Duality-Guided Splitting Algorithms

• Duality-guided splitting algorithms are methods that leverage primal–dual correspondences (e.g., Attouch–Théra and Fenchel–Rockafellar dualities) to derive explicit fixed-point equations.
• They decompose complex monotone inclusions and composite convex problems by processing individual operators via tailored proximal or resolvent mappings.
• These algorithms ensure convergence under monotonicity and paramonotonicity assumptions and have practical applications in image restoration, signal recovery, and decentralized optimization.

A duality-guided splitting algorithm is an operator-splitting scheme where the iteration, the decomposition, and the algorithmic guarantees are explicitly steered by the dual structure of the monotone inclusion or optimization problem. Instead of treating primal and dual quantities separately or formulating monolithic resolvent steps for the entire composite operator, these methods systematically leverage Attouch–Théra, Fenchel–Rockafellar, or saddle-point dualities to derive explicit primal–dual fixed-point equations. This allows each constituent operator or function to be processed via its individual proximal or resolvent mapping, enabling distributed, parallel, and modular algorithmic realizations. Duality-guided splitting is the prevailing framework for the class of scalable algorithms solving large-scale monotone inclusions, convex composite minimization, and best-approximation problems, and it encompasses, generalizes, and clarifies the geometry and convergence theory of all classical splitting methods.

1. Duality Foundations of Splitting Schemes

The backbone of duality-guided splitting is the rich correspondence between primal inclusions and their duals, notably as codified by Attouch–Théra duality and Fenchel–Rockafellar theory. For monotone inclusions $0 \in A x + B x$, with $A, B$ maximally monotone on a Hilbert space $X$, Attouch–Théra associates to $(A,B)$ a unique dual pair and shows that the dual inclusion $0 \in A^{-1}k + B^{\mathcal{A}} k$ encodes all relevant structure for algorithmic design (Bauschke et al., 2011). Analogously, for composite convex minimization with linear compositions, the Fenchel dual and the associated saddle-point system are leveraged to build splitting algorithms maintaining primal–dual symmetry and facilitating separate treatment of nonsmooth, smooth, and composite terms (Condat et al., 2020, Condat et al., 2019).

A unifying perspective is to represent both the primal and dual solution sets as the Kuhn–Tucker set for the monotone inclusion and to formulate fixed-point equations that parametrize all primal–dual pairs via mappings such as the Douglas–Rachford operator, or product-space monotone inclusions (Bauschke et al., 2011, Combettes, 2012, Briceno-Arias et al., 2010).

2. General Problem Class and Operator Decomposition

Duality-guided splitting algorithms address broad classes of problems, such as monotone inclusions of the form

$$0 \in \sum_{i=1}^n A_i(x) + \sum_{k=1}^r L_k^* B_k(L_k x) + \sum_{j=1}^p C_j(x)$$

with $A_i$ and $B_k$ maximally monotone operators, $L_k$ bounded linear, and $C_j$ (possibly) single-valued monotone or Lipschitz operators (Dao et al., 11 Dec 2025, Becker et al., 2013, Combettes, 2012). The key principle is that each block of operators is handled individually—resolvents/proximity steps are performed per $A, B$0, $A, B$1, and explicit steps per $A, B$2, $A, B$3—which notably avoids computation of the composite resolvents for sums or intricate compositions. Parallel sum and parallel composition constructions (e.g., $A, B$4, $A, B$5) extend the class of admissible operator structures (Becker et al., 2013).

This block-wise decomposition is essential for practical scalability and underpins the modularity of the framework. Dual variables naturally arise for linear composition and coupling terms, and the corresponding dual system can complexly interleave primal and dual components, as in the inclusions of (Combettes, 2012).

3. Algorithmic Templates and Detailed Iteration Structure

A typical duality-guided splitting scheme constructs a product-space fixed-point operator whose blocks correspond to primal and dual variables. The algorithm alternates between explicit forward steps (evaluations of Lipschitz, cocoercive, or linear mappings) and backward/proximal steps (resolvents of set-valued monotone operators). A general template for a two-block structure $A, B$6 is the Douglas–Rachford iteration: $A, B$7 with $A, B$8, $A, B$9 (Bauschke et al., 2011). The primal–dual mapping $X$0 yields the solution pair immediately from any fixed point.

