Edge-Weighted Maximum Clique Problem

• Edge-Weighted Maximum Clique Problem is the task of identifying a clique in a graph whose sum of edge weights is maximized, generalizing the classic maximum clique problem with pairwise interactions.
• Optimization methods such as a quadratic 0–1 formulation, linearization, and continuous relaxations drive algorithmic advances, particularly through effective branch-and-bound and coloring strategies.
• Applications in graph matching and structural pattern recognition leverage adaptive bounds, with recent studies comparing combinatorial bounds (UB2) to LP relaxations (UB1) for efficient clique search.

Searching arXiv for recent and foundational papers on the Edge-Weighted Maximum Clique Problem. The Edge-Weighted Maximum Clique Problem (EWMCP) asks for a clique in a simple undirected graph whose total edge weight is maximal. In the formulation emphasized in recent comparative work, an instance is a graph $G=(V,E)$ with nonnegative edge weights $w_{ij}$ on $(i,j)\in E$, and the objective is to maximize

$w(C)=\sum_{\{i,j\}\subseteq C,\ (i,j)\in E} w_{ij}$

over all cliques $C\subseteq V$; the optimal value is denoted $\omega(G,w)$ and is sometimes called the edge-weighted clique number (Ciccarelli et al., 9 Jul 2025). EWMCP is a special case of the more general maximum weight clique framework with both vertex and edge weights, obtained by omitting vertex weights or setting them to a constant that does not affect the argmax (0912.4584). The problem occupies an important position at the intersection of combinatorial optimization and graph matching, because weighted association-graph constructions reduce a broad class of matching objectives to maximum clique search, and purely pairwise matching objectives specialize that reduction to EWMCP (0912.4584).

1. Formalization and relation to neighboring clique problems

EWMCP is defined on a simple undirected graph with symmetric, nonnegative edge weights. A clique is a subset of vertices that is pairwise adjacent, and its value is the sum of weights on all induced clique edges (Ciccarelli et al., 9 Jul 2025, Shimizu et al., 2018). In the notation used by Shimizu and coauthors, non-edges may be assigned weight $0$ for convenience in formulas, but feasibility is still determined by clique adjacency constraints rather than by allowing arbitrary vertex subsets (Shimizu et al., 2018).

The problem generalizes the classical Maximum Clique Problem (MCP). If all edge weights are uniform, maximizing the clique’s total edge weight is equivalent to maximizing clique cardinality up to a monotone transformation, so EWMCP inherits NP-hardness from MCP (Shimizu et al., 2018). It also differs from the standard Maximum Vertex-Weight Clique Problem (MWCP), where weights are attached only to vertices. In EWMCP, the contribution of a candidate vertex depends on the current partial clique through pairwise interactions, so many vertex-weighted techniques require adaptation rather than direct reuse (Shimizu et al., 2018).

A broader weighted-graph formalism places weights on both vertices and edges. In that setting, a weighted graph can be written as $Z=(V,\alpha)$ with an attribute function $\alpha:V\times V\to\mathbb{R}\cup\{\epsilon\}$, vertex weights on the diagonal, and edge weights off the diagonal. The clique weight is then

$\omega(C)=\sum_{i,j\in C}\alpha(i,j),$

which includes both unary and pairwise terms (Jain et al., 2011). EWMCP is the edge-only specialization of this objective. In the notation of association graphs, this specialization is obtained by setting $w_{ij}$0 for all vertices, yielding

$w_{ij}$1

so only edge weights remain (0912.4584).

2. Optimization formulations and continuous relaxations

A standard quadratic $w_{ij}$2–$w_{ij}$3 formulation introduces binary variables $w_{ij}$4 indicating vertex selection and maximizes

$w_{ij}$5

subject to clique-enforcing non-edge constraints $w_{ij}$6 for all $w_{ij}$7 (Shimizu et al., 2018). This is the direct quadratic representation of the pairwise objective.

A common linearization augments vertex variables with edge variables $w_{ij}$8, where $w_{ij}$9 iff both endpoints are chosen. The objective becomes $(i,j)\in E$0, together with constraints $(i,j)\in E$1, $(i,j)\in E$2, and either non-edge clique constraints or, more generally, independent-set inequalities $(i,j)\in E$3 for all independent sets $(i,j)\in E$4 (Ciccarelli et al., 9 Jul 2025). For nonnegative weights, the strengthening $(i,j)\in E$5 is redundant for optimality, although it yields an exact linearization (Ciccarelli et al., 9 Jul 2025).

Within the more general MWCP formalism, the weighted adjacency matrix $(i,j)\in E$6 supports a quadratic formulation

$(i,j)\in E$7

under clique constraints and binary $(i,j)\in E$8 (0912.4584). If vertex weights are present, one may write $(i,j)\in E$9 and obtain $w(C)=\sum_{\{i,j\}\subseteq C,\ (i,j)\in E} w_{ij}$0; for EWMCP, $w(C)=\sum_{\{i,j\}\subseteq C,\ (i,j)\in E} w_{ij}$1 (0912.4584).

