Vertex r-Triangle s-Club: Complexity & Algorithms

• Vertex r-Triangle s-Club is a graph model where a vertex set induces a subgraph of diameter at most s and each vertex lies in at least r triangles.
• It exhibits a sharp complexity dichotomy: being fixed-parameter tractable for r=1 with s≥4 while largely W[1]-hard for other settings.
• Research leverages Turing kernels, data reductions, and structural parameterizations to optimize solution size and analyze robustness in dense network graphs.

Vertex $r$-Triangle $s$-Club is a triangle-constrained variant of $s$-Club. Given an undirected graph $G=(V,E)$ and an integer $k$, the decision problem asks whether there exists a vertex set $S\subseteq V$ with $|S|\ge k$ such that the induced subgraph $G[S]$ has diameter at most $s$ and every selected vertex belongs to at least $r$ triangles in $s$0. In the notation of Garvardt et al., this is the paper’s Vertex Triangle $s$1-Club with $s$2; the triangles are counted inside the chosen induced subgraph, not in the ambient graph, and the constraint is an “at least” lower bound (Garvardt et al., 2022).

1. Formal definition and problem variants

Formally, a set $s$3 is feasible for Vertex $s$4-Triangle $s$5-Club if

$s$6

and

$s$7

For $s$8, an equivalent formulation of the diameter condition is that every two selected vertices are adjacent or have a common neighbor in $s$9. In the optimization form studied for $s$0, the objective is to maximize $s$1 subject to the same feasibility constraints. A useful local count is

$s$2

so feasibility is exactly $s$3 for all $s$4 together with $s$5 when $s$6 (Grüttemeier et al., 2022).

The vertex-triangle model is distinct from two nearby families of constraints. In Edge $s$7-Triangle $s$8-Club, the requirement is imposed on edges of a suitable spanning subgraph $s$9, not directly on vertices of the induced graph. Every edge-$G=(V,E)$0-triangle solution is also a vertex-$G=(V,E)$1-triangle solution, but not conversely. Seeded $G=(V,E)$2-Club adds a prescribed set of seed vertices $G=(V,E)$3 that must be included in the solution; those results are treated separately and do not transfer automatically to the vertex-triangle setting (Garvardt et al., 2022).

2. Parameterized complexity by solution size

With parameter $G=(V,E)$4, the required solution size, the central classification is a dichotomy in the pair $G=(V,E)$5. The positive region is extremely narrow: Vertex $G=(V,E)$6-Triangle $G=(V,E)$7-Club is fixed-parameter tractable only for $G=(V,E)$8 and $G=(V,E)$9. It is W[1]-hard for every $k$0 and every $k$1, and also W[1]-hard for $k$2 when $k$3.

Regime	Status for parameter $k$4	Notes
$k$5	FPT	Via Turing kernels
$k$6	W[1]-hard	Hard already at low radius
$k$7	W[1]-hard	Hard for all nontrivial radii

The same line of work also notes classical NP-hardness for all $k$8 and all $k$9. This places the problem in sharp contrast with unconstrained $S\subseteq V$0-Club: ordinary $S\subseteq V$1-Club is W[1]-hard only for $S\subseteq V$2, while for every $S\subseteq V$3 it is FPT parameterized by $S\subseteq V$4. The per-vertex triangle requirement therefore removes most of the positive parameterized behavior enjoyed by the unconstrained problem (Garvardt et al., 2022).

For the special case $S\subseteq V$5, later work on structural parameterizations restates the same negative picture for solution-size parameterization: Vertex $S\subseteq V$6-Triangle 2-Club remains W[1]-hard when parameterized by the target size, even though several other parameters yield positive results (Jacob et al., 19 Sep 2025).

3. The FPT regime $S\subseteq V$7: structural preprocessing and Turing kernels

The tractable case $S\subseteq V$8, $S\subseteq V$9 is based on an adaptation of the classical Turing-kernel approach for $|S|\ge k$0-Club. The key preprocessing step is Reduction Rule 1: delete all vertices that are not part of any triangle. This is safe because every vertex in a feasible vertex-$|S|\ge k$1-triangle solution must lie in a triangle of the induced solution, hence also in a triangle of the input graph.

