Dissipative Cat Qubits: Noise Bias & Stabilization

• Dissipative cat qubits are bosonic qubits encoded in Schrödinger-cat superpositions and autonomously stabilized by engineered two-photon (or multi-photon) dissipation.
• They employ a Lindblad equation framework where controlled dissipation confines the system to a logical manifold, yielding exponentially suppressed bit-flip errors with increasing cat size.
• Architectural implementations in circuit QED and autoparametric designs illustrate how engineered dissipation and noise bias enable robust quantum control and fault-tolerant operations.

Searching arXiv for recent and foundational papers on dissipative cat qubits and related developments. Dissipative cat qubits are bosonic qubits encoded in Schrödinger-cat superpositions of a single harmonic mode and autonomously stabilized by engineered multi-photon dissipation, most commonly a two-photon process that makes the manifold spanned by $|\pm\alpha\rangle$ or, equivalently, by even and odd cat states, the asymptotically stable logical space. In this setting dissipation is both a noise source and a control resource: it confines the oscillator to a protected manifold, produces a pronounced noise bias with exponentially suppressed logical bit flips as the cat size grows, and supports bias-preserving control protocols relevant to fault-tolerant quantum computation [(Mirrahimi et al., 2013); (Putterman et al., 2024)].

1. Physical realization and open-system description

The canonical reduced description is a Lindblad equation for a single bosonic mode $a$ with an engineered two-photon dissipator,

$$\dot\rho=\kappa_2\,\mathcal D[a^2-\alpha^2]\rho, \qquad \mathcal D[L]\rho=L\rho L^\dagger-\tfrac12\{L^\dagger L,\rho\},$$

for which $|\pm\alpha\rangle$ are dark states because $(a^2-\alpha^2)|\pm\alpha\rangle=0$. A more explicit driven-dissipative form, used for phase-sensitive control and spectral analysis, is

$$\dot{\rho}=-i[H,\rho]+\kappa\,\mathcal D[a]\rho+\kappa_2\,\mathcal D[a^2]\rho,$$

with

$$H=\Delta a^\dagger a+(\varepsilon_2 a^{\dagger2}+\varepsilon_2^*a^2)+\varepsilon(a^\dagger+a),$$

where $\Delta$ is the detuning, $\varepsilon_2=|\varepsilon_2|e^{-i\theta}$ is the two-photon drive amplitude, and $\varepsilon$ is a single-photon drive (Yang et al., 22 Apr 2026).

In circuit QED this effective single-mode description is commonly realized with a two-mode “memory + buffer” architecture. A high-Q memory mode $a$0 stores the logical information, while a lossy auxiliary mode $a$1 mediates reservoir engineering through a three-wave-mixing interaction such as

$a$2

or, in a driven form,

$a$3

Adiabatic elimination of the strongly damped buffer yields an effective two-photon drive and two-photon loss on the memory, with $a$4 in the standard elimination regime (Yang et al., 22 Apr 2026, Ferrari et al., 1 Jul 2026).

Several hardware realizations of this paradigm have been analyzed. One route uses a driven lossy buffer with colored dissipation and a multi-pole filter to preserve the memory’s phase coherence and linearity while maintaining strong two-photon stabilization; another uses an autoparametric architecture in which a buffer at frequency $a$5 passively generates strong two-photon exchange without a separate pump tone (Putterman et al., 2024, Marquet et al., 2023). These implementations differ in circuit details but share the same operational principle: a memory mode with weak natural loss is continuously pulled toward a cat-code manifold by an engineered nonlinear reservoir.

2. Logical encoding, parity structure, and stabilized manifolds

The basic cat states are

$a$6

with $a$7 supported on even Fock states and $a$8 on odd Fock states. Two common logical conventions coexist in the literature. In the parity basis one identifies

$a$9

whereas in the coherent-state basis one uses

$$\dot\rho=\kappa_2\,\mathcal D[a^2-\alpha^2]\rho, \qquad \mathcal D[L]\rho=L\rho L^\dagger-\tfrac12\{L^\dagger L,\rho\},$$0

up to $$\dot\rho=\kappa_2\,\mathcal D[a^2-\alpha^2]\rho, \qquad \mathcal D[L]\rho=L\rho L^\dagger-\tfrac12\{L^\dagger L,\rho\},$$1 corrections (Yang et al., 22 Apr 2026, Guillaud et al., 2022).

