Continuous-Variable Quantum Computation

Continuous-variable quantum computation uses bosonic modes as computational systems and lets quadrature amplitudes, rather than a finite list of basis labels, carry the state. Its circuit language includes Gaussian state preparation, affine symplectic gates and channels, continuous measurements, classical feedforward, and explicitly identified non-Gaussian resources. A complete proposal must also say which finite-energy inputs are allowed, what continuous or discrete output is decoded, how approximation is measured, and which physical and numerical resources are charged.

This page owns that device-independent computation model. It does not turn an ideal quadrature eigenstate into a physical source, identify a Gaussian circuit with a classical device, equate one oscillator with one protected logical qubit, or infer efficient compilation, fault tolerance, or computational advantage from algebraic universality.

Bosonic and Encoded Computation Models owns finite logical programs induced on declared oscillator encodings, including complete physical composition, decoded logical channels, leakage, rejection, recovery, frames, and truncation. This page retains direct-mode quadrature computation, Gaussian and non-Gaussian resources, continuous outcomes, and CV cluster patterns.

Required background. Continuous-Variable Systems supplies mode Hilbert spaces, Fock and quadrature representations, canonical commutators, generalized eigenstates, normalizability, and finite-energy domain cautions. Gaussian States and Wigner Functions supplies first moments, covariance matrices, the uncertainty condition, Gaussianity, symplectic propagation, and Wigner-function interpretation.

Continuous-Variable Computation as a Mode Model

For an instance $x$, a model begins with an input state $\rho_x$ on $m$ modes, a declared sequence of allowed transformations and measurements, a classical control rule, and an output decoder. If $B$ is a measurable output event, its probability has the operational form

$$P_x(B) = \operatorname{Tr} \left[ M_x(B)\, \Phi_x(\rho_x) \right],$$

where $\Phi_x$ includes all outcome-forgotten processing before the terminal POVM $M_x$. When intermediate outcomes remain selected, the model instead records the corresponding instrument branch and its probability. A probability density at one real value is not a discrete event probability.

This definition separates a computation from the notation used to describe it. The Circuit Model supplies the general grammar of wires, composition, measurement, and logical resource counts. Here a wire denotes an infinite-dimensional mode, and an allowed operation must be stated in quadrature, phase-space, channel, or measurement language. Bits, Qubits, Qudits, and Modes owns the more general distinction between a physical carrier and an encoding.

Direct mode computation and encoded computation are different contracts. A quadrature displacement can be a direct continuous-variable output, while a grid-code decoder can turn oscillator data into a logical qubit. Mode count, logical dimension, mean energy, cutoff, and decoder must therefore be reported separately.

The learner exit is concrete: given a proposed program, identify its input family and domains; translate every covariance and gate into one convention; propagate the Gaussian part; locate every departure from Gaussian closure; track measurements and feedforward by branch; state an energy-limited approximation metric; compute the output event and error; total the mode, operation, precision, energy, record, and repetition resources; and stop at the first claim owned by hardware, error correction, compilation, or benchmarking.

The Ten-Field Continuous-Variable Computation Record

Use this record before propagating a covariance matrix or invoking a universality result. It is vertical so that assumptions remain visible on narrow screens.

1. Computational task, input family, and licensed claim. State the fixed fixture or instance family, requested output, allowed input energy or cutoff, and the narrow conclusion the calculation can support.
2. Modes, quadrature convention, operator domains, and energy constraint. Give the mode order, commutators, vacuum variance, common domain of unbounded operators, and mean or peak energy restriction.
3. Initial state, ancillas, preparation, and encoding. Specify every system and ancilla state, how it is assumed to be prepared, and whether the computation is direct in modes or encoded.
4. Gaussian operations, channels, and affine symplectic map. List the physical order of Gaussian unitaries and channels, their affine data, and all convention translations.
5. Non-Gaussian states, gates, measurements, and universality role. Name every element outside Gaussian closure and state whether it is merely present, consumed, postselected, or used in a licensed generation theorem.
6. Measurements, classical records, feedforward, and branch convention. Give POVMs or measured quadratures, measurable bins, retained and forgotten records, feedforward functions, and conditioning or postselection.
7. Output system, decoder, success event, and error metric. State the remaining quantum system or classical record, the decoder, the accepted event, and the metric against a declared target.
8. Finite squeezing, loss, noise, precision, cutoff, and resource ledger. Charge finite squeezing, channels, control precision, mode and operation counts, adaptive depth, cutoff, repetitions, and any physical cost actually licensed.
9. Verification data, uncertainty, and reproducibility. Provide analytic invariants, numerical method and convergence data, or sampling uncertainty, together with the conventions needed to reproduce every value.
10. Conclusion, stopping point, and canonical handoff. Say exactly what follows, what does not, and which canonical owner receives the next theorem, implementation, encoding, or evidence question.

Every field needs a value or a specific reason that it is not applicable. A gate list without an input domain, output event, and resource boundary is not a continuous-variable computation record.

Quadrature Conventions, Operator Domains, and Energy Constraints

This page uses $\hbar=1$ and

$$[a_j,a_k^\dagger] = \delta_{jk}, \qquad q_j = \frac{a_j+a_j^\dagger}{\sqrt2}, \qquad p_j = \frac{a_j-a_j^\dagger}{i\sqrt2}.$$

Thus $[q_j,p_k]=i\delta_{jk}$ and the vacuum variances are $\operatorname{Var}(q_j)=\operatorname{Var}(p_j)=1/2$. A covariance quoted in a convention with vacuum variance $1/4$ or $1$ must be rescaled before it is combined with formulas here.

