The Gain-Engineered Transmon

AI summary

Overview

Research area: Superconducting quantum computing — specifically dissipation engineering and noise-biased qubits for hardware-efficient quantum error correction (QEC).

Technical level: Intermediate. The core idea is intuitive (use a pump tone to undo a specific decay error), but full appreciation of the calibration, decoherence modelling and randomized benchmarking requires familiarity with circuit QED, master equations and transmon level structure.

Scope: The paper introduces, calibrates and benchmarks a noise-biased qubit — the gain-engineered transmon — encoded in the |g⟩ and |f⟩ states of a standard transmon–cavity circuit, using a single added microwave pump tone to autonomously reverse single-photon loss.

What This Paper Is About

Quantum error correction normally needs many physical qubits per logical qubit, which is a severe hardware burden. If one error channel can be made to dominate all others ("noise bias"), tailored codes can achieve the same logical error rate with far fewer physical components. The problem is that previous noise-biased qubits relied on specialised circuit elements whose complexity undermines scalability. This paper asks whether noise bias can instead be engineered in one of the most common superconducting circuits — a standard transmon with a readout resonator — by adding only a single microwave tone.

Key Contributions

1. A new qubit concept with minimal hardware cost. The gain-engineered transmon is realized in a standard transmon–readout resonator circuit by adding one microwave pump tone, with the qubit encoded in the transmon |g⟩ and |f⟩ states rather than |g⟩ and |e⟩.
2. A frequency-selective, autonomous loss-correction channel. A single-photon parametric pump at ω_gain = (ω_a − α + ω_b − 2χ)/2 activates four-wave mixing that converts two pump photons into one transmon photon on the |e⟩–|f⟩ transition and one cavity photon; cavity loss at κ_b makes the exchange unidirectional, producing an engineered gain dissipator κ_gain 𝒟[|f⟩⟨e|] with rate κ_gain = 8g_gain²/κ_b.
3. Experimental demonstration of order-of-magnitude relaxation enhancement. At g_gain/2π = 133 kHz (κ_gain/2π = 165 kHz), T₁^f and T₁^g are extended by ×16.7 and ×12.6 respectively (average ×14.7) relative to the |g⟩–|e⟩ encoding, at the cost of only a 2.1× reduction in echo coherence time.
4. Compatibility with fast, high-fidelity control. A two-photon drive at ω_a − α/2 implements |g⟩–|f⟩ single-qubit gates during gain operation, with randomized-benchmarking fidelities of 99.58 ± 0.01% at the operating point and above 99.55% across the full explored range of g_gain.

Main Findings

• Device parameters and baseline coherence: The transmon has ω_a/2π = 4.872 GHz, anharmonicity α/2π = 172 MHz and residual thermal population n_th,a = (2.5 ± 0.08) × 10⁻². The |g⟩–|e⟩ transition has T₁^e = 133.6 (± 2.2) µs, Ramsey T₂^ge = 107.2 (± 2.8) µs and echo T_2E^ge = 138 (± 6) µs; the |e⟩–|f⟩ lifetime is T₁^ef = 59.1 (± 3) µs, approximately following κ_a^ef ≈ 2κ_a^ge. The cavity has ω_b/2π = 7.442 GHz and κ_b/2π = 858 kHz, serving as both readout resonator and lossy buffer mode.

• Gain calibration and model validation: Sweeping pump frequency and amplitude maps the resonance condition ω_p = ω_gain, which shifts with pump amplitude due to the Stark shift. Time-domain traces at g_gain/2π = 31 kHz, 261 kHz and 1.5 MHz agree with master-equation simulations that take the predicted g_gain and independently measured parameters as input, with only readout contrast left free; at 1.5 MHz the interaction is strong enough to drive coherent |e⟩–|f⟩ population oscillations.