For multi-operator and composite inclusions, primal–dual forward–backward–forward (FBF) algorithms and primal–dual three-operator splitting (PD3O) schemes are deployed. In the general model (Briceno-Arias et al., 2010, Condat et al., 2020, Dao et al., 11 Dec 2025): $X$1 with extrapolation, possibly variable-stepsizes, and relaxation (Dao et al., 11 Dec 2025). All operator evaluations and linear maps are performed separately. For parallel-sum or parallel composition cases, blocks are processed in parallel with communication only through shared primal and dual aggregates (Becker et al., 2013, Combettes, 2012).

The relaxation and step-size parameters are chosen to ensure averagedness or quasi-averagedness of the update operator under the appropriate metric induced by the operator splitting, with conditions such as $X$2 (Dao et al., 11 Dec 2025, Condat et al., 2019).

4. Convergence Guarantees and Paramonotonicity

Rigorous convergence theory for duality-guided splitting relies on monotonicity and paramonotonicity. For maximally monotone (and in particular, paramonotone) operators, weak convergence to solution pairs is guaranteed under minimal assumptions (Bauschke et al., 2011, Combettes, 2012, Briceno-Arias et al., 2010). Strong convergence obtains when one operator is uniformly monotone at the limit or under additional regularity (Briceno-Arias et al., 2010, Becker et al., 2013).

Paramonotonicity plays a central role in full solution recovery: if $X$3 paramonotone, recovery formulas like

$X$4

guarantee that any one dual solution suffices to generate the full set of primal solutions, and best-approximation projectors enjoy explicit formulas $X$5 on the sum set $X$6 (Bauschke et al., 2011). These properties generalize best-approximation results and extend the geometric understanding of splitting fixed-point sets.

For composite convex minimization, sublinear, accelerated $X$7, and linear convergence rates in function suboptimality and iterate gap are available under step-size conditions and additional convexity/smoothness (Condat et al., 2020). Ergodic/non-ergodic convergence and metric preconditioning via product space geometry further broaden the guarantees (Condat et al., 2019).

5. Algorithmic Variants, Applications, and Distributed Realizations

Duality-guided splitting provides a flexible blueprint encompassing numerous algorithmic variants:

• Douglas–Rachford and Peaceman–Rachford for two-operator inclusions.
• Chambolle–Pock (PDHG), Condat–Vũ, Davis–Yin (PD3O), Loris–Verhoeven, and forward–backward (proximal gradient) for various monotone sum and saddle points (Condat et al., 2019, Condat et al., 2020).
• Parallel Dykstra splitting for best-approximation with product-set or general Fenchel dual problems, implementing full or partial block parallelism (Pang, 2017).

Applications span image restoration (structured TV-$X$8 models), signal recovery, large-scale lasso and fused lasso, distributed and decentralized optimization on network graphs (Becker et al., 2013, Dao et al., 11 Dec 2025, Condat et al., 2020). The ability to work in reduced-dimension product spaces, or to exploit graph-based structure, leads to highly scalable implementations.

Distributed and parallel variants emerge naturally: splitting steps are performed independently on each operator block, synchronization is carried out only via linear combinations or primal/dual aggregations, and communication–computation trade-offs can be managed explicitly (Pang, 2017, Condat et al., 2020).

6. Choice of Parameters, Relaxation, and Preconditioning

Practical deployment of duality-guided splitting mandates careful selection of step-sizes, relaxation, and (optionally) diagonal preconditioning metrics (Dao et al., 11 Dec 2025, Condat et al., 2019). For composite inclusions, parameter ranges are explicitly determined by operator Lipschitz/cocoercivity constants and the norms of linear couplings, e.g., $X$9. Relaxation parameters up to $(A,B)$0 are possible for firmly nonexpansive cases (Douglas–Rachford, ADMM), or appropriately reduced when the composite operator is not firmly nonexpansive.

Aggressive preconditioning (e.g., diagonal or block-diagonal metrics) can enlarge allowable step-sizes and accelerate convergence. Monitoring the primal–dual gap and residuals provides adaptive stopping and diagnostics.

7. Impact and Theoretical Unification of Operator Splitting

The duality-guided splitting paradigm establishes a rigorous, modular, and unifying framework for operator splitting in convex optimization and monotone inclusion problems. It demystifies the geometry underlying classical algorithms, extends the reach of splitting methods to wide classes of structured problems (including non-cocoercive and multi-block settings), and offers convergence guarantees and algorithmic templates that are scalable, distributed, and parallelizable (Dao et al., 11 Dec 2025, Becker et al., 2013, Combettes, 2012, Condat et al., 2019).

The central insight—algorithmic design and analysis directly informed by the dual structure—fuels continued developments in primal–dual methods, distributed optimization, and beyond. Duality-guided splitting remains a cornerstone for modern, large-scale optimization and monotone operator theory.