Cat Qubits and Bosonic Encodings

An oscillator supplies a larger code space#

Define coherent states, phase-space separation, encoding, and the physical elements counted as one bosonic mode.

Most qubits are carved out of a two-level physical system: two spin states, two charge states, two energy levels. Bosonic encodings take a different route. They store quantum information in an oscillator mode — a microwave resonator, for instance — whose state space is infinite-dimensional, and they spend that extra room on built-in error resistance.

Bosonic encodings use oscillator Hilbert space to encode quantum information and can produce structured error channels. [bosonic-review] [gkp-2001]

The code space lives inside a physical oscillator#

An oscillator supplies many energy levels, but an encoding occupies a structured subset of that space. GKP states use a phase-space lattice; binomial codes use selected Fock-state superpositions; cat codes use separated coherent-state components. Each construction turns particular small physical disturbances into detectable syndromes or a biased logical channel. The mechanism is only the first layer. Finite energy, preparation, nonlinear control, ancilla coupling, measurement, and loss determine the realized code.

Count the physical system consistently. One stored logical qubit may require a high-quality cavity or resonator, one or more nonlinear elements, pumps, couplers, readout modes, and an ancillary two-level system. If a surface-code baseline counts every data and syndrome qubit, a bosonic proposal cannot count only storage modes. Report modes, nonlinear elements, control channels, measurements, cycle duration, and reusable versus dedicated ancillas [full-stack-review].

Cycle evidence: Publish time-ordered syndrome and operation records for enough repeated cycles to estimate correlation and leakage persistence. Compare a decoder using hard decisions with one using available analog information, while keeping the data split and tuning policy fixed. Report conditional and unconditional logical failure, not only accepted-shot fidelity. Then remove one idealization at a time from the resource model: finite ancilla reset, pump bandwidth, measurement duration, decoder delay, and calibration downtime. If the projected advantage survives, it is attached to a plausible system. If one omission reverses it, that missing primitive becomes the next proof gate. This workflow respects the genuinely different physics of bosonic storage while preventing its largest idle-state number from standing in for complete computation.

Noise bias is a ratio with an operation label#

Define bit- and phase-like logical error probabilities for idle, gate, and readout separately.

The cat encoding is the canonical example. An oscillator driven into a superposition of two large, opposite-amplitude states—traditionally written $\lvert+\alpha\rangle$ and $\lvert-\alpha\rangle$ for coherent-state amplitude $\alpha$—holds a qubit whose logical states are widely separated in phase space. Small environmental disturbances tend to rotate or shrink each component, producing phase-like errors; a bit-like transition must cross the separation and can be engineered to be much rarer. This is biased noise: one error channel is suppressed relative to another.

Cat-state encodings can exhibit biased noise whose usefulness depends on physical operations preserving that bias. [bosonic-review]

Noise bias needs an operation subscript#

Let $p_X^{(o)}$ and $p_Z^{(o)}$ denote bit- and phase-like logical errors for operation $o$. The bias $\eta_o=p_Z^{(o)}/p_X^{(o)}$, or its inverse depending on convention, is meaningless unless the convention and operation are named. Measure separate channels for idle, preparation, entangling gate, measurement, and reset. A long-lived idle manifold does not guarantee that a control pulse preserves the same asymmetry.

Use confidence intervals when the suppressed event is rare. Observing zero bit flips in $N$ trials is not an infinite bias; it gives an upper bound determined by $N$ and the confidence level. Also inspect leakage from the encoded manifold and events that produce both error types. An outer decoder trained on a simple biased Pauli channel may fail when leakage persists across cycles.

Photon loss and leakage reshape the channel#

Trace dominant physical mechanisms into encoded error or detectable syndrome under explicit assumptions.

The strategic promise follows directly. Quantum error correction is expensive largely because it must fight bit flips and phase flips at once. If physics has already crushed one channel, the correction layer — often a much simpler code — only handles the survivor, and the physical-qubit overhead per logical qubit could fall dramatically. That is a real mechanism, not a marketing line. The question this chapter trains is whether the saving survives the whole machine.

Error-correction overhead comparisons require a declared physical error model, operation set, and target logical reliability. [bosonic-binomial-qec] [bosonic-binomial-qec]

Photon loss can be correctable, detectable, or destructive#

The action of loss depends on the encoding and measurement scheme. A parity measurement may reveal a loss event; a recovery operation may restore the code; repeated losses or unobserved timing can produce a logical error. Dephasing, thermal excitation, Kerr distortion, ancilla faults, and pump noise add other channels. Build the channel from measured mechanisms rather than assigning one “bosonic fidelity.”

Analog syndrome information can improve decoding. GKP correction, for example, can use the residual displacement within a lattice cell rather than only a hard nearest-cell decision [gkp-2001]. Preserve measurement resolution and calibration in the data path. Digitizing too early can discard precisely the information behind a projected overhead advantage.

The outer code prices the remaining error#

Compare a bias-tailored code with a symmetric baseline only at matched target logical failure and operation set.

Every fault-tolerance plan reduces to one accounting identity: $Q_{\mathrm{physical}}\approx Q_{\mathrm{logical}}q_{\mathrm{per\ logical}}$ . Bosonic and cat-encoding claims are arguments that the second factor can shrink under the right assumptions. The only useful version of that argument states the assumptions out loud: the physical error rates, the gate set, the measurement scheme, and the baseline being beaten.

