Contradiction Graphs Determine VC Dimension

Abstract: We study the contradiction graphs associated with binary concept classes. For a class $H \subseteq {0,1}^X$, the order-$m$ contradiction graph $G_m(H)$ has as vertices the $H$-realizable labeled sequences of length $m$, with two vertices adjacent when the two sequences assign opposite labels to some common domain point. Our main result is that the single graph $G_m(H)$ determines the threshold predicate $\mathrm{VCdim}(H)\ge m$. Consequently, the full sequence $(G_m(H))_{m \ge 1}$ determines the exact VC dimension and, in particular, detects finite versus infinite VC dimension, answering a question posed by Alon et al. (2024).

• The paper introduces cube-trace cliques in contradiction graphs as a certificate for determining if VCdim(H) is at least m.
• It employs a novel graph-theoretic approach that links shattering phenomena with intrinsic structural properties of binary concept classes.
• The findings reveal that the full sequence of contradiction graphs uniquely encodes the finite versus infinite nature of VC dimension, impacting learning theory and private learning.

Contradiction Graphs as Determinants of VC Dimension

Introduction

The paper "Contradiction Graphs Determine VC Dimension" (2605.20434) addresses the fundamental relationship between the sequence of contradiction graphs derived from a binary concept class $H \subseteq\{0,1\}^X$ and the VC dimension of $H$. The order-$m$ contradiction graph $G_m(H)$, with vertices as $H$-realizable labeled sequences of length $m$ and edges connecting pairs that assign contradicting labels to the same domain point, is the central construct. Previous investigations of contradiction graphs focused on their relation to private learning, raising the question of whether such a sequence or even a single graph $G_m(H)$ contains sufficient structural information to reveal the VC dimension of $H$. The paper provides a complete, explicit, and robust characterization: for every $m \geq 1$, the graph $G_m(H)$ determines whether $H$0; the entire sequence $H$1 determines the exact VC dimension, distinguishing finite from infinite VC classes.

Main Contributions and Technical Results

The main theorem asserts that $H$2 contains, in a purely graph-theoretic sense, a certificate of $H$3, and that no two classes with distinct VC dimensions (one finite, one infinite) can have isomorphic contradiction graphs at every level. This resolves a question posed by Alon et al. concerning possible indistinguishability of finite and infinite VC dimension from the contradiction graph sequence [AlonMoranScheflerYehudayoff2024].

The key combinatorial device is the concept of a cube-trace clique within $H$4: a clique of size $H$5 with a bijective identification to the Boolean cube $H$6, such that for any vertex in the graph, the pattern of non-neighbors within the clique forms a Boolean subcube. The main technical result is that $H$7 if and only if $H$8 contains such a cube-trace clique of size $H$9.

Formally, for every $m$0 and binary class $m$1, one has:

• $m$2 iff $m$3 contains a cube-trace clique of size $m$4.
• If for some $m$5 the graphs $m$6 for all $m$7, then $m$8.

The proof critically leverages the trace structure: it is not merely the clique number or the presence of a large clique, but the intersection pattern of the clique’s neighborhoods that encodes shattered sets, thus fully characterizing shattering from the external graph perspective. Examples make clear that large cliques in $m$9 can exist without corresponding to shattering unless the trace condition holds, particularly when clique supports are not concentrated.

Preliminary Results and Examples

Several illustrative phenomena demonstrate that the sequence of contradiction graphs is the sharpest-possible structural signature.

• For example, the difference in VC dimension between the full class $G_m(H)$0 and the even-parity subclass $G_m(H)$1 is invisible to all graphs $G_m(H)$2 for $G_m(H)$3. Thus, no finite prefix of the sequence suffices to determine the VC dimension.
• The presence of large cliques devoid of cube-trace structure in classes of VC dimension one demonstrates that merely the clique number is insufficient.

These phenomena corroborate that the proposed cube-trace certificate is both necessary and sufficient, and that the negative instance for finite prefix distinguishability is an inherent limitation.

Implications

The findings close the possibility of indistinguishable contradiction graph sequences for concept classes of different (finite vs. infinite) VC dimension. The explicit cube-trace certificate provides a concrete, intrinsic method for extracting the VC threshold property directly from graph-theoretic data. This result has several implications:

• For private and online learnability theory: Since contradiction graphs are essential in characterizations of private learnability, these results illuminate how classical combinatorial dimensions manifest in privacy-sensitive structural encodings.
• For computational learning theory: The result underscores the universality of VC dimension—it emerges robustly, even under maximal structural abstraction as a sequence of contradiction graphs.
• For graph-theoretic methods in combinatorial geometry and learning: Cube-trace cliques provide a precise bridge between shattering phenomena and clique structure, suggesting that further combinatorial parameters might admit similar exact translations in graph-based encodings.

The result does not address open questions concerning other dimensions (e.g., Littlestone, clique, or fractional clique dimensions), or the extraction of VC dimension from variants of contradiction graphs (e.g., those precluding repeated domain points).

Prospects for Future Research

Future research directions may include the extension of the contradiction graph framework to multiclass and real-valued concept classes, or to situations where data is restricted to finite domains or given by probability measures. Further, the cube-trace structure may support computational algorithms for VC dimension estimation when access is restricted to graphs derived from data or learning oracles. Finally, analogues for Littlestone and other online or adversarial dimensions might follow by appropriate modifications of the contradiction graph concept and trace structure.

Conclusion

The paper rigorously establishes that the full sequence of contradiction graphs associated with a binary concept class $G_m(H)$4 encodes—and indeed, precisely determines—the VC dimension of $G_m(H)$5. The identification and characterization of cube-trace cliques as the critical structure enables this inference, directly connecting combinatorial shattering to intrinsic graph properties. The investigation clarifies the limits of this equivalence, both highlighting the necessity of the trace condition and refuting the sufficiency of finite prefixes or clique number alone. This work strengthens the conceptual bridge between structural graph theory and statistical learning theory.