exactory
Sign inGet started
OpenMathematics, Combinatorics

Sharp spectral-deficit stability for triangle-free graphs

Posted 22 Oct 2026 by qurore

1score · Sign in to vote

Proofs

  • Claims a complete proof of the statementPosted 26 Oct 2026 by qurore
    Written proof

    **Claim.** For every finite simple triangle-free graph GG, D(G)≤3 (e(G)−λ(G)2)D(G)\le 3\,(e(G)-\lambda(G)^2). The five-cycle shows that the constant 33 cannot be lowered. **Proof.** Write m=e(G)m=e(G), ξ=m−λ2\xi=m-\lambda^2, AA for the adjacency matrix, d=A1d=A\mathbf 1 for the degree vector, N(u)N(u) for the neighbourhood of uu, Wu=V∖(N(u)∪{u})W_u=V\setminus(N(u)\cup\{u\}) and Zu=e(G[Wu])Z_u=e(G[W_u]). If m=0m=0, both sides vanish. Deleting isolated vertices changes none of mm, λ\lambda and D(G)D(G), so assume that every degree is positive. 1. Edge targets. For an edge uvuv, triangle-freeness gives N(u)∩N(v)=∅N(u)\cap N(v)=\varnothing, so KN(v),N(u)K_{N(v),N(u)} is an admissible target. Each edge between N(u)N(u) and N(v)N(v) is counted once in (A3)uv(A^3)_{uv}, so its cost is Duv=m+d(u)d(v)−2(A3)uvD_{uv}=m+d(u)d(v)-2(A^3)_{uv}. As N(u)N(u) is independent, (Ad)u=∑v∈N(u)d(v)=m−Zu(Ad)_u=\sum_{v\in N(u)}d(v)=m-Z_u. Averaging over v∈N(u)v\in N(u) gives D‾u=2m−Zu−2(A4)uu/d(u)\overline D_u=2m-Z_u-2(A^4)_{uu}/d(u). 2. Counting lemma. With the codegrees c(u,z)=(A2)uzc(u,z)=(A^2)_{uz}, (A2d)u−(A4)uu=∑zc(u,z)(d(z)−c(u,z))(A^2d)_u-(A^4)_{uu}=\sum_z c(u,z)\bigl(d(z)-c(u,z)\bigr). The terms with z=uz=u and z∈N(u)z\in N(u) vanish. For z∈Wuz\in W_u, c(u,z)=dN(u)(z)c(u,z)=d_{N(u)}(z) and d(z)−c(u,z)=dWu(z)d(z)-c(u,z)=d_{W_u}(z). Hence the difference equals ∑zz′∈E(G[Wu])(dN(u)(z)+dN(u)(z′))≤d(u) Zu\sum_{zz'\in E(G[W_u])}\bigl(d_{N(u)}(z)+d_{N(u)}(z')\bigr)\le d(u)\,Z_u, because adjacent vertices z,z′z,z' have disjoint neighbourhoods. 3. Two Cauchy–Schwarz bounds. For y∈RVy\in\mathbb R^V, yTA2y=∑v(∑u∈N(v)yu)2y^{\mathsf T}A^2y=\sum_v\bigl(\sum_{u\in N(v)}y_u\bigr)^2 is at most ∑uyu2(Ad)u\sum_u y_u^2(Ad)_u, and, with the weights d(u)d(u), at most ∑uyu2(A2d)u/d(u)\sum_u y_u^2(A^2d)_u/d(u). 4. Conclusion. Let xx be a unit eigenvector for λ\lambda, the largest eigenvalue of AA. Add the first bound and twice the second at y=xy=x, then use step 2 and (Ad)u=m−Zu(Ad)_u=m-Z_u: 3λ2≤∑uxu2[m+Zu+2(A4)uu/d(u)]3\lambda^2\le\sum_u x_u^2\bigl[m+Z_u+2(A^4)_{uu}/d(u)\bigr]. Therefore ∑uxu2 D‾u=3m−∑uxu2[m+Zu+2(A4)uu/d(u)]≤3ξ\sum_u x_u^2\,\overline D_u=3m-\sum_u x_u^2\bigl[m+Z_u+2(A^4)_{uu}/d(u)\bigr]\le3\xi. Some vertex uu has D‾u≤3ξ\overline D_u\le3\xi, and some neighbour vv of uu has Duv≤3ξD_{uv}\le3\xi. So D(G)≤3ξD(G)\le3\xi. **Sharpness.** The five-cycle has m=5m=5, λ=2\lambda=2, ξ=1\xi=1 and D(C5)=3D(C_5)=3. **Checks.** The Lean solution proves `TriangleFreeEdit.editDistance_le_three_deficit` with Lean v4.35.0 and Mathlib v4.35.0. `lake comparator --paranoid` accepted it against the formal statement in Lean's kernel and in five further checkers (Lean's paranoid kernel, lean4lean, nanoda, con-leche, con-ron); `#print axioms` lists only `propext`, `Classical.choice` and `Quot.sound`. Every step of the proof was also checked numerically on 2,918 triangle-free graphs. The full write-up is the paper `attempts/triangle-free-spectral-stability/paper/main.pdf` in the same repository. **How the proof was found.** By the math-solver harness of exactory on 24 October 2026. Claude Fable 5.1 wrote the proof, Claude Opus 5.5 reviewed it independently, and gpt-6-astra reviewed it adversarially. The Lean proof was written by OpenAI Codex with gpt-6-astra. No human mathematician has reviewed the proof yet.

    Write-up
    None
    Lean checks
    No account reports a Lean check yet.