Neighborhood Complexity and Radius-1 Merge-Width in Monadically Dependent Graph Classes

Abstract: Monadic dependence is a proposed structural dividing line for fixed-parameter tractability of first-order model checking on hereditary graph classes. A graph class is \emph{monadically dependent} if the class of all graphs cannot be interpreted in its vertex-colored members using a fixed first-order formula. We prove two structural consequences of monadic dependence. First, every monadically dependent class has \emph{almost linear neighborhood complexity}: for every graph $G$ in the class and every set $A\subseteq V(G)$, the family ${N_G(v)\cap A : v\in V(G)}$ has size $|A|<sup>{1+o(1)}$. Second, every $n$-vertex graph in a monadically dependent class has radius-1 merge-width $n<sup>{o(1)}$. Here, merge-width is the decomposition parameter of Dreier and Toruńczyk based on construction sequences; its radius-$r$ version measures local reachability among parts through already resolved pairs. This settles the radius-1 case of the conjectured connection between monadic dependence and almost bounded merge-width and provides the first decomposition-based structural description of monadically dependent graph classes. Our proof is algorithmic: we give an $\mathcal{O}(n<sup>5)$-time algorithm that, given an $n$-vertex graph $G$ such that $|{N_G(v)\cap A : v\in V(G)}|\le O(|A|<sup>d)$ for every $A\subseteq V(G)$, computes a construction sequence witnessing radius-1 merge-width $\mathcal{O}(n^{1-1/d}\log n)$.

• The paper proves that every monadically dependent graph class has neighborhood complexity at most |A|^{1+o(1)}, using bounded VC-dimension, sparsification, and set-system techniques.
• The paper gives an O(n^5)-time algorithm that constructs radius-1 merge-width at most k·n^{1−1/d} log n for graphs with neighborhood complexity bounded by c|A|^d, yielding n^{o(1)} width for monadically dependent classes.
• These results establish the radius-1 case of the conjectured equivalence between monadic dependence and almost bounded merge-width while enabling consequences such as compact adjacency labeling, sparse spanners, and efficient shortest-path algorithms.

This paper, "Neighborhood Complexity and Radius-1 Merge-Width in Monadically Dependent Graph Classes" (2607.10941), establishes two structural theorems about monadically dependent graph classes and provides the first decomposition-based structural description of this class hierarchy. Its two main results are: (i) every monadically dependent class has almost linear neighborhood complexity, and (ii) every $n$-vertex graph in a monadically dependent class has radius-1 merge-width $n^{o(1)}$, witnessed by a construction sequence computable in polynomial time. The second result settles the radius-1 case of the conjectured equivalence between monadic dependence and almost bounded merge-width, which is the proposed route toward the central open problem of first-order model checking on hereditary graph classes.

Context: model checking and monadic dependence

First-order model checking — deciding whether a sentence $\varphi$ holds in an input graph $G$ — captures problems such as $k$-Clique, $k$-Dominating Set, and $k$-Independent Set. The landmark result of Grohe, Kreutzer, and Siebertz established fixed-parameter tractability on nowhere dense classes, and nowhere denseness is the exact dividing line among monotone classes. Dense hereditary classes such as bounded clique-width and bounded twin-width fall outside this classification, and the search for the exact tractability boundary has converged on monadic dependence: a class is monadically dependent if the class of all graphs cannot be first-order interpreted (transduced) in its vertex-colored members. Monadic dependence coincides with nowhere denseness on monotone classes, with monadic stability on orderless classes, and with bounded twin-width on ordered classes. The hardness direction of the conjectured equivalence — FPT model checking if and only if monadic dependence on hereditary classes — was recently confirmed: model checking is AW[$*$]-hard on every hereditary class that is not monadically dependent (Przybyszewski et al., 22 May 2025).

Dreier and Toruńczyk [merge-width paper, STOC 2025] introduced merge-width, a radius-indexed family of decomposition parameters generalizing treewidth, degeneracy, twin-width, clique-width, and generalized coloring numbers, and showed FPT model checking on bounded merge-width classes given a construction sequence. Their conjecture is that monadically dependent classes are exactly those with almost bounded merge-width. One direction is known (almost bounded merge-width implies monadic dependence); this paper proves the other direction for radius $r=1$.

Result I: almost linear neighborhood complexity

For a graph $G$ and $n^{o(1)}$0, the neighborhood complexity of $n^{o(1)}$1 is $n^{o(1)}$2. The main theorem states that for every monadically dependent class $n^{o(1)}$3, every $n^{o(1)}$4, and every $n^{o(1)}$5, this quantity is at most $n^{o(1)}$6 — equivalently, for every $n^{o(1)}$7 there is a constant $n^{o(1)}$8 bounding it by $n^{o(1)}$9. This answers affirmatively the open problem posed in the monadically stable model checking paper [(Dreier et al., 2023) / FOCS 2024], and the bound is essentially tight: nowhere dense classes can already exhibit superlinear neighborhood complexity (e.g., 1-subdivisions of high-girth, high-degree graphs).

The proof structure parallels the monadically stable argument, which reduced to the nowhere dense case via Shelah's branching index; here the induction is on VC-dimension instead. Since monadically dependent classes have bounded VC-dimension $\varphi$0, and since a monadically dependent class excluding $\varphi$1 is nowhere dense (hence has almost linear neighborhood complexity), it suffices to handle bipartite graphs $\varphi$2 with no twins in $\varphi$3 and show $\varphi$4.

