Enum Vertex Cover & Kernelization

• Enum Vertex Cover is a parameterized enumeration problem that lists all vertex covers of size at most k with output-sensitive complexity guarantees.
• It employs polynomial-delay kernelization methods—using crown decompositions and Nemhauser–Trotter reductions—to compress graphs efficiently.
• Recent advances introduce both optimal 2k-vertex PDE-kernels and flexible 3k-vertex PD kernels, extending to structural parameterizations like feedback vertex set and treedepth.

Enum Vertex Cover is the parameterized enumeration problem that, given a graph $G$ and an integer $k$, asks for all vertex covers of $G$ of size at most $k$, without duplicates and under output-sensitive complexity guarantees. It sits at the intersection of parameterized complexity, kernelization, and enumeration algorithms: the decision version of Vertex Cover has long served as a canonical fixed-parameter benchmark, while the enumeration version requires additional machinery to reconstruct the full solution space from compressed instances. Recent work establishes both an optimal-size polynomial-delay enumeration kernel with $2k$ vertices for the standard parameterization by solution size and a more flexible kernelization framework that extends to several structural parameterizations (Bougeret et al., 10 Sep 2025, Bougeret et al., 25 Apr 2026).

1. Problem formulation and output-sensitive complexity

In its standard form, Enum Vertex Cover is defined as follows: input a simple undirected graph $G$ and an integer $k$; enumerate all subsets $S \subseteq V(G)$ such that $S$ is a vertex cover and $|S| \le k$. The required output condition is exhaustive and duplicate-free enumeration. The standard performance measure is polynomial delay, meaning that the time before the first output, between consecutive outputs, and after the last output is bounded by a polynomial in $k$0, independently of the number of solutions (Bougeret et al., 10 Sep 2025).

The problem is naturally parameterized by $k$1, the maximum allowed size of the cover. In the broader parameterized-enumeration framework, one also considers FPT-delay, where delay is bounded by $k$2 for some computable function $k$3. This distinction is central in kernelization models for enumeration, because a compressed instance must not only preserve the existence of solutions, but must also support efficient reconstruction of all original solutions (Bougeret et al., 25 Apr 2026).

Enum Vertex Cover is dual to an independent-set enumeration problem. Specifically, a set $k$4 is a vertex cover of $k$5 of size at most $k$6 if and only if $k$7 is an independent set of size at least $k$8. This complementarity is used repeatedly in structural parameterizations, especially when modulators to forests, bounded treedepth graphs, or minor-closed classes are given explicitly (Bougeret et al., 25 Apr 2026).

2. Polynomial-delay enumeration kernels of optimal size

A major development in the area is the existence of a polynomial-delay enumeration kernel with at most $k$9 vertices for Enum Vertex Cover parameterized by $G$0. In the PDE-kernel model, a compression algorithm produces an equivalent instance $G$1 of size bounded by a function of $G$2, and a lifting algorithm takes a kernel solution $G$3 and outputs, with polynomial delay, a non-empty block of original solutions. These blocks must partition the entire solution space, which guarantees completeness and duplicate-freedom (Bougeret et al., 10 Sep 2025).

For Enum Vertex Cover, the compression reuses the classical kernelization backbone of decision Vertex Cover. Isolated vertices are removed by a degree-0 rule, and large graphs are reduced via crown decompositions derived from Nemhauser–Trotter structure. If a graph without isolated vertices has at least $G$4 vertices, then either no half-integral vertex cover of weight at most $G$5 exists, or a crown decomposition of width at most $G$6 can be found in polynomial time. Repeated application of these reductions yields a kernel graph $G$7 with $G$8 (Bougeret et al., 10 Sep 2025).

What distinguishes the enumeration setting from the decision setting is the lifting layer. In decision kernelization, deleting the crown and decrementing the parameter suffices. In enumeration, each kernel solution can lift to many original solutions because deleted vertices may be optionally reintroduced while respecting the size bound. The 2025 result resolves this by designing a non-trivial lifting algorithm compatible with polynomial delay and shows that the $G$9-vertex bound known from decision kernelization can also be achieved for enumeration (Bougeret et al., 10 Sep 2025).

The same work identifies the resulting kernel as extension-only: every lifted solution is obtained from a kernel solution by adding vertices from the deleted part of the graph, never by removing kernel vertices. This extension-only property immediately implies disjointness of lifted solution families for different kernel solutions and is therefore especially well suited to subset problems such as Vertex Cover (Bougeret et al., 10 Sep 2025).

3. Crown structure, propagation, and lifting

The technical core of the $k$0-vertex PDE-kernel lies in the enumeration of solutions in crowned graphs. A crowned graph is a graph whose vertex set is partitioned into $k$1, where $k$2 is an independent set and there exists a matching between $k$3 and $k$4 saturating $k$5. A crown decomposition $k$6 of a general graph further requires $k$7, with $k$8 the remaining body. In such structures, every minimum vertex cover of size $k$9 intersects each matching edge in exactly one endpoint (Bougeret et al., 10 Sep 2025).

The lifting problem for a single crown reduction is reduced to an intermediate enumeration problem, Enum Crown, which asks for all vertex covers of a crowned graph of size exactly $2k$0. A further special case, Enum Small Crown, considers crowned graphs with $2k$1 and enumerates covers of size exactly $2k$2. The full lifting mechanism proceeds by showing that polynomial-delay enumeration of Enum Small Crown implies polynomial-delay enumeration of Enum Crown, which in turn validates crown reduction as a kernel rule in the PDE sense (Bougeret et al., 10 Sep 2025).

The crucial combinatorial device is a propagation procedure on small crowned graphs. Given a matching $2k$3 saturating $2k$4 and a prescribed subset $2k$5, the procedure constructs a directed bipartite graph in which matching edges are oriented from $2k$6 to $2k$7 and non-matching edges from $2k$8 to $2k$9. Reachability from $G$0 yields a set of forced head vertices $G$1 and forbidden crown vertices $G$2. Any solution extending $G$3 must include all of $G$4 and avoid all of $G$5, while the remaining instance is again a smaller small-crown graph (Bougeret et al., 10 Sep 2025).

This yields a recursive branching algorithm for Enum Small Crown. One branches on a head vertex $G$6, either forcing $G$7 into the solution or forcing its avoidance; each branch invokes propagation, shrinks the crowned graph, and recurses. Because the recursion depth is at most $G$8 and each leaf corresponds to a solution, the algorithm has polynomial delay. Combined with the crown reduction and Nemhauser–Trotter-based compression, this produces the full $G$9-vertex PDE-kernel for Enum Vertex Cover (Bougeret et al., 10 Sep 2025).

4. PD kernels and a more flexible kernelization model

A subsequent development proposes polynomial-delay kernels (PD kernels) as a more permissive alternative to the strong PDE-kernel model. Earlier enum-kernels were shown to be too permissive, in the sense that every problem with an FPT-delay algorithm admits a constant-size enum-kernel. Strong PDE-kernels corrected this by requiring polynomial-delay lifting and insisting that every kernel solution lift to at least one original solution, but this requirement was argued to be too restrictive in practice. PD kernels relax exactly this point: some kernel solutions may be “bad” and lift to the empty set, provided that the non-empty lifted blocks form a partition of the original solution space (Bougeret et al., 25 Apr 2026).

The model preserves the essential complexity-theoretic correspondence: a parameterized enumeration problem admits an FPT-delay algorithm if and only if it admits a PD kernel, and it admits a polynomial-delay algorithm if and only if it admits a constant-size PD kernel. This places PD kernels between strong enumeration kernels and the older enum-kernels: more expressive than the former, but still preserving meaningful kernel size as an algorithmic invariant (Bougeret et al., 25 Apr 2026).

The paper also develops a generic framework for graph vertex-subset problems. A decision kernel can be lifted to a PD kernel of the same size if three ingredients are available: a preserved core with a trace map from kernel to original instance, a polynomial-time test that identifies a unique canonical solution of the kernel for each good trace, and a polynomial-delay enumerator that expands a trace into all original solutions sharing that trace. In effect, the framework separates compression from trace-sensitive reconstruction (Bougeret et al., 25 Apr 2026).

