Rotation invariants can blur highly symmetric molecules: an extended HIP-HOP basis gives the lowest error among HIP-HOP variants on methane and still runs faster at inference
Synopsis
Rotation-invariant representations can map geometrically distinct atomic environments onto the same descriptor in highly symmetric settings, limiting the accuracy of machine-learned interatomic potentials. A lightweight benchmark built around configurations with multiple-fold symmetry exposes degeneracies shared by invariant and equivariant architectures, and guides an extension of the invariant HIP-HOP basis to higher angular order. On a randomly displaced methane dataset of more than seven million configurations, the extended basis gives the lowest error among HIP-HOP variants and, at about 36K parameters, nearly matches EquiformerV3 while running faster at inference; both still fail on the (3/4) ring configuration.
Interpretation
Rotation-invariant representations lose the information needed to separate distinct geometries in highly symmetric local environments, and the failure appears in both invariant and equivariant models. Invariant models previously traded directional detail for efficiency by using scalar quantities such as distances, angles and inner products, and their adequacy depends on whether those invariants distinguish all configurations; equivariant models keep directional information at higher computational cost. On a synthetic rings benchmark of a central node with two neighbor rings and multiple-fold symmetry, the original HIP-HOP basis cannot distinguish four-fold symmetric configurations, and EquiformerV3 also fails on the (3/4) ring configuration.
Projecting the local density on a sphere onto a polynomial basis of higher angular order and taking scalar contractions of moment-tensor products yields a more complete invariant description. Moment-invariant theory treats the local atomic environment as a density on a sphere, where the angular cutoff plays the role of the truncation order in a Fourier or Taylor series and low orders keep only coarse angular features. On the randomly displaced methane dataset the extended basis reaches the lowest values on all four energy and force error metrics among HIP-HOP variants; at about 36K parameters it nearly matches EquiformerV3 while running faster at inference.
Perspective
The demonstrated scope is the synthetic rings benchmark and the randomly displaced methane dataset, which contains more than seven million configurations with an 80,000-configuration test set; comparisons cover HIP-HOP basis variants and EquiformerV3 at roughly 36K and 1.1M parameters.
Both the extended basis and EquiformerV3 still fail on the (3/4) ring configuration; the extra contraction added to remove that degeneracy is theoretically necessary but gave no measurable benefit in later methane experiments. At about 1.1M parameters EquiformerV3 has lower error, and HIP-HOP is only faster at inference.
