arXiv++ Combinatorics

We study the change in the trace norm and the centred energy of the $A_α$-matrix of a graph under vertex deletion. A pinching argument proves that the trace norm is nondecreasing under passage from an induced subgraph to the original graph, and gives a lower bound in terms of the edges induced by the deleted vertices. Exact perturbation spectra yield sharp upper bounds for deleting a vertex or an independent set of vertices with a common neighbourhood. For $0<α<1/2$, equality in these upper bounds is characterized by star and complete bipartite components, respectively. For the centred energy, we derive explicit deletion bounds and classify all connected equality graphs for $0\leα<1$: the extremal graphs are stars and complete bipartite graphs for single-vertex and independent-twin deletion, respectively. We also characterize equality in the general-set trace-norm lower bound and prove strictness of the total centred-energy inequality for connected nonregular graphs when $0<α<1$. Examples show that centred local energy can be negative and that the degree deviation term cannot be omitted in general.

We establish local and global spectral bounds for Hermitian $A_α$-matrices of mixed graphs. Using the first three spectral moments, we obtain an upper bound for the largest eigenvalue as the largest real zero of an explicit cubic polynomial. Vertexwise estimates for the extreme eigenvalues yield new lower bounds for the spectral spread, including a bound that strictly improves an existing degree-based estimate for ordinary graphs. A local two-dimensional compression produces a spread bound involving neighbour-degree data and gain-weighted triangles. This bound is exact for every mixed orientation of a star and is independent of a known Zagreb-index bound. We also derive two complementary upper bounds for sums of the smallest eigenvalues. As further consequences, the spectral estimates provide a computable convergence guarantee for Richardson graph filtering and a stability certificate for residual graph-neural-network layers on directed networks. Numerical examples illustrate the sharpness and mutual incomparability of the proposed bounds.

Nikiforov\cite{nikiforov2017merging} introduced the concept of the $A_{\p}$-matrix as a convex linear combination of a graph's adjacency matrix and its diagonal matrix of vertex degrees. In this paper, we introduce a new variant of the $A_{\p}$-matrix, which is a linear combination of the Hermitian adjacency matrix of the second kind, $H^{\om}$, and the degree-diagonal matrix. This study provides new insights into the $A_{\p}$-matrix as a Hermitian matrix, leading to novel bounds in spectral graph theory.

This paper presents new lower bounds for the first general Zagreb index $Z_α(G)$ involving two, three, and four arbitrary degrees of vertices of a simple graph $G$. For the special cases $α= 2$ and $α= -2$, the results give sharper bounds for the first Zagreb index $M_1(G)$ and the modified first Zagreb index ${}^{m}M_1(G)$, thereby improving several well-known inequalities in the literature. Furthermore, some applications of the derived bounds for $M_1(G)$ are demonstrated, establishing new bounds for the second Zagreb index, the spectral radius, Nordhaus-Gaddum type bounds, and their corresponding coindices.