Graphs with three eigenvalues from group divisible designs

By Theorem 1.2 the degree-partition is equitable. Hence, B\mathbf1_{n_2}=k_{12}\mathbf1_{n_1} and B^ {\mathpalette\@transpose{}} \mathbf1_{n_1}=k_{21}\mathbf1_{n_2}. The (1,1)-block of (1.1) is

where \alpha_1^2=k_{11}+k_{12}+\theta_1\theta_2. In the complete multipartite case, A_1=J-I_m\otimes J_n and A_1^2=(n_1-2n)J+n(I_m\otimes J_n). In the disjoint-union-of-cliques case, A_1=I_m\otimes J_n-I and A_1^2=I+(n-2)(I_m\otimes J_n). Substitution in (Proof) yields

with the values of \lambda_1 and \lambda_2 given by (2.1). Since B is a \{0,1\}-matrix with constant row and column sums, this proves the GDD assertion.

For the rest of the proof, assume that \mathcal D is semiregular. On the (m-1)-dimensional space of vectors that are constant on each group and orthogonal to \mathbf1_{n_1}, the eigenvalue of BB^ {\mathpalette\@transpose{}} is

Using semiregularity and (?), we have \beta=k_{12}k_{21}-n_1\lambda_2=0. On the other hand, substituting (2.1) into (?) yields

By Lemma 1.1, we have \theta_1\geqslant0 and \theta_2<0. Furthermore, since n\geqslant2, the factors n+\theta_1 and n-1-\theta_2 are positive. Consequently, n=-\theta_2 in the first case and n=\theta_1+1 in the second.

Since n_1k_{12}=n_2k_{21}, we have

The vanishing of \beta shows that B^ {\mathpalette\@transpose{}} annihilates every vector orthogonal to \mathbf1_{n_1} that is constant on each group. Therefore, the sum of the incidence rows in any group is independent of the group, whence (I_m \otimes J_n)B=(k_{21}/m)J and hence (I_m \otimes J_n)\bigl(B-\frac{k_{21}}{n_1}J_{n_1,n_2}\bigr)=0. Thus,

where a=0 in the complete multipartite case and a=-1 in the disjoint-union-of-cliques case. The (1,2)-block of (1.1) is

Using (?), we obtain

and transposing yields,

Thus A_2 acts as the scalar \theta_1+\theta_2-a on \operatorname{Im}\bigl((B-\frac{k_{21}}{n_1}J_{n_1,n_2})^ {\mathpalette\@transpose{}} \bigr). Moreover,

is an orthogonal decomposition. Indeed, k_{21}>0 by connectedness, and the row space of B is \langle\mathbf1_{n_2}\rangle\oplus \operatorname{Im}\bigl((B-\frac{k_{21}}{n_1}J_{n_1,n_2})^ {\mathpalette\@transpose{}} \bigr). These subspaces are invariant under the symmetric matrix A_2. Since \ker B\subseteq\mathbf1_{n_2}^{\perp} the (2,2)-block of (1.1) restricted to \ker B yields

The eigenvalues of A_2 therefore belong to \{k_{22},\theta_1+\theta_2-a,\theta_1,\theta_2\}.

Since \mathcal D is semiregular, (?) becomes r-\lambda_1=n(\lambda_2-\lambda_1)>0. In particular, \lambda_2-\lambda_1>0 and k>0. We can rearrange (2.3) as

Since (B-\frac{k}{v}J_{v,b}) is nonzero, \lambda_2-\lambda_1-n is an eigenvalue of A(\Gamma), and \lambda_2-\lambda_1-n \leqslant k_\Gamma. Therefore,

Since v-n+\lambda_1>0, we apply (2.3) to deduce d\coloneqq\frac{1}{b} \big((k_\Gamma-\lambda_2+\lambda_1)(k_\Gamma+n)+rk\big)>0. We may therefore set

and (2.3) implies \alpha_1\alpha_2=C. Let A=\left [\begin{smallmatrix}X_{\mathcal D}&B\\ B^ {\mathpalette\@transpose{}} &A(\Gamma)\end{smallmatrix}\right ]. We claim that

We have X_{\mathcal D}=J_v-I_m\otimes J_n, so X_{\mathcal D}^2=(v-2n)J_v+nI_m\otimes J_n and

Using r-\lambda_1=n(\lambda_2-\lambda_1), the (1,1)-block of the left-hand side of (Proof) is therefore (v-n+\lambda_1)J_v=\alpha_1^2J_v. By Remark 2.2 and (2.3), the (1,2)-block of left-hand side of (Proof) is

The (2,1)-block follows by transposition, and the (2,2)-block is exactly (2.3). This proves (Proof).

