Efficient benchmarking of logical magic state

High-fidelity logical magic states are a critical resource for fault-tolerant quantum computation, enabling non-Clifford logical operations through state injection. However, benchmarking these states presents significant challenges: one must estimate the infidelity $ε$ with multiplicative precision, while many quantum error-correcting codes only permit Clifford operations to be implemented fault-tolerantly. Consequently, conventional state tomography requires $\sim1/ε^2$ samples, making benchmarking impractical for high-fidelity states. In this work, we show that any benchmarking scheme measuring one copy of the magic state per round necessarily requires $Ω(1/ε^2)$ samples for single-qubit magic states. We then propose two approaches to overcome this limitation: (i) Bell measurements on two copies of the twirled state and (ii) single-copy schemes leveraging twirled multi-qubit magic states. Both benchmarking schemes utilize measurements with stabilizer states orthogonal to the ideal magic state and we show that $O(1/ε)$ sample complexity is achieved, which we prove to be optimal. Finally, we demonstrate the robustness of our protocols through numerical simulations under realistic noise models, confirming that their advantage persists even at moderate error rates currently achievable in state-of-the-art experiments.

💡 Research Summary

The paper addresses the problem of efficiently benchmarking logical magic states, which are indispensable non‑stabilizer resources for fault‑tolerant quantum computation (FTQC). A logical magic state |ψ⟩ is characterized by its fidelity F=⟨ψ|ρ|ψ⟩ to a prepared (possibly noisy) state ρ, and the infidelity ε=1‑F must be estimated with multiplicative precision, i.e., the estimator ˆε should satisfy |ˆε‑ε|≤r ε for a constant 0<r<1. Because most quantum error‑correcting codes only allow fault‑tolerant Clifford gates and Pauli measurements, the benchmarking protocol is restricted to these operations.

The authors first formalize a very general “single‑copy benchmarking scheme”: in each of N rounds a single copy of ρ is prepared, a Clifford unitary Ui is applied, and all qubits are measured in the computational basis. This framework includes standard state tomography and direct fidelity estimation. They prove a fundamental limitation (Theorem 1): if the target magic state |ψ⟩ has no orthogonal stabilizer state (which is true for every single‑qubit non‑stabilizer state such as |T⟩ or |H⟩), then any such scheme requires Ω(1/ε²) copies of ρ to achieve multiplicative precision. The intuition is that every stabilizer measurement yields outcomes drawn from a Bernoulli distribution with a constant bias c∈(0,1); the variance therefore does not shrink with ε, leading to the 1/ε² scaling.

To break this barrier, the paper proposes two distinct strategies that exploit joint measurements on multiple copies or on multi‑qubit magic states.

1. Bell‑measurement protocol (two‑copy scheme).
   The authors observe that while no single‑qubit stabilizer is orthogonal to |T⟩, the antisymmetric Bell state |Ψ⁻⟩=(|01⟩‑|10⟩)/√2 is orthogonal to |T⟩⊗|T⟩. By preparing two copies of ρ, twirling them (random Clifford from the stabilizer group of |ψ⟩), applying CNOT₁,₂·(H⊗I), and measuring both qubits in the computational basis, the outcome “11” occurs with probability ε(1‑ε)=Θ(ε). Hence the measurement statistics are Bernoulli(Θ(ε)), and the estimator’s variance scales as ε/N. Consequently, O(1/ε) samples suffice, which matches a lower bound and is therefore optimal. Theorem 2 generalizes this result to any n‑qubit magic state whose twirling group Gψ decomposes the Hilbert space into irreducible representations H₀=span{|ψ⟩} and H_j (j≥1) that are not equivalent to H₀. In this case the Bell‑type circuit implements a SWAP test, and the probability of obtaining an odd parity outcome is exactly ε+O(ε²). The sample complexity becomes N=O((log 1/δ)/(r² ε)) for confidence 1‑δ.

2. Single‑copy protocol using multi‑qubit magic states.
   Preparing two high‑fidelity copies simultaneously can be experimentally demanding, especially when the distillation success probability is low. The authors therefore consider multi‑qubit magic states such as |CZ⟩ and |CCZ⟩. After twirling, the noisy state takes the form ρ′=(1‑ε)|ψ⟩⟨ψ|⊕εσ, where σ lives in the orthogonal subspace ⊕{j≥1}H_j. Each H_j contains at least one stabilizer state |s_j⟩. Measuring the overlap ⟨s_j|ρ′|s_j⟩ directly yields λ_j, the weight of ρ′ on H_j. Since ε=∑{j≥1}λ_j·dim(H_j), estimating the λ_j’s with standard Bernoulli sampling gives an O(1/ε) overall sample cost. Theorem 3 formalizes the conditions under which this works and proves optimality.

The paper backs the theoretical claims with extensive numerical simulations that incorporate realistic error channels: depolarizing gate noise, state‑preparation errors, and measurement mis‑assignments. For logical infidelities around 10⁻³–10⁻² (fidelities 0.99–0.995), the Bell‑measurement scheme reduces the required number of magic‑state copies by two to three orders of magnitude compared with conventional tomography. Even when only a single copy is available, the multi‑qubit single‑copy protocol still achieves a factor‑≈100 improvement. The simulations demonstrate robustness: the advantage persists up to moderate physical error rates (≈1 %) that are already achievable in neutral‑atom arrays and superconducting qubit platforms.

In conclusion, the work establishes a fundamental Ω(1/ε²) lower bound for any single‑copy, Clifford‑only benchmarking of single‑qubit magic states, and then provides two optimal O(1/ε) protocols that circumvent this bound. The Bell‑measurement approach is broadly applicable to any magic state whose twirling group satisfies the representation‑theoretic condition, while the multi‑qubit single‑copy method offers a practical alternative when simultaneous preparation of two copies is infeasible. These results give experimentalists concrete, resource‑efficient tools for real‑time quality assessment of logical magic states, a critical step toward scalable fault‑tolerant quantum computers.