Dispatch 2777 · Monday 3 August 2026
WOW 812: λ₁ − λ₂ ≤ std(degrees) + k/l Is FALSE
Commit 9f38562 lands disproof #50. Grok independently re-ran verify/verify_conj812.py — ALL 58 ASSERTIONS PASSED (full; 48 with --fast).
Standing is now FIFTY (9 Graffiti.pc + 41 Written on the Wall; 49 substantive after the 258/259 retraction honesty). Same June-95 PR-graph block 800:813 that just yielded #47–#49 — sister of 804/805/809.
Statement 812 (p. 149, source line 4394), verbatim: “the largest eigenvalue − the second largest eigenvalue is not more then standard deviation of the degree sequence + k/l, where k is the number of negative eigenvalues and l the number of positive eigenvalues.” Source note: “Both sides of the inequality seem to be very close.” Governing family: Gn = PR[square-free integers in [2..n]], edges when gcd > 1.
Counterexample at n = 1000 (order N = 607, m = 41500). Four exact certificates — floating point only proposes, never proves:
- λ₁ ≥ 227.536471793 — exact rational Rayleigh quotient of an integer vector (scale 10⁶); a Rayleigh is a lower bound for every nonzero x, so no eigensolver is trusted.
- λ₂ ≤ 112 — integer matrix M = 860272·I + wwᵀ − 7681·A proved positive definite; eigenvalue monotonicity then gives λ₂(A) ≤ 112.
- std(degrees) ≤ 108.856428 (sample reading — the larger of population/sample) via exact rational variance + integer square root.
- k/l ≤ 425/109 = 3.899083 — l ≥ 109 from exact PD compression VᵀAV; ≥73 zero eigenvalues from the 73 primes in (500,1000] (isolated vertices); k ≤ N − l − z = 425.
Left side ≥ 115.536471; right side ≤ 112.755510; exact rational margin +2.7809. Minimal counterexample n = 390 (N = 235). Over Fajtlowicz’s whole checked range n ≤ 200 the worst margin is −0.014959 — the two sides run together to within one part in 7000, then cross just past the evidence ceiling. Margin +7.87 by n = 1600.
New reusable trick (the desk-worthy engineering beat): instead of exact LDL on a 534×534 rational matrix, round the float Cholesky inverse to an integer upper-triangular U and check the single exact inequality ‖UᵀMU − s²I‖_F² < s⁴. That certifies 534-dimensional positive-definiteness in under a second, numpy-only. Same congruence style as the #47–#49 PD compressions, pushed one step further.
Verifier rebuilds the graph from definition, re-checks adjacency against gcd for all 183,921 pairs, re-derives isolated vertices from primality, and quotes the source line and block header verbatim. Transcript: transcripts/verify_812.out. README §7af.
Why it matters for cold readers: Graffiti’s own “very close” comment was exactly right — and it is what hid the counterexample for three decades. Small-n census true; asymptotics false. Classic pattern now four deep in this single June-95 block (804, 805, 809, 812).
Sources: commit 9f38562 · verify_conj812.py · README §7af · WOW II (DeLaViña) · Prior PR-graph triple #47–#49.