Degeneracy Graphs: Concepts and Insights

• Degeneracy graphs are defined by the maximum minimum degree in any subgraph, computed via k-core decompositions and serving as a key hereditary sparsity measure.
• The spectral variant bounds eigenvalues by relating the spectral radius to maximum degrees, influencing extremal results and coloring bounds.
• Weak and strict degeneracy variants offer practical algorithmic insights for efficient coloring and ordering in planar, sparse, and random graph models.

Degeneracy is a hereditary sparsity parameter that measures the largest minimum degree attained by a subgraph. A graph $G$ is $d$-degenerate when every subgraph of $G$ contains a vertex of degree at most $d$; equivalently, $d(G)=\max_{H\subseteq G}\delta(H)$, and this quantity is the largest $k$ for which the $k$-core is non-empty (Kim et al., 2016, Fomin et al., 2019). Around this classical notion, a substantial literature studies spectral degeneracy, weak degeneracy, strict and fractional variants, and applications to extremal graph theory, coloring, random graph models, and exponential random graph models (Dvorak et al., 2010, Bernshteyn et al., 2021).

1. Classical degeneracy, cores, and hereditary sparsity

For a simple graph $G=(V,E)$, the $k$-core is the unique maximal subgraph whose minimum degree is at least $k$, equivalently obtained by iteratively deleting vertices of degree less than $d$0 until all remaining vertices have degree at least $d$1. The degeneracy is

$d$2

and $d$3 is $d$4-degenerate precisely when every subgraph has a vertex of degree at most $d$5 (Kim et al., 2016). This is equivalent to the existence of a degeneracy ordering $d$6 such that each $d$7 has at most $d$8 neighbors among $d$9, or, after reversing the order, at most $G$0 earlier neighbors (Dvorak et al., 2010, Fomin et al., 2019).

The parameter is closely tied to peeling and orientation. A $G$1-degenerate graph admits an orientation with maximum indegree $G$2, and a uniform indegree orientation with maximum indegree $G$3 implies $G$4-degeneracy (Dvorak et al., 2010). Degeneracy also bounds density: if $G$5 is $G$6-degenerate, then for every subgraph $G$7 one has $G$8, so the average degree of $G$9 is at most $d$0, and $d$1 (Dvorak et al., 2010). The coloring number satisfies $d$2 (Li et al., 6 Mar 2026).

These formulations make degeneracy a canonical hereditary sparsity measure. It is hereditary because it is defined by quantifying over all subgraphs; it is algorithmically tractable because the peeling process computes the full core decomposition in linear time in practice, namely $d$3 for $d$4 (Kim et al., 2016, Fomin et al., 2019). This explains why degeneracy orderings underlie sparse graph routines for coloring, ordering, and subgraph listing.

2. Spectral and extremal formulations

A central spectral variant is spectral degeneracy. Writing $d$5 for the spectral radius and $d$6 for the maximum degree, a graph $d$7 is spectrally $d$8-degenerate if every subgraph $d$9 satisfies

$d(G)=\max_{H\subseteq G}\delta(H)$0

This notion is hereditary by definition and is motivated by a common pattern in classical upper bounds: if a tree has maximum degree $d(G)=\max_{H\subseteq G}\delta(H)$1, then $d(G)=\max_{H\subseteq G}\delta(H)$2; if $d(G)=\max_{H\subseteq G}\delta(H)$3 is planar with maximum degree $d(G)=\max_{H\subseteq G}\delta(H)$4, then $d(G)=\max_{H\subseteq G}\delta(H)$5; and if $d(G)=\max_{H\subseteq G}\delta(H)$6 is $d(G)=\max_{H\subseteq G}\delta(H)$7-degenerate with maximum degree $d(G)=\max_{H\subseteq G}\delta(H)$8, then $d(G)=\max_{H\subseteq G}\delta(H)$9 (Dvorak et al., 2010).

