Experimental Observation of Error-Resilience Phase Transition on Superconducting Quantum Processor

AI summary

Overview

Research area: Quantum information science — specifically the interplay of random (scrambling) unitary dynamics with coherent noise, and the experimental detection of a noise-driven phase transition on superconducting quantum hardware.

Technical level: Advanced. The paper assumes familiarity with Haar-random unitaries, Rényi entropies, stabilizer Rényi entropy (non-stabilizerness), randomized measurement protocols, and finite-size scaling.

Scope (one sentence): The paper reports the first experimental observation of the predicted "error-resilience" phase transition between an error-protecting phase and an error-vulnerable phase, using IBM's ibm_pittsburgh Heron r3 processor for system sizes up to N = 26 qubits, and extracts a critical noise angle of α_c ≈ 1.01 via a joint finite-size scaling collapse of three order parameters.

What This Paper Is About

Random unitaries scramble quantum information into highly nonlocal degrees of freedom, which can protect it from local noise — but theory predicts this protection abruptly fails once the noise strength crosses a critical threshold. Earlier theoretical work by Turkeshi and Sierant described this as a phase transition in an encoding–decoding circuit model, occurring at a critical coherent noise angle α_c ≈ 1.14 (for code rate r = 1/2). This paper's goal is to actually observe that transition on real superconducting hardware, and to determine how far a realistic, non-ideal implementation deviates from the idealized theoretical model.

Key Contributions

1. First experimental observation of the error-resilience phase transition. The authors implement the encoding–decoding model of Turkeshi and Sierant on IBM's ibm_pittsburgh Heron r3 device for system sizes up to N = 26 qubits, detecting the transition through three distinct order parameters: fidelity, Rényi-2 entanglement entropy, and stabilizer Rényi-2 entropy.

2. Randomized-measurement estimators for entanglement and non-stabilizerness, with a variance-reduction scheme. Because multi-copy observables are prohibitively complex at these qubit counts, the authors use single-copy randomized measurement protocols, and they incorporate their measured fidelity data into an importance sampling (IS) variance-reduction protocol (also comparing against a common randomized measurements, or CRM, approach) built on a simple ansatz state.

3. A joint Gaussian process regression fit of the finite-size scaling collapse. Rather than fitting each order parameter separately, the authors jointly model the scaling collapse across all three order parameters, producing posterior distributions for the critical parameters and characterizing the deviation of real-device behavior from theory.

4. A diagnosis of the theory–experiment discrepancy, plus a demonstration that error mitigation is indispensable. The fitted critical angle is α_c ≈ 1.01, well below the theoretical 1.14; feeding noiseless simulated data from the same circuit ensemble into the same model yields α_c ≈ 1.05, implicating the encoding circuits (which are not deep enough to be truly Haar-random) rather than hardware noise as the source of the gap. The same model cannot adequately identify critical behavior in unmitigated raw data.

Main Findings

• The transition is observed, but at a lower critical angle than theory predicts. Joint scaling collapse across fidelity, Rényi-2 entanglement entropy, and stabilizer Rényi-2 entropy gives α_c ≈ 1.01, "well below" the theoretical value of α_c ≈ 1.14 (i.e., ≈ 1.1437 as quoted in the finite-size analysis).

• The gap is attributed to circuit depth, not hardware. Running noiseless simulated data through the identical Gaussian process model yields α_c ≈ 1.05, implying that the discrepancy between experiment and theory stems from the encoding circuits being not deep enough to be truly Haar-random.

• The experimental circuit ensemble appears to converge toward Haar-random behavior as N grows. At α = 1.1437, the annealed and quenched fidelities of the experimental ensemble initially dip below 1/2 but approach 1/2 as N increases (simulations from N = 6 to N = 30), consistent with the Haar-random result F̃(α_c) ≥ 1/2 for all N, approaching 1/2 as N → ∞. The authors therefore expect the effective α_c to increase toward 1.1437 as larger N are included in the dataset.

