Induced-Universal Graphs

• Induced-universal graphs are graphs that contain every graph in a specified family as an induced subgraph, preserving both edges and non-edges.
• Key research focuses on achieving tight size bounds and developing explicit constructions through methods like expander-based products, algebraic encodings, and random graphs.
• These constructions have practical implications in adjacency labeling schemes, extremal graph theory, and universal embedding across diverse graph families.

An induced-universal graph for a family $\mathcal F$ is a graph $U$ such that every $H\in\mathcal F$ occurs as an induced subgraph of $U$; equivalently, there is an injective embedding of $V(H)$ into $V(U)$ that preserves both edges and non-edges. The notion ranges from finite $n$-universal constructions for all graphs on $n$ vertices to proper universal objects that remain inside a restricted graph class, and to countable strongly universal graphs for infinite classes. Across these settings, induced universality is closely tied to adjacency labeling, extremal lower bounds, and explicit constructions based on expanders, permutations, minor decompositions, algebraic encodings, and random graphs (Alon et al., 2016, Atminas et al., 2013, Alstrup et al., 2014).

1. Formal notion and basic variants

For a family $\mathcal F$ of finite simple graphs, a graph $U$ is induced-universal for $U$0 if for every $U$1 there exists an injective map $U$2 such that, for all distinct $U$3,

$U$4

Equivalently, $U$5 for some $U$6. This is the formulation used both for fixed-$U$7 families such as $U$8, the set of all simple graphs on $U$9, and for structured classes such as bounded-degree graphs, split permutation graphs, or minor-free graphs (Chappelon, 9 Mar 2026).

The induced condition is strictly stronger than ordinary universality. In the non-induced version, one only requires that edges of $H\in\mathcal F$0 map to edges of $H\in\mathcal F$1; extra edges among the image vertices are allowed. Ferber, Kronenberg, and Luh make this distinction explicit when contrasting $H\in\mathcal F$2-universal graphs with induced-$H\in\mathcal F$3-universal graphs, and note that their random-graph results concern only the non-induced notion (Ferber et al., 2016).

Several terminological refinements are standard in the literature represented here. A graph $H\in\mathcal F$4 is $H\in\mathcal F$5-universal for a class $H\in\mathcal F$6 if it contains every graph in $H\in\mathcal F$7 as an induced subgraph; it is proper $H\in\mathcal F$8-universal if $H\in\mathcal F$9 as well (Atminas et al., 2013). In the countable setting, the wheel-minor-free literature distinguishes weakly universal, meaning every graph in the class appears as a not necessarily induced subgraph, from strongly universal, meaning every graph appears as an induced subgraph; in that terminology, “strongly universal” coincides with induced-universal (Krill, 2023). The random-graph literature on graphs on $U$0 uses “universal” for containing every countable graph as an induced subgraph and “weakly universal” for containing every finite graph as an induced subgraph (Brian, 2016).

2. Size parameters, labeling equivalence, and the all-graphs problem

A central quantitative parameter is the minimum order of an induced-universal host. For bounded-degree families this is often denoted $U$1, while for the family of all $U$2-vertex graphs one writes

$U$3

The labeling viewpoint is equivalent: a family $U$4 has a $U$5-bit adjacency labeling scheme with unique labels if and only if $U$6 (Abrahamsen et al., 2016).

For all undirected graphs on $U$7 vertices, Moon proved a lower bound $U$8 on the size of an induced-universal graph, and later Alon sharpened the asymptotic upper bound to

$U$9

Alstrup, Kaplan, Thorup, and Zwick then constructed adjacency labels of length $V(H)$0, yielding induced-universal graphs with $V(H)$1 vertices, which is optimal up to a multiplicative constant and solves an open problem of Vizing from 1968 (Alstrup et al., 2014, Trimble, 2021).

For small orders, exhaustive computation is possible. The exact values known for the family of all $V(H)$2-vertex graphs are

$V(H)$3

and for $V(H)$4 one has

$V(H)$5

For $V(H)$6, the numbers of non-isomorphic minimal induced-universal graphs were also computed: $V(H)$7, respectively (Trimble, 2021).

The same labeling equivalence produces sharp asymptotics for other ambient categories. For directed graphs, the optimal label length is $V(H)$8, hence the optimal induced-universal size is $V(H)$9. For tournaments, the optimal label length is $V(U)$0, giving $V(U)$1. For bipartite graphs, the optimal label length is $V(U)$2, giving $V(U)$3 (Alstrup et al., 2014).