The converse is only partial. If $k$0 is spectrally $k$1-degenerate and $k$2, then $k$3 contains a vertex of degree at most

$k$4

more precisely at most $k$5 (Dvorak et al., 2010). The same paper shows that the dependence on $k$6 cannot be eliminated if the dependence on $k$7 is subexponential, and that deciding whether a graph is spectrally $k$8-degenerate is co-NP-complete (Dvorak et al., 2010). Spectral degeneracy therefore enforces low-degree structure, but more weakly than classical degeneracy.

Degeneracy also supports sharp spectral extremal results. For $k$9-degenerate graphs of order $k$0, the split graph

$k$1

simultaneously maximizes the adjacency spectral radius $k$2 and the signless Laplacian spectral radius $k$3. Explicitly,

$k$4

and

$k$5

with equality only for $k$6 (Nikiforov, 2013). In the same direction, a general upper bound for $k$7 in terms of $k$8 yields exact equality criteria involving regular graphs or components of order $k$9 whose degrees are $G=(V,E)$0 or $G=(V,E)$1 (Nikiforov, 2013).

A non-spectral extremal manifestation appears in layout theory. If $G=(V,E)$2 denotes cutwidth, then every graph satisfies

$G=(V,E)$3

and if $G=(V,E)$4 is triangle-free, then

$G=(V,E)$5

More generally, $G=(V,E)$6-free graphs satisfy

$G=(V,E)$7

(0907.5138). Degeneracy thus controls not only coloring and density but also spectral growth and linear layout complexity.

3. Weak, strict, local, and fractional variants

Weak degeneracy replaces pure deletion by a deletion-with-saving operation. Given a budget function $G=(V,E)$8, Delete removes a vertex and decreases the budgets of all neighbors by $G=(V,E)$9, whereas DelSave removes a vertex $k$0 and spares one adjacent vertex $k$1 from the decrement, provided $k$2 and the updated budget remains nonnegative. A graph is weakly $k$3-degenerate if all vertices can be removed by a legal sequence of these operations, and the weak degeneracy $k$4 is the least constant $k$5 such that $k$6 is weakly $k$7-degenerate (Bernshteyn et al., 2021, Han et al., 2023). Every $k$8-degenerate graph is weakly $k$9-degenerate, but the converse fails in general (Bernshteyn et al., 2021).

Weak degeneracy is tailored to coloring. If $k$0 is weakly $k$1-degenerate, then

$k$2

and Bernshteyn–Lee proved that planar graphs are weakly $k$3-degenerate (Bernshteyn et al., 2021). A corrected proof of this planar bound was later given, together with a strengthened boundary-induction theorem for plane graphs (Bernshteyn et al., 2024). The same framework sharpens in special families: planar graphs of girth at least $k$4 are weakly $k$5-degenerate, and locally planar graphs are weakly $k$6-degenerate (Han et al., 2023). A local refinement uses the vertex-wise girth parameter $k$7, the length of a shortest cycle containing $k$8, and proves that every planar graph is weakly $k$9-degenerate whenever

$d$00

which in turn implies the corresponding local correspondence-coloring theorem (Davies et al., 30 Apr 2025).

Weak degeneracy also admits structural analogues of Brooks-type phenomena. A connected graph of maximum degree $d$01 is weakly $d$02-degenerate unless it is $d$03, and more generally, if $d$04, then either $d$05 contains a $d$06-clique or

$d$07

(Bernshteyn et al., 2021). For regular graphs, the lower bound is exact in a different direction: every $d$08-regular graph satisfies

$d$09

and this is tight for every $d$10 (Yang, 2023).

A further extension replaces unweighted deletions by arc-weighted acyclic orientations. In the type-$d$11 strict variant, $d$12 is $d$13-$d$14-degenerate if it has an arc-weighted orientation $d$15 with weighted outdegree $d$16 for every vertex and such that every nonempty sub-digraph contains a dominating arc $d$17 with

$d$18

This implies that $d$19 is $d$20-paintable and $d$21-AT (Zhou et al., 2023).

