Continuous-Variable Quantum Computing

In every other chapter of this curriculum, a qubit has been a discrete two-level system. A superconducting transmon lives in its two lowest energy eigenstates. A trapped ion lives in two hyperfine levels. A photon in polarisation encoding lives in |H\rangle vs |V\rangle. Even in measurement-based photonics, the logical basis is discrete — photon present here versus photon present there. The whole edifice of quantum gates, algorithms, and error correction has been built on the Hilbert space \mathbb{C}^2 of a single qubit and its tensor products.

There is a completely different way to build a quantum computer, and it has been pursued in parallel with the qubit approach for over two decades.

Forget the qubit. Start instead with a single harmonic oscillator — a mode of the electromagnetic field in a resonator, a mechanical oscillator, a collective atomic excitation, anything whose Hamiltonian looks like \hat{H} = \tfrac{1}{2}(\hat{p}^2 + \hat{x}^2). You met this system already in class 12 physics. Its state is a wavefunction \psi(x) on the real line; its position operator \hat{x} has continuous spectrum (-\infty, \infty); its momentum operator \hat{p} likewise. These are called the quadratures of the mode — two continuous-valued observables linked by the canonical commutation [\hat{x}, \hat{p}] = i (we use \hbar = 1 throughout).

Now make two bold choices.

• Encode quantum information directly in the wavefunction, not in a two-level subspace of it. A "state" in continuous-variable (CV) quantum computing is a full oscillator wavefunction — Gaussian, squeezed, superposed, entangled with another mode's wavefunction.
• Gates act on the quadratures, not on discrete spin states. A displacement slides \psi(x) sideways; a squeezing stretches it along \hat{x} and compresses along \hat{p}; a rotation mixes \hat{x} and \hat{p} like a phase-space spinner.

This is continuous-variable quantum computing. It is not a curiosity. It is a mature, experimentally active paradigm pioneered by Seth Lloyd, Samuel Braunstein, Peter van Loock, and others in the late 1990s, and it is the architecture Xanadu (Canada) uses for their photonic quantum processors. It even has its own universality theorem, its own error-correcting codes, and its own distinctive failure mode (Gaussian simulability).

This chapter is the CV playbook. Quadratures. Gaussian and non-Gaussian. GKP. Xanadu.

The quadratures of a bosonic mode

A single bosonic mode — for example, one mode of the electromagnetic field in a laser cavity at frequency \omega — is described by the annihilation operator \hat{a} and creation operator \hat{a}^\dagger, with [\hat{a}, \hat{a}^\dagger] = 1. The number operator \hat{n} = \hat{a}^\dagger \hat{a} has integer eigenvalues 0, 1, 2, \ldots — these are the Fock states |n\rangle, each corresponding to exactly n photons in the mode.

The quadrature operators are defined as

Why this combination: inverting, \hat{a} = (\hat{x} + i\hat{p})/\sqrt{2} and \hat{a}^\dagger = (\hat{x} - i\hat{p})/\sqrt{2}. The commutator [\hat{x}, \hat{p}] = \tfrac{1}{2}[\hat{a} + \hat{a}^\dagger, -i(\hat{a} - \hat{a}^\dagger)] = -i \cdot \tfrac{1}{2}([\hat{a}, -\hat{a}^\dagger] + [\hat{a}^\dagger, \hat{a}]) = -i \cdot \tfrac{1}{2}(-1 - 1) = i as required. The \sqrt{2} factors are convention.

Think of \hat{x} as "position" and \hat{p} as "momentum" of an effective particle — the language inherited from mechanical oscillators. For a photonic mode, "position" is actually the in-phase component of the electric field, and "momentum" is the out-of-phase (quadrature) component. The physics is identical because mathematically they are the same oscillator.

Now the basic states.

• Vacuum |0\rangle: zero photons, \langle \hat{n} \rangle = 0. Its wavefunction is the ground state of the harmonic oscillator, a Gaussian centred at x = 0: \psi_0(x) = \pi^{-1/4} e^{-x^2/2}. In phase space it sits at the origin with equal spread in x and p: \Delta x = \Delta p = 1/\sqrt{2}, saturating the uncertainty relation.
• Coherent state |\alpha\rangle: a Gaussian wavefunction displaced away from the origin by a complex number \alpha = (\alpha_x + i\alpha_p)/\sqrt{2}. Same shape as the vacuum, just translated: \psi_\alpha(x) = \pi^{-1/4} e^{-(x - \alpha_x)^2/2} e^{i \alpha_p x} (up to a phase). This is what a laser produces.
• Squeezed state |r\rangle: a Gaussian whose width along \hat{x} is reduced by factor e^{-r} and along \hat{p} increased by e^r — uncertainty relation still saturated, variance redistributed. r is the squeezing parameter; r > 0 squeezes the x-quadrature. The wavefunction is \psi_r(x) = (e^r/\pi)^{1/4} e^{-e^{2r} x^2 / 2}.
• Fock state |n\rangle with n \geq 1: a non-Gaussian state with wavefunction proportional to H_n(x) e^{-x^2/2} where H_n is the Hermite polynomial. Rings of probability in phase space, definitely not a Gaussian.

