Quantum error correction well below the color code threshold

AI summary

Overview

Research area: Quantum error correction (QEC) on hardware — specifically topological color codes implemented on a trapped-ion quantum processor.

Technical level: Advanced. The paper assumes familiarity with stabilizer codes, code distance, syndrome extraction, and decoders, though the core results (error rates falling as the code gets larger) are stated plainly enough to follow without that background.

Scope (one sentence): An experimental demonstration that triangular color codes of distance 3, 5, and 7 run on Quantinuum's Helios-1 trapped-ion processor operate below the color-code threshold, with logical error rates suppressed by a factor of roughly 12 per increase in distance for memories and roughly 9 for Clifford gates.

What This Paper Is About

Quantum error correction only helps if the physical error rate is below a critical "threshold" — otherwise adding more physical qubits to encode one logical qubit makes things worse rather than better. The problem this paper addresses is that color codes, despite having attractive features (a full transversal Clifford gate set), have demanding syndrome-extraction circuits and are highly vulnerable to leakage, so below-threshold operation had been hard to demonstrate cleanly in hardware.

The goal is to show, on a real trapped-ion device, that logical error rates fall steeply and consistently as the color code distance grows from 3 to 5 to 7, and to show that fast neural-network decoders can do this decoding in real time.

Key Contributions

1. Below-threshold distance scaling for color-code memory. Logical error rates per syndrome extraction cycle fall from 5.5⁻⁰·⁴₊₀.₄ × 10⁻⁴ (d = 3) to 3.8⁻¹₊₁ × 10⁻⁵ (d = 5) to 5⁻²₊₄ × 10⁻⁶ (d = 7), giving a suppression factor of Λ_Mem. = 12⁻²₊₃.

2. Below-threshold encoded Clifford two-qubit randomized benchmarking. Using color codes at d = 3 and d = 5, the authors report 7⁻¹₊₁ × 10⁻⁴ and 7⁻³₊₃ × 10⁻⁵ error per logical Clifford gate, corresponding to Λ = 9⁺⁸₋₃ (the abstract's value; the results text quotes 9⁺⁸₋₄).

3. A hardware-level leakage repump scheme for ¹³⁷Ba⁺ ions that optically returns leaked population to the computational subspace, combined with a circuit-level SWAP-based leakage-reduction unit (LRU). The repump alone improves observed logical error rate by 4× at d = 3 and 8× at d = 5 compared with the circuit-level LRU alone.

4. Two neural-network decoders (a 3D CNN and a round-token transformer, RTT) trained only on emulator data and applied to hardware data without fine-tuning, plus a streaming WebAssembly implementation of the Frontier decoder — all three achieving latencies compatible with real-time decoding.

Main Findings

• Error suppression of over an order of magnitude per distance step. Fitting the memory data to log(LER(d)) = −(log Λ)/2 · d + c yields Λ = 12⁺³₋₂ (Tesseract), 11⁺²₋₂ (Frontier), 7⁺¹₋₁ (CNN), and 8⁺¹₋₁ (RTT).

• Algorithmic decoders are statistically indistinguishable. Tesseract and Frontier give, for example, 178 versus 166 failures out of 15,840 shots at d = 3 and 17 versus 18 out of 21,600 at d = 5; the two neural decoders are slightly worse but comparable (171 and 172 failures at d = 3; 37 and 29 at d = 5; 3 and 5 at d = 7 out of 29,000 shots).

• Leakage mitigation is the enabling ingredient. With three repump cycles the leaked population is suppressed by 6.5(1)×, with a per-cycle suppression of 1.89(1)× and a fitted decay factor r = 0.529(3). At the operating point the repump adds no time to the circuit and negligible gate error because it overlaps with ground-state cooling and ion transport.

• Leakage accounts for 41% of the total 1/Λ error budget. The simulated error budget attributes 43% to two-qubit gate error, 36% to memory errors (mostly transport-related), and 13% to mid-circuit measurement and reset cross-talk.

• Real-time decoding is viable. Per-round GPU compute times on an NVIDIA H100 80GB are approximately 0.26, 0.36, 0.52 ms for the CNN and 0.22, 0.25, 0.35 ms for the RTT at d = 3, 5, 7 — described as an order of magnitude faster than the syndrome extraction rate.

• No significant time-correlated errors. Pearson correlations between detector firings in the d = 7 data show each syndrome correlates only with its own round and the immediately neighboring rounds.

• Syndrome data doubles as hardware diagnostics. "Error spectroscopy" on the syndrome density revealed inaccuracies in qubit phase tracking on Helios, which were fixed in software; the authors include data both with and without that improvement, noting it did not significantly change logical error rates.

Methodology in Plain English

The code. The team uses triangular color codes, which pack (3d² + 1)/4 physical qubits into one logical qubit at distance d — so 13, 37, and 73 physical qubits for d = 3, 5, 7. Colleagues on a trapped-ion device can implement every Clifford gate transversally (bit-by-bit, without decoding), because the code is self-dual and the ion trap offers all-to-all connectivity.

Syndrome extraction. They use a "SWAP-superdense" circuit adapted from superdense syndrome extraction, in which ancilla qubits are swapped in and out so that leaked qubits get cycled away and measured. On Helios the SWAP itself is free — it is implemented as a software relabeling of qubits — which removes one layer of depth. The three different face orderings rotate over successive rounds so that every data qubit eventually gets swapped.