A bosonic element remains embedded in a larger control, coupling, readout, and manufacturing stack. [full-stack-review] [bosonic-break-even]

A full cycle is the minimum systems experiment#

Trace preparation, stabilization, one representative idle interval, one gate, syndrome extraction, decoding, recovery or frame update, and readout. Repeat the cycle enough times to expose drift and correlated events. Record pump duty, ancilla resets, controller latency, and rejected runs. A break-even logical memory experiment is a meaningful milestone because it compares an encoded lifetime with a physical constituent under declared conditions [bosonic-break-even]; it does not automatically demonstrate gates or repeated fault-tolerant computation.

A binomial-code experiment with error correction and a universal gate set addresses a broader operation set under its specific apparatus [bosonic-binomial-qec]. Read its denominators carefully: which operations are corrected, how process fidelity is estimated, what ancillas are counted, and which faults remain. The evidence should populate distinct rows for memory, one-mode gates, two-mode gates, measurement, and repeated cycles.

A full cycle tests whether the bias survives#

Require prepare-operate-measure-correct repetitions with latency, ancillas, and uncertainty.

Score such claims the way this book scores all of them: decision score = claim quality + proof progress − risk − kill criteria pressure . The score should penalize overhead promises that arrive without a denominator and reward comparisons against a clear baseline — typically a surface-code-style architecture — run under comparable error assumptions. An encoding that wins only against a straw-man baseline has proven something about its marketing, not its physics.

Current bias-preserving operation and overhead claims require primary experimental/resource papers beyond the registered review. [bosonic-review]

The outer code prices the residual channel#

Suppose an inner bosonic encoding strongly suppresses $X$-like errors while leaving $Z$-like errors. A repetition-style outer code may then be cheaper than a symmetric two-dimensional code. To compare, feed both architectures matched operation-level channels, target sequence failure, and timing. Include syndrome rounds, decoder, and ancillas. The outer code's check duration can expose the oscillator to more of the dominant channel, partly consuming the bias advantage.

Perform sensitivity sweeps over idle bias, gate bias, measurement error, leakage lifetime, and cycle duration independently. The claim is robust only if it wins over a plausible region rather than at one selected point. Mark discontinuities where code distance changes. A lower count at one target may reverse at a stricter reliability or different gate mix.

Resource accounting prevents denominator games#

List modes, transmons/nonlinear elements, readout channels, controls, cycle time, and factory-like overhead consistently.

Watch the bias itself as a measurable quantity with a lifetime. Preparation, gates, and readout all act on the oscillator, and each can leak population out of the encoded subspace or mix the two error channels. The claim to check is never "the noise is biased" but "the noise is still biased after the operations the computer actually performs."

Bosonic encodings use oscillator Hilbert space to encode quantum information and can produce structured error channels. [bosonic-review] [gkp-2001]

Do not compare incompatible physical units#

“Physical qubits per logical” can count transmons, oscillators, modes, modules, or a mixture. Publish both a functional bill of materials and a platform-specific footprint: storage modes, nonlinear elements, readout chains, pump tones, control bandwidth, cryogenic power, and volume where relevant. Then report logical throughput and failure. A one-mode logical memory and a complete surface-code logical processor are not equal rows simply because both contain one logical qubit.

For the synthetic worksheet, keep the units deliberately explicit. If one row reports modes per logical and another transmons per logical, the comparison engine should refuse to rank them until a common system boundary is supplied. This is a feature. It stops a smaller numerator produced by narrower accounting from masquerading as architectural savings.

The worksheet should make a fragile advantage visible#

Idle

error channel per unit time with stabilization active.

Entangling operation

operation-labeled $X$, $Z$, and leakage probabilities.

Measurement

assignment channel, duration, and rejected records.

Outer cycle

decoder input, correction cadence, and delivered logical failure.

Give each operation its own $X$-like error, $Z$-like error, leakage, duration, and confidence interval. Feed preparation, idle, entangling gate, measurement, and reset through the declared sequence. Apply the outer decoder model and recompute run-level failure. The worksheet must reject an idle-bias value reused for a gate without an explicit assumption.

Compare against a symmetric baseline at the same delivered operation, target failure, and wall-clock boundary. Count oscillator modes, nonlinear ancillas, pumps, readout chains, and outer-code elements on one side; count data, syndrome, control, and correction resources on the other. Where units cannot be normalized, show two resource columns rather than an invented scalar score.

Sweep the weakest operation. If a tenfold worsening of measurement bias or entangling leakage eliminates the advantage, the recommendation should fund that primitive before scaling storage. If the advantage survives uncertainty and repeated cycles, the result supports an architecture study. Sensitivity, not the largest headline bias, determines the next experiment.

Price the bias across a complete logical primitive#

The useful object is not an idle cat state but an operation set whose dominant error remains compatible with the outer code. Measure or model preparation, idle, entangling interaction, syndrome extraction, readout, and reset separately. For each operation $o$, retain $p_X^{(o)}$, $p_Z^{(o)}$, leakage or loss, duration, and the convention for the reported bias. A single ratio averaged across unlike operations can hide the gate that destroys the advantage.

Map those channels through the intended schedule. Repeated parity checks introduce ancilla faults and dead time; couplings between oscillators can convert a suppressed bit-like process into correlated phase faults; recovery can return population to the code manifold with a different logical channel. The outer-code estimate should consume this operation-labeled channel, not the lifetime of the best idle state.

Compare against a baseline with the same logical task and inventory boundary. Count oscillator modes, nonlinear elements, ancillary two-level systems, pumps, readout chains, controller bandwidth, cycle seconds, and accepted-run probability. Break-even can mean a longer memory, a lower error per operation, or less total hardware at a target failure rate; choose one and keep the others visible. The conclusion reverses when bias-preserving gates, measurement, or manufacturing cost no longer close the matched budget. A sensitivity sweep over those operation-specific channels should identify the first crossover, rather than assuming that a favorable idle-bias measurement survives the processor schedule.