3. Bounded-degree, degree-2, forest, and arboricity regimes

The bounded-degree problem asks for the minimum size of an induced-universal graph for

$V(U)$4

For every fixed integer $V(U)$5, there exists a constant $V(U)$6 such that for every $V(U)$7 there is a graph $V(U)$8 with $V(U)$9 that is induced-universal for $n$0. Since Butler had shown $n$1, this is optimal up to a constant factor; for odd $n$2 it closes the earlier gap left by $n$3 bounds (Alon et al., 2016).

For the family of graphs of maximum degree $n$4, much more precise statements are available. An induced-universal graph with $n$5 vertices exists for all $n$6-vertex graphs of maximum degree $n$7, resolving a conjecture of Esperet in a stronger form. For acyclic graphs of maximum degree $n$8, the exact value is

$n$9

For larger degree bounds $n$0, the same work gives

$n$1

and, when $n$2,

$n$3

These estimates are paired with the first adjacency labeling schemes that for any $n$4 are at most $n$5 bits from optimal (Abrahamsen et al., 2016).

For forests and constant-arboricity classes, the picture is polynomial rather than exponential. A $n$6-bit adjacency labeling scheme for forests implies $n$7, so there exists a graph with $n$8 nodes containing every $n$9-node forest as a node-induced subgraph. By a reduction due to Chung, every $\mathcal F$0-node graph of constant arboricity $\mathcal F$1 then has an induced-universal graph of size $\mathcal F$2, matching the $\mathcal F$3 lower bound (Alstrup et al., 2015).

Family	Best bound stated in the cited work	Source
$\mathcal F$4, fixed $\mathcal F$5	$\mathcal F$6	(Alon et al., 2016)
Maximum degree $\mathcal F$7	$\mathcal F$8 vertices	(Abrahamsen et al., 2016)
Forests on $\mathcal F$9 nodes	$U$0 vertices	(Alstrup et al., 2015)
Constant arboricity $U$1	$U$2, matching $U$3	(Alstrup et al., 2015)

These results show that induced universality for sparse families is governed by structural decompositions: thin subgraphs and expander-based products in the bounded-degree case, and forest decompositions in the arboricity case. A plausible implication is that the dominant complexity parameter is not sparsity alone, but how efficiently the family decomposes into highly constrained components.

4. Explicit structured families and algebraic constructions

Induced universality also appears inside highly structured graph families. One explicit algebraic example is given by Steinhaus graphs. Writing $U$4, Delahan’s theorem, reproved by Chappelon, states that for every positive integer $U$5, the linear map

$U$6

with

$U$7

is an isomorphism of vector spaces over $U$8. In particular, every simple graph on $U$9 vertices is isomorphic to the induced subgraph $U$00 of some Steinhaus graph on

$U$01

vertices. Equivalently, Steinhaus graphs of order $U$02 form an $U$03-induced-universal family (Chappelon, 9 Mar 2026).

A combinatorial example is the class of split permutation graphs. There exists a split permutation graph with $U$04 vertices containing all split permutation graphs with $U$05 vertices as induced subgraphs, and this construction is order-optimal. The proof proceeds through a proper $U$06-universal 321-avoiding permutation $U$07 of length $U$08, a bijection between labelled 321-avoiding permutations and labelled symmetric split permutation graphs, and a completion step that embeds every split permutation graph on $U$09 vertices into a symmetric split permutation graph on $U$10 vertices (Atminas et al., 2013).

A third line of work is algebraic but broader in scope. The class of zero-divisor graphs of finite commutative rings with unity is universal in the induced sense: every finite graph is isomorphic to an induced subgraph of a zero-divisor graph. This remains true for Boolean rings, products of fields, and local rings. By contrast, if a local ring has principal maximal ideal, then its zero-divisor graph is a threshold graph, and every threshold graph is embeddable in the zero-divisor graph of such a ring; in this restricted setting the class is threshold-universal rather than universal for all finite graphs. The same paper also constructs a countable local ring whose zero-divisor graph embeds the Rado graph and hence every finite or countable graph as an induced subgraph (Arunkumar et al., 2022).

These examples illustrate two distinct mechanisms. One is exact algebraic parametrization, where the host family has enough linear or ring-theoretic degrees of freedom to realize every adjacency pattern. The other is structural translation, where a well-understood encoding object such as a permutation class is turned into an induced-universal graph family.

