From one-edge bounds to the full transfer inequality
Fix a finite-generator Coxeter system \(W\), \(u\le_W b\), a finite real basis \(\Sigma\) of its extended realization, and a boundary sheaf \(\mathcal B\) with top \(b\) on the corresponding polynomially labelled graph of \([u,b]_W\). The coefficient ring is \(A=\mathbb R[X_i:i\in\Sigma]\).
Let \(p\) be a finite list of graph edges. A suffix of \(p\) is a list \(q\) such that \(p=r+\!\!+q\) for some list \(r\); write \(q\preceq_{\mathrm{suf}}p\). For every suffix \(e::q\), put \(x=s(e)\), \(y=r(e)\), where \(s,r\) denote source and target, and define the actual edge images
\[
I_e(q)=\rho_e^-\bigl(K_x(F_q(x))\bigr),\qquad
J_e(q)=\rho_e^+\bigl(K_y(F_q(y))\bigr).
\]
Here \(F_q\) is the remaining-condition set, and \(K\) its selected kernel. Let \(\mu_e:\mathcal B^e\to\mathcal B^e\) be the edge-factor multiplication map edgeFactorMap. Assume
\[
I_e(q)\subseteq\mu_e(J_e(q))
\qquad\text{for every }e::q\preceq_{\mathrm{suf}}p.
\tag{1}
\]
For \(0<t<1\), set \(T=t^{-1}-t>0\), and let \(D_q(t)\) be the Hilbert-state vector. Then \[ D_{[]}(t)\le\mathcal D_{p^{\mathrm{rev}}}(T)D_p(t) \tag{2} \] coordinatewise. The hypothesis (1) is specified for each actual tail \(q\); the conclusion applies to any finite edge list satisfying these hypotheses.
Relation to the OpenAI notes
Actual suffix conditions and the conditional whole-list bound.
OpenAI notes: The comparison processes the increasing edge schedule, retains the equations on unprocessed outgoing edges, and propagates the single-edge estimate through the complete matrix product. eq:state, eq:step, eq:comparison, edge:bound.
Lean (primary): The numerical theorem first accepts any finite list with image inclusion at every nonempty suffix. A separate theorem takes the actual chronological schedule and freeness of each upper suffix kernel. Exact suffix-to-ambient kernel identities let it apply the global edge inclusion before invoking the numerical theorem. PositiveTransfer:1089, PositiveTransfer:1117, PositiveTransfer:1148, TerminalKernels:1126, TerminalKernels:1171.
Role in the proof: The Blueprint spells out the set-valued remaining conditions, source and non-source coordinate changes, suffix quantifiers and reversed product. The source-specific upper-freeness step has its own entry.
Proof
Step 1. The states before and after one removal
Fix a suffix \(e::q\), and abbreviate \(x=s(e)\), \(y=r(e)\), \(F=F_q(x)\), \(H=F_q(y)\). The remaining-condition identities give \[ F_{e::q}(x)=F\cup\{e\},\qquad F_{e::q}(z)=F_q(z)\quad(z\ne x). \] The first state is \(D_{e::q}(t)\). After processing \(e\), it is \(D_q(t)\). Since the edge is strictly increasing, \(x\ne y\), and the second identity applies in particular at \(y\).
Step 2. The inequality at the source
Apply the one-edge Hilbert comparison to \(e,F,H\) and inclusion (1). It gives \[ D_x(t,F)\le D_x(t,F\cup\{e\})+T D_y(t,H). \] The first two values are \(D_q(t)_x\) and \(D_{e::q}(t)_x\). At \(y\), the selected set is unchanged, so \(D_y(t,H)=D_{e::q}(t)_y\). Substitution yields \[ D_q(t)_x \le D_{e::q}(t)_x+T D_{e::q}(t)_y =\bigl(U_e(T)D_{e::q}(t)\bigr)_x. \]
Step 3. Equality at every other coordinate
For \(z\ne x\), the selected sets agree, so
\[
D_q(t)_z=D_{e::q}(t)_z.
\]
The elementary transfer also leaves coordinate \(z\) unchanged. Thus
\[
D_q(t)_z=\bigl(U_e(T)D_{e::q}(t)\bigr)_z.
\]
Together with Step 2 this proves the vector inequality
\[
D_q(t)\le U_e(T)D_{e::q}(t).
\tag{3}
\]
This is scheduleState_step.
Step 4. The induction with all suffix hypotheses
The generic remaining_state_le applies to any function \(D\) from finite edge lists to real vectors and any \(T\ge0\). Its hypothesis is
\[
D_q\le U_e(T)D_{e::q}
\quad\text{for every }e::q\preceq_{\mathrm{suf}}p.
\]
We prove \(D_{[]}\le\mathcal D_{p^{\mathrm{rev}}}(T)D_p\) by induction on \(p\), keeping \(D\) fixed and carrying these suffix hypotheses.
For \(p=[]\), its reverse is empty and the product is the identity, so the conclusion is \(D_{[]}\le D_{[]}\).
Step 5. Propagation through the tail
Let the list be \(p=e::h\), and suppose the result holds for \(h\). Every suffix \(a::q\) of \(h\) is also a suffix of \(e::h\): concatenate the initial segment witnessing the first suffix relation with the initial edge \(e\). Thus the original hypotheses supply all the hypotheses for the shorter list, giving \[ D_{[]}\le\mathcal D_{h^{\mathrm{rev}}}(T)D_h. \tag{4} \] The full list \(e::h\) is its own suffix. Its one-step hypothesis says \[ D_h\le U_e(T)D_{e::h}. \tag{5} \] By the transfer properties, the tail product is monotone for \(T\ge0\). Apply it to (5) and combine with (4): \[ D_{[]}\le \mathcal D_{h^{\mathrm{rev}}}(T)D_h \le\mathcal D_{h^{\mathrm{rev}}}(T)\bigl(U_e(T)D_{e::h}\bigr). \tag{6} \]
Step 6. The reversed product
Reversing \(e::h\) gives \(h^{\mathrm{rev}}+\!\!+[e]\). The concatenation formula gives \[ \mathcal D_{(e::h)^{\mathrm{rev}}}(T)v =\mathcal D_{h^{\mathrm{rev}}}(T) \bigl(\mathcal D_{[e]}(T)v\bigr) =\mathcal D_{h^{\mathrm{rev}}}(T)(U_e(T)v). \] Use \(v=D_{e::h}\) in (6). This is precisely the induction conclusion for \(e::h\), completing the generic proof.
For the sheaf states, Steps 1–3 supply its hypotheses at every suffix, and \(0<t<1\) gives \(T=t^{-1}-t>0\). Substitution proves (2), the source's scheduleState_le.
Step 7. A two-edge instance
For chronological list \([e_1,e_2]\), (3) gives \[ D_{[e_2]}\le U_{e_1}D_{[e_1,e_2]},\qquad D_{[]}\le U_{e_2}D_{[e_2]}. \] Applying \(U_{e_2}\) to the first inequality and using the second gives \[ D_{[]}\le U_{e_2}D_{[e_2]} \le U_{e_2}U_{e_1}D_{[e_1,e_2]}. \] This is dispatch on the reversed list \([e_2,e_1]\). Each intermediate vector is still the Hilbert value of its actual remaining kernels.
References
The local source declarations below give the construction and proof.
Lean formalization
OAI.KLInvariance.SchedulePaths.remaining_state_le: PositiveTransfer.lean:866.
OAI.KLInvariance.BruhatGraph.scheduleState_step: PositiveTransfer.lean:1117.
OAI.KLInvariance.BruhatGraph.scheduleState_le: PositiveTransfer.lean:1148.