---
title: Verification in Upper-bound Graph
url: https://www.emergentmind.com/topics/verification-in-upper-bound-graph
type: topic
---

# Verification in Upper-bound Graph

Searching arXiv for the cited papers to ground the article in current records.
Verification in “upper-bound graph” (*Editor’s term*) denotes work in which the central object of certification is an upper bound on a graph parameter, a graph-derived invariant, or a graph property. In the cited literature, this appears in several technically distinct forms: extremal characterization of a unique maximizing graph or simplicial complex, spectral inequalities with explicit equality cases, local and global certification schemes for graph properties, and witness-based procedures in which a verified structure certifies that a quantity cannot exceed a stated value. The most explicit combinatorial instance is the upper bound theorem for odd-dimensional flag triangulations of homology manifolds, where the balanced join of \(r\) cycles is proved to be the unique maximizer of the \(f\)-, \(h\)-, \(g\)-, and \(\gamma\)-vector entries for sufficiently many vertices [1503.05961].

## 1. Extremal verification for odd-dimensional flag triangulations

For a \((d-1)\)-dimensional flag triangulation \(M\) of a homology manifold, with \(d\) even and \(n\) sufficiently large, the central extremal statement is: for every face function \(F\) of dimension \(\ell\) with \(1\le \ell \le d/2-1\), there is a constant \(n_0\) such that, if \(M\) has at least \(n_0\) vertices, then
\[
F(M)\le F\!\left(\mathbf{J}_{d/2}(n)\right),
\]
with equality if and only if
\[
M \cong \mathbf{J}_{d/2}(n).
\]
In graph form, for every \(r\ge 2\) and every clique function \(F\) of order \(k\) with \(2\le k\le r\), there is \(n_0\) such that any graph \(G\) on \(n\ge n_0\) vertices that triangulates a \((2r-1)\)-dimensional homology manifold satisfies
\[
F(G)\le F(J_r(n)),
\]
and equality holds iff \(G\cong J_r(n)\) [1503.05961].

The extremal object is the balanced join of \(r\) cycles, denoted \(\mathbf{J}_r(n)\) on the simplicial-complex side and \(J_r(n)\) on the graph side. It is built from the Turán graph \(T_r(n)\) with parts \(V_1,\dots,V_r\) of sizes \(\lfloor n/r\rfloor\) or \(\lceil n/r\rceil\), then each part is made to induce a cycle. When \(n\ge 4r\), each part is a cycle of length at least \(4\), so the clique complex \(X(J_r(n))\) is a flag simplicial \((2r-1)\)-sphere [1503.05961].

The theorem is not limited to a single face-counting statistic. The corresponding corollary states
\[
f_i(M)\le f_i(\mathbf{J}_{d/2}(n)),\qquad
h_i(M)\le h_i(\mathbf{J}_{d/2}(n)),\qquad
g_i(M)\le g_i(\mathbf{J}_{d/2}(n)),\qquad
\gamma_i(M)\le \gamma_i(\mathbf{J}_{d/2}(n)),
\]
for the relevant ranges of \(i\), and equality in any of these implies \(M\cong \mathbf{J}_{d/2}(n)\) [1503.05961].

The underlying definitions are standard in flag-complex combinatorics. A flag simplicial complex is one in which every clique in the \(1\)-skeleton spans a face; equivalently, it is the clique complex of its graph. A homology manifold of dimension \(2r-1\) is a simplicial complex whose local homology looks like that of \(S^{2r-1}\), and in particular every vertex link is a homology \((2r-2)\)-sphere. The balanced join is the join of \(r\) cycle graphs, with the vertex set split as evenly as possible among the \(r\) factors [1503.05961].

The face-number vectors appearing in the theorem are defined by
\[
\sum_{i=0}^{d} f_{i-1}x^i(1-x)^{d-i}=\sum_{i=0}^{d} h_i x^i,
\]
with \(g_0=1\) and \(g_i=h_i-h_{i-1}\), and, for Eulerian complexes,
\[
\sum_{i=0}^{d} h_i x^i=\sum_{i=0}^{\lfloor d/2\rfloor}\gamma_i x^i(x+1)^{d-2i}.
\]
These formulas place the theorem within the standard upper-bound tradition for simplicial manifolds, but the flag condition makes the extremal object a graphically defined cycle join rather than a neighborly complex [1503.05961].

## 2. Extremal graph theory, stability, and rigidity

The proof strategy in the odd-dimensional flag case translates the face-counting problem into a graph problem on clique counts. The stated ingredients are a middle Dehn–Sommerville estimate to bound the number of large cliques, the graph removal lemma to make the graph \(K_{r+1}\)-free after deleting few edges, Zykov’s inequalities to compare clique counts in \(K_{r+1}\)-free graphs with the Turán graph, and the Erdős–Simonovits stability theorem to conclude that a near-extremal graph must be close to \(T_r(n)\) [1503.05961].