The first three — vacuum, coherent, squeezed — are all Gaussian states: their Wigner functions (the phase-space quasi-probability distributions) are 2D Gaussians. The Fock states beyond n = 0 and any superposition that is not Gaussian-weighted are non-Gaussian states.

This Gaussian / non-Gaussian distinction is not a matter of aesthetics. It is the deep structural divide of CV quantum computing.

Gaussian operations — the linear-optical toolkit

A Gaussian operation is a unitary \hat{U} that maps Gaussian states to Gaussian states. Equivalently, \hat{U} maps the quadrature operators (\hat{x}, \hat{p}) to linear combinations of themselves, plus a displacement. In phase-space language, Gaussian operations are exactly the affine symplectic transformations of the quadrature plane.

There are three elementary Gaussian operations, and every CV Gaussian circuit is built from them.

Displacement

Shifts the quadrature by a complex number \alpha — slides \psi(x) rightward by \alpha_x \sqrt{2} and multiplies it by a phase e^{i \alpha_p x \sqrt{2}}. Physically, \hat{D}(\alpha) applied to the vacuum produces a coherent state: \hat{D}(\alpha) |0\rangle = |\alpha\rangle. In the lab, displacement is implemented by mixing the mode with a strong coherent-state reference at a highly-asymmetric beam splitter.

Squeezing

Stretches the quadrature plane — compresses \hat{x} by e^{-r} while stretching \hat{p} by e^r (or vice versa for negative r). In the lab, squeezing is generated by parametric amplification: pump a non-linear crystal (like \chi^{(2)} PPKTP) with an intense laser beam at frequency 2\omega, and the crystal non-linearity couples the vacuum mode at frequency \omega to a two-photon process that produces squeezed light.

The current-record single-mode squeezing is about -15 dB below vacuum noise, achieved by the Schnabel group in Hannover around 2016. "Decibels below vacuum" means e^{-2r} = 10^{-15/10}, so r \approx 1.7.

Beam splitter (two-mode Gaussian)

Mixes two modes — a 50/50 beam splitter (\theta = \pi/4) sends input modes \hat{a}, \hat{b} to (\hat{a} + \hat{b})/\sqrt{2}, (-\hat{a} + \hat{b})/\sqrt{2}. Together with phase shifters, beam splitters generate any two-mode Gaussian unitary. A network of beam splitters and phase shifters realises any multi-mode passive linear-optical transformation.

Together: all Gaussian unitaries

Theorem (Bloch–Messiah / Euler decomposition). Any multi-mode Gaussian unitary can be written as a product of

• a multi-mode passive linear-optical network (beam splitters + phase shifters),
• single-mode squeezers on each mode,
• another multi-mode passive linear-optical network,
• displacements on each mode.

This is an extraordinary structural result. It means that the entire Gaussian toolkit is linear optics + single-mode squeezing + displacements — all building blocks experimentally available in any optics lab. The expressive power of Gaussian CV is exactly the power of these three ingredients combined.

Homodyne detection — the Gaussian measurement

A CV circuit must end with a measurement. The standard CV measurement is homodyne detection: interfere the mode of interest with a strong coherent-state local oscillator at a balanced beam splitter, photodetect both output ports, and subtract the counts. The output is a classical real number x \in (-\infty, \infty) with probability density |\psi(x)|^2 — a direct measurement of \hat{x} (or, by choosing the local oscillator phase, any linear combination \cos\phi \, \hat{x} + \sin\phi \, \hat{p}).

Homodyne is fast, efficient, and Gaussian: if the pre-measurement state is Gaussian, the measurement outcome is drawn from a Gaussian distribution. Crucially, the outcome is continuous — no discretisation into 0 and 1 as in qubit measurement.

The Gaussian trap — why Gaussian alone is not enough

You have a full Gaussian toolkit. Squeezing, beam splitters, displacements, homodyne. You can entangle, you can process, you can measure. Is this enough for a universal quantum computer?

No. This is the central theorem of CV quantum computing.

Bartlett–Sanders–Braunstein–Nemoto theorem (2002). Any quantum circuit consisting of Gaussian state preparation + Gaussian unitary operations + homodyne measurement can be classically simulated in polynomial time.

