Abstract
Developing equivariant neural networks for the E(3) group plays an importantrole in modeling 3D data across real-world applications. Enforcing thisequivariance primarily involves the tensor products of irreduciblerepresentations (irreps). However, the computational complexity of suchoperations increases significantly as higher-order tensors are used. In thiswork, we propose a systematic approach to substantially accelerate thecomputation of the tensor products of irreps. We mathematically connect thecommonly used Clebsch-Gordan coefficients to the Gaunt coefficients, which areintegrals of products of three spherical harmonics. Through Gaunt coefficients,the tensor product of irreps becomes equivalent to the multiplication betweenspherical functions represented by spherical harmonics. This perspectivefurther allows us to change the basis for the equivariant operations fromspherical harmonics to a 2D Fourier basis. Consequently, the multiplicationbetween spherical functions represented by a 2D Fourier basis can beefficiently computed via the convolution theorem and Fast Fourier Transforms.This transformation reduces the complexity of full tensor products of irrepsfrom $\mathcal{O}(L^6)$ to $\mathcal{O}(L^3)$, where $L$ is the max degree ofirreps. Leveraging this approach, we introduce the Gaunt Tensor Product, whichserves as a new method to construct efficient equivariant operations acrossdifferent model architectures. Our experiments on the Open Catalyst Project and3BPA datasets demonstrate both the increased efficiency and improvedperformance of our approach.