The integrability of sesqui-regular graphs with smallest eigenvalue at least $-3$

Throughout the proof, Assumptions 3.3 and 3.5 are in force.

(i) Since $\mathfrak{g}$ is a factor of the fat Hoffman graph $\mathfrak{h}$, Lemma 2.3 implies that every fat neighbor in $\mathfrak{h}$ of a slim vertex of $\mathfrak{g}$ also belongs to $\mathfrak{g}$. Thus $\mathfrak{g}$ is fat. By Lemma 2.2,

If some slim vertex of $\mathfrak{g}$ had at least two fat neighbors, Theorem 2.5 would give an integral reduced representation of norm $3$ for $\mathfrak{g}$, contrary to Assumption 3.5. Since $\mathfrak{g}$ is fat, every slim vertex therefore has exactly one fat neighbor.

(ii) ($a$) The indecomposability of $\mathfrak{g}$ is equivalent to connectedness of its special graph, so $D$ is connected.

($b$) Let $x$ and $y$ be distinct slim vertices of $\mathfrak{g}$. A negative special edge could occur only if $x$ and $y$ were nonadjacent in the slim graph and shared a fat neighbor. In that case both slim vertices would lie in the same quasi-clique of $\mathfrak{h}$, which is a clique by Assumption 3.3, a contradiction. Thus the special graph has no negative edges.

($c$) Since every slim vertex has exactly one fat neighbor, each diagonal entry of $\operatorname{Sp}(\mathfrak{g})$ equals $-1$, while an off-diagonal entry equals $1$ precisely on an edge of $D$. This proves $A(D)=\operatorname{Sp}(\mathfrak{g})+I$, and it follows immediately that

($d$) Suppose, to the contrary, that $D$ is $1$-integrable. If $D\cong K_{1}$, then $\mathfrak{g}$ has a single slim vertex with one fat neighbor. Assigning the integral vector $\mathbf{e}_{1}+\mathbf{e}_{2}$ to the slim vertex gives $\mathfrak{g}$ an integral reduced representation of norm $3$, a contradiction. Thus $|V(D)|\geq 2$.

If $\lambda_{\min}(D)=-1$, then $D$ is a clique, because every connected graph with smallest eigenvalue at least $-1$ is complete. Assigning $\mathbf{e}_{0}+\mathbf{e}_{i}$ to the $i$th vertex of $D$ gives an integral matrix $N$ such that

If $-2\leq\lambda_{\min}(D)<-1$, the assumed $1$-integrability of $D$ again gives an integral matrix $N$ satisfying (6). Since $A(D)+2I=\operatorname{Sp}(\mathfrak{g})+3I$, the columns of $N$ give an integral reduced representation of norm $3$ for $\mathfrak{g}$, a contradiction. Hence $D$ is not $1$-integrable. Since $D$ is connected and has smallest eigenvalue at least $-2$, Theorem 3.6 now implies (5).

($e$) The sets $F_{f}$ clearly partition $V(D)$. For two distinct vertices $x,y\in F_{f}$, they lie in the clique $Q_{\mathfrak{h}}(f)$, and hence they are adjacent in the slim graph. Their corresponding special matrix entry is nevertheless $1-1=0$, because they share the fat vertex $f$. Thus $\{x,y\}\notin E(D)$, and every set $F_{f}$ is independent. ∎

Let $v\sim_{\Gamma}u$. If $v\notin V_{\mathrm{slim}}(\mathfrak{g})$, then $u$ and $v$ belong to different factors of $\mathfrak{h}$. Lemma 2.3 therefore implies that their unique common fat neighbor is $f_{u}$, so $v\in C(f_{u})$. If $v\in V_{\mathrm{slim}}(\mathfrak{g})$, then either $f_{v}=f_{u}$, in which case $v\in C(f_{u})$, or $f_{v}\neq f_{u}$, in which case $v\in N_{D}(u)$.

