Selected kernels and scaled images under sheaf isomorphisms
Let \(R\) be a commutative ring, \(G\) an ordered graph, and \(j:B\simeq C\) an \(R\)-linear sheaf isomorphism. Write \(j_x\) for its vertex maps and \(j_e\) for its edge maps.
For any vertex \(x\) and outgoing selection \(F\), \[ j_x(K_x^B(F))=K_x^C(F). \] For an edge \(e:x\to y\) and selections \(F,H\) at its endpoints, \[ j_e(L_e^B(F))=L_e^C(F),\qquad j_e(U_e^B(H))=U_e^C(H). \] Consequently, for every scalar \(r\in R\), \[ L_e^B(F)=rU_e^B(H) \quad\Longleftrightarrow\quad L_e^C(F)=rU_e^C(H). \]
Relation to the OpenAI notes
The same embedded images through one edge component.
OpenAI notes: The model decomposition carries both localized kernels and their endpoint images into the model summands. Its single edge component carries both comparison submodules, and linearity carries the correction scalar through that same map. edge:bound.
Lean (primary): Selected-kernel membership and both image equalities are proved for arbitrary sheaf isomorphisms over a commutative ring. ScaledImages is an elementwise witness equivalence, identified with submodule equality. Coordinatewise witness transport proves the product and model-sum statements with one common scalar. PositiveTransfer:362, PositiveTransfer:366, CharacterReciprocity:443, GradedSheaves:631, GradedSheaves:639, TerminalKernels:19, CharacterReciprocity:548.
Role in the proof: The Blueprint defines the actual submodules and witnesses before transporting them. The generic ungraded transport is the one applied to the local model decomposition.
Proof
Step 1. Transport a selected zero restriction
For an outgoing edge \(f\) from \(x\), the lower restriction square is \[ \rho^{C,-}_f(j_xv)=j_f(\rho^{B,-}_fv). \] If \(\rho^{B,-}_f(v)=0\), linearity makes the right side and hence the left restriction zero.
Conversely, if the left restriction is zero, injectivity of \(j_f\) forces \(\rho^{B,-}_fv=0\). Applying this to every \(f\in F\) gives \[ j_xv\in K_x^C(F)\quad\Longleftrightarrow\quad v\in K_x^B(F). \]
Step 2. Prove equality of the transported kernel submodules
Step 1 gives the containment \(j_x(K_x^B(F))\subseteq K_x^C(F)\).
For the reverse containment, take \(w\in K_x^C(F)\) and put \(v=j_x^{-1}w\). Step 1 puts \(v\) in \(K_x^B(F)\), and \(j_xv=w\). Thus the two submodules are equal.
Step 3. Transport both endpoint images
A value in \(L_e^B(F)\) has the form \(\rho^{B,-}_e(v)\) with \(v\in K_x^B(F)\). The lower square gives \[ j_e(\rho^{B,-}_e(v))=\rho^{C,-}_e(j_xv). \] Step 2 places \(j_xv\) in the corresponding selected kernel of \(C\). This proves one containment of lower images. Using \(j_x^{-1}\) for a lower witness in \(C\) proves the reverse containment.
The upper square \[ j_e(\rho^{B,+}_e(w))=\rho^{C,+}_e(j_yw) \] gives the same argument at \(y\), proving the upper-image equality. Both equalities use the single edge map \(j_e\).
Step 4. Commute a scalar image with a linear isomorphism
For any \(R\)-linear isomorphism \(T:M\simeq N\), scalar \(r\) and submodule \(J\subseteq M\), \[ T(rJ)=rT(J). \] Indeed \(T(rw)=rT(w)\) gives the forward containment. Every value \(rT(w)\), with \(w\in J\), is the image of \(rw\in rJ\), giving the reverse containment.
For another submodule \(I\subseteq M\), if \(T(I)=rT(J)\), apply \(T^{-1}\) and the same identity to obtain \(I=rJ\). The forward implication follows by applying \(T\). These are the generic EmbeddedModule scalar-transport formulas.
Step 5. Transfer the scaled equality in both directions
Apply Step 4 to \(T=j_e\). If \(L_e^B(F)=rU_e^B(H)\), then Step 3 gives \[ L_e^C(F) =j_e(L_e^B(F)) =j_e(rU_e^B(H)) =rj_e(U_e^B(H)) =rU_e^C(H). \] Apply the same argument to the inverse sheaf isomorphism for the converse. This proves invariance of the ScaledImages proposition.
For the local model decomposition, \(j_e\) is the actual edge component of its sheaf isomorphism. Its \(R_P\)-linearity is exactly the scalar compatibility used in Step 4.
References
The local source declarations below give the construction and proof.
Lean formalization
OAI.KLInvariance.MomentGraph.Sheaf.Iso.upwardKernel_iff: GradedSheaves.lean:609.
OAI.KLInvariance.MomentGraph.Sheaf.Iso.map_upwardKernel: GradedSheaves.lean:620.
OAI.KLInvariance.MomentGraph.Sheaf.Iso.map_lowerImage: GradedSheaves.lean:631.
OAI.KLInvariance.MomentGraph.Sheaf.Iso.map_upperImage: GradedSheaves.lean:639.
OAI.KLInvariance.MomentGraph.Sheaf.Iso.scaledImages: TerminalKernels.lean:19.
OAI.KLInvariance.EmbeddedModule.map_scaled: GradedSheaves.lean:657.
OAI.KLInvariance.EmbeddedModule.map_eq_scaled_iff: GradedSheaves.lean:665.