On the Adjacency Spectra of Hypertrees

We extend the results of Zhang et al. to show that $\lambda$ is an eigenvalue of a $k$-uniform hypertree $(k \geq 3)$ if and only if it is a root of a particular matching polynomial for a connected induced subtree. We then use this to provide a spectral characterization for power hypertrees. Notably, the situation is quite different from that of ordinary trees, i.e., $2$-uniform trees. We conclude by presenting an example (an $11$ vertex, $3$-uniform non-power hypertree) illustrating these phenomena.


Introduction
The following beautiful result was shown in [7]: the set of roots of a certain matching polynomial of a k-uniform hypertree (an acyclic k-uniform hypergraph) is a subset of its homogeneous adjacency spectrum. In the present work, we show how to obtain all of the eigenvalues of a hypertree, and use this description to give a spectral characterization of "power" hypertrees (defined below). Here the notion of a hypergraph's eigenpairs is the homogeneous adjacency spectrum,à la Qi [6], Lim [4], and Cooper-Dutle [1]. We extend Theorem 1 as follows to describe the spectrum of a hypertree, answering the main open question in [7]. Theorem 2 Let H be a k-uniform hypertree, for k ≥ 3; λ is a nonzero eigenvalue of H if and only if there exists an induced subtree H ⊆ H such that λ is a root of the polynomial ϕ(H).
We note that Theorem 2 is not true for (2-uniform) trees (c.f. Cauchy's interlacing theorem), making it an unusual example of a result in spectral hypergraph theory that fails in the graph case. The necessity for k ≥ 3 is established by Theorem 7 ([7], [8]).
Below, we reserve the use of the term "hypergraph" for the case of k-uniform hypergraphs with k ≥ 3, and use the language of "graphs" exclusively for k = 2. We also make use of the nomenclature k-graph and k-tree to mean k-uniform hypergraph and k-uniform hypertree, respectively. We maintain much of the notation of [7] and refer the interested reader to their paper. In the next section, we provide necessary definitions and prove Theorem 2. We then use this result to show that power trees are characterized by their spectra being cyclotomic (Theorem 9) and provide an example demonstrating these phenomena.

Proof of Theorem 2
A vector is totally nonzero if each coordinate is nonzero and an eigenpair (λ, x) is totally nonzero if λ = 0 and x is a totally nonzero vector. Given a vector x ∈ C n the support supp(x) is the set of all indices of non-zero coordinates of x. Let x • denote the totally nonzero projection (by restriction) of x onto C |supp(x)| . For ease of notation, we assume that the coordinate indices of vectors agree with the vertex labeling of the hypergraph under consideration. We denote the induced subgraph of H on U ⊆ V (H) by and write H H to mean H = H[U ] for some U ⊆ V (H).
The following establishes the forward direction of Theorem 2. Proof: As (λ, x) is an eigenpair of H, The following result, from [1], shows how the eigenvalues of a disconnected hypergraph arise from the eigenvalues of its components.
The matching polynomial ϕ(H) of H is defined by where M i is the collection of all t-matchings of H. We now show that ϕ is multiplicative over connected components.
. Proof: Clearly, the result follows inductively if it is true for t = 2. Denote matching numbers m(H 1 ) = m 1 and m(H 2 ) = m 2 . Indexing the sum of H 1 and H 2 by i and j, respectively, we compute Since matchings of H are unions of matchings of H 1 and H 2 , m = m 1 + m 2 is the size of the the largest matching of H. Furthermore, for any 0 ≤ ≤ m, since a matching of edges in H consists of a matching of i edges in H 1 and j edges in H 2 , where i + j = . Substituting yields as desired.
Recall from [7] that a pendant edge is a k-uniform edge with exactly k − 1 vertices of degree 1. If H is connected we apply Theorem 1 to conclude that λ is a root of ϕ(H) as desired. If instead H is disconnected, then we may write As λ is a root of ϕ(H), it is true that λ is a root of ϕ(H i ) for some i. Appealing to Theorem 1, once more we have that (λ, x • ) is a totally nonzero eigenpair of H i ; therefore, (λ, x) is an eigenpair of H by Claim 6 and Theorem 7 as desired.

