Source-linked AI summary
On Weisfeiler-Leman Invariance: Subgraph Counts and Related Graph Properties
V. Arvind, Frank Fuhlbrück, Johannes Köbler, Oleg Verbitsky
TL;DR
The paper asks how Weisfeiler-Leman equivalence relates to subgraph-count and occurrence invariance, questions relevant to graph isomorphism and graph similarity. It investigates these patterns for dimensions 1 and 2, completely characterizes the 1-WL case, and extends known 2-WL results while leaving broader 2-WL characterizations open.
Problem
The paper asks which subgraph counts and occurrence properties are invariant under k-WL equivalence, a question motivated by the use of k-WL and subgraph frequencies in graph isomorphism and graph similarity.
Source: §1 · pp. 2–3, Our results · p. 3Method
The paper studies count-invariant and occurrence-invariant subgraph patterns for dimensions 1 and 2, denoted C(k) and R(k).
Source: Our results · p. 3Results
The paper completely characterizes C(1) and R(1) and extends 2-WL results, including that sK2 has 2-WL-invariant counts exactly for s ≤ 5.
Source: Our results · p. 4Takeaways & Limitations
For 1-WL, the invariant pattern classes are fully described; for 2-WL, the results cover standard graph sequences, while a general characterization remains open.
Source: Our results · p. 4, §3 · p. 13, §4 · p. 20Whether homomorphism-hereditary treewidth at most 2 is necessary for membership in C(2) remains open, as does whether 2-WL preserves the matching number.
Source: §3 · p. 13, §4 · p. 20
Abstract
from arXiv · showhide
The $k$-dimensional Weisfeiler-Leman algorithm ($k$-WL) is a fruitful approach to the Graph Isomorphism problem. 2-WL corresponds to the original algorithm suggested by Weisfeiler and Leman over 50 years ago. 1-WL is the classical color refinement routine. Indistinguishability by $k$-WL is an equivalence relation on graphs that is of fundamental importance for isomorphism testing, descriptive complexity theory, and graph similarity testing which is also of some relevance in artificial intelligence. Focusing on dimensions $k=1,2$, we investigate subgraph patterns whose counts are $k$-WL invariant, and whose occurrence is $k$-WL invariant. We achieve a complete description of all such patterns for dimension $k=1$ and considerably extend the previous results known for $k=2$.
1 Introduction
The paper studies how k-WL equivalence relates to subgraph counts and occurrence, questions relevant to isomorphism testing, graph similarity, and logic. It completely characterizes the patterns invariant under 1-WL and extends known results for 2-WL.
1 Introduction
- Color refinement initializes vertices by degree and refines colors using neighborhood colors; 1-WL is this routine, while 2-WL is the original pair-coloring algorithm.The paper situates these algorithms within graph isomorphism testing and notes their role in practical algorithms.Source: §1 · pp. 1–2
- The paper asks how k-WL invariance relates to subgraph-pattern counts and occurrence, distinguishing count-invariant patterns C(k) from occurrence-invariant patterns R(k), with C(k) ⊆ R(k).The motivation includes subgraph frequencies used in graph-similarity testing and the significance of k-WL for isomorphism testing and logic.Source: §1 · pp. 2–3, Our results · p. 3
Our results
- For 1-WL, C(1) consists, up to isolated vertices, of stars and 2K2, while R(1) adds exactly P3 + P2, P3 + 2P2, and 2P3.Every count-invariant pattern has a count determined by the degree sequence.Source: Our results · p. 4
- For 2-WL, the matching graphs sK2 have invariant subgraph counts exactly when s ≤ 5, extending the matching-polynomial result from strongly regular graphs to all 2-WL-equivalent graphs.The extension is tight and cannot be extended to larger s.Source: Our results · p. 4
- C7 is the largest cycle in C(2), and C(2) contains P1 through P7 but no other path graphs.More generally, 2-WL-equivalent G and H have equal counts of Cs for 3 ≤ s ≤ 2g(G) + 1; neither the factor 2 nor the additive 1 can be improved.Source: Our results · p. 4
- For 2-WL occurrence invariance, R(2) excludes every graph with a unique 4-clique, and contains only finitely many cycle graphs.For every k, R(k) contains only finitely many cycle graphs; the broader conjecture that R(2) excludes all graphs with clique number above 3 remains a conjecture.Source: Our results · pp. 4–5
2 Color refinement invariance
The section completely characterizes the subgraph patterns whose counts or occurrence are invariant under color refinement, and establishes the structural restrictions behind the occurrence classification. Its results also show that the three additional occurrence-invariant forests have only narrowly specified exceptions in the amenability argument.