Two-photon processes preserve photon-number parity, so in the ideal stabilization limit an initial even state relaxes to the even cat and an initial odd state relaxes to the odd cat. In the strong-stabilization regime $$\dot\rho=\kappa_2\,\mathcal D[a^2-\alpha^2]\rho, \qquad \mathcal D[L]\rho=L\rho L^\dagger-\tfrac12\{L^\dagger L,\rho\},$$2, the oscillator is dynamically confined to the logical manifold with confinement rate

$$\dot\rho=\kappa_2\,\mathcal D[a^2-\alpha^2]\rho, \qquad \mathcal D[L]\rho=L\rho L^\dagger-\tfrac12\{L^\dagger L,\rho\},$$3

and single-photon processes then act as weak perturbations within that manifold (Yang et al., 22 Apr 2026).

Within the cat subspace, annihilation approximately exchanges even and odd cats. In the even/odd encoding, a single-photon jump therefore flips parity; in the coherent-state encoding, the same process is interpreted differently at the logical level, because the computational states are the two phase-space wells rather than the parity eigenstates. This basis dependence is central to the error-bias discussion and explains why different papers label the dominant error channel as logical $$\dot\rho=\kappa_2\,\mathcal D[a^2-\alpha^2]\rho, \qquad \mathcal D[L]\rho=L\rho L^\dagger-\tfrac12\{L^\dagger L,\rho\},$$4 or logical $$\dot\rho=\kappa_2\,\mathcal D[a^2-\alpha^2]\rho, \qquad \mathcal D[L]\rho=L\rho L^\dagger-\tfrac12\{L^\dagger L,\rho\},$$5 without contradiction (Yang et al., 22 Apr 2026, Guillaud et al., 2022).

The stabilized manifold can also be deformed. A dissipatively stabilized squeezed cat qubit replaces the physical mode $$\dot\rho=\kappa_2\,\mathcal D[a^2-\alpha^2]\rho, \qquad \mathcal D[L]\rho=L\rho L^\dagger-\tfrac12\{L^\dagger L,\rho\},$$6 by the squeezed-mode operator

$$\dot\rho=\kappa_2\,\mathcal D[a^2-\alpha^2]\rho, \qquad \mathcal D[L]\rho=L\rho L^\dagger-\tfrac12\{L^\dagger L,\rho\},$$7

and stabilizes the manifold with

$$\dot\rho=\kappa_2\,\mathcal D[a^2-\alpha^2]\rho, \qquad \mathcal D[L]\rho=L\rho L^\dagger-\tfrac12\{L^\dagger L,\rho\},$$8

This produces squeezed coherent components $$\dot\rho=\kappa_2\,\mathcal D[a^2-\alpha^2]\rho, \qquad \mathcal D[L]\rho=L\rho L^\dagger-\tfrac12\{L^\dagger L,\rho\},$$9 and squeezed cat states $|\pm\alpha\rangle$0, retaining parity structure while modifying phase-space geometry and error bias (Hillmann et al., 2022).

3. Noise bias, logical errors, and switching theory

The defining operational advantage of dissipative cat qubits is biased noise. In the coherent-state basis used by several works, single-photon loss predominantly produces logical phase flips, while logical bit flips correspond to rare transitions between the two wells $|\pm\alpha\rangle$1. The resulting bit-flip rate is exponentially suppressed with the cat size, whereas the phase-flip rate grows roughly linearly with $|\pm\alpha\rangle$2 or $|\pm\alpha\rangle$3 (Guillaud et al., 2022, Putterman et al., 2024).

For standard dissipative cats, the phase-flip rate under single-photon loss and gain can be written as

$|\pm\alpha\rangle$4

which is linear in $|\pm\alpha\rangle$5 at large photon number (Rousseau et al., 11 Feb 2025). By contrast, in the hidden-time-reversal-symmetric regime analyzed with Keldysh path integrals, the non-perturbative switching theory yields

$|\pm\alpha\rangle$6

with the bit flip identified as inter-basin switching between two metastable attractors of the Lindbladian flow. The same framework shows that dephasing breaks the hidden time-reversal symmetry and invalidates the time-reversed instanton construction (Carde et al., 24 Jul 2025).

The literature also emphasizes that the cat-qubit error model is not purely abstract but tied to physical reservoir engineering. In dissipative stabilization, single-photon loss is the dominant unwanted channel, while the engineered two-photon process continuously restores the state to the logical manifold. This is why dissipative cats are often described as noise-biased or autonomously error-corrected bosonic qubits: the reservoir does not eliminate all errors, but it reshapes them into a form that is strongly asymmetric and therefore favorable for concatenation with outer codes [(Mirrahimi et al., 2013); (Putterman et al., 2024)].