The complete multipartite graph on the point vertices is connected, since m\geqslant2, and every block vertex has k>0 point neighbours. Hence the graph with adjacency matrix A is connected. Set \bm\alpha^ {\mathpalette\@transpose{}} = [\alpha_1\mathbf1_v^ {\mathpalette\@transpose{}} ,\alpha_2\mathbf1_b^ {\mathpalette\@transpose{}} ]. Then (Proof) becomes

Since A commutes with the right-hand side, it preserves \langle\bm\alpha\rangle. The Perron–Frobenius theorem [BrouwerHaemers2012, Theorem 2.2.1] therefore identifies \bm\alpha as a Perron eigenvector, with simple eigenvalue \theta_0. On \bm\alpha^\perp, we have (A-(\lambda_2-\lambda_1) I)(A+nI) = O, so all remaining eigenvalues belong to \{\lambda_2-\lambda_1,-n\}. Both must occur, otherwise A would have a quadratic minimal polynomial, forcing (A^2)_{xy}=0 for every pair of distinct nonadjacent vertices x,y. However, two distinct point vertices in the same group are nonadjacent and have at least v-n>0 common neighbours. Moreover,

and \theta_0>0, so \theta_0>\lambda_2-\lambda_1.

The point–block partition is equitable, with quotient matrix

Its positive eigenvector (\alpha_1,\alpha_2)^ {\mathpalette\@transpose{}} has eigenvalue \theta_0, which is therefore the larger root of its characteristic polynomial. Thus

The row sums of Q are v-n+r and k_\Gamma+k, respectively. These give the two vertex degrees, proving the final assertion.

By (?), the matrix B^{ {\mathpalette\@transpose{}} }B has the spectrum [rk]^1,[r-\lambda_1]^{m(n-1)},[0]^{b-m(n-1)-1}.

Multiplying (2.3) on the left by B^ {\mathpalette\@transpose{}} and taking transposes, we obtain A(\Gamma)B^{ {\mathpalette\@transpose{}} }B = (\lambda_2 - \lambda_1 - n)B^{ {\mathpalette\@transpose{}} }B + \frac{k^2}{v}(k_{\Gamma} + \lambda_1 - \lambda_2 + n) J. We consider the eigenvalue r-\lambda_1 of B^{ {\mathpalette\@transpose{}} }B which has multiplicity m(n-1). Any eigenvector in this eigenspace is orthogonal to \mathbf{1}. Hence, A(\Gamma) has the eigenvalue \lambda_2 - \lambda_1 - n on this eigenspace.

Now we use (2.3) on the nullspace of B^{ {\mathpalette\@transpose{}} }B. This eigenspace (for eigenvalue 0) is orthogonal to \mathbf{1}. Hence, the eigenvalue of A(\Gamma) on this eigenspace satisfies the quadratic equation \theta^2 - (\lambda_2 - \lambda_1 - n)\theta - n(\lambda_2 - \lambda_1) = 0. Thus, the corresponding eigenvalues of A(\Gamma) are \lambda_2 - \lambda_1 and -n.

Therefore, A(\Gamma) has the spectrum [k_\Gamma]^1,[\lambda_2-\lambda_1-n]^{m(n-1)},[\lambda_2-\lambda_1]^x,[-n]^y which implies 1 + m(n-1) + x + y = b. Furthermore, since \operatorname{tr}{A(\Gamma)} = 0, we have k_{\Gamma} + m(n-1)(\lambda_2 - \lambda_1 - n) + x(\lambda_2 - \lambda_1) + y(-n) = 0. One can then solve to find x = \frac{n(b-1) - k_{\Gamma} - m(\lambda_2 - \lambda_1)(n-1)}{\lambda_2 - \lambda_1 + n} and y = \frac{(\lambda_2 - \lambda_1)(b-1) + k_{\Gamma} - nm(n-1)}{\lambda_2 - \lambda_1 + n}, as required. Finally, \operatorname{tr}(A(\Gamma)^2)=bk_\Gamma because \Gamma is k_\Gamma-regular. Evaluating this trace from the spectrum proves (2.5).

