OctoInformBeta

OpenAI preprint, September 2026 · Section 6. Scalar tests and incidence averaging

Section 6 (proof of Theorem 1.2): signed pairing and incidence averaging: On the blocks MpM_p for integers p≥1p\geq1, norm at most 2p2p forces error at least one at an anchor

Finite anchor sets obstruct supported operators.

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

The question is approximation of evaluation vectors at a finite anchor set in a metric block by operators whose values have a common finite support. The scalar estimates concern signed coefficients on that support and exceptional sets indexed by graph incidences.

For a pointed metric space (M,d,o)(M,d,o), the space Lip⁡0(M)\operatorname{Lip}_0(M) consists of real Lipschitz functions vanishing at oo, with the Lipschitz norm.✦ Evaluation at xx is the bounded functional δx:f↦f(x)\delta_x:f\mapsto f(x).✦ The Lipschitz-free space ℱ(M)\mathcal F(M) over (M,d,o)(M,d,o) is the closed span of the evaluation functionals in Lip⁡0(M)*\operatorname{Lip}_0(M)^*, with canonical dual pairing given by (1). ℱ(M)=span¯{δx:x∈M}⊆Lip⁡0(M)*,⟨∑xaxδx,f⟩=∑xaxf(x).\mathcal F(M)=\overline{\operatorname{span}}\{\delta_x:x\in M\} \subseteq\operatorname{Lip}_0(M)^*, \qquad \left\langle\sum_x a_x\delta_x,f\right\rangle=\sum_x a_xf(x).(1) The Dirac embedding δ:M→ℱ(M)\delta:M\to\mathcal F(M) sends xx to its evaluation functional δx\delta_x, regarded as an element of the closed evaluation span.

The Metric Blocks

Fix an integer p≥1p\geq1. The scales, the number kk of generators and the number mm of colors are defined by (2): L=10p,R=3L,K=10p(R+1),D=K+2L+5,B=4R+D=Bp=300p2+150p+5,V=1+4pB,k=20V+2,Δ=2k,Q=Δ2L+1,H=ΔD+1,m=2ΔQH+1.\begin{aligned} L&=10p,& R&=3L,& K&=10p(R+1),& D&=K+2L+5,\\ B&=4R+D=B_p=300p^2+150p+5,& V&=1+4pB,\\ k&=20V+2,& \Delta&=2k,& Q&=\Delta^{2L+1},& H&=\Delta^{D+1},\\ m&=2\Delta QH+1. \end{aligned}(2) The block generator set is A={0,…,k−1}A=\{0,\ldots,k-1\}, and its color set is C={0,…,m−1}C=\{0,\ldots,m-1\}. The local radius LL is 10p10p, and the scale KK is 10p(R+1)10p(R+1). The label cap DD is K+2L+5K+2L+5, giving 4R+D4R+D for the scale B=BpB=B_p and 1+4pB1+4pB for the auxiliary bound VV. Twice the number kk of generators gives the graph degree bound Δ\Delta. The local ball bound QQ is Δ2L+1\Delta^{2L+1}, while the ball bound HH is ΔD+1\Delta^{D+1}.

Write 𝔽(A)\mathbb F(A) for the free group on AA, and |g||g| for reduced word length. For each nonnegative integer NN, the word ball WNW_N in 𝔽(A)\mathbb F(A) is the set of words of length at most NN, and the word ball permutation group ΓN\Gamma_N is the symmetric group of WNW_N, as in (3): WN={g∈𝔽(A):|g|≤N},ΓN=Sym⁡(WN).W_N=\{g\in\mathbb F(A):|g|\leq N\}, \qquad \Gamma_N=\operatorname{Sym}(W_N).(3) For a∈Aa\in A, an extended generator translation σN,a∈ΓN\sigma_{N,a}\in\Gamma_N is a permutation of WNW_N that agrees with left multiplication by aa at every g∈WNg\in W_N for which ag∈WNag\in W_N. Choose such translations and let πN:𝔽(A)→ΓN\pi_N:\mathbb F(A)\to\Gamma_N be the homomorphism sending each generator aa to σN,a\sigma_{N,a}.✦

A graph with reversible ports on a vertex set XX, with ports PP, is a move map G:X×P→XG:X\times P\to X and an involution r:P→Pr:P\to P such that G(G(x,e),r(e))=xG(G(x,e),r(e))=x.✦ Here P=A×{+,−}P=A\times\{+,-\}; the port reversal changes the sign, and the signed generator s(a,+)s_{(a,+)} is aa, while s(a,−)=a−1s_{(a,-)}=a^{-1}. The finite Cayley graph on ΓN\Gamma_N has move h↦hπN(se)h\mapsto h\pi_N(s_e).✦ Taking all colorings of each such graph and then their disjoint union defines the block vertices XpX_p, the block graph GpG_p and the block coloring cpc_p in (4): Xp=∐N≥0(CΓN×ΓN),Gp((N,c,h),e)=(N,c,hπN(se)),cp(N,c,h)=c(h).\begin{aligned} X_p&=\coprod_{N\geq0}\bigl(C^{\Gamma_N}\times\Gamma_N\bigr),\\ G_p((N,c,h),e)&=(N,c,h\pi_N(s_e)), &c_p(N,c,h)&=c(h). \end{aligned}(4)