Order the quadratures of $m$ modes as

$$R = (q_1,p_1,\ldots,q_m,p_m)^{\mathsf T}, \qquad [R_j,R_k] = i\Omega_{jk},$$

with

$$\Omega = \bigoplus_{j=1}^{m} \begin{pmatrix} 0&1\\ -1&0 \end{pmatrix}.$$

The first moments and covariance are

$$d = \langle R\rangle, \qquad V_{jk} = \frac12 \left\langle \{R_j-d_j,R_k-d_k\} \right\rangle,$$

and a physical covariance satisfies

$$V+\frac{i}{2}\Omega\succeq0.$$

These moment data characterize a Gaussian state, but not a general state. Higher moments can distinguish two non-Gaussian states with identical $d$ and $V$.

Quadratures, number operators, and polynomial Hamiltonians are unbounded. Algebraic identities involving them are statements on a declared common dense domain before closure. Generalized quadrature eigenvectors and ideal homodyne projectors belong to a rigged-Hilbert-space description; infinitely squeezed cluster states and ideal grid combs are not normalizable finite-energy preparations.

For finite second moments, the total number operator and its expectation are

$$\widehat N = \sum_{j=1}^{m}a_j^\dagger a_j = \frac{R^{\mathsf T}R-m}{2},$$ $$\bar n = \langle\widehat N\rangle = \frac{\operatorname{Tr}V+d^{\mathsf T}d-m}{2}.$$

Mean energy controls a useful but limited cutoff bound. If $\Pi_{<N}$ projects onto total photon numbers $0,\ldots,N-1$, Markov’s inequality gives

$$\operatorname{Tr} \left[ (I-\Pi_{<N})\rho \right] \leq \frac{\bar n}{N}.$$

This is discarded state weight. By itself it is not a gate error, a channel distance, an observable error, or a proof that a truncated numerical sequence has converged.

Gaussian Gates and Symplectic Propagation

A Gaussian unitary has affine Heisenberg action

$$U^\dagger R U = SR+s, \qquad S\Omega S^{\mathsf T} = \Omega.$$

It propagates moments by

$$d \longmapsto Sd+s, \qquad V \longmapsto SVS^{\mathsf T}.$$

Composition follows physical time order. If map 1 is followed by map 2, then

$$S_{\rm tot} = S_2S_1, \qquad s_{\rm tot} = S_2s_1+s_2.$$

For one mode, a phase rotation and a $q$-squeezer can be represented by

$$R(\theta) = \begin{pmatrix} \cos\theta&-\sin\theta\\ \sin\theta&\cos\theta \end{pmatrix}, \qquad S(r) = \begin{pmatrix} e^{-r}&0\\ 0&e^r \end{pmatrix}.$$

For two modes, the controlled-phase gate

$$C_Z(g) = e^{igq_1q_2}$$

leaves both $q$ quadratures fixed and sends

$$p_1 \longmapsto p_1+gq_2, \qquad p_2 \longmapsto p_2+gq_1.$$

Displacements, rotations, squeezers, passive mode mixing, and controlled-phase gates all fit the affine symplectic record, but the record does not imply that they have equal physical cost.

A Gaussian channel has affine data $(X,Y,d_0)$:

$$d \longmapsto Xd+d_0, \qquad V \longmapsto XVX^{\mathsf T}+Y.$$

Complete positivity in this convention requires

$$Y + \frac{i}{2} \left( \Omega-X\Omega X^{\mathsf T} \right) \succeq0.$$

Gaussian Channels owns the proof, dilation theory, named-channel classification, and capacities. A computation record applies a declared channel and charges its noise; it does not rederive that taxonomy.

Measurements, Feedforward, and Output Channels

Ideal homodyne at phase $\theta$ measures

$$q_\theta = q\cos\theta+p\sin\theta.$$

Its result $x\in\mathbb R$ has a density $p_\theta(x)$. Only a measurable set $B$ has an event probability,

$$P(B) = \int_B p_\theta(x)\,dx.$$

Detector bins, saturation, digitization, and finite resolution change $B$ or the measurement channel; they cannot be recovered from the phrase “measure $q$.” Homodyne and Heterodyne Detection owns the quantum-optical detector theory and physical readout conventions.

An adaptive program must distinguish four objects:

• the normalized state selected by a particular record;
• the subnormalized branch map whose trace is that record’s probability;
• the outcome-forgotten channel obtained by summing or integrating branches; and
• a postselected output conditioned on a declared accepted set.

For Gaussian inputs and Gaussian measurements, conditional means depend affinely on the real outcome while the conditional covariance is fixed by the declared measurement model. Affine displacements or frame updates preserve a joint Gaussian description. An arbitrary nonlinear function of the outcome does not enter that closure theorem silently.

The final output may be a mode, an encoded subsystem, a vector of real records, or a decoded discrete label. A valid error statement compares the resulting state or probability law with a named target, using a declared trace distance, fidelity, total-variation distance, moment tolerance, or accepted-event probability. Selected and forgotten records must not be interchanged during that comparison.

Gaussian Closure and Non-Gaussian Resources

Gaussian input states processed by Gaussian unitaries, Gaussian channels, Gaussian measurements, and affine classical feedforward remain Gaussian. Under a polynomial-size finite real description and polynomial requested precision, their moments and Gaussian outcome densities can be propagated efficiently on a classical computer. The simulation results of Bartlett and coauthors, and the positive-Wigner sufficient conditions studied by Mari and Eisert, have explicit hypotheses; neither is a blanket equivalence between “positive Wigner function” and “easy.”

Gaussian does not mean classical, separable, noiseless, or experimentally available. Gaussian states can be entangled and operationally useful. The closure statement says that first and second moments remain a complete description within the declared Gaussian model.