After this reduction, the crucial structural lemma states that for $|S|\ge k$2, if a vertex $|S|\ge k$3 satisfies

$|S|\ge k$4

then the instance is immediately a yes-instance. The proof constructs a feasible set by starting from $|S|\ge k$5 and, for each vertex on the outer layer, adding two witnesses that form a triangle with it. This expansion keeps all selected vertices within distance at most $|S|\ge k$6 from $|S|\ge k$7, so the resulting induced subgraph remains an $|S|\ge k$8-club while ensuring that every selected vertex lies in a triangle.

This yields case-specific Turing kernels:

• a $|S|\ge k$9-vertex Turing kernel for $G[S]$0 or $G[S]$1,
• a $G[S]$2-vertex Turing kernel for $G[S]$3,
• a $G[S]$4-vertex Turing kernel for $G[S]$5 or $G[S]$6.

The method is inherently tied to the threshold $G[S]$7. For $G[S]$8, adding witnesses for one vertex may create new vertices whose own triangle demands require further additions outside the available radius budget, producing a cascade that destroys the $G[S]$9-club argument. This explains the sharp divide between the $s$0 and $s$1 regimes (Garvardt et al., 2022).

The W[1]-hardness proofs use parameterized reductions from Clique. Their common structure replaces each original vertex by a gadget $s$2 with an “empty-or-complete” property: $s$3 Distances between gadgets are then engineered so that two gadgets can coexist in an $s$4-club only if the corresponding original vertices are adjacent. For $s$5, a construction based on cliques of size $s$6, where $s$7, yields hardness for every $s$8. For $s$9, and for $r$0 with $r$1, a more elaborate gadget family shows hardness even on graphs where each input vertex lies in exactly $r$2 triangles (Garvardt et al., 2022).

4. Exact optimization for $r$3

For $r$4, the optimization problem has received a dedicated exact-algorithmic treatment. The central observation is local: every 2-club is contained in the closed 2-neighborhood $r$5 of each of its vertices. Accordingly, the global instance can be decomposed into local subinstances

$r$6

solve each local problem optimally, and return the largest solution.

The exact solver is a branch-and-bound algorithm with a marked set $r$7 of vertices forced into the solution. Two vertices are called compatible if their distance in the current graph is at most $r$8; otherwise they are incompatible. If an incompatible pair exists, at least one endpoint must be excluded from any feasible 2-club. Rather than branching symmetrically on the pair, the algorithm branches on an unmarked vertex $r$9: either delete $s$00, or mark $s$01. In the second branch, every vertex incompatible with $s$02 can be deleted immediately.

A substantial part of the method is a suite of safe data reductions and upper bounds.

Rule	Effect
LDR	Delete vertices whose degree is too small to support $s$03 triangles
LTR	Delete vertices in fewer than $s$04 triangles
IRR / MIR	Remove vertices incompatible with marked ones; reject branches with incompatible marked pairs
CR / NCR	Mark vertices forced by triangle support or unique distance-2 paths
2-NR / LCR	Delete vertices with too few compatible partners
Matching Rule	Prune if a conflict-graph matching gives an upper bound no better than the incumbent

The conflict graph $s$05 has the same vertex set as the current graph and edges between incompatible pairs. It supports pruning rules such as the Low-Compatibility Rule and a matching-based upper bound. Lower bounds are equally important. The Neighborhood Lower Bound (N-LB) works on closed neighborhoods $s$06 and is exact whenever an optimal solution has a universal vertex. The Greedy Lower Bound (G-LB) works on larger 2-neighborhood instances and greedily repairs incompatibilities. Their combination, Multi-LB, first computes N-LB, applies reductions, and then uses G-LB on the remaining candidates. Theoretical running times reported for core components are $s$07 for exhaustive LTR, $s$08 for conflict-graph construction, $s$09 for N-LB, $s$10 for G-LB, and $s$11 for the decision problem parameterized by $s$12 (Grüttemeier et al., 2022).

The implementation is highly engineered and was evaluated on 67 social, biological, and technical network instances, from fewer than 100 to about 300,000 vertices, for 27 values of $s$13. On sparse real-world graphs, the solver finds optimal solutions for instances with up to roughly $s$14 vertices and $s$15 edges. For Vertex Triangle 2-Club, N-LB is the fastest overall variant, while Multi-LB gives the strongest lower bounds. Relative to the best compared ILP formulation, denoted Fxkany after Almeida and Brás’s formulations, the paper reports that Fxkany and the basic branch-and-bound solve roughly $s$16 of instances within one hour, whereas N-LB solves $s$17 in less than one minute and is reported as more than 60 times faster. Increasing $s$18 tends to accelerate computation because LDR and LTR remove more of the graph early; it also increases density, the global clustering coefficient, and the minimum local clustering coefficient of the returned solutions (Grüttemeier et al., 2022).