The proof is computationally explicit. Gaussian states are described by a finite vector of means and a finite covariance matrix — 2n means and (2n)(2n+1)/2 covariance entries for n modes. Gaussian unitaries are symplectic matrices acting on this data — polynomial-size objects evolved in polynomial time. Homodyne measurements are Gaussian integrals, computable in closed form.

The classical simulation is not merely "there exists an algorithm" — it is a concrete, constructive procedure that tracks the Gaussian state's mean and covariance throughout the circuit. A laptop running Xanadu's Strawberry Fields simulator can simulate millions of Gaussian modes exactly.

The conclusion for CV quantum computing: Gaussian is to CV what Clifford is to qubits. The analogy is deep:

• Clifford gates on qubits (Hadamard, phase, CNOT) compose the stabilizer formalism, which is classically simulable by the Gottesman–Knill theorem.
• Gaussian gates on modes (squeezing, beam splitter, displacement) compose the phase-space formalism, which is classically simulable by Bartlett–Sanders–Braunstein–Nemoto.

And the remedy is analogous. For Clifford qubits, you inject a non-Clifford magic state (the T state) to reach universality. For Gaussian CV, you inject a non-Gaussian resource — either a non-Gaussian unitary (the cubic phase gate) or a non-Gaussian state (a Fock state, a cat state, a GKP state).

The non-Gaussian ingredient

Three standard ways to escape the Gaussian trap.

The cubic phase gate

The first fully general prescription is to augment the Gaussian set with the cubic phase gate:

This unitary is generated by a Hamiltonian cubic in \hat{x} — the first non-quadratic case after the Gaussians. Gaussian + cubic-phase is provably universal: Lloyd and Braunstein (1999) proved that any polynomial-in-quadrature Hamiltonian can be approximately implemented by alternating Gaussian unitaries with cubic-phase unitaries, up to arbitrary precision.

The catch: e^{i\gamma \hat{x}^3} is not realisable by any linear optical process. It requires a third-order optical non-linearity (\chi^{(3)}) strong enough at the single-photon level, which in conventional materials is many orders of magnitude too weak. Alternative implementations use measurement-induced non-linearities — consume a non-Gaussian resource state in a measurement-based protocol to enact an effective cubic phase gate. This is the route Xanadu and most experimental groups take.

Photon-number-resolving measurement

Instead of homodyne (continuous outcome, Gaussian-friendly), measure using a photon-number-resolving (PNR) detector — count how many photons arrive in a mode. The outcome is a discrete integer, and the conditional state on the other modes of an entangled CV state given a specific photon count is generically non-Gaussian.

PNR detection with high efficiency is hard but now routinely available: superconducting nanowire arrays (SNSPDs), transition-edge sensors (TES) at ~100 mK, and photon-number-resolving threshold detectors. A single PNR outcome in a Gaussian circuit breaks Gaussianity and, with the right architecture, delivers universality.

Non-Gaussian input states

Prepare a Fock state, a cat state (|\alpha\rangle + |-\alpha\rangle)/\mathcal{N}, or a GKP state, then run Gaussian operations on the Fock-plus-Gaussian input. Again, the non-Gaussianity of the input is injected into the circuit and lifts it out of the simulable regime.

This is the route for many hybrid and measurement-based CV proposals. Preparing high-quality non-Gaussian states is itself challenging — single photons have been generated at high efficiency but cat states at amplitude \alpha > 2 and GKP states with multiple clean peaks remain the cutting edge of experimental CV.

GKP encoding — a qubit inside a qumode

Now the payoff. With non-Gaussian states available, you can encode a discrete qubit inside a single bosonic mode, with built-in error correction against small displacement errors.

The GKP codewords

In 2001, Daniel Gottesman, Alexei Kitaev, and John Preskill proposed a remarkable code — now universally called the GKP code — for encoding a qubit into an oscillator. The ideal codewords are:

Read this: |\bar{0}\rangle is a sum of position eigenstates at every even multiple of \sqrt{\pi}; |\bar{1}\rangle is the same, but shifted — at every odd multiple. Equivalently, |\bar{0}\rangle is a Dirac comb of infinitely many delta functions spaced by 2\sqrt{\pi}, and |\bar{1}\rangle is the same comb shifted by \sqrt{\pi}.

Both are unnormalisable (infinite in extent) and unphysical in their ideal form, but approximate GKP states use finite-width Gaussian peaks with a Gaussian envelope — normalisable and realisable in the lab.

Logical gates from Gaussian operations

Here is the beautiful part. The logical Pauli operators on the GKP qubit are:

• \bar{X} = e^{-i\sqrt{\pi} \hat{p}} — a displacement by \sqrt{\pi} along \hat{x} — flips |\bar{0}\rangle \leftrightarrow |\bar{1}\rangle exactly because the codewords are shifts of each other.
• \bar{Z} = e^{i\sqrt{\pi} