Enum Feedback Vertex Set

• Enum Feedback Vertex Set is a parameterized enumeration problem that lists every vertex subset of size at most k whose removal results in an acyclic graph.
• Recent approaches leverage polynomial-delay enumeration kernels, search-tree strategies, and iterative compression to efficiently traverse the solution space.
• Compact representations and structural decompositions in undirected, directed, mixed, and hypergraph settings underpin effective Delay-FPT algorithms.

Enum Feedback Vertex Set denotes the family of enumeration problems centered on listing vertex subsets whose deletion destroys every cycle. In the standard parameterized formulation on an undirected multigraph $G=(V,E)$ and integer $k$, the task is to enumerate every feedback vertex set $S\subseteq V(G)$ with $|S|\le k$, where $G-S$ is a forest. Closely related literatures study minimal or minimum feedback vertex sets, directed feedback vertex sets, mixed-graph variants, and hypergraph variants, and the resulting algorithmics combine kernelization, search trees, iterative compression, dynamic programming, and compact solution representations (Bougeret et al., 10 Sep 2025, Arhire et al., 2023, Bonsma et al., 2010, Diao et al., 2018).

1. Formal models and problem variants

In the undirected multigraph setting, a feedback vertex set is a subset $S\subseteq V(G)$ such that $G-S$ is acyclic, and the canonical enumeration problem asks to output every such $S$ with $|S|\le k$. The multigraph model used in the enumeration-kernel literature allows multiple edges, while loops are allowed only transiently before being reduced. A closely related notion used in reconfiguration-style work is that of a minimal feedback vertex set, meaning that no proper subset is itself a feedback vertex set (Bougeret et al., 10 Sep 2025, Saito et al., 28 Nov 2025).

Search-tree formulations often enrich the instance with an undeletable set. In the extended-instance model $(G,k,F)$, one seeks a solution $k$0 such that $k$1 is a forest and $k$2 is already a forest. This viewpoint is particularly natural for branching and for iterative compression, because “include $k$3” and “exclude $k$4” become “delete $k$5” and “move $k$6 into $k$7” (Cao, 2017).

The same deletion paradigm extends beyond undirected graphs. For directed graphs, the directed feedback vertex set problem asks for $k$8 such that the residual digraph $k$9 is acyclic, where cycles are directed cycles and loops count as cycles. For mixed graphs $S\subseteq V(G)$0, one must hit every undirected, directed, or mixed cycle. For 3-uniform hypergraphs, the relevant cycle notion is Berge-type: a feedback vertex set is a set $S\subseteq V(G)$1 such that $S\subseteq V(G)$2 has no cycle, and the minimum size is denoted $S\subseteq V(G)$3 (Arhire et al., 2023, Bonsma et al., 2010, Diao et al., 2018).

These variants are not interchangeable. A plausible implication is that “Enum Feedback Vertex Set” is best viewed as an umbrella term whose core object is the acyclic remainder, while the exact meaning of cycle, minimality, and admissible outputs depends on the ambient combinatorial model.

2. Parameterized enumeration and enumeration kernels

The most explicit formalization of Enum Feedback Vertex Set is as a parameterized enumeration problem $S\subseteq V(G)$4 in which an instance is a multigraph $S\subseteq V(G)$5 with an integer $S\subseteq V(G)$6, the parameter is $S\subseteq V(G)$7, and $S\subseteq V(G)$8 is the set of all feedback vertex sets of size at most $S\subseteq V(G)$9. The target complexity notion is polynomial delay: precomputation, delay between two consecutive outputs, and postcomputation are all required to be polynomial in the input size; in the parameterized setting this yields Delay-FPT behavior. The central kernel notion is the polynomial-delay enumeration kernel (PDE-kernel) of Golovach et al.: a compression algorithm $|S|\le k$0 maps the instance to a bounded-size instance $|S|\le k$1, and a lifting algorithm $|S|\le k$2 maps each kernel solution $|S|\le k$3 to a non-empty set $|S|\le k$4, with the family $|S|\le k$5 forming a partition of the original solution set. For graph subset problems, an extension-only PDE-kernel is one in which lifting only adds vertices outside the kernel (Bougeret et al., 10 Sep 2025).

For Enum Feedback Vertex Set, the current benchmark is a polynomial-delay enumeration kernel with $|S|\le k$6 vertices; more precisely, the reduced instance $|S|\le k$7 satisfies

$|S|\le k$8

and $|S|\le k$9. The lifting algorithm enumerates, with polynomial delay, a non-empty set $G-S$0 for each kernel solution $G-S$1, and the family $G-S$2 partitions $G-S$3. This yields a Delay-FPT algorithm for listing all feedback vertex sets of size at most $G-S$4 (Bougeret et al., 10 Sep 2025).

