The Number That Matters in Quantum Computing Is the Error Threshold

Every quantum announcement leads with a qubit count, and that number tells you almost nothing. What decides whether a machine can run a long algorithm is the error-correction threshold: the physical error rate below which adding more qubits makes a computation more reliable instead of less. Above it, a bigger machine is a bigger mistake.

The threshold theorem is the load-bearing result in the whole field. Work published in the mid-nineties — Aharonov and Ben-Or, and separately Knill, Laflamme and Zurek — showed that if the error rate per physical operation is below a constant, then arbitrarily long quantum computations are possible with overhead that grows only polylogarithmically in the size of the computation. That constant is the threshold. Below it, error correction wins; each additional layer of encoding makes the logical error rate smaller. Above it, encoding adds more errors than it removes, and the failure probability climbs with every qubit you bolt on.

The theorem assumes things worth naming, because real machines are judged against them: noise that is local and roughly independent, fresh ancilla qubits available on demand, measurements that are fast relative to gate times, and the ability to feed a measurement result forward into a later gate. Each assumption costs hardware. The threshold number quoted in a paper is therefore a property of a noise model as much as of a machine.

What One Logical Qubit Costs

The surface code is the workhorse because it needs only nearest-neighbour connectivity on a two-dimensional grid and tolerates a high error rate — around one percent under idealized circuit-level noise, and less once you model measurement error, leakage out of the computational states, and crosstalk. Its structure is simple and its price is not. A distance-d patch uses on the order of d² data qubits: 49 at distance 7, 225 at distance 15, 729 at distance 27. Add the ancillas that measure the stabilizers and the real footprint is closer to twice that.

Distance is set by the error budget. A logical error rate falls roughly as (p/p_th)^(d+1)/2, so operating at a tenth of threshold and wanting one logical error in a billion operations pushes you to large d and therefore to hundreds or thousands of physical qubits per logical one. That is where the phrase hundreds of physical qubits per logical qubit comes from, and it is a consequence of physics and arithmetic, not of anyone's engineering being sloppy.

Then come the T gates. The surface code protects Clifford operations cheaply and protects the non-Clifford gates that make a computer universal only through magic state distillation, which consumes batches of noisy magic states to produce a smaller number of cleaner ones. A distillation factory is a permanent, substantial allocation of physical qubits sitting off to the side of the data block. For algorithms dominated by T gates — most of chemistry and much of cryptography — the factories, not the data, are the resource bill. This is why a machine described as having a thousand qubits might support a few dozen logical qubits, or only a handful if it is doing serious non-Clifford work.

Why the Count Says Nothing

A qubit count is a size, and size is only meaningful next to a fidelity and a clock. Two-qubit gate error dominates the budget on every technology I know of, so the number that matters most is the error rate of the entangling gate, followed by measurement error, connectivity, and how long a gate takes relative to how long the qubits stay coherent. A hundred-qubit machine with good two-qubit fidelity and all-to-all coupling can be closer to a useful device than a thousand-qubit machine with a worse gate and a sparse lattice, because the second one cannot even run the code that would fix it.

• Two-qubit gate error rate, per gate, on the actual device.
• Measurement and reset fidelity — the surface code measures constantly.
• Connectivity: nearest-neighbour lattices are cheap to lay out, expensive to route around.
• Gate time and coherence: error per operation is error rate multiplied by how often you operate.
• Whether the machine can do a non-Clifford gate at all, at any cost.

Announcements rarely put those five on one line, which is precisely the problem. Qubit count is the one metric that is easy to count, easy to grow, and almost independent of the thing you care about.

The Result That Made the Threshold Empirical

For a long time the threshold was a theoretical claim with simulation behind it. That changed in December 2024, when Google published surface-code results on its Willow processor showing the logical error rate dropping as the code distance increased — distances 3, 5 and 7, with each step in distance cutting the error roughly in half. That is the signature of operating below threshold: the curve bends the right way, and the exponent is measurable rather than assumed.

It is a real milestone and a small one. A distance-7 logical qubit is a demonstration that the mechanism works, not a logical qubit you would build an algorithm on, and the experiment ran on a code designed around one kind of noise. What it established is the more important thing: the threshold is not an abstraction, it is a place you can stand, and once you are on the right side of it, the engineering problem becomes scale and yield rather than impossibility.

What I Think

I would like to see one number on every quantum press release: measured two-qubit gate error, and whether the demonstrated logical qubit improved when the code got bigger. That single trend line separates the machines that are on the right side of the threshold from the ones that are accumulating qubits to accumulate a chart. A company with 200 qubits and a falling logical error curve has a roadmap. A company with 2,000 qubits and no such curve has a fabrication facility.

The uncomfortable part is that the threshold is not a finish line. Crossing it converts one hard problem — can we compute reliably at all — into a much more boring and much more expensive set of problems: packaging, cryogenics, wiring density, yield, and the factory floor. That is what scaling a real technology looks like, and the industry is only now arriving at it. The qubit count was always the easy number to publish. The threshold is the one that decides whether anything is coming.