Singular Irreducible M-Matrices

• Singular irreducible M-matrices are real square matrices with nonpositive off-diagonals and a one-dimensional kernel spanned by a positive vector.
• They are characterized by equivalent algebraic and combinatorial criteria, leveraging Perron–Frobenius theory and strong connectivity in their associated digraphs.
• Their applications span stability analysis in dynamical systems, Markov chain modeling, neural networks, and robust numerical preconditioning in graph-based computations.

A singular irreducible M-matrix is a real square matrix with nonpositive off-diagonal entries which satisfies a set of equivalent characterizations associated with both its algebraic and combinatorial structure, as well as powerful spectral properties. Singular irreducible M-matrices play a pivotal role in dynamical systems, Markov chain theory, numerical linear algebra, and stability analysis, particularly when standard invertibility assumptions are not tenable. Their analysis draws from the Perron–Frobenius theory for nonnegative and Metzler matrices, operator theory, and graph connectivity, and has driven recent extensions and applications in the study of complex networked systems, infinite-delay neural networks, and sparse linear algebra preconditioning.

1. Definitions and Characterizations

A real matrix $A\in\mathbb{R}^{n\times n}$ is a Z-matrix if all its off-diagonal entries are nonpositive: $a_{ij}\le 0$ for all $i\ne j$. An M-matrix is a special class of Z-matrices, defined by the existence of a decomposition

$A = sI - B, \quad B\ge 0, \quad s\ge\rho(B)$

where $\rho(B)$ is the spectral radius of $B$. If $s>\rho(B)$, $A$ is a nonsingular M-matrix; if $s=\rho(B)$, $A$ is a singular M-matrix.

Irreducibility (in the sense of nonnegative matrices) means there is no simultaneous row–column permutation that transforms $A$ into a block-upper-triangular form. For $A=sI-B$, irreducibility is equivalent to the strong connectivity of the digraph associated with $B$ (i.e., every index communicates with every other via a sequence of nonzero off-diagonal connections).

Key algebraic properties of singular irreducible M-matrices include:

• $\operatorname{rank}(A)=n-1$
• There exists $d>0$ such that $A d=0$
• Every principal $(n-1)\times(n-1)$ block is a nonsingular M-matrix (Bick et al., 25 Dec 2025)

Algorithmic and spectral criteria for singularity include the row-diagonal dominance test: an irreducible, diagonally dominant M-matrix $A$ is singular if and only if

$$a_{ii} = \sum_{j\ne i} |a_{ij}| \quad \text{for all } i$$

i.e., every row achieves equality in the diagonal-domination inequality (Jin, 29 Nov 2025). Geometric conditions for singularity in the more general complex case are given in terms of a solvable angle equation system associated with the phases of the off-diagonal entries.

2. Foundational Theorems and Perron–Frobenius Structure

The spectral theory underpinning singular irreducible M-matrices is closely tied to the Perron–Frobenius theorem for nonnegative irreducible matrices. If $A=\alpha I - B$ is a singular irreducible M-matrix ($\alpha=\rho(B)$), then:

• $\rho(B)$ is a simple eigenvalue of $B$ with strictly positive right and left eigenvectors.
• $A d=0$ and $d>0$ (nullspace is spanned by a positive vector).
• The zero eigenvalue is simple, and the group inverse $A^\#$ exists.

This guarantees that the kernel and image of $A$ intersect trivially on the nonnegative cone, implying that the only nonnegative vector in the range of $A$ is the zero vector. This property (trivial range monotonicity) has operator-theoretic analogues in the context of the Lyapunov and Stein operators (Encinas et al., 2023).

Rescaling variables using the positive null eigenvector yields a transformation in which the original (possibly non-invertible) dynamics can be reduced to a globally contracting system, a key step in stability theory for delay and networked systems (Oliveira et al., 16 Feb 2025).

3. Singular Irreducible M-Matrices in Dynamical Systems

Singular irreducible M-matrices are essential in stability theory for both continuous- and discrete-time dynamical systems subject to constraints, time delays, and nonautonomous evolution:

• In discrete-time, nonautonomous Hopfield neural networks with infinite distributed delays (including delays in leakage terms), a matrix $\mathcal{M}^+$ forms from system coefficients and activation function bounds. If $\mathcal{M}^+$ is a singular irreducible M-matrix and the activation functions are sublinear, then all solutions are globally attractive:

  $\lim_{m\to\infty} \|x(m)\|=0$

• The irreducibility condition ensures that the directed interaction graph of the system is strongly connected, enforcing a single global mode of decay (Oliveira et al., 16 Feb 2025).