A standard non-Gaussian example is the cubic-phase family

$$U_3(\gamma) = e^{i\gamma q^3}.$$

On a common polynomial domain,

$$U_3(\gamma)^\dagger qU_3(\gamma) = q, \qquad U_3(\gamma)^\dagger pU_3(\gamma) = p+3\gamma q^2.$$

The quadratic term in the transformed momentum leaves affine symplectic closure. Photon counting is likewise a non-Gaussian measurement; Photon Counting owns its detector and POVM physics.

Breaking Gaussian closure is necessary in the usual model for universal polynomial continuous-variable computation, but it is not sufficient for a complete claim. The model must state which Gaussian controls coexist with the resource, whether strengths, signs, or inverses are available, which domains are preserved, how accurately commutator constructions are implemented, and how every cost scales. Resource Theories owns the general free-operation, monotone, and conversion-law framework. Non-Gaussianity is not assigned one context-free scalar value here.

Cluster Nullifiers and Measurement-Based Computation

Let $A$ be a real symmetric weighted adjacency matrix. Prepare independently momentum-squeezed modes and apply the commuting entanglers

$$U_A = \prod_{j<k} e^{iA_{jk}q_jq_k}.$$

The output quadratures satisfy

$$q_j^{\rm(out)} = q_j^{\rm(in)}, \qquad p_j^{\rm(out)} = p_j^{\rm(in)} + \sum_k A_{jk}q_k^{\rm(in)}.$$

The corresponding nullifiers are

$$\delta_j := p_j^{\rm(out)} - \sum_k A_{jk}q_k^{\rm(out)} = p_j^{\rm(in)}.$$

If the input modes are independent and momentum squeezed by $r_j$, then

$$\operatorname{Var}(\delta_j) = \frac{e^{-2r_j}}{2},$$

and distinct nullifiers have zero covariance. The formal zero-nullifier graph state requires $r_j\to\infty$ and is not a normalizable finite-energy state. Finite nullifier variance is a logical noise source, not merely an experimental inconvenience.

Continuous-variable measurement-based computation consumes such a graph by homodyne or non-Gaussian measurements, propagates real records, and applies feedforward or a tracked displacement frame. A Gaussian cluster with homodyne and affine feedforward implements Gaussian logical transformations. Universal CV cluster computation needs a declared non-Gaussian state, gate, or measurement and an approximation contract.

Measurement-Based Quantum Computation owns finite-qubit open graphs, equatorial commands, Pauli byproducts, flow, gflow, and qubit pattern accounting. This page owns the mode-valued cluster model, continuous records, finite-squeezing noise, and its Gaussian/non-Gaussian boundary. Continuous-Variable Platforms owns physical cluster generation, mode functions, loss, detectors, delay lines, calibration, and experimental evidence.

Universality, Approximation, and Compilation

The Lloyd–Braunstein generation argument shows how Gaussian controls plus a suitable available nonquadratic Hamiltonian can generate polynomial Hamiltonians through commutators. This is an ideal algebraic statement. A computational claim must additionally specify the number of modes, allowed input family, common operator domain, energy budget or cutoff, target transformation, approximation metric, accuracy, continuous-parameter precision, and resource scaling.

No finite-dimensional cutoff approximates an oscillator operation uniformly over the entire Hilbert space. If a channel-level comparison is useful, one possible declared metric is the energy-constrained diamond norm

$$\lVert\Phi-\Psi\rVert_{\diamond E} = \sup_{ \rho_{AR}:\, \operatorname{Tr}(\widehat N_A\rho_A)\leq E } \left\lVert \bigl[ (\Phi-\Psi)\otimes I_R \bigr](\rho_{AR}) \right\rVert_1.$$

Using this definition does not supply a convergence theorem. A numerical claim still needs a cutoff sequence, a state or channel metric, tolerances, and observed convergence. A fixed-fixture state error can instead be reported directly in trace distance or fidelity.

Universal Gate Sets owns the general exact-versus-dense universality and synthesis criteria. A finite-dimensional Solovay–Kitaev statement cannot be imported unchanged for unbounded continuous-variable generators. Algebraic generation also does not imply an efficient compiler, bounded energy, robust implementation, or advantage over a matched classical method.

Direct CV computation must also remain distinct from logical qubits encoded in oscillators. GKP Codes owns grid-state codewords, displacement recovery, decoders, logical gates, finite-energy approximations, and threshold conditions. Fault-Tolerant Gates owns encoded gadgets and overhead. A non-Gaussian mode resource is not by itself an error-corrected logical gate.

Finite Squeezing, Noise, and Resource Accounting

Finite squeezing replaces ideal delta-function constraints with Gaussian widths. In the present convention, a momentum-squeezed vacuum has

$$\operatorname{Var}(q) = \frac{e^{2r}}{2}, \qquad \operatorname{Var}(p) = \frac{e^{-2r}}{2}.$$

If squeezing is quoted relative to vacuum variance, its positive reduction in decibels is

$$s_{\rm dB} = -10\log_{10}(e^{-2r}) = \frac{20r}{\ln 10}.$$

The sign and reference convention must accompany every decibel value. Squeezed Light owns optical generation, calibration, and realistic loss limits.

Pure loss with transmissivity $\eta$ acts on $m$ modes as

$$X = \sqrt{\eta}\,I, \qquad Y = \frac{1-\eta}{2}I,$$

so that

$$V \longmapsto \eta V+\frac{1-\eta}{2}I.$$

The vacuum term is required by complete positivity. Loss can erase squeezing, reduce displacement, mix the state, and change a cluster’s logical channel. An abstract ideal calculation must mark omitted loss and detector effects as assumptions, not as zero-cost verified components.

