The short version. The error threshold in quantum computing is a phase boundary. For the surface code, decoding maps onto an order-disorder transition in a disordered magnet. Below the critical noise level, larger codes suppress logical errors exponentially; above it, they make things worse.
A number that decides everything
Quantum bits are fragile. In the second half of the 1990s, several groups proved the threshold theorem: if the error rate per operation stays below a constant value, the threshold, error-correcting codes can make a quantum computation as reliable as you like, at a cost that grows only slowly with the size of the computation [1], [2]. Above the threshold, more error correction only adds more opportunities for error.
My own background is in statistical mechanics and Markovian models rather than in building qubits, so I will make the case I can make: for the surface code, the threshold is a phase transition in the precise sense physicists use, with an ordered phase, a disordered phase and a critical point between them.
The surface code in one paragraph
The surface code, introduced by Alexei Kitaev, stores quantum information in the topology of a two-dimensional lattice [3], [4]. Some qubits on the grid hold data; others repeatedly check the parity of their neighbours, and a check that changes value flags a nearby error. A classical program, the decoder, reads this pattern of flags, called the syndrome, and infers which errors most plausibly caused it. The code distance, d, is the smallest number of physical errors that can corrupt the stored information undetected.
The mapping onto a magnet
In 2002, Dennis, Kitaev, Landahl and Preskill showed that decoding the surface code is a problem in statistical mechanics [4]. In the simplest case, each qubit suffers a bit flip with independent probability p, and the checks are measured perfectly. Recovery then becomes a question about the two-dimensional random-bond Ising model: spins pointing up or down, with neighbours preferring to align except across a random fraction p of “wrong-sign” bonds. The actual errors fix that pattern, like impurities frozen into a material. The decoder’s uncertainty plays the role of thermal fluctuation: alternative error patterns consistent with the same syndrome behave like domain walls between regions of up and down spins. A logical error is a domain wall that spans the whole system.
The elegant part is that the magnet’s temperature is not a free parameter. A relation Hidetoshi Nishimori found in 1981 ties it to the error probability, so that thermal fluctuations and frozen disorder are drawn from the same distribution [5], [7]. Changing p moves the system along a single path through the phase diagram, the Nishimori line.
In the ordered phase, a domain wall costs free energy in proportion to its length, so a wall spanning a system of size d is exponentially unlikely, and so is a logical error. In the disordered phase, domain walls are cheap and appear at every scale, and no decoder can tell which class of error occurred. The threshold is the critical point between the two. For this idealized noise, numerical studies place it at p = 0.1094 ± 0.0002, about 10.9% [6].
What realistic noise does to the number
Real devices also misread the checks, so the syndrome is measured repeatedly. Time becomes a third dimension, and the mapping leads to a three-dimensional gauge model with random “wrong-sign” plaquettes [4], [7]. The zero-temperature version of each model describes a decoder that seeks the single most likely error pattern, using a classical matching algorithm, rather than weighing them all. Because minimizing energy is not the same as minimizing free energy, its threshold is slightly lower [7]. A distinction every student of statistical mechanics meets early reappears as the gap between a good decoder and an optimal one.
| Noise and decoder | Statistical model | Threshold |
|---|---|---|
| Bit flips, perfect checks, optimal | 2D random-bond Ising | about 10.9% [6] |
| Same, matching decoder | 2D model at zero temperature | about 10.3% [7] |
| Faulty checks too, matching | 3D gauge model at zero temperature | about 2.9% [7] |
| Every operation can fail | Circuit simulation | 0.57% per step [8] |
The exact value depends on the noise model; the existence of a threshold does not.
The equation behind the experiments
Below threshold, the logical error rate of a distance-d surface code is commonly approximated as
ε_d ∝ (p / p_th)^((d + 1) / 2)
Here ε_d is the probability of a logical error per round of error correction, p is the physical error rate, p_th is the threshold, and the exponent (d + 1) / 2 is the number of physical errors it takes to defeat a code of distance d [9]. The ratio p / p_th is the whole story. Below 1, each step up in distance multiplies the logical error rate by a number smaller than 1; above 1, each step up makes things worse. Experimentalists summarize this with the suppression factor Λ (lambda), the logical error rate at distance d divided by the rate at distance d + 2, roughly p_th / p [9].
