141
Part Three · D · Error correction and fault tolerance

qLDPC codes and bosonic codes (cat, GKP)

The surface code is robust and local, but wasteful: each logical qubit costs hundreds of physical ones, and the overhead grows with the square of the distance. Two very different escapes are being built. One uses cleverer codes whose checks reach across the chip, storing 12 logical qubits in 288 physical ones. The other stores a qubit in the many energy levels of a single oscillator, so that much of the protection comes from physics rather than numbers of qubits.

On the plate: the cost of storing 12 logical qubits, two ways.

Quantum LDPC codes

The surface code’s parameters are [[∼2d2,1,d]][[\sim2d^2, 1, d]]: rate falling to zero. Classical LDPC codes (Chapter 136) achieve constant rate and growing distance. Their quantum versions were long elusive, because X and Z checks must commute. Jean-Pierre Tillich and Gilles Zémor’s 2009 hypergraph product combined two classical codes into a quantum one with constant rate and distance n\sqrt n. In 2022 Pavel Panteleev and Gleb Kalachev proved that good quantum LDPC codes exist, with both rate and relative distance constant, settling a twenty-year question; their constructions also underpin the NLTS proof (Chapter 129).

IBM’s 2024 gross code, a bivariate bicycle code, is [[144,12,12]][[144, 12, 12]]: 144 data qubits and 144 check qubits hold 12 logical qubits at distance 12. Each check touches six qubits, two of them far away, so the chip needs long-range couplers. Protecting 12 logical qubits at distance 12 with surface codes would need about 12×(2⋅122−1)=344412\times(2\cdot12^2 - 1) = 3444 qubits. IBM’s roadmap targets “Starling”, 200 logical qubits running 100 million gates, by 2029. Established for the code; Speculative for the roadmap.

Cat qubits

A cat qubit stores 0 and 1 as coherent states ∣α⟩|\alpha\rangle and ∣−α⟩|-\alpha\rangle of a microwave resonator (Chapter 36), stabilized by engineered two-photon loss. A bit flip needs the field to jump between two well-separated blobs, which becomes exponentially unlikely as the mean photon number nˉ=∣α∣2\bar n = |\alpha|^2 grows, roughly as e−2nˉe^{-2\bar n}. Phase flips grow only linearly with nˉ\bar n. The noise becomes strongly biased, so a simple repetition code against phase flips suffices. Set α=\alpha = 2.0 and see the Wigner function of the even cat ∣α⟩+∣−α⟩|\alpha\rangle + |-\alpha\rangle: two blobs, with interference fringes between them, negative in places, the mark of a non-classical state. In February 2025 AWS reported Ocelot, a chip with five cat data qubits in a repetition code. Established

GKP grid states

Daniel Gottesman, Kitaev and Preskill proposed in 2001 encoding a qubit in an oscillator as a grid of sharp peaks in position: ∣0L⟩|0_L\rangle has peaks at even multiples of π\sqrt\pi, ∣1L⟩|1_L\rangle at odd multiples. A small shift in position or momentum, the oscillator’s natural error, is detected by measuring position modulo π\sqrt\pi and undone. In 2023 Volodymyr Sivak and colleagues at Yale ran GKP error correction in a superconducting cavity that extended the qubit’s lifetime beyond the best uncorrected encoding in the same system by a factor of 2.27: beyond break-even. The plate shows the grid schematically. Established

What you now have

The surface code’s overhead, hypergraph products and good qLDPC codes, the gross code, cat qubits and biased noise, GKP grid states, and the 2023–25 experiments.

· · ·

Three questions to carry with you.

What is the price of qLDPC codes’ efficiency?tap
Non-local checks, needing long-range connections on the chip.
Why are cat qubits’ bit flips suppressed?tap
The two basis states are separated coherent states; a flip requires an exponentially unlikely jump between them.
What error does GKP correct?tap
Small displacements in position and momentum of the oscillator.

And one to work out. 3444 surface-code qubits versus 288 for the gross code: by what factor is the gross code smaller (nearest whole number)?

factor ≈

Sources: J.-P. Tillich and G. Zémor, IEEE Trans. Inf. Theory 60, 1193 (2014) (arXiv 2009). P. Panteleev and G. Kalachev, Proc. 54th ACM STOC, 375 (2022). S. Bravyi et al., Nature 627, 778 (2024). D. Gottesman, A. Kitaev and J. Preskill, Phys. Rev. A 64, 012310 (2001). V. V. Sivak et al., Nature 616, 50 (2023). H. Putterman et al., Nature (February 2025) (Ocelot). IBM, “Large-scale fault-tolerant quantum computing” (2025 roadmap). Checked by code: 12 × 287 = 3444.

Next · Chapter 142
Compilation: gate synthesis, routing, optimization

The same algorithm can cost 10× more on the wrong chip layout.

Turn the page