Historically, global attractivity criteria in neural networks or multi-agent systems have required nonsingular M-matrix conditions, i.e., invertibility, to guarantee contractiveness and exponential decay. The relaxation to allow singular irreducible M-matrices—leveraging the simplicity and positivity of the kernel—enables analysis in settings where standard invertibility is precluded (e.g., in systems with mass conservation or network Laplacians).

4. Applications: Markov Chains, Laplacians, and Preconditioning

Singular irreducible M-matrices naturally arise as generators of Markov chains and as graph Laplacians. In these cases, the row sums of $A$ are zero, enforcing conservation laws. This structure is exploited in numerical linear algebra for constructing robust preconditioners:

• The Factorized Sparse Approximate Inverse (FSAI) method applies to singular irreducible M-matrices by careful management of the nonzero pattern, specifically omitting one off-diagonal entry to avoid inconsistency arising from the rank deficiency (Bick et al., 25 Dec 2025).
• The resulting factors $L_G$ and $U_G$ are nonsingular, entrywise nonnegative, and the preconditioned matrix remains a singular M-matrix with strictly positive diagonal.
• A (1,2)-inverse of $A$ can be explicitly constructed, providing stable pseudo-inverse preconditioning even under singularity.

In practical computations, these techniques retain the qualitative properties (e.g., positivity, singular M-matrix structure) needed for stable iterative solution of linear systems arising in Markov modeling and spectral graph theory.

5. Operator Theory: Lyapunov and Stein Analogues

The fundamental monotonicity property—only the zero vector in the intersection of the nonnegative cone and the image of a singular irreducible M-matrix—has direct analogues at the operator level:

• For the Lyapunov operator $\mathcal{L}_A(X)=AX+XA^T$ and Stein operator $\mathcal{S}_A(X)=X-AXA^T$, operator-level trivial range monotonicity holds under specific spectral hypotheses, such as $A^2=-I$ or $A^T=-A$ (Encinas et al., 2023).
• The key structural ingredients remain: index-1 (group-invertible) operator, strong connectivity (irreducibility), and complementarity of the cone and the kernel.
• Outside these specific spectral classes, counterexamples demonstrate the necessity of the classical irreducibility and spectral radius conditions.

These results further connect M-matrix theory to stability analysis of matrix equations, control theory, and symmetric cone complementarity.

6. Examples and Algorithmic Criteria

The distinction between reducible and irreducible singular M-matrices is critical:

• Reducible singular M-matrices correspond to systems with disconnected dynamics or multiple invariant subspaces; global attractivity and monotonicity may fail for unbounded activations or when the coupling structure is not strongly connected.
• Example: A $3\times3$ cycle graph Laplacian

$$A = \begin{pmatrix} 2 & -1 & -1 \ -1 & 2 & -1 \ -1 & -1 & 2 \end{pmatrix}$$

is a singular irreducible M-matrix, as it is exactly diagonally dominant and irreducible; its kernel is spanned by the all-ones vector.

• The singularity test for irreducible diagonally dominant M-matrices is algorithmic: check if all rows satisfy equality in the dominance inequality (Jin, 29 Nov 2025).

Property	Singular Irreducible M-Matrix	Reducible Singular M-Matrix
Nullspace	1-dimensional, positive vector	Higher-dimensional, may lack positivity
Graph theoretic structure	Strongly connected digraph	Block-diagonalizable, disconnected graph
Stability in unbounded case	Ensured by monotonicity property	Typically fails

7. Broader Implications and Research Directions

The theory of singular irreducible M-matrices broadens the scope of classical M-matrix methods to include systems exhibiting conservation, mass neutrality, or lack of dissipation, without sacrificing qualitative stability or robust computation:

• The relaxation of invertibility allows rigorous global attractivity results for discrete-time, non-autonomous, and infinite-delay systems previously inaccessible (Oliveira et al., 16 Feb 2025).
• Operator-theoretic perspectives deepen connections to Lyapunov/Stein complementarity and matrix equation theory (Encinas et al., 2023).
• Sparse preconditioning algorithms such as FSAI remain applicable, permitting efficient numerical solution of systems with singular M-matrix structure, thus supporting applications in Markov modeling, network dynamics, and spectral graph analysis (Bick et al., 25 Dec 2025).
• Necessary and sufficient conditions for singularity in diagonally dominant (irreducible) matrices facilitate structural analysis in multi-agent, networked, and control systems (Jin, 29 Nov 2025).

The continued synthesis of algebraic, spectral, and combinatorial theories around singular irreducible M-matrices is central to modern developments in stability theory, operator analysis, and computational mathematics.