2 Color refinement invariance
- Up to isolated vertices, R(1) consists of the stars, 2K2, and exactly P3 + P2, P3 + 2P2, and 2P3.For the P3 + 2P2-free case, the amenability argument has exceptions 2C3 and C6; both are handled because their six vertices cannot contain P3 + 2P2.Source: §2 · p. 6, §2.3 · p. 7
2.1 Membership in C(1)
- Up to isolated vertices, C(1) consists exactly of the stars K1,s for s ≥1 and the 2-matching 2K2.The section relates star counts to the degree sequence and derives 2K2 counts by separating pairs of edges into disjoint or adjacent pairs.Source: §2 · p. 6, §2.1 · pp. 6–7
2.2 Non-membership in C(1)
- For each of P3 + P2, 2P3, and P3 + 2P2, the section uses 1-WL-equivalent witness graphs with differing occurrence to establish non-membership in C(1).The table caption specifies that each witness pair has the same number of vertices and the same degree.Source: §2.2 · p. 7, Table 1 · p. 7
2.3 Membership in R(1)
- Every (P3 + P2)-free or 2P3-free graph is amenable, while every P3 + 2P2-free graph is amenable except 2C3 and C6.Amenable means color refinement distinguishes the graph from every nonisomorphic graph; the two exceptions still satisfy the needed occurrence-invariance implication because they have six vertices.Source: §2.3 · p. 7
2.4 Non-membership in R(1)
- R(1) contains only forests of stars, and the section excludes every star forest outside its classified list using basic counterexamples and two reduction rules.The rules remove a component or cut one leaf from each component; repeated reductions reach a basic star forest already shown not to belong to R(1).Source: §2.4 · pp. 8–10
3 Weisfeiler-Leman invariance
The section investigates when subgraph-pattern counts and occurrence properties are Weisfeiler-Leman invariant, giving bounds and exact cases across several graph families. It also uses a cloning construction to preserve 2-WL indistinguishability and produce equivalent graphs that differ in whether they contain a specified graph.
Subgraph counts
- For every k, C(k) contains patterns with htw(F)≤k; C(1) consists exactly of stars and 2K2, up to isolated vertices.Here htw is the maximum treewidth among homomorphic images of the pattern graph.Source: §3 · p. 12
- htw(F)>2 if and only if some partition of V(F) has quotient graph isomorphic to K4.Source: §3 · p. 12
- C(2) contains exactly matchings through 5K2, cycles C3–C7, and paths P1–P7 within those respective graph families.The theorem states there are no other matching, cycle, or path graphs in C(2).Source: §3 · p. 13
Cycle counts
- G and H have equal girth and equal subgraph counts, sub(Cs, G) = sub(Cs, H), for each 3 ≤ s ≤ 2g(G) + 1.Source: §3 · p. 15
Occurrence properties
- For each k, R(k) contains only finitely many cycle graphs; this follows in part because no R(k) contains all path graphs.A first-order reduction from graph isomorphism to Long Path establishes the path non-invariance used in the argument.Source: §3 · pp. 16–17
- No graph containing a unique 4-clique belongs to R(2).Source: §3 · p. 17
3 Weisfeiler-Leman invariance
- Cloning vertices into same-color copies preserves 2-WL indistinguishability when Duplicator transfers a winning 3-pebble bijection strategy.The projection from clones to original vertices lets Duplicator simulate the game on the original graphs.Source: §3 · pp. 18–19
- The constructed graphs G′ and H′ are 2-WL-equivalent, yet G′ contains F while H′ contains no copy of F.The construction assumes F contains exactly one K4; it attaches F′ to the cloned vertices according to their colors.Source: §3 · pp. 19–20
4 Concluding discussion
The paper leaves the full description of C(k) open in general, while giving a complete description for k=1; R(2) remains unresolved. It also points to connections with subgraph-count algorithms and calls for study of induced subgraph invariance.
