REVIEW 4 major objections 5 minor 2 cited by
Strengthened upper bound on the third eigenvalue of graphs
T0 review · 4 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash
Pith's one-line read This paper proves that the supremum of the third eigenvalue ratio λ3/n over all n-vertex graphs is strictly smaller than the classical threshold 1/(2√2).
desk verdict A serious proof of a Nikiforov claim, with one load-bearing numerical gap that a referee can ask to be fixed. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The central object is Operation ∗: given a graph G and an orthonormal pair of eigenvectors x,y for λ_{n−1}(G), λ_n(G), define G* by i∼j iff x_i x_j + y_i y_j < 0. The spectral minimisation principle in Theorem 2.2 shows λ_{n−1}(G*)+λ_n(G*) ≤ λ_{n−1}(G)+λ_n(G), so a minimal graph can be assumed invariant, G=G*. Invariance gives a circular-arc structure (Theorem 2.4), clique number ≤3 and chromatic number ≤4. For the critical n/2-regular invariant case the adjacency matrix splits as [[Q, J−Q],[J−Q, Q]], and the spectrum of G reduces to the spectra of J and 2Q−J; the eigenvectors of 2Q−J are shown to come in two monotone Types (front-increasing-then-decreasing nonnegative, and always-decreasing), from which a series of boundary inequalities defines a feasible region in (a,c,ν)-space. The proof of Theorem 3.7 traces the minimum of ν1+ν2 across phases of these inequalities and rules out ν1+ν2=−√2.
What would settle it
Evaluate the polynomial P(ν1)=$ν1^{3}$+$ν1^{2}$+4(T−S)ν1+4(√T(T+S)−S) from Intersection IIa with interval arithmetic over the stated ranges T∈[1/9,1/8], S∈[T+7/400,1/4] and check that P(−0.7)>0; likewise verify P(−√2+2√T)>0.001 for the second polynomial at the prescribed S-endpoints over T∈[0.055,1/8]. A single counterexample to either inequality, or an explicit sequence of n/2-regular invariant graphs with (λ_{n−1}+λ_n)/n approaching −√2/2, would settle the claim either way.
Extended reading notes
Core claim
The central claim is Theorem 1.2: there exists ε>0 such that inf{ (λ_{n−1}(G)+λ_n(G))/n : |V(G)|=n≥3 } > −√2/2 + ε. Consequently, via the inequality λ3+λ_{n−1} ≤ λ2(K_n) = −1, the third eigenvalue ratio is bounded away from 1/(2√2). The proof eliminates the equality case of the elementary AM-QM bound λ_{n−1}^2 + $λ_n^{2}$ ≤ $n^{2}$/4: a sequence of graphs with the sum converging to −√2/2 would force a specific limit state — parameters X=Z=1/2, T=1/8 and eigenvector components a=c=√2/2 — and the paper shows that this state cannot be approached simultaneously by the two structural types of eigenvectors that the invariant graphs admit. The near-equality graphs are then shown to be close, up to o($n^{2}$) edge changes, to n/2-regular graphs invariant under the new operation, so the contradiction transfers to the general problem.
Load-bearing premise
The proof that the infimum is strictly above −√2/2 depends on two numerical inequalities in Stage 4 of Theorem 3.7 — that the cubic P is positive at −0.7 and that at −√2+2√T it is bounded below by 0.001 — which are asserted from direct computation rather than proved or machine-verified; if either sign were wrong, the contradiction forcing strict inequality would fail.
Editorial extensions
If this is right
- There is a constant ε3>0 such that every n-vertex graph satisfies λ3(G)/n < 1/(2√2) − ε3, settling the long-claimed strengthening for k=3.
- The infimum of (λ_{n−1}+λ_n)/n over all graphs of order n is bounded below by −√2/2 + ε, so the trivial AM-QM bound is not tight.
- The same argument gives c_{−2} < 1/(2√2) − ε3, a strengthened upper bound on the second smallest eigenvalue in absolute value.
- Conjecture 5.1, that λ_{n−1}+λ_n ≥ −2n/3, is verified for all graphs on at most 9 vertices and is compatible with the new bound; if true it would imply c3 = 1/3.
Reading between the lines
- The direct numerical checks in Stage 4 (positivity of P(−0.7) and the 0.001 lower bound for P(−√2+2√T)) could be replaced by rigorous interval arithmetic or a computer-certified proof, so those two evaluations are the first place to audit the argument.