A central methodological distinction separates enumeration kernels from classical decision kernels. Decision kernelization for Feedback Vertex Set can use reductions based on $G-S$5-expansion and virtual edges, as in Thomassé’s quadratic kernel, because it only needs to preserve existence of a solution. Enumeration cannot use the same machinery directly: the compression must preserve the entire solution space together with a lifting map that reconstructs every solution exactly once. The cubic enumeration kernel arises precisely because the paper avoids virtual edges and instead restricts itself to reductions with explicit, local lifting procedures (Bougeret et al., 10 Sep 2025).

3. Exact search, kernelization, and compressed search spaces

A basic exact-search foundation is the highest-degree branching algorithm on extended instances $G-S$6. After deleting vertices of degree $G-S$7 and forcing vertices that close a cycle inside $G-S$8, the algorithm branches on an undecided vertex $G-S$9 of maximum degree: either $S\subseteq V(G)$0 is included in the feedback vertex set or $S\subseteq V(G)$1 is moved into $S\subseteq V(G)$2. The improved analysis shows running time $S\subseteq V(G)$3, while an earlier simpler variant gives $S\subseteq V(G)$4. The same include/exclude search tree is directly compatible with enumeration, because exploring both branches exhaustively yields a backtracking structure over all solutions of size at most $S\subseteq V(G)$5 (Cao, 2017).

A more refined structural basis comes from Disjoint Feedback Vertex Set. Here the vertex set is partitioned into $S\subseteq V(G)$6, with both $S\subseteq V(G)$7 and $S\subseteq V(G)$8 forests, and the task is to find a feedback vertex set entirely inside $S\subseteq V(G)$9. The associated branch-and-search algorithm uses the measure

$G-S$0

where $G-S$1 is the number of nice $G-S$2-vertices and $G-S$3 is the number of connected components of $G-S$4. It combines safe degree-based reductions, a $G-S$5-vertex kernel for DISJOINT-SMALLER-FVS, and a polynomial-time algorithm for the degree-$G-S$6 case via cographic matroid parity. This yields an $G-S$7-time algorithm for general undirected Feedback Vertex Set through iterative compression (Cao et al., 2010).

The enumeration-kernel framework builds on a further layer of local reductions. Rules FVS.1–FVS.8 cover easy cases, local contractions, pending double-edges, multiflags, flowers, auxiliary double-edge insertion, and edge deletion. Their analysis includes the lower bound

$G-S$8

for reduced instances of maximum degree $G-S$9, and a degree bound

$S$0

once the high-degree reductions are exhausted. Together these imply the cubic bound on the kernel size. This suggests that exact enumeration on undirected graphs is now organized around a layered pipeline: local reductions to preserve the whole solution set, a bounded kernel, and then kernel-level enumeration with explicit lifting (Bougeret et al., 10 Sep 2025).

4. Compact representations and implicit enumeration of minimal solutions

A different route to enumeration-relevant structure arises in Feedback Vertex Set Discovery, where one is given a connected graph $S$1, an initial token configuration $S$2 of size $S$3, and a budget $S$4, and asks whether $S$5 can be transformed into a feedback vertex set using at most $S$6 token slides. This is not classical enumeration, since only the final configuration must be a feedback vertex set, but its algorithmic core is directly informative for enumerative work on small minimal feedback vertex sets (Saito et al., 28 Nov 2025).

The key obstacle is that the number of minimal feedback vertex sets of size at most $S$7 can be $S$8. The FPT algorithm therefore does not enumerate them explicitly. Instead it uses $S$9-compact representations of minimal feedback vertex sets: a family $|S|\le k$0 of pairwise disjoint subsets of $|S|\le k$1 with $|S|\le k$2, such that every choice of exactly one vertex from each $|S|\le k$3 forms a minimal feedback vertex set. Misra et al.’s algorithm computes a complete list $|S|\le k$4 of such representations in time $|S|\le k$5, with $|S|\le k$6 (Saito et al., 28 Nov 2025).

For each representation $|S|\le k$7, the algorithm constructs a complete bipartite graph $|S|\le k$8 with bipartition $|S|\le k$9 and edge weights

$(G,k,F)$0

A minimum-weight matching saturating $(G,k,F)$1 has total weight at most $(G,k,F)$2 if and only if some minimal feedback vertex set represented by $(G,k,F)$3 is reachable from the initial configuration within budget $(G,k,F)$4. The overall running time is $(G,k,F)$5 (Saito et al., 28 Nov 2025).

For enumeration, this suggests a precise distinction between explicit and implicit listing. Explicit enumeration outputs each minimal feedback vertex set individually. Compact representations instead enumerate families of minimal feedback vertex sets in FPT number, and each family can then be expanded by choosing one vertex from each slot $(G,k,F)$6. A plausible implication is that compact representations form a natural interface between parameterized decision, parameterized enumeration, and constrained generation of minimal solutions.