• Relaxation enhancement followed by saturation: T₁^g,f rise with g_gain and initially follow the analytical prediction T₁^f ≈ κ_gain/(κ_a^ge κ_a^ef), then peak and decrease. At g_gain/2π = 133 kHz the measured values are T₁^f = 2.22 (± 0.06) ms and T₁^g = 1.67 (± 0.34) ms, compared with a simulated 3 ms limited mainly by thermal excitation from |g⟩ to |e⟩ (which the gain converts into bit flips). The gap between experiment and simulation is attributed to processes outside the model, such as pump-induced heating or multi-photon transitions.

• Coherence cost is smaller than the relaxation gain: T₂^gf falls sharply with g_gain while T_2E^gf decreases more slowly, reaching 65 (± 4) µs at the operating point (÷2.1 relative to T_2E^ge). Simulations show no g_gain dependence of coherence, so the degradation is attributed to low-frequency noise (for example drift in gain-pump amplitude) sampled during the interleaved four-day measurement campaign; averaging over individual four-minute runs gives a median T₂^gf of 22.9 µs at this g_gain.

• Gate performance is maintained under gain: The |g⟩–|f⟩ Rabi oscillations scale as cos(π(A_gf/A_π)²), consistent with a two-photon process. Gates use 6σ Gaussian pulses with σ = 20 ns, and the conceptual speed limit set by 2/α is 1.85 ns. Randomized benchmarking across the range of g_gain gives fidelities that decrease only slightly, with ℱ = 99.58 ± 0.01% at g_gain/2π = 133 kHz; the authors note these are lower bounds because the gates are amplitude-Rabi calibrated and not otherwise optimized.

• Direct comparison of encodings (Table 1): Against |g⟩–|e⟩ values of T₁^i = 133.6 (± 2.2) µs, T₂ = 107.2 (± 2.8) µs and T_2E = 138 (± 6) µs, the |g⟩–|f⟩ encoding at g_gain/2π = 133 kHz gives T₁^f = 2.22 (± 0.06) ms (×16.7), T₁^g = 1.67 (± 0.34) ms (×12.6), T₂ = 7.49 (± 0.47) µs (÷14) and T_2E = 65 (± 4) µs (÷2.1).

• In-situ tunability: The noise bias, coherence and gate fidelity can be traded off by choosing g_gain; smaller values give lower bias but higher coherence and fidelity.

Methodology in Plain English

The team started from an ordinary superconducting transmon coupled to a microwave cavity — the same circuit used in many standard readout schemes. Rather than building a new type of circuit element, they stored the qubit in the transmon's ground and second-excited states (|g⟩ and |f⟩) and added one continuous microwave pump tone.

That tone is tuned to a special frequency where two pump photons can be converted, via the Josephson junction's nonlinearity, into one transmon photon on the |e⟩–|f⟩ transition plus one cavity photon. Because the transmon's anharmonicity is large (172 MHz), this conversion happens only on that one transition and does not disturb the computational states. The cavity's intrinsic loss then absorbs the cavity photon irreversibly, so the process is one-way: it acts as a continuous "gain" that pushes population from |e⟩ back up to |f⟩. If a loss event knocks |f⟩ down to |e⟩, the pump automatically restores it — correcting the error without measurement or feedback.

The team then characterized the system step by step: they mapped the pump resonance versus pump amplitude, verified the dynamics against master-equation simulations (run with QuTiP) using only independently measured parameters, measured relaxation, Ramsey and echo times as a function of pump strength, and finally performed randomized benchmarking of two-photon |g⟩–|f⟩ gates applied while the pump was on. All data in the main figures and the qubit characterization were taken in an interleaved fashion over a continuous four-day period. The device itself is a tantalum-on-sapphire transmon with an aluminium/aluminium-oxide Josephson junction inside a 3D microwave cavity.

Why This Matters

The work shows that a useful form of noise bias — the property that makes certain QEC codes dramatically cheaper in hardware — can be obtained without exotic circuit elements, using a tone on a standard architecture. This directly addresses the scalability objection that has limited previous noise-biased qubits, and it does so while preserving gate fidelities above 99.55%.

