Weakly Matrix Monotone Functions

• Weakly matrix monotone functions are scalar functions that relax classical matrix monotonicity via an α-scaled order preservation, applicable in both univariate and multivariate contexts.
• They underpin a weak diminishing returns property in spectral optimization, allowing greedy algorithms to achieve approximation ratios like 1 - e^(-α).
• Their extensions to noncommutative and operator system settings enable efficient eigenvalue updates, facilitating practical large-scale data subset selection and spectral clustering.

Weakly matrix monotone functions generalize classical matrix monotonicity by introducing a quantitative attenuation of the order-preserving property familiar from operator theory and matrix analysis. These functions and their set-function analogues appear in matrix spectral optimization, submodular maximization, and noncommutative analysis. The notion accommodates both one-variable and several-variable contexts and is closely tied to relaxed versions of the diminishing returns property, with precise approximation guarantees in optimization. The theory connects deeply with the Loewner order, analytic continuations into upper half-plane domains, and the structure of submodular matrix spectral objectives.

1. Formal Definitions and Variants

The archetypal definition, as presented in "How Much Is a Dataset Worth? Scaling Laws, the Vendi Score, and Matrix Spectral Functions" (Bilmes et al., 28 May 2026), considers a scalar function $\psi\colon \mathbb{R}_+ \to \mathbb{R}$ and introduces the notion of $\alpha$-weak matrix monotonicity. For $\alpha \in [0,1]$, $\psi$ is called $\alpha$-weakly matrix-monotone if for every pair of positive-semidefinite matrices $A, B$ with $0 \preceq A \preceq B$,

$\alpha \psi(A) \preceq \psi(B)$

where $\preceq$ denotes the Loewner order (i.e., $B - A$ is positive semidefinite). When $\alpha$0, this coincides with classical matrix monotonicity; for $\alpha$1, the requirement is relaxed, allowing a scaled violation of the monotonicity property.

For spectral set-functions $\alpha$2, where $\alpha$3 is a concave nondecreasing function with $\alpha$4 and $\alpha$5 is a principal submatrix or Gram matrix restriction, the weakly matrix monotone property is used for $\alpha$6, which being $\alpha$7-weakly matrix-monotone gives $\alpha$8 weakly submodular structure and controls the diminishing returns constant (Bilmes et al., 28 May 2026).

In the context of several variables, locally (or weakly) matrix monotone functions are defined along positive commuting directions. A real function $\alpha$9 (with $\alpha \in [0,1]$0 open) is locally matrix monotone if for every $\alpha \in [0,1]$1 and every C$\alpha \in [0,1]$2-path $\alpha \in [0,1]$3 in the space of $\alpha \in [0,1]$4-tuples of commuting self-adjoint matrices with joint spectrum in $\alpha \in [0,1]$5, and with $\alpha \in [0,1]$6, $\alpha \in [0,1]$7 is positive semidefinite for $\alpha \in [0,1]$8 (Agler et al., 2010, Pascoe, 2017, Pascoe, 2014).

2. Core Properties and Weak Diminishing Returns

Unlike fully matrix-monotone functions, which require $\alpha \in [0,1]$9 whenever $\psi$0, the weak variant only guarantees this up to a multiplicative factor $\psi$1. In matrix spectral objectives, this directly yields a notion of $\psi$2-weak diminishing returns (DR)-submodularity: for a set-function $\psi$3,

$\psi$4

for all $\psi$5 and $\psi$6. This generalized submodularity is strict for $\psi$7, while for $\psi$8, the standard greedy maximization procedure still enjoys provable approximation: $\psi$9 for monotone objectives. The practical effect is that, for $\alpha$0 near 1, the approximation guarantee remains strong; for example, $\alpha$1 yields a bound of $\alpha$2 (Bilmes et al., 28 May 2026).

3. Constructions and Examples

Explicit examples of $\alpha$3-weakly matrix-monotone functions $\alpha$4 include [(Bilmes et al., 28 May 2026), Lemma 3.7]:

• $\alpha$5 with derivative $\alpha$6, which gives weak matrix monotonicity for $\alpha$7 (spectrum dependent);
• $\alpha$8 with $\alpha$9, with $A, B$0, so for a normalized spectrum $A, B$1, $A, B$2;
• $A, B$3 with $A, B$4, again with $A, B$5.

Classical spectral objectives such as log-determinant and entropy scores correspond to fully matrix-monotone $A, B$6, while saturating and feature-based functions generally only achieve weak matrix monotonicity, hence offering only weakly submodular guarantees (Bilmes et al., 28 May 2026). By parameter adjustment, $A, B$7 can often be made arbitrarily close to 1 for practical settings.

