Primal-Dual Greedy Algorithms

• Primal-dual greedy algorithms are meta-frameworks that combine greedy selection with simultaneous updates of both primal and dual solutions to efficiently approximate complex optimization problems.
• They exploit the structure of linear or convex relaxations by iteratively raising dual variables until constraints tighten, then committing corresponding primal updates to maintain feasibility.
• These methods are applied across weighted covering, online routing, submodular maximization, and variational problems, achieving provable approximation and competitive performance guarantees.

A primal-dual greedy algorithm is a meta-algorithmic framework that leverages both primal and dual formulations of an optimization problem, combining greedy steps with primal-dual coupling to efficiently yield approximate or competitive solutions across a wide range of combinatorial, nonlinear, and convex problems. Such algorithms exploit the structure of the primal and dual linear or convex relaxations, often using greedy selection or augmentation steps guided by dual variable updates, and are analyzed by showing that the resulting primal and dual solutions satisfy approximate or combinatorial slackness properties which yield provable approximation or competitive ratios.

1. Formal Structure and General Principles

Primal-dual greedy algorithms operate by tightly coupling the primal and dual (or Lagrangian) solutions during their iterative construction. For a given optimization problem, the algorithm maintains feasibility (or bounded infeasibility) of at least one side (primal or dual) while greedily augmenting the current solution based on local or global optimality criteria, typically driven by either primal constraints becoming tight or by maximizing or minimizing certain reduced costs or marginal values defined in terms of the dual variables.

A canonical setting is weighted integer covering or packing:

• Primal: $\min\{c^T x : A x \geq r,~x\geq 0\}$,
• Dual: $\max\{y^T r : A^T y \leq c,~y \geq 0\}$.

The key algorithmic routine involves:

1. Growing dual variables greedily (along some lattice or partial order of constraints) until some dual constraint becomes tight (“saturates” a variable).
2. Committing to a corresponding augmentation in the primal (e.g., increasing a variable) to maintain or improve feasibility.
3. Repeating, often adapting the set of active rows or columns, until a global feasibility criterion is achieved.

Primal-dual greedy algorithms are not restricted to linear programs but extend to configuration LPs, convex programs, or combinatorial structures with submodular, nonlinear, or supermodular objective or constraint functions (Peis et al., 2017, Thang, 2017, Zhang, 2018, Chakrabarty et al., 2023, Gupta et al., 2011, Fielbaum et al., 2019).

2. Greedy Systems and Lattice-Based Formulations

The generalization of “greedy systems” formalizes a class of integer covering systems for which the primal-dual greedy approach yields tractable and well-analyzed guarantees (Peis et al., 2017). A greedy system is defined relative to a partial order (typically a lattice) of constraints, with properties:

• Monotonicity of the right-hand side $r$ and matrix $A$,
• Lattice structure ensuring unique supports and join/meet operations,
• Weighted supermodularity, a twisted supermodular inequality binding $r$ and $A$.

Truncation of the constraint matrix, by capping each coefficient to the necessary coverage required to satisfy the rank difference between elements, further closes integrality gaps while preserving the structure that the greedy algorithm exploits.

The standard dual-based greedy phase raises the dual variable corresponding to the most “hard” (maximal rank) unsatisfied constraint until some variable becomes tight, selects the associated primal variable, and iterates on the reduced system. The overall process is captured by an algorithm featuring strict monotonicity, well-defined bottlenecks, and telescoping summation arguments in the approximation proofs.

3. Extension to Nonlinear and Configuration LP Settings

In nonlinear or configuration-LP-based problems, primal-dual greedy algorithms generalize by handling more complex cost functions, often requiring exponentially large LPs or duals indexed by resource configurations or demand vectors (Thang, 2017, Fielbaum et al., 2019).

The greedy step assigns each new request (in online or batched settings) to the strategy or resource set that minimizes the marginal cost, where marginal cost is evaluated as the difference in (possibly convex or nonconvex) objective increments.

Dual variables are updated in tandem—each assignment step contributes “charges” to the dual, which are then certified as feasible by exploiting properties such as (λ, μ)-smoothness (in the sense of set functions or their multilinear extensions):

• For set function costs, (λ, μ)-smoothness bounds the sum of marginal increments over all chains relative to the global and final costs,
• For covering LPs, local smoothness (w.r.t. multilinear extensions) ensures feasibility of dual constraints (Thang, 2017).

Key analytical result: If all cost functions satisfy (λ, μ)-smoothness with $\mu < 1$, the competitive ratio is $\lambda/(1-\mu)$. Many polynomial and superlinear cost settings can be cast in this form.

4. Primal-Dual Greedy Methods in Submodular and Variational Problems

Primal-dual greedy algorithms extend beyond covering and allocation to submodular maximization and variational model reduction.