A useful model-level ledger reports:

• input, output, ancilla, and cluster-mode counts;
• mean energy and any declared peak-energy or cutoff constraint;
• squeezing strengths and conventions;
• Gaussian gate and channel counts, weighted edges, and parallel depth;
• non-Gaussian state, gate, or detector counts, strengths, and preparation yields;
• continuous measurement records, bins, resolution, feedforward gains, and adaptive depth;
• loss, noise, postselection, repetitions, decoder, and error metric; and
• numerical precision, convergence evidence, comparator boundary, and every explicitly omitted physical cost.

Continuous-Variable Platforms owns source rates, optical and microwave architectures, delay and switching costs, calibration, and device evidence. Photonic Qubits owns discrete photonic encodings and their source-to-detector accounting. Algorithmic Benchmarking owns matched end-to-end comparators and advantage claims. A symplectic gate count is not an end-to-end runtime.

Worked Audit: A Two-Mode Finite-Squeezing Cluster

This audit prepares a finite-energy two-node graph state and checks its covariance, nullifiers, energy, and one measurable verification event. It is a resource-state calculation, not a claim about a logical algorithm or device.

Computational task, input family, and licensed claim. Prepare the unit-weight two-node continuous-variable graph from two independent momentum-squeezed vacua. License exact Gaussian state propagation, finite nullifier noise, purity and energy checks, and the probability of one declared joint nullifier window for this fixed fixture.

Modes, quadrature convention, operator domains, and energy constraint. Use two modes ordered as $R=(q_1,p_1,q_2,p_2)^{\mathsf T}$, with $[q_j,p_k]=i\delta_{jk}$ and vacuum covariance $I_4/2$. All states have finite second moments. The nullifiers below are commuting self-adjoint linear combinations on the shared oscillator domain.

Initial state, ancillas, preparation, and encoding. Prepare two independent momentum-squeezed vacua with

$$r = \frac{\ln2}{2}.$$

Their mean and covariance are

$$d_{\rm in} = 0, \qquad V_{\rm in} = \operatorname{diag} \left( 1,\frac14,1,\frac14 \right).$$

There are no ancillas beyond the two resource modes and no logical-qubit encoding.

Gaussian operations, channels, and affine symplectic map. Apply one ideal unit-weight controlled phase,

$$U_{12} = e^{iq_1q_2}.$$

Its symplectic matrix in the declared coordinate order is

$$S = \begin{pmatrix} 1&0&0&0\\ 0&1&1&0\\ 0&0&1&0\\ 1&0&0&1 \end{pmatrix}.$$

Therefore $d_{\rm out}=0$ and

$$\begin{aligned} V_{\rm out} &= SV_{\rm in}S^{\mathsf T} \\ &= \begin{pmatrix} 1&0&0&1\\ 0&\frac54&1&0\\ 0&1&1&0\\ 1&0&0&\frac54 \end{pmatrix}. \end{aligned}$$

No Gaussian noise channel or displacement is inserted.

Non-Gaussian states, gates, measurements, and universality role. Every preparation and entangling operation is Gaussian. No non-Gaussian resource is present, so this state alone does not supply universal continuous-variable computation.

Measurements, classical records, feedforward, and branch convention. The preparation has no measurement or feedforward. On independent verification copies, jointly measure

$$\delta_1 = p_1^{\rm(out)}-q_2^{\rm(out)}, \qquad \delta_2 = p_2^{\rm(out)}-q_1^{\rm(out)}.$$

The observables commute. Record both real outcomes and accept the measurable rectangle

$$|\delta_1| \leq \frac12, \qquad |\delta_2| \leq \frac12.$$

The verification measurement consumes its copy. It is not postselection inside the resource-preparation channel, and no probability is assigned to one exact real outcome.

Output system, decoder, success event, and error metric. The quantum output is the two-mode state. The Heisenberg relations give

$$\delta_1 = p_1^{\rm(in)}, \qquad \delta_2 = p_2^{\rm(in)}.$$

Thus the nullifiers are independent centered normals with

$$\operatorname{Var}(\delta_1) = \operatorname{Var}(\delta_2) = \frac14, \qquad \operatorname{Cov}(\delta_1,\delta_2) = 0.$$

The declared window probability is

$$\begin{aligned} P_{\rm win} &= \operatorname{erf} \left( \frac{1}{\sqrt2} \right)^2 \\ &= 0.4660649426743922670\ldots. \end{aligned}$$

The two nullifier variances and this bin probability are verification metrics, not a logical success probability or threshold.

Finite squeezing, loss, noise, precision, cutoff, and resource ledger. The ideal ledger contains two modes, two squeezers at $r=\ln2/2$, one unit-weight entangler, no non-Gaussian resource, no loss channel, no feedforward, and one optional joint verification read per tested copy. The input and output mean photon numbers are

$$\bar n_{\rm in} = \frac14, \qquad \bar n_{\rm out} = \frac54.$$

No Fock cutoff is used. Physical preparation yield, gate time, detector efficiency, calibration uncertainty, and experimental repetition count are not applicable to this ideal model fixture.

Verification data, uncertainty, and reproducibility. Direct multiplication checks

$$S\Omega S^{\mathsf T} = \Omega, \qquad \det V_{\rm out} = \frac1{16}.$$

The two symplectic eigenvalues are

$$\nu_1 = \nu_2 = \frac12,$$

as required for a pure two-mode Gaussian state. Transforming the nullifiers back to the input and integrating two independent normal densities reproduces $P_{\rm win}$. All values are analytic, so this record has no empirical or numerical uncertainty.

Conclusion, stopping point, and canonical handoff. The finite-squeezing two-node graph state is internally consistent, pure, finite energy, and has nullifier variance $1/4$. The audit establishes no universal program, hardware source, detector model, threshold, or advantage. Covariance theory stops with its prerequisite owner; physical generation stops with Continuous-Variable Platforms; qubit flow and gflow stop with Measurement-Based Quantum Computation; encoded protection stops with GKP Codes.