Squeezing changes this bias quantitatively. For dissipatively stabilized squeezed cat qubits, moderate squeezing reduces the bit-flip rate significantly relative to the ordinary cat while leaving the phase-flip rate essentially unchanged. In the experimental “moon cat” implementation, the bit-flip scaling exponent reached $|\pm\alpha\rangle$7, corresponding to a factor of 74 improvement per added photon, and a squeezed cat with $|\pm\alpha\rangle$8 achieved a bit-flip time of 22 seconds for a phase-flip time of $|\pm\alpha\rangle$9s, a 160-fold improvement over a standard cat at the same phase-flip time (Hillmann et al., 2022, Rousseau et al., 11 Feb 2025).

4. Control, gates, and architectural integration

A central theme since the original dynamically protected cat-qubit proposal is that logical control should not reintroduce the very bit flips that the cat encoding suppresses. In the two-photon dissipative scheme, weak resonant drives implement $(a^2-\alpha^2)|\pm\alpha\rangle=0$0-rotations through quantum Zeno projection, beam-splitter-type couplings generate two-qubit $(a^2-\alpha^2)|\pm\alpha\rangle=0$1 interactions, and Kerr evolution provides a logical $(a^2-\alpha^2)|\pm\alpha\rangle=0$2 $(a^2-\alpha^2)|\pm\alpha\rangle=0$3-rotation; analogous constructions exist for four-photon cat encodings (Mirrahimi et al., 2013).

More recent dissipative implementations preserve the same principle while changing the hardware layer. In the autoparametric cat qubit, an additional drive on the memory mode changes the phase of a superposition between $(a^2-\alpha^2)|\pm\alpha\rangle=0$4 and $(a^2-\alpha^2)|\pm\alpha\rangle=0$5 while the engineered dissipation remains active. In that platform the passive two-photon process reached $(a^2-\alpha^2)|\pm\alpha\rangle=0$6, and bit-flip errors were suppressed up to $(a^2-\alpha^2)|\pm\alpha\rangle=0$7 with only a mild impact on phase-flip errors (Marquet et al., 2023).

Strong stabilization is also compatible with pulsed operation. A planar architecture with a filtered lossy buffer and optimized ATS buffer design showed that stabilization can remain off for a significant fraction of a $(a^2-\alpha^2)|\pm\alpha\rangle=0$8 cycle without degrading bit-flip times; specifically, the stabilization can remain off for two thirds of the cycle. In the same system, near-ideal enhancement of bit-flip times with photon number was observed, with bit-flip times exceeding $(a^2-\alpha^2)|\pm\alpha\rangle=0$9 s at $$\dot{\rho}=-i[H,\rho]+\kappa\,\mathcal D[a]\rho+\kappa_2\,\mathcal D[a^2]\rho,$$0, while the effective phase-flip lifetime remained $$\dot{\rho}=-i[H,\rho]+\kappa\,\mathcal D[a]\rho+\kappa_2\,\mathcal D[a^2]\rho,$$1s, comparable to the bare oscillator lifetime (Putterman et al., 2024).

At the architecture level, dissipative cats can serve as data qubits while transmons act as syndrome ancillas. A hybrid cat-transmon proposal introduces a transmon-controlled CX and a cat-controlled rotation gate enabling a thin rectangular surface code in which the large cat-qubit bias is exploited by choosing $$\dot{\rho}=-i[H,\rho]+\kappa\,\mathcal D[a]\rho+\kappa_2\,\mathcal D[a^2]\rho,$$2. In that setting, current state-of-the-art coherence supports physical error rates of $$\dot{\rho}=-i[H,\rho]+\kappa\,\mathcal D[a]\rho+\kappa_2\,\mathcal D[a^2]\rho,$$3 and noise biases in the range $$\dot{\rho}=-i[H,\rho]+\kappa\,\mathcal D[a]\rho+\kappa_2\,\mathcal D[a^2]\rho,$$4; with this performance, the qubit overhead needed to reach algorithmically relevant logical error rates matches that of an unbiased-noise architecture with physical error rates in the range $$\dot{\rho}=-i[H,\rho]+\kappa\,\mathcal D[a]\rho+\kappa_2\,\mathcal D[a^2]\rho,$$5 (Hann et al., 2024).

