OctoInformBeta

OpenAI preprint, September 2026 · §4. Deletion operators and injectivity of the leading projection

Lemma 4.2 [The edge table and exact allocation]: Weight bounds for complete deletion and the characteristic-two cut of the pentagon rows

Deletion weights control scalar tests.

Hover over a blue or green term in the text, or a node in the graph, to find it in the other. Select it to read its declaration both in Lean and in words.

The Setting

Deleting a letter with a coefficient determined by its neighbors gives scalar tests on associative words in the pentagon substitutions. The coefficients come from integral hyperplane weights and four deletion kernels over 𝔽2\mathbb F_2.

The deletion alphabet is 𝒟={o,e,p,w}\mathscr D=\{o,e,p,w\}. The type B3B_3 hyperplane family ℋ\mathscr H consists of the coordinate labels a1,a2,a3a_1,a_2,a_3 and the diagonal labels p12,m12,p13,m13,p23,m23p_{12},m_{12},p_{13},m_{13},p_{23},m_{23}.✦ Write H*H^* for the coordinate covectors of the free ℤ\mathbb Z-module on ℋ\mathscr H. For a hyperplane label H∈ℋH\in\mathscr H and deletion letters a,b∈𝒟a,b\in\mathscr D, the integral deletion weights νH(a,b)\nu_H(a,b) are symmetric integer edge weights, with coefficient vector 𝝂(a,b)=∑H∈ℋνH(a,b)H*\boldsymbol\nu(a,b)=\sum_{H\in\mathscr H}\nu_H(a,b)H^* specified in (12).

The target alphabet E=ℱ⊔ℬE=\mathscr F\sqcup\mathscr B is the disjoint union of the fiber alphabet ℱ={f0,…,f4}\mathscr F=\{f_0,\ldots,f_4\} and the base alphabet ℬ={b0,…,b3}\mathscr B=\{b_0,\ldots,b_3\}.✦ Each target letter j∈Ej\in E has a list Λ(j)\Lambda(j) of hyperplane labels, defined by (1). (Λ(b0),Λ(b1),Λ(b2),Λ(b3))=([a1],[a2],[p12],[p12,m12]),(Λ(f0),Λ(f1),Λ(f2),Λ(f3),Λ(f4))=([a3],[p13],[p23],[p13,m13],[p23,m23]).\begin{aligned} (\Lambda(b_0),\Lambda(b_1),\Lambda(b_2),\Lambda(b_3)) &=([a_1],[a_2],[p_{12}],[p_{12},m_{12}]),\\ (\Lambda(f_0),\Lambda(f_1),\Lambda(f_2),\Lambda(f_3),\Lambda(f_4)) &=([a_3],[p_{13}],[p_{23}],[p_{13},m_{13}],[p_{23},m_{23}]). \end{aligned}(1) The target deletion weight νj(a,b)∈𝔽2\nu_j(a,b)\in\mathbb F_2 is the sum of the integral weights on this list, reduced modulo two, as in (2); the row edge weights εj(a,b)\varepsilon_j(a,b) use the same deletion letters in (13). νj(a,b)=∑H∈Λ(j)νH(a,b)¯.\nu_j(a,b)=\sum_{H\in\Lambda(j)}\overline{\nu_H(a,b)}.(2)

Weighted Deletion and Word Weights

For a commutative ring RR and an alphabet α\alpha, the free associative algebra 𝒜R(α)=R⟨α⟩\mathcal A_R(\alpha)=R\langle\alpha\rangle is the monoid algebra of the words on α\alpha.✦ Write 𝒔\mathbf s for the basis monomial associated with a word ss, and 11 for the empty word.✦ For an RR-algebra TT, associative evaluation is the algebra homomorphism ev⁡f:𝒜R(α)→T\operatorname{ev}_f:\mathcal A_R(\alpha)\to T extending an assignment f:α→Tf:\alpha\to T.✦

