Dual-rail superconducting qubits generate high-fidelity logical entanglement, study finds

A team at the Shenzhen International Quantum Academy has demonstrated multi-qubit entanglement using a superconducting processor built from dual-rail erasure qubits — an architecture that treats the dominant error type in superconducting systems as detectable rather than invisible. The result, published in Nature Physics on March 6, 2026, achieved a logical controlled-NOT (CNOT) gate with 98.1 percent process fidelity at a 13 percent erasure rate, alongside a three-logical-qubit Greenberger-Horne-Zeilinger (GHZ) state at 93.9 percent fidelity. The work represents a genuine step in error-corrected quantum computing — not a qubit count press release wearing a lab coat.

Erasure qubits are built on a specific idea: in superconducting circuits, the dominant error is amplitude damping, where a qubit decays toward its ground state. In a dual-rail encoding, a logical qubit is stored across a pair of transmons, and the logical states correspond to which rail holds a single excitation. When a rail loses its excitation to a non-computational state — that is, when leakage occurs — the error becomes detectable via mid-circuit ancilla measurement. The location is known. Conventional superconducting qubits cannot say the same: leakage to non-computational states in a standard transmon is invisible to syndrome extraction, which is why most error correction schemes treat it as a background hazard.

The processor integrates four dual-rail qubits, each comprising a pair of tunable transmons and an associated ancilla for erasure detection. Logical relaxation time (T1) reached 0.98 milliseconds and dephasing time (T2) hit 0.66 milliseconds — numbers that put millisecond coherence within reach for logical qubits. Single-qubit gate errors were measured at 2.6 times ten to the negative five per Clifford gate, placing the error rate at the ten-to-the-negative-five level. These are the best single-qubit gate fidelities reported in a multi-qubit superconducting processor with this encoding.

With those foundations in place, the team generated entangled logical states. A logical Bell state — a two-qubit maximally entangled state — reached 98.8 percent fidelity. The three-logical-qubit GHZ state, a generalized entangled state used as a benchmark for multi-qubit coherence, came in at 93.9 percent fidelity. The logical CNOT gate — the two-qubit operation that, together with single-qubit gates, constitutes a universal quantum gate set — achieved 98.1 percent process fidelity via quantum process tomography. Logical entanglement lifetime extended to over 100 microseconds while maintaining above 70 percent fidelity, roughly an order of magnitude longer than the physical Bell state counterpart.

The limiting factor is coupler-induced decoherence. The tunable couplers used to activate two-qubit interactions between logical qubits introduce additional noise channels during gate operations. Erasure rates during the gate, and residual dephasing from the coupling infrastructure, constrain how much the gate fidelity can be pushed before the coupling architecture itself becomes the bottleneck. This is not a solved problem — it is the next problem.

The result comes from researchers led by Youpeng Zhong and Dapeng Yu at the Shenzhen International Quantum Academy, with Xiayu Linpeng as corresponding author. The institution sits within Southern University of Science and Technology and operates under Guangdong Provincial Key Laboratory of Quantum Science and Engineering. A parallel institution, Shenzhen Branch of Hefei National Laboratory, also appears on the author list.

The framing matters. This is not a demonstration of quantum advantage, nor does it claim fault-tolerant computation at scale. It is a demonstration that dual-rail erasure qubits — previously shown to work at the single-qubit level — can generate and sustain multi-qubit entangled states with error-biased protection baked into the hardware. That is a meaningful distinction from a surface code demonstration running on a conventional superconducting qubit, where leakage remains an untreated background process.

The field has been here before in spirit: Google Quantum AI showed below-threshold surface code performance in 2024, and earlier results from IBM, Amazon, and academic groups have pushed error-corrected logical qubits to levels that outperform their physical counterparts. What the dual-rail approach offers is not a new threshold but a different resource equation. If erasure errors are the dominant channel — and in superconducting systems, they are — then building an architecture that detects and corrects them more efficiently could require fewer physical qubits per logical qubit than a conventional approach. That efficiency argument has been theoretical for years. This is the experimental data supporting it.

The gap between this result and fault-tolerant quantum computing is not small. Four logical qubits is a demonstration, not a system. The coupler-induced decoherence identified in the paper is a known scaling problem that the authors acknowledge needs further engineering. Concatenated codes layering additional error correction on top of the dual-rail substrate are proposed in the paper as the next step — which is the right way to read it: this is a building block, not a product.

What to watch next: whether the coherence times and gate fidelities hold as the team adds more qubits and deeper entanglement, and whether the erasure conversion efficiency — currently sitting at 13 percent erasure rate during the CNOT — can be pushed lower. The architecture's promise is real. The engineering is not done.