Cat Qubits 101

Updated 28 October 2026.

Introduction

A cat qubit stores quantum information in specially controlled states of an oscillator. In superconducting implementations, that oscillator is a microwave resonator. Instead of using just its lowest two energy levels, the encoding combines many photon-number states to create two distinguishable oscillations and their quantum superpositions. Those superpositions are related to Schrödinger’s famous cat thought experiment.

The attraction is a favorable imbalance between errors. With suitable stabilization, changing one encoded computational state into the other can become exceptionally unlikely. Disturbing the relative phase between them remains much easier. A quantum computer can exploit this imbalance, called noise bias, by concentrating much of its error correction on the more common faults.

This article focuses on the two-component, biased-noise cat qubit used in many current superconducting proposals. The broader family of cat codes also includes other encodings with different error-correction properties. The distinction matters: a two-component cat, a four-component photon-loss code and a logical qubit built from several cats are different objects.

The essential promise is that deliberately engineered physics can reduce the overhead of reliable computation. A cat is still a noisy qubit, and a long lifetime against one error does not mean that all of its quantum information survives for that duration. The useful machine emerges only when the encoding, stabilization, gates and error-correcting code work together.

How Cat Qubits Work

Two Oscillations and Their Superpositions

Imagine two microwave oscillations with the same amplitude but opposite phase: when one field points in one direction, the other points in the opposite direction. Quantum mechanically, the corresponding coherent states are written |α⟩ and |−α⟩. The complex number α describes the field amplitude and phase. It is not a photon count; each coherent component has an average of |α|² photons.

The two states appear as separated lobes in phase space, a plot of the oscillator’s two field quadratures. Those quadratures are analogous to position and momentum. An ordinary mixture would mean that the oscillator is in one lobe or the other with some classical uncertainty. A quantum superposition retains a relative amplitude and phase between both possibilities. Interference between the components is what makes the encoding quantum.

For this article, the computational states are approximately |0_c⟩ ≈ |α⟩ and |1_c⟩ ≈ |−α⟩. The subscript c labels the qubit encoded in one oscillator. The approximation is necessary because coherent states have a small overlap at finite amplitude; the mathematical section below gives an exact orthonormal definition.

Two other useful states are the even and odd cats, proportional to |α⟩ + |−α⟩ and |α⟩ − |−α⟩. In our convention, these are the qubit’s plus and minus states, |+_c⟩ and |−_c⟩. They contain only even or only odd photon numbers, respectively. They do not turn into individual coherent states when α becomes large: each remains a superposition of both components.

Which Error Is a Bit Flip?

A bit flip, or X error, exchanges the two computational states. A phase flip, or Z error, leaves those basis states in place but changes their relative sign in a superposition. For example, it turns the plus cat into the minus cat. Both types of error are quantum-information errors, but the physical processes that produce them can be very different.

These names depend on the chosen basis. Some papers define the even and odd cats themselves as computational zero and one. In that convention, the X and Z labels are exchanged. The underlying physics has not changed. Keeping one convention throughout an explanation prevents a photon-loss event from being called a bit flip in one paragraph and a phase flip in the next.

In the coherent-state convention used here, single-photon loss produces predominantly a phase error. The photon-loss operator acts with opposite signs on |α⟩ and |−α⟩, changing their relative phase. It does not normally swap the lobes. With suitable stabilization, the unwanted lobe-switching process can be strongly suppressed while photon loss remains a substantial source of phase noise. This is the basis of repetition-cat fault-tolerance proposals.

Why Stabilization Is Essential

Separating two states in phase space is not enough by itself. An ordinary, undriven resonator loses energy and its coherent amplitude shrinks. Protection requires engineered dynamics that keep the relevant two-dimensional space available while quantum information is stored or manipulated.

Dissipative stabilization combines coherent two-photon driving with an engineered process that removes photons in pairs. A lossy buffer mode can carry away the engineered dissipation. The combined process confines the storage oscillator to the cat manifold, which contains both coherent states and all their superpositions. Mirrahimi et al.’s 2014 paper explains this construction and its distinction from four-component cat codes.

Confinement does not automatically correct the relative-sign change caused by losing one photon. A phase error can occur entirely within the protected space. This is why saying that a cat “returns to its protected state” is insufficient: the oscillator may return to the allowed manifold while the encoded information still carries an error.