Let q:α→ℕq:\alpha\to\mathbb N be a letter weight. For a word ss on α\alpha, its word weight dq(s)d_q(s) is the nonnegative integer given by the sum in (3). For a predicate PP on words, the prescribed word span 𝒜R(P)\mathcal A_R(P) is the RR-span of their monomials satisfying PP.✦ For a nonnegative integer nn, the homogeneous component 𝒜R(α)nq\mathcal A_R(\alpha)^q_n and the strictly lower weight space 𝒜R(α)<nq\mathcal A_R(\alpha)^q_{<n} are the submodules of 𝒜R(α)\mathcal A_R(\alpha) spanned respectively by monomials of weight nn and of weight less than nn, as in (3). dq(a1⋯am)=∑j=1mq(aj),𝒜R(P)=span⁡R{𝒔:P(s)},𝒜R(α)nq=𝒜R(dq=n),𝒜R(α)<nq=𝒜R(dq<n).\begin{aligned} d_q(a_1\cdots a_m)&=\sum_{j=1}^m q(a_j),& \mathcal A_R(P)&=\operatorname{span}_R\{\mathbf s:P(s)\},\\ \mathcal A_R(\alpha)^q_n&=\mathcal A_R(d_q=n),& \mathcal A_R(\alpha)^q_{<n}&=\mathcal A_R(d_q<n). \end{aligned}(3)

For deletion labels β\beta, let κ(b,l,a,t)∈R\kappa(b,l,a,t)\in R be a coefficient for a label bb and three letters of α\alpha. For a deletion label b∈βb\in\beta and boundary letters l,r∈αl,r\in\alpha, the weighted deletion operator Dκ,bl,rD_{\kappa,b}^{l,r} is the RR-linear endomorphism of 𝒜R(α)\mathcal A_R(\alpha) given on basis monomials by the sum in (4). Here a0=la_0=l and am+1=ra_{m+1}=r, and the empty word has image zero. Dκ,bl,r(𝒂𝟏⋯𝒂𝒎)=∑j=1mκ(b,aj−1,aj,aj+1)𝒂𝟏⋯𝒂𝒋ˆ⋯𝒂𝒎.D_{\kappa,b}^{l,r}(\mathbf{a_1\cdots a_m}) =\sum_{j=1}^m\kappa(b,a_{j-1},a_j,a_{j+1}) \mathbf{a_1\cdots\widehat{a_j}\cdots a_m}.(4) The weighted deletion representation ρκl,r\rho_\kappa^{l,r} is the algebra homomorphism from 𝒜R(β)\mathcal A_R(\beta) to End⁡R(𝒜R(α))\operatorname{End}_R(\mathcal A_R(\alpha)) that sends each deletion label bb to Dκ,bl,rD_{\kappa,b}^{l,r}, as in (5). For a word ss on α\alpha, the complete deletion coefficient Sκl,r(s)S_\kappa^{l,r}(s) is the linear functional on 𝒜R(β)\mathcal A_R(\beta) defined there by extracting the coefficient [1][1] of the empty word. ρκl,r:𝒜R(β)→End⁡R(𝒜R(α)),ρκl,r(𝒃)=Dκ,bl,r,Sκl,r(s)(p)=[1](ρκl,r(p)(𝒔)).\begin{aligned} \rho_\kappa^{l,r}:\mathcal A_R(\beta) &\longrightarrow\operatorname{End}_R(\mathcal A_R(\alpha)),& \rho_\kappa^{l,r}(\mathbf b)&=D_{\kappa,b}^{l,r},\\ S_\kappa^{l,r}(s)(p)&=[1]\bigl(\rho_\kappa^{l,r}(p)(\mathbf s)\bigr). \end{aligned}(5) For costs c:β→ℕc:\beta\to\mathbb N, the weight-matched kernel κ[q,c]\kappa^{[q,c]} is the function that retains precisely the coefficients whose deleted letter has weight equal to the label cost, as defined in (6). κ[q,c](b,l,a,t)={κ(b,l,a,t),q(a)=c(b),0,q(a)≠c(b).\kappa^{[q,c]}(b,l,a,t)= \begin{cases} \kappa(b,l,a,t),&q(a)=c(b),\\ 0,&q(a)\ne c(b). \end{cases}(6)✦

The Pentagon Row Kernels

The substitution alphabet is Σ={A,C,B}\Sigma=\{A,C,B\}, with A=x2A=x^2, C=[x,y]C=[x,y], B=yB=y for free associative letters x,yx,y in characteristic two, of weights 2,2,12,2,1, respectively. The counting weights ωo\omega_o on 𝒟\mathscr D and ωB\omega_B on Σ\Sigma are the indicator functions in (7), where 𝟏P\mathbf1_P denotes the indicator of a condition PP. The resulting BB-count differs from ordinary yy-depth: CC has depth one and BB-count zero. ωo(a)=𝟏a=o,ωB(k)=𝟏k=B.\omega_o(a)=\mathbf1_{a=o},\qquad \omega_B(k)=\mathbf1_{k=B}.(7)