The role of the stability input is structural rather than merely numerical. After obtaining a graph that is close to the Turán graph, the argument analyzes “almost extremal” graphs and proves that the only local structure compatible with maximality is the balanced join of cycles. A final rigidity argument shows that equality forces the graph to be \(r\)-radical and hence exactly \(J_r(n)\) [1503.05961].

A useful formula characterizes the clique counts of graphs in the extremal class. If the parts are \(V_1,\dots,V_r\), then
\[
e_\ell(G)=\sum_{j=0}^{\ell}\binom{j}{\ell-j}\,\sigma_j\!\left(|V_1(G)|,\ldots,|V_r(G)|\right),
\]
where \(\sigma_j\) is the \(j\)-th elementary symmetric polynomial. This is used to show that the objective is maximized when the parts are as equal as possible, and hence when \(G\) is exactly \(J_r(n)\) [1503.05961].

This equality theory is unusually rigid. It does not merely identify an asymptotic extremal profile; it gives a unique maximizer. A plausible implication is that, in this setting, “verification” is inseparable from classification: once the upper bound is sharp, the homology-manifold condition and flag condition together force the graph to have the cycle-join structure.

## 3. Spectral upper bounds as verification statements

A second major strand verifies upper bounds by converting graph invariants into spectral inequalities. For chromatic number, one result proves that if \(G\) is a simple graph on \(n\) vertices with chromatic number \(\chi\) and least adjacency eigenvalue \(\lambda_n\), then for every graph with \(3\le \chi \le n-1\),
\[
\chi\le \left(\frac{n}{2}+1+\lambda_n\right)+\sqrt{\left(\frac{n}{2}+1+\lambda_n\right)^2-4(\lambda_n+1)\left(\lambda_n+\frac{n}{2}\right)},
\]
with equality if and only if
\[
G \cong \left(K_{\frac{\chi}{2}\cup \frac{\chi}{2}K_1}\right)\vee \left(K_{\frac{n-\chi}{2}\cup \frac{n-\chi}{2}K_1}\right),
\]
where both \(n\) and \(\chi\) are even [2511.07712]. The proof reduces to graphs of the form
\[
G(a,a_0)= (K_a\cup O_{a_0}) \vee (K_b\cup O_{b_0}),
\]
uses an explicit quartic characteristic polynomial for the least eigenvalue, and establishes extremality by an algebraic comparison with the symmetric choice \(a=\chi/2\), \(a_0=(n-\chi)/2\) [2511.07712].

For the Hodge \(1\)-Laplacian of a graph’s clique complex, the graph case of the Helmholtzian problem yields the unconditional estimate
\[
\lambda_{\max}\!\bigl(L_1^{\mathrm{up}}(G)\bigr)\le \mu_1(G)+\frac13\bigl(n-\mu_1(G)\bigr),
\]
where \(\mu_1(G)\) is the largest ordinary graph Laplacian eigenvalue. The argument localizes the problem on the cycle space \(Z_1\) of \(K_n\) and rewrites it as a lower-bound problem for the up Laplacian \(\bar L(G)\) of the missing triangles:
\[
\lambda_{\max}\!\bigl(L_1^{\mathrm{up}}(G)\bigr)=n-\lambda_{\min}\!\bigl(\bar L(G)\big|_{Z_1}\bigr),
\]
with
\[
a(\overline{G})=n-\mu_1(G).
\]
The stated obstruction is the comparison
\[
\bar L(G)\succeq \frac13 Q,
\]
which is too crude on triangles with two or three missing edges [2606.19742].

For adjacency spectra, a general upper bound on the \(k\)-th eigenvalue states that for every graph \(G\) on \(n\) vertices and every \(k\ge 2\),
\[
\lambda_k(G)\le \frac{(k-2)\sqrt{k+1}+2}{2k(k-1)}\,n - 1.
\]
The proof combines Weyl’s inequality, the complement identity \(A(G)+A(\overline G)=J-I\), and a projection estimate derived from the positivity of the kernels
\[
f_2^r(t)=t^2-\frac1r,\qquad f_4^r(t)=t^4-\frac{3}{r(r+2)}
\]
on \(S^{r-1}\). The bound is stated to be tight for \(k\in\{2,3,4,8,24\}\) [2603.28738].