This is also how thresholds are measured. Plot the logical error rate against the physical error rate for several distances, and the curves cross at one point: below it larger codes do better, above it they do worse [8]. Statistical physicists locate critical points in magnets with the same finite-size crossing, because at a critical point the system looks alike at every size.
Experiments on the ordered side
For years, experiments sat close to the crossing. In 2023, Google Quantum AI reported a distance-5 surface code at 2.914% logical error per cycle against 3.028% for distance 3: better, but only just [10]. That is life near a critical point, where size barely matters. At the end of 2024, the same group reported Λ = 2.14 ± 0.02 between distances 5 and 7, with a 101-qubit distance-7 code at 0.143% error per cycle [9]. The device has moved into the ordered phase: each step in distance cuts the logical error rate by a factor of about two.
Where the analogy strains
The mapping assumes local, independent errors. Correlated errors are the statistical-mechanics equivalent of long-range interactions, which can reshape a phase diagram. In superconducting processors, high-energy radiation such as cosmic rays can disrupt many qubits across a chip at once [11]. In the 2024 study, a distance-29 repetition code, a simpler code that guards against only one type of error, hit a floor of about one logical error in ten billion cycles, set by rare correlated bursts that arrived roughly once an hour [9].
Established: for independent local noise, the surface-code threshold is the critical point of a well-defined statistical-mechanics model [4], [6], [7]. Also established: at least one hardware platform now operates with Λ above 2 [9]. Not yet established: that exponential suppression continues unchanged to the much larger distances long algorithms need, given correlated errors the simple mapping leaves out. Speculative, and therefore not something I will offer: a timeline.
For anyone weighing a quantum roadmap, ask for Λ, and ask how it was measured across distances. A physical error rate tells you where a device is. Λ tells you which phase it is in.
References
- D. Aharonov and M. Ben-Or, “Fault-tolerant quantum computation with constant error,” in Proc. 29th Annual ACM Symposium on Theory of Computing (STOC), 1997.
- E. Knill, R. Laflamme and W. H. Zurek, “Resilient quantum computation,” Science, vol. 279, no. 5349, pp. 342-345, 1998.
- A. Y. Kitaev, “Fault-tolerant quantum computation by anyons,” Annals of Physics, vol. 303, no. 1, pp. 2-30, 2003.
- E. Dennis, A. Kitaev, A. Landahl and J. Preskill, “Topological quantum memory,” Journal of Mathematical Physics, vol. 43, no. 9, pp. 4452-4505, 2002.
- H. Nishimori, “Internal energy, specific heat and correlation function of the bond-random Ising model,” Progress of Theoretical Physics, vol. 66, no. 4, pp. 1169-1181, 1981.
- A. Honecker, M. Picco and P. Pujol, “Universality class of the Nishimori point in the 2D ±J random-bond Ising model,” Physical Review Letters, vol. 87, no. 4, art. 047201, 2001.
- C. Wang, J. Harrington and J. Preskill, “Confinement-Higgs transition in a disordered gauge theory and the accuracy threshold for quantum memory,” Annals of Physics, vol. 303, no. 1, pp. 31-58, 2003.
- A. G. Fowler, M. Mariantoni, J. M. Martinis and A. N. Cleland, “Surface codes: Towards practical large-scale quantum computation,” Physical Review A, vol. 86, no. 3, art. 032324, 2012.
- Google Quantum AI and Collaborators, “Quantum error correction below the surface code threshold,” Nature, vol. 638, pp. 920-926, 2025.
- Google Quantum AI, “Suppressing quantum errors by scaling a surface code logical qubit,” Nature, vol. 614, pp. 676-681, 2023.
- M. McEwen et al., “Resolving catastrophic error bursts from cosmic rays in large arrays of superconducting qubits,” Nature Physics, vol. 18, pp. 107-111, 2022.