- The same ∗_k generalisation in Subsection 2.2 may give analogous strengthened bounds for higher k, although the paper notes that for k≥3 there is no canonical ordering of the vectors, which currently blocks that route.
- The equality-state analysis suggests that any minimising sequence, if it existed, would concentrate on block-type constructions; one testable extension is to check whether explicit pivalous or circulant blow-up families achieve the −2/3 infimum in the limit rather than −√2/2.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper proves a strengthened upper bound for the third eigenvalue of a graph, namely that c3 is strictly below the classical bound 1/(2√2). The strategy is to study the closely related quantity λ_{n-1}+λ_n and to show a uniform spectral gap: inf(λ_{n-1}+λ_n)/n > −√2/2 + ε. The proof introduces a new graph operation G*, shows that minimising graphs can be assumed invariant under it, derives structural restrictions (ω ≤ 3, χ ≤ 4), then reduces a hypothetical worst-case sequence to n/2-regular invariant graphs. For those graphs the paper develops a lengthy finite-parameter extremal argument with two eigenvector types and rules out the limiting equality case, completing the proof of Theorems 3.1, 1.2, and 1.1.
Significance. If the argument is correct, this is a significant result: it resolves Nikiforov's omitted case k = 3 and gives the first rigorous proof of a uniform gap below 1/(2√2) for c3. The new operation G* is natural and the structural theorems (clique number ≤ 3, chromatic number ≤ 4, and the canonical front/middle/back decomposition) are elegant and appear to be correct. The proof is self-contained in the sense that no numerical constants are fitted and no external computational evidence is needed at the level of the main theorem. However, the paper is not yet a complete proof: several load-bearing inequalities, especially in Stage 4 of Theorem 3.7, are asserted on the basis of ad hoc Desmos computations rather than proved or certified, and the phase analysis in Stage 5 is similarly visual and informal. These gaps are local in the sense that they can likely be repaired with written-out polynomial bounds or interval arithmetic, but as submitted they prevent the central claim from being fully verified.
major comments (4)
- [§3.1, Stage 4 (Intersections IIa and IIb)] The exclusion of the equality state ν1+ν2 = −√2 depends on numerical assertions that are never proved or certified. In Intersection IIa the proof states 'Directly computing with Desmos, we get that as functions of T, P(−0.7) is always positive, contradiction', and in Intersection IIb it states that P(−√2+2√T) 'is bounded below by a positive number (0.001 suffices)' while the upper endpoint has 'the only root at T=1/8'. These are claims about cubic polynomials in parameters T and S, but the polynomials are not written out, no analytic proof is given, and no code or interval-arithmetic certificate is supplied. These checks are load-bearing: they are exactly what forces ν1 > −0.7 in IIa and what leaves only X=Z=1/2, T=1/8 in IIb. If either check is false or too coarse, the contradiction in Stage 4 fails and Theorems 3.7, 3.1, and 1.2 are unsupported. This is the central gap and must be closed by a rigorous proof or a machine-checkable certificate.
- [§3.1, Stage 2 (Smoothing inequalities)] The displayed smoothing inequalities are obtained by an unproved extremal heuristic. The text says that the minimum of Σ d_a x_a under fixed t and A occurs when d_a = n−k−l for the first t/(n−k−l) indices and x_a = x_b for the remaining indices, 'by considering the continuous generalisation'. This is not a proof, and the resulting inequalities are used in Intersections I, II, IIa and IIb to restrict the feasible region. Since the smoothing inequalities are load-bearing for the strict-gap argument, a rigorous derivation of these bounds must be supplied.
- [§3.1, Stage 5 and the final casework of Theorem 3.7] The phase analysis is presented largely through visual Desmos plots and informal assertions about hyperbola branches. For example, the proof says 'we claim that smoothing(c) eliminates anything below the lower intersection' and 'the feasible region lies inside the region bounded by the two intersection points of the branches', but no analytic verification of the relevant convexity, monotonicity, and branch-selection facts is provided. These phase claims determine which of Intersections Ia, Ib, IIa, IIb is active, so they are not merely illustrative. The forward reference 'For reasons justified in Stage 5, we require the second root of this cubic' only compounds the problem, since Stage 5 itself is not formal. A complete proof needs explicit inequalities proving the claimed shape of the feasible regions in each phase.