Conversely, every vertex of $C(f_{u})\setminus\{u\}$ is adjacent to $u$ because $C(f_{u})$ is a clique, and every vertex of $N_{D}(u)$ is adjacent to $u$ in $\Gamma$. The two sets are disjoint: the vertices of $C(f_{u})\cap V_{\mathrm{slim}}(\mathfrak{g})$ belong to $F_{f_{u}}$, whereas $N_{D}(u)$ is disjoint from $F_{f_{u}}$. This proves (7). Since $d_{D}(u)\leq|V(D)|-1\leq 35$, (8) follows. ∎

The independence of the sets $F_{f}$ gives $f_{x}\neq f_{y}$. More generally, if $f$ and $f^{\prime}$ are distinct fat vertices of $\mathfrak{g}$, then

Indeed, the sets $F_{f}$ and $F_{f^{\prime}}$ are nonempty because every fat vertex of $\mathfrak{g}$ has a slim neighbor. Lemma 3.4 gives $|C(f)\cap C(f^{\prime})|\leq 2$, while Lemma 3.8, applied to vertices of $F_{f}$ and $F_{f^{\prime}}$, gives $|C(f)|,|C(f^{\prime})|\geq k-34$. If the intersection were nonempty, each vertex in it would have valency at least

a contradiction. In particular, $C_{x}\cap C_{y}=\varnothing$.

Because $y\in N_{D}(x)\cap F_{f_{y}}$ and $|V(D)|\leq 36$,

Every fat vertex in $\mathcal{F}_{x,y}\cup\{f_{x}\}$ is distinct from $f_{y}$, so (15) implies (11). By (7) and the independence of $F_{f_{y}}$,

This proves (12).

Let $z\in Z_{x,y}$. Since $y,z\in C_{y}$, the vertices $y$ and $z$ are adjacent in $\Gamma$. Moreover, $z\notin V_{\mathrm{slim}}(\mathfrak{g})$ by (11). If $x\sim_{\Gamma}z$, then Lemma 2.3 would imply $z\in C_{x}$, contrary to $C_{x}\cap C_{y}=\varnothing$. Hence $\operatorname{dist}_{\Gamma}(x,z)=2$.

First,

Indeed, every vertex of $W_{x}$ is adjacent to $x$ because $C_{x}$ is a clique. No vertex of $W_{x}$ belongs to $V_{\mathrm{slim}}(\mathfrak{g})$: otherwise its unique fat neighbor would be $f_{x}$, and its adjacency to $z$ across two factors would imply $z\in C_{x}$. Conversely, if $v\in N_{\Gamma}(x,z)\setminus V_{\mathrm{slim}}(\mathfrak{g})$, then the adjacency of $x$ and $v$ across two factors implies that their common fat neighbor is $f_{x}$. Hence $v\in C_{x}$, and therefore $v\in W_{x}$. This proves (16).

Next,

Let $u$ belong to the left-hand side. If $f_{u}=f_{x}$, then the adjacency of $u$ and $z$ across two factors implies $z\in C_{x}$, a contradiction. Thus $f_{u}\neq f_{x}$. Since $u\sim_{\Gamma}x$ and $u,x$ have distinct fat neighbors, $u\in N_{D}(x)$. If $f_{u}\neq f_{y}$, then (15) gives $C(f_{u})\cap C_{y}=\varnothing$, whereas the adjacency of $u$ and $z$ across two factors implies $z\in C(f_{u})$ and then $z\in C(f_{u})\cap C_{y}$. Thus $f_{u}=f_{y}$ which proves the inclusion from left to right.

Conversely, let $u\in N_{D}(x)\cap F_{f_{y}}$. Then $u\sim_{\Gamma}x$. Moreover, $u,z\in C_{y}$, so $u\sim_{\Gamma}z$. This proves (17). Combining (16) and (17) gives (13); (14) follows from sesqui-regularity. ∎

Fix a pair of vertices $x$ and $y$ which are adjacent in $D$, and put

Both $m$ and $m^{\prime}$ are positive. In $\Gamma$, every vertex of $Z_{x,y}$ has exactly $c-m$ neighbors in $C_{x}$ by (14). These neighbors lie outside $\mathfrak{g}$ by (16), and hence they belong to $Z_{y,x}=C_{x}\setminus F_{f_{x}}$. Interchanging $x$ and $y$, shows that every vertex of $Z_{y,x}$ has exactly $c-m^{\prime}$ neighbors in $Z_{x,y}=C_{y}\setminus F_{f_{y}}$. Consequently,

