O(N2) Universal Antisymmetry in Fermionic Neural Networks

Tianyu Pang, Shuicheng Yan, Min Lin

Introduction

Simulating many-electron systems is a fundamental challenge in quantum chemistry, primarily due to the Pauli Exclusion Principle. This principle requires the electronic wavefunction to be antisymmetric—flipping sign upon the exchange of any two electrons. Traditional state-of-the-art methods like the Fermionic Neural Network (FermiNet) rely on Slater determinants to enforce this property. However, calculating determinants scales at $O(N^3)$ with respect to the number of electrons $N$, creating a significant computational bottleneck for large-scale molecular simulations.

The Pairwise Antisymmetry Approach

To address this scaling issue, researchers have proposed a "determinant-free" architecture that reduces complexity to $O(N^2)$. The core innovation is a pairwise construction where the total wavefunction is represented as a product of terms for every pair of electrons $(i, j)$:

$$\psi_{pair}(x) = \prod_{1 \le i < j \le N} \mathcal{A} \circ F(x_i, x_j; \{x \setminus \{i,j\}\})$$

Here, $\mathcal{A}$ is a simple antisymmetrizer, and $F$ is a function that incorporates information about the entire electronic environment. By leveraging permutation-equivariant backbones—similar to those used in FermiNet—the model ensures that the pairwise terms are aware of the global configuration without sacrificing efficiency.

Theoretical Universality and Continuity

A critical contribution of this work is the formal proof that this $O(N^2)$ construction is a universal approximator for antisymmetric functions. Unlike previous attempts at $O(N^2)$ scaling which required discontinuous functions that are difficult for neural networks to learn, this method can represent ground-state wavefunctions using continuous, learnable functions. The authors demonstrate that by including multi-electron dependence within the pairwise terms, the "cusp" conditions and topological constraints of the wavefunction can be satisfied smoothly.

Experimental Results

The proposed architecture was evaluated on a series of atomic systems, ranging from Lithium ($N=3$) to Nitrogen ($N=7$). The researchers tested two variants: $\psi'_{pair}$ (a basic version) and $\psi''_{pair}$ (a more flexible variant).

Ground-state energy comparisons for small atoms using the proposed O(N²) pairwise models versus exact benchmarks.

As shown in the table above, the $O(N^2)$ models achieve high accuracy for smaller systems. For Lithium (Li) and Beryllium (Be), the calculated energies are remarkably close to the exact benchmarks. However, as the number of electrons increases, a performance gap begins to emerge. For Nitrogen (N), the model achieves an energy of -53.1855 compared to the exact value of -54.5892. This suggests that while the $O(N^2)$ construction is theoretically universal, capturing the complex correlations in larger atoms may require increased model capacity or deeper architectures.

Conclusion

This work provides a theoretical and empirical foundation for scaling fermionic neural networks beyond the $O(N^3)$ limit. By proving that pairwise constructions can maintain both universality and continuity, it opens the door for more efficient simulations of large molecules and materials. Future research will likely focus on optimizing these architectures to close the accuracy gap in larger systems while fully realizing the speed advantages of $O(N^2)$ scaling.

Summary updated

BibTeX
@inproceedings{pang2022o,
  author = {Tianyu Pang and Shuicheng Yan and Min Lin},
  title = {{O(N2) Universal Antisymmetry in Fermionic Neural Networks}},
  booktitle = {AI4Science @ ICML},
  year = {2022},
  url = {https://openreview.net/forum?id=d23FsidO9s0},
}
Figure