- [§3.1, final paragraph before Claims 3.8 and 3.9] The transition from the equality-state analysis to the asymptotic contradiction relies on the statement that, by continuity, the matrices get arbitrarily close to the equality state. This presupposes a compactness or subsequence argument for the normalized parameters X, Y, Z, T, a, c, and ν as n grows. The later claims do use averages, but they do not directly prove the needed parameter convergence. This is a more localized gap than the numerical checks, but it is still part of the strict-gap argument and should be made precise.
minor comments (5)
- [Corollary 2.9] The proof of Corollary 2.9 cites Leonida and Li [7], which is an unpublished preprint. If this result is used only as motivation and for examples, this should be stated explicitly; if it is needed in the proof, the argument should be made self-contained or the citation should be to a published source.
- [Theorem 2.4] The statement that the neighbours of vertex i are {a_i, a_i+1, ..., b_i} mod n is ambiguous; please specify the circular-interval convention and the intended ranges of a_i and b_i.
- [§3.1, Stage 4] In the sentence 'mean(c) and smoothing(c) intersect until T = 1/8, similarly with mean(c) and smoothing(c)', the second clause should presumably refer to mean(a) and smoothing(a); please correct this typo.
- [Throughout §3.1] The proof of Theorem 3.7 is very long but has almost no equation numbering; references such as 'smoothing(c)' and 'Extrema' would be much easier to check if the key displayed formulas and inequalities were numbered.
- [Theorem 3.2] The phrase 'the complement of Q is in a perfect elimination ordering' is used before its connection to Theorem 2.4 is explained; a short definition or reference would improve readability.
Circularity Check
No significant circularity; the central derivation is self-contained and does not reduce to its inputs.
full rationale
The central claim, Theorem 1.2, is proved by a self-contained extremal argument. The operation G* is introduced via the variational characterization of the two smallest eigenvalues and is used only to impose structural restrictions, not to encode the target bound. The reduction to n/2-regular invariant graphs proceeds through explicit matrix block structure and eigenvector component inequalities, with no fitted parameter renamed as a prediction. The exclusion of the limiting value -sqrt(2)/2 in Theorem 3.7 is obtained by analyzing intersections of inequalities and ruling out the equality state X = Z = 1/2, T = 1/8 through a further contradiction between Type 1 and Type 2 eigenvectors. Section 4 reduces a hypothetical worst-case sequence to this constrained family using o(n^2) edge modifications, which preserves the spectrum up to o(n); this is a structural reduction, not a circular redefinition. The self-citation to Leonida and Li [7] appears in Corollary 2.9 and as motivation, but it is not load-bearing for Theorem 3.1 or Theorem 1.2: those proofs do not rely on the pivalous-graph eigenvalue results. Nikiforov's prior bounds are used as baselines and as external facts, not as the strengthened conclusion being derived. The Desmos-based numerical checks in Stage 4 are a verification gap and a correctness risk, not a circularity: they assert inequalities for explicit cubic polynomials rather than importing the target result. Overall, the derivation chain is independent of its conclusions, so the circularity score is minimal.
Assumptions & free parameters
assumptions (7)
- standard math Weyl's inequalities for Hermitian matrices
- domain assumption Nikiforov's bound λ_{n-1}^2 + λ_n^2 ≤ n^2/4, equivalently λ_{n-1}+λ_n ≥ -√2 n/2
- domain assumption Nikiforov's degree-regularity bound s(G)=Σ|d(i)-2e(G)/n|=o(n^2) when λ_1 is close to n/2
- standard math Chebyshev sum inequality for monotone sequences
- standard math Strong perfect graph theorem (Chudnovsky et al., 2006)
- standard math Rankin's packing lemma (Lemma 8 of Rankin, 1947)
- domain assumption Results from Leonida-Li [7] on abelian Cayley graphs and pivalous graphs
Cite this review
Pith. "Pith review of Strengthened upper bound on the third eigenvalue of graphs." pith.science (2026). https://pith.science/paper/EN2B5TSZ
@misc{pith2026250107494,
author = {Pith},
title = {Pith review of: Strengthened upper bound on the third eigenvalue of graphs},
year = {2026},
howpublished = {\url{https://pith.science/paper/EN2B5TSZ}},
note = {Machine review of arXiv:2501.07494}
}
abstract
Let $G$ be a graph on $n \ge 3$ vertices, whose adjacency matrix has eigenvalues $\lambda_1 \ge \lambda_2 \ge \dots \ge \lambda_n$. The problem of bounding $\lambda_k$ in terms of $n$ was first proposed by Hong and was studied by Nikiforov, who demonstrated strong upper and lower bounds for arbitrary $k$. Nikiforov also claimed a strengthened upper bound for $k \ge 3$, namely that $\frac{\lambda_k}{n} < \frac{1}{2\sqrt{k-1}} - \varepsilon_k$ for some positive $\varepsilon_k$, but omitted the proof due to its length. In this paper, we give a proof of this bound for $k = 3$. We achieve this by instead looking at $\lambda_{n-1} + \lambda_n$ and introducing a new graph operation which provides structure to minimising graphs, including $\omega \le 3$ and $\chi \le 4$. Then we reduce the hypothetical worst case to a graph that is $n/2$-regular and invariant under said operation. By considering a series of inequalities on the restricted eigenvector components, we prove that a sequence of graphs with $\frac{\lambda_{n-1} + \lambda_n}{n}$ converging to $-\frac{\sqrt{2}}{2}$ cannot exist.
