One Logical Qubit Needs a Thousand Bodyguards: How Quantum Error Correction Finally Works
In December 2024, Google released Willow, a 105-qubit chip, and most headlines focused only on speed. But the real headline was buried in the error correction section: each two-step increase in code distance cut the logical error rate roughly in half. The more qubits you add, the bigger the system gets—and the fewer errors it makes. That seemingly unremarkable statement took nearly thirty years for the quantum computing field to realize.
To understand what it was waiting for, you have to see how fragile qubits are.
Fragile Bits
A classical bit on a hard drive is the collective will of hundreds of millions of electrons; written once, it sits stable for decades. Qubits are different—their information lives in the spin of a single atom or the oscillation of a superconducting circuit. Thermal noise at room temperature, stray electromagnetic fields, material defects, even cosmic rays can knock it off course. Today's best superconducting qubits start forgetting after tens to hundreds of microseconds: the phase information of the superposition state quietly leaks away, a process called decoherence.
Real computations require thousands of gate operations to complete before a qubit forgets; running Shor's algorithm demands astronomically more. A gap of more than a dozen orders of magnitude, bridged by exactly one thing: error correction.
Why the Old Trick Fails
How do classical computers handle errors? Copying. Store the same data three times, read it out, take a majority vote. This simple scheme underpins all the reliability of modern computing.
The quantum world doesn't allow copying. In 1982, physicists proved the no-cloning theorem: unknown quantum states cannot be copied. This isn't an engineering shortcoming—it's a law of nature. To learn what α and β are, you'd have to measure, and measurement collapses the superposition. The act of copying is itself destruction.
The Way Out: Don't Copy, Spread Out
In 1995, Shor found a path around the prohibition: encode the information of one "logical qubit" into the entanglement structure of 9 physical qubits. The information doesn't reside on any single qubit but is spread across their correlations. If any one qubit gets knocked off course, the overall relational structure shows a crack—it can be detected and repaired.
The clever part is how you "check." You can't read the data directly—one read and it collapses. Instead you perform an indirect measurement: check whether the relationships between neighboring qubits have changed. It's like checking whether a book has been tampered with without opening the text—just verify the page numbers connect and the table of contents matches. The text remains untouched, yet the intruder is caught. Each round's detected "symptoms" point to the error's location, and a quantum gate flips it back.
This idea grew into the surface code: spreading a logical qubit across a 2D grid and repeatedly checking symptoms. It's tolerant of hardware imperfections—physical error rates only need to be pushed below 1%, and in theory, the more qubits you add, the lower the logical error rate goes. This "threshold exists" result, the threshold theorem, was proven in the late 1990s and is the theoretical bedrock behind the field's sustained investment.
What Willow Actually Measured
The threshold theorem says "below threshold, more is better," but between theory and measurement lay more than two decades of engineering. The embarrassment of all previous chips: as the code grew, logical error rates rose instead of falling—the newly hired bodyguards were making mistakes themselves, and more people meant more chaos.
Willow bent the curve for the first time: measured step by step with distance-3, 5, and 7 codes, the logical error rate dropped to about 1/2.14 of the previous level at each step. One hundred and five physical qubits ultimately safeguard roughly one logical qubit. Below threshold, experimentally confirmed—"more error correction, more errors" went from prophecy to data.
As a side note, in mid-October Google added another move: magic state cultivation—cutting the error rate of a scarce resource in error correction ("magic states," needed to execute certain critical quantum gates) by about 40 times. One more of the bottlenecking hands loosened.
How Far From Useful?
Honest numbers: the latest engineering estimates put breaking RSA-2048 with Shor's algorithm at under one million physical qubits. At Willow's current exchange rate—105 physical qubits for 1 logical qubit—even if the ratio is optimized to a thousand to one, a million physical qubits yield only about a thousand logical qubits. "A thousand bodyguards protecting one" is the exchange rate a decade of engineering must deliver.
IBM's roadmap promises Starling by 2029: a fault-tolerant machine with 200 logical qubits. From 1 to 200 and beyond, every step is an expansion of the bodyguard corps.
So when looking at any quantum computer today, the number to watch isn't total qubit count—it's the logical error rate. It determines whether the machine is a toy or a prototype. Quantum computing's enemy was never slowness of calculation, but speed of forgetting. Whoever suppresses the forgetting first turns the exponential space into programmable reality.
*(Third in a quantum intro series. The previous post covered interference—how quantum computers get work done; this one covers what kind of bodyguard team is needed to preserve those results. The next returns to earth: what this machine can and cannot actually do.)*