Continuous relaxations are also part of the methodological landscape. One relaxation maximizes $w(C)=\sum_{\{i,j\}\subseteq C,\ (i,j)\in E} w_{ij}$2 over the probability simplex

$w(C)=\sum_{\{i,j\}\subseteq C,\ (i,j)\in E} w_{ij}$3

and, under suitable transformations, local maxima often correspond to maximal cliques (0912.4584). In the graph-matching literature, this viewpoint motivates projected gradient ascent, replicator dynamics, and fractional programming for clique-based formulations (0912.4584). By contrast, the 2025 upper-bound comparison study explicitly restricts attention to a coloring-based LP relaxation and does not analyze semidefinite, spectral, or Motzkin–Straus-type relaxations (Ciccarelli et al., 9 Jul 2025).

3. Association graphs and the graph-matching reduction

A central reason EWMCP is studied beyond stand-alone clique optimization is its role in graph matching. A generic graph matching problem (GMP) can be written as

$w(C)=\sum_{\{i,j\}\subseteq C,\ (i,j)\in E} w_{ij}$4

where $w(C)=\sum_{\{i,j\}\subseteq C,\ (i,j)\in E} w_{ij}$5 is a partial morphism between graphs $w(C)=\sum_{\{i,j\}\subseteq C,\ (i,j)\in E} w_{ij}$6 and $w(C)=\sum_{\{i,j\}\subseteq C,\ (i,j)\in E} w_{ij}$7, $w(C)=\sum_{\{i,j\}\subseteq C,\ (i,j)\in E} w_{ij}$8 is the feasible set of morphisms, and $w(C)=\sum_{\{i,j\}\subseteq C,\ (i,j)\in E} w_{ij}$9 provides compatibility values on matched items (0912.4584). The 2009 equivalence result gives a necessary and sufficient condition for such a GMP to be equivalent to an MWCP in a derived association graph: the feasible set must be exactly the set of all $C\subseteq V$0-morphisms for some local property $C\subseteq V$1 on item pairs, i.e. $C\subseteq V$2 (0912.4584).

Under this $C\subseteq V$3-closure condition, the association graph $C\subseteq V$4 has vertices

$C\subseteq V$5

and edges

$C\subseteq V$6

with weights $C\subseteq V$7 on compatible vertex and edge pairs (0912.4584). Each clique corresponds bijectively to a feasible morphism, and the clique weight equals the matching objective:

$C\subseteq V$8

This establishes exact equivalence between clique optimization in the association graph and the original matching problem (0912.4584).

EWMCP arises in this framework when the matching objective has only pairwise terms, so all unary compatibilities vanish, $C\subseteq V$9, and the resulting association graph has zero vertex weights (0912.4584). More generally, if unary terms are constant across feasible morphisms, omitting them preserves the argmax and again yields an EWMCP instance (0912.4584). This explains why edge-weighted clique models are natural in structural pattern recognition and graph similarity: pairwise compatibility dominates, and the clique captures mutually consistent correspondences.

The same line of work also links weighted clique search to graph kernels. In the Bron–Kerbosch extension for weighted cliques, an association graph $\omega(G,w)$0 is built with weights

$\omega(G,w)$1

so a maximum-weight clique corresponds to an optimal partial injective matching under the induced kernelized similarity measure (Jain et al., 2011). This places EWMCP and its mixed vertex-edge generalization inside a broader family of kernelized graph comparison methods.

4. Exact algorithms and bounding strategies

Exact EWMCP algorithms are predominantly branch-and-bound methods. The 2018 MECQ algorithm is representative: in a subproblem $\omega(G,w)$2 with current clique $\omega(G,w)$3 and candidate set $\omega(G,w)$4, it dynamically assigns pseudo vertex weights

$\omega(G,w)$5

to each candidate vertex $\omega(G,w)$6 (Shimizu et al., 2018). This converts the remaining search space into a vertex-and-edge-weighted subproblem so that coloring-based upper bounds, originally associated with MWCP, can be applied on the fly (Shimizu et al., 2018).

Given a proper coloring $\omega(G,w)$7 of the candidate subgraph, MECQ defines a lifted score

$\omega(G,w)$8

where $\omega(G,w)$9 is the color index of $0$0 (Shimizu et al., 2018). Because a clique can contain at most one vertex from each color class, the bound

$0$1

holds for any clique in the colored graph (Shimizu et al., 2018). Branches are pruned whenever the current clique weight plus the corresponding upper bound does not exceed the incumbent.