A port walk ww is a finite list of ports followed in order, with length |w||w| and with the empty list ending at its starting vertex. A bounded color coordinate vector vv on the color set CC is an element of {0,…,D}C\{0,\ldots,D\}^{C}, equipped with the supremum norm.✦ For a vertex x∈Xpx\in X_p, the distance profile ℓ(x)∈{0,…,D}C\ell(x)\in\{0,\ldots,D\}^{C} is the vector of truncated distances to the colors defined by (5); Gp(x,w)G_p(x,w) denotes the endpoint of the port walk ww: ℓ(x)(i)=min⁡({D}∪{|w|:cp(Gp(x,w))=i})(x∈Xp,i∈C).\ell(x)(i)=\min\Bigl(\{D\}\cup \{|w|:c_p(G_p(x,w))=i\}\Bigr) \qquad (x\in X_p,\ i\in C).(5) A realized label anchor aa of the block graph GpG_p is a vector in ℓ(Xp)\ell(X_p).✦ The metric block MpM_p, with attachment scale RR, is the graph with label anchors obtained by adjoining these vectors to XpX_p. Let ι\iota be the anchor inclusion, and let ax=ℓ(x)a_x=\ell(x) be the vertex label anchor. The finite block anchor set 𝒜p\mathcal A_p and the block basepoint opo_p are specified in (6), where x*=(0,c0,1)x_*=(0,c_0,1) is the distinguished graph vertex and c0c_0 is the constant zero coloring: Mp=Xp⊔ℓ(Xp),𝒜p=ι(ℓ(Xp)),op=ι(ax*).M_p=X_p\sqcup\ell(X_p), \qquad \mathcal A_p=\iota(\ell(X_p)), \qquad o_p=\iota(a_{x_*}).(6)

Finite Support and Signed Coefficients

Let (M,d,o)(M,d,o) have d(x,y)≥1d(x,y)\geq1 for distinct points. The Lipschitz point indicator χj\chi_j at j∈Mj\in M is the 11-Lipschitz function equal to 11 at jj when j≠oj\ne o, and zero elsewhere.✦ The coefficient functional cjc_j at j∈Mj\in M on ℱ(M)\mathcal F(M) is the functional defined by the canonical dual pairing cj(u)=⟨u,χj⟩c_j(u)=\langle u,\chi_j\rangle.✦

For a finite subset S⊆MS\subseteq M containing oo, the finite support projection PSP_S on ℱ(M)\mathcal F(M) is the continuous linear operator defined in (7). The affine weights μu\mu_u of u∈ℱ(M)u\in\mathcal F(M) are the signed coefficients on SS defined there: PSu=∑j∈Scj(u)δj,μu(j)=cj(u)+𝟏{o}(j)(1−∑k∈Sck(u)).P_Su=\sum_{j\in S}c_j(u)\delta_j, \qquad \mu_u(j)=c_j(u)+\mathbf1_{\{o\}}(j) \left(1-\sum_{k\in S}c_k(u)\right).(7) For an operator TT on ℱ(M)\mathcal F(M), write μx=μTδx\mu_x=\mu_{T\delta_x}.

For signed coefficients μ\mu and a real function ff on a finite set SS, the pairing ⟨μ,f⟩S\langle\mu,f\rangle_S is the finite sum in (8). The total variation ∥μ∥TV\|\mu\|_{\mathrm{TV}} and the variation |μ|(E)|\mu|(E) on a subset E⊆SE\subseteq S are given by (8): ⟨μ,f⟩S=∑j∈Sμ(j)f(j),∥μ∥TV=∑j∈S|μ(j)|,|μ|(E)=∑j∈E|μ(j)|.\langle\mu,f\rangle_S=\sum_{j\in S}\mu(j)f(j), \qquad \|\mu\|_{\mathrm{TV}}=\sum_{j\in S}|\mu(j)|, \qquad |\mu|(E)=\sum_{j\in E}|\mu(j)|.(8)

