Kl.Invariance
Hilbert comparison and combinatorial invariance
The proof uses an established inclusion of edge images to obtain a numerical bound. Exact endpoint identities later force these bounds to be equalities, which recover equality of the images and freeness of additional kernels. In the schedule application, the upper selected kernel freeness needed for the image inclusion is supplied by induction.
We will calculate the whole process for the one-edge sheaf from moment-graph sheaves. The general argument uses the same operations at each edge, with the hypotheses supplied by the simultaneous invariance induction.
One outgoing condition and its actual image
Let \(e:x\to y\) be an edge, and let \(F\) be the other selected outgoing conditions at \(x\). Put \[ K'=K_x(F\cup\{e\}),\qquad K=K_x(F),\qquad I_e=\rho_e^-(K). \] Restriction with codomain \(I_e\) is surjective. Its kernel consists precisely of the vectors already satisfying \(F\) whose \(e\)-restriction is also zero. Thus \[ 0\longrightarrow K'\longrightarrow K\longrightarrow I_e\longrightarrow0. \] The one-condition exact-sequence theorem checks both kernel directions and surjectivity in each homogeneous degree.
Hilbert values use all integer degrees: \[ H_K(q)=\sum_{n\in\mathbb Z}\dim_{\mathbb R}K_nq^n. \] The graded module data used by the sheaf impose \(M_n=0\) for \(n<0\). Therefore the displayed sum also equals its sum over \(n\ge0\). Finite polynomial modules have finite-dimensional pieces and convergent values for \(0<q<1\). The generic Hilbert image lemmas are stated for integer gradings; the sheaf application retains the additional nonnegative-support field.
Degreewise exactness and convergence give \[ H_K(q)-H_{K'}(q)=H_{I_e}(q). \] For the one-edge example over \(A=\mathbb R[\alpha]\), \(K'=\alpha A\), \(K=A\), and \(I_e=A/(\alpha)\), so the identity is \[ \frac1{1-q}-\frac q{1-q}=1. \]
The two factors in the upper-image value
Let \(H\) be an outgoing selection at the upper vertex \(y\), and \(J_e=\rho_e^+(K_y(H))\). The current label \(\alpha_e\) acts injectively on every selected outgoing edge module. Therefore \[ \alpha_ev\in K_y(H)\quad\Longrightarrow\quad v\in K_y(H): \] apply each selected restriction and cancel \(\alpha_e\) on its target edge module.
This saturation identifies the restricted upper-map kernel with \(\alpha_eK_y(H)\). The saturated-image Hilbert theorem, with the degree-one current label, gives \[ H_{J_e}(q)=(1-q)H_{K_y(H)}(q). \] For an injective correction map of degree \(k\), the homogeneous-image shift theorem gives \[ H_{f_eJ_e}(q)=q^kH_{J_e}(q) =q^k(1-q)H_{K_y(H)}(q). \] It constructs the piece isomorphism by taking the degree-\((n-k)\) component of a preimage in degree \(n\).
The raw edge-increment theorem now uses the embedded inclusion \(I_e\subseteq f_eJ_e\) to compare their piece dimensions. This inclusion is supplied by the global edge theorem when the required upper selected kernel is free. The Hilbert calculation starts from the inclusion and the actual label/correction regularity.
Equality at one positive parameter
The difference between the upper bound and the lower increment is \[ \sum_{n\in\mathbb Z} \bigl(\dim(f_eJ_e)_n-\dim(I_e)_n\bigr)q^n. \] Every term is nonnegative, and every \(q^n\) is positive. If the sum is zero, the dimensions agree in every degree. Inclusion then gives equality of the finite-dimensional pieces; homogeneity of the larger image gives equality of the full embedded submodules.
The raw edge-increment theorem proves this equivalence at the chosen \(q\). Its one-edge example uses lower image \((X,Y)\) inside the edge quotient \(\mathbb R[X,Y]\). The constant quotient contributes a defect of one: at \(q=1/2\), the lower increment is \(3\) and the upper bound is \(4\).
Normalization of the one-edge coefficient
Let \(A\) have \(N\) degree-one variables, with \(N\) the cardinality of the actual finite coefficient basis. Choose \(0<t<1\), put \(q=t^2\), and define \[ s_x(t)=(1-t^2)^N t^{\ell(x)-\ell(b)},\qquad D_x(t,F)=s_x(t)H_{K_x(F)}(t^2). \] The exponent is an integer. The state scale is strictly positive throughout the comparison range.
