Error resilience and unitary 2-designs

The paper in detail

The protocol starts from a product state, applies \(U\), a coherent error \(\hat K=\exp[-i(\alpha/2)\sum_jZ_j]\), and \(U^{\dagger}\), then postselects the ancillas on their initial state. The fidelity \(F\) is the probability that the logical state is recovered given successful postselection. For Haar encoders at code rate \(1/2\), \(F\to1\) for \(\alpha<\alpha_c=2\arccos(2^{-1/4})\approx1.14\) and \(F\to0\) above.

The threshold comes from a competition inside the postselection probability: the chance that the error leaves the encoded state alone, which decays like \(\cos^{2N}(\alpha/2)\), against the chance it is rotated into another logical state, about \(2^{-N(1-r)}\). The question is which encoders keep this competition, and so the protected phase.

The annealed fidelity \(\tilde F\), the ratio of averages, depends on the ensemble only through its second-moment twirl, so any exact 2-design reproduces the Haar value. The inequality quantifies the converse: a deviation from Haar behaviour costs design excess. For matchgates the twirl is spanned by \(2N+1\) Majorana shell operators, giving \(\tilde F\) in closed form, which tends to 0 for every fixed \(\alpha>0\). Each non-Gaussian doping gate acts on the shells through a transfer matrix, and as doping grows the shells mix towards the Haar answer.

Thermalisation of the encoded state is not enough: the Ising chain's half-chain entropy reaches the Page value but the protected phase does not appear, because local Hamiltonian ensembles are not 2-designs. An experimentally motivated all-to-all disordered spin model recovers resilience as it approaches a 2-design. Open questions include incoherent errors and a proof of the doping scale.

Transferable methods

1. Second moments decide the annealed fidelity

   To connect error resilience with randomness, both must be expressed through the same object. The annealed fidelity is a ratio of two linear functionals of the second-moment twirl, and the frame potential is the squared distance of that twirl from Haar, so Cauchy–Schwarz links them.

   With \(\Phi^{(2)}_{\mathcal E}(O)=\mathbb E_U[(U^{\dagger})^{\otimes2}OU^{\otimes2}]\), the annealed fidelity is \[\tilde F=\frac{\mathrm{Tr}[\Phi^{(2)}(\hat K\otimes\hat K^{\dagger})B_{\mathrm{num}}]}{\mathrm{Tr}[\Phi^{(2)}(\hat K\otimes\hat K^{\dagger})B_{\mathrm{den}}]}.\] For any ensemble with mean postselection probability \(\bar p>0\), \[\Delta_{\mathcal E}\ \ge\ \frac{\bar p^{\,2}\,|\tilde F-\tilde F_{\mathrm{Haar}}|^{2}}{d^{2}\bigl(1+D\tilde F_{\mathrm{Haar}}^{2}-2\tilde F_{\mathrm{Haar}}\bigr)},\] with \(d=2^{N}\), \(D=2^{k}\), \(c=|\mathrm{Tr}\hat K|^{2}\) and \[\tilde F_{\mathrm{Haar}}=\frac{(c+d)(d-1)}{(c+d)(d-1)+(D-1)(d^{2}-c)}.\]

   Editorial checks for \(N=3\), \(k=1\), \(\alpha=0.9\): Monte Carlo over Haar unitaries gives \(\tilde F=0.9081\) against the formula's \(0.9080\). For encoders that are products of single-qubit Haar unitaries (frame potential \(2^{N}=8\), so \(\Delta=6\); sampled 7.8), \(\tilde F=0.874\) and the right-hand side is \(1.3\times10^{-5}\), so the inequality holds with a wide margin, as a sufficient condition should.

   How to reuse it

   Any figure of merit that is a ratio of linear functionals of the \(t\)-th moment of an ensemble can be bounded against its Haar value by the \(t\)-th frame potential excess; this turns operational tasks into design certificates.

2. Majorana shells for matchgate twirls

   Matchgate circuits are far from 2-designs, but their second-moment twirl is small: it is spanned by one operator per Majorana degree. That makes the annealed fidelity exactly computable and shows how the protected phase fails without doping.

   The two-copy commutant of the orthogonal Gaussian group is spanned by \(\Upsilon_\ell\propto\sum_{|S|=\ell}\gamma_S\otimes\gamma_S\), \(\ell=0,\dots,2N\). Since \(\hat K=\prod_j[\cos\tfrac\alpha2-\sin\tfrac\alpha2\,\gamma_{2j-1}\gamma_{2j}]\), only even shells contribute, with weights \[s_{2m}\propto\binom{N}{m}\Bigl(\cos^{2}\tfrac{\alpha}{2}\Bigr)^{N-m}\Bigl(\sin^{2}\tfrac{\alpha}{2}\Bigr)^{m},\] and \(\tilde F_{\mathrm M}=\sum_ms_{2m}N_{2m}/\sum_ms_{2m}D_{2m}\) in terms of explicit binomial overlaps \(N_{2m},D_{2m}\).

   Doping inserts a transfer matrix between shells, \(\tilde F_{\mathrm D}=\sum N_{2r}(\Gamma^{n_d})_{rm}s_{2m}/\sum D_{2r}(\Gamma^{n_d})_{rm}s_{2m}\), whose subleading eigenvalue sets the approach to Haar. Editorial remark: this is why the fidelity deviation and the frame-potential excess decay with the same exponent, \(e^{-4n_d/N}\), as the End Matter explains.

   How to reuse it

   For circuit families with a small commutant (matchgates, number-conserving or symmetric circuits), express second-moment quantities in the commutant basis; doping or symmetry breaking then becomes a transfer matrix on that basis, and its spectral gap gives the convergence rate to Haar behaviour.

Source: Eqs. (1)–(3); Eqs. (4)–(7); End Matter. Based on the linked paper version. Derivations marked as editorial, and suggested uses in other problems, are editorial guidance rather than results of the paper.

Context & limitations

Why it matters. The result separates two notions often run together: thermalising a state is not enough for error resilience, because the postselection probability probes the randomness of the unitary itself at second order. Recovery fidelity thus becomes an operational diagnostic of how close a many-body circuit or Hamiltonian evolution is to a unitary 2-design, measurable on hardware.

Limitations. The inequality and most analysis concern the annealed fidelity, which averages numerator and denominator separately; the physically measured quenched fidelity is compared numerically and agrees in the cases shown. Errors are coherent \(Z\) rotations only. The \(N\log N\) doping scale is inferred from a data collapse, not proved for the fidelity. Hamiltonian results are numerical, at small sizes. The authors used Claude Opus 5.5 to improve the writing.

Description based on the full arXiv HTML of v1, read 2026-11-11: the main text (protocol, inequality (3), matchgate formulas (4)–(7), Hamiltonian encoders) and the AI statement; the End Matter proof and Supplementary Material were read for structure only. The Haar threshold, the Haar annealed-fidelity formula and inequality (3) for a non-design ensemble were checked numerically.