Maximally Monotone Operators

• Maximally monotone operators are set-valued maps in Banach spaces whose graphs cannot be extended without losing monotonicity.
• They utilize Fitzpatrick functions and variational representations to connect convex analysis with practical optimization and PDE solutions.
• Their properties are key for designing proximal algorithms and splitting methods, especially under Rockafellar’s constraint qualifications and in nonreflexive spaces.

A maximally monotone operator is a central object in nonlinear analysis, convex optimization, and partial differential equations. Given a Banach space $X$, a set-valued operator $A : X \rightrightarrows X^*$ is called monotone if, for all $(x, x^*), (y, y^*) \in \operatorname{gra}(A)$, the inequality $\langle x - y, x^* - y^* \rangle \geq 0$ holds; $A$ is maximally monotone if its graph is not properly contained in the graph of any other monotone operator—in other words, the monotonicity property cannot be extended beyond its domain without loss. Maximal monotonicity underpins the existence and uniqueness of solutions to variational inequalities, the convergence of iterative algorithms, and the stability of inclusions defined by subdifferentials of convex functions.

1. Definitions, Types, and Foundational Properties

A monotone operator $A$ is maximally monotone if for all $(x, x^*) \notin \operatorname{gra}(A)$, there exists $(y, y^*) \in \operatorname{gra}(A)$ such that $\langle x - y, x^* - y^* \rangle < 0$. The canonical examples are subdifferentials of proper, lower semicontinuous, convex functions, and normal cone operators to closed convex sets.

An important subclass is operators of type (FPV), characterized by a local criterion: for any open convex $U \subseteq X$ with $U \cap \operatorname{dom}A \neq \emptyset$, if $(x, x^*) \in U \times X^*$ is monotonically related (i.e., $\langle x-y,x^* - y^* \rangle \geq 0$ for all $(y,y^*)\in \operatorname{gra}A$ with $y \in U$), then $(x,x^*) \in \operatorname{gra}A$. This property enables "detection" of graph membership via local monotonicity (Yao, 2010).

Other notable types are type (D) (dense type) and type (NI) (negative infimum type), which are connected through the Fitzpatrick function—a convex representation tool—and coincide under maximal monotonicity (Bauschke et al., 2011). In reflexive Banach spaces, all maximally monotone operators are of type (FPV), (D), and (NI).

Domain notation is standard: $$\operatorname{dom}A = \{ x \in X : A x \neq \emptyset \}$$; $\operatorname{int}\operatorname{dom}A$ indicates the topological interior.

2. Fitzpatrick Functions and Variational Representation

The Fitzpatrick function $F_A:X \times X^*\to [−\infty, +\infty]$ associated to a monotone operator $A$ is defined as

$$F_A(x, x^*) = \sup_{(a, a^*) \in \operatorname{gra}A} \lbrace \langle x, a^* \rangle + \langle a, x^* \rangle - \langle a, a^* \rangle \rbrace.$$

$F_A(x,x^*)$ always dominates the duality product $\langle x, x^* \rangle$, with equality if and only if $(x,x^*) \in \operatorname{gra}A$.

Many analytic and algebraic properties of maximal monotonicity, sum theorems, and operator inclusions are encoded in inequalities via Fitzpatrick functions. The representability formalism—where $$A = \{ (x, x^*) : f(x,x^*) = \langle x, x^* \rangle \}$$ for some convex $f$—bridges convex analysis and monotone operator theory.

The variational sum and composition leverage Yosida approximations and Fitzpatrick functions to build representable monotone operators that may otherwise not be maximal (García et al., 2011).

3. Sum Theorems and Rockafellar’s Constraint Qualification

The classical sum problem investigates when $A+B$ is maximally monotone, given $A,B$ maximally monotone. Rockafellar’s constraint qualification ($$\operatorname{dom}A \cap \operatorname{int}\operatorname{dom}B \neq \emptyset$$) is necessary. In nonreflexive spaces, additional regularity is essential; for instance, if $A$ is of type (FPV) and $B$ is maximally monotone (often with full domain), the maximality of $A+B$ is restored under CQ (Yao, 2010, Borwein et al., 2013, Yao, 2014, Pattanaik et al., 2015).

A typical result: if $A$ and $B$ are maximally monotone, $A+N_{\overline{\operatorname{dom}B}}$ is (FPV), and

$$\operatorname{dom}A \cap \operatorname{int}\operatorname{dom}B \neq \emptyset, \quad \operatorname{dom}A \cap \overline{\operatorname{dom}B} \subseteq \operatorname{dom}B,$$

then $A+B$ is maximally monotone (Yao, 2010). When both are subdifferentials or one is a linear relation, the same result generally holds (Borwein et al., 2012).