5. Liouvillian topology, erasure signatures, and saturation phenomena

Recent work has extended dissipative cat-qubit theory beyond error-bias phenomenology to the full non-Hermitian spectral structure of the Liouvillian. For a cat qubit stabilized by two-photon drive and engineered two-photon loss, and perturbed by single-photon drive and single-photon loss, the Liouvillian spectrum in the $$\dot{\rho}=-i[H,\rho]+\kappa\,\mathcal D[a]\rho+\kappa_2\,\mathcal D[a^2]\rho,$$6 plane exhibits both second- and third-order Liouvillian exceptional points, LEP2s and LEP3s. The phase $$\dot{\rho}=-i[H,\rho]+\kappa\,\mathcal D[a]\rho+\kappa_2\,\mathcal D[a^2]\rho,$$7 of the two-photon drive acts as a coherent control knob: the LEP3 remains finite and tunable for $$\dot{\rho}=-i[H,\rho]+\kappa\,\mathcal D[a]\rho+\kappa_2\,\mathcal D[a^2]\rho,$$8, but diverges and vanishes at $$\dot{\rho}=-i[H,\rho]+\kappa\,\mathcal D[a]\rho+\kappa_2\,\mathcal D[a^2]\rho,$$9. A topological invariant based on the winding number of a resultant vector assigns unit topological charge to the LEP3, and full master-equation simulations show that the dynamics remains confined to the logical subspace with fidelity above $$H=\Delta a^\dagger a+(\varepsilon_2 a^{\dagger2}+\varepsilon_2^*a^2)+\varepsilon(a^\dagger+a),$$0 for representative experimental parameters (Yang et al., 22 Apr 2026).

Another recent development concerns the observability of logical failures. In a dissipatively stabilized memory-buffer system, logical bit flips are accompanied by a strong, time-localized burst of photons emitted by the dissipative buffer. Using quantum trajectories, past quantum states, and number-resolved master equations, one can show that bit flips are therefore not silent logical errors but erasures: photon counting and homodyne monitoring of the output line can herald the loss of logical information without interrupting the autonomous stabilization (Ferrari et al., 1 Jul 2026).

A complementary line of work addresses the empirically observed saturation of bit-flip suppression at large cat size. In a realistic memory-buffer model, exponential suppression fails when two conditions are met: the adiabatic approximation must break down, and nonlinear terms such as cross-Kerr together with dephasing must be present. In that regime a reflection symmetry and associated phase-locking condition are lost, memory fluctuations are amplified by the buffer during bit flips, and the switching events appear as chaotic bursts embedded in otherwise regular dynamics. The saturation is therefore not captured by single-mode effective models and reflects intrinsically two-mode dissipative dynamics (Ferrari et al., 22 May 2026).

6. Limitations, variants, and long-term prospects

Dissipative cat-qubit analyses almost always rely on some controlled approximation, and the limits of those approximations are now comparatively well understood. Projection to the cat manifold is accurate only when the engineered confinement dominates leakage channels; large detunings, residual Kerr terms, or nonadiabatic memory-buffer coupling can spoil that picture. For dissipatively stabilized squeezed cats, good performance requires residual Kerr satisfying $$H=\Delta a^\dagger a+(\varepsilon_2 a^{\dagger2}+\varepsilon_2^*a^2)+\varepsilon(a^\dagger+a),$$1, and the required pump engineering is more involved than in unsqueezed devices (Hillmann et al., 2022).

The theory of switching rates also identifies a structural limitation: hidden time-reversal symmetry underlies the exact instanton construction for a broad class of single-mode dissipative cat models, but dephasing and other symmetry-breaking dissipators invalidate that simplification and alter the bit-flip exponent (Carde et al., 24 Jul 2025). At a more microscopic level, quasiparticle tunneling in driven superconducting cat-qubit circuits induces effective single-photon loss, single-photon gain, and dephasing channels. A microscopic derivation shows that present cat-qubit experiments are not yet limited by quasiparticles, but that drive-assisted quasiparticle excitation will become relevant as other noise sources are reduced (Dubovitskii et al., 2024).

Several variants broaden the scope of the subject. Kerr-cat qubits emphasize Hamiltonian confinement and fast gates; hybrid “critical cat” schemes combine Kerr nonlinearity, two-photon loss, and detuned two-photon drive to exploit a first-order dissipative phase transition and a metastable cat manifold with enhanced bit-flip suppression over broad detuning windows (Gravina et al., 2022). Dissipative information-processing proposals use stabilized Kerr cats as repeated-interaction elements for quantum classification and homogenization tasks, illustrating that the same reservoir-engineering machinery can support both protected storage and dissipative computation (Korkmaz et al., 2023).

The long-term prospect is therefore not a single device recipe but a family of bosonic qubits defined by the same organizing principle: the logical manifold is encoded in nonclassical phase-space structure and protected by an engineered open-system dynamics. Within that family, current research has already connected autonomous stabilization, non-perturbative switching theory, squeezed-manifold engineering, Liouvillian topology, heralded erasures, pulsed control, and hardware-efficient outer codes. This suggests that dissipative cat qubits are best understood not merely as long-lived oscillator memories, but as a programmable class of open bosonic qubits whose logical behavior is set jointly by phase-space geometry and Liouvillian design (Yang et al., 22 Apr 2026, Hann et al., 2024).