Bosonic Qubits

A bosonic qubit stores a finite-dimensional logical system in the Hilbert space of one or more harmonic-oscillator modes. The mode may be a microwave cavity field, an optical field, trapped-ion motion, an acoustic resonance, or another approximately bosonic degree of freedom. Its many levels supply redundancy inside each mode: photon loss, gain, dephasing, or small phase-space displacements can move a state into a distinguishable error sector without immediately revealing the logical information.

That compact statement can be misleading in two opposite directions. One oscillator can replace many two-level data carriers at the encoding layer, but a working bosonic module is not one passive resonator. It commonly needs a nonlinear ancilla, pumps, couplers, a readout channel, reset or engineered dissipation, and real-time control. Conversely, the ancillary system need not be more coherent than the storage mode if the control protocol detects, confines, or prevents propagation of its faults.

This page explains the physical implementation of oscillator encodings: storage, nonlinear control, measurement, and entropy removal. Bosonic Codes develops the correction conditions, while the cat, binomial, and GKP articles derive the individual codes. A bosonic qubit selects a finite logical space inside an oscillator; Continuous-Variable Platforms also considers computation that retains quadratures as continuous variables. Experimental milestones and open problems are discussed in Bosonic Codes: Research Directions.

Required background. Bits, Qubits, Qudits, and Modes distinguishes a carrier from its encoding. Number States supplies ladder operators, and Quantum Channels and Noise introduces damping and completely positive maps.

From an Oscillator to a Protected Carrier

The ideal mode and its useful imperfections

An ideal mode has Hamiltonian

$$H_0 = \hbar\omega \left( a^\dagger a+\frac{1}{2} \right), \qquad [a,a^\dagger]=1.$$

Its equally spaced spectrum is excellent for long-lived storage and poor for selective control: a resonant linear drive addresses every neighboring Fock transition at the same frequency. Bosonic hardware therefore combines two seemingly incompatible ingredients:

• a storage mode made as harmonic and isolated as possible;
• a controllable nonlinearity that distinguishes states or conditions operations on the oscillator.

In circuit QED, a dispersively coupled transmon can provide an effective Hamiltonian of the form

$$\frac{H}{\hbar} ={} \omega_c a^\dagger a +\frac{\omega_q}{2}\sigma_z -\chi a^\dagger a\,|e\rangle\langle e| -\frac{K}{2}a^{\dagger 2}a^2+\cdots .$$

Here $n=a^\dagger a$, $\sigma_z=|e\rangle\langle e|-|g\rangle\langle g|$, and $\omega_c,\omega_q,\chi,K$ have angular-frequency units. The signed coefficient $\chi$ gives the ancilla transition $\omega_q(n)=\omega_q-\chi n$. The dispersive approximation requires sufficiently small mode–ancilla hybridization over the occupied number range; large occupation or strong drives can invalidate it. The shift makes the transition number selective. The inherited self-Kerr $K$ can be useful for control, but it also shears phase-space states during idle periods. Higher-order terms, drive-induced shifts, and ancilla-state dependence matter once a state spans many number levels.

The architectural problem is to turn the nonlinearity on where useful without letting the nonlinear element spoil the storage mode continuously.

Loss, gain, and phase diffusion

A common reduced model is

$$\dot\rho={}-\frac{i}{\hbar}[H,\rho] +\kappa_\downarrow\mathcal D[a]\rho +\kappa_\uparrow\mathcal D[a^\dagger]\rho +\kappa_\phi\mathcal D[a^\dagger a]\rho +\sum_j\gamma_j\mathcal D[L_j^{\rm anc}]\rho ,$$

with

$$\mathcal D[L]\rho = L\rho L^\dagger -\frac{1}{2} \left\{ L^\dagger L,\rho \right\}.$$

The rates describe oscillator loss, thermal gain, oscillator dephasing, and ancilla faults. This model is a starting point, not a complete device description. Pump-induced heating, quasiparticle bursts, telegraph switching, mode collisions, correlated ancilla–cavity dephasing, reset transients, and slow parameter drift can violate its stationary and Markovian assumptions.

With no drive and a Hamiltonian commuting with $n$, this convention gives

$$\frac{d\langle n\rangle}{dt} =-(\kappa_\downarrow-\kappa_\uparrow)\langle n\rangle +\kappa_\uparrow.$$

A thermal bath has $\kappa_\downarrow=\kappa(\bar n_{\rm th}+1)$ and $\kappa_\uparrow=\kappa\bar n_{\rm th}$: $\kappa$ is the energy-relaxation rate toward $\bar n_{\rm th}$, not the downward-jump coefficient at finite temperature. A finite stationary mean requires $\kappa_\downarrow>\kappa_\uparrow$. The pure-dephasing term alone gives

$$\dot\rho_{mn} =-\frac{\kappa_\phi}{2}(m-n)^2\rho_{mn}.$$

Widely separated Fock components therefore lose phase coherence faster in this model. For a Fock state $|n\rangle$, the initial probability per unit time of a single loss is $n\kappa_\downarrow$. Increasing the mean occupation can make logical states more distinguishable while exposing them to more loss. Every finite-energy bosonic code negotiates that tradeoff.

