Sample compression size versus VC dimension

Determine whether every concept class C ⊆ {0,1}^n admits a sample compression scheme whose size is polynomial in its VC dimension, and in particular whether a scheme of size O(vc(C)) exists for all such classes.

Background

Sample compression schemes aim to summarize a realizable labeled sample by a small subsample and auxiliary bits so that a reconstructor produces a hypothesis consistent with the full sample. The long-standing Sample Compression Conjecture of Floyd and Warmuth posits a scheme of size O(vc(C)) for every concept class, but only exponential-size schemes are known in general.

This paper shows barriers to one prominent approach (embedding into extremal classes) but does not resolve whether linear or polynomial-size compression schemes exist universally. The authors explicitly note the question remains open and reference Floyd (1995) and Warmuth (2003).

References

Yet, it remains an open question whether every class has a compression scheme of size linear or polynomial in its VC dimension, as explicitly asked in .

Dual VC Dimension Obstructs Sample Compression by Embeddings  (2405.17120 - Chase et al., 2024) in Section 1.2 (Sample Compression Schemes)