The four pentagon rows are indexed by i∈{1,2,3,4}i\in\{1,2,3,4\}.✦ For a substitution letter k∈Σk\in\Sigma, the row substitution list Si(k)S_i(k) consists of target letters in EE and is defined in slot order (A,C,B)(A,C,B) by (8). iSi(A)Si(C)Si(B)1[b0][b3,f3][b2,f1]2[b1,b3][f4][f2]3[b0,b1,b3][f3,f4][f1,f2]4[b0][b3][b2]\begin{array}{c|ccc} i&S_i(A)&S_i(C)&S_i(B)\\ \hline 1&[b_0]&[b_3,f_3]&[b_2,f_1]\\ 2&[b_1,b_3]&[f_4]&[f_2]\\ 3&[b_0,b_1,b_3]&[f_3,f_4]&[f_1,f_2]\\ 4&[b_0]&[b_3]&[b_2] \end{array}(8) The slot edge weight gi(k;a,b)∈𝔽2g_i(k;a,b)\in\mathbb F_2 is the sum of row edge weights along Si(k)S_i(k).✦ For a row ii, a substitution letter k∈Σk\in\Sigma, and deletion letters l,a,b∈𝒟l,a,b\in\mathscr D, the raw deletion kernel κ0i(k;l,a,b)\kappa_0^i(k;l,a,b) is the difference in 𝔽2\mathbb F_2 of the slot edge weights at the two edges meeting the deleted letter aa, as in (9). gi(k;a,b)=∑j∈Si(k)εj(a,b),κ0i(k;l,a,b)=gi(k;a,b)−gi(k;l,a).\begin{aligned} g_i(k;a,b)&=\sum_{j\in S_i(k)}\varepsilon_j(a,b),\\ \kappa_0^i(k;l,a,b)&=g_i(k;a,b)-g_i(k;l,a). \end{aligned}(9)

The deletion weights λk(l,a,b)∈𝔽2\lambda_k(l,a,b)\in\mathbb F_2 of the second row are the functions in (10). λA(l,a,b)=𝟏a=e,b=w,λC(l,a,b)=𝟏a=p(𝟏l∈{e,w}+𝟏b∈{e,w}),λB(l,a,b)=𝟏a=o(𝟏l=p+𝟏b=p).\begin{aligned} \lambda_A(l,a,b)&=\mathbf1_{a=e,\ b=w},\\ \lambda_C(l,a,b)&=\mathbf1_{a=p} (\mathbf1_{l\in\{e,w\}}+\mathbf1_{b\in\{e,w\}}),\\ \lambda_B(l,a,b)&=\mathbf1_{a=o} (\mathbf1_{l=p}+\mathbf1_{b=p}). \end{aligned}(10) The four row deletion kernels κi(k;l,a,b)∈𝔽2\kappa_i(k;l,a,b)\in\mathbb F_2 are given by (11), with the arguments (l,a,b)(l,a,b) of λk\lambda_k suppressed and all sums taken in 𝔽2\mathbb F_2. iκi(A;l,a,b)κi(C;l,a,b)κi(B;l,a,b)1λAλA𝟏a=o,b=w2λAλC+𝟏a=e(𝟏l=p+𝟏b=p)λB30𝟏a=e(𝟏l=p+𝟏b=p)+𝟏a=p(𝟏l=e+𝟏b=e)λB4λA𝟏a∈{e,p},b=w𝟏a=o,b=w\begin{array}{c|ccc} i&\kappa_i(A;l,a,b)&\kappa_i(C;l,a,b)&\kappa_i(B;l,a,b)\\ \hline 1&\lambda_A&\lambda_A&\mathbf1_{a=o,\ b=w}\\ 2&\lambda_A&\lambda_C+\mathbf1_{a=e} (\mathbf1_{l=p}+\mathbf1_{b=p})&\lambda_B\\ 3&0&\begin{gathered} \mathbf1_{a=e}(\mathbf1_{l=p}+\mathbf1_{b=p})\\ {}+\mathbf1_{a=p}(\mathbf1_{l=e}+\mathbf1_{b=e}) \end{gathered}&\lambda_B\\ 4&\lambda_A&\mathbf1_{a\in\{e,p\},\ b=w}&\mathbf1_{a=o,\ b=w} \end{array}(11)✦

