Cat Codes and Biased Noise
Cat codes store a qubit in superpositions of coherent states, engineering the hardware so that one type of error becomes exponentially rare.
Coherent-state code words
A cat code encodes a logical qubit into superpositions of coherent states of an oscillator, such as |alpha> plus |-alpha> and |alpha> minus |-alpha>. As the amplitude alpha grows, the two coherent states become nearly orthogonal and well separated in phase space. A stabilizing mechanism, typically two-photon dissipation or a Kerr nonlinearity, confines the state to this cat manifold.
Engineered noise bias
The point of cat codes is bias. In a well-designed cat qubit, phase-flip errors (Z) occur at a rate that only grows linearly with alpha^2, while bit-flip errors (X) are suppressed exponentially in alpha^2 because flipping between the two far-apart coherent states requires a large jump. Real devices have shown bit-flip times orders of magnitude longer than phase-flip times.
- Bit-flip suppression is exponential in the mean photon number.
- Phase flips remain and must be corrected by an outer code.
- Two-photon dissipation autonomously stabilizes the cat manifold.
- The dominant error becomes essentially one-dimensional, simplifying the outer code.
This bias reshapes the whole error-correction stack. Instead of correcting X and Z symmetrically, one can pick an outer code tuned for a single dominant error type. A repetition code against phase flips, layered on bit-flip-protected cat qubits, needs far fewer qubits than an unbiased surface code for the same protection, because it only has to fight one error.
The catch is that the bias must survive the gates. A two-qubit gate that reintroduces bit flips as a side effect would destroy the advantage, so bias-preserving gate design is central to the cat-code program, and is the subject of active experimental work on superconducting cavities.