5. Structural parameterizations

A later line of work studies Vertex $s$19-Triangle $s$20-Club through structural parameters of the input graph. For $s$21, the main positive theorem is fixed-parameter tractability by treewidth: $s$22 The dynamic program runs on a nice tree decomposition and stores, for each bag $s$23, a subset $s$24 of selected bag vertices, a truncated triangle-count function $s$25, and a family $s$26 encoding bag-neighborhood signatures of forgotten vertices. The state is large because $s$27 ranges over families of subsets of $s$28, producing the doubly exponential dependence on treewidth. The proof then removes the explicit dependence on $s$29 by showing that any feasible solution satisfies

$s$30

The same paper gives an XP algorithm for Vertex $s$31-Triangle 2-Club parameterized by the $s$32-index: $s$33 Its strategy isolates a high-degree core, guesses the solution’s intersection with that core, groups the remaining vertices by identical neighborhoods into the guessed part, and localizes difficult interactions within bounded-radius regions of the low-degree remainder.

For all $s$34, not just $s$35, the strongest kernelization result is for the feedback edge number $s$36. Vertex $s$37-Triangle $s$38-Club admits a kernel with at most

$s$39

vertices and

$s$40

edges. The underlying reason is that every triangle must use a feedback edge, and each feedback edge can support at most one triangle-relevant vertex outside the set of feedback-edge endpoints. The same study also records an MSO/Courcelle-based FPT result for parameter $s$41, polynomial-time solvability for Vertex $s$42-Triangle 2-Club when deleting one vertex makes the graph bipartite, and an FPT algorithm parameterized by vertex cover together with the lower bound that no polynomial kernel exists for that parameter unless $s$43 (Jacob et al., 19 Sep 2025).

A structurally useful lemma in this context is the twin augmentation property: if $s$44 is an $s$45-club satisfying the vertex $s$46-triangle property, $s$47, $s$48, and $s$49, then $s$50 is also feasible. This supports algorithms that compress vertices into identical-neighborhood classes (Jacob et al., 19 Sep 2025).

6. Density, robustness, and open directions

The triangle requirement enforces local density, but it does not by itself guarantee robustness. In a vertex-$s$51-triangle $s$52-club, the minimum degree is larger than $s$53, yet this lower bound is insufficient to prevent fragile solutions. A concrete example is two $s$54-cliques connected by a single edge: the resulting graph is a vertex-$s$55-triangle $s$56-club, but deleting that single inter-clique edge disconnects it. This clarifies a common misconception: high per-vertex triangle support is not equivalent to edge robustness or fault tolerance (Garvardt et al., 2022).

That distinction is precisely why the edge-triangle model is treated separately. The proposition concerning robustness under up to $s$57 edge deletions applies to edge-$s$58-triangle $s$59-clubs, not to the vertex version. Likewise, the stronger positive statement that the case $s$60 is fixed-parameter tractable for each $s$61 belongs to the edge variant, not to Vertex $s$62-Triangle $s$63-Club. For the vertex version, the exact boundary under parameter $s$64 remains the narrow tractable region $s$65, $s$66 versus W[1]-hardness elsewhere (Garvardt et al., 2022).

The most developed exact and heuristic machinery is also strongly tied to $s$67. Incompatibility testing there reduces to checking whether two vertices are adjacent or share a common neighbor, neighborhood-based lower bounds exploit the fact that a closed neighborhood is automatically a 2-club, and the No-Choice Rule depends on unique common neighbors. These ingredients do not carry over directly to $s$68 (Grüttemeier et al., 2022).

Open problems remain substantial. For every $s$69, the FPT status of Vertex $s$70-Triangle $s$71-Club parameterized by treewidth is open. Beyond the special case where distance to bipartite graph is $s$72, the status for larger distances is also open. Suggested further parameters include distance to cluster, distance to co-cluster, and degeneracy. Collectively, these questions indicate that the interaction between bounded diameter and per-vertex triangle support is understood in several sharp special cases, but not yet in a general structural theory (Jacob et al., 19 Sep 2025).