These theorems are essential for the theory and algorithms of splitting, operator decomposition, and convex programming.

4. Pathological Cases and Nonreflexive Spaces

In nonreflexive Banach spaces, pathologies emerge. Not all maximally monotone operators have dense-type graphs; some spaces, such as those containing $c_0$, $\ell^1$, the James space $J$, or its dual, admit maximally monotone operators that fail (D) and the Brønsted–Rockafellar property (Bauschke et al., 2011).

Explicit "predual constructions" use operators $A:X^* \to X^{**}$, together with vectors $e \in X^{**}\setminus X$, to manufacture maximally monotone operators with almost arbitrary behaviors regarding approximation, closure, and inf-convolution properties (Bauschke et al., 2011). These constructions inform about the necessity of reflexivity in classical results and caution about extrapolation from Hilbert space settings.

5. Structure, Faces, and Decomposition

Recent work establishes fine structural results for maximally monotone operators. When the domain has nonempty interior (as in the subdifferential of a convex function), one can explicitly reconstruct the operator as a sum of a recession or normal cone part and a convex hull of values on a dense set of interior points:

$$A(x) = N_{\operatorname{dom}A}(x) + \overline{\operatorname{conv}} A_s(x)$$

where $A_s(x)$ collects weak*-limits of values $A(s)$ as $s \to x$ in the norm and $s$ ranges over a dense subset of $\operatorname{int}\operatorname{dom}A$ (Borwein et al., 2012).

More generally, faces and support functions of $A(x)$ are obtainable by variational limits along directions, and the operator can be locally and globally decomposed via selections and normal cones (see (Nguyen et al., 2019)): | Representation | Formula/Description | Context | |--------------------------|-----------------------------------------------|------------------------| | Face $A(x; v)$ | $$\operatorname{Lim\,sup}_{w\to v, t\downarrow 0} A(x+tw)$$ | Uniformly convex spaces | | Support function $\sigma_{A_x}(v)$ | $$\lim\inf_{w\to v, t\downarrow 0} \langle A^{x+tw}, w \rangle$$ | Minimal-norm selection |

These formulas have implications for uniqueness, local determination from selections, and algorithmic approximation of monotone inclusions.

6. Practical Algorithms and Applications

Maximally monotone operators underpin many proximal algorithms, primal-dual splitting methods, and projection-type schemes. Recent advances include primal-dual algorithms that solve zero-finding problems for sums of monotone operators composed with linear maps via alternating resolvent evaluations, applicable to composite nonsmooth optimization, image reconstruction, and facility location (Bot et al., 2012). The iterates generally use explicit splitting of operators and linear compositions to avoid costly resolvents of the sum.

The averaged alternating modified reflections algorithm extends classical projection methods to compute the resolvent of the sum via parallel cuts, using modified reflected resolvents and product space reformulations; strong convergence is guaranteed under mild qualification (Artacho et al., 2018).

Inertial and accelerated variants, such as those employing Yosida regularizations and Nesterov-type momentum, offer faster fixed-point residual decay rates, with worst-case $O(1/i^2)$ convergence and automatic acceleration (Attouch et al., 2017, Kim, 2019).

7. Open Problems and Further Directions

Outstanding questions pertain to relaxing CQ and (FPV) conditions in sum theorems—especially in nonreflexive settings—and to the broader applicability of Fitzpatrick function constructs. Alternative approaches to constructing maximally monotone operators with prescribed graphical or range properties remain active (Bauschke et al., 2019, Wachsmuth, 2021).

Counterexamples demonstrate that graphical limits of maximally monotone operators may lose maximality (Wachsmuth, 2021), motivating refined analysis of limiting behavior, proto-differentiability, and directionally differentiable resolvents.

A central avenue is studying the precise boundaries between operator subclasses (linear relations, subdifferentials, nonexpansive retractions) and their impact on decomposition, paramonotonicity, rectangularity, and representability—key for optimization, equilibrium problems, and large-scale computational schemes.

---

In sum, maximally monotone operators constitute a foundational and versatile framework for nonlinear analysis, with deep connections to variational structure, convexity, and computation. Advances in their characterization, sum theorems, structure formulas, and algorithmic exploitation continue to shape both theory and applications.