The code condition is an operational target

Let $P$ project onto the logical code space and let $\{E_\mu\}$ be the errors to be corrected during one cycle. Exact correction requires the Knill–Laflamme conditions

$$P E_\mu^\dagger E_\nu P = c_{\mu\nu}P.$$

Isolating zero-temperature oscillator loss over a short interval, the leading error set is

$$E_0 \simeq I-\frac{\kappa_\downarrow\Delta t}{2}a^\dagger a, \qquad E_1 \simeq \sqrt{\kappa_\downarrow\Delta t}\,a.$$

The no-jump term matters: when loss probabilities depend on the logical state, the absence of a jump also carries information about it and changes the conditional amplitudes. Equal mean occupations remove this leading distinguishability only when the corresponding off-diagonal code matrix elements also satisfy the correction condition. Matching the first moment does not establish exact correction of the finite-time damping channel.

Real hardware satisfies the conditions only approximately and only for a declared error set. The residual logical channel includes uncorrectable multiple events, imperfect syndrome extraction, recovery error, and deformation of the finite-energy code manifold.

Encoding Families and Their Hardware Consequences

Fock and dual-rail encodings

The simplest single-mode encoding selects two number states, such as $|0_L\rangle=|0\rangle$ and $|1_L\rangle=|1\rangle$. It is a qubit in an oscillator but not, by itself, an error-correcting code. Loss maps the excited logical state to the ground logical state and is not distinguishable without learning logical information.

A two-mode dual-rail encoding,

$$|0_L\rangle=|1,0\rangle, \qquad |1_L\rangle=|0,1\rangle,$$

turns a single photon loss into the orthogonal vacuum state $|0,0\rangle$. It is therefore naturally error detecting. Passive beam-splitter interactions implement logical rotations, and total number can be a syndrome. The lost photon carries away the unknown rail superposition. Once only $|0,0\rangle$ remains, reloading an excitation cannot restore that unknown state. Recovery requires additional encoded information, for example an outer code across several dual-rail qubits. Equal rail losses leave the surviving one-photon state unchanged after normalization; unequal losses filter its amplitudes, as calculated in Photonic Qubits.

Cat and rotation-symmetric encodings

A coherent state satisfies $a|\alpha\rangle=\alpha|\alpha\rangle$. For $\alpha\ne0$, normalized even and odd cat states are

$$|C_\alpha^\pm\rangle = \mathcal N_\pm \left( |\alpha\rangle\pm|-\alpha\rangle \right),$$

where

$$\mathcal N_\pm = \left[ 2\left(1\pm e^{-2|\alpha|^2}\right) \right]^{-1/2}.$$

They occupy opposite photon-number parity sectors, and one loss flips parity:

$$a|C_\alpha^\pm\rangle \propto |C_\alpha^\mp\rangle.$$

If these even and odd cats are the two logical basis states, parity is logical $Z$, not an error syndrome: measuring it resolves the logical state and destroys a general superposition. A loss acts inside this two-dimensional space and has no new parity sector to flag independently of the logical information. A loss-correcting multi-component cat code instead places both logical words in the same initial parity sector.

Larger rotation-symmetric codes use several coherent components around a circle in phase space. Their discrete rotational symmetry turns loss modulo a chosen integer into a syndrome (Grimsmo, Combes, and Baragiola). The same symmetry can make selected logical operations simple.

There are two related but different hardware idioms:

1. A loss-correcting cat code uses several coherent components and a syndrome that does not distinguish the logical basis states.
2. A stabilized cat qubit engineers two-photon drive and dissipation, or a Kerr potential, so two separated coherent states form a protected dynamical manifold.

In the stabilized-qubit convention the logical basis is usually chosen as $(|C_\alpha^+\rangle\pm|C_\alpha^-\rangle)/\sqrt2$, approaching $|\pm\alpha\rangle$ at large amplitude. The latter coherent states are not exactly orthogonal at finite amplitude. In this basis, increasing $|\alpha|^2$ can suppress transitions between the two coherent basins approximately exponentially while phase-flip exposure grows roughly with photon number. This creates a strong noise bias rather than protection from every error. Gates, measurement, pump noise, and leakage must preserve that bias if an outer repetition code is to exploit it (Guillaud and Mirrahimi).

Binomial encodings

Binomial codes use finite superpositions of Fock states (Michael and colleagues). A lowest-order single-loss code is

$$|0_L\rangle = \frac{|0\rangle+|4\rangle}{\sqrt 2}, \qquad |1_L\rangle=|2\rangle.$$

Both codewords have mean photon number two, and

$$a\bigl(u|0_L\rangle+v|1_L\rangle\bigr) =\sqrt2\bigl(u|3\rangle+v|1\rangle\bigr).$$

A parity measurement distinguishes this odd error space without resolving $u,v$; a conditional isometry can return it to the code. Finite-time loss also has a no-loss operator $E_0=\eta^{n/2}$, with $\eta=e^{-\kappa t}$ at zero temperature. Its probabilities for the two basis words are

$$p_0^{(0)}=\frac{1+\eta^4}{2},\qquad p_0^{(1)}=\eta^2.$$

Their difference is $(1-\eta^2)^2/2$, second order in $\kappa t$ at short times. Moreover, the no-loss image of $|0_L\rangle$ has unequal vacuum and four-photon amplitudes. The simple loss correction therefore needs no-loss deformation control and does not exactly undo a finite interval of amplitude damping. Binomial Codes derives that correction condition, its finite-time limitations, and the higher-order loss, gain, and dephasing family.

The hardware price is number-selective control over a deliberately broad Fock manifold. State preparation and recovery require calibrated nonlinear operations, while ancilla faults during a long number-selective pulse can imprint uncontrolled phases on the storage mode.