• Error mitigation is essential. The same scaling-collapse model cannot adequately identify critical behavior in unmitigated, raw experimental data. Only after zero-noise extrapolation (ZNE) and M3 SPAM mitigation is the transition resolvable.

• Both mitigation techniques help, and combining them is best. At α = 0 (where every encoding–decoding circuit ideally implements the identity, giving F = 1), fidelities improve with M3 alone and with ZNE alone, and are highest with both applied together.

• Dynamical decoupling did not help. The authors considered dynamical decoupling to suppress correlated idling noise but found it did not improve circuit performance, citing two reasons: tunable couplers on IBM Heron devices significantly reduce crosstalk, and the brickwork structure means only boundary qubits have very short idle periods.

• Universality class matches theory. The finite-size scaling ansatz O = f(N^{1/ν}(α − α_c)) with ν = 1 applies to all three order parameters (fidelity, entanglement, non-stabilizerness) in the theoretical treatment.

Methodology in Plain English

The circuit. Qubits are split into a "logical" set X and an ancilla set. All qubits start in |0⟩, a random encoding unitary U is applied, then identical single-qubit coherent noise acts on every qubit, then U† decodes. The ancillas are measured and the run is kept (post-selected) only when they return to their initial state; what remains is the final logical state ρ_X. In the experiment, the noise is injected deliberately as virtual R_Z(α) gates after encoding — IBM devices implement R_Z as classical phase offsets requiring no extra microwave pulse, so the injected error is negligible.

Building approximate Haar randomness. True Haar-random circuits require exponentially many two-qubit gates, so the authors build brickwork circuits of two-qubit-gate depth exactly equal to N: each layer is random single-qubit SX and R_Z(φ) gates (φ uniformly random) followed by the device-native CZ gate on alternating qubit pairs. They generate four circuits per system size and reuse the same four encodings at every α, relying on self-averaging to capture the average behavior while remaining implementable.

Error mitigation. Gate noise is addressed with ZNE using global folding at scale factors λ ∈ {1, 3, 5}, with quadratic Richardson extrapolation to estimate the zero-noise measurement distribution; the authors wrote their own codebase so that only encoding and decoding layers are folded and the injected coherent noise is left untouched. SPAM errors are addressed with IBM's matrix-free measurement error mitigation (M3), with an M3 calibration run performed immediately before each data collection to capture current device behavior.

Measuring the order parameters. Fidelity is the fraction of post-selected outcomes equal to the all-zero bit string, using 30,000 shots per circuit. Entanglement entropy and non-stabilizerness are estimated from single-copy randomized measurements: random single-qubit unitaries are applied (sampled so that θ has density ½ sin θ on [0, π] and φ uniform on [0, 2π]), the purity estimator built from Hamming distances between bit-string pairs is averaged, and then S₂ = −log₂ of the result. The authors note that ρ_X is mixed on noisy hardware, so calling S₂(ρ_{|X₁|}) an entanglement entropy is technically incorrect (they keep the name for consistency) and they set S₂(ρ_X) = 0 when computing M₂ from experimental data. On the "smaller subsystem" they use n ≡ |X₁| = N/4.

Variance reduction. Using an approximate state σ(p) = (1−p)|0_{X₁}⟩⟨0_{X₁}| + p·I/2^{|X₁|}, the authors implement both a CRM-style estimator (with a = 1) and an importance-sampling estimator, exploiting the fact that X_σ(u) is a normalized probability density over measurement settings that is efficient to sample from directly for this ansatz. They sweep p from 0 to 1 and compare against the observed fidelity with the all-zero state.

Analysis. A Gaussian process regression jointly models the finite-size scaling collapse across all three order parameters, yielding posterior distributions for the critical parameters.

Why This Matters

This work converts a theoretical prediction about scrambling, noise, and recoverability into a measurable hardware phenomenon, and in doing so it establishes a concrete benchmark for how close current devices come to idealized random-circuit behavior. It also delivers a practical methodological lesson: the transition is invisible in raw data and only appears after ZNE and M3 are applied, which makes the case that error mitigation is not cosmetic but enabling for this class of physics experiment. The α_c ≈ 1.01 versus 1.14 versus 1.05 comparison is a rare, clean attribution of an experimental deviation to circuit-ensemble depth rather than to device noise.