The spectra of power trees
The following generalizes the definition of powers of a hypergraph from [2].
In other words, one adds exactly enough new vertices (each of degree 1) to each edge of H so that H k is k-uniform. Note that, if k = r, then H k = H. Adhering to this nomenclature we refer to a power of a 2-tree simply as a power tree. In this section we prove the following characterization of power trees.
Theorem 9 Let H be a k-tree. Then σ(H) ⊆ R[ζ k ] if and only if H is a power tree, where ζ k is a principal k th root of unity.
We recall the following Theorem from Cooper-Dutle.
Theorem 10 [1] The (multiset) spectrum of a k-cylinder is invariant under multiplication by any k th root of unity.
One can show by straightforward induction that a k-tree is a k-cylinder, so its spectrum is symmetric in the above sense. The following result, from [8], shows that power trees have spectra which satisfy a much more stringent condition: they are cyclotomic, in the sense that they belong to R[ζ k ].
Theorem 11 [8] If λ = 0 is an eigenvalue of any subgraph of G, then λ 2/k is an eigenvalue of G k for k ≥ 4.
We restate Theorem 11 with the additional assumption that the underlying graph is a tree; the proof is easily obtained by applying Claim 6 to the proof of Theorem 11 appearing in [8].
Corollary 12 If λ = 0 is an eigenvalue of any subgraph of a tree T , then λ 2/k is an eigenvalue of T k for k ≥ 3.
Note that Theorem 9 provides a converse to Corollary 12 in the case of power trees. In particular, appealing to Theorem 2, Corollary 12, and the fact that the spectrum of a graph is real-valued, we have that the spectrum of a power tree is a subset of R[ζ k ]. All that remains to be shown is that if a k-tree is not a power tree then it has a root in C \ R[ζ k ]. To that end, we introduce the k-comb.
Let comb k be the k-graph where We refer to comb k as the k-comb. By the definition of power tree, a non-power tree H must contain an edge e incident to a family F consisting of at least three other edges which are mutually disjoint. This edge e, together with F, form a connected induced subgraph H of H which is the k th power of a t-comb for t = |F| ≥ 3. It is straightforward to see that ϕ(H )(x) = ϕ(comb k t )(x) = ϕ(comb t )(x k/t ), since matchings in H are simply k th powers of matchings in comb t ; therefore, roots of ϕ(H ) are k th roots of reals if and only if the roots of ϕ(comb t ) are t th roots of reals. We presently show that the spectrum of the k-comb is not contained within the k th cyclotomic extension of R, completing the proof of Theorem 9.
Lemma 13 There exists a root λ of ϕ(comb k ) for k ≥ 3 such that λ ∈ C \ R[ζ i ].
Proof: Let H be a k-comb where k ≥ 3. By a simple counting argument, Appealing to the binomial theorem we have It is easy to see that (1) has precisely one solution when k ≥ 3 is odd and precisely two solutions when it is even. In either case, as the number of solutions is strictly less than k it follows that there must be a non-real solution and the claim follows.

Concluding Remarks
We conclude our note by presenting an example demonstrating Theorems 2 and 9.
Consider the 3-uniform hypergraphs We have computed Let P n and S n denote the 3-uniform loose path and star with n edges, respectively. We list the non-trivial induced subgraphs of H 3 and their matching polynomials below.
H H 3 ϕ(H) x 9 − 5x 6 + 5x 3 − 2 Figure 1 gives a drawing of H 3 (the striped subgraph is H 1 ) and a plot of the roots of φ(H 3 ), with a circle centered at each root in the complex plane whose area is proportional to the multiplicity of the root. Notice that, despite the rotational symmetry (turning by a third), the cubes of the roots are not all real, i.e., some of the roots do not lie on the the rays with argument 0, 2π/3, or 4π/3. Observe that each matching polynomial divides wholly into the characteristic polynomials. A priori, this is a symptom of the matching polynomials having distinct roots. A preliminary question: what can one say about the roots of ϕ(H 1 ) and ϕ(H 2 ) for k-trees H 1 H 2 ? With this question in mind we conjecture the following.