Gottesman–Kitaev–Preskill grid encodings

For dimensionless quadratures

$$\begin{gathered} q=\frac{a+a^\dagger}{\sqrt 2}, \qquad p=\frac{a-a^\dagger}{i\sqrt 2}, \\ [q,p]=i, \end{gathered}$$

the ideal square-lattice GKP qubit is stabilized by

$$S_q=e^{i2\sqrt{\pi}q}, \qquad S_p=e^{-i2\sqrt{\pi}p}.$$

The ideal states are infinite combs and are not normalizable. Laboratory states replace each peak by a finite-width packet and impose a finite-energy envelope. Small displacements in $q$ and $p$ shift the syndrome continuously; a decoder rounds the inferred displacement to a lattice cell while retaining the analog measurement value when useful.

Suitable Gaussian transformations implement logical operations when they preserve the appropriate code lattice; an arbitrary rotation or squeezing does not. Another useful simplification is that pure loss followed by a quantum-limited phase-insensitive amplifier of intensity gain $1/\eta$ has unity net gain. In these quadrature units its added noise is Gaussian with variance $(1-\eta)/\eta$ per quadrature. Both the lost vacuum and the amplifier’s independent vacuum contribute. Reversing their order changes this variance. The equivalence does not replace finite-energy state or non-Gaussian device errors by a universal displacement-noise model. The hardware burden of a GKP encoding lies in producing high-quality non-Gaussian grid states, measuring both stabilizers repeatedly, controlling finite-energy deformation, and preventing ancilla faults from becoming large unheralded displacements.

“GKP squeezing” must specify a convention. Peak width, envelope width, mean energy, and an effective noise variance are related only after a state model is chosen. A single decibel number does not define the logical channel.

Multimode and concatenated encodings

One mode is not a sacred boundary. Logical information can be spread across two or more oscillators to make loss erasures easier to identify, add confidence information, simplify syndrome structure, or reduce sensitivity to one ancilla fault. A bosonic inner code can also be concatenated with an outer qubit code:

$$\begin{gathered} \text{oscillator physics} \\ \downarrow \\ \text{bosonic logical qubit} \\ \downarrow \\ \text{outer-code logical qubit} \end{gathered}$$

Stabilized cat qubits concatenated with a repetition code are one example. GKP qubits concatenated with a surface code are a prominent theoretical architecture (Noh and Chamberland). The outer decoder should use the inner decoder’s analog or erasure information when available; hard-decision syndromes can discard a substantial part of the bosonic advantage.

Multimode proposals can reduce the number of nonlinear ancillas per logical qubit through frequency multiplexing. They can also create mode crowding, correlated control errors, larger calibration graphs, and common-mode loss. “One logical qubit per cavity” is not a complete resource statement when one cavity contains many modes and shares nonlinear control hardware.

No family wins without a noise model

Encoding idiom	Main structure	Natural syndrome	Hardware opportunity	Characteristic cost
Dual rail	one excitation in two modes	total-number loss erasure	passive linear control	matched modes and outer redundancy for recovery
Cat or rotation symmetric	separated coherent components	code-dependent number residue; parity is logical for two-component cats	biased noise or loss correction	pumps, phase stability, and suitable logical gates
Binomial	finite Fock superpositions	parity or generalized number residue	tailored correction with finite support	number-selective nonlinear control
GKP grid	lattice in quadrature phase space	modular quadratures	analog syndromes and Gaussian logical operations	finite-energy grid preparation and precise displacements
Multimode or concatenated	code across modes or code layers	joint checks plus inner confidence	lower outer-code overhead in favorable regimes	couplers, correlated faults, and a larger decoder

The table is a design map, not a ranking. The relevant question is which code, module, and recovery minimize task-level logical error under the measured noise and complete resource budget.

Physical Implementations

Superconducting microwave resonators

Superconducting implementations use high-quality microwave storage modes coupled to Josephson nonlinearities (Blais and colleagues). Storage can reside in a three-dimensional cavity, a coaxial resonator, or a planar resonator. Transmons, fluxonia, superconducting nonlinear asymmetric inductive elements, and pumped couplers can provide number selectivity, conditional displacements, conversion, and engineered multiphoton processes.

This division of labor is powerful:

• the cavity can have a much longer energy lifetime than the ancilla;
• the ancilla supplies preparation, gates, parity or modular measurement;
• a readout resonator converts the ancilla state into a classical record;
• a reset channel removes entropy between rounds;
• buffer modes and pumps stabilize a desired cat or grid manifold;
• tunable couplers connect storage modes while suppressing idle interaction.

It also creates the central circuit-QED failure mode: the long-lived cavity inherits noise from the short-lived nonlinear controller. Ancilla thermal occupation shifts the cavity frequency; ancilla relaxation during a dispersive interaction can leave an unknown cavity phase; strong drives can activate higher levels and spurious modes. Better bare cavity lifetime alone does not guarantee better logical lifetime.

Trapped-ion motion

A trapped ion’s quantized motion is a mechanical oscillator, while internal electronic states provide an ancilla. State-dependent optical forces generate conditional displacements, and optical pumping resets the spin. Grid states and dissipative error-correction maps have been demonstrated in this setting (Flühmann and colleagues; de Neeve and colleagues).

The platform offers precise spin–motion control and a clean oscillator model, but motional heating, frequency drift, anharmonicity, spectator modes, laser noise, and the need to couple several logical modes constrain scaling. The motion often serves as a shared gate bus in ion processors; using it as persistent logical storage changes scheduling and cooling requirements.