Worked Audit: A Cubic-Phase Gate outside the Gaussian Sector

This audit applies one ideal cubic-phase unitary to the vacuum. It uses a centered third moment and the covariance-purity mismatch to prove that the output has left Gaussian closure.

Computational task, input family, and licensed claim. Apply $U_3=\exp(iq^3/3)$ to one vacuum mode. License the exact output moments, an analytic non-Gaussian witness, its mean energy, and a state-level cutoff bound for this fixed input. Do not license a physical gate, compiler, complete universal set, or protected computation.

Modes, quadrature convention, operator domains, and energy constraint. Use one mode with $[q,p]=i$ and vacuum variances $1/2$. The vacuum wavefunction and its cubic-phase image lie in the Schwartz space, which supplies a common domain for every polynomial moment used below. The output has finite mean photon number.

Initial state, ancillas, preparation, and encoding. Prepare the normalized vacuum $|0\rangle$. There are no ancillas, displacements, squeezers, or logical-qubit encoding.

Gaussian operations, channels, and affine symplectic map. No nontrivial Gaussian unitary or channel is applied. The identity affine data describe only the absent Gaussian part; they cannot represent the nonlinear transformed momentum or characterize the output state.

Non-Gaussian states, gates, measurements, and universality role. Apply

$$U_3 = \exp \left( \frac{i}{3}q^3 \right).$$

The exact Heisenberg action is

$$U_3^\dagger qU_3 = q, \qquad U_3^\dagger pU_3 = p+q^2.$$

This gate leaves Gaussian closure. One formal cubic gate does not by itself specify compatible Gaussian controls, sign or inverse access, approximation accuracy, compilation cost, implementation, or fault tolerance.

Measurements, classical records, feedforward, and branch convention. There is no intermediate measurement, record, branch, postselection, or feedforward. The final moments are analytic expectation values. An empirical test would require separately declared quadrature measurements, copies, estimators, and uncertainty.

Output system, decoder, success event, and error metric. The output is the single transformed mode. Its position-space wavefunction is

$$\psi_{\rm out}(q) = \pi^{-1/4} \exp \left( -\frac{q^2}{2} + \frac{iq^3}{3} \right).$$

Direct evaluation gives

$$\langle q\rangle = 0, \qquad \langle p\rangle = \frac12, \qquad \langle q^2\rangle = \frac12, \qquad \langle p^2\rangle = \frac54,$$

and

$$\langle p^3\rangle = \frac{17}{8}, \qquad \left\langle \left( p-\frac12 \right)^3 \right\rangle = \frac12.$$

The nonzero centered third moment is the declared non-Gaussian witness. There is no binary decoder or success event.

Finite squeezing, loss, noise, precision, cutoff, and resource ledger. The ideal record uses one mode, one cubic-phase gate with coefficient $1/3$, no Gaussian gate beyond identity, no ancilla, no loss or noise channel, and no measurement branch. Its mean photon number is

$$\bar n = \frac38.$$

For a cutoff retaining number states $0,\ldots,N-1$,

$$\operatorname{Tr} \left[ (I-\Pi_{<N})\rho_{\rm out} \right] \leq \frac{3}{8N}.$$

This is not a gate or channel approximation bound. Physical cubic-state yield, a finite peak-energy cap, gate time, coefficient precision cost, and protected overhead are not applicable because the audit licenses no implementation.

Verification data, uncertainty, and reproducibility. The exact first two moments give

$$d = \begin{pmatrix} 0\\[2pt] 1/2 \end{pmatrix}, \qquad V = \begin{pmatrix} 1/2&0\\ 0&1 \end{pmatrix}.$$

Hence

$$\det V = \frac12, \qquad \sqrt{\det V} = \frac1{\sqrt2}.$$

The output is pure because a unitary maps the vacuum to it. A one-mode pure Gaussian state would instead have $\det V=1/4$, so a Gaussian state with these first two moments would be mixed. This covariance-purity mismatch and the centered third moment independently expose the failure of a Gaussian description. All values are analytic and have no empirical uncertainty.

Conclusion, stopping point, and canonical handoff. The ideal cubic phase produces a pure non-Gaussian state with $\bar n=3/8$, and its first two moments are not a complete description. The audit establishes no source, detector, efficient compilation, threshold, or advantage. Implementation stops with Continuous-Variable Platforms, general resource monotones with Resource Theories, and encoded protection with GKP Codes.

Common Failure Modes and Canonical Handoffs

Combining conventions without translating them. Vacuum covariance can be $I/4$, $I/2$, or $I$ in common conventions. Translate quadratures, means, covariances, displacements, noise, gate coefficients, and the vacuum reference together before comparing squeezing or energy.

Treating generalized eigenstates as prepared states. Exact quadrature eigenvectors, zero-width nullifiers, infinite squeezing, and ideal grid combs are distributional limits. A physical or numerical claim needs finite energy, finite squeezing, a cutoff or convergence sequence, and an error metric.

Reversing affine composition. Matrices multiply in physical time order, and a later displacement is not generally equivalent to an earlier one. Write the action on $R$ before multiplying matrices.

Assigning probability to one continuous outcome. A density value is not an event probability. State a measurable bin, integrate the density, and say whether the branch is selected, forgotten, coarse-grained, or postselected.

Calling Gaussian quantum mechanics classical. Gaussian mode states can be entangled and useful. Efficient moment propagation under a finite-description model is a conditional simulation result, not a denial of quantumness or a statement about every detector and representation.

Calling any non-Gaussian element universal. Photon counting, a cubic gate, or a non-Gaussian state breaks closure, but universality additionally needs compatible controls, domains, signs or inverses, approximation, precision, and scaling. Universality still supplies neither fault tolerance nor advantage.