The edge degree is \(k=(\ell(y)-\ell(x)-1)/2\). Put \(m=\ell(x)-\ell(b)\) and \(n=\ell(y)-\ell(b)\), so \(m+2k=n-1\). The edge-normalization theorem computes \[ t^m(t^2)^k(1-t^2) =t^{n-1}(1-t^2) =(t^{-1}-t)t^n. \] Multiplying by \((1-t^2)^N\) gives \[ s_x(t)(t^2)^k(1-t^2)=T(t)s_y(t), \qquad T(t)=t^{-1}-t>0. \] A cover has \(k=0\); an edge of length gap three has \(k=1\). The scale compensates for either gap and leaves the same coefficient \(T(t)\).
Multiplying the raw inequality by \(s_x(t)>0\) therefore yields \[ D_x(t,F)\le D_x(t,F\cup\{e\})+T(t)D_y(t,H). \] The normalized edge-increment theorem also proves the equality criterion: cancel the nonzero scale to recover raw equality, then apply the embedded-image criterion above.
A complete one-edge state calculation
Take \(A=\mathbb R[\alpha]\), \(x=1\), \(y=b=s\), so \(N=1\), \(\ell(x)=0\), \(\ell(y)=1\). Initially the edge condition is still present. The kernels are \(\alpha A\) at \(x\) and \(A\) at \(y\). Their normalized state is \[ D_{\mathrm{initial}} =\left((1-t^2)t^{-1}\frac{t^2}{1-t^2},\quad (1-t^2)\frac1{1-t^2}\right) =(t,1). \] After removing the edge condition, both kernels are entire stalks: \[ D_{\mathrm{final}}=(t^{-1},1). \] The elementary transfer adds \(T\) times the upper coordinate to the lower one: \[ U_e(a,b)=(a+Tb,b). \] It sends \((t,1)\) to \((t+T,1)=(t^{-1},1)\). At \(t=1/2\), this is \[ (1/2,1)\longmapsto(2,1),\qquad T=3/2. \]
The calculation is exact in this example. In the general argument we first have only \[ D_{\mathrm{next}}\le U_eD_{\mathrm{current}} \] coordinatewise. The possible gap measures the failure of the lower image to fill the corrected upper image.
Remaining conditions and the ordered transfer bound
Let \(P=[e_1,\ldots,e_m]\) be the chronological edge list. After its first \(i\) edges have been removed, the remaining list is the suffix \(p_i=[e_{i+1},\ldots,e_m]\). Thus \(p_0=P\) and \(p_m=[]\). At vertex \(x\), the outgoing conditions are \[ F_{p_i}(x)=\{e:s(e)=x,\ e\text{ occurs in }p_i\}. \] The schedule state \(D_{p_i}(t)\) uses the Hilbert value of exactly this selected kernel. At a removal \(e_i:x\to y\), the source selection loses its current edge, while every other selection stays unchanged. In particular the upper coordinate at \(y\) has the same value before and after removal.
Write \(U_e(T)\) for the elementary transfer, with \(T=t^{-1}-t>0\). At the changed coordinate the one-edge Hilbert comparison gives the inequality, and at the other coordinates the states are equal. Hence \[ D_{p_i}(t)\le U_{e_i}(T)D_{p_{i-1}}(t). \] Each transfer preserves coordinatewise order. For the first two steps this gives \[ D_{p_2}\le U_{e_2}D_{p_1} \le U_{e_2}U_{e_1}D_{p_0}. \] The full transfer inequality repeats this induction with every suffix hypothesis explicit, obtaining \[ D_{[]}(t)\le U_{e_m}(T)\cdots U_{e_1}(T)D_P(t) =\mathcal D_{P^{\mathrm{rev}}}(T)D_P(t). \] The reversal comes from the definition of dispatch: its first edge supplies the outermost operator.
Upper suffix kernels supply the inclusions
For the schedule transported by \(\phi:[u,b]\simeq[u',b']\), the required hypothesis is freeness of \[ K_{r(e)}(F_q(r(e))) \quad\text{for every suffix }e::q\preceq_{\mathrm{suf}}P. \] The chronological suffix identification matches these conditions with the ambient strict-later selections at both endpoints. Upper-interval restriction identifies the selected modules inside the same actual stalks. The free upper module therefore supplies precisely the freeness hypothesis of the global edge inclusion.
The comparison under upper-kernel freeness performs this substitution at every suffix and then applies the full transfer inequality. This is the step where the geometric image theorem enters the numerical process.