Optical modes

Optical bosonic encodings can propagate through room-temperature channels and support deterministic Gaussian transformations with beam splitters, phase shifters, squeezers, homodyne detection, and feed-forward. Their main error is loss. Cat and grid resources can convert optical continuous-variable hardware into finite logical carriers, and temporal or frequency multiplexing can create many modes in one spatial path.

The severe bottleneck is deterministic, scalable preparation of high-quality non-Gaussian resource states. Finite squeezing, source impurity, detector inefficiency, mode mismatch, and switching loss affect the logical channel. They need not all reduce to independent Gaussian displacements; heralding, non-Gaussian contamination, and correlated faults require their own model. A state reconstructed only after heralding is not a deterministic logical source; its success probability and multiplexing cost determine the available preparation rate.

Photonic Qubits explains flying encodings and source-to-detector loss; Bosonic Codes: Research Directions discusses open questions about optical resource preparation and protected operations.

Acoustic, mechanical, and hybrid modes

Bulk-acoustic, surface-acoustic, nanomechanical, and magnonic modes offer long lifetimes, compact wavelengths, or coupling to otherwise incompatible systems. A superconducting ancilla can prepare and measure nonclassical mechanical states; a spin can use motion for control or transduction.

These are promising oscillator carriers, but a nonclassical state or coherent swap is not yet a bosonic logical module. A QEC architecture additionally needs repeated nondemolition syndrome extraction, reset, logical readout, fault handling, and a benchmark against the best physical storage option. Thermal occupation is especially consequential because a mode at frequency $\omega$ and temperature $T$ has

$$\bar n_{\rm th} = \frac{1}{e^{\hbar\omega/(k_{\rm B}T)}-1}.$$

Low-frequency mechanical modes can require deep refrigeration, sideband cooling, or both before gain errors are rare enough for the intended code.

Preparation, Control, and Readout

Gaussian controls are necessary but not always sufficient

A resonant linear drive implements a displacement

$$D(\alpha) = e^{\alpha a^\dagger-\alpha^*a}.$$

Frequency conversion implements beam-splitter interactions, while parametric driving can generate squeezing and two-mode squeezing. These Gaussian operations are fast and structurally simple. They cannot by themselves prepare arbitrary non-Gaussian cat, Fock-superposition, or grid resources from vacuum.

A non-Gaussian resource is needed: nonlinear dynamics, a non-Gaussian ancilla, or a non-Gaussian measurement such as photon counting. Gaussian homodyne measurement and Gaussian feed-forward alone do not create a non-Gaussian state from vacuum. Circuit-QED modules use number-selective ancilla rotations, selective number-dependent arbitrary phase operations, echoed conditional displacements, optimal control, and measurement-based feedback (Krastanov, Heeres, and Eickbusch and colleagues). Optical platforms use photon counting, non-Gaussian ancilla states, conditional preparation, or nonlinear light–matter interfaces.

Preparation fidelity should be separated from logical survival. A protocol can prepare a state with high conditional fidelity but low success probability, or it can reach a stable manifold reliably while the logical phase remains poorly controlled.

Ancilla-mediated control is a fault path

Suppose the dispersive interaction is

$$H_{\rm disp} = -\hbar\chi a^\dagger a\,|e\rangle\langle e|.$$

If the ancilla begins in $|e\rangle$ and relaxes at an unknown time $t_j$ during an otherwise free dispersive interval, the cavity acquires a random number-dependent phase, relative to its bare evolution,

$$U_j = e^{i\chi t_j a^\dagger a}.$$

For codewords spanning many Fock states, this can be more damaging than the ancilla’s own bit flip. Four design strategies recur:

1. detect the ancilla fault and reject or decode it as an erasure;
2. shape the interaction so the propagated error is correctable;
3. use error-transparent Hamiltonians whose action agrees in code and error sectors;
4. reduce ancilla participation through autonomous or direct bosonic interactions.

Fault-tolerant control is a property of the complete noisy operation, not of the ideal logical unitary.

Syndrome extraction is a quantum instrument

Parity, modular quadrature, or a joint multimode stabilizer must be measured without resolving the logical state. One round is a quantum instrument

$$\begin{aligned} \mathcal I_s(\rho)&=\sum_\nu M_{s\nu}\rho M_{s\nu}^\dagger,\\ p(s)&=\operatorname{Tr}\mathcal I_s(\rho). \end{aligned}$$

Here $s$ is the recorded outcome and $\nu$ labels unobserved alternatives, such as an ancillary fault. For discrete outcomes, $\sum_{s,\nu}M_{s\nu}^\dagger M_{s\nu}=I$ when every outcome, including failure, is retained. One Kraus operator per outcome is a special efficient measurement model, not the general noisy case. Characterization includes:

• the assignment matrix for the syndrome;
• measurement-induced dephasing within the logical subspace;
• leakage and ancilla-reset errors;
• correlations between successive rounds;
• latency from measurement to recovery;
• whether the decoder retains analog confidence.

Repeated high-fidelity ancilla readout does not imply a quantum-nondemolition logical measurement. The backaction channel must be measured on logical states.