The Claim

Theorem 1. The integral deletion weights are given by (12), with the listed pairs unordered. 𝝂(a,b)={p23*−m23*,{a,b}={o,p},p12*−m12*,{a,b}={o,w},a3*−m23*,{a,b}={e,p},a1*−m12*,{a,b}={e,w},a2*+p13*−m12*−m23*,{a,b}={p,w},0,otherwise.\boldsymbol\nu(a,b)= \begin{cases} p_{23}^*-m_{23}^*,&\{a,b\}=\{o,p\},\\ p_{12}^*-m_{12}^*,&\{a,b\}=\{o,w\},\\ a_3^*-m_{23}^*,&\{a,b\}=\{e,p\},\\ a_1^*-m_{12}^*,&\{a,b\}=\{e,w\},\\ a_2^*+p_{13}^*-m_{12}^*-m_{23}^*,&\{a,b\}=\{p,w\},\\ 0,&\text{otherwise}. \end{cases}(12)✦ For j∈Ej\in E and a,b∈𝒟a,b\in\mathscr D, the row edge weight is (13). εj(a,b)=νj(a,b).\varepsilon_j(a,b)=\nu_j(a,b).(13)✦ For a row ii, a slot kk, and letters l,a,b∈𝒟l,a,b\in\mathscr D with l≠wl\ne w and a≠wa\ne w, the cut identity is (14). (κ0i)[ωo,ωB](k;l,a,b)=κi(k;l,a,b).\bigl(\kappa_0^i\bigr)^{[\omega_o,\omega_B]}(k;l,a,b) =\kappa_i(k;l,a,b).(14)✦

Let RR be a commutative ring, let α,β\alpha,\beta be alphabets, and give them weights q:α→ℕq:\alpha\to\mathbb N and costs c:β→ℕc:\beta\to\mathbb N. Suppose every nonzero coefficient κ(b,l,a,t)\kappa(b,l,a,t) satisfies q(a)≤c(b)q(a)\le c(b). For boundary letters l,rl,r, a word ss on α\alpha, and p∈𝒜R(β)<dq(s)cp\in\mathcal A_R(\beta)^c_{<d_q(s)}, the scalar vanishing statement is (15). Sκl,r(s)(p)=0.S_\kappa^{l,r}(s)(p)=0.(15)✦

Suppose instead that every nonzero coefficient κ(b,l,a,t)\kappa(b,l,a,t) satisfies c(b)≤q(a)c(b)\le q(a). For boundary letters l,rl,r, a word vv on β\beta, and p∈𝒜R(α)dc(v)qp\in\mathcal A_R(\alpha)^q_{d_c(v)}, the exact allocation statement is (16). [1](ρκl,r(𝒗)(p))=[1](ρκ[q,c]l,r(𝒗)(p)).[1]\bigl(\rho_\kappa^{l,r}(\mathbf v)(p)\bigr) =[1]\bigl(\rho_{\kappa^{[q,c]}}^{l,r}(\mathbf v)(p)\bigr).(16)✦

In the paper, the result reads: with ziz_i denoting the coordinates and tHt_H the generator associated with a hyperplane HH, in the integral basis ai=tzi=0a_i=t_{z_i=0}, uij=tzi−zj=0u_{ij}=t_{z_i-z_j=0}, vij=tzi−zj=0+tzi+zj=0v_{ij}=t_{z_i-z_j=0}+t_{z_i+z_j=0}, the edge weights are (17), with oo denoting the paper’s 00, edgea,v partu partop0u23*ow0u12*oe00epa3*−v23*0ewa1*−v12*0pwa2*+v13*−v12*−v23*u13*\begin{array}{c|cc} \text{edge}&a,v\text{ part}&u\text{ part}\\ \hline op&0&u_{23}^*\\ ow&0&u_{12}^*\\ oe&0&0\\ ep&a_3^*-v_{23}^*&0\\ ew&a_1^*-v_{12}^*&0\\ pw&a_2^*+v_{13}^*-v_{12}^*-v_{23}^*&u_{13}^* \end{array}(17) equal-letter edges have weight zero, and after reduction modulo two a word containing rr zeros has zero scalar test against every product with fewer than rr factors uiju_{ij}, while with exactly rr such factors every one must delete a zero in a nonzero contribution.

What Is Checked

The statements concern integral hyperplane weights and 𝔽2\mathbb F_2 row kernels with l≠wl\ne w and a≠wa\ne w, while the scalar assertions range over arbitrary commutative coefficient rings and alphabets under their respective inequalities on nonzero deletion weights.