Acyclic Monotone Operators Are Not Closed Under Addition

Abstract: Borwein and Wiersma [SIAM J. Optim. 18(3) (2007), 946-960] asked if the set of acyclic monotone operators is closed under addition. We answer this question in the negative.

• The paper demonstrates that the sum of two nonlinear acyclic monotone operators can fail to be acyclic, challenging previous structural assumptions.
• It utilizes convex-concave functions and their saddle gradients to construct an explicit counterexample where acyclic components sum to a non-acyclic operator.
• The result has major implications for monotone operator theory and algorithm design in optimization and variational inequalities.

Acyclic Monotone Operators and Non-Closure Under Addition

Background and Motivation

Monotone operator theory is foundational in modern convex analysis and optimization. Given an operator $F$, monotonicity is characterized by the inequality $\langle F(z) - F(z'), z - z' \rangle \geq 0$ for all $z, z'$ in the domain. Central to the structure theory for monotone operators is the Asplund decomposition: every single-valued, maximally monotone operator $F$ on $\mathbb{R}^d$ can be uniquely represented as

$F = \nabla g + S$

where $g$ is a convex function and $S$ is an acyclic monotone operator. Acyclic monotone operators, defined as those monotone $F$ for which in any decomposition $F = \partial g + S$ the convex part must be affine, are essential because they represent the "non-conservative" component of $\langle F(z) - F(z'), z - z' \rangle \geq 0$0.

Borwein and Wiersma [SIAM J. Optim. 18(3):946–960, 2007] posed a fundamental structural question: Is the set of acyclic monotone operators closed under addition? While closure holds if at least one summonend is linear, the general case was unresolved until the present work.

Main Result and Technical Contribution

The paper provides a negative answer, demonstrating that the sum of two nonlinear acyclic monotone operators need not themselves be acyclic. This contradicts any presumption of convexity or closure under addition and provides a precise structural limitation. The key construction utilizes the following two convex-concave functions: $\langle F(z) - F(z'), z - z' \rangle \geq 0$1 which induce, via their saddle gradients, the monotone operators

$\langle F(z) - F(z'), z - z' \rangle \geq 0$2

Applying the Asplund decomposition to each, they write

$\langle F(z) - F(z'), z - z' \rangle \geq 0$3

and show that $\langle F(z) - F(z'), z - z' \rangle \geq 0$4 are both acyclic monotone operators.

The crucial observation is that

$\langle F(z) - F(z'), z - z' \rangle \geq 0$5

for $\langle F(z) - F(z'), z - z' \rangle \geq 0$6, meaning the sum $\langle F(z) - F(z'), z - z' \rangle \geq 0$7 must equal the gradient of a convex (and generally non-affine) function, specifically

$\langle F(z) - F(z'), z - z' \rangle \geq 0$8

They prove that $\langle F(z) - F(z'), z - z' \rangle \geq 0$9 is non-affine, and therefore, by the definition of acyclicity, $z, z'$0 is not acyclic. This provides a constructive counterexample and establishes non-closure under addition for the set of acyclic monotone operators.

Theoretical Implications

The result refutes any claim that acyclic monotone operators form a convex cone (or even a vector space). This adds a subtlety to monotone operator calculus analogous to classical issues in the theory of non-conservative vector fields. The finding also implies that operator classes relevant to optimization such as the conservative, acyclic, and skew-symmetric classes have fundamentally different algebraic structures.

The construction's generality is enhanced in the remarks, noting that the argument extends to other $z, z'$1 with bounded second partials and nonconstant mixed partial, indicating that the phenomenon is robust and not reliant on the trigonometric form.

Relation to Prior Work and Algorithms

The operators $z, z'$2 are connected to contemporary research on variational inequalities and operator splitting methods, such as those studied by Lee et al. (Lee et al., 13 Apr 2026), where explicit decompositions of monotone operators inform algorithmic acceleration strategies and the analysis of convergence rates in min-max optimization.

In algorithmic design for saddle point problems, monotone operator calculus is integral to convergence analysis of first-order and primal-dual algorithms [Bauschke and Combettes 2017; Ryu and Yin 2022]. The loss of closure hampers any effort to decompose classes of operators using only acyclic components. Consequently, algorithms relying on such structural decomposability (for example, for adaptive preconditioning or operator splitting) face intrinsic limitations.

Practical and Future Directions

While the paper’s contribution is primarily structural, it may inform the design and theoretical guarantees of algorithms for saddle-point or variational inequality problems. The sharp non-closure result circumscribes the utility of decomposing complex monotone operators as sums of acyclic components, suggesting that alternative structural properties (such as resolvent or cocoercivity decompositions) may need emphasis.

In the broader context of optimization and learning theory, especially with the increasing use of operator-theoretic frameworks in large-scale games and learning systems, a fine-grained understanding of operator algebra is crucial. Future work may examine whether analogous non-closure statements hold for other operator classes or under additional restrictions (e.g., single-valuedness, Lipschitz continuity). Additionally, the negative result signals challenges for generalizing algorithmic acceleration techniques that exploit Asplund decomposition.

Conclusion

This work rigorously establishes that acyclic monotone operators are not closed under addition by explicit counterexample, answering an open question from Borwein and Wiersma. The findings reveal a fundamental non-convexity in the structure of monotone operator classes, with implications for operator analysis and the theory and design of first-order algorithms in convex optimization and game theory (2604.22570).