Readout can be destructive or logical

Oscillator tomography often measures displaced parity,

$$\begin{aligned} W(\alpha) &= \frac{2}{\pi} \operatorname{Tr} \left[ \rho\,\Pi(\alpha) \right], \\ \Pi(\alpha) &= D(\alpha)\Pi D(-\alpha), \\ \Pi &= e^{i\pi a^\dagger a}. \end{aligned}$$

Here $W$ is normalized with measure $d^2\alpha$; for vacuum, $W(\alpha)=2e^{-2|\alpha|^2}/\pi$. Repeated displaced-parity measurements reconstruct this quasiprobability but are too expensive for routine processor readout. Operational schemes instead map a logical Pauli or modular observable to an ancilla, release the field into a detector, or decode the oscillator into a two-level system. Reported readout fidelity should state whether it includes decoding, heralding, leakage classification, and state preparation.

Active, Autonomous, and Hybrid Protection

An active correction cycle

An active round has four stages:

$$\begin{gathered} \text{interact} \longrightarrow \text{measure syndrome} \\ \downarrow \\ \text{decode} \longrightarrow \text{recover or update frame} \end{gathered}$$

The cycle time $t_{\rm cyc}$ must be short compared with accumulation of uncorrectable errors, yet making it too short can increase ancilla-induced error, readout load, and reset overhead. If a simplified per-cycle model has uncorrectable storage probability $A t_{\rm cyc}^2$ and correction overhead $B$, the logical error rate per unit time is

$$\Gamma_L(t_{\rm cyc}) \simeq A t_{\rm cyc} +\frac{B}{t_{\rm cyc}}.$$

Its optimum is

$$t_{\rm cyc}^{\star} = \sqrt{\frac{B}{A}}, \qquad \Gamma_L^\star=2\sqrt{AB}.$$

This short-cycle approximation requires $At_{\rm cyc}^2\ll1$ and $B\ll1$. It also assumes an attainable cycle duration: if the apparatus imposes $t_{\rm cyc}\ge t_{\min}$, the optimum within this model is $\max(t_{\min},\sqrt{B/A})$. A predicted optimum outside the small-error regime needs a more complete model.

Autonomous correction and confinement

Autonomous schemes engineer a reservoir whose steady manifold is the desired code or whose jumps repair a selected error. A target master equation might contain

$$\dot\rho \supset \kappa_2 \mathcal D[a^2-\alpha^2]\rho .$$

For this ideal dissipator, every density operator supported on $\operatorname{span}\{|\alpha\rangle,|-\alpha\rangle\}$ is stationary, because $a^2-\alpha^2$ annihilates that entire span. It stabilizes a two-dimensional manifold, not one chosen logical state. A single loss acts inside it, so this process alone does not correct the resulting logical error. Additional reservoirs can repair errors for suitable encodings or pump finite-energy grid states; their action must be checked on arbitrary logical superpositions (Mirrahimi, Leghtas, and Gertler and colleagues).

Autonomous protection removes measurement and digital-feedback latency, but it does not remove control hardware. Pumps, lossy buffer modes, filters, and reservoir calibration become part of the correction apparatus. The engineered dissipation must remove entropy faster than harmful errors while avoiding logical dephasing, heating, and undesired steady states.

Hybrid protection is often the practical architecture

Real modules can stabilize a manifold continuously, measure occasional syndromes, track likely errors in software, and concatenate modes with an outer code. For example:

• two-photon dissipation suppresses cat bit flips;
• an outer repetition code corrects remaining phase flips;
• ancilla outcomes flag dangerous control faults;
• a decoder combines flags with parity history.

The distinction between active and autonomous is therefore not binary. Architectures should state which entropy is removed by engineered physics, which information reaches a classical controller, and which residual channel the outer code sees.

Storage and control errors

Photon or phonon loss

For a zero-temperature amplitude-damping channel with transmissivity $\eta=e^{-\kappa t}$, a number state evolves through

$$\begin{aligned} \mathcal E_\eta \left( |n\rangle\langle n| \right) &= \sum_{\ell=0}^{n} p_\ell |n-\ell\rangle\langle n-\ell|, \\ p_\ell &= \binom{n}{\ell} \eta^{n-\ell} (1-\eta)^\ell . \end{aligned}$$

Codes can identify low-order loss number $\ell$, but the probability of two or more losses grows with occupation and correction interval. Loss also shrinks coherent amplitudes and blurs finite-energy grids.

Thermal gain

Residual thermal population adds $a^\dagger$ jumps. A code optimized only for loss may not correct gain with the same distance. Thermal photons in an ancilla or readout mode can be equally damaging because dispersive coupling converts their occupation into storage-mode dephasing.

Dephasing and frequency noise

Oscillator frequency fluctuations couple through $a^\dagger a$. They rotate phase-space states and apply number-dependent phases to Fock superpositions. The same noise can act very differently on different codes: coherent cat components rotate in phase space, binomial states span selected number levels, and a GKP lattice rotates away from its decoding axes.

A Ramsey-derived $T_2$ is not enough when noise is nonstationary or when control pumps change the spectrum. Logical characterization should include idle, driven, and correction-cycle conditions.

Kerr and deterministic deformation

Self-Kerr evolution

$$U_K(t) = e^{iKt\,a^{\dagger 2}a^2/2}$$

is coherent rather than stochastic. Its phase is quadratic in photon number, so an ordinary linear phase rotation cannot undo it on a general Fock superposition. Compensation requires the corresponding nonlinear operation, a suitable sign-reversal sequence, or control pulses designed with this known evolution included. Uncertainty in $K$, ancilla-state-dependent Kerr, and drive-induced nonlinear terms turn it into logical error. Calling all deformation “decoherence” hides a potentially calibratable mechanism.

