The usual quantum-error-correction story assumes you can pause the processor, measure a syndrome, and let a classical computer decide the correction. A joint team at the University of Innsbruck, RWTH Aachen University, Forschungszentrum Jülich and Alpine Quantum Technologies has now run a complete logical algorithm without that pause. The algorithm is Grover’s search. The register is three logical qubits, encoded in eight physical qubits of a trapped-ion processor. The paper is Friederike Butt, Ivan Pogorelov and colleagues, “Demonstration of measurement-free universal logical quantum computation,” Nature Communications 17:995 (2026), DOI 10.1038/s41467-026-68533-x. Innsbruck posted the news on 7 April 2026. Nature’s version of record is 27 January 2026.
That is a real first, and it is still a proof of concept. The code is error-detecting, not error-correcting. The experimental success probability sits below the best classical guess. Three logical qubits on eight ions is not a product.
What happened
Most fault-tolerant recipes still need mid-circuit measurements and classical feed-forward: stop, read error information, decide a correction, resume. On trapped ions that pause is slow, heats the atoms, and lets idling qubits dephase. Butt and Markus Müller, at RWTH Aachen and Forschungszentrum Jülich’s Institute for Theoretical Nanoelectronics (PGI-2), designed a toolbox that never leaves the quantum circuit. Stabilizer information is copied onto auxiliary qubits; the correction is applied with ordinary gates. Entropy leaves later, when those auxiliaries are reset or swapped for fresh ions.
The hardware run was in Innsbruck. Pogorelov, Robert Freund, Alex Steiner, Marcel Meyer and Thomas Monz, at the Institute for Experimental Physics, implemented the circuits on a 16-ion chain of 40Ca+ in a linear Paul trap. Monz is also at Alpine Quantum Technologies (AQT). Butt and Pogorelov share first authorship.
Innsbruck’s 7 April note quotes the leads more broadly than the paper does. Butt: the coherent feedback “happens entirely within the quantum computation itself, using only standard quantum gate operations,” and is “particularly well-suited to hardware platforms where measurements are especially costly.” Pogorelov: “For the first time, we have shown that a complete fault-tolerant quantum algorithm can be executed without mid-circuit measurements with feed-forward control.” Monz calls it “a first, important step.” The paper’s closer is narrower: a first experimental fault-tolerant universal gate set without mid-circuit measurements, and a first fault-tolerant logical algorithm of that kind, on a search space of N = 8.
How the experiment worked
The device already can measure mid-circuit. The authors chose not to. A selective reset on this trap takes 1.7 ms; their current mid-circuit measurement is ≈30 ms. Process tomography on the data qubits gives reset fidelity 0.955(9) against 0.908(12) for the measurement path. Native operations are optically addressed single-qubit rotations at 729 nm, virtual Z gates, and Mølmer–Sørensen two-qubit gates with all-to-all connectivity. In the noise model fitted to this machine, single-qubit depolarizing is p1 = 3.6 × 10−3, two-qubit p2 = 2.5 × 10−2, idle T2 = 50 ms. Gates run sequentially, so idle dephasing accumulates.
Two codes do the logical work.
The [[4, 1, 2]] code encodes one logical qubit in four physical qubits at distance 2: any single error is detectable. Instead of lattice surgery with measurements, the team maps joint logical operators onto an auxiliary register and applies coherent CZ and CNOT feedback, so source and target never couple directly. Logical state tomography, 40,000 shots per state per basis, reaches up to 93(2)% for teleportation and 95(3)% for a logical Hadamard — both below a physical gate here (≈0.996).
The algorithm itself runs on the [[8, 3, 2]] code — the smallest three-dimensional color code, three logical qubits in eight physical qubits, again distance 2. It has a transversal non-Clifford: a logical CCZ from single-qubit T and T†. Logical CNOTs inside the block are relabelings of ion pairs, not extra pulses. The missing piece is a logical Hadamard, injected from an auxiliary [[4, 2, 2]] block prepared in |+0⟩L, with the usual measure-and-correct step replaced by two CNOTs. That gadget uses four auxiliary qubits and 26 two-qubit gates. Tomography (7,500 shots per logical state, qubit and basis) gives up to 81(3)% fidelity for HL on logical qubit 0 after keeping 10% of runs, and 65(6)% to 99.89(14)% on the two idling logical qubits. The authors estimate idle dephasing accounts for almost two-thirds of the logical error on that Hadamard.
