Let \(\alpha(G)\), \(\mu(G)\) and prk\((G)\) denote the independence number, the matching number and the permanental rank of \(G\), respectively. Here prk\((G)\) is the maximum order of a principal submatrix with nonzero permanent of the adjacency matrix of \(G\). Let \(d(G)=\max_{S\subseteq V(G)}\{|S|-|N(S)|\}\) be the critical difference of \(G\). Let core\((G)\) and ker\((G)\) be the intersection of all maximum independent sets and all critical independent sets, respectively. In this note we use Larson’s critical independence decomposition to split the graph into two induced subgraphs, \(L_G\) and \(L_G^c\), where \(L_G\) is Kőnig–Egerváry and \(L_G^c\) is 2-bicritical. We prove that for every graph \(G\) one has \(\alpha(G)-\mu(G) = |L_G|-prk(L_G)+\alpha(L_G^c)-\mu(L_G^c) = d(L_G)+\alpha(L_G^c)-\mu(L_G^c).\) Moreover, we show that \(\alpha(L_G^c)\le \mu(L_G^c)\) and establish the refined kernel bound \(d(L_G)+k\le |ker(G)|,\) where \(k\) is the number of nontrivial connected components of \(L_G\) without a perfect matching. Consequently, \(\alpha(G)-\mu(G)+k\le |ker(G)|.\) In particular, when \(\alpha(G)>\mu(G)\), one has \(|L_G|>prk(L_G)\). The bound is sharp for every prescribed value of \(k\). Since ker\((G)\subseteq core(G)\) for every graph, we recover as a consequence the known Boros–Golumbic–Levit inequality \(\alpha(G)-\mu(G)+1\le |core(G)|\) for connected graphs with at least two vertices and \(\alpha(G)>\mu(G)\). This result improves on related results by Hammer et al. (1982) and by Levit and Mandrescu (1999).

Adversarial perturbations measured by ${\mathsf{\ell }}_p$ norms do not reflect key structural constraints in tabular security data, including anisotropic geometry, feature dependence, and distributional plausibility. We introduce a composite attacker cost functional $C_{\mathrm{atk}}(\mathit{x},{\mathit{x}}^{\prime })=\tau max\{0,m\left({\mathit{x}}^{\prime }\right)\}+{\lambda }_1{\mathit{\delta }}^{\top }{\mathit{G}}_{\perp }^{\left(\gamma \right)}\left(\mathit{x}\right)\mathit{\delta }+\lambda {\displaystyle \sum _j}{\omega }_j|{\delta }_j|+{\lambda }_2{\mathit{\delta }}^{\top }{\mathit{L}}_H\mathit{\delta }+{\lambda }_3\left[log{\widehat{f}}_1^{\left(ϵ\right)}\left(\mathit{x}\right)-log{\widehat{f}}_1^{\left(ϵ\right)}\left({\mathit{x}}^{\prime }\right)\right]+{\displaystyle \sum _{j\in supp\left(\mathit{\delta }\right)}}{c}_j+\beta {\left|\mathcal{M}(supp\left(\mathit{\delta }\right))\right|}^{\nu },$ which integrates a spectrally truncated geometric term, a graph-based coherence penalty, a smooth copula density barrier, and a superlinear module-spread term. Under spectral degeneracy of the legitimate-class covariance, we establish nonnegativity under density dominance, exact zero self-cost, lower semicontinuity, and ${\lambda }_3{\kappa }_K$-weak convexity of the continuous component on compact convex sets, for both affine and ${\rho }_m$-weakly convex scoring functions. These properties yield existence of constrained minimizers. The continuous component is locally Lipschitz, whereas the full functional is not due to the support-counting term. A component feasibility result shows that each term eliminates a distinct class of degenerate perturbations. Limiting regimes and refined evasion cost bounds are derived. An empirical instantiation on PHIUSIIL indicates that perturbations with identical ${\mathsf{\ell }}_2$ norm can incur costs differing by an order of magnitude.

We relate two decompositions that isolate König–Egerváry structure: the SD–KE decomposition, defined by the classical Edmonds–Sterboul–Deming obstructions, and Larson’s critical independence decomposition, whose complement is \(2\)-bicritical. Let \(I\) be a maximum critical independent set, set \(L(G)=I\cup N_G(I)\), \(L^c(G)=V(G)\setminus L(G)\), and \(B=N_G(I)\). We first prove a normal form for maximum matchings: each maximum matching is obtained from a maximum matching of \(G[L^c(G)]\), a matching crossing from a subset \(X\subseteq B\) into exposed vertices on the \(L^c(G)\) side, and a maximum matching of \(G[L(G)-X]\). The possible sets \(X\) are the crossing profiles. For a crossing profile \(X\) and a maximum matching \(P\) of \(G[L(G)-X]\), view \(P\) as a matching of \(G[L(G)]\). Let \(\Lambda_I(G)\) be the union of the vertices on \(P\)-augmenting paths in \(G[L(G)]\), over all crossing profiles and all residual maximum matchings. We prove \(SD(G)=L^c(G)\cup\Lambda_I(G),\) and \(KE(G)=L(G)\setminus\Lambda_I(G).\) Thus the SD vertices on the Larson König–Egerváry side are precisely those reached by augmenting corridors forced by admissible crossings.