Ancilla relaxation, dephasing, and leakage

Ancilla faults can propagate through dispersive or conditional operations. Leakage into higher ancilla levels changes dispersive shifts and can survive reset designed for a two-level model. A useful module reports ancilla temperature, $T_1$, $T_2$, leakage, reset error, and the storage channel conditioned on each detected ancilla event.

Pump and mode errors

Parametric pumps can heat chips, create quasiparticles, mix unwanted modes, or amplify phase noise. Multimode cavities introduce spectral crowding and cross-Kerr couplings. Filters and lossy buffers create extra fabrication and packaging constraints. Correction performance measured with one module does not establish that many pumped modules can operate simultaneously.

Metrics That Survive Comparison

A logical channel, not one lifetime

A logical memory maps an input state through encoding, storage and correction, then decoding. Its effective channel $\mathcal E_L(t)$ can be summarized by entanglement fidelity, average fidelity, Pauli error rates, or worst-axis survival. One exponential lifetime is meaningful only if the state ensemble, fit model, SPAM treatment, and logical observable are declared.

For a qubit channel, the average fidelity and entanglement fidelity obey

$$F_{\rm avg} = \frac{2F_e+1}{3},$$

provided both refer to the same trace-preserving logical channel. Postselected data arise from a trace-decreasing accepted map $\mathcal A$. The normalized update $\mathcal A(\rho)/\operatorname{Tr}\mathcal A(\rho)$ is generally nonlinear when acceptance depends on $\rho$, and is then not a quantum channel. Report the accepted map and its acceptance probabilities; the trace-preserving fidelity identity cannot be applied blindly to normalized postselected data.

Break-even requires a named denominator

A common memory gain is

$$G = \frac{\tau_L}{\tau_{\rm ref}}.$$

Possible references include the storage mode’s best two-level subspace, the bare ancilla, the best physical qubit in the module, or the best unencoded memory available in the architecture. These denominators answer different questions. The comparison should match:

• input-state ensemble and error metric;
• total elapsed time and duty cycle;
• encoding and decoding;
• active control and accepted runs;
• hardware available to the reference;
• uncertainty and drift interval.

Crossing memory break-even establishes a useful logical memory under those conditions. It does not establish fault-tolerant state preparation, gates, measurement, or scaling.

Per-round error needs a time and a circuit

An error probability per correction round can fall while the round duration grows. Reports should provide both $p_L$ per round and $\Gamma_L$ per unit time, the number and type of ancilla interactions, and the decoder. In concatenated experiments, logical improvement with distance is the key scaling evidence; one low-error distance is not a threshold demonstration.

Storage and correction use different components

Reducing the number of data modes need not reduce the number of nonlinear ancillas, buffers, couplers, pump tones, or readout chains. Shared controls can save components while limiting concurrency or creating correlated faults. Error-Correction Case Studies compares protected tasks and their physical references; current experimental questions are discussed in Bosonic Codes: Research Directions.

Practical advantages and limitations

Useful physical features

• Large local Hilbert space: redundancy can live inside one mode instead of requiring a separate data carrier for every inner-code degree of freedom.
• Favorable storage hierarchy: a high-quality oscillator can outlive the nonlinear ancilla that controls it.
• Noise tailoring: rotation symmetry, grid translation symmetry, and dual-rail number checks can convert dominant faults into detectable syndromes or strongly biased channels.
• Analog information: GKP and multimode measurements can return confidence values rather than only hard syndrome bits.
• Engineered protection: pumps and dissipation can remove selected entropy continuously, reducing digital-feedback latency.
• Native multimode operations: displacements, beam splitters, conversion, and squeezing can be fast and structurally simple in suitable hardware.

Central bottlenecks

• Nonlinearity without contamination: every useful ancilla or pump can shorten storage coherence, induce Kerr, or propagate faults.
• Preparation overhead: finite-energy non-Gaussian states are costly to create deterministically and repeatedly.
• Fault-tolerant gates: memory protection must survive logical operations, measurement, reset, and inter-module coupling.
• Mode and pump scaling: spectral crowding, common-mode noise, heating, and calibration grow with concurrent modules.
• Auxiliary components: compact data encoding still needs buffers, couplers, readout chains, classical control, and outer-code overhead.
• Manufacturing and packaging: three-dimensional cavities provide excellent lifetimes but are not automatically dense; planar devices are denser but can expose storage modes to additional interfaces and loss.
• Verification: finite-energy envelopes, analog decoders, postselection, and anisotropic logical channels make one-number comparisons fragile.

Common pitfalls

Parity is not always a syndrome. For an even/odd two-component cat qubit, it measures a logical observable. For a code with both words in one parity sector, a changed parity can flag loss without revealing the logical amplitudes; recovery still has to preserve those amplitudes.

Confinement is not complete correction. A dark manifold can protect against escape while allowing logical errors inside it. An outer code must handle the residual channel, including rare errors that break a favorable noise bias.

More energy is not uniformly better. Greater separation or grid squeezing can improve distinguishability while increasing loss exposure, control difficulty, and finite-energy deformation. Optimize the complete operation under its noise model, not a single state parameter.

Memory protection and protected gates are separate tests. A gate that maps the ideal code to itself may spread one ancilla fault into an uncorrectable data error. Universality alone does not establish fault tolerance.

Exercises

1. Bosonic enhancement of loss

A storage mode has energy-decay rate $\kappa$. Compare the no-loss probabilities of $|1\rangle$ and $|4\rangle$ after time $t$. Expand both to first order in $\kappa t$.

