from the actual character
Related character-identification conditions
- Balanced perfect duality on global sections
- The full outgoing-kernel freeness condition
- The local intersection map and its matrix
- Image, symmetry, and grading of the local intersection map
- Positive-ideal coordinates and strict diagonal degree
- A nonzero local diagonal forces a strict half-degree bound
- The stalk-character identification condition
- KL endpoint formulas under character identification