Applied to Enum Vertex Cover parameterized by $k$0, this yields a particularly simple PD kernel based on a heavy-crown decomposition. The compression exhaustively removes isolated vertices and, if the graph has more than $k$1 vertices, either finds a matching of size $k$2 and concludes NO or finds a heavy crown and deletes unmatched crown vertices. The resulting kernel has at most $k$3 vertices. Unlike the $k$4-vertex strong kernel, its lifting algorithm is short: given a kernel solution, one adds exactly those deleted vertices that are forced to cover edges left uncovered by the trace and then enumerates subsets of the remaining undecided deleted vertices subject to the size bound (Bougeret et al., 25 Apr 2026).

This comparison is methodologically significant. The $k$5-vertex result shows that strong PDE-kernels can match the optimal decision-kernel size. The $k$6-vertex PD kernel shows that relaxing the requirement that every kernel solution be productive can simplify lifting substantially, while still preserving polynomial-delay reconstruction and a linear-size kernel (Bougeret et al., 10 Sep 2025, Bougeret et al., 25 Apr 2026).

5. Structural parameterizations and kernelization beyond $k$7

The newer PD-kernel framework extends Enum Vertex Cover beyond the standard parameterization by solution size. One line of results starts from feedback vertex number. Via the dual Enum Independent Set formulation, if a feedback vertex set $k$8 is given and $k$9 is a forest, then a cubic-size PD kernel exists. The reduction adapts the decision kernel of Jansen and Bodlaender, uses conflict measures on the forest part, and relies on lexicographic polynomial-delay enumeration in forests to lift traces back to full solutions (Bougeret et al., 25 Apr 2026).

A second line concerns modulators to bounded treedepth. For every fixed $S \subseteq V(G)$0, Enum Independent Set parameterized by the size of a modulator to treedepth-$S \subseteq V(G)$1 graphs admits a polynomial-size PD kernel, obtained through an annotated independent-set variant and an enumerative bikernelization argument. By complementarity, this yields a polynomial-size PD kernel for Enum Vertex Cover under the same parameterization (Bougeret et al., 25 Apr 2026).

A third line concerns modulators to bounded bridgedepth. If $S \subseteq V(G)$2 has bridgedepth at most $S \subseteq V(G)$3, then Enum Independent Set and hence Enum Vertex Cover admit polynomial-size PD kernels parameterized by $S \subseteq V(G)$4. The compression adapts the decision-kernel machinery for minor-closed classes, while lifting again works through traces on the original modulator and bounded-width enumeration in the residual structure (Bougeret et al., 25 Apr 2026).

These results culminate in a structural dichotomy for minor-closed classes. Let $S \subseteq V(G)$5 be minor-closed, and parameterize by the size of a given modulator $S \subseteq V(G)$6 such that $S \subseteq V(G)$7. Assuming $S \subseteq V(G)$8, Enum Vertex Cover admits a polynomial-size PD kernel if and only if $S \subseteq V(G)$9 has bounded bridgedepth. This is an enumeration analogue of the decision-kernel dichotomy of Bougeret, Jansen, and Sau and identifies bounded bridgedepth as the structural boundary for polynomial PD kernels in this setting (Bougeret et al., 25 Apr 2026).

The following summary captures the principal kernelization results explicitly stated in the literature.

Parameterization	Enumeration-kernel result	Source
Solution size $S$0	PDE-kernel with at most $S$1 vertices	(Bougeret et al., 10 Sep 2025)
Solution size $S$2	PD kernel with at most $S$3 vertices	(Bougeret et al., 25 Apr 2026)
Given feedback vertex set $S$4	Cubic-size PD kernel	(Bougeret et al., 25 Apr 2026)
Modulator to treedepth-$S$5	Polynomial-size PD kernel for fixed $S$6	(Bougeret et al., 25 Apr 2026)
Modulator to $S$7-bridgedepth	Polynomial-size PD kernel for fixed $S$8	(Bougeret et al., 25 Apr 2026)

