Bi-Lipschitz Ansatz for Anti-Symmetric Functions

Technical Summary: Bi-Lipschitz Ansatz for Antisymmetric Functions

Problem Statement
The paper addresses the challenge of constructing neural network ansatzes for antisymmetric functions, a requirement central to simulating quantum many-body systems (specifically fermionic wave functions) where the Pauli exclusion principle dictates total antisymmetry. Existing approaches face significant limitations:

• Backflow ansatzes: While universal, they often require an exponential number of determinants to achieve polynomial approximation accuracy or result in discontinuous models that are impractical for algorithms.
• Recent continuous models: Approaches like those in [CL26] and [YLG+24] use combinations of symmetric functions and Vandermonde determinants to achieve continuity and polynomial complexity. However, these models suffer from numerical instability (scaling issues) and, when normalized to fix stability, lose continuity.
• The Gap: There is a need for ansatzes that are simultaneously continuous, antisymmetric by construction, computationally efficient (polynomial complexity), and capable of providing quantitative universal approximation guarantees.

Methodology
The authors propose two distinct frameworks to achieve these goals:

1. Bi-Lipschitz Embedding Approach
This method leverages the relationship between antisymmetric functions and functions invariant under the alternating group $A_n$ (permutations with positive sign).

• Core Concept: Construct a bi-Lipschitz embedding $\mu$ from the quotient space $\mathbb{R}^{d \times n} / A_n$ into a Euclidean space $\mathbb{R}^m$.
• Construction:
  • For $d=1$, the authors define a symmetric bi-Lipschitz function using sorting ($\text{sort}(x)$) and an antisymmetric function $Q(x)$ based on the minimum distance between sorted elements and their sign.
  • For $d > 1$, they apply linear projections to the input and aggregate the $d=1$ embeddings.
  • A simple projection operator $\pi[f](x) = \frac{1}{2}(f(x) - f(\tau_0 x))$ (where $\tau_0$ is a negative permutation) converts an $A_n$-invariant function into an antisymmetric one.
• Approximation: An antisymmetric function is approximated as $\pi \circ f \circ \mu$, where $f$ is a standard Multi-Layer Perceptron (MLP).
• Theoretical Guarantee: The bi-Lipschitz property ensures that the approximation error of the antisymmetric function is bounded by the approximation error of the underlying invariant function, yielding quantitative rates for Lipschitz functions.

2. Antisymmetric Frame Averaging Approach
This method generalizes group averaging to avoid the factorial complexity of summing over all $n!$ permutations.

• Core Concept: Utilize "frame averaging," where averaging is performed over a subset of the group (a frame) rather than the entire group.
• Continuity Challenge: Standard frame averaging often fails to preserve continuity at points where coordinates coincide (the set $\Omega_{d \times n}$).
• Solution:
  • The authors define a Weighted Permutation Frame based on projections of the input onto specific vectors, assigning weights that vanish when permutations are ill-defined.
  • They introduce a Weakly Stabilizing (WS) Operator $\Phi$. This operator modifies the input function to ensure that as points approach the singularity set $\Omega_{d \times n}$, the function value smoothly transitions to zero (consistent with the property that antisymmetric functions must vanish when coordinates coincide).
  • The final ansatz is an averaged projection: $E[f](x) = \sum w_i(x) (-1)^{\tau_i} \Phi[f](\tau_i^{-1} x)$.
• Complexity: This approach requires averaging over $O(n^3)$ permutations (specifically $O(n)$ frame elements composed with $O(n^2)$ transpositions), significantly reducing complexity compared to full group averaging.

Key Contributions

1. Two Novel Ansatzes: The introduction of a bi-Lipschitz embedding method and a weighted frame averaging method, both of which are continuous and antisymmetric by construction.
2. Quantitative Approximation Rates: The paper derives explicit bounds on the number of parameters required to approximate Lipschitz antisymmetric functions to an accuracy $\epsilon$. For the bi-Lipschitz approach, the complexity scales as $O(\epsilon^{-nd/2} |\log \epsilon|)$, comparable to standard ReLU network approximation rates for Lipschitz functions.
3. Continuity and Stability: Unlike previous continuous models that required normalization (breaking continuity) or suffered from scaling instability, these models maintain continuity and stability without ad-hoc fixes.
4. Theoretical Proofs: The authors prove universality and provide bounds for the bi-Lipschitz distortion and the boundedness of the frame averaging operator.

Experimental Results
The authors conducted preliminary experiments approximating the determinant function $f(A) = \det(A)$ for matrices of order $n \in \{10, 15, 20\}$.

• Comparison: They compared their two methods against a standard MLP (which does not enforce antisymmetry) and the Vandermonde-based approach from [YLG+24].
• Performance:
  • Accuracy: Both proposed methods achieved lower Mean Absolute Error (MAE) and Mean Absolute Relative Error (MARE) than the competing models, particularly for larger $n$.
  • Convergence: The proposed methods demonstrated significantly faster convergence during training.
  • Parameter Efficiency: The proposed ansatzes required fewer parameters than the [YLG+24] approach while maintaining or improving accuracy.
  • MLP Baseline: The standard MLP performed poorly for larger $n$ (e.g., $n=20$), highlighting the necessity of explicitly enforcing antisymmetry.

Significance and Claims
The paper claims to resolve the trade-off between computational efficiency, continuity, and universality in antisymmetric neural network ansatzes. By establishing that antisymmetric functions can be approximated with polynomial complexity and quantitative rates while maintaining continuity, the work provides a theoretically sound and practically viable foundation for simulating quantum many-body systems. The authors emphasize that their bi-Lipschitz construction is expected to lead to improved generalization and learning stability, while the frame averaging approach offers a scalable alternative to full group symmetrization. The preliminary experimental evidence supports the theoretical advantages, showing improved performance in learning antisymmetric functions compared to existing state-of-the-art methods.