Using a finite-dimensional theorem without an energy contract. Unbounded generators and infinite Hilbert spaces require a finite-energy input set, common domains, an accuracy metric, and convergence evidence. Full-space operator-norm approximation and an unqualified finite-dimensional synthesis bound are not available by default.

Turning a cutoff tail into algorithm error. The mean-energy inequality bounds discarded state weight. A gate, channel, observable, or decoded-output claim needs an additional continuity or convergence argument.

Replacing qubit MBQC by real-valued labels. CV clusters have quadrature nullifiers, continuous records, displacement frames, and finite-squeezing noise. Qubit open graphs instead have Pauli byproducts, flow, and gflow. The two owners should be linked, not conflated.

Equating a mode with a protected logical qubit. Direct mode computation, bosonic encoding, grid-state recovery, and fault-tolerant logical gates are different layers. The GKP and fault-tolerance owners retain the latter contracts.

Promoting an ideal map to hardware evidence. A matrix does not specify a source rate, nonlinear interaction, loss budget, detector resolution, feedforward latency, calibration, or uncertainty. Those questions stop with Continuous-Variable Platforms and the relevant measurement owners; matched performance and advantage stop with Algorithmic Benchmarking.

Exercises

Exercise 1: Translate a Quadrature Convention

An author uses $[Q,P]=i/2$, vacuum covariance $I_2/4$, and reports

$$d_{QP} = \begin{pmatrix} 1\\ -1/2 \end{pmatrix}, \qquad V_{QP} = \operatorname{diag} \left( \frac14,1 \right).$$

Translate the state into this page’s convention, compute its mean photon number in both conventions, and identify one convention-invariant squeezing quantity.

Solution

The conventions are related by

$$q = \sqrt2\,Q, \qquad p = \sqrt2\,P.$$

First moments scale once and covariances twice:

$$d = \sqrt2\,d_{QP} = \begin{pmatrix} \sqrt2\\ -1/\sqrt2 \end{pmatrix}, \qquad V = 2V_{QP} = \operatorname{diag} \left( \frac12,2 \right).$$

The vacuum covariance correspondingly becomes $I_2/2$. In the first convention,

$$\bar n = \operatorname{Tr}V_{QP} + d_{QP}^{\mathsf T}d_{QP} - \frac12 = 2.$$

In the page convention,

$$\bar n = \frac{ \operatorname{Tr}V+d^{\mathsf T}d-1 }{2} = 2.$$

The ratio of a quadrature variance to the vacuum variance is invariant. Therefore a squeezing value in decibels is also invariant after the state and vacuum reference are translated together. A coefficient multiplying an unbounded gate generator would need its own dimensional translation and is not inferred here.

Exercise 2: Compose an Affine Gaussian Circuit

Begin in vacuum. Apply a $q$-squeezer with $r=\ln2$, then rotate by $\pi/4$ using

$$R\left(\frac\pi4\right) = \frac1{\sqrt2} \begin{pmatrix} 1&-1\\ 1&1 \end{pmatrix},$$

and finally displace by $(1,-1/2)^{\mathsf T}$. Find the final mean, covariance, purity determinant, mean photon number, and probability that an ideal $q$-homodyne result lies in $[0,2]$.

Solution

The squeeze matrix is

$$S_{\rm sq} = \operatorname{diag} \left( \frac12,2 \right).$$

The total linear map is $R(\pi/4)S_{\rm sq}$ because the squeeze occurs first. The input mean is zero, and the final displacement occurs last, so

$$d = \begin{pmatrix} 1\\ -1/2 \end{pmatrix}.$$

Starting from $V_0=I_2/2$,

$$\begin{aligned} V &= R\left(\frac\pi4\right) S_{\rm sq}V_0S_{\rm sq}^{\mathsf T} R\left(\frac\pi4\right)^{\mathsf T} \\ &= \frac1{16} \begin{pmatrix} 17&-15\\ -15&17 \end{pmatrix}. \end{aligned}$$

Its determinant is $1/4$, as expected for a pure one-mode Gaussian state. The energy is

$$\bar n = \frac{ \operatorname{Tr}V+d^{\mathsf T}d-1 }{2} = \frac{19}{16}.$$

The $q$ outcome is normal with mean $1$ and variance $17/16$. The interval is symmetric about the mean, so

$$\begin{aligned} P(0\leq q\leq2) &= \operatorname{erf} \left( \frac{1}{\sqrt{2(17/16)}} \right) \\ &= \operatorname{erf} \left( \frac4{\sqrt{34}} \right) \\ &= 0.6680245329172630578\ldots. \end{aligned}$$

Reversing the matrix product would describe a different physical circuit.

Exercise 3: Derive Three-Mode Line-Cluster Nullifiers

Prepare three independent momentum-squeezed vacua with $r=\ln2/2$. Apply unit-weight controlled phases on edges $(1,2)$ and $(2,3)$. Derive the three nullifiers, their covariance, the input and output mean photon numbers, and the abstract preparation ledger.

Solution

The two entanglers commute because they are functions of $q$ quadratures. Their Heisenberg action is

$$\begin{aligned} p_1' &= p_1+q_2, \\ p_2' &= p_2+q_1+q_3, \\ p_3' &= p_3+q_2, \end{aligned}$$

while every $q_j$ is unchanged. Hence

$$\begin{aligned} (\delta_1,\delta_2,\delta_3) &= (p_1'-q_2',\, p_2'-q_1'-q_3',\, p_3'-q_2') \\ &= (p_1,p_2,p_3). \end{aligned}$$

The input momenta are independent, so

$$\operatorname{Var}(\delta_j) = \frac14, \qquad \operatorname{Cov}(\delta_j,\delta_k) = 0 \quad (j\neq k).$$

Each input mode has mean photon number $1/8$, giving

$$\bar n_{\rm in} = \frac38.$$

For these inputs, each unit edge adds one mean photon. With two edges,

$$\bar n_{\rm cluster} = \frac38+2 = \frac{19}{8}.$$

The abstract preparation uses three modes, three squeezers, and two controlled-phase gates. It has no non-Gaussian resource and licenses no universal-computation claim.