Kerr-cat confinement uses a nonlinear oscillator and a two-photon drive to engineer a protected Hamiltonian structure. Grimm et al.’s 2020 experiment demonstrated this approach and single-qubit control. Hamiltonian confinement is different from an engineered dissipative process that attracts a state back into a manifold. Real hardware may combine both kinds of technique.

For appropriate noise processes, increasing the cat size improves bit-flip protection exponentially in the photon-number parameter |α|². The phase-error rate from ordinary photon loss, however, generally increases with that parameter. Hardware imperfections can also impose a floor on bit-flip suppression. The useful operating point therefore balances both errors with the correction cycle and gate schedule.

Comparison with Transmon Qubits

A transmon is also an oscillator, but one made sufficiently anharmonic that its lowest two energy levels can be addressed as a qubit. A cat uses a different two-dimensional encoding within a larger oscillator state space. Neither description means that every photon is a separate qubit or that a cat stores an unlimited amount of information.

It is misleading to describe all transmon errors as symmetric. Energy relaxation is also not literally an isolated Pauli X: amplitude damping is a different quantum channel. Likewise, a transmon’s T₁ energy-relaxation time, its T₂ coherence time and a cat’s bit-flip lifetime are not interchangeable measures. T₂ includes the effects of relaxation as well as pure dephasing.

The most useful comparison is the logical reliability achieved for the same task and total resources. A cat may need more circuitry per data qubit yet fewer data qubits for a particular error-corrected workload. The outcome depends on measured operations and architectural assumptions; it does not follow from a single lifetime record or a universal physical-to-logical-qubit ratio.

Controlling and Reading a Cat

In our basis convention, a weak resonant drive can generate a Z rotation. A full X operation can exchange the lobes by moving them around phase space with the stabilization appropriately controlled. The lobes need not merge at the origin. Other rotations can use different control sequences, sometimes temporarily relaxing the protection. Work on Zeno gates examines how dissipative-cat control can balance speed and errors.

For multi-qubit operations such as a CNOT, preserving noise bias means preventing common phase faults from becoming substantial bit-flip faults. Cat–cat gates and cat–transmon gates are different implementations, and a transmon mediator is not required by every cat architecture. Experiments must also characterize correlated faults and leakage.

A universal computer needs more than a set of protected memory operations. A Hadamard exchanges X and Z, illustrating why every native unitary in a universal gate set cannot preserve the same dephasing bias. Bias-preserving gate proposals therefore need to be understood within complete logical protocols, with encoded operations, measurements and suitable resource-state preparation. A proposed Toffoli construction is not evidence of a demonstrated fault-tolerant universal processor.

Readout is also basis-dependent. Measuring which coherent lobe is occupied probes the computational basis. Measuring photon-number parity probes the plus/minus basis. Either can be useful for characterization or intentional qubit measurement. Whether a measurement is a safe error syndrome depends on how the information has been encoded.

Mathematical Background

Coherent-State Overlap

Let a be the oscillator’s photon-annihilation operator. A coherent state satisfies a|α⟩ = α|α⟩. The two opposite coherent states have overlap:

⟨α|−α⟩ = e^−2|α|².

For α = 2, each coherent component has mean photon number four, and the overlap amplitude is e⁻⁸ ≈ 3.35 × 10⁻⁴. The states are nearly orthogonal, but not exactly orthogonal. This overlap is an amplitude, not directly an experimentally measured bit-flip probability. Actual protection depends on the noise operators and stabilization.

Even, Odd and Computational Bases

The normalized even and odd cats are:

|C_α⁺⟩ = N₊(|α⟩ + |−α⟩),
|C_α⁻⟩ = N₋(|α⟩ − |−α⟩),
N_± = [2(1 ± e^−2|α|²)]^−1/2.

For nonzero α, these cats are exactly orthogonal: one contains only even photon-number states, the other only odd ones. Their normalization factors differ at finite amplitude. The mean photon number of each cat is therefore not exactly |α|² in general, although that approximation becomes good for large cats.

An exact orthonormal computational basis that approaches the coherent lobes is:

|0_c⟩ = (|C_α⁺⟩ + |C_α⁻⟩)/√2,
|1_c⟩ = (|C_α⁺⟩ − |C_α⁻⟩)/√2.

When the coherent-state overlap is small, these approach |α⟩ and |−α⟩, respectively. Conversely, the exact plus and minus states of this computational basis are |+_c⟩ = |C_α⁺⟩ and |−_c⟩ = |C_α⁻⟩. It is the sum and difference of even and odd cats that approach the individual lobes.