The same paper reports that MECQ consistently achieved optimal solutions faster than EWCLIQUE and MIP on random graphs and many DIMACS-derived EWMCP benchmarks, and that it typically explored only $0$2–$0$3 of the nodes traversed by EWCLIQUE (Shimizu et al., 2018). Representative DIMACS results include brock200_1 with MECQ about $0$4 s versus EWCLIQUE about $0$5 s, and c-fat200-5 with MECQ about $0$6 s versus EWCLIQUE about $0$7 s, although there are exceptions such as hamming8-2 where EWCLIQUE was faster (Shimizu et al., 2018). On structured instances, the use of an initial clique from Phased Local Search (PLS) significantly improved pruning (Shimizu et al., 2018).

A distinct exact-search lineage extends Bron–Kerbosch to weighted clique search. In the general weighted setting with continuous-valued vertex and edge weights, branch-and-bound pruning relies on an admissible estimate $0$8 satisfying

$0$9

for all cliques $Z=(V,\alpha)$0 (Jain et al., 2011). Proposed estimates include a max-weight-degree bound $Z=(V,\alpha)$1 and, for association graphs, a Cauchy–Schwarz bound $Z=(V,\alpha)$2 (Jain et al., 2011). Experiments on IAM graph datasets found that pivoting strategies substantially improved runtime over basic Bron–Kerbosch, and “max-weight clique selection” provided the best trade-off among the tested pivot rules (Jain et al., 2011).

These exact methods illustrate a persistent algorithmic theme: EWMCP search is governed not only by graph structure but by how edge weights can be aggregated into safe, computationally cheap upper bounds. Coloring, pseudo vertex-weight assignment, and admissible kernel-based residual bounds all serve this purpose.

5. Upper bounds, coloring dependence, and incomparability results

A major recent development concerns the comparative strength of the two principal upper bounds used in the EWMCP literature. The first, denoted here as UB1, is LP-based. Starting from the linearized formulation, integrality on vertex variables is relaxed, edge variables remain continuous, and only independent-set constraints corresponding to a fixed coloring $Z=(V,\alpha)$3 are kept (Ciccarelli et al., 9 Jul 2025). In the dual, edge weights are split between their endpoints via nonnegative variables $Z=(V,\alpha)$4, inducing per-vertex accumulations

$Z=(V,\alpha)$5

and the bound takes the form

$Z=(V,\alpha)$6

minimized over feasible splittings (Ciccarelli et al., 9 Jul 2025).

The second bound, UB2, is combinatorial and depends on both a coloring and an ordering of its color classes. If $Z=(V,\alpha)$7 is the index of the class containing $Z=(V,\alpha)$8, define

$Z=(V,\alpha)$9

Then

$\alpha:V\times V\to\mathbb{R}\cup\{\epsilon\}$0

is a valid upper bound on $\alpha:V\times V\to\mathbb{R}\cup\{\epsilon\}$1 (Ciccarelli et al., 9 Jul 2025). Unlike UB1, UB2 changes under reordering of the same color classes (Ciccarelli et al., 9 Jul 2025).

The 2025 study proves that neither bound admits a relative performance guarantee. It constructs a family $\alpha:V\times V\to\mathbb{R}\cup\{\epsilon\}$2 for which both $\alpha:V\times V\to\mathbb{R}\cup\{\epsilon\}$3 and $\alpha:V\times V\to\mathbb{R}\cup\{\epsilon\}$4 are unbounded (Ciccarelli et al., 9 Jul 2025). It also shows strict incomparability between the bounds by giving one family $\alpha:V\times V\to\mathbb{R}\cup\{\epsilon\}$5 where $\alpha:V\times V\to\mathbb{R}\cup\{\epsilon\}$6 and another family $\alpha:V\times V\to\mathbb{R}\cup\{\epsilon\}$7 where $\alpha:V\times V\to\mathbb{R}\cup\{\epsilon\}$8 (Ciccarelli et al., 9 Jul 2025). This rules out any universal dominance claim for either method.

Empirically, however, UB2 was stronger on the tested benchmarks. On DIMACS instances, the average percentage difference was about $\alpha:V\times V\to\mathbb{R}\cup\{\epsilon\}$9 in favor of UB2, with extreme cases exceeding $\omega(C)=\sum_{i,j\in C}\alpha(i,j),$0; average gaps to optimum were approximately $\omega(C)=\sum_{i,j\in C}\alpha(i,j),$1 for UB1 and $\omega(C)=\sum_{i,j\in C}\alpha(i,j),$2 for UB2 over instances with known optimum (Ciccarelli et al., 9 Jul 2025). On RANDOM graphs, average gaps were about $\omega(C)=\sum_{i,j\in C}\alpha(i,j),$3 for UB1 and $\omega(C)=\sum_{i,j\in C}\alpha(i,j),$4 for UB2, with the worst behavior around moderate densities and much tighter values on very dense graphs such as $\omega(C)=\sum_{i,j\in C}\alpha(i,j),$5, where both could be about $\omega(C)=\sum_{i,j\in C}\alpha(i,j),$6 (Ciccarelli et al., 9 Jul 2025). UB1 was substantially slower to compute, while UB2 was near-instantaneous; UB2 also varied under reordering by up to about $\omega(C)=\sum_{i,j\in C}\alpha(i,j),$7 on the same coloring (Ciccarelli et al., 9 Jul 2025).

