Selected kernels under outgoing-complete graph pullback
Let \(R\) be a commutative ring, \(f:G\to H\) a graph map, and \(B\) an \(R\)-linear sheaf on \(H\). Fix a vertex \(x\in V_G\) and an arbitrary set \(T\subseteq E_H\).
Choose \(L\subseteq E_H\) with \(B^e=0\) for every \(e\notin L\). Suppose each edge of \(L\) starting at \(f_V(x)\) has a lift starting at \(x\): \[ e\in L,\quad s_H(e)=f_V(x) \Longrightarrow \exists a,\quad s_G(a)=x,\quad f_E(a)=e. \]
For a set of ambient edges, let \(K_y^B(T)\) impose vanishing only on those members of \(T\) starting at \(y\). Then, in the common stalk \((f^*B)^x=B^{f_V(x)}\), \[ K_x^{f^*B}(f_E^{-1}(T))=K_{f_V(x)}^B(T). \]
Relation to the OpenAI notes
Relative extension through an explicit graph map.
OpenAI notes: The coset intersection is a finite relative lower ideal. A relative upper subset is closed under the surviving plane-labelled edges. Descending ambient extension gives relative flabbiness after localization. loc:relative-flabbiness.
Lean (primary): The same descending extension is factored into extend_closedUnder and a graph-pullback theorem. The coset map supplies vertex injectivity and explicit outgoing lifts; the root/unit criterion makes all outside edge modules zero. A separate theorem records selected-kernel equality under that pullback. Sheaves:106, MinimalExtensions:1520, PlaneLocalization:497, PlaneLocalization:591, TerminalKernels:393.
Role in the proof: Relative order determines the upper subsets on the coset, while the ambient graph supplies the extension. The localized stalks and edge maps remain the actual tensor-extended ambient modules, so selected kernels and their edge images retain their embeddings.
Proof
Step 1. Write both selected systems
A vector \(v\in B^{f_V(x)}\) belongs to the left kernel exactly when \[ \rho^{f^*B,-}_a(v)=0 \] for every edge \(a\) starting at \(x\) with \(f_E(a)\in T\).
The pullback formula identifies this restriction with \(\rho^{B,-}_{f_E(a)}(v)\), through \(s_H(f_E(a))=f_V(x)\). The right kernel requires this same vanishing for every target-graph edge in \(T\) starting at \(f_V(x)\).
Step 2. Pull ambient conditions back
Suppose \(v\) lies in the right kernel. For any selected outgoing edge \(a\) at \(x\), the endpoint identity makes \(f_E(a)\) an outgoing edge at \(f_V(x)\), and it belongs to \(T\) by selection. Its restriction vanishes, so the pullback restriction vanishes. Thus \(v\) belongs to the left kernel.
Step 3. Recover every surviving ambient condition
Conversely, suppose \(v\) lies in the left kernel. Let \(e\in T\) start at \(f_V(x)\).
If \(e\in L\), take an outgoing lift \(a\) with \(f_E(a)=e\). It is selected in the pullback, so \[ 0=\rho^{f^*B,-}_a(v)=\rho^{B,-}_e(v). \]
If \(e\notin L\), its edge module is zero and the restriction vanishes. Every required ambient condition is therefore satisfied.
Step 4. Identify the underlying stalk vector
Both containments used the same vector \(v\) in \(B^{f_V(x)}\), proving equality as submodules.
The source expresses the endpoint identifications by the assignment supported at \(f_V(x)\) with value \(v\). Its pullback has value \(v\) at \(x\), and stalkMap_assignment rewrites each component restriction. The resulting equations are exactly those checked above.
References
The local source declarations below give the construction and proof.
Lean formalization
OAI.KLInvariance.MomentGraph.Sheaf.stalkMap_assignment: MinimalExtensions.lean:1590.
OAI.KLInvariance.MomentGraph.Sheaf.pullback_kernel: MinimalExtensions.lean:1596.