6. Refined parameterizations and near-optimal enumeration

Enum Vertex Cover is also informed by work on decision versions of Vertex Cover parameterized above or below tight combinatorial bounds. One study analyzes degree-based, matching-based, and maximal-matching-based bounds and classifies the resulting decision problems as fixed-parameter tractable or W-hard. In particular, VCL1—Vertex Cover Above Tight Lower Bound-1 on graphs of maximum degree $S$9—is solvable in time $|S| \le k$0. The proof establishes that if a cover of size at most $|S| \le k$1 exists, then deleting at most $|S| \le k$2 edges makes the graph bipartite; enumeration can then proceed by brute-forcing the small non-bipartite core and solving the remainder exactly in polynomial time. The same work shows that the capacitated analogue above $|S| \le k$3 is W[2]-hard, that the matching-based parameterization above the maximum matching lower bound is FPT, and that parameterization below the maximal-matching upper bound $|S| \le k$4 is W[1]-hard (0907.4488).

These classifications do not themselves yield canonical polynomial-delay algorithms, but they delimit the algorithmic geography for near-optimal enumeration. A plausible implication is that “above lower bound” formulations such as $|S| \le k$5 or $|S| \le k$6 are substantially better candidates for FPT-total-time or FPT-delay enumeration than “below upper bound” formulations such as $|S| \le k$7, where even decision is already W-hard (0907.4488).

A related line of work studies Vertex Cover above structural guarantees such as $|S| \le k$8-index, degeneracy, clique number, chromatic number, feedback vertex number, cluster deletion number, and treewidth. It proves that Vertex Cover parameterized by $|S| \le k$9 is FPT, which by parameter dominance also yields FPT for parameterizations above degeneracy, clique number, and chromatic number. It further proves that Vertex Cover above treewidth is FPT on planar graphs. On the hardness side, Vertex Cover above feedback vertex number is W[1]-hard, Vertex Cover above cluster deletion number is NP-hard even for parameter value $k$00, and Vertex Cover above minimum degree or clique number has no polynomial kernel unless $k$01 (Kellerhals et al., 2022).

For enumeration, these results function primarily as feasibility indicators. The paper explicitly notes that it does not study enumeration, but its branching and dynamic-programming arguments suggest that high-degree structure, degeneracy, and planar treewidth may support output-sensitive enumeration schemes, while feedback vertex number and cluster deletion number are poor generic “above guarantee” parameters because the corresponding decision problems are already hard (Kellerhals et al., 2022).

7. Limitations, variants, and open directions

Despite the recent progress, several aspects of Enum Vertex Cover remain open. The $k$02-vertex PDE-kernel is optimal in size relative to the best known decision kernel, but lower-bound theory for enumeration kernels is still described as being in its infancy. The same paper emphasizes that its techniques enumerate all vertex covers of size at most $k$03, not minimal vertex covers, and that extending the lifting machinery to preserve minimality is non-trivial (Bougeret et al., 10 Sep 2025).

The PD-kernel framework simplifies many constructions, but it also leaves a number of structural directions unresolved. The general theory of lower bounds for PD kernels remains limited, and the attraction of the model is precisely that it allows “bad” kernel solutions to be ignored during lifting. This suggests that future work will likely concentrate on identifying which decision-kernel techniques can be transferred to enumeration through trace-based lifting and which require fundamentally new enumerative ideas (Bougeret et al., 25 Apr 2026).

Open complexity questions also arise from refined parameterizations. In the above/below-tight-bounds setting, the parameterization denoted VCU, based on the bound $k$04, is left open in general. In the structural-guarantee setting, Vertex Cover above treewidth is proved FPT only for planar graphs, while its status on general graphs and general minor-free classes remains open (0907.4488, Kellerhals et al., 2022).

A final limitation concerns the relation between exact enumeration and heuristic optimization. Quantum-walk-based heuristics for minimum vertex cover have recently been proposed, but they are explicitly single-solution greedy heuristics rather than enumeration algorithms; they may generate high-quality candidate covers, yet they provide neither coverage guarantees over the solution space nor polynomial-delay enumeration (Luiz et al., 2 Dec 2025). This sharpens the role of Enum Vertex Cover as a distinctly combinatorial problem: the central issues are not merely finding one good cover, but compressing the instance while preserving the entire bounded-size solution space and reconstructing that space with rigorous output-sensitive guarantees.