Quantum computers are powerful in theory but fragile in practice. A stray electromagnetic field, a tiny vibration, even a cosmic ray can scramble the quantum states they depend on, corrupting a calculation before it finishes. The answer is error correction, but error correction has always had a problem of its own: it adds so many extra operations and qubits that, for most of the field’s history, it has made qubits less reliable overall, not more. IonQ has just demonstrated, across nine different error-correcting codes on a single device, that this has changed. Its error-corrected qubits now last longer than the raw physical qubits they are built from.
The Noise Problem Blocking Useful Quantum Computing
A quantum computer’s power comes from qubits holding multiple states simultaneously and interacting in ways that classical bits cannot. But qubits are extraordinarily sensitive to the environment. Any interaction with the outside world, heat, vibration, stray fields, tends to push them into definite states, erasing the quantum properties that make them useful. This process, called decoherence, means that without error correction a quantum computer can only run short circuits before errors overwhelm the result. The algorithms that would genuinely beat classical computers at drug discovery, materials simulation, logistics, and cryptography all require circuits far longer than today’s physical qubits can sustain unaided.
Error correction solves this by spreading one logical qubit across many physical qubits, using redundancy to detect and fix errors as they occur. The challenge is that error correction itself introduces overhead. Measuring the error syndromes requires extra qubits, extra gates, and extra time, all of which introduce new opportunities for errors. For most of the past decade, those costs have outweighed the benefits: a logical qubit built from error correction has been shorter-lived than the physical qubits it was built from. Reaching the breakeven point, where the logical layer finally outlasts the physical one, is one of the key milestones on the road to fault-tolerant quantum computing.
What IonQ Built
IonQ’s experiment used a stationary chain of 40 barium-133 ions as its physical qubits. Using steerable laser beams, any two ions in the chain can interact directly, giving the device effectively all-to-all connectivity. The team ran nine structurally different error-correcting codes across three families: five quantum low-density parity-check (qLDPC) codes, two toric codes, and one concatenated code, all without any hardware reconfiguration between experiments. The codes ranged from 16 to 30 physical qubits per logical block.
The biggest technical advance is a new implementation of the optical-metastable-ground (OMG) architecture. In most trapped-ion systems, measuring ancilla qubits mid-circuit requires physically moving ions to a measurement zone, consuming most of the circuit runtime and requiring dedicated coolant ions. The OMG approach keeps all ions in place. It temporarily shelves every qubit into a metastable energy level, selectively measures only the target ancillae in place, uses that same measurement step to cool the chain, then returns all ions to normal operation. No ions move. No dedicated coolant ions are needed. Each measurement round also catches any qubit that has leaked out of its proper energy manifold, providing an early-warning flag rather than silent corruption.
What Each Code Achieved
The table below shows logical error rate per logical qubit per syndrome cycle for each of the nine codes, alongside the superconducting benchmark from the only prior experimental demonstration of the same BB code class. Lower is better.
The standout result is the GB [[26,2,5]] code, which encodes two logical qubits into 26 physical qubits and achieved 0.96% Z-error rate per cycle, the lowest single figure in the experiment. The BB [[18,4,3]] code, encoding four logical qubits into 18 physical qubits, is the direct comparison to the superconducting experiment: IonQ hit 2.01% for X errors and 1.08% for Z errors, against 8.67% and 9.15% on the custom superconducting chip. The toric codes both came in below 1.6% per cycle. The concatenated [[16,4,4]] code had the highest error rates at 3.10% and 2.35%, reflecting the extra circuit depth that concatenation requires.
All nine codes performed well below the superconducting baseline of roughly 9% per cycle. More importantly, every code family reached logical qubit lifetimes within error bars of the physical qubit lifetime, and several exceeded it. The physical qubits in this system have a T2* coherence time of around 1.1 seconds. Using a lifetime metric that averages across all qubit state preparations, the best logical qubit result reached 3.95 seconds, against 3.3 seconds for the underlying physical layer. This is the first time breakeven has been demonstrated with qLDPC codes. Google’s earlier breakeven result used a surface code, which requires far more physical qubits per logical qubit and has a lower encoding efficiency.
Why qLDPC Codes Are the Efficient Route Forward
Surface codes, the most widely used error-correcting codes today, require each qubit to interact only with its immediate neighbours. This makes them straightforward to build on 2D chip architectures, but their encoding efficiency is poor: a single logical qubit at useful error rates will need hundreds of physical qubits. Building a machine with thousands of useful logical qubits, the scale needed for practically relevant algorithms, therefore requires millions of physical qubits. That scale is far beyond anything that exists today.
qLDPC codes relax the locality constraint. Each qubit still interacts with only a bounded number of others, but those partners can be anywhere in the device. The result is a far better encoding rate. IonQ’s smallest code instance here packs four logical qubits into 18 physical qubits. The surface code would need hundreds to match that logical capacity. Trapped-ion hardware is uniquely suited to this because any two ions can interact directly via laser beams, removing the routing bottleneck that makes qLDPC codes impractical on most other platforms. The superconducting demonstration of the equivalent code required a purpose-built chip with custom long-range couplers designed for that specific code’s connectivity graph. IonQ ran the same code, plus eight others, on the same general-purpose hardware.
Source: https://arxiv.org/pdf/2606.06455