By Lemma 2.5 with \theta_1 = n and \theta_2 = -n, we have the stated spectrum, since \theta_1+\theta_2 = 0 and \theta_1-\theta_2 = 2n. Adding and subtracting the expressions for x and y in Lemma 2.5, one obtains the stated sum and difference. Since x and y are integers, so is y-x = k_\Gamma/n.

By Remark 2.4, k_\Gamma=\lambda_2-n=\lambda_1 or

In the latter case, Remark 2.4 also implies v-n>n, so v=mn implies m>2.

Multiplicativity implies that \theta_1+\theta_2 = 0 and \theta_1^2 = \theta_2^2 = n^2, so the trace identity of Lemma 2.5 reads k_\Gamma^2 + n^2(x+y) = bk_\Gamma; substituting x+y = b-1-m(n-1) and solving for m yields the stated formula.

Using (?), we find that \lambda_1=0, \lambda_2=n, r=n^2, k=m and b=n^3. Suppose that k_\Gamma\ne0. By Lemma 3.1, we have m>2 and k_\Gamma=2n(m-1)/(m-2), and (?) becomes

Its discriminant is (n^2-n-3)^2-8. This is negative for n=2; for n\geqslant4 it lies strictly between the consecutive squares (n^2-n-4)^2 and (n^2-n-3)^2. Thus n=3, and (Proof) implies m=3 or 4. If m=3, then k_\Gamma=12 and both full-graph degrees are 15, contrary to biregularity. If m=4, then k_\Gamma=9; by Lemma 3.1, we have x+y=18 and y-x=3, contrary to integrality of the multiplicities.

Hence k_\Gamma=0. By Lemma 3.1, we have m=(n^3-1)/(n-1)=n^2+n+1. Substitution provides the remaining parameters.

Suppose first that the nonprincipal quotient eigenvalue is s^2. Taking n=s^2 and \lambda_2=s^2(s+1) in (?), we have m^2-(s^3-2s+1)m+2s(s^2-1)=0. The discriminant is D_s=s^6-4s^4-6s^3+4s^2+4s+1. For s=2 it is negative. For s\geqslant3, we have (s^3-2s-4)^2<D_s<(s^3-2s-3)^2, since the differences are 2s^3-12s-15>0 and 8(s+1)>0, respectively. Thus D_s is not a square, contradicting m\in\mathbb{Z}.

The nonprincipal quotient eigenvalue is therefore -s^2. By Remark 2.4, k_\Gamma=\lambda_2-s^2=s^3. By Lemma 3.1, we have m=s^3+1 and

These yield the stated spectrum.

Since A(G_q)\mathbf1=q^3\mathbf1, (4.2) implies \left(B-\frac1qJ\right)\mathbf1=0, and hence B\mathbf1=q^3(q+1)\mathbf1. For the indicator \mathbf g of any one of the m groups, (4.2) implies

Hence,

Furthermore, expanding (4.2) using the row sums of B yields

These identities give the claimed GDD parameters, which satisfy r>\lambda_1 and rk=v\lambda_2. Thus \mathcal D is semiregular.

Let E_{\ker} be the orthogonal projection onto \ker A(G_q). By (4.2),

The matrix in parentheses is an orthogonal projection of rank m(q^2-1). Consequently,

and all nonzero squared singular values of B-\frac1qJ equal q^4. By (4.2) and the symmetry of A(G_q), the row space of B-\frac1qJ is contained in \ker A(G_q). By Lemma 4.1, we have \dim\ker A(G_q)=m(q^2-1). Thus, \ker A(G_q) is equal to the row space of B-\frac1qJ. It follows that

Since B^ {\mathpalette\@transpose{}} \mathbf1=qm\mathbf1 and v=mq^2, we have \left(B-\frac1qJ\right)^ {\mathpalette\@transpose{}} \mathbf1=0. Hence,

By Lemma 4.1, we have

Together with (4.2) and the q^3-regularity of G_q, these identities imply

Here \lambda_2-\lambda_1=n=q^2, so these are precisely (2.3) and (2.3). Finally, (2.3) becomes

which is satisfied. Thus G_q is an auxiliary graph for \mathcal D.