4 Concluding discussion
- The complete-description question for C(k) remains open generally, although Theorem 2.3 answers it affirmatively for k=1 and completely describes R(1).For R(2), whether matching graphs sK2 occur for only finitely many s or for all s remains unknown.Source: §4 · p. 20
- The parameter htw(F) governs both k-WL invariance of sub(F,G) and a cited algorithm's running-time exponent, suggesting a connection between invariance and algorithmics.The paper identifies this connection as an area for further exploration.Source: §4 · p. 20
- The invariance of induced subgraphs and their counts for different k remains a topic for further study.The passage notes prior study of induced-subgraph counts in finite model theory.Source: §4 · p. 20
A Proof of Lemma 2.4
The proof of Lemma 2.4 is divided into three lemmas and relies on explicit descriptions of graphs avoiding three specified forbidden patterns.
A Proof of Lemma 2.4
- The proof splits Lemma 2.4 into Lemmas A.5, A.9, and A.8, using explicit characterizations of F-free graphs for three choices of F.The patterns are P3 + P2, P3 + 2P2, and 2P3.Source: A Proof of Lemma 2.4 · p. 24
A.1 Forbidden forests
This appendix characterizes several forbidden-pattern graph classes, including a line-graph criterion for 3P2-free graphs of sufficient size and structural families for related forbidden patterns.
A.1 Forbidden forests
- For v(H)≥6, H is 3P2-free exactly when its line graph has a clique cover of size at most 2.The appendix also characterizes 3P2-free graphs up to isolated vertices using specified structural families.Source: §A.1 · p. 24
- Up to isolated vertices, 3P2-free graphs are subgraphs of 2K3, K1∗(K3+sK1), or K2∗sK1 for the stated ranges of s.K2∗sK1 is a complete split graph with a clique part of size 2.Source: §A.1 · pp. 24–25
- The six-vertex threshold in the line-graph criterion is necessary: the complement of P3 + 2K1 is 3-matching-free but has line-graph clique cover number 3.Thus the criterion does not extend to all smaller graphs.Source: §A.1 · p. 25
- For 2P3-free graphs, the listed cases include a six-vertex net plus isolated edges and subgraphs of K1∗sK2.The latter family is also known as friendship graphs.Source: §A.1 · p. 27
- For a connected (P3 + P2)-free graph with at least five vertices and no isolated vertices, the proof forces a star K1,s.The proof rules out maximum degree 2 and shows all external vertices must be adjacent to the 3-star's center.Source: §A.1 · p. 27
- For connected (P3 + 2P2)-free graphs with at least seven vertices, the characterization places H as a subgraph of K1∗(K3+sK1) or K2∗sK1.The proof derives this case from the 3P2-free characterization.Source: §A.1 · p. 28
- In the connected 2P3-free proof, a tree's diameter is at most 4, and the diameter-4 case is a subgraph of a windmill graph K1∗sK2.A longer diameter would contain a P6 and hence a 2P3.Source: §A.1 · pp. 28–29
- A graph has vertex cover number at most 2 exactly when it is a subgraph of K2∗sK1 for some s≥1.Source: §A.1 · p. 29
A.2 Amenability
Amenability means that 1-WL distinguishes a graph from every nonisomorphic graph. The section proves amenability for several forbidden-subgraph classes, with explicit exceptions for two of them.
- Every (P3 + P2)-free graph is amenable, and H + K is amenable for a forest K exactly when H is amenable.Source: §A.2 · p. 29
- In the proof for the vertex-cover case, stable color cells and their bipartite connections let a cell-preserving bijection yield an isomorphism.When the two exclusive-neighbor sets have equal size, the proof separately matches the two-star structure between cells.Source: §A.2 · pp. 30–31
- Every 3P2-free graph is amenable except 2C3.The proof reduces the remaining cases to subgraphs of K1 ∗ (K3 + sK1).Source: §A.2 · p. 31
- For 3P2-free graphs, the proof handles large pendant-vertex sets by checking that cell-respecting bijections preserve the graph.If one of the three non-pendant vertices becomes pendant, it joins the pendant cell and the same argument applies.Source: §A.2 · p. 31
- Every (P3 + 2P2)-free graph is amenable except 2C3 and C6.Source: §A.2 · p. 31
- Every 2P3-free graph is amenable.For windmill subgraphs, the proof uses stable color cells and shows that the within-cell and between-cell structures are determined up to isomorphism.Source: §A.2 · pp. 31–32