Exercise 4: Propagate a Squeezed State through Pure Loss

A zero-mean one-mode state has $V_{\rm in}=\operatorname{diag}(1/8,2)$. Send it through a vacuum pure-loss channel with transmissivity $\eta=3/4$. Find the output covariance, determinant, mean photon number, and remaining $q$ squeezing.

Solution

For vacuum pure loss,

$$X = \sqrt{\eta}\,I_2, \qquad Y = \frac{1-\eta}{2}I_2.$$

Therefore

$$\begin{aligned} V_{\rm out} &= \frac34 \operatorname{diag} \left( \frac18,2 \right) + \frac18 I_2 \\ &= \operatorname{diag} \left( \frac7{32},\frac{13}{8} \right). \end{aligned}$$

The determinant is

$$\det V_{\rm out} = \frac{91}{256} > \frac14,$$

so the output is mixed. Its mean photon number is

$$\bar n_{\rm out} = \frac{ \operatorname{Tr}V_{\rm out}-1 }{2} = \frac{27}{64} = \frac34 \left( \frac9{16} \right).$$

The final $q$ variance $7/32$ remains below the vacuum value $1/2$, so some squeezing survives. Gaussian Channels retains the general attenuator classification and dilation.

Exercise 5: Separate Gaussian Closure from Universality

Classify the following three model claims:

1. Gaussian input states, Gaussian gates and channels, homodyne, and affine feedforward.
2. The same model with photon counting added.
3. Gaussian controls plus an available cubic interaction.

State what each modification licenses and what it does not.

Solution

The first model remains in Gaussian closure. Under a polynomial-size finite real description and polynomial precision, its means, covariances, and Gaussian outcome densities admit efficient classical propagation. This does not make the states classical, but the declared operation set is not universal for arbitrary continuous-variable transformations.

Photon counting is a non-Gaussian measurement, so the second model leaves Gaussian closure. That fact alone does not give a universal gate set, efficient state preparation, useful postselection rate, fault tolerance, or an advantage result.

The third model meets the ideal algebraic premise used to generate polynomial Hamiltonians only after the record also specifies available strengths, signs or inverses, common domains, energy-limited inputs, approximation accuracy, continuous-parameter precision, and commutator construction costs. Even then, algebraic universality does not supply an efficient compiler, physical implementation, or fault tolerance.

Exercise 6: Bound a Fock-Space Truncation from Mean Energy

Use Audit B’s output energy $\bar n=3/8$. Bound the probability outside a cutoff retaining levels $0$ through $63$, and find the smallest cutoff parameter guaranteed by the same inequality to make the tail at most $10^{-3}$. Explain the stopping point.

Solution

Retaining levels $0,\ldots,63$ means $N=64$. Markov’s bound gives

$$\Pr(n\geq64) \leq \frac{3/8}{64} = \frac3{512} = 0.005859375.$$

For a guarantee of at most $10^{-3}$,

$$\frac{3}{8N} \leq 10^{-3},$$

so

$$N \geq 375.$$

The bound may be loose. It controls state weight outside the retained Fock sector, not a gate, channel, observable, decoded-output, or algorithm error without an additional continuity or convergence argument.

Exercise 7: Count a Twenty-Five-Mode Cluster Wire

Prepare a path of twenty-five independent momentum-squeezed vacua with $r=\ln2/2$, join nearest neighbors by unit-weight controlled phases, measure the first twenty-four modes by homodyne, and retain one output. Count the model-level resources and compute the energy before and after entangling.

Solution

The path has twenty-five vertices and twenty-four edges. Its preparation ledger contains twenty-five modes and squeezers and twenty-four controlled-phase gates. A path’s edges can be colored with two colors, so the entangling preparation has two gate layers if disjoint gates are parallel and shared-mode gates are not. The measurement stage consumes twenty-four modes, produces twenty-four real records, and leaves one quantum output. A fully sequential dependency declaration permits worst-case adaptive measurement depth twenty-four.

Each squeezed input has mean photon number $1/8$, hence

$$\bar n_{\rm in} = \frac{25}{8}.$$

Every unit path edge adds one mean photon for these input variances. Therefore

$$\bar n_{\rm cluster} = \frac{25}{8}+24 = \frac{217}{8}.$$

There is no non-Gaussian resource. Because no injected logical input, homodyne angles, or feedforward gains were given, these counts license only a Gaussian path-resource and measurement ledger, not a particular logical channel or fidelity. Physical loss, detector bandwidth, feedforward latency, and hardware parallelism are not applicable to the ideal count.

Exercise 8: Complete a Ten-Field Gaussian Program Record

Start in vacuum, squeeze $q$ with $r=\ln2$, displace $q$ by $+1$, measure $q$ ideally, and output the bit $1$ exactly when the result is nonnegative. Complete the full record and verify its output law.

Solution

Computational task, input family, and licensed claim. Prepare one displaced squeezed Gaussian mode from a fixed vacuum input and return one thresholded classical bit. License only the exact ideal output law, its error relative to target bit $1$, and the abstract resource ledger for this fixture.

Modes, quadrature convention, operator domains, and energy constraint. Use one mode with $[q,p]=i$, vacuum covariance $I_2/2$, and ideal homodyne outcomes $x\in\mathbb R$. The premeasurement state is normalizable and finite energy; the generalized homodyne outcome is interpreted through measurable bins.