Let W have the vectors of an h-clique as its rows, and let D be the adjacency matrix of the graph recording inner product -54. Then

This matrix is positive semidefinite and has rank at most \dim E_3=18, so every eigenvalue of D is at most 2. In each connected component of D, an eigenvalue 2, if present, is the Perron eigenvalue and is simple. Such a component has at least three vertices. The nullity of 108I_h-54D is therefore at most h/3, whence

Thus h\leqslant27. If h=27, equality forces nine connected components of order three, each with spectral radius 2; each is therefore a triangle.

Within each triangle, the sum of its three vectors has squared norm 3\cdot108+6(-54)=0, so that sum is zero. The three corresponding rows of (W+J)/3 therefore sum to \mathbf1_{54}^ {\mathpalette\@transpose{}}. Thus every column has sum 9. Every row has sum 18, and the prescribed inner products give intersection 0 within each group and 6 between groups. These are the stated GDD parameters. Finally, 18>0 and 18\cdot9=27\cdot6, so the design is semiregular.

By direct computation

The first two lines give the stated GDD parameters, with the indicated groups. The design is semiregular since 18>0 and 18\cdot9=27\cdot6. The last two lines are the auxiliary-graph identities; the coefficient of J_{54} in (2.3) is ((15-6)(15+3)+18\cdot9)/54=6. Theorem 2.3 now applies. The quotient matrix \left[\begin{smallmatrix}24&18\\9&15\end{smallmatrix}\right] has eigenvalues 33 and 6, and its row sums give the two degrees. The eigenvalue multiplicities follow from the order of \Gamma_{81} and the trace of its adjacency matrix.

Set W=2B-J and H=I_{15}\otimes J_4. The row constraints give

For the indicator \mathbf g of any group of four rows, the last identity gives \|W^ {\mathpalette\@transpose{}} \mathbf g\|^2=0. Thus each group of sign rows sums to zero, and every column of B contains two ones in each group. Consequently,

These give the GDD parameters and semiregularity. The rank of WW^ {\mathpalette\@transpose{}} is 45, so the rows of W span E_4=\ker(A(G)-4I). The nonzero eigenvalues of WW^ {\mathpalette\@transpose{}} all equal 128, so W^ {\mathpalette\@transpose{}} W=128P_4, where P_4 is the orthogonal projection onto E_4. Since W^ {\mathpalette\@transpose{}} \mathbf1=0 and B=(W+J)/2, we have

Together with (?), this gives A(G)^2-4A(G)+B^ {\mathpalette\@transpose{}} B=32I+18J. The coefficient in (2.3) is ((20+16-24)(20+4)+48\cdot30)/96=18, so Theorem 2.3 applies. The quotient matrix \left[\begin{smallmatrix}56&48\\30&20\end{smallmatrix}\right] has eigenvalues 80,-4 and row sums 104,50; the multiplicities follow from the order and trace of the full adjacency matrix.

By Lemma 7.1, the graphs L_{q,a} are strongly regular with parameters independent of a. Hence products of the within-copy relations have the same coefficients in every copy. Products involving \mathsf A_{\neq} depend only on the common order and valency of these graphs. Thus the span of \mathcal A_q^\prime is closed under multiplication, so \mathcal A_q^\prime is a symmetric three-class association scheme.

Its common eigenspaces are the span of \mathbf1; the space \{\mathbf g\otimes\mathbf1_{q^4}:\sum_a g_a=0\} of dimension q; and the two spaces

On the copy-constant space orthogonal to \mathbf1, the matrices \mathsf A_{\neq} and \mathsf A_3 act as -q^4 and q^3-q, respectively. On both the spaces \mathfrak U_1 and \mathfrak U_2, the matrix \mathsf A_{\neq} acts as zero. For C\in\mathcal P, let R_{a,C} join distinct points of \Omega_a on a common copied line \{a\}\times L, where L\in C. Then A(L_{q,a})=\sum_{C\in\mathcal S_a}R_{a,C}. On U_a(C'), the matrix R_{a,C} acts as q^2-1 when C=C' and as -1 otherwise: lines of distinct classes meet in one point, so a vector in U_a(C') sums to zero on every line of C. Summing over the q used classes, \mathsf A_3 therefore acts as -q on the free-class space and as q^2-q on the used-class space. Their dimensions are (q+1)(q^2-q+1)(q^2-1)=(q^2-1)(q^3+1) and (q+1)q(q^2-1), respectively. The entries for \mathsf A_4 follow from I + \mathsf A_{\neq}+\mathsf A_3+\mathsf A_4 = J.