Actual-character endpoints of the schedule
For the actual source boundary-sheaf character \(C_{x,b}\), define \[ c_x(t)=t^{\ell(x)-\ell(b)}C_{x,b}(t^2). \] The empty schedule selects the whole stalk. Its Hilbert denominator \((1-t^2)^N\) cancels against the state scale, giving \(D_{[]}(t)=c(t)\). The full schedule selects the full outgoing kernel. The normalized endpoint identities use homogeneous section duality to give its Hilbert value \[ \frac{(t^2)^{\ell(b)-\ell(x)}C_{x,b}(t^{-2})}{(1-t^2)^N}. \] Multiplication by the same scale leaves \(t^{\ell(b)-\ell(x)}C_{x,b}(t^{-2})=c_x(t^{-1})\). Hence \(D_P(t)=c(t^{-1})\).
Substituting these vectors into the transfer bound gives \[ c(t)\le\mathcal D_{P^{\mathrm{rev}}}(t^{-1}-t)c(t^{-1}). \] The geometric path identity identifies this ordered product with the target normalized R-matrix. That identification yields the matrix inequality used below; both endpoint vectors are constructed from the actual source sheaf.
Endpoint equality and vanishing intermediate deficits
Suppose a later argument establishes equality of the two endpoints. Let \(X_j\) be the ideal state obtained by applying transfers exactly, and \(D_j\) the actual state. Initially they agree. Put \(E_j=X_j-D_j\).
At the next step let \[ \eta_j=U_jD_j-D_{j+1}\ge0. \] Then \[ E_{j+1}=U_jE_j+\eta_j. \] Every coordinate of \(E_j\) is nonnegative. A transfer retains those coordinates and adds a nonnegative multiple of one to another. Therefore a positive deficit cannot disappear in a later step. If \(E_{\mathrm{final}}=0\), every \(\eta_j\) must have been zero.
The no-deficit theorem proves the corresponding general statement using monotonicity and injectivity. The displayed calculation explains its mechanism for the actual elementary transfers.
Equality of numbers recovers equality of submodules
At one step, zero deficit means \[ H_{I_e}(q)=H_{f_eJ_e}(q) \] for the actual homogeneous inclusion \(I_e\subseteq f_eJ_e\). In each degree, the smaller space has dimension at most the larger. If the inclusion were strict, homogeneity would give some degree with a positive quotient dimension. Since \(q^n>0\), that degree would contribute a strictly positive amount to the difference of convergent Hilbert values.
Thus equality of values forces \(I_e=f_eJ_e\). For comparison, the proper inclusion \((\beta,\gamma)\subset\mathbb R[\beta,\gamma]\) from localization and edge-image comparison has Hilbert deficit \(1\), despite equal rank and equal height-one localizations. This is the information the Hilbert step retains.
Recovering freeness by reversing the schedule
At the end of the forward process, no conditions remain and the kernels are the original free stalks. Traverse the edges backward, adding one condition at a time. Assume the current lower and upper kernels \(M,N\) are free.
Image equality gives surjections \[ f:M\to I,\qquad g:N\to I \] onto the same embedded image. The second map includes the regular correction factor. The upper quotient rule and saturation identify its kernel with \(\alpha N\). Multiplication by \(\alpha\) is injective and identifies \(N\) with \(\ker g\), so this kernel is projective.
Projectivity supplies lifts \(a:M\to N\), \(b:N\to M\) with \(ga=f\), \(fb=g\). The source constructs \[ i:\ker f\to M\oplus\ker g,\quad i(x)=(x,a(x)), \] \[ r:M\oplus\ker g\to\ker f,\quad r(m,k)=m-ba(m)+b(k). \] The relation \(fr(m,k)=f(m)-ga(m)+g(k)=0\) verifies the codomain, and \(ri(x)=x\). Thus \(\ker f\) is a direct summand of a projective module and is projective.
The graded projective-to-free theorem applies with the polynomial coefficient ring, finite generation and homogeneous nonnegative pieces. It makes the new kernel free. Repeating the step proves all the suffix kernels free. The image-equality theorem constructs the common-image maps and checks these hypotheses; the two-presentation theorem gives the explicit retraction.
Inputs supplied by induction
The upper-freeness hypothesis supplies comparison, and exact endpoints recover free suffix kernels. The simultaneous invariance induction supplies the upper freeness from smaller intervals and proves the R-matrix equality that makes the endpoints exact.
The order is therefore essential: \[ \text{image inclusion}\ \Longrightarrow\ \text{Hilbert inequality} \ \Longrightarrow\ \text{exact endpoints} \ \Longrightarrow\ \text{image equality} \ \Longrightarrow\ \text{projectivity and freeness}. \] The simultaneous induction supplies the endpoint equality in the middle of this sequence.