Grover is compiled into the fault-tolerant set {HL, CNOTL, CCZL} inside one [[8, 3, 2]] block. Three logical qubits search a database of size N = 8. A phase oracle marks two solutions, |011⟩ and |101⟩. Noise-free, one Grover iteration succeeds with probability 1. The classically optimal strategy — one query, then a random guess among the rest — succeeds with probability ≈0.46. The experiment, 37,500 shots per measurement basis, reports (Fig. 3, 9 and 10):
| Quantity | Value |
|---|---|
| Search space N | 8 (three logical qubits) |
| Marked states | |011⟩ and |101⟩ |
| Experimental p_success = p011 + p101 | 0.40(4) |
| Optimal classical success | ≈0.46 |
| Same circuit on physical qubits | 76(2)% exp. / 77(1)% sim. |
| Simulated if p2 ≈ 0.015 | ≈0.52 |
| Simulated if T2 = 100 ms | ≈0.67 |
The two marked peaks are visible. The logical run does not beat classical, and it does not beat the same circuit on bare physical qubits. The authors say so. Their simulations argue that a 1% drop in two-qubit error, to a rate other ion traps have already shown, or a doubling of T2 to 100 ms, would push p_success above 0.46. That is a projection, not a second experiment.
Why measurement-free matters
Syndrome measurement is not a detail. In atomic hardware, fluorescence readout is orders of magnitude slower than a gate and usually needs recooling. Superconducting processors have the same time-scale gap. While you measure, every idle logical qubit is a decoherence clock. Measurement-free protocols move that work onto extra qubits and extra two-qubit gates. The cost is circuit depth and post-selection, not a 30-millisecond freeze.
The paper is explicit that the scheme is built for all-to-all trapped ions and is a candidate for neutral atoms, where long-range gates are already good and mid-circuit measurement with real-time feedback is still expensive. That is a second toolbox for platforms that cannot, or should not, stop — not a claim that measurement-based surface-code roadmaps are obsolete.
Limits
Keep the code name in view. [[8, 3, 2]] is an error-detecting code. Distance 2 means a single error can be flagged; it does not suppress errors as you add qubits the way a distance-3 or distance-5 surface code is supposed to. Accepted fractions after the logical Hadamard fall as low as 0.1 in the Z basis. Concatenating the teleportation gadgets, the authors note, would detect more errors and discard still more runs.
The Grover number that matters is 0.40(4), not the press line that the experiment “clearly identified the correct solutions.” Identification is real: the two solution bins stand above the other six. The integrated probability is still short of classical, and well short of the physical-qubit control at 76(2)%. The logical algorithm is running above break-even with its unencoded twin. Idle dephasing, including global magnetic-field noise on an eight-qubit GHZ-like |000⟩L, is the dominant error the paper can name.
AQT is a company. Monz’s competing-interest statement says he is connected to it. The remaining authors declare none. Eight ions are not a product.
What to watch next
Three questions, all in the paper’s discussion. First: does coherent feedback survive on a higher-distance code, where you correct instead of post-select? The authors sketch a surface-code route using d disjoint logical-operator representations and d-qubit GHZ auxiliaries, and they flag concatenation and a distance-3 construction without large GHZ states as open. Until someone runs that, the verified object is Grover on [[8, 3, 2]]. Second: do idle errors fall fast enough for the simulated crossover — p2 ≈ 0.015 or T2 = 100 ms — to show up in a repeat of this circuit, not only in Monte Carlo? Third: will a neutral-atom group compile this gate set? A measurement-free logical algorithm on atoms would test whether the pause was the bottleneck.
Until then the sourced claim is small and specific: a fault-tolerant Grover search on three logical qubits, eight physical 40Ca+ ions, no mid-circuit measurements, p_success = 0.40(4), Nature Communications 17:995 (2026).