We first verify the products involving \mathsf A_1. These are instances of the matched parallel-class calculations in [Muzychuk2007, Proposition 2.2]:

Every vertex has a line of q^2 neighbours in each of the q other copies, so its degree is q^3. Two distinct vertices in the same copy have common neighbours in another copy precisely when they lie on the same line of the class assigned to that copy. Each selected class is used once, so their number of common neighbours is q^2 when their joining line is in a selected class and zero otherwise. For vertices in distinct copies, each of the q-1 remaining copies contributes the intersection of two lines from distinct parallel classes, hence one common neighbour. This proves (Proof).

For (Proof), take vertices x\in\Omega_a and y\in\Omega_b, where a\ne b. The neighbours of x in \Omega_b form a line L in a selected class. If y\in L, then the q^2-1 other points of L are adjacent to y in L_{q,b}, and there are no further contributions. If y\notin L, each of the other q-1 selected classes supplies exactly one point of L adjacent to y. These points are distinct. The product is zero when the two vertices lie in the same copy, proving the identity.

Since \mathsf A_2 = \mathsf A_{\neq} - \mathsf A_1 and \mathsf A_4 = J - I - \mathsf A_{\neq} - \mathsf A_3, and since \{I, \mathsf A_{\neq}, \mathsf A_3, \mathsf A_4\} spans the Bose–Mesner algebra of \mathcal{A}_q^\prime, (Proof) and (Proof), together with

show that the span of \mathcal A_q is closed under multiplication. Here, (?) follows from the fact that every vertex has q^2 neighbours in each other copy of \mathbb{F}_{q^2}^2. Thus \mathcal{A}_q is a four-class symmetric association scheme.

Now we determine its first eigenmatrix. By (Proof), each eigenvalue \theta of \mathsf A_1 satisfies \theta^2 = q^3 + (q-1)\alpha + q^2\beta, where \alpha and \beta are the eigenvalues of \mathsf A_{\neq} and \mathsf A_3 on the corresponding idempotent of \mathcal A_q'. Using the values of Lemma 7.2, we find

Thus the eigenvalues of \mathsf A_1 on the column spaces of \mathsf E_1 and \mathsf E_3^\prime lie in \{q^2,-q^2\}.

To determine the sign on \mathsf E_1, let \mathbf x = \mathbf g \otimes \mathbf{1}_{q^4} with \sum_{a=0}^q g_a = 0, as in Lemma 7.2. By construction, for fixed distinct a,b and fixed \mathbf v there are exactly q^2 vectors \mathbf w with \mathsf A_1[(a,\mathbf v),(b,\mathbf w)] = 1, so

Hence \mathsf A_1 \mathsf E_1 = - q^2\mathsf E_1, and \mathsf E_1 does not split.

Consequently the image of \mathsf E_3^\prime decomposes into its q^2- and -q^2-eigenspaces, of dimensions m_3,m_4, initially possibly zero. Since \operatorname{tr} \mathsf A_1 = 0,

so m_3 = m_4; and m_3 + m_4 = \operatorname{rank}(\mathsf E_3^\prime) = q(q+1)(q^2-1) implies that m_3 = m_4 = \frac{q(q+1)(q^2-1)}{2}.

The dimensions are positive, so both eigenspaces occur. Their projections are the remaining primitive idempotents \mathsf E_3,\mathsf E_4, since the algebra has dimension five. The remaining entries follow by linearity: since \mathsf A_2 = \mathsf A_{\neq} - \mathsf A_1, its eigenvalues are \alpha - \theta; and since \sum_{j=0}^4 \mathsf A_j = J, the eigenvalues of \mathsf A_4 on the non-principal idempotents are -1 - \theta - (\alpha - \theta) - \beta = -1 - \alpha - \beta.