Simultaneous induction on interval rank
The first induction proves R-invariance together with the suffix-kernel freeness needed in later comparisons. A second induction proves equality of the canonical KL polynomials.
We compare \([a,b]\) and \([a',b']\) through an order isomorphism \(\phi\). Their common rank is \[ d=\ell(b)-\ell(a)=\ell(b')-\ell(a'). \] Coxeter combinatorics proved equality of these rank differences. For now assume that the simple-generator sets are finite, so the real realizations and module constructions are available.
The simultaneous induction assertions
For each interval isomorphism, prove simultaneously:
- \(\widetilde R_{a,b}=\widetilde R'_{a',b'}\).
- Every suffix kernel for the schedule transported by \(\phi\) is free.
- Every suffix kernel for the schedule transported by \(\phi^{-1}\) is free.
The two freeness assertions matter because we will run the comparison in both directions.
At \(d=0\), each interval has one vertex. The R-polynomial is the diagonal value \(1\). There are no edge conditions, and the relevant kernels are free stalks. This proves all three assertions in the base case.
Now suppose the assertions are known for every smaller rank. The proof must obtain two different kinds of information from that hypothesis.
Smaller intervals supplying the induction inputs
If \(z<b\), the initial interval \([a,z]\) has smaller rank. The restriction of \(\phi\) identifies it with \([a',\phi(z)]\). Induction supplies \[ \widetilde R_{a,z}=\widetilde R'_{a',\phi(z)}. \] These are the already-known entries in the bottom matrix row.
At a current edge \(x\to y\), we have \(y>a\). Hence the upper interval \([y,b]\) also has smaller rank. Induction supplies its suffix-kernel freeness, which must be matched with the upper kernel required at that step.
Restriction to \([y,b]\) retains every outgoing edge at \(y\). The transported scores are unchanged. The remaining outgoing edges are exactly the ones past the current score, and hence form the appropriate suffix conditions in the restricted schedule. An outgoing edge at \(y\) cannot have the same reflection label as the incoming edge from \(x\): that reflection sends \(y\) back to \(x\), not upward.
The selected-kernel identification transports the free basis supplied by the smaller upper interval to the required upper module.
Cancellation of the known matrix entries
Fix \(0<t<1\), and put \(T=t^{-1}-t\). Let \(c(t)\) be the normalized actual boundary-character vector of the source system. It is the same vector in both formulas below.
Actual-character reciprocity gives the exact source identity \[ c_a(t)=\sum_{a\le z\le b}\widetilde R_{a,z}(T)c_z(t^{-1}). \] The upper freeness from Step 2 permits the transported comparison from Hilbert comparison: \[ c_a(t)\le \sum_{a\le z\le b}\widetilde R'_{a',\phi(z)}(T)c_z(t^{-1}). \] The transported comparison applies the target R-matrix to the same actual source-character vector used by source reciprocity.
For every \(z<b\), the two R-entries agree by induction. Let \(S(t)\) denote the sum of all those common terms. The normalized top character is \(c_b(t^{-1})=1\). The two formulas reduce to \[ S(t)+\widetilde R_{a,b}(T) \le S(t)+\widetilde R'_{a',b'}(T). \] Subtract \(S(t)\), obtaining the desired one-sided inequality for the new entry.
Polynomial equality from the two comparisons
Apply the same reasoning to \(\phi^{-1}\), now using the target's own actual-character vector and the reverse freeness hypotheses. It gives \[ \widetilde R'_{a',b'}(T)\le\widetilde R_{a,b}(T). \] Therefore the two evaluations agree for every \(T=t^{-1}-t\), \(0<t<1\).
There are infinitely many such values: if \(0<s<t<1\), then \[ (s^{-1}-s)-(t^{-1}-t) =(t-s)\left(\frac1{st}+1\right)>0. \] A nonzero polynomial over the reals has only finitely many roots, so the difference polynomial is zero. Injectivity of the integer-to-real coefficient map gives equality of the original integer polynomials.
The strict half-degree bound enters the subsequent canonical P-induction.
Recovery of freeness at the current rank
Every proper subinterval already has R-invariance, and Step 4 gives the one new full-interval entry. Thus the entire R-matrices on the two intervals agree. Noncomparable entries vanish in both matrices.
Recall the endpoints of the state process: \[ D_{\mathrm{initial}}=c(t^{-1}),\qquad D_{\mathrm{final}}=c(t). \] The target transfer product is now the source R-matrix. Actual-character reciprocity therefore makes the product applied to the initial state equal the final state exactly.