The Claim

Theorem 1. Let SS be finite, let μ,ν,f,b,c:S→ℝ\mu,\nu,f,b,c:S\to\mathbb R, and let E,F⊆SE,F\subseteq S. Let L∈ℝL\in\mathbb R and suppose |f|≤L|f|\leq L, 0≤b,c≤L0\leq b,c\leq L, with f=bf=b off EE and f=−cf=-c off FF. The opposing pairing estimate is (9): ⟨μ,b⟩S+⟨ν,c⟩S−2L(|μ|(E)+|ν|(F))≤⟨μ,f⟩S−⟨ν,f⟩S.\langle\mu,b\rangle_S+\langle\nu,c\rangle_S -2L\bigl(|\mu|(E)+|\nu|(F)\bigr) \leq\langle\mu,f\rangle_S-\langle\nu,f\rangle_S.(9)✦

Let (M,d,o)(M,d,o) be as above, let S⊆MS\subseteq M be finite and contain oo, and let TT be a bounded linear operator on ℱ(M)\mathcal F(M) with PST=TP_ST=T and ∥T∥≤t\|T\|\leq t, where t∈ℝt\in\mathbb R. For any 11-Lipschitz real function gg on MM, the affine pairing bound (10) holds for all x,y∈Mx,y\in M: |⟨μx,g|S⟩S−⟨μy,g|S⟩S|≤td(x,y).\bigl|\langle\mu_x,g|_S\rangle_S-\langle\mu_y,g|_S\rangle_S\bigr| \leq t\,d(x,y).(10)✦

Let PP be finite and let (Ee)e∈P(E_e)_{e\in P} be subsets of a finite set SS such that each j∈Sj\in S belongs to at most one EeE_e. For any signed coefficients μ\mu on SS, the disjoint variation bound is (11): ∑e∈P|μ|(Ee)≤∥μ∥TV.\sum_{e\in P}|\mu|(E_e)\leq\|\mu\|_{\mathrm{TV}}.(11)✦

Let XX and PP be finite nonempty sets, let SS be finite, and set Δ=|P|\Delta=|P|. For every x∈Xx\in X, let μx\mu_x be signed coefficients on SS, with ∥μx∥TV≤V\|\mu_x\|_{\mathrm{TV}}\leq V, where V≥0V\geq0. Let ℬ⊆X\mathcal B\subseteq X be the set of bad vertices, and let Ex,e⊆SE_{x,e}\subseteq S be exceptional sets such that, at each x∉ℬx\notin\mathcal B, every j∈Sj\in S belongs to at most one Ex,eE_{x,e}. Suppose 20V<Δ20V<\Delta and 20V|ℬ|<|X|20V|\mathcal B|<|X|. For any bijection ρ:X×P→X×P\rho:X\times P\to X\times P and any real p>0p>0, put L=10pL=10p. There is an incidence i=(x,e)i=(x,e), with ρ(i)=(x′,e′)\rho(i)=(x',e'), satisfying the signed incidence obstruction (12): 3L2−2L(|μx|(Ex,e)+|μx′|(Ex′,e′))>2p.\frac{3L}{2}-2L\bigl(|\mu_x|(E_{x,e}) +|\mu_{x'}|(E_{x',e'})\bigr)>2p.(12)✦

For the block MpM_p at an integer p≥1p\geq1, let S⊆MpS\subseteq M_p be finite and contain opo_p. Let T:ℱ(Mp)→ℱ(Mp)T:\mathcal F(M_p)\to\mathcal F(M_p) be bounded and linear, with PST=TP_ST=T and ∥T∥≤2p\|T\|\leq2p. Some anchor a∈𝒜pa\in\mathcal A_p satisfies the finite support obstruction (13): ∥Tδa−δa∥≥1.\|T\delta_a-\delta_a\|\geq1.(13)✦

In the paper, the result reads: a supported operator of norm at most 2p2p and anchor error less than one gives opposing tests fef_e, for directed edges e=(x,y)e=(x,y) with reversals e‾\bar e, satisfying ⟨μx−μy,fe⟩S≥3L/2−2L(|μx|(Ee)+|μy|(Ee‾))\langle\mu_x-\mu_y,f_e\rangle_S\geq3L/2-2L(|\mu_x|(E_e)+|\mu_y|(E_{\bar e})) and |⟨μx−μy,fe⟩S|≤2p|\langle\mu_x-\mu_y,f_e\rangle_S|\leq2p, while total variation is at most VV, each good state is exceptional at most once, V/Δ<1/20V/\Delta<1/20, and the bad fraction is less than 1/(20V)1/(20V), so incidence averaging gives a contradiction for L=10pL=10p and the finite anchor obstruction follows.

What Is Checked

The claims are stated over finite signed coefficients and finite incidence sets, and over the block metric spaces with their anchor sets and the finite support projections on their Lipschitz-free spaces.

The block color count is positive.✦ The block metric satisfies the separation bound (14): d(x,y)≥1(x,y∈Mp,x≠y).d(x,y)\geq1\qquad(x,y\in M_p,\ x\ne y).(14)✦

For every radius NN and generator aa, there is a permutation of WNW_N agreeing with left multiplication by aa at every g∈WNg\in W_N for which ag∈WNag\in W_N.✦