Lie-algebraic incompleteness of symmetry-adapted VQE for non-Abelian molecular point groups

1. Problem Statement

The Variational Quantum Eigensolver (VQE) is a leading algorithm for electronic structure simulation on NISQ devices. To reduce the exponential parameter count of the Unitary Coupled-Cluster Singles and Doubles (UCCSD) ansatz, researchers use Symmetry-adapted UCCSD (SymUCCSD), which filters excitation operators based on molecular point group symmetries.

• The Issue: While SymUCCSD works efficiently for Abelian point groups (e.g., $C_{2v}$), recent benchmarks revealed a catastrophic failure for non-Abelian groups (e.g., $C_{3v}$ in Ammonia, $T_d$ in Methane).
• The Mystery: Previous empirical fixes (like HiUCCSD) restored accuracy but lacked a theoretical explanation. It was unclear why the symmetry-adapted filter caused the optimizer to converge to a high-energy plateau despite the ansatz appearing to span the necessary symmetry space.

2. Methodology

The authors employ a rigorous Lie-algebraic and geometric framework to diagnose the failure. Their approach involves:

• Algebraic Formulation: Defining the action of the molecular point group $G$ on Fock space and the creation/annihilation operators.
• Pool Filtering Analysis: Comparing the $G$-equivariant pool (operators invariant under the full group $G$) against the SymUCCSD pool (operators invariant only under a maximal Abelian subgroup $H \leq G$).
• Dynamical Lie Algebra (DLA) Analysis: Investigating the Lie algebra generated by the retained excitation operators to determine the topology of the reachable state manifold.
• Numerical Validation: Simulating the Ammonia ($NH_3$) molecule ($C_{3v}$ symmetry) using the STO-3G basis set (16 qubits) to test theoretical predictions against Full Configuration Interaction (FCI) benchmarks.

3. Key Contributions and Theoretical Findings

A. Linear Incompleteness (Spurious Irrep Splitting)

The paper proves that restricting SymUCCSD to an Abelian subgroup $H$ causes a spurious splitting of multidimensional irreducible representations (irreps).

• Mechanism: A $d_\lambda$-dimensional irrep of the non-Abelian group $G$ decomposes into $d_\lambda$ distinct 1D irreps under the Abelian subgroup $H$.
• Consequence: The SymUCCSD filter retains only "diagonal" excitations (where the component index $\mu$ matches $\nu$) and discards "off-diagonal" cross-component excitations ($\mu \neq \nu$).
• Result: For a pair of shells in a $d_\lambda$-dimensional irrep, the method discards $d_\lambda(d_\lambda - 1)$ single excitations and a significant fraction of double excitations that are mathematically required for $G$-equivariance.

B. Lie-Algebraic Incompleteness (Torus Confinement)

The most critical finding is that the remaining operators do not just reduce the parameter count; they fundamentally alter the topology of the reachable state space.

• Commutativity: The retained diagonal generators commute with each other.
• DLA Collapse: The Dynamical Lie Algebra (DLA) collapses from the full unitary algebra $\mathfrak{u}(d_\lambda)$ to a purely Abelian subalgebra isomorphic to $\mathfrak{u}(1)^{d_\lambda}$.
• Geometric Confinement: Consequently, the reachable state manifold is confined to a torus $T^{d_\lambda}$ (a measure-zero submanifold) within the full unitary group $U(d_\lambda)$. The ansatz physically cannot access the off-diagonal orbital rotations required to describe the full electron correlation, regardless of optimization parameters.

C. The "Abelian Basis Trap" (Gradient Plateau)

The authors identify a secondary, equally fatal numerical obstruction independent of the DLA structure:

• Vanishing Integrals: In standard quantum chemistry, molecular orbitals (MOs) are often diagonalized within Abelian subgroups. In such an Abelian-adapted basis, cross-component Hamiltonian integrals (e.g., coupling different components of a degenerate irrep) vanish identically due to orthogonality.
• Zero Gradients: Even if one were to manually add the missing off-diagonal generators to the ansatz, the energy gradient along these directions would be zero at the Hartree-Fock initialization ($\theta=0$).
• Optimizer Trap: The optimizer becomes "blind" to the non-Abelian directions, trapping the system in an artificial energy plateau.

4. Numerical Results

The authors validated their theory on Ammonia ($NH_3$) with $C_{3v}$ symmetry:

• Setup: 16 qubits, STO-3G basis, BFGS optimizer.
• SymUCCSD Performance: The ansatz (75 parameters) converged fully (gradient norm $< 10^{-4}$) but stalled at an energy 21.8 mHa above the FCI energy.
• Comparison:
  • Standard UCCSD (135 parameters, no symmetry filter) achieved $\Delta E \approx 0.1$ mHa.
  • The SymUCCSD failure was confirmed to be structural (topological confinement) rather than an optimization artifact.
• Verification: The results confirmed the DLA confinement to a $T^2$ torus and the existence of the gradient plateau.

5. Significance and Implications

• Theoretical Diagnosis: The paper provides the first rigorous explanation for the failure of symmetry-adapted VQE in non-Abelian systems, moving beyond empirical observations to a geometric and algebraic proof.
• Distinction between Linear and Lie-Algebraic Completeness: It highlights that having a linearly complete set of operators is insufficient; the set must also be closed under commutators to generate the full Lie group orbit.
• Algorithmic Requirements: To successfully apply symmetry-adapted VQE to non-Abelian systems, two conditions must be met simultaneously:
  1. Lie-Algebraic Completeness: The operator pool must include all off-diagonal generators derived from the full point group $G$, not just the Abelian subgroup.
  2. Gradient Recovery: The molecular orbital basis must be rotated or parameterized to ensure non-zero cross-component integrals, preventing the optimizer from getting stuck in a zero-gradient plateau.
• Future Directions: The findings suggest that adaptive algorithms (like ADAPT-VQE) relying on Abelian filters will fail for non-Abelian systems unless they incorporate integral-based selection rules (like HiUCCSD) or explicitly break the Abelian symmetry in the orbital basis.

In summary, the paper demonstrates that Abelian restrictions in non-Abelian systems create a "double trap": they structurally remove the necessary operators to explore the full state space (Lie-algebraic incompleteness) and, through standard basis choices, eliminate the gradients needed to discover those missing operators (numerical plateau).