The core machinery combines classical results from computational learning theory on set systems of bounded VC-dimension: the Sauer-Shelah lemma, the Hamming graph of a set system (whose edge count is at most $\varphi$5 times the number of sets, by Haussler, Littlestone, and Warmuth), and a lemma from the monadically stable paper guaranteeing a large subfamily in which each set has a unique "witness" element. The proof introduces the notion of a $\varphi$6-sparsification: sets $\varphi$7, $\varphi$8, and definable functions $\varphi$9 such that $G$0 is injective. A deletion argument on the ground set (a harmonic-sum weighting argument in the spirit of Welzl's partition trees) shows that a nonterminal $G$1-sparsification of size $G$2 and dimension $G$3 can be refined to a $G$4-sparsification of dimension $G$5, losing only a $G$6 factor in size and increasing first-order complexity by a constant. Iterating at most $G$7 times terminates with a sparsification whose associated partition has size at most $G$8, via the $G$9-free case. Combining the two bounds yields $k$0.

This theorem has immediate corollaries, previously known only in more restrictive regimes (nowhere dense, monadically stable, almost bounded merge-width): every $k$1-vertex graph in a monadically dependent class admits a Welzl ordering (spanning path) with crossing number $k$2, a neighborhood cover of overlap $k$3 and strong radius 2, a spanner of stretch 4 with $k$4 edges, an $k$5-bit adjacency labeling scheme, and an $k$6-time All-Pairs Shortest Path algorithm.

Result II: radius-1 merge-width

The paper's second theorem is algorithmic and applies to all graphs, not only those from monadically dependent classes. There is an $k$7-time algorithm that, given an $k$8-vertex graph whose neighborhood complexity is at most $k$9 for every $k$0, computes a construction sequence of radius-1 merge-width at most $k$1, where $k$2. Instantiating with the almost linear neighborhood complexity of monadically dependent classes yields radius-1 merge-width $k$3; instantiating with the linear complexity of bounded merge-width classes yields $k$4.

The algorithm is a greedy construction sequence builder over $k$5 rounds, each performing resolve operations followed by one merge. It maintains a leader for each part, the leader graph (induced subgraph on leaders), and integer weights. The key combinatorial input is a corollary of Haussler's Packing Lemma: any graph with $k$6 neighborhood complexity admits a pair of $k$7-fractional twins for $k$8, meaning a pair $k$9 with $k$0 for every weight function $k$1. The algorithm finds such a pair simultaneously with respect to the current weight function and the uniform weight function (via a weighted-average trick, at factor 3 loss), merges the smaller part into the larger's leader, resolves all pairs between the merged part and parts whose leaders lie in $k$2, and applies a Welzl-style multiplicative weight update.

Three claims complete the analysis. First, the procedure produces a valid construction sequence, maintained via an invariant that unresolved pairs agree with the adjacency of their leaders. Second, each vertex has at most $k$3 distinct leaders over the run — this follows from always promoting the leader of the larger part, which keeps the leader-change tree at logarithmic height. Third, each vertex's leader lies in $k$4 in at most $k$5 rounds, by the multiplicative weight analysis: the total weight grows as $k$6, while each "hit" doubles the vertex's leader weight. Summing the contributions of the two resolution cases gives the claimed width bound $k$7.

Notably, the algorithm is oblivious to the constants $k$8 and $k$9; it always runs in $*$0 time and its output quality adapts to the actual neighborhood complexity of the input. A construction sequence of radius-1 merge-width $*$1 yields a signed tree model with $*$2 transversal pairs, connecting to recent work on shortest paths in graphs with sparse signed tree models.

Limitations and open questions

The paper is explicit about several boundaries of its contribution. The radius-1 case is only the first step: the conjecture of Dreier and Toruńczyk concerns radius-$*$3 merge-width for all $*$4, and the tractability direction of the main model checking conjecture for all monadically dependent classes remains open. The $*$5 running time is not optimized; the authors leave open whether construction sequences of width $*$6 can be computed in near-linear time for all graphs of linear neighborhood complexity, noting that for bounded twin-width graphs such signed tree models are obtainable in randomized $*$7 time. The $*$8 terms throughout hide constants depending on the class (and radius), and the constant in the width bound inherits the (potentially large) constant from Haussler's Packing Lemma. Finally, the neighborhood complexity bound, while essentially tight in exponent, is proved non-constructively in the sense that it relies on the model-theoretic definition of monadic dependence; the algorithmic result applies only to graphs whose neighborhood complexity is polynomially bounded, and recognizing whether a given graph satisfies this hypothesis is not addressed.

Conclusion

This paper delivers the first structural characterizations of monadically dependent graph classes via quantitative combinatorial measures: almost linear neighborhood complexity and almost bounded radius-1 merge-width, the latter with a polynomial-time decomposition algorithm. The proof techniques — VC-dimension induction through Hamming and merge graphs on the structural side, and fractional twins with multiplicative weight updates on the algorithmic side — transfer tools from computational learning theory and computational geometry into structural graph theory. The radius-$*$9 case for $r=1$0, and the resulting full tractability direction for first-order model checking on monadically dependent classes, remain open.