For independence-type parameters, the inertia bound extends from classical independence to quantum independence. If \(A\) has inertia \((n^+,n^0,n^-)\), then
\[
\alpha_q(G)\le n^0(A)+\min\{n^+(A),n^-(A)\},
\]
and more generally
\[
\alpha_q^k(G)\le \min\Bigl\{\,\#\{i: p_k(\lambda_i)\ge w(p_k)\},\ \#\{i: p_k(\lambda_i)\le W(p_k)\}\Bigr\}.
\]
The proof passes through the quantum \(k\)-projective packing number and an interlacing argument for projectors encoded as vectors in a tensor product space [1910.07339].

Taken together, these results verify upper bounds by showing that extremal spectral data force highly constrained graph structure. This suggests that spectral verification is strongest when it is accompanied by an equality classification or an explicit obstruction.

## 4. Approximate, local, and global certification of graph properties

In distributed verification, the notion of upper-bound certification changes form. An approximate proof labeling scheme for a graph property \(P\subseteq \mathrm{Gra}\), where \(\mathrm{Gra}\) is the class of finite simple graphs of maximum degree at most \(d\), is defined using the edit distance
\[
e(H,P) := \inf_{G \in P,\; V(G)=V(H)} \frac{|E(G)\triangle E(H)|}{|V(H)|}.
\]
A property \(P\) can be verified by an approximate proof labeling scheme in constant time if, for every \(\varepsilon>0\), there exist a finite label set \(Q_\varepsilon\), a constant radius \(r_\varepsilon\), and a local verifier \(V_\varepsilon\subseteq B^{Q_\varepsilon}_{r_\varepsilon,d}\) such that every \(G\in P\) admits an accepted proof, while every graph \(H\) with \(e(H,P)>\varepsilon\) rejects every proof. Equivalently,
\[
P \subseteq L_{V_\varepsilon} \subseteq P_\varepsilon.
\]
The main theorem states that planarity, and more generally all monotone hyperfinite properties that are closed under disjoint unions, can be verified by an approximate proof labeling scheme in constant time for bounded-degree graphs; the listed examples include planar graphs, outer-planar graphs, bounded genus graphs, knotlessly embeddable graphs, and, more generally, minor-closed families [2006.11869].

The verification mechanism for approximate planarity proceeds through Property A and strong hyperfiniteness. Property A requires a probability-measure-valued map
\[
f:V(G)\to \mathrm{Prob}(V(G))
\]
with local support
\[
\operatorname{Supp}(f(x)) \subseteq B_r(x,G)
\]
and edgewise \(\ell^1\)-variation
\[
\|f(x)-f(y)\|_1 < \varepsilon \quad \text{whenever } x\sim y.
\]
The local certificates encode discretized probability distributions and symmetry-breaking color information, so that each node inspects only a constant-radius neighborhood [2006.11869].

For exact local certification, one upper bound concerns \(P_5\)-freeness. In the radius-\(1\) model, there is a certification scheme for \(P_5\)-free graphs with certificates of size \(O(n^{3/2})\). The proof uses the structural theorem of Bacsó and Tuza that every connected \(P_5\)-free graph has a dominating set that is either a clique or an induced \(P_3\), then recursively builds a valid tree partition \((T,\beta)\) whose parts are cliques or \(P_3\)s and whose children correspond to connected components after removing the dominating part [2410.14658].

For global certification, graph homomorphism to a fixed graph \(H\) admits a global certificate of size
\[
O(n\log n' + \log\log M),\qquad n'=|V(H)|.
\]
The certificate is
\[
c=(n,h,L),
\]
where \(h\) is chosen from a \((k,\ell)\)-perfect hash family \(H_{k,\ell}\) of size \(\lceil k e^k\log \ell\rceil\), and \(L[i]=\varphi(v)\) when \(h(Id(v))=i\). For bipartiteness, obtained by taking \(H=K_2\), this becomes \(O(n+\log\log M)\), matching the lower bound \(\Omega(n+\log\log M)\) [2402.03849].

These certification results show that upper-bound verification need not mean a bound on a numerical invariant alone. It can also mean a controlled acceptance region: exact acceptance for the target class, or acceptance for graphs within a prescribed edit-distance neighborhood.

## 5. Kernelization, algebraic proofs, and extremal graph forms

Upper-bound verification also appears as a decision problem near a known ceiling. For the independence number, with
\[
p := \left\lfloor \frac{1}{2} + \sqrt{\frac{1}{4} + n^2 - n - 2m} \right\rfloor,
\]
one asks whether \(\alpha(G)\le p-k\). If \(p\ge 2k+1\), there is an \(O(n^2)\) algorithm that returns an induced subgraph \(G_{p,k}\) with \(n_0\le p+2k+1\) vertices such that
\[
\alpha(G) \le p-k \iff \alpha(G_{p,k}) \le p-k,
\]
and the reduced instance can be decided in time
\[
O(1.2738^{3k} + kn).
\]
The reduction deletes vertices of degree at least \(n-p+k\), and the resulting kernel is then handled through a vertex-cover formulation [1709.02475].