Photon Loss in the Encoded Qubit

Because a|α⟩ = α|α⟩ and a|−α⟩ = −α|−α⟩, a photon-loss jump changes the relative sign in a coherent-state superposition. Up to an overall factor and normalization:

a(c₀|α⟩ + c₁|−α⟩) = α(c₀|α⟩ − c₁|−α⟩).

Within the exact orthonormal basis, the action approaches αZ_c as the overlap becomes small. At finite amplitude, the exact action on the parity states is:

a|C_α⁺⟩ = α(N₊/N₋)|C_α⁻⟩,
a|C_α⁻⟩ = α(N₋/N₊)|C_α⁺⟩.

The loss flips parity but remains within the ideal two-component manifold. It is approximately a Z error in our computational basis, and would be approximately an X error if even and odd cats were chosen as computational zero and one. The finite-amplitude factors explain why the exact channel is not simply a perfect Pauli operation in every regime.

Two loss jumps satisfy a²|±α⟩ = α²|±α⟩. They therefore do not generically exchange the lobes; within this ideal manifold, a² acts as the same scalar on both components. Real evolution also includes what happens between jumps and the device’s other noise channels. Photon gain, described by a†, is different from photon loss and can create leakage outside the encoded space.

If κ₁ denotes the single-photon energy-loss rate, the large-cat phase-flip jump rate scales approximately as κ₁|α|². A quoted phase-coherence decay time can differ by a factor of two from the inverse jump rate, depending on the convention. Rates, correlation-decay times and mean waiting times between jumps should therefore be compared only after checking their definitions.

The Engineered Protection Operator

An ideal dissipative-cat model uses the engineered jump operator:

L₂ = √κ₂(a² − α²).

Here κ₂ sets the strength of the engineered process. Since L₂|α⟩ = L₂|−α⟩ = 0, their span is a dark manifold: any superposition within it is left undisturbed by that ideal process. This is a useful mathematical description of confinement, but it does not imply that ordinary single-photon loss has disappeared. Lecture notes on quantum computing with cat qubits develop the model and its limitations.

The name of a physical noise process also needs care. Oscillator dephasing, generated by fluctuations associated with photon number, is not automatically the same thing as a Z error on the encoded qubit. Encoding changes the effective action of physical noise. One should identify the induced logical operation, rather than carrying error labels unchanged from the oscillator to the qubit.

Why Parity Alone Does Not Correct This Qubit

The photon-number parity operator is P = e^iπa†a. Its eigenvalues distinguish even and odd photon numbers:

P|C_α⁺⟩ = +|C_α⁺⟩,
P|C_α⁻⟩ = −|C_α⁻⟩.

Within our two-component encoding, parity is the qubit’s X_c operator. It is not a stabilizer of the whole codespace, because states in that space can have either parity. Measuring a single cat’s parity therefore measures an encoded qubit observable. On an arbitrary unknown state, that can reveal and disturb the information we wanted to protect.

Parity monitoring does provide a useful error syndrome in other encodings. In a four-component photon-loss-correcting cat code, both logical basis states can occupy the same parity sector. A photon loss then moves them to an error sector with opposite parity, allowing detection without revealing the logical amplitudes. This is the relevant distinction when reading about Ofek et al.’s 2016 bosonic error-correction experiment.

Correcting Phase Flips with Several Cats

For the two-component encoding, an outer phase-flip repetition code can supply the missing redundancy. For three cats, one possible encoded state is:

|ψ_L⟩ = A|+_c+_c+_c⟩ + B|−_c−_c−_c⟩.

This is one quantum state distributed across three oscillators, not three independent copies of an unknown qubit. Measuring the joint checks X₁X₂ and X₂X₃ identifies disagreements while preserving the amplitudes A and B. A Z error on the middle cat changes the sign of both checks, providing a syndrome that distinguishes it from a phase error on either end.

With ideal measurements, a phase repetition code containing d cats, for odd d, can correct up to (d − 1)/2 phase flips. Noisy measurement requires repeated checks and a decoder that interprets their history. Ancillas may be cats, transmons or other suitable systems, depending on the implementation.

The code does not correct an arbitrary physical bit flip. Indeed, an X on one cat can change the relative sign between the two terms in |ψ_L⟩ without changing those checks. That is why finite bias eventually limits a simple repetition code. Larger or more demanding machines may use additional coding, such as asymmetric surface codes or Elevator Codes, to address the residual error channel.