These results clarify a common misconception. It is tempting to treat LP-based relaxations as categorically stronger than combinatorial color-order bounds, or conversely to regard the cheaper combinatorial bound as uniformly preferable. The proven separations show that neither position is correct in general (Ciccarelli et al., 9 Jul 2025). A plausible implication is that bound selection should be instance-aware rather than doctrinal.

6. Applications, assumptions, and scope boundaries

The graph-matching reduction covers exact and inexact matching problems whenever the objective decomposes additively over matched items and the feasible set is $\omega(C)=\sum_{i,j\in C}\alpha(i,j),$8-closed (0912.4584). Examples explicitly discussed include maximum common induced subgraph, maximum common homomorphic subgraph, best common subgraph, probabilistic graph matching, and graph edit distance (GED) (0912.4584).

For GED, edit operations on vertices and edges—substitution, deletion, insertion—are encoded through dummy extensions of the input graphs. Dummy vertices and dummy edges are added, a modified substitution cost is defined, and a compatibility function $\omega(C)=\sum_{i,j\in C}\alpha(i,j),$9 is constructed so that maximizing

$w_{ij}$00

over total monomorphisms on the dummy-extended graphs is equivalent to minimizing edit cost on the original pair (0912.4584). The resulting association graph yields an MWCP instance, and this reduces to EWMCP when unary terms vanish or can be normalized away (0912.4584).

The principal assumptions and limitations are explicit. The 2009 equivalence theorem requires that the feasible morphism set be describable entirely by a local property $w_{ij}$01 on item pairs; globally constrained sets such as “$w_{ij}$02” are not generally $w_{ij}$03-closed, and equivalence may fail (0912.4584). Likewise, objectives with higher-order consistency terms beyond pairwise item compatibilities are outside the pairwise association-graph construction and would require an extended hypergraph-style model (0912.4584).

Most algorithmic and comparative results for EWMCP also assume nonnegative weights. The 2018 branch-and-bound algorithm is designed for nonnegative symmetric edge weights and identifies negative-weight handling as future work rather than a feature of the presented method (Shimizu et al., 2018). The 2025 upper-bound study likewise states that all results and proofs assume $w_{ij}$04, and that the linearization and dominance claims rely on this assumption (Ciccarelli et al., 9 Jul 2025). Although the weighted Bron–Kerbosch framework allows real weights in the more general MWCP setting, the identification of maximal weight cliques with inclusion-wise maximal cliques ceases to hold when negative weights are present, so plain enumeration is no longer sufficient (Jain et al., 2011).

The distinction between EWMCP and general MWCP is therefore not merely terminological. If unary terms are essential and vary with the feasible solution, they must be retained as vertex weights in MWCP rather than discarded (0912.4584). EWMCP is the correct model only when the structure of the compatibility function justifies an edge-only objective.

7. Research directions and practical interpretation

Current research presents EWMCP as a mature but still actively refined NP-hard optimization problem. Exact search has benefited from dynamic weight reassignment, coloring-based pruning, and weighted variants of Bron–Kerbosch (Shimizu et al., 2018, Jain et al., 2011). At the same time, recent work shows that the principal upper bounds used in exact algorithms are theoretically incomparable and can both be arbitrarily weak relative to the optimum (Ciccarelli et al., 9 Jul 2025).

Practically, the recent comparative evidence favors UB2 as the default bound because it is usually tighter and dramatically cheaper on DIMACS and random benchmarks, while UB1 remains relevant as a complementary tool on structures where LP weight-splitting better captures the instance geometry (Ciccarelli et al., 9 Jul 2025). The same study points to hybrid strategies, better ordering heuristics for UB2, coloring strategies tuned specifically for bounding, and stronger convex relaxations as open directions (Ciccarelli et al., 9 Jul 2025). This suggests that progress is likely to come less from a single universally superior bound than from adaptive combinations of color-aware relaxations and structure-sensitive branching.

Within graph matching, the principal enduring significance of EWMCP is representational. The necessary-and-sufficient equivalence criterion based on $w_{ij}$05-closure turns reduction to clique search from an ad hoc modeling trick into a theorem with explicit scope conditions (0912.4584). When those conditions hold and unary terms disappear, EWMCP becomes the exact combinatorial core of the matching problem. When they do not, the general MWCP—or a richer higher-order model—remains necessary.