Fermionic Gaussianity can be tested with mode-independent sample-complexity

Technical Summary: Fermionic Gaussianity Testing with Mode-Independent Sample Complexity

Problem Statement
The paper addresses the fundamental problem of quantum property testing: deciding whether an unknown pure quantum state $|\psi\rangle$ belongs to a specific family of states or is at least $\epsilon$-far (in trace distance) from that family. The authors focus specifically on two families:

1. Fermionic Gaussian States: States obtained by evolving a computational basis state under Hamiltonians quadratic in fermionic creation and annihilation operators. These states are central to understanding quantum advantage and are used in tomography and benchmarking.
2. Slater Determinant States: States formed by the exterior action of a unitary on a single occupation-basis vector, representing non-interacting fermions.

Prior to this work, the best-known sample complexities for testing fermionic Gaussianity were dependent on the number of modes $n$, ranging from $O(n^5/\epsilon^4)$ down to the recent $O(n/\epsilon^2)$. The Bell sampling test, which measures Majorana operators in the Bell basis, was known to be effective but its sample complexity was thought to scale with $n$. The central question is whether the sample complexity can be made independent of the system size $n$, achieving the information-theoretic lower bound of $\Theta(\epsilon^{-2})$.

Methodology
The authors develop a general framework for upper-bounding the sample complexity of property testing for arbitrary families of pure states. The approach relies on a two-copy projective POVM $\{P, Q\}$ where $P$ accepts states in the target family $\mathcal{M}$ with certainty (perfect completeness). The core challenge is to lower-bound the rejection probability $r(\psi) = \text{tr}[Q(|\psi\rangle\langle\psi|)^{\otimes 2}]$ for states far from $\mathcal{M}$.

The methodology proceeds in three logical steps:

1. Geometric Decomposition and Witness Vectors:
   For a state $|\psi\rangle$ far from the compact manifold $\mathcal{M}$, the authors decompose it relative to its closest state $|g\rangle \in \mathcal{M}$:
   $$|\psi\rangle = \sqrt{F} |g\rangle + \sqrt{1-F} |\chi\rangle$$
   where $F$ is the fidelity and $|\chi\rangle$ is orthogonal to the tangent space $T_g\mathcal{M}$. They derive conditions (Lemma 1) under which a "witness vector" $|w\rangle$ exists in the range of the rejection projector $Q$ such that the rejection probability is bounded by $r(\psi) \geq c F(1-F)$. This bound, however, vanishes for states with very low fidelity ($F \approx 0$).

2. The Path Property:
   To handle low-fidelity states, the authors introduce the Path Property (Definition 1). A family has this property if any state $|\psi\rangle$ can be continuously deformed to a state in $\mathcal{M}$ via a path $\gamma(t)$ such that the acceptance probability never drops below its initial value $f(\psi)$.
   By combining the Path Property with the geometric bound, they prove (Theorem 1) that if a state is $\epsilon$-far, its rejection probability is bounded away from zero by a constant independent of $n$, specifically $r(\psi) \geq c(1-F)/4$. This implies a sample complexity of $O(\epsilon^{-2})$.

3. Recursive Construction of Paths:
   To establish the Path Property for specific cases, the authors employ a recursive strategy. They decompose the Hilbert space based on the occupation of specific modes (or orbitals). By showing that the acceptance probability is a quadratic function of the interpolation parameter between occupation sectors, they demonstrate that one can always deform a state toward a target state (e.g., a vacuum or a specific Slater determinant) without decreasing the acceptance probability.

Key Results

• Fermionic Gaussianity Testing (Theorem 2):
  The authors apply their framework to the Bell sampling test for fermionic Gaussian states.

  • They prove that for any state $|\psi\rangle$, the rejection probability satisfies $r_n(\psi) \geq \epsilon^2/8$.
  • This establishes that the Bell sampling test distinguishes Gaussian states from $\epsilon$-far states using $\Theta(\epsilon^{-2} \log(1/\delta))$ copies.
  • Crucially, this sample complexity is mode-independent (independent of $n$), matching the optimal lower bound. This holds even against protocols using arbitrary collective and adaptive measurements.

• Slater Determinant Testing (Theorem 3):
  The framework is extended to Slater determinants using a specific two-copy POVM based on the operator $T_{n,\eta} = \sum a_i \otimes a_i^\dagger$.

  • They establish a similar rejection bound with a constant $c = 2/3$.
  • The sample complexity is also shown to be $\Theta(\epsilon^{-2} \log(1/\delta))$, independent of the number of modes $n$ and particle number $\eta$ (in non-trivial regimes).

Significance and Claims
The paper claims that the Bell sampling procedure, previously analyzed with $n$-dependent bounds, is in fact optimal in terms of sample complexity. The primary contribution is the demonstration that the "mode-independence" of the sample complexity arises from the specific geometric structure of the Gaussian and Slater manifolds, which allows for the construction of the required deformation paths.

The authors emphasize that their analysis applies to the standard Bell sampling test without requiring new experimental protocols; rather, it provides a rigorous theoretical justification for the test's optimality. They note that while the constant factors in the sample complexity are settled, the exact optimal constants (e.g., whether Bell tests achieve the absolute minimum constant) remain an open question.

The work also highlights that the technique is generalizable, successfully applying to Slater determinants, suggesting potential for broader application to other families of quantum states. The authors acknowledge that concurrent work [33, 34] has established similar results for bosons and fermions using gradient flow techniques, noting that while the proof techniques differ, the conclusions regarding constant sample complexity are consistent.

Note: The technical details leading to the proofs of the results in this paper were largely developed by GPT-6, with minimal human input, as stated by the authors Maxwell West and Martín Larocca.