5. Countable universality, random graphs, and minor-closed settings

In the countable setting, the canonical induced-universal graph is the Rado graph, the countable random graph on $U$11. It satisfies the extension property: for every finite set $U$12 of vertices and every $U$13, there is a vertex adjacent to all of $U$14 and to none of $U$15. As a consequence, it contains every finite and every countable graph as an induced subgraph (Brian, 2016).

The same paper studies which subsets of a random graph on $U$16 remain induced-universal. For almost every graph on $U$17, every set of positive upper density is universal, and every set $U$18 with

$U$19

is weakly universal. The second statement is sharp: almost surely there are non-universal sets with divergent reciprocal sums. More generally, many partition-regular largeness notions do not force universality or even weak universality (Brian, 2016).

For minor-closed classes, the strongest positive result in the supplied material concerns forbidden wheels. If $U$20 is any wheel graph and $U$21 is the class of all countable graphs not containing $U$22 as a minor, then there exists a countable graph in $U$23 that contains every graph in $U$24 as an induced subgraph. The construction uses a tree-decomposition of adhesion at most $U$25, universal bricks $U$26 for graphs excluding both $U$27 and a path minor $U$28, and a recursive gluing scheme producing a graph $U$29 that is itself $U$30-minor-free and strongly universal for the class (Krill, 2023).

The random-graph threshold problem is much better understood for ordinary universality than for induced universality. For $U$31, the family of all $U$32-vertex graphs of maximum degree at most $U$33, there exists a constant $U$34 such that if

$U$35

then $U$36 is typically $U$37-universal. More generally, for

$U$38

if

$U$39

then $U$40 is typically $U$41-universal (Ferber et al., 2016). This serves as a benchmark rather than an induced-universal result. A plausible implication is that induced universality in random models should require additional control beyond the subgraph thresholds that suffice for ordinary universality.

6. Lower bounds, conflicting families, and extremal density

Recent lower-bound work has made the finite-size theory substantially sharper for sparse graph classes. For planar graphs, the minimum size $U$42 of an induced-universal graph for all $U$43-vertex planar graphs satisfies

$U$44

and any attempt to beat this lower bound must use a conflicting family of at least $U$45 planar graphs: any family of less than $U$46 planar graphs of $U$47 vertices has an induced-universal graph with less than $U$48 vertices. Analogous linear lower bounds are established for several other classes (Gavoille et al., 15 Aug 2025).

Family $U$49	Lower bound on $U$50	Minimum conflicting-family size stated
Forests	$U$51	$U$52
Outerplanar graphs	$U$53	$U$54
Series-parallel graphs	$U$55	$U$56
$U$57-minor-free graphs	$U$58	$U$59
Planar graphs	$U$60	$U$61
$U$62-minor-free graphs	$U$63	$U$64 or $U$65

The same paper gives a converse-type upper bound: any family of $U$66 graphs of $U$67 vertices having small chromatic number and sublinear pathwidth, like any proper minor-closed family, has an induced-universal graph with less than

$U$68

vertices. The proof builds a bridge between equitable colorings, combinatorial designs, and path-decompositions (Gavoille et al., 15 Aug 2025).

A complementary extremal parameter is inducibility. For a graph $U$69 on $U$70 vertices, $U$71 is the maximum possible probability that choosing $U$72 vertices uniformly at random from a large graph $U$73 induces a copy of $U$74. If

$U$75

then

$U$76

with equality for the star graph $U$77, the graph with a single edge on $U$78 vertices, and their complements. For all other graphs $U$79, one has

$U$80

for an absolute constant $U$81. Moreover, $U$82 is bounded away from zero exactly for graphs $U$83 for which there is a bounded-size set $U$84 with the property that all permutations of $U$85 extend to an automorphism of $U$86 (Ueltzen, 2024). This suggests that nonvanishing induced density is reserved for highly symmetric patterns, whereas the induced-universal problem itself only asks for existence of one copy of each member of the family.

Taken together, these lower bounds and extremal density results show that induced universality is constrained in two distinct ways. The first is global: universal hosts for rich families must be linearly or exponentially larger than the target graphs, depending on the family. The second is local: even inside a universal host, most specific induced patterns cannot occur with large density unless they possess exceptional symmetry.