Weighted Matching Graphs

• Weighted matching graphs are undirected graphs with nonnegative edge weights, central to solving the maximum weight matching problem with applications in network design and operations research.
• Key methodologies include combinatorial algorithms like the blossom and scaling algorithms, as well as linear programming relaxations and message passing techniques to achieve near-optimal matchings.
• Recent advancements focus on dynamic, streaming, and stochastic settings, extending classical matching models to connected matchings, b-matchings, and f-factors for improved practical efficiency.

A weighted matching graph is an undirected graph equipped with nonnegative weights on its edges; its study encompasses combinatorial optimization, polyhedral theory, algorithm engineering, and probabilistic analysis. The prototypical problem—maximum weight matching—asks for a vertex-disjoint subset of edges of maximum total weight, with applications spanning operations research, statistics, network design, and stochastic optimization. Recent developments encompass deterministic and stochastic settings, fully dynamic maintenance, large-scale computation, and the extension to structures such as connected matchings, $b$-matchings, and general $f$-factors.

1. Formal Problem Definition and Polyhedral Structure

Let $G = (V, E)$ be an undirected graph, and $w : E \to \mathbb{R}_{\ge 0}$ an edge-weight function. A matching $M \subseteq E$ is a set of edges no two of which share a vertex. The maximum weight matching (MWM) problem is

$$\max_{M \subseteq E\ \text{matching}} \sum_{e \in M} w_e,$$

equivalently posed through the integer program: $$\begin{array}{rl} \max_{x \in \{0,1\}^E} & \sum_{e \in E} w_e x_e \ \text{s.t.} & \sum_{e \ni v} x_e \le 1 \quad \forall\ v \in V. \end{array}$$ The convex hull of the incidence vectors of matchings— the matching polytope— is instrumental in linear programming relaxations, polyhedral algorithms, and the analysis of belief-propagation heuristics (0705.0760, Samer et al., 2023).

Extensions include:

• Perfect matchings: Each vertex is matched.
• $b$-matchings: Vertex $v$ may be incident to up to $b(v)$ matching edges (Gabow, 2016).
• $f$0-factors: Vertex $f$1 is incident to exactly $f$2 matching edges.

Weighted connected matching further constrains the induced subgraph on covered vertices to be connected, dramatically increasing combinatorial complexity.

2. Exact Combinatorial and Scaling Algorithms

Foundational approaches for weighted matchings in general graphs are built on Edmonds' blossom algorithm, primal-dual augmentations, and dual corrections. Modern variants optimize for time complexity and extendability:

Scaling Algorithms: By processing the binary representation of weights from most to least significant bits, scaling algorithms iteratively refine dual and primal solutions. In each scale, dual variables $f$3 (vertices) and $f$4 (odd sets/blossoms) are maintained, preserving near-complementary slackness. The best-known algorithm in general graphs achieves $f$5 time, where $f$6, $f$7, and $f$8 bounds edge weights (Duan et al., 2014). For bipartite graphs, improved bounds are possible and recent algorithmic innovations exploit sparsity and component decomposition (Kwok, 28 Feb 2025, Lahn et al., 2019).

Data Structures and Engineering: The analytic performance of scaling and blossom algorithms depends critically on union-find schemes, dynamic nearest common ancestor data structures, and efficient heap implementations (Gabow, 2016). For dynamic settings, level-based graph partitioning, maximal submatchings, and greedy re-matching per level allow efficient update handling with polylogarithmic amortized time per edge insertion or deletion (Anand et al., 2012).

3. Fractional Relaxations, LP Duality, and Message Passing

The integer programming formulation leads to a natural fractional relaxation: $f$9 with dual as a minimum weight vertex cover. The tightness of this LP is essential: in bipartite graphs, it always produces integral solutions; in general graphs, integrality can fail due to odd set constraints ("blossoms").

Max-product belief propagation, viewed as message passing on a factor graph encoding the matching constraints and edge weights, is deeply connected to this LP. It is shown that:

• If the LP relaxation is tight, then max-product converges to the correct (maximum weight) matching.
• If the LP is loose (fractional solution), max-product does not converge.

This establishes a precise equivalence between combinatorial algorithms, convex relaxations, and local iterative methods (0705.0760).

4. Streaming, Parallel, and Approximate Weighted Matchings

When computation is distributed or data arrives as a stream, algorithms must be adapted to limited memory and pass constraints:

Streaming and MPC: Reduction frameworks allow one to convert an unweighted matching algorithm into an efficient weighted matching routine by searching for augmenting paths in filtered, layered, or weight-bucketed unweighted subgraphs. For instance, multi-pass streaming and MPC implementations derive $G = (V, E)$0-approximate weighted matchings from corresponding unweighted subroutines, with only constant-factor overhead in resource usage (Gamlath et al., 2018).

