Theorem Endpoint equality forces equality at every transfer step
✓ theorem proved

Endpoint equality forces equality at every transfer step

Let \(V,E\) be sets with source and target maps \(s,r:E\to V\), let \(T\ge0\), and let \(D\) assign a real vector \(D_q:V\to\mathbb R\) to each finite edge list \(q\). Fix a finite list \(p\). Assume:

  1. Every edge in \(p\) has distinct endpoints.
  2. For every suffix \(e::q\preceq_{\mathrm{suf}}p\), \[ D_q\le U_e(T)D_{e::q}. \tag{1} \]
  3. The final state equals the exact transfer product: \[ D_{[]} =\mathcal D_{p^{\mathrm{rev}}}(T)D_p. \tag{2} \]

Then every inequality (1) is an equality. This is SchedulePaths.remaining_steps_eq.

For the Hilbert states of a boundary sheaf on a finite-generator Coxeter interval with a finite real coefficient basis, assume \(0<t<1\), \(T=t^{-1}-t\), and the embedded image inclusion at every suffix as in the full transfer inequality. If (2) holds, then for every such suffix \[ \rho_e^-\bigl(K_{s(e)}(F_q(s(e)))\bigr) =\mu_e\rho_e^+\bigl(K_{r(e)}(F_q(r(e)))\bigr) \tag{3} \] inside the actual edge module. Here \(\mu_e\) is the edge-factor multiplication map. This is edgeImages_eq_of_scheduleState_eq.

Relation to the OpenAI notes

Cancellation through the remaining transfer product.

OpenAI notes: Exact endpoint comparison forces the positive Hilbert defects to vanish and hence makes the embedded lower and corrected upper images equal. eq:step, eq:comparison, eq:image-equality.

Lean (primary): The finite-index statement removes the last transfer. The actual sheaf application uses a remaining-list statement: sandwich the tail-product values between equal endpoints, use injectivity of the tail product to recover exactness of the head removal, and recurse on the tail. Source-coordinate equality then invokes the normalized edge-increment equality criterion. Comparison:105, Comparison:119, PositiveTransfer:880, PositiveTransfer:1161.

Role in the proof: Both induction interfaces are explained. The actual embedded-image conclusion retains every suffix inclusion and the positive comparison-parameter hypotheses.

Proof

Step 1. Induction on the remaining list

We prove the first statement for every finite \(p\), carrying all three assumptions. When \(p=[]\), no nonempty list \(e::q\) is a suffix: its positive length would have to be at most zero. Thus there are no step equalities to prove.

Now write \(p=e::h\). Every suffix of \(h\) is also a suffix of \(e::h\), so the hypotheses (1) restrict to all steps of \(h\). Put \[ A=\mathcal D_{h^{\mathrm{rev}}}(T),\qquad v=D_h,\qquad w=U_e(T)D_{e::h}. \] The current one-step inequality is \(v\le w\).

Step 2. The tail-product values between equal endpoints

Apply the generic full transfer inequality to the shorter list \(h\). It gives \[ D_{[]}\le A v. \tag{4} \] The transfer properties give monotonicity of \(A\), so \(v\le w\) implies \[ A v\le A w. \tag{5} \] The reversed-list identity \((e::h)^{\mathrm{rev}}=h^{\mathrm{rev}}+\!\!+[e]\) and the concatenation formula turn the endpoint equality (2) into \[ D_{[]}=A w. \tag{6} \] Combine (4)–(6): \[ D_{[]}\le A v\le A w=D_{[]}. \] In each coordinate, the intermediate real number is bounded above and below by the same real number. Consequently \[ A v=A w,\qquad D_{[]}=A v. \tag{7} \]

Step 3. Cancellation of the tail product

Every edge of \(h^{\mathrm{rev}}\) belongs to \(p\), so it has distinct endpoints. The transfer properties make \(A\) injective. Apply injectivity to \(A v=A w\) in (7): \[ D_h=v=w=U_e(T)D_{e::h}. \tag{8} \] This is the exactness of the head removal. The second equality in (7) supplies the endpoint equality for the shorter list: \[ D_{[]}=\mathcal D_{h^{\mathrm{rev}}}(T)D_h. \tag{9} \]

Step 4. Exactness at later suffixes

The distinct-endpoint condition and all one-step inequalities restrict to \(h\); (9) is its required endpoint equality. Therefore the induction hypothesis proves every step indexed by a suffix of \(h\) exact.

A nonempty suffix of \(e::h\) is either the entire list \(e::h\), handled by (8), or a suffix of \(h\), handled by this induction hypothesis. These two cases cover all required steps and complete the proof of the first statement.

Step 5. The source-coordinate equality

For a sheaf, the hypotheses of the full transfer inequality provide (1). Bruhat edges have distinct endpoints because they are strictly increasing. The assumed endpoint identity is (2), so the first statement gives \[ D_q(t)=U_e(T)D_{e::q}(t). \] At \(x=s(e)\), set \(y=r(e)\), \(F=F_q(x)\), \(H=F_q(y)\). The remaining-condition identities identify the source selection before removal as \(F\cup\{e\}\), and leave the target selection \(H\) unchanged. Evaluating the vector equality at \(x\) therefore gives \[ D_x(t,F)=D_x(t,F\cup\{e\})+T D_y(t,H). \tag{10} \]

Step 6. Equality of the actual edge images

The one-edge Hilbert comparison proves that (10), under the given inclusion \(I_e(q)\subseteq\mu_eJ_e(q)\) and \(0<t<1\), is equivalent to equality of the actual images. Its positive state scale allows cancellation to obtain equality of the raw Hilbert values. Degree by degree, the inclusion gives a nonnegative dimension difference; every weight \((t^2)^n\) is positive. A zero total difference therefore forces equal homogeneous pieces, and hence equal embedded submodules. This yields (3).

Step 7. The natural-number-indexed formulation

The source also states the transfer result for states \(D_0,\ldots,D_n\) and an edge sequence \(e_0,\ldots,e_{n-1}\). Define ideal states \[ E_0=D_0,\qquad E_{j+1}=U_{e_j}(T)E_j. \] From \(D_{j+1}\le U_{e_j}(T)D_j\), induction gives \(D_j\le E_j\): it starts with equality at zero, and the induction step applies monotonicity to \(D_j\le E_j\), followed by the given one-step bound.

If \(D_n=E_n\) and \(n>0\), these bounds give \[ D_n\le U_{e_{n-1}}D_{n-1} \le U_{e_{n-1}}E_{n-1}=E_n=D_n. \] Coordinatewise antisymmetry makes both inequalities equalities. Injectivity of \(U_{e_{n-1}}\) then gives \(D_{n-1}=E_{n-1}\). Induction on \(n\) proves exactness of all earlier steps, with no steps in the base case \(n=0\). This is Comparison.every_step_eq_of_endpoint_eq; the remaining-list proof above is the form invoked by the actual sheaf theorem.

References

The local source declarations below give the construction and proof.

Lean formalization

OAI.KLInvariance.Comparison.idealState: Comparison.lean:98.

OAI.KLInvariance.Comparison.state_le_ideal: Comparison.lean:105.

OAI.KLInvariance.Comparison.every_step_eq_of_endpoint_eq: Comparison.lean:119.

OAI.KLInvariance.SchedulePaths.remaining_steps_eq: PositiveTransfer.lean:880.

OAI.KLInvariance.BruhatGraph.edgeImages_eq_of_scheduleState_eq: PositiveTransfer.lean:1161.

Used by