A Geometric Approach to Sample Compression

The Sample Compression Conjecture of Littlestone & Warmuth has remained unsolved for over two decades. This paper presents a systematic geometric investigation of the compression of finite maximum concept classes. Simple arrangements of hyperplanes in Hyperbolic space, and Piecewise-Linear hyperplane arrangements, are shown to represent maximum classes, generalizing the corresponding Euclidean result. A main result is that PL arrangements can be swept by a moving hyperplane to unlabeled d-compress any finite maximum class, forming a peeling scheme as conjectured by Kuzmin & Warmuth. A corollary is that some d-maximal classes cannot be embedded into any maximum class of VC dimension d+k, for any constant k. The construction of the PL sweeping involves Pachner moves on the one-inclusion graph, corresponding to moves of a hyperplane across the intersection of d other hyperplanes. This extends the well known Pachner moves for triangulations to cubical complexes.

💡 Research Summary

The paper tackles the long‑standing Sample Compression Conjecture (SCC) by giving a systematic geometric treatment of finite maximum concept classes. It begins by recalling that a maximum class of VC‑dimension d can be represented as a d‑dimensional cubical complex, i.e., the cell structure induced by an arrangement of d‑dimensional hyperplanes in Euclidean space. The authors then extend this representation in two directions. First, they show that simple arrangements of hyperplanes in hyperbolic space also generate exactly the same class of cubical complexes; despite the non‑Euclidean curvature, the intersection pattern of any d hyperplanes still yields a d‑dimensional cell that corresponds to a concept in the maximum class. Second, they move to piecewise‑linear (PL) hyperplane arrangements, where each hyperplane is a union of linear facets glued together. By constructing an explicit homeomorphism between the PL cell decomposition and the cubical complex, they prove that every finite maximum class can be realized by a PL arrangement.

The central technical contribution is a sweeping procedure for PL arrangements. A single “moving” hyperplane is translated continuously across the arrangement. Whenever it passes through the intersection of d fixed hyperplanes, the authors perform a Pachner move on the one‑inclusion graph of the class. In the language of cubical complexes, this move replaces a (d‑1)‑dimensional face by its complementary face, exactly mirroring the bistellar flips known from triangulations but adapted to cubes. Each such move eliminates one vertex of the one‑inclusion graph, i.e., one sample from the concept class, while preserving the maximum‑class property. By iterating the sweep until the moving hyperplane exits the arrangement, exactly d vertices remain; thus the class admits an unlabeled d‑compression. This construction realizes the “peeling” scheme conjectured by Kuzmin and Warmuth and provides an explicit algorithmic procedure for compressing any finite maximum class.

Beyond the positive result, the paper derives a negative corollary: there exist d‑maximal classes that cannot be embedded into any maximum class of VC‑dimension d + k for any fixed constant k. The proof exploits the rigidity of PL arrangements: the number and configuration of intersection points of d hyperplanes impose combinatorial constraints that cannot be satisfied by a higher‑dimensional maximum class unless the dimension increase is unbounded. Consequently, while every maximum class can be compressed, not every maximal class can be “lifted” into a larger maximum class with only a bounded increase in VC‑dimension.

Methodologically, the work bridges combinatorial learning theory, topological combinatorics, and geometric group theory. The authors give a detailed analysis of the one‑inclusion graph, showing that Pachner moves preserve its connectivity and the Euler characteristic of the underlying cubical complex. They also verify that the sweeping hyperplane never creates degenerate cells, ensuring that the process terminates after a finite number of moves. The extension to hyperbolic space is handled by exploiting the fact that hyperbolic hyperplanes are still geodesic codimension‑one submanifolds, so their intersection patterns are combinatorially identical to the Euclidean case.

In summary, the paper provides (1) a geometric unification of Euclidean, hyperbolic, and PL hyperplane arrangements as models for maximum concept classes; (2) a constructive sweeping algorithm that yields an unlabeled d‑compression for any finite maximum class, thereby confirming the peeling conjecture; and (3) a structural limitation showing that d‑maximal classes may resist embedding into bounded‑increase higher‑dimensional maximum classes. These results advance the understanding of SCC, demonstrate the power of topological moves (Pachner flips) in learning‑theoretic contexts, and open new avenues for applying non‑Euclidean geometry to sample compression problems.