Chapter 38B: P-matrices and H-matrices¶

Prerequisites: M-matrices (Ch38A) · Positive Definite Matrices (Ch16) · Determinants (Ch03)

Chapter Outline: Generalization of M-matrices → Definition of P-matrices (All Principal Minors Positive) → Solving Linear Complementarity Problems (LCP) → H-matrices (Comparison Matrix is an M-matrix) → Ostrowski Criterion → Hierarchical Relationships: Positive Definite \(\subset\) M-matrices \(\subset\) H-matrices → Applications in Stability and Numerical Computation: Circuit Analysis and Linearization of Non-linear Systems

Extension: P-matrices and H-matrices are the advanced algebraic language for "non-symmetric stability"; they preserve many key properties of positive definite matrices without requiring symmetry, serving as the core for solving complementarity problems in operations research.

If we relax the "non-positive off-diagonals" restriction of M-matrices, we obtain the more general H-matrices; if we only retain the "all principal minors positive" characteristic, we get P-matrices. These two classes of matrices are critical in non-linear programming and circuit simulation, as they guarantee the existence and uniqueness of equilibrium points in complex systems. This chapter reveals the hierarchical connections between these structures.

38B.1 P-matrices and the LCP¶

Definition 38B.1 (P-matrix)

A square matrix \(A\) is a P-matrix if every principal minor (not just leading ones) is strictly positive. Properties: Symmetric positive definite matrices and non-singular M-matrices are always P-matrices.

Application: Linear Complementarity Problem (LCP)

P-matrices provide the necessary and sufficient condition for the LCP (\(w - Az = q, w \ge 0, z \ge 0, w^T z = 0\)) to have a unique solution for any \(q\). This is invaluable in game theory and contact mechanics.

38B.2 H-matrices and Comparison Matrices¶

Definition 38B.2 (H-matrix)

A square matrix \(A\) is an H-matrix if its Comparison Matrix \(\mathcal{M}(A)\) is a non-singular M-matrix. The comparison matrix is defined as: - Diagonal: \((\mathcal{M}(A))_{ii} = |a_{ii}|\) - Off-diagonal: \((\mathcal{M}(A))_{ij} = -|a_{ij}|\) (\(i \neq j\))

38B.3 Criteria and Hierarchies¶

Theorem 38B.1 (Ostrowski Criterion)

If \(A\) is generalized strictly diagonally dominant (i.e., there exists a positive vector \(d\) such that \(|a_{ii}| d_i > \sum_{j \neq i} |a_{ij}| d_j\)), then \(A\) is a non-singular H-matrix.

Exercises¶

1. [Basics] Determine if \(A = \begin{pmatrix} 1 & -2 \\ 0 & 1 \end{pmatrix}\) is a P-matrix.

Solution

Calculate Principal Minors: 1. \(1 \times 1\) minors: \(|1|=1, |1|=1\) (both positive). 2. \(2 \times 2\) minor: \(\det(A) = 1 - 0 = 1 > 0\). Conclusion: Since all principal minors are positive, it is a P-matrix. Note: It is neither symmetric nor positive definite.

2. [Comparison] Find the comparison matrix for \(A = \begin{pmatrix} 2 & i \\ -1 & 3 \end{pmatrix}\).

Solution

Construction: 1. Absolute values of diagonals: \(2, 3\). 2. Negative absolute values of off-diagonals: \(-|i| = -1, -|-1| = -1\). Result: \(\mathcal{M}(A) = \begin{pmatrix} 2 & -1 \\ -1 & 3 \end{pmatrix}\).

3. [H-matrix Check] Is the matrix \(A\) from the previous problem an H-matrix?

Solution

Determination: 1. Check if \(\mathcal{M}(A)\) is an M-matrix. 2. It is a Z-matrix, and its principal minors are \(2 > 0\) and \(6-1=5 > 0\). Conclusion: Since the comparison matrix is a non-singular M-matrix, \(A\) is a non-singular H-matrix.

4. [Properties] Can the eigenvalues of a P-matrix have negative real parts?

Solution

Conclusion: Yes. Reasoning: P-matrices only guarantee positive minors, not necessarily positive real parts for eigenvalues (which is a stronger property held by M-matrices or PD matrices). Example: \(A = \begin{pmatrix} 1 & 2 \\ -2 & 1 \end{pmatrix}\) is a P-matrix (minors 1, 1, 5), but its eigenvalues are \(1 \pm 2i\). While these have positive real parts, in higher dimensions, more complex distributions are possible.

5. [Hierarchy] Briefly state the containment relationship between PD, M, and P matrices.

Solution

Inclusion Chain: (Symmetric PD) \(\cup\) (Non-singular M-matrices) \(\subset\) (P-matrices). P-matrices represent the broadest generalization of "determinantal positivity."

6. [Inversion] Is the inverse of an H-matrix always non-negative?

Solution

Conclusion: Not necessarily. Analysis: Only M-matrices guarantee \(A^{-1} \ge 0\). For a general H-matrix (e.g., with complex entries), the inverse is usually complex. However, H-matrices satisfy \(\|A^{-1}\| \le \|\mathcal{M}(A)^{-1}\|\), which is useful for error bounds.

7. [Calculation] Determine if \(\begin{pmatrix} 1 & 2 \\ 2 & 1 \end{pmatrix}\) is a P-matrix.

Solution

Calculation: Principal minors: 1, 1. But \(\det = 1 - 4 = -3 < 0\). Conclusion: It is not a P-matrix.

8. [Application] In circuit analysis, what type of matrix is the conductance matrix formed by resistors?

Solution

Typically an M-matrix (due to diagonal dominance from Kirchhoff's Current Law and negative conductances on off-diagonals). If dependent sources are present, it may generalize to an H-matrix.

9. [Stability] Why are H-matrices important in neural network stability?

Solution

Because non-linear activation functions are often bounded in slope. Using the Ostrowski criterion for H-matrices allows one to determine if the weight matrix guarantees a unique global equilibrium.

10. [Limit] As the degree of diagonal dominance increases, what does an H-matrix approach?

Solution

It approaches a diagonal matrix. Diagonal dominance is the core numerical trait of H-matrices; the stronger the dominance, the more the inverse resembles that of a diagonal matrix, increasing system stability.

Chapter Summary¶

P-matrices and H-matrices define the boundaries of generalized stability:

Dominance of Minors: P-matrices prove that even without symmetry, the overall positivity of principal minors ensures unique solutions to linear complementarity problems—an algebraic cornerstone of operations research.
Abstraction of Magnitudes: Through the comparison matrix technique, H-matrices simplify complex numerical (even complex-valued) operations into magnitude analysis using M-matrices, providing unified bounds for error propagation.
Evolution of Structure: From Positive Definite to M to H, this hierarchy demonstrates how linear algebra generalizes stability logic to non-linear and non-symmetric domains by step-wise relaxation of constraints.