Initial state, ancillas, preparation, and encoding. Start in vacuum, apply a $q$-squeezer with $r=\ln2$, and then displace $q$ by $+1$. There are no ancillas and no logical-qubit encoding.

Gaussian operations, channels, and affine symplectic map. The squeeze and following displacement have

$$S = \operatorname{diag} \left( \frac12,2 \right), \qquad s = \begin{pmatrix} 1\\ 0 \end{pmatrix}.$$

There is no noisy channel. In physical order,

$$d = \begin{pmatrix} 1\\ 0 \end{pmatrix}, \qquad V = \operatorname{diag} \left( \frac18,2 \right).$$

Non-Gaussian states, gates, measurements, and universality role. No non-Gaussian resource is used. The program remains in Gaussian closure and is efficiently tractable under the exact finite-description model. It licenses no universality or computational advantage.

Measurements, classical records, feedforward, and branch convention. Perform one ideal $q$ homodyne. The real record obeys

$$x \sim \mathcal N \left( 1,\frac18 \right).$$

There is no feedforward or postselection. The branches are the measurable bins $x\geq0$ and $x<0$, not individual point outcomes.

Output system, decoder, success event, and error metric. The output is the classical bit

$$b = \mathbf 1[x\geq0].$$

The accepted event is $b=1$, with

$$\begin{aligned} P(b=1) &= \Phi(2\sqrt2) \\ &= \frac{1+\operatorname{erf}(2)}{2} \\ &= 0.9976611325094763671\ldots. \end{aligned}$$

The reject probability is

$$P(b=0) = 0.0023388674905236329\ldots.$$

It is the program-error probability relative to target bit $1$, equivalently the total-variation distance between the output Bernoulli law and a point mass at $1$.

Finite squeezing, loss, noise, precision, cutoff, and resource ledger. The record uses one mode, one $r=\ln2$ squeezer, one displacement, one ideal homodyne, one real record, no non-Gaussian resource, no loss, and no postselection. The squeezing reduction relative to vacuum variance is

$$\frac{20\ln2}{\ln10} = 6.020599913\ldots\,\mathrm{dB}.$$

Immediately before readout,

$$\bar n = \frac{17}{16}.$$

An optional cutoff retaining $0,\ldots,15$ has only the loose guarantee

$$\Pr(n\geq16) \leq \frac{17}{256}.$$

A physical energy cap, preparation yield, duration, detector resolution, and experimental repetition count are not applicable to this ideal fixture. Its comparator is exact Gaussian propagation, not a speedup baseline.

Verification data, uncertainty, and reproducibility. Recompute the affine data in physical order, check $\det V=1/4$, evaluate the normal tail as $(1+\operatorname{erf}2)/2$, verify that the two bins normalize, reproduce $17/16$, and keep the optional cutoff tail distinct from program error. All values are analytic and have no empirical uncertainty.

Conclusion, stopping point, and canonical handoff. The program has a complete ideal Gaussian record and the stated output probability. It establishes no non-Gaussian universality, grid encoding, platform performance, detector fidelity, fault tolerance, or advantage. Gaussian-channel variants stop with Gaussian Channels; physical preparation and readout stop with Continuous-Variable Platforms.

References

• S. D. Bartlett and B. C. Sanders, “Universal continuous-variable quantum computation: Requirement of optical nonlinearity for photon counting,” Physical Review A 65, 042304 (2002), doi:10.1103/PhysRevA.65.042304.
• S. D. Bartlett, B. C. Sanders, S. L. Braunstein, and K. Nemoto, “Efficient Classical Simulation of Continuous Variable Quantum Information Processes,” Physical Review Letters 88, 097904 (2002), doi:10.1103/PhysRevLett.88.097904.
• S. L. Braunstein and P. van Loock, “Quantum Information with Continuous Variables,” Reviews of Modern Physics 77, 513–577 (2005), doi:10.1103/RevModPhys.77.513.
• 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.
• M. Gu, C. Weedbrook, N. C. Menicucci, T. C. Ralph, and P. van Loock, “Quantum computing with continuous-variable clusters,” Physical Review A 79, 062318 (2009), doi:10.1103/PhysRevA.79.062318.
• S. Lloyd and S. L. Braunstein, “Quantum Computation over Continuous Variables,” Physical Review Letters 82, 1784–1787 (1999), doi:10.1103/PhysRevLett.82.1784.
• A. Mari and J. Eisert, “Positive Wigner Functions Render Classical Simulation of Quantum Computation Efficient,” Physical Review Letters 109, 230503 (2012), doi:10.1103/PhysRevLett.109.230503.
• N. C. Menicucci, “Fault-Tolerant Measurement-Based Quantum Computing with Continuous-Variable Cluster States,” Physical Review Letters 112, 120504 (2014), doi:10.1103/PhysRevLett.112.120504.
• N. C. Menicucci, P. van Loock, M. Gu, C. Weedbrook, T. C. Ralph, and M. A. Nielsen, “Universal Quantum Computation with Continuous-Variable Cluster States,” Physical Review Letters 97, 110501 (2006), doi:10.1103/PhysRevLett.97.110501.
• R. Takagi and Q. Zhuang, “Convex resource theory of non-Gaussianity,” Physical Review A 97, 062337 (2018), doi:10.1103/PhysRevA.97.062337.
• P. van Loock, C. Weedbrook, and M. Gu, “Building Gaussian cluster states by linear optics,” Physical Review A 76, 032321 (2007), doi:10.1103/PhysRevA.76.032321.
• C. Weedbrook et al., “Gaussian quantum information,” Reviews of Modern Physics 84, 621–669 (2012), doi:10.1103/RevModPhys.84.621.