Natural graph spectra

In 2003, van Dam and Haemers posed a fundamental question in spectral graph theory: does there exist a ``sensible’’ matrix whose spectrum determines a random graph up to isomorphism? This paper introduces the class of {\em natural graph matrices}, which are matrices defined by applying a fixed sequence of elementary operations to the adjacency matrix. This class includes many standard matrices such as the adjacency matrix, the Seidel matrix, the Laplacian matrix, and the distance matrix. We give an affirmative answer to the question of van Dam and Haemers by proving the existence of a natural graph matrix whose spectrum determines random graphs up to isomorphism. The proof introduces a new algebraic framework called {\em double algebras}, which provides a simple sufficient condition for spectral determination. This sufficient condition is then shown to hold for random graphs.

💡 Research Summary

The paper addresses a long‑standing question raised by van Dam and Haemers (2003): does there exist a “sensible” matrix whose spectrum uniquely determines a random graph up to isomorphism? The authors answer affirmatively by introducing the class of natural graph matrices and developing a novel algebraic framework called double algebras.

Natural graph matrices are defined as matrices obtained from the adjacency matrix (A_G) of a graph (G) by applying a fixed sequence of three elementary operations: (1) linear combinations, (2) ordinary matrix multiplication (denoted ( \bullet )), and (3) entry‑wise (Hadamard) multiplication (denoted ( \circ )). The requirement that the same sequence be used for every graph guarantees uniformity; examples include the adjacency matrix itself, the Seidel matrix, the Laplacian, the signless Laplacian, and distance matrices for fixed‑size graphs. The spectrum of such a matrix, taken over the algebraic closure of the base field, is called a natural graph spectrum.

To formalize the notion of a fixed sequence and to study the resulting spectra, the authors introduce double algebras. A double algebra (R) over a field (\mathbb{F}) is a vector space equipped with two bilinear products ( \bullet ) and ( \circ ). Each product separately makes (R) into an ordinary algebra ((R_{\bullet}) and (R_{\circ})). The paper focuses on associative double algebras where (R_{\circ}) is commutative and split semisimple; under these conditions, (R_{\circ}) possesses a basis consisting of primitive (\circ)-idempotents (by the Wedderburn‑Artin theorem).

The authors construct a free double algebra (\mathbb{F}\langle!\langle x\rangle!\rangle) to model the fixed sequence of operations as a “double polynomial” in a formal variable (x). Evaluation at a concrete matrix (a) yields a homomorphism (\operatorname{ev}a) that substitutes (x) by (a). This machinery allows one to treat any natural graph matrix as the image of a double polynomial (p(x)) under (\operatorname{ev}{A_G}).

A central technical achievement is the existence of a universal (\circ)-idempotent basis for any finite family of (\circ)-subalgebras of a double algebra. Concretely, there exists a finite set (B) of double polynomials such that, for each graph (G) in a given family, the non‑zero evaluations ({b(A_G) : b\in B}) are distinct (\circ)-idempotents forming a basis of the subalgebra generated by the evaluated polynomials. This universal basis enables the definition of a natural graph spectrum (\operatorname{Spec}(G)=\operatorname{Spec}_\circ(p(A_G))) that satisfies a simple sufficient condition: if the double algebra generated by (A_G) equals the full matrix double algebra (M_n(\mathbb{F})), then (\operatorname{Spec}(G)) determines (G) up to isomorphism. This is Theorem 1.2.

The second major component of the paper proves that the sufficient condition holds asymptotically almost surely for Erdős–Rényi random graphs (G(n,1/2)). Using probabilistic arguments about the rank and invertibility of random matrices, the authors show that with probability tending to 1 as (n\to\infty), the double algebra (\mathbb{F}\langle!\langle A_G\rangle!\rangle) generated by the adjacency matrix of a random graph is the entire matrix algebra (M_n(\mathbb{F})). Consequently, for almost all large random graphs, the constructed natural graph spectrum uniquely identifies the graph. This is Theorem 1.3, and together with Theorem 1.2 it yields the main result (Theorem 1.1).

The paper also discusses why natural spectra cannot determine all graphs: if they did, one could solve the Graph Isomorphism problem via spectral comparison, contradicting current complexity expectations. Moreover, the authors note that more “complicated” natural spectra tend to be more discriminating, suggesting a research direction toward finding spectra that are both computationally tractable and highly informative (e.g., closer to classical adjacency or Laplacian spectra).

In summary, the authors provide:

1. A rigorous definition of “sensible” matrices via natural graph matrices.
2. A new algebraic framework (double algebras) that captures the interplay of matrix multiplication and Hadamard product.
3. A universal idempotent basis yielding a sufficient condition for spectral determination.
4. A probabilistic proof that this condition holds for almost all Erdős–Rényi graphs.
5. An affirmative answer to the van Dam–Haemers DS problem for random graphs, opening avenues for further exploration of natural spectra in graph theory, coding theory, and design theory.