Self-supervised learning on 3D point clouds often relies on masked autoencoding, where parts of the input are hidden and the model learns to reconstruct them. But when the backbone is an SE(3)-equivariant network—one that respects rotations and translations—a new theoretical result identifies a fundamental incompatibility. The authors prove that standard masking causes a topological collapse, effectively forcing the model to hallucinate the missing data rather than reason about it.

The proof shows that the equivariance constraint and the masking objective cannot be simultaneously satisfied in a meaningful way. As a result, the learned representations are distorted, and the model's reconstructions become unreliable. This is not a practical tuning issue but a structural limitation of combining these two popular techniques.

To address this, the paper introduces SIEVE, a method designed to overcome the topological obstruction. While the abstract does not detail the mechanism, the title and framing indicate that SIEVE modifies the masking or equivariant structure to preserve both equivariance and reconstruction quality. The work provides a theoretical foundation for why certain self-supervised approaches fail on point clouds and points toward a fix.