For spanning trees, an upper bound is verified directly from the degree sequence. If \(G\) is a connected simple graph on \(n\) vertices with degree sequence
\[
d_1 \le d_2 \le \cdots \le d_n,
\]
then
\[
T(G)\le \frac{(d_1+1)(d_2+1)\cdots(d_n+1)}{n^2}.
\]
The proof uses the Laplacian spectrum
\[
0=\mu_1<\mu_2\le \cdots \le \mu_n,
\]
Kirchhoff’s Matrix Tree Theorem,
\[
T(G)=\frac{1}{n}\mu_2\mu_3\cdots \mu_n,
\]
the complement graph, Schur’s inequality, and Karamata’s inequality. Equality holds exactly for \(K_n\), recovering Cayley’s formula \(T(K_n)=n^{n-2}\) [2103.00310].

For generalized Hamiltonian-type quantities, if \(f:V(H)\to V(G)\) is a pseudoordering and
\[
s_H(f,G)=\sum_{\{x,y\}\in E(H)} \rho_G\bigl(f(x),f(y)\bigr),
\]
then the upper \(H\)-Hamiltonian number is
\[
h_H^+(G)=\max\{s_H(f,G)\mid f \text{ pseudoordering of } G \text{ by } H\}.
\]
For connected finite graphs \(G,H\) with \(|V(G)|=|V(H)|\),
\[
h_H^+(G)\le h_H^+\bigl(P_{|V(G)|-1}\bigr),
\]
and, when \(H\) is connected, equality holds if and only if \(G\) is a path. The proof passes through spanning trees and a local tree-transformation argument that decreases a degree statistic until a path is reached [2003.07652].

These examples illustrate three common verification formats. One format reduces the problem to a smaller equivalent graph, as in near-upper-bound independence. A second proves a universal numerical inequality with a complete equality case, as in spanning trees. A third identifies the unique extremal graph under a distance-based objective, as in the upper \(H\)-Hamiltonian number.

## 6. Clarifications, limitations, and recurrent misconceptions

A persistent misconception is that a successful verification scheme must be exact. In bounded-degree graph certification, planarity cannot be verified by a constant-time exact proof labeling scheme, but it can be verified approximately: every planar graph can be accepted, every graph more than \(\varepsilon\)-far from planar must be rejected, and graphs with \(e(H,P)\le \varepsilon\) may still be accepted [2006.11869]. Approximate verification therefore certifies a robust neighborhood of the property rather than exact membership.

Another misconception is that an upper bound for one coloring parameter should automatically control related parameters. The least-eigenvalue upper bound on \(\chi\) is specific to the ordinary chromatic number. The same expression is not an upper bound for the list chromatic number or the coloring number; the stated example is \(G=K_{3,3}\), where \(n=6\) and \(\lambda_n=-3\), so the formula yields \(2\), but
\[
\chi_\ell(K_{3,3})=\operatorname{col}(K_{3,3})\ge 3
\]
[2511.07712].

A third misconception is that rough structural information is enough to force equality in a sharp upper bound. For the Helmholtzian problem, the estimate
\[
\lambda_{\max}\!\bigl(L_1^{\mathrm{up}}(G)\bigr)\le \mu_1(G)+\frac13(n-\mu_1(G))
\]
is sharp exactly when the ceiling \(n\) is attained, but being a join is not sufficient to force \(\lambda_{\max}(L_1^{\mathrm{up}}(G))=n\). The explicit example
\[
G^*=K_1\vee\bigl(K_2\cup (n-3)K_1\bigr)
\]
satisfies \(\mu_1(G^*)=n\), yet
\[
\lambda_{\max}(L_1^{\mathrm{up}}(G^*))=3
\]
[2606.19742].

A fourth misconception is that heuristic search alone certifies an upper bound. In the witness-based study of APM-LDPC codes with active Tanner graphs of girth \(8\), search is used only to generate candidates. The certified upper bound begins only after exact verification of the two algebraic conditions
\[
H_Zx^{\mathsf T}=0,\qquad x\notin Row(H_X),
\]
or the \(X\leftrightarrow Z\) analogue. Once these are checked, the witness is a non-stabilizer logical representative, and its weight is a valid upper bound on the corresponding minimum distance [2604.15307].

Across these settings, the common theme is not a single formalism but a repeated logical pattern: an upper bound is made verifiable by coupling a numerical inequality with a structural certificate, an equality classification, or an explicit rejection criterion. The strongest results are those in which the bound, the extremal object, and the obstruction to equality are all identified explicitly.

Source: https://www.emergentmind.com/topics/verification-in-upper-bound-graph