By Theorem 7.3 the coherent closure \mathcal W(G_q) is contained in \operatorname{span}_{\mathbb{R}}\{I,\mathsf A_1,\mathsf A_2, \mathsf A_3,\mathsf A_4\}. To prove the reverse inclusion, put H=\mathsf A_1^2-q^3I. By (Proof), H=(q-1)\mathsf A_{\neq}+q^2\mathsf A_3. Since \mathsf A_{\neq} and \mathsf A_3 have disjoint binary supports,

Both matrices therefore belong to \mathcal W(G_q). The identities

show that the remaining basis matrices also belong to the coherent closure. This proves equality.

Write A=A(\Gamma) and let d_1\ne d_2 be the two degrees. The diagonal matrix D=A^2\circ I belongs to \mathcal W(\Gamma), so the diagonal indicator matrices

of V_1,V_2 belong to it. Restricted to V_i\times V_i, the corner F_i\mathcal W(\Gamma)F_i is a coherent algebra containing A(\Gamma[V_i]). Its dimension is therefore at least \operatorname{cr}(\Gamma[V_i]). Each off-diagonal corner contains the two linearly independent matrices B,J_{|V_1|,|V_2|} or their transposes. The four corners have disjoint supports, so their dimensions add, giving the stated bound.

Let m=q^3+1, v=mq^2, b=q^4(q+1). The adjacency matrix of \Gamma_q is

By Corollary 4.3, its point and block sets are its two degree classes. We exhibit a coherent algebra of dimension 12 containing A, then apply Lemma 7.5 to prove that it is the coherent closure.

Let

Here \mathcal L is the rank-three Bose–Mesner algebra of the complete multipartite graph on the points, and \mathcal R is the rank-five coherent closure of the auxiliary graph by Corollary 7.4.

Since G_q is an auxiliary graph of \mathcal D_q by Lemma 4.2, we can apply (2.3) to obtain

Using (Proof) and J=I+\mathsf A_{\neq}+\mathsf A_3+\mathsf A_4, we obtain

To check multiplication across the two degree classes, we use

The first identity follows from (2.3). Every row of M_q has q^3 ones in each copy, so M_q\mathsf A_{\neq}=q^4J, giving the second. For the third, each of the q used-class lines through a fixed vertex meets the row’s q selected free-class lines in q points. Removing the vertex itself subtracts q precisely when its incidence entry is one, giving q^2-qM_q. Since I+\mathsf A_3+\mathsf A_4 is the block-diagonal matrix with one all-ones block for each copy, and every row of M_q has q^3 ones in each copy, we have

Using the third identity, we obtain

These four identities show that M_qR\in\mathcal U for every R\in\mathcal R. Since each such R has constant column sum, we also have JR\in\operatorname{span}_{\mathbb R}\{J\}. As \mathcal U=\operatorname{span}_{\mathbb R}\{M_q,J\}, it follows that \mathcal U\mathcal R\subseteq\mathcal U.

By Remark 2.2, we have (I_m\otimes J_{q^2})M_q=qJ. Together with the constant row and column sums of M_q, one obtains \mathcal L\mathcal U\subseteq\mathcal U. By (?), we have \mathcal U\mathcal U^ {\mathpalette\@transpose{}} \subseteq\mathcal L, and by (?), we have \mathcal U^ {\mathpalette\@transpose{}} \mathcal U\subseteq\mathcal R. Thus

with its four blocks chosen independently, is closed under ordinary multiplication and transpose. It is also closed under Schur multiplication: its disjoint binary basis consists of I_v,I_m\otimes J_{q^2}-I_v,J_v-I_m\otimes J_{q^2} on the point side, I_b,\mathsf A_1,\ldots,\mathsf A_4 on the block side, and M_q,J-M_q and their transposes between the two sides, each extended by zero outside its corner. These twelve matrices are nonzero, and \mathcal C contains I,J,A. Hence \mathcal W(\Gamma_q)\subseteq\mathcal C and \dim\mathcal C=12.

The induced point graph has coherent rank 3, and the induced block graph has coherent rank 5 by Corollary 7.4. Also M_q\notin\{O,J\} by its row sums. By Lemma 7.5, we have \operatorname{cr}(\Gamma_q)\geqslant3+5+4=12. It follows that \mathcal W(\Gamma_q)=\mathcal C. Its two diagonal basis relations are the identities on the point and block sets, giving the asserted fibres and type.