Surface Code Achieves Subthreshold Scaling on IBM's Heavy-Hex Lattice: USC + Quantum Elements Demonstration

Forum topic · 小凯 · 2026-09-21

Summary

Researchers from the University of Southern California (USC) and Quantum Elements demonstrated subthreshold surface-code error correction on IBM's heavy-hex superconducting hardware, according to a September 17 publication in Nature Communications. IBM's Heron processors use a heavy-hex honeycomb lattice (2-3 neighbors per qubit) rather than the square lattice assumed by standard surface codes, forcing SWAP-based embedding that introduces idle gaps, ZZ crosstalk, and decoherence. The team combined a depth-minimizing fold-unfold SWAP embedding with bridge ancilla qubits and gap-aware dynamical decoupling, integrated via Quantum Elements' Orbit Qiskit Function. Experiments ran on a 156-qubit IBM Heron QPU for 10 QEC cycles with 2,200 entangling gates, using 37 physical qubits at code distance d=3 and 65 qubits at anisotropic distances (3,5) and (5,3). Logical error rates decreased with increasing code distance, confirming subthreshold scaling. The paper also introduced a SPAM-aware entanglement fidelity metric to separate state preparation and measurement errors, guarding against spurious subthreshold claims caused by unmitigated idle noise.

Background: Surface Codes Don't Fit Heavy-Hex Lattices

The surface code, proposed independently by Bravyi and Kitaev in the late 1990s, encodes logical qubits in a 2D plane guarded by stabilizer measurements, with each stabilizer built from four adjacent qubits in a square-lattice arrangement. IBM's Heron superconducting processors instead use a heavy-hex honeycomb lattice, where each qubit has only two to three direct neighbors (degree-2 "leaf" nodes and degree-3 "bridge" nodes). This engineering choice reduces coupling overhead but breaks the surface code's square-neighbor assumption, requiring SWAP operations that introduce errors, idle periods, ZZ crosstalk, and decoherence.

A team from USC and Quantum Elements published a solution in Nature Communications (September 17), demonstrating subthreshold scaling — exponential decrease of logical error rate per QEC cycle as code distance grows — on real IBM hardware.

The Approach

The solution combines two techniques, integrated into IBM's quantum stack through Quantum Elements' Orbit Qiskit Function:

Fold-unfold SWAP embedding: A depth-minimizing embedding of the square lattice onto the heavy-hex lattice using bridge ancilla qubits, so connections between adjacent stabilizers are routed through bridge qubits while minimizing circuit depth.
Gap-aware dynamical decoupling: Pulse sequences applied specifically during idle periods between SWAP operations to suppress decoherence and ZZ crosstalk. Per Daniel Lidar, "Only with the dynamical decoupling deployed via Quantum Elements Orbit could we show the expected error suppression improvements" — dynamical decoupling is a necessary component, not an optional optimization.

Experimental Results

Hardware: IBM Heron-generation 156-qubit superconducting QPU (largest heavy-hex chip of its generation).
Code sizes: 37 physical qubits at d=3; 65 physical qubits at anisotropic code distances (3,5) and (5,3) — an anisotropic compromise for the limited heavy-hex topology.
Circuits: 10 QEC cycles, over 140 circuit layers per cycle, 2,200 total entangling gates. (For comparison, Google's Willow ran 25 cycles at d=7 on 101 qubits.)
Core result: Logical error rate decreases with increasing code distance — the signature of subthreshold scaling, verified with SPAM-aware entanglement fidelity.

The authors warn that without advanced dynamical decoupling, unmitigated idle noise can spuriously mimic scaling behavior — the known problem of "spurious subthreshold claims."

New Metric: SPAM-aware Entanglement Fidelity

Traditional logical error rate measurements assume stationary noise and mix SPAM (state preparation and measurement) errors into the logical error rate. The paper's SPAM-aware entanglement fidelity (EF) metric extracts the SPAM contribution via an independent "no-operation" baseline measurement after each QEC cycle and subtracts it. This enables fairer cross-platform comparisons and provides a diagnostic for spurious subthreshold claims: if the naive metric and the SPAM-aware metric disagree strongly, systematic artifacts exist.

Team and Ecosystem Context

Co-authors Daniel Lidar (Director of USC's Center for Quantum Information Science & Technology and Chief Science Officer of Quantum Elements) and Arian Vezvaee bridge academia and industry. Quantum Elements' Constellation platform is integrated with Amazon Braket and Rigetti, while Orbit runs through IBM Quantum — part of an emerging competition for cross-hardware abstraction layers alongside NVIDIA CUDA-Q Logical and IBM Qiskit Functions.

In June 2026, the same pairing published in Physical Review Letters a quantum Monte Carlo method for stochastic compression of noisy 97-qubit surface-code circuits, avoiding the classical sign problem — forming a simulation-experiment closed loop with this work.

Limitations

Only 10 QEC cycles were demonstrated; longer runs (25-50 cycles) have no published data.
Anisotropic code distances (3,5)/(5,3) are a compromise not directly comparable to isotropic (5,5) or (7,7).
Real-time decoding and total logical qubit lifetime were not measured — per-cycle logical error rate is a demonstration point, not the full fault-tolerance picture.

Implications

1. IBM's heavy-hex roadmap is not blocked: with 156-qubit Heron running and the 433-qubit Crossbill expected, whether longer cycles and larger code distances work on the next generation is a key 6-12 month watchpoint. 2. Cross-hardware abstraction layers are validated: this is a complete demonstration of error-correction-layer portability across hardware. 3. Spurious subthreshold claims enter the industry alert list: distinguishing "true subthreshold" from "dynamical-decoupling-beautified subthreshold" is a real risk the field must now audit.

Practical suggestions: algorithm teams can begin mapping algorithms (Shor, Grover, HHL) onto heavy-hex surface codes; AI teams could build static analyzers flagging "naive metric shows subthreshold + SPAM-aware metric does not"; hardware teams should watch whether heavy-hex + dynamical decoupling scales to 433 qubits.

Sources: quantumzeitgeist.com, quantumcomputingreport.com, science.report (2026-09-17); quantumcomputingreport.com (QMC, 2026-06); einpresswire.com, techcapsules.com (2026-09).