Apply Hilbert comparison in its established order. Exact endpoints remove every numerical deficit. Hilbert detection turns the inclusions of images into equalities. Reverse traversal starts at the free stalks and proves all suffix kernels free. Repeat in the inverse direction.
This proves both new freeness assertions after proving the new R-entry. The induction closes because each freeness hypothesis used before Step 4 came from a strictly smaller interval.
A separate induction for canonical KL polynomials
The R-stage is complete. We now use the canonical family \(P\), constructed in Hecke algebra and polynomial invariants with its strict degree bound.
Induct on the interval rank with fixed top endpoints. The diagonal polynomials are both \(1\). For the current lower endpoint \(a\), induction gives equality of \(P_{z,b}\) and \(P'_{\phi(z),b'}\) whenever \(z>a\).
Define the normalized canonical vectors \[ k_z(t)=t^{-d(z,b)}P_{z,b}(t^2),\qquad k'_z(t)=t^{-d(z,b)}P'_{\phi(z),b'}(t^2). \] Rank preservation justifies using the same exponent. Both vectors satisfy reciprocity with the now common R-matrix: \[ k(t)=\widetilde R(T)k(t^{-1}),\qquad k'(t)=\widetilde R(T)k'(t^{-1}). \]
Subtract their bottom equations. Every term indexed by \(z>a\) vanishes by induction. The diagonal coefficient at \(a\) is \(1\). Thus, writing \(\Delta(t)=k_a(t)-k'_a(t)\), we obtain \[ \Delta(t)=\Delta(t^{-1}). \]
Vanishing of the reciprocal difference below the middle degree
Put \(p=P_{a,b}-P'_{a',b'}\) and \(d=d(a,b)>0\). Undoing the normalization gives \[ t^{-d}p(t^2)=t^dp(t^{-2}). \] Multiply by \(t^d\) and set \(q=t^2\). Then \[ p(q)=q^dp(q^{-1})\qquad(0<q<1). \]
Suppose \(p\ne0\) and let its highest degree be \(m\). Both original polynomials have degree below \(d/2\), so \(m<d/2\). Every exponent in \(p(q)\) is at most \(m\), while every exponent in \(q^dp(q^{-1})\) is at least \(d-m>m\). Their nonzero terms occupy disjoint degree ranges. They cannot be the same polynomial.
The identity on infinitely many \(q\) makes it a polynomial identity, so this contradiction proves \(p=0\). This establishes canonical KL equality in finite-generator systems from the common R-matrix and the canonical degree bound.
Transport from finite-generator models
Choose a reduced word for each top endpoint \(b,b'\). Each word uses finitely many simple generators. By the subword criterion, every element of the corresponding lower ideal uses only those generators.
The standard parabolic construction supplies finite-generator Coxeter systems and inclusions into the original groups. Its transport theorem preserves lengths, Bruhat comparisons, and the canonical polynomial values. Hence it gives interval isomorphisms from finite-support models to the original intervals.
Transport \(\phi\) into these models, apply the finite-generator theorem, and transport the polynomial equality back. The groups in the models may still be infinite; it is the generating sets that have become finite.
The two inductions and their inputs
The first induction proves R-invariance together with freeness in both directions. Its decisive operation is cancelling known R-row entries against one actual-character vector, then using the inverse interval isomorphism.
The second induction compares canonical KL vectors using the common R-matrix. Its decisive operation is eliminating a reciprocal polynomial difference by the strict degree bound.
The source-comparison entries relate this R-first route in Lean to the P-first presentation in the OpenAI notes. The upper freeness used before a new R-step comes from smaller intervals; the new R-equality then gives freeness at the current rank.
Definitions and proofs by construction
A single edge comparison
- The exact sequence for adding one outgoing condition
- Kernel Hilbert values and normalized states
- The Hilbert increment from an embedded edge-image inclusion
- The edge degree gives a uniform positive transfer coefficient
- The normalized single-edge inequality and its equality criterion
A complete schedule
- Remaining conditions and the Hilbert-state vector
- From one-edge bounds to the full transfer inequality
- The complete schedule comparison under upper-kernel freeness
- The transported comparison for actual characters
Endpoint identities
Equality and reverse reconstruction of freeness
- Endpoint equality forces equality at every transfer step
- Recovering a projective kernel from scaled image equality
- Recovering terminal-kernel freeness by reverse traversal
Simultaneous induction for R and suffix kernels
- Freeness of all terminal kernels in a transported schedule
- Extracting the bottom R-tilde row from character comparison
- The R-tilde induction step from smaller intervals
- Simultaneous R-tilde invariance and terminal-kernel freeness