Origins and Milestones

Cat-qubit ideas developed through several strands of quantum optics, circuit QED and quantum error correction. There is no need to assign all of those advances to one “first” paper.

Early cat-state encoding. Cochrane, Milburn and Munro’s 1999 paper proposed cat-state encodings for correcting amplitude-damping errors. Jeong and Kim’s 2002 work developed quantum-computing ideas using coherent-state superpositions in optical systems. Oscillator-based quantum information therefore predates the recent superconducting implementations.

Hardware-efficient circuit-QED protection. Leghtas et al.’s proposal, submitted in 2012 and published in 2013, described a practical route to protecting a cavity-encoded qubit using the nonlinearity of a coupled ancillary qubit. The 2014 Mirrahimi et al. paper then developed dynamically protected encodings and operations using multi-photon driven-dissipative processes.

Break-even bosonic error correction. In 2016, Ofek et al. demonstrated a corrected four-component cat-code lifetime of 320 microseconds, exceeding the uncorrected |0⟩/|1⟩ Fock-state encoding in the same oscillator by about 10%. It was a major error-correction milestone, using a different encoding from the two-component biased cat explained above.

Biased cats as a computing architecture. Guillaud and Mirrahimi’s 2019 repetition-cat proposal, Grimm et al.’s 2020 Kerr-cat experiment, and work on protected gates helped establish the modern architecture: suppress one error channel in the oscillator, then correct the remaining faults with a suitable multi-qubit code.

What Has Been Demonstrated by October 2026?

There is now experimental evidence beyond isolated cat memories. AWS’s Ocelot work demonstrated repeated correction in a five-cat logical memory with transmon ancillas. Its below-threshold result concerns the phase-correcting layer; finite bit errors still limit the complete memory. The companion architecture article explains the measurements and resource accounting in detail.

On 20 October 2026, Alice & Bob reported cat–cat CNOTs, repeated distance-3 correction and a non-Clifford operation across experimental demonstrators. The company also stated that the logical error rate was not yet below threshold. Demonstrating the building blocks is progress; their combined logical performance remains the decisive test.

The squeezing-inspired cat experiment published on 22 October 2026 gives another useful lesson. It increased the bit-flip time from 138 milliseconds to 22 seconds while comparing operating points with the same 1.3-microsecond phase-flip time. Reshaping the protection can therefore improve a meaningful tradeoff, while leaving a clear need for phase-error correction.

More generally, long idle lifetimes must be supplemented by gate measurements. A 2026 detuned Kerr-cat benchmarking study found that the noise bias inferred from lifetime measurements alone could overestimate the bias during operations. This is why an architecture cannot be validated simply by dividing its longest and shortest reported lifetimes.

Challenges and Future Prospects

Correcting phase errors quickly enough. Phase faults accumulate while the qubits wait, interact and undergo syndrome extraction. Useful logical performance depends on the entire correction cycle, including measurements, resets and classical decoding. There is no single logical-error target that suffices for every application: longer circuits generally require lower error per operation.

Keeping the bias during operations. The noise distribution of a driven gate can differ substantially from that of an idle qubit. Leakage and correlations can also defeat an otherwise attractive code. Experiments need to establish the error channels during the operations that the actual logical architecture will use.

Managing rare bit errors. Exponential suppression does not mean that the residual rate is zero. At sufficiently large scale or long runtime, even rare faults matter. The 2026 Elevator Codes proposal, for example, explicitly addresses phase and bit errors in different coding layers. It is an architectural study rather than a completed hardware demonstration.

Building universal logical operations. Error-corrected memory is only one ingredient. A useful processor needs reliable logical gates, including non-Clifford resources. Unfolded distillation, published in 2026, proposes a way to reduce magic-state costs under an assumed biased-noise model. Such modeled savings must still be implemented and validated with physical devices.

Scaling the hardware and its control. Storage-mode quality, pump-induced imperfections, thermal excitations, nonlinear couplings, fabrication variation and crosstalk can all affect protection. Large registers also require calibration and decoding that operate reliably alongside the quantum hardware. A few exceptionally stable cats do not establish uniform performance across a processor.

Cat qubits provide a concrete way to move part of the error-correction burden into the physics of an oscillator. The strongest evidence of success will be progressively more reliable logical operations at a defensible total cost, with every relevant error channel included. That is what connects impressive individual cat states to a practical fault-tolerant quantum computer.

From the author