5. Directed, mixed, and hypergraph generalizations

In directed graphs, the solver submitted as UAIC_ANDREIARHIRE to the PACE 2022 heuristic track does not target enumeration, but several of its ingredients are immediately relevant to it. The solver has three stages: construction of an initial DFVS using reductions and greedy selection, pruning of redundant vertices, and local search through repeated subgraph reconstructions. The degree scores

$(G,k,F)$7

and

$(G,k,F)$8

prioritize cycle-critical vertices. Stage 2 tests whether a chosen vertex is essential by reinserting it and checking acyclicity via an optimized Tarjan SCC algorithm, which is very close to a minimality filter. Stage 3 repeatedly perturbs the current best solution and re-solves subgraphs; the paper explicitly notes that this inherently explores multiple solutions and is useful to think about for enumeration, even though only the best solution is retained. The stated running time is

$(G,k,F)$9

where $k$00 is the number of local-search iterations (Arhire et al., 2023).

For mixed graphs, the first FPT algorithm covering both undirected and directed edges uses iterative compression, S-Disjoint FVS, a backbone graph on the undirected part, a reduction to FVS/UMC, and finally a reduction to the Skew Separator problem on directed graphs. The exact running time is

$k$01

and the algorithm is constructive: if a solution exists, it returns one. The paper does not present enumeration guarantees, but its backbone decomposition and bounded branching on connection paths show how a small mixed-graph feedback vertex set can be localized before any systematic listing procedure is attempted (Bonsma et al., 2010).

For 3-uniform hypergraphs, the main contribution is structural rather than algorithmic. If $k$02 is a 3-uniform hypergraph with $k$03 edges, then $k$04; if $k$05 is linear and 3-uniform, then $k$06; and equality in the $k$07 bound holds if and only if every component of $k$08 is a 2-cycle. The paper also proves $k$09 for feedback edge sets. These bounds do not enumerate feedback vertex sets, but they sharply constrain the search space and identify extremal components that can be handled separately (Diao et al., 2018).

6. Structural regimes, tractable classes, and hardness boundaries

Difference-parameter results show that the tractability of Feedback Vertex Set depends strongly on how the solution size is measured against graph structure. Parameterizing by $k$10, where $k$11 is degeneracy, yields an FPT algorithm with running time

$k$12

The branching is driven by a core $k$13 with minimum degree $k$14, and each branch commits many core vertices to the solution. By contrast, Feedback Vertex Set above clique number has no polynomial kernel unless $k$15, Feedback Vertex Set below vertex cover is W[1]-hard even for $k$16, and parameterization by $k$17 is the Maximum Induced Forest problem, known W[1]-hard. For enumeration, this delineates promising and unpromising parameter regimes: above-degeneracy search appears structurally compressible, whereas below-guarantee regimes do not (Kellerhals et al., 2022).

On special graph classes, optimization can be dramatically easier. On chordal bipartite graphs, Feedback Vertex Set is solvable in polynomial time. The structural basis is unusually rigid: minimal separators induce complete bipartite graphs, the number of minimal separators is $k$18, neighborhoods can be treated as hyperedges in a maximal chordal bipartite embedding, and a replacement lemma states that if $k$19 satisfy $k$20, then any feedback vertex set containing $k$21 and not $k$22 can be transformed by swapping $k$23 for $k$24. In chains of nested neighborhoods this yields a threshold property for minimum solutions. The resulting algorithm is a dynamic program over a hierarchy of $k$25-trees on one bipartition side, and its structure is strongly suggestive of enumeration by DP backtracking, even though the paper itself only proves polynomial-time optimization (Kloks et al., 2011).

A common misconception is that structural tractability for classical Feedback Vertex Set transfers automatically to configuration-sensitive or richer solution-generation variants. It does not. Feedback Vertex Set Discovery is NP-complete on chordal graphs and on graphs of diameter $k$26, yet all three discovery problems for vertex cover, independent set, and feedback vertex set become polynomial-time solvable on split graphs; for split graphs, the positive result proceeds by polynomial-time enumeration of minimal solutions followed by a minimum-weight matching test (Saito et al., 28 Nov 2025).

Overall, Enum Feedback Vertex Set sits at the intersection of three methodological themes. The first is enumerative kernelization, where the current undirected benchmark is a cubic PDE-kernel. The second is compressed representation of solution families, most explicitly through compact representations of minimal feedback vertex sets. The third is structural decomposition, whether through iterative compression, bounded-degree branching, SCC-based minimality filters, clique-width dynamic programming, or hyperedge decompositions in specialized graph classes. The present landscape is therefore less a single algorithm than a set of interoperable paradigms, each tailored to a different notion of cycle, output family, and parameter regime.