Fractional analogues transfer these ideas to fractional DP-coloring. The parameters $d$22 and $d$23 are defined via the ShadeSave process and satisfy

$d$24

For cycles,

$d$25

and for unicyclic graphs one has $d$26 (Dominik et al., 14 Apr 2026). These results show that degeneracy-type control persists even in fractional DP-coloring.

4. Algorithmic consequences and restricted planar families

Classical degeneracy is computationally simple: the smallest-last or peeling procedure computes $d$27 and the core decomposition in linear time (Fomin et al., 2019). That tractability makes degeneracy a natural parameter for above-guarantee algorithms. If $d$28, then every graph contains a cycle of length at least $d$29 by the Erdős–Gallai argument, but deciding whether a graph contains a cycle of length at least $d$30 is NP-complete, even for connected graphs. The complexity changes on 2-connected graphs: deciding whether such a graph contains a cycle of length at least $d$31 is solvable in time

$d$32

and for connected graphs the analogous longest-path problem is solvable in time $d$33 (Fomin et al., 2019).

Weak and strict degeneracy have been especially productive on sparse surface families. For planar graphs without any configuration from Fig. 2 of Wang–Wang–Yang, toroidal graphs without any configuration from Fig. 5, and planar graphs without intersecting $d$34-cycles, the graphs are weakly $d$35-degenerate; moreover, they admit strictly $d$36-degenerate transversals, have DP-paint number at most four, and have list vertex arboricity at most two (Wang et al., 2021). A different planar theorem shows that planar graphs without $d$37- and $d$38-cycles satisfy weak $d$39-degeneracy under two separate additional hypotheses: either no $d$40-cycles are normally adjacent to $d$41-cycles, or no $d$42-cycles are normally adjacent to $d$43-cycles. In each case the consequence is $d$44-DP-colorability, $d$45-DP-paintability, and near-bipartiteness (Wang, 2023).

Other local restrictions produce the same threshold. Planar graphs with distance of $d$46-cycles greater than $d$47 and no cycles of lengths $d$48 are weakly $d$49-degenerate, and they admit a partition $d$50 in which $d$51 is independent and $d$52 is a forest (Wang et al., 25 Feb 2025). Planar graphs of girth at least $d$53 are weakly $d$54-degenerate, while locally planar graphs are weakly $d$55-degenerate (Han et al., 2023). These are stronger structural conclusions than merely $d$56-choosable or $d$57-choosable.

Arc-weighted strict degeneracy leads to truncated-degree choosability. A graph is $d$58-truncated-degree-choosable when it is $d$59-choosable for $d$60. Here the picture is mixed. There exists a $d$61-connected non-complete planar graph that is not $d$62-truncated-degree-choosable, answering a question of Richter in the negative, but every $d$63-connected non-complete planar graph is $d$64-$d$65-truncated-degree-degenerate and hence $d$66-truncated-degree-choosable (Zhou et al., 2023).

Taken together, these results show that degeneracy variants are not merely abstract refinements of peeling. They furnish constructive reduction schemes, parameterized algorithms, and stronger coloring consequences on graph classes where ordinary $d$67-degeneracy or $d$68-degeneracy is too coarse.

5. Statistical, probabilistic, and compositional settings

Degeneracy also appears as a sufficient statistic in network models. The edge-degeneracy ERGM is the two-parameter exponential family

$d$69

with sufficient statistics $d$70 and $d$71 (Kim et al., 2016). Its convex support is the polytope

$d$72

whose extremal edge counts at fixed degeneracy are

$d$73

Every lattice point in $d$74 is realizable, the boundary has exactly $d$75 realizable lattice points and all of them are vertices, and the normalized limit shape is the lens

$d$76

(Kim et al., 2016). This geometry explains when the model concentrates on the empty graph, the complete graph, or nontrivial extremal core structures.