• Hardware-efficient quantum error correction: The paper proposes the device as a noise-biased ancilla qubit for fault-tolerant error detection, where the gain suppresses ancilla decay errors propagating into the logical qubit, and as a possible data qubit for noise-biased codes such as the XZZX surface code.
• Stronger interactions per encoding: Because the |g⟩–|f⟩ encoding gives a two-fold increase in interaction strength compared with |g⟩–|e⟩, decoherence during error-correction protocols is reduced.
• Applications cited for quantum computing more broadly: The introduction lists simulation of complex quantum systems, cryptography and optimization as motivations for quantum computation. The paper does not report device-specific demonstrations of these applications.
• Scalable fabrication: The device uses thin-film tantalum on sapphire with a Dolan-bridge Josephson junction, i.e. standard fabrication, and a single added microwave tone rather than new on-chip elements.

Industry relevance: The transmon-plus-readout-resonator circuit is among the most widely deployed superconducting architectures. A modification that requires only an extra microwave tone — rather than a redesigned chip — is far easier to insert into existing fabrication and control stacks, which matters for organizations pursuing hardware-efficient error correction at scale.

Future Directions

• Reducing thermal population: The authors identify n_th,a as a significant limitation on T₁^g and T₁^f, and point to improved thermalization and infrared filtering, noting that thermal populations as low as ≈0.1% have been reported in the literature.
• Replacing the readout cavity with a dedicated lossy element: The finite gain linewidth κ_b will, at larger g_gain, off-resonantly drive the |g⟩–|e⟩ and |f⟩–|h⟩ transitions and reduce T₁^g,f. A dedicated multi-stage lossy filter could allow higher κ_gain while preserving frequency selectivity and could also delay breakdown of the adiabatic elimination regime.
• Better simulation-guided design: Including Floquet simulations of the transmon–resonator system in the design process to avoid spurious multi-photon resonances.
• Extending the encoding and the gate set: Substituting higher transmon levels |n⟩ and engineering both gain to |n⟩ and loss to |g⟩ is predicted to give exponential suppression of bit flips in n while the phase-flip rate grows only linearly. The authors also note that a bias-preserving CX gate is prohibited for two-level qubits but no such no-go theorem exists for multilevel systems, so such a gate could be enabled via an additional transmon level or replaced by multi-qubit quantum-non-demolition measurements. Dissipative stabilization of qudit states is suggested as a further direction.

Target Audience

Experimental superconducting-qubit researchers and quantum error-correction theorists, particularly those working on dissipation engineering, noise-biased qubits, bosonic codes and hardware-efficient QEC. Graduate students entering circuit QED will find the paper a clear worked example of calibrating an engineered dissipator in a transmon, and hardware architects evaluating scalable routes to noise bias will find the parameter-by-parameter accounting of gains and costs most relevant. Readers without a background in superconducting circuits will need to consult the Supplementary Information, which covers the experimental setup, full system parameters, the decoherence model and the ancilla-qubit analysis.

Authors’ abstract

The interaction between a qubit and its environment can be engineered such that one error channel dominates over all others, resulting in noise bias. This property enables error correction codes to focus on the dominant error type, thereby significantly reducing the number of physical systems required for fault-tolerant quantum computation. However, engineering noise bias typically introduces complexity at the physical system level, which decreases its usefulness by limiting scalability. Here, we introduce and experimentally realize a noise-biased qubit in a standard transmon-readout resonator circuit, one of the most common superconducting architectures, by only adding a single microwave tone. We encode the qubit in the transmon $|\mathrm{g}\rangle$- and $|\mathrm{f}\rangle$-states, and engineer a frequency-selective gain channel that counteracts single-photon loss errors between the computational states. We demonstrate an order-of-magnitude enhancement in relaxation time compared to the $|\mathrm{g}\rangle-|\mathrm{e}\rangle$ encoding, conceding only a factor-of-two decrease in the echo-coherence time. Furthermore, we show that this qubit is compatible with fast, high-fidelity operations. Our results open a path towards using this system as a simple building-block for hardware-efficient quantum error detection and correction schemes.