Memory experiment. They prepare logical states in six Pauli bases, run 20 syndrome-extraction rounds, and measure out. To keep each shot using a comparable number of physical qubits, they run 6, 2, and 1 experiments in parallel per shot at d = 3, 5, 7 respectively, totaling 15,840 / 21,600 / 29,000 shots. Because failures at d = 7 are rare enough that fitting a round-count curve is impractical, they estimate the per-round logical error rate from the single r = 20 point using LER = (n_F/n)/(r + 1/2), where the 1/2 accounts for state-preparation error.

Leakage handling. Hardware side: eight state-selective 1762 nm pulses per cycle move population out of the six leakage levels, 614 nm light then drives it through ²P₃/₂ to decay back, and 650 nm light clears the ²D₃/₂ shelving state. Circuit side: the SWAP-LRU sends leaked qubits into ancillas that get measured and reset.

Decoding. Tesseract (an A*-search most-likely-error decoder) is the reference; Frontier (pruned dynamic programming) and two neural networks are evaluated for speed. The neural nets are trained on Helios-1 emulator data generated with noise intensity increased by a factor of four, then applied to hardware data with no fine-tuning.

Benchmarking gates. For the Clifford two-qubit randomized benchmarking, logical gates are applied transversally with a syndrome extraction round inserted after every 7 logical Clifford gates (matching the CNOT-depth of 7 of the extraction circuit). Sequence lengths of m = 69, 139, 209 with 2,220, 1,110, and 666 sampled circuits give 9, 19, and 29 syndrome extraction rounds respectively. d = 7 is excluded here because those circuits need more qubits than Helios-1 has.

Why This Matters

Impact on research. Color codes offer a transversal Clifford group, which is a strong practical advantage over surface codes, but they carry a heavier syndrome-extraction overhead and are notoriously sensitive to leakage — so they have lagged behind surface codes in hardware demonstrations. Showing Λ ≈ 12 for memory and ≈ 9 for gates at d ≤ 7 places color codes on firm below-threshold footing. The paper also argues that scaled trapped-ion processors could reach application-relevant performance "with substantially smaller code blocks than indicated by previous results in other qubit modalities," which reframes the resource estimates for color-code-based fault tolerance.

Real-world applications (these are the broader implications of fault-tolerant quantum computing rather than claims tested in the paper):

• Simulating molecular electronic structure for drug and catalyst discovery
• Materials design, including strongly correlated and magnetic systems
• Optimization and logistics problems beyond classical reach
• Cryptanalytic tasks that motivate post-quantum migration timelines

Industry relevance. The system-level engineering matters as much as the physics: hardware-level leakage repumping that costs no circuit time, decoder inference in a few hundred microseconds on commodity GPUs, a WebAssembly Frontier implementation reachable from Quantinuum's Guppy real-time control stack, and error spectroscopy turning QEC syndrome data into a hardware debugging tool. The paper explicitly frames the results as approaching the logical error rates needed for "megaquop-scale quantum computation."

Future Directions

• Close the remaining gap to full fault tolerance: demonstrate below-threshold scaling of high-fidelity color-code magic state preparation, the one non-trivial circuit component (besides gates and syndrome extraction) this work does not cover.
• Run a complete Clifford + T circuit encoded in a color code with real-time decoding of syndromes during magic state teleportation.
• Leakage-aware decoding: determine how much of the benefit of repumping could instead be reclaimed by a decoder that explicitly models leakage errors, which would also reduce decoding latency by avoiding a search over higher-order faults.
• Fine-tune the neural decoders on hardware data or on small-data fine-tuning schemes, and move beyond stabilizer-simulation training data, which only approximates non-stabilizer effects such as coherent memory errors.
• Replace direct-sampling logical benchmarking with sample-efficient alternatives and learn circuit error models from syndrome data, which would also help characterize error floors on Helios.

Target Audience

Quantum error correction researchers and experimentalists working on fault-tolerant quantum computing, especially those focused on color codes, trapped-ion architectures, and leakage mitigation. It is also valuable to quantum-hardware engineers interested in real-time decoding pipelines and to computer scientists working on machine-learning decoders for quantum error correction. Readers without a background in stabilizer codes will find the headline results accessible but the methodology sections demanding.

Authors’ abstract

Quantum error correction enables scalable fault-tolerant quantum computation by suppressing logical errors exponentially with code distance once physical error rates fall below a critical threshold. Here we demonstrate below-threshold operation of triangular color codes with distances $d=3,5,$ and $7$ on Quantinuum's Helios trapped-ion quantum processor. Leveraging both circuit- and hardware- level leakage reduction schemes, we achieve logical error rates as low as $5_{-2}^{+4}\times 10^{-6}$, and demonstrate error suppression by over an order of magnitude, $Λ=12_{-2}^{+3}$, with each successive increase in code distance. Moreover, we demonstrate below threshold scaling of error rates for Clifford two-qubit randomized benchmarking circuits when encoded in $d=3,5$ color codes, achieving logical error rate of $7^{+3}_{-3}\times 10^{-5}$ per logical two-qubit Clifford at distance $5$ with a suppression factor of $Λ=9_{-3}^{+8}$. We compare logical performance for existing near-optimal algorithmic decoders, and both convolutional neural network and transformer-based neural network decoders capable of real-time decoding an order of magnitude faster than the syndrome extraction rate. These results approach the logical error rates required for megaquop-scale quantum computation and indicate that scaled trapped-ion processors could reach application-relevant performance with substantially smaller code blocks than indicated by previous results in other qubit modalities.