Suppose first that $c\geq 5$ and, contrary to the desired conclusion, $m\leq 2$. Choose $u\in Z_{y,x}$. Every vertex of $C_{x}$ has at most $35$ neighbors outside $C_{x}$ by (8), whereas $|Z_{x,y}|\geq k-35\geq 63$. Hence some $z\in Z_{x,y}$ is nonadjacent to $u$. Since $|N_{\Gamma}(z)\cap C_{x}|=c-m\geq 1$, the vertices $u$ and $z$ have a common neighbor; hence $\operatorname{dist}_{\Gamma}(u,z)=2$. The two disjoint sets

consist of common neighbors of $u$ and $z$. Their sizes are $c-m$ and $c-m^{\prime}$, respectively. Sesqui-regularity therefore gives

and therefore

The left-hand side of (19) is positive, so $m^{\prime}\leq c-1$. Moreover, by (12),

because $d_{D}(x)\geq m$ and $|F_{f_{x}}|\geq m^{\prime}$. Thus

If $m=1$, then (20) and $m^{\prime}\leq c-1$ give $m^{\prime}=c-1$; (21) then gives $k\leq 45$ when $c\in\{5,6,7,8\}$. If $m=2$, then $m^{\prime}\in\{c-2,c-1\}$, and (21) gives $k\leq 97$. Both alternatives contradict $k\geq 98$. Therefore $m\geq 3$ whenever $c\geq 5$.

Finally, suppose that $c\in\{3,4\}$ and $m=1$. The preceding argument again gives $m+m^{\prime}\geq c$ and $m^{\prime}\leq c-1$, so $m^{\prime}=c-1$. Here (21) gives $k\leq 68$ for $c=3$ and $k\leq 51$ for $c=4$, again a contradiction. Hence $m\geq 2$ in this case. This proves (18). ∎

Let $H$ be a nontrivial connected component of $D[F_{f_{i}}\cup F_{f_{j}}]$ with minimum valency $\delta(H)$ and largest eigenvalue $\lambda_{\max}(H)$. Every vertex of $H$ is incident with an edge of $H$. Lemma 3.10 shows that each vertex of $H$ has at least two neighbors. Thus $\delta(H)\geq 2$. Since $H$ is an induced subgraph of $D$, Lemma 3.1 and (4) give

The graph $H$ is bipartite by Lemma 3.7 (ii) $(e)$, and hence $\lambda_{\max}(H)=-\lambda_{\min}(H)\leq 2$. On the other hand, the Rayleigh quotient of the all-ones vector yields

Equality holds throughout. Thus the average valency and the minimum valency of $H$ are both equal to $2$, so $H$ is a cycle; since it is bipartite, the cycle has even length. The final assertion follows immediately. ∎

Fix a fat vertex $f_{i}$ of $\mathfrak{g}$ such that $F_{f_{i}}\neq\varnothing$. (7) shows that $d_{D}(x)=k+1-|C(f_{i})|$ for every $x\in F_{f_{i}}$. Thus $d_{D}(x)$ is constant on $F_{f_{i}}$; denote its value by $d_{f_{i}}$.

Suppose that $D[F_{f_{i}}\cup F_{f_{j}}]$ has an edge joining $F_{f_{i}}$ and $F_{f_{j}}$. Let $H$ be the component of $D[F_{f_{i}}\cup F_{f_{j}}]$ containing that edge. By Lemma 3.11, $H$ is a cycle of even length. Let

Define $\mathbf{q}\in\mathbb{R}^{V(D)}$ coordinatewise by

Since $H$ is a cycle of even length with bipartition $(X,Y)$, $|X|=|Y|$ and $|E(H)|=|X|+|Y|$. Therefore

By (4), the matrix $A(D)+2I$ is positive semidefinite; hence $(A(D)+2I)\mathbf{q}=0$. Multiplying by the all-ones row vector gives

Thus $d_{f_{i}}=d_{f_{j}}$ whenever an edge of $D$ joins $F_{f_{i}}$ and $F_{f_{j}}$. The auxiliary graph whose vertices are the nonempty sets $F_{f}$, with two such sets adjacent whenever there is an edge in $D$ joining them, is connected because $D$ is connected. Hence all values $d_{f}$ are equal, and $D$ is regular. Finally, Lemma 3.11 shows that every set $F_{f}$ not containing a given vertex contributes either zero or two of its neighbors, so the valency of $D$ is even. ∎