• For monotone submodular maximization under cardinality or matroid constraints, a primal-dual greedy procedure maintains both a growing solution and a dual certificate distribution over chains of subsets, yielding the optimal $(1-1/e)$ approximation (empirically sometimes much better), with an explicit dual upper bound on the instance optimum (Chakrabarty et al., 2023).
• For reduced basis methods in parametric PDEs, particularly symmetric coercive variational problems, the primal-dual greedy approach constructs low-dimensional spaces for both primal and dual problems, guided by a robust primal-dual gap estimator which exactly equals the normed sum of primal and dual errors. Greedy selection is used to adaptively build bases (with or without mesh adaptation), balancing finite element, reduced basis, and mesh refinement errors (Zhang, 2018).

5. Application Domains and Performance Guarantees

Primal-dual greedy algorithms have broad applicability:

• Weighted Covering and Packing: Includes set cover, knapsack cover, subset cover, and flow/multicut/Steiner variants with proven 2- and $O(k)$-approximation, tightly related to lattice-width and chain parameters (Peis et al., 2017).
• Online Scheduling and Routing: Online load balancing, vector scheduling, and network routing problems with convex or polynomial objectives, yielding $O(\alpha^\alpha)$-competitive solutions for $\alpha$-power cost structures via dual fitting (Gupta et al., 2011, Thang, 2017).
• Nonlinear Covering with Complex Cost Structures: Nonlinear knapsack-cover, unsplittable flow-cover, and related instances using water-filling dual progress and segment-based greedy selection for $(2+\epsilon)$-approximation, or $(4+\epsilon)$ in the line/segment covering case (Fielbaum et al., 2019).
• Ad Allocation: Primal-dual greedy matches in Adwords with the small-bid assumption yield $1/2$-competitiveness through tight dual analysis and assignment-budgeting (Li, 2019).
• Wireless Scheduling under Uncertainty: Dynamic primal-dual greedy algorithms for network scheduling track slowly varying (stochastic) system parameters, achieving near-optimal cost in the fluid limit under strong law of large numbers (Li et al., 2010).

Approximation bounds are typically derived via telescoping sums capturing chain-based residuals or submodular increments, or by exploiting complementary slackness between primal and dual progress phases. For instance, the primal-dual greedy for weighted covering on a greedy system achieves a cost at most $2\delta + 1$ times the LP optimum, where $\delta$ is a problem-dependent truncation width (Peis et al., 2017).

6. Algorithmic Patterns and Notable Variants

Several algorithmic schemata recur:

• Dual-Greedy Raising: Iteratively increase (groups of) dual variables until some primal variable or constraint is tight, then update the primal accordingly (as in primal-dual schemes for covering, allocation, or matroid intersection).
• Bucket-Filling: Water-filling and bucket-oriented variants (“greedy bucket-filling”) manage resource allocation incrementally by simultaneously raising dual potentials until limit conditions trigger a primal update (Fielbaum et al., 2019).
• Balanced and Adaptive Greedy Bases: In reduced basis model reduction for PDEs, three major variants—fixed mesh/adaptive tolerance, adaptive mesh/fixed tolerance, and bi-adaptivity under degree-of-freedom budget—are orchestrated by a greedy sequence on primal-dual error gap (Zhang, 2018).

The following table categorizes several representative algorithmic settings:

Problem Domain	Greedy Step	Dual Update Type
Weighted Covering	Max residual (lattice)	Raise dual for max rank
Submodular Maximization	Max marginal gain	Progress on dual chain
Nonlinear Load Balancing	Min marginal cost	Dual fit via Lagrangian
Nonlinear Covering	Water-filling (buckets)	Incremental over segments
Variational PDE RBM	Max primal-dual gap	Error estimator driven

7. Theoretical Insights and Extensions

Primal-dual greedy algorithms’ analysis often involves approximate complementary slackness, chain/rank-based telescoping inequalities, and smoothness-based relaxations. The specific guarantees rely on problem structure:

• For greedy systems or systems admitting truncations, tight bounds on integrality gap and approximation ratio follow from lattice width or block parameters (Peis et al., 2017).
• For configuration LP-based online setting with non-convex or superlinear objectives, competitive ratios are governed by explicit smoothness coefficients (Thang, 2017).
• In submodular and convex minimization, explicit dual certificates provide instance-optimal bounds, not just worst-case guarantees (Chakrabarty et al., 2023, Zhang, 2018).
• Robustness to parameter changes, mesh adaptivity, and composite error control are achievable when the primal-dual gap accurately separates contributions from different discretization or parametric errors (Zhang, 2018).

The recursive, greedy progress, whether by chain, bucket, or error gap selection, is underpinned by dual feasibility at all steps and explicit tracking of how primal augmentations pay for or are paid by dual progress.

---

References:

(Peis et al., 2017, Thang, 2017, Zhang, 2018, Chakrabarty et al., 2023, Gupta et al., 2011, Fielbaum et al., 2019, Li, 2019, Li et al., 2010)