Real-world applications (as directions this line of work feeds into):

• Processor benchmarking and validation — using phase-transition order parameters and scaling collapse as a diagnostic of how close a device's executed circuit ensemble is to an ideal random ensemble.
• Error-mitigation development — ZNE with custom folding plus M3 calibration timed against drift form a reusable pipeline for NISQ experiments requiring many circuit variants.
• Quantum error correction design — the encoding/decoding-plus-post-selection structure is a toy encoded-memory protocol, so understanding where scrambling-based protection breaks down informs code and decoder design.
• Efficient characterization of entanglement and magic — the single-copy randomized-measurement estimators with importance-sampling variance reduction lower the shot overhead for measuring non-stabilizerness and purity on near-term hardware.

Industry relevance: The experiment runs on IBM's ibm_pittsburgh Heron r3 hardware and uses IBM's M3 package, so the results speak directly to the capabilities and limits of commercially available superconducting processors.

Future Directions

• Deepen the encoding circuits. The authors' own diagnosis is that depth-N brickwork is not Haar-random, so going to greater depth (or otherwise improving the ensemble's t-design properties) should push the effective α_c toward the theoretical 1.1437 — the paper explicitly predicts the fitted α_c will rise as larger N enter the dataset.
• Extend to larger N and confirm the trend. Simulations already span N = 6 to 30, but experiment stops at N = 26; checking whether the experimental α_c tracks 1.1437 as N grows is the natural test of the depth explanation.
• Handle the mixedness of ρ_X properly. The authors set S₂(ρ_X) = 0 when computing M₂ from experimental data and only partially correct for impurity, leaving a rigorous treatment of the noisy-state entanglement and magic observables as an open problem.
• Test robustness beyond the idealized noise model. The model assumes identical, site-independent coherent noise on a single Bloch axis; how the transition behaves under correlated, site-dependent device noise — and whether the axis choice matters for a non-Haar ensemble — remains to be characterized.

Target Audience

Quantum information researchers and graduate students studying scrambling, random circuits, and measurement-induced or noise-induced phase transitions; experimental superconducting-qubit teams interested in benchmarking random-circuit ensembles; and error-mitigation practitioners who want a concrete case study in which ZNE and M3 determine whether a physical effect is observable at all. Readers without a background in Rényi entropies, stabilizer Rényi entropy, and randomized measurements will find the paper demanding.

Authors’ abstract

Random unitaries scramble quantum information into highly nonlocal degrees of freedom, thereby providing protection from local noise. However, as demonstrated by Turkeshi and Sierant (Phys. Rev. Lett. 132, 140401, 2024 and arXiv:2603.00235), this protection may fail once local noise grows sufficiently strong. These works describe a phase transition in a noisy random encoding-decoding circuit model between an error-protecting phase, in which encoded information is recoverable, and an error-vulnerable phase, in which encoded information is lost, with the transition occurring at a critical coherent noise angle $α_c \approx 1.14$. Here, we present the first experimental observation of this transition, conducted on IBM's ibm_pittsburgh Heron r3 device for system sizes of up to $N = 26$ qubits. We detect this transition via three order parameters: fidelity, Rényi-2 entanglement entropy, and stabilizer Rényi-2 entropy. For the latter two, we employ randomized measurement protocols to construct estimators, and for all three, we rely on zero-noise extrapolation and M3 SPAM error mitigation. We then jointly model the finite-size scaling collapse across all three order parameters with a Gaussian process regression, yielding a critical angle $α_c \approx 1.01$, well below the theoretical value. Inputting noiseless simulated data into the same model yields $α_c \approx 1.05$, implying that the discrepancy between experiment and theory is due to our encoding circuits, which are not deep enough to be truly Haar-random. Finally, we observe that the same model cannot adequately identify critical behavior in unmitigated, raw experimental data. This indicates that error mitigation techniques are crucial for detecting the transition on hardware.