In random graph theory, degeneracy undergoes a sharp trichotomy in random intersection graphs. For the model $d$77 with

$d$78

one has, with high probability,

$d$79

$d$80

and bounded expansion, hence $d$81, if $d$82 (Farrell et al., 2014). Under the condition $d$83, the same model has hyperbolicity at least $d$84 asymptotically almost surely (Farrell et al., 2014). Degeneracy therefore separates the structurally sparse regime from the somewhere-dense regime in this model.

Compositional graph operations exhibit further controlled behavior. For the $d$85-sum operations $d$86, if $d$87 is $d$88-degenerate and $d$89 is $d$90-degenerate, then the resulting degeneracies satisfy explicit formulas: $d$91

$d$92

$d$93

while for $d$94-sums, $d$95 is $d$96-degenerate when $d$97 or $d$98, and $d$99-degenerate when $G$00 (Li et al., 6 Mar 2026). Through the inequality $G$01, these formulas immediately imply upper bounds on the Alon–Tarsi number of the corresponding constructions (Li et al., 6 Mar 2026).

These lines of work enlarge degeneracy from a static graph invariant into a modeling coordinate. In ERGMs it describes core intensity; in random intersection graphs it marks a phase transition in structural sparsity; in $G$02-sum operations it predicts coloring behavior through explicit compositional rules.

6. Terminological boundaries, family-level notions, and open directions

A recurring source of confusion is that several nearby terms are not equivalent. In hereditary classes defined by a finite forbidden induced set $G$03, bounded degeneracy occurs if and only if $G$04 includes a complete graph, a complete bipartite graph, and a forest (Atminas et al., 2022). This is a theorem about the classical parameter $G$05 on every graph in the class. By contrast, in the spectral extremal paper of Wu, Kang, and Ni, a family $G$06 is called degenerate when it contains at least one bipartite graph; the associated parameters are the covering number $G$07 and the independent covering number $G$08, and the spectral extremizers are characterized by joins with a core of size $G$09 (Wu et al., 16 Jul 2025). That usage is family-theoretic and unrelated to a peeling bound on individual graphs.

A second terminological divergence appears in work on graphs realized by finite abelian groups. There, degeneration is a homomorphism-based relation: an edge or a graph degenerates to another if a graph homomorphism maps the former to the latter, and group homomorphisms induce such graph homomorphisms on annihilator graphs (Raja, 2022). This notion produces partial orders, threshold graphs, and Young-diagram lattices, but it is not the same as $G$10-degeneracy.

Current open problems are concentrated in recognition, locality, and sharp constants. For spectral degeneracy, natural questions include improving the rough converse

$G$11

understanding whether the logarithmic dependence on $G$12 can be improved under additional assumptions, and deciding whether polynomial-time approximate recognition exists for spectrally $G$13-degeneracy (Dvorak et al., 2010). For weak degeneracy, open directions include the complexity of computing $G$14 and the conjectural behavior for clique-free or triangle-free graphs of large maximum degree (Bernshteyn et al., 2021). In local weak degeneracy, one explicit conjecture asks for a universal constant $G$15 such that every planar graph with a local girth function admits a legal sequence with average availability at least $G$16, which would imply exponentially many local-girth correspondence colorings (Davies et al., 30 Apr 2025).

The modern study of degeneracy graphs is therefore best understood as a family of related but non-identical theories. Classical degeneracy supplies hereditary sparsity, cores, and greedy orderings; spectral degeneracy translates sparse structure into eigenvalue control; weak and strict degeneracy sharpen coloring and online coloring; probabilistic and statistical models use degeneracy as a phase parameter; and family-level or algebraic uses of “degenerate” and “degeneration” require separate interpretation. The common thread is structural control by recursive low-complexity subgraphs, but the mechanisms and consequences differ substantially across these settings.