Solution

For a number state, every one of the $n$ excitations can be lost, so

$$P_0(n,t)=e^{-n\kappa t}.$$

Therefore,

$$P_0(1,t)=e^{-\kappa t}, \qquad P_0(4,t)=e^{-4\kappa t}.$$

At short times,

$$\begin{aligned} P_0(1,t)&\simeq1-\kappa t, \\ P_0(4,t)&\simeq1-4\kappa t. \end{aligned}$$

Higher occupation can increase code separation or redundancy, but it also increases the raw loss-event rate. Code performance depends on whether those more frequent events remain correctable.

2. Accepted rounds and retained memories

A recovery module uses one storage oscillator, one nonlinear ancilla, and one readout oscillator. Sequential parity extraction, recovery, and reset take $1.2$, $0.5$, and $0.8\,\mu\mathrm{s}$. Each round is independently accepted with probability $0.97$, and the entire memory run is discarded at the first rejection. Find the cycle time and the fraction of runs retained after 100 rounds. Does the single-round acceptance imply that $97\%$ of those memories survive?

Solution

The cycle takes $2.5\,\mu\mathrm{s}$, so 100 rounds take $250\,\mu\mathrm{s}$ under the stated sequential schedule. The retained fraction is

$$p_{100}=0.97^{100}\simeq0.0476.$$

Only about $4.8\%$ of runs remain. A high conditional fidelity of these survivors would omit most attempted memories. An outer code may instead recover a flagged erasure, but that is a different protocol from discarding the run. There are two explicitly listed oscillator modes plus the nonlinear ancilla; whether the latter is itself modeled as an oscillator depends on the hardware.

3. Noise bias and arbitrary-state memory

Consider a logical qubit with independent stochastic Pauli jumps,

$$\dot\rho=\Gamma_X(X\rho X-\rho)+\Gamma_Z(Z\rho Z-\rho),$$

where $\Gamma_X=0.1\,\mathrm{s}^{-1}$ and $\Gamma_Z=2\times10^6\,\mathrm{s}^{-1}$. Derive the three Bloch-component decays and the state-averaged memory fidelity. Interpret the ten-second mean waiting time between $X$ jumps.

Solution

Conjugation by $Z$ changes the signs of $X,Y$, while conjugation by $X$ changes the signs of $Y,Z$. Thus

$$\begin{aligned} r_x(t)&=e^{-2\Gamma_Zt}r_x(0),\\ r_y(t)&=e^{-2(\Gamma_X+\Gamma_Z)t}r_y(0),\\ r_z(t)&=e^{-2\Gamma_Xt}r_z(0). \end{aligned}$$

Writing these decay factors as $\lambda_x,\lambda_y,\lambda_z$, the input/output overlap averaged over pure qubit states is

$$F_{\rm avg}(t)=\frac12+\frac16(\lambda_x+\lambda_y+\lambda_z).$$

The jump-rate bias is $2\times10^7$. The waiting time $1/\Gamma_X=10$ s is not a Bloch decay time: the $z$ component decays with time constant $1/(2\Gamma_X)=5$ s. Transverse coherence decays on a $0.25\,\mu\mathrm{s}$ scale. After that rapid decay but before appreciable $z$ decay, $F_{\rm avg}$ is approximately $2/3$, not near unity. Protecting one axis for a long time does not preserve an arbitrary qubit.

4. Thermal occupation of a microwave mode

Estimate $\bar n_{\rm th}$ for a $5\ {\rm GHz}$ mode at:

1. $T=20\ {\rm mK}$;
2. $T=100\ {\rm mK}$.

Use $h f/k_{\rm B}\simeq0.240\ {\rm K}$.

Solution

The thermal occupation is

$$\bar n_{\rm th} = \frac{1}{e^{hf/(k_{\rm B}T)}-1}.$$

At $20\ {\rm mK}$, the exponent is $0.240/0.020=12$, so

$$\bar n_{\rm th} \simeq \frac{1}{e^{12}-1} \simeq 6.1\times10^{-6}.$$

At $100\ {\rm mK}$, the exponent is $2.4$, giving

$$\bar n_{\rm th} \simeq \frac{1}{e^{2.4}-1} \simeq 0.10.$$

The equilibrium estimate is extremely temperature sensitive. A measured mode can still be hotter than the refrigerator because of imperfect filtering, pump heating, readout backaction, or a hot ancillary system.

5. Ideal square-grid correction radius

In the ideal square GKP code with $[q,p]=i$, nearest-lattice decoding corrects an isolated $q$ displacement when

$$|\delta q|<\frac{\sqrt\pi}{2}.$$

Classify displacements $\delta q=0.60$ and $\delta q=1.00$. What is omitted by this binary classification?

Solution

Numerically,

$$\frac{\sqrt\pi}{2}\simeq0.886.$$

Thus $\delta q=0.60$ lies inside the nearest cell and is ideally correctable. The displacement $\delta q=1.00$ lies beyond the boundary and is rounded to a neighboring lattice point, producing a logical error under this decoder.

The statement omits finite peak width, the finite-energy envelope, imperfect syndrome measurement, simultaneous $p$ displacement, analog likelihood, loss-induced attenuation, and correlations between rounds. Near the boundary, a confidence-aware decoder can be more informative than a hard correct/error label.

6. Optimize a correction interval

For the schematic rate

$$\Gamma_L(t) = A t+\frac{B}{t}, \qquad A,B>0,$$

find the optimal cycle time and minimum rate. Explain the two terms, and then impose a hardware lower bound $t\ge t_{\min}$.