4. Loewner Theorem, Analytic Structure, and Characterizations

Weakly matrix monotone (locally matrix monotone) functions in several variables possess deep analytic properties. The Agler–McCarthy–Young theorem establishes that real-valued $A, B$8 (or merely locally monotone) functions $A, B$9 are real-analytic and extend holomorphically to the poly-upper-half-plane, mapping $0 \preceq A \preceq B$0 into $0 \preceq A \preceq B$1, with $0 \preceq A \preceq B$2 the $0 \preceq A \preceq B$3-fold product of upper half-planes (Agler et al., 2010, Pascoe, 2014). This analytic continuation ensures that local (directional) monotonicity is sufficient for global analytic and positivity structure, which is essential for operator theory and matrix functional calculus.

For rational functions in two variables, operator (matrix) monotonicity is equivalent to explicit transfer function realizations, with the general form

$0 \preceq A \preceq B$4

where $0 \preceq A \preceq B$5, $0 \preceq A \preceq B$6, $0 \preceq A \preceq B$7, $0 \preceq A \preceq B$8 (Agler et al., 2010). Mollification arguments show that every locally matrix monotone function (with no differentiability a priori) is necessarily real-analytic (Pascoe, 2014), harmonizing multivariate and univariate theories.

The Loewner matrix criterion offers an alternative: For $0 \preceq A \preceq B$9, conditional positive/negative definiteness of matrices of divided differences of $\alpha \psi(A) \preceq \psi(B)$0, $\alpha \psi(A) \preceq \psi(B)$1, $\alpha \psi(A) \preceq \psi(B)$2 characterizes matrix monotonicity and its "weakening" (Hiai et al., 2010).

5. Secular Equation Algorithms and Computational Utility

Greedy maximization of spectral functions derived from weakly matrix-monotone constructions relies on efficient marginal gain calculations. A key innovation is the use of secular-equation–based eigenvalue updates: After maintaining the eigendecomposition $\alpha \psi(A) \preceq \psi(B)$3, the candidate rank-1 update $\alpha \psi(A) \preceq \psi(B)$4 yields new eigenvalues as roots of the secular equation

$\alpha \psi(A) \preceq \psi(B)$5

This enables $\alpha \psi(A) \preceq \psi(B)$6 updates (versus $\alpha \psi(A) \preceq \psi(B)$7 for complete decomposition) and allows spectral submodular objectives, including those using weakly matrix-monotone $\alpha \psi(A) \preceq \psi(B)$8, to be optimized tractably on large-scale datasets—providing $\alpha \psi(A) \preceq \psi(B)$9 empirical speedup for typical embedding dimensions (Bilmes et al., 28 May 2026).

6. Noncommutative and Operator System Extensions

The notion of weakly matrix monotone functions extends to the noncommutative setting. For real free functions over operator systems, monotonicity in the above sense (for all $\preceq$0 in the matrix order) is equivalent to the existence of a free analytic extension to a noncommutative upper half-plane, mapping into another noncommutative upper half-plane (Pascoe, 2017). This outcome generalizes classical Loewner's theorem to the framework of operator systems and noncommutative functions, with key applications in multivariate operator theory.

7. Connections, Applications, and Performance Guarantees

Weakly matrix-monotone functions enable construction of matrix-spectral set-functions that interpolate between classical (fully) matrix monotone submodular objectives and more general, potentially non-convex or saturating spectral scores. When the associated derivative $\preceq$1 is only $\preceq$2-weakly matrix-monotone, the family of associated set-functions is weakly DR-submodular, and greedy selection attains approximation ratio $\preceq$3. Empirically, this is sufficient for practical data subset selection, spectral clustering, and information-theoretic screening (Bilmes et al., 28 May 2026).

Exploiting secular-equation methods ensures computational viability in high-dimensional scenarios, such as large-scale dataset curation or feature subset selection. Weak monotonicity thus bridges the gap between theoretical submodularity and the limitations of physical objectives common in real-world combinatorial optimization tasks.

---

Selected References:

• "How Much Is a Dataset Worth? Scaling Laws, the Vendi Score, and Matrix Spectral Functions" (Bilmes et al., 28 May 2026)
• "Operator monotone functions and Löwner functions of several variables" (Agler et al., 2010)
• "The noncommutative Löwner theorem for matrix monotone functions over operator systems" (Pascoe, 2017)
• "Note on Löwner's theorem on matrix monotone functions in several commuting variables of Agler, McCarthy and Young" (Pascoe, 2014)
• "Loewner matrices of matrix convex and monotone functions" (Hiai et al., 2010)
• "Characterization of the monotonicity by the inequality" (Hoa et al., 2012)