Noise, Error Rates And Why Error Correction Decides Everything

If you remember one thing about the state of quantum computing, make it this: the number of qubits in a headline matters far less than how often those qubits make mistakes. NIST’s explainer, updated in May 2026, says the best current machines err roughly once in every thousand operations.¹ A useful algorithm, such as one that could threaten today’s encryption, needs billions of operations to run without a mistake: a widely cited 2025 estimate for breaking 2048-bit RSA counts about 6.5 billion of one expensive type of gate alone.² Closing that gap is the job of quantum error correction.

This article explains how error correction works, what the “threshold” is and why the field treated Google’s 2024 Willow result as a turning point. It also gives you a short checklist for reading error correction claims, because the term “logical qubit” now appears in almost every vendor announcement and does not always mean the same thing.

Why Qubits Are So Error-Prone

A qubit stores information in a delicate balance of amplitudes. Stray heat, electromagnetic noise, imperfect control pulses and interactions with neighbouring qubits all nudge that balance. Two measures capture how well a qubit resists this. Coherence time is how long a qubit holds its state before the environment scrambles it. Gate fidelity is how often an operation does exactly what it was meant to.

Both have improved sharply. Google reported that its Willow chip, announced in December 2024 with 105 qubits, holds its state for close to 100 microseconds, about five times longer than its previous generation.³ Even so, a long algorithm would accumulate errors far faster than it could finish. Better hardware alone cannot close the gap.

Physical Qubits, Logical Qubits And The Threshold

Error correction spreads one unit of information across many physical qubits, creating what is called a logical qubit. Extra helper qubits are measured repeatedly to check for signs of an error, without reading the protected information itself. These check results are called syndromes. A classical decoder interprets them and works out which correction to apply, and the cycle repeats continuously.

There is a catch. Adding more physical qubits also adds more places for errors to occur. Error correction only pays off if the physical error rate is below a critical value called the threshold. Below it, making the code larger suppresses logical errors exponentially. Above it, making the code larger makes things worse.⁴ The size of a code is described by its distance: a distance-7 code can withstand more simultaneous errors than a distance-3 code.

What Willow Showed, And What Others Have Added

Google’s result, published in Nature, used Willow to run surface code memories at distances 3, 5 and 7. Each step up cut the logical error rate by a factor of about 2.14. The largest code used 101 qubits, and its logical qubit outlived the best physical qubit on the chip by a factor of about 2.4. In a separate distance-5 run, a real-time decoder kept pace with the hardware for up to a million correction cycles.⁴ This was the first clear demonstration that a superconducting surface code operates below threshold, a goal pursued for decades.

Other groups have since shown related results on different hardware. Harvard, MIT and QuEra reported a neutral-atom architecture with up to 448 atoms that combined the main ingredients of fault tolerance and performed 2.14 times below threshold in a four-round test, helped by detecting lost atoms and by machine learning decoders.⁵ Quantinuum, in a February 2026 preprint on its 98-qubit Helios trapped-ion machine, reported up to 94 logical qubits protected by error-detecting codes and up to 48 protected by error-correcting codes, with logical gate error rates around 1 in 10,000 and performance better than the same circuits run without encoding.⁶ In September 2026 Infleqtion announced 30 entangled logical qubits on 80 neutral atoms; that is the company’s own claim, and its release does not name the code or say whether errors were corrected or only detected.⁷

These results are not directly comparable. The table below shows why.

The Overhead Problem

Even when error correction works, it is expensive. IBM researchers estimated in a 2024 Nature paper that protecting 12 logical qubits with a surface code to a given standard would need almost 3,000 physical qubits. They proposed a different family of codes, quantum low-density parity-check codes, that would do the same job with 288.⁸ That was a design study, not an experiment, and it needs longer-range connections between qubits than a flat grid provides. IBM’s experimental Loon chip, designed to supply them, was announced in November 2025.⁹¹⁰ It still explains why roadmaps now differ so much in how many physical qubits they say a useful machine needs.

Overhead is also why estimates for breaking RSA-2048 are quoted in physical qubits. Craig Gidney’s 2025 estimate of under a million physical qubits assumes every physical operation fails no more than 0.1 percent of the time, with error correction running throughout.² A February 2026 preprint that uses quantum low-density parity-check codes brings the figure under 100,000 under the same error assumption, in exchange for a run of about a month, which is the overhead saving described above put to work.¹¹

How To Read An Error Correction Claim

When a press release announces a logical qubit count, three questions tell you most of what you need. First, what is the code distance, and does the logical error rate fall as distance grows? Second, are errors corrected, or only detected and discarded? Third, was the result peer reviewed, or is it a preprint or company announcement? The article on reading vendor roadmaps builds on these questions.