Solution

Differentiate:

$$\frac{d\Gamma_L}{dt} = A-\frac{B}{t^2}.$$

The positive optimum is

$$t^\star=\sqrt{\frac{B}{A}}.$$

Substitution gives

$$\Gamma_L(t^\star) = 2\sqrt{AB}.$$

The $At$ term represents uncorrectable multiple storage errors becoming more likely as the interval grows. The $B/t$ term represents a roughly fixed syndrome, reset, or recovery error paid every cycle. Faster correction reduces the first contribution while applying the faulty correction apparatus more often. With a hardware lower bound, the optimum becomes $t_{\rm opt}=\max(t_{\min},\sqrt{B/A})$; the minimum is then $At_{\rm opt}+B/t_{\rm opt}$. Both estimates require the underlying per-cycle probabilities to remain small.

7. Propagate a break-even uncertainty

A logical memory has

$$\begin{aligned} \tau_L &= (1.82\pm0.03)\ {\rm ms}, \\ \tau_{\rm ref} &= (0.800\pm0.010)\ {\rm ms}. \end{aligned}$$

Assuming independent Gaussian uncertainties, estimate $G=\tau_L/\tau_{\rm ref}$ and its uncertainty.

Solution

The central value is

$$G = \frac{1.82}{0.800} = 2.275.$$

For independent uncertainties,

$$\left( \frac{\sigma_G}{G} \right)^2 = \left( \frac{0.03}{1.82} \right)^2 + \left( \frac{0.010}{0.800} \right)^2.$$

This gives

$$\sigma_G\simeq0.047,$$

so a suitable rounded result is

$$G=2.28\pm0.05.$$

The calculation quantifies statistical separation from unity. It does not check whether the state ensemble, SPAM treatment, hardware access, and drift conditions were matched; those are part of the benchmark definition.

8. Conditional fidelity as a measurement probability

A heralded preparation has acceptance probability $p_{\rm acc}=0.60$ and conditional overlap $F_{\rm cond}=\langle\psi|\rho_{\rm acc}|\psi\rangle=0.98$ with a pure target. What is the probability that an attempt is accepted and then passes an ideal projector test onto $|\psi\rangle$? What changes if rejected attempts are stored as a flag orthogonal to the target?

Solution

The joint Born probability of acceptance and passing the target-state test is

$$0.60\times0.98=0.588.$$

For an orthogonal rejection flag, the unconditional target overlap is also $0.588$. This is a precise measurement probability, not a decomposition of each quantum output into a hidden label “correct” or “incorrect.” The conditional $0.98$ alone omits rejected attempts and their time cost.

References

• A. Blais, A. L. Grimsmo, S. M. Girvin, and A. Wallraff, “Circuit quantum electrodynamics,” Reviews of Modern Physics 93, 025005 (2021), doi:10.1103/RevModPhys.93.025005.
• B. de Neeve et al., “Error correction of a logical grid state qubit by dissipative pumping,” Nature Physics 18, 296–300 (2022), doi:10.1038/s41567-021-01487-7.
• A. Eickbusch et al., “Fast universal control of an oscillator with weak dispersive coupling to a qubit,” Nature Physics 18, 1464–1469 (2022), doi:10.1038/s41567-022-01776-9.
• C. Flühmann et al., “Encoding a qubit in a trapped-ion mechanical oscillator,” Nature 566, 513–517 (2019), doi:10.1038/s41586-019-0960-6.
• J. M. Gertler et al., “Protecting a bosonic qubit with autonomous quantum error correction,” Nature 590, 243–248 (2021), doi:10.1038/s41586-021-03257-0.
• D. Gottesman, A. Kitaev, and J. Preskill, “Encoding a qubit in an oscillator,” Physical Review A 64, 012310 (2001), doi:10.1103/PhysRevA.64.012310.
• A. L. Grimsmo, J. Combes, and B. Q. Baragiola, “Quantum computing with rotation-symmetric bosonic codes,” Physical Review X 10, 011058 (2020), doi:10.1103/PhysRevX.10.011058.
• J. Guillaud and M. Mirrahimi, “Repetition cat qubits for fault-tolerant quantum computation,” Physical Review X 9, 041053 (2019), doi:10.1103/PhysRevX.9.041053.
• R. W. Heeres et al., “Cavity State Manipulation Using Photon-Number Selective Phase Gates,” Physical Review Letters 115, 137002 (2015), doi:10.1103/PhysRevLett.115.137002.
• S. Krastanov et al., “Universal control of an oscillator with dispersive coupling to a qubit,” Physical Review A 92, 040303(R) (2015), doi:10.1103/PhysRevA.92.040303.
• Z. Leghtas et al., “Confining the state of light to a quantum manifold by engineered two-photon loss,” Science 347, 853–857 (2015), doi:10.1126/science.aaa2085.
• M. H. Michael et al., “New class of quantum error-correcting codes for a bosonic mode,” Physical Review X 6, 031006 (2016), doi:10.1103/PhysRevX.6.031006.
• M. Mirrahimi et al., “Dynamically protected cat-qubits: a new paradigm for universal quantum computation,” New Journal of Physics 16, 045014 (2014), doi:10.1088/1367-2630/16/4/045014.
• K. Noh and C. Chamberland, “Fault-tolerant bosonic quantum error correction with the surface–Gottesman–Kitaev–Preskill code,” Physical Review A 101, 012316 (2020), doi:10.1103/PhysRevA.101.012316.