Figures
Figures from the paper (5 more)
Forward citations
Cited by 2 Pith papers
-
Graph Eigenvalues and Projection Constants
λ_k(G) ≤ ((k−2)√(k+1)+2)n/(2k(k−1)) − 1 for all graphs, tight for k ∈ {2,3,4,8,24}, resolving c₃ = 1/3 and Nikiforov's Conjecture 4.2.
-
Generalized Nordhaus--Gaddum Inequalities for Eigenvalues
The asymptotic maximum of λ₁(G)+λ₂(complement of G) is exactly 8/7 per vertex, with new general bounds for all pairs and a short proof of Terpai's spectral-radius bound.
Reference graph
Works this paper leans on
-
[1]
Efficient testing of large graphs
Noga Alon, Eldar Fischer, Michael Krivelevich, and Mario Szegedy. Efficient testing of large graphs. Combinatorica, 20(4):451–476, 2000
2000
-
[2]
C. Borgs, J.T. Chayes, L. Lov´ asz, V.T. S´ os, and K. Vesztergombi. Convergent se- quences of dense graphs I: Subgraph frequencies, metric properties and testing. Ad- vances in Mathematics , 219(6):1801–1851, 2008
work page 2008
-
[3]
Chudnovsky, N
M. Chudnovsky, N. Robertson, P. Seymour, and R. Thomas. The strong perfect graph theorem. Annals of Mathematics , 164(1):51–229, 2006
2006
-
[4]
P´ eter Csikv´ ari. On a conjecture of V. Nikiforov.Discrete Mathematics, 309(13):4522– 4526, 2009
work page 2009
-
[5]
On the sum of two largest eigenvalues of a symmetric matrix
Javad Ebrahimi B, Bojan Mohar, Vladimir Nikiforov, and Azhvan Sheikh Ahmady. On the sum of two largest eigenvalues of a symmetric matrix. Linear Algebra and its Applications, 429(11):2781–2787, 2008
work page 2008
-
[6]
Bounds of eigenvalues of graphs
Yuan Hong. Bounds of eigenvalues of graphs. Discrete Mathematics , 123(1):65–74, 1993
work page 1993
-
[7]
On graphs with large third eigenvalue, 2025
Giacomo Leonida and Sida Li. On graphs with large third eigenvalue, 2025
work page 2025
-
[8]
Improved lower bounds on the extrema of eigenvalues of graphs
William Linz. Improved lower bounds on the extrema of eigenvalues of graphs. Graphs and Combinatorics , 39, 07 2023
work page 2023
Show all 13 references
-
[9]
Linear combinations of graph eigenvalues
Vladimir Nikiforov. Linear combinations of graph eigenvalues. ELA. The Electronic Journal of Linear Algebra , 15:329–336, 2006
2006
-
[10]
Eigenvalue problems of Nordhaus–Gaddum type
Vladimir Nikiforov. Eigenvalue problems of Nordhaus–Gaddum type. Discrete Math- ematics, 307(6):774–780, 2007
2007
-
[11]
Extrema of graph eigenvalues
Vladimir Nikiforov. Extrema of graph eigenvalues. Linear Algebra and its Applications, 482:158–190, 2015
2015
-
[12]
R. A. Rankin. On the closest packing of spheres in n dimensions. Annals of Mathe- matics, 48(4):1062–1081, 1947
1947
-
[13]
Proof of a conjecture of V
Tam´ as Terpai. Proof of a conjecture of V. Nikiforov.Combinatorica, 31:739–754, 2011. Appendix We deal with the block construction mentioned in the proof of Theorem 3.7 and prove that C[t] 6 is optimal over all closed vertex multiplications of C6 by [a, b, c, a, b, c]. Theore...
2011
Reviewed August 10, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.