Fully Dynamic Algorithms: Approximate maintenance under arbitrary edge insertions and deletions exploits local updating, random walks, and dynamic programming. Partitioning edges into exponentially scaled weight levels and maintaining per-level maximal matchings yields $G = (V, E)$1-approximation deterministically or $G = (V, E)$2 in expectation via random scaling, with $G = (V, E)$3 expected amortized update time, where $G = (V, E)$4 is the weight spread (Anand et al., 2012). Random walk/short augmenting path methods, coupled with dynamic programming, can maintain a $G = (V, E)$5-approximation with high probability after every update, performing only polylogarithmic work (Angriman et al., 2021).

5. Stochastic and Query-Efficient Weighted Matching

The stochastic weighted matching problem arises when each edge is only realized (present) with a known independent probability. The canonical algorithmic challenge is, given a query budget (number/degree of edges that can be selected for probing), to find in expectation a matching close to the omniscient optimum (that knows the realized instance):

• Non-adaptive selection: Select a degree-bounded subgraph $G = (V, E)$6 prior to seeing edge realizations, output the matching in its surviving subgraph. Early results struggled to surpass a $G = (V, E)$7-approximation for weighted graphs. Major advances include a $G = (V, E)$8-approximation with $G = (V, E)$9 probes per vertex (breaking the $w : E \to \mathbb{R}_{\ge 0}$0 barrier) (Behnezhad et al., 2018), and more recently a $w : E \to \mathbb{R}_{\ge 0}$1-approximation with only $w : E \to \mathbb{R}_{\ge 0}$2 queries per vertex, via sophisticated variance-bounding reductions and random-order contention resolution (Derakhshan et al., 2023).
• Adaptive and query-commit settings: In bipartite graphs, the classical $w : E \to \mathbb{R}_{\ge 0}$3 approximation barrier (achievable by LP-based rounding) was recently marginally exceeded; a two-phase algorithm, combining a load-slack transformation on the LP solution and exploitation of "light" edges in a second pass, yields a $w : E \to \mathbb{R}_{\ge 0}$4-approximation for some explicit constant $w : E \to \mathbb{R}_{\ge 0}$5 (Derakhshan et al., 2022).
• Fractional matching analysis: These results leverage a division of edges into crucial and non-crucial based on their marginal probability of being included in the optimal matching, constructing intricate fractional assignments to balance coverage and variance while observing per-vertex degree constraints.

6. Connected Weighted Matching and Polyhedral Approaches

Connected matching imposes the additional constraint that the vertices covered by the matching induce a connected subgraph. While the maximum-cardinality connected matching can be solved in polynomial time, the weighted version is NP-hard even under restrictions such as planar bipartite graphs with binary edge weights (Samer et al., 2023).

Two principal mixed-integer programming (MIP) formulations have been proposed:

• Flow-based extended formulation: Introduces edge selection and auxiliary arborescence/flow variables, enforcing matching and connectivity via explicit constraints on flows from a super-source to matched vertices.
• Inequality-based original formulation: Relies on the intersection of degree, minimal separator, indegree, and blossom inequalities to describe the convex hull of connected matchings or tight approximations.

Branch-and-cut implementations using these formulations have demonstrated strong root-node LP bounds and practical solvability for large benchmarks, with the full formulation (degree, separator, indegree, blossom inequalities) conjectured to closely approximate the true connected-matching polytope.

7. Practical Considerations and Algorithmic Engineering

Efficient implementations depend on tailored data structures and numerical stability:

• Dynamic NCA/link for blossom algorithms: Supports $w : E \to \mathbb{R}_{\ge 0}$6 time blossom creation (with $w : E \to \mathbb{R}_{\ge 0}$7 the inverse-Ackermann function), crucial to enabling $w : E \to \mathbb{R}_{\ge 0}$8-time weighted matching in general graphs (Gabow, 2016).
• Dual and slack management: Accurate and efficient dual adjustments, slack updates, and priority queue event scheduling are core to scaling and primal-dual methods (Duan et al., 2014).
• Streaming heuristics and tuning: In streaming and dynamic contexts, heuristics such as early stopping of random walks, weight bucketing, and hash-based adjacency access yield significant speed-ups with minimal approximation loss (Angriman et al., 2021).

The design and analysis of algorithms in weighted matching graphs thus interconnect rich combinatorial, polyhedral, and probabilistic methods, buttressed by advanced algorithm engineering for both offline and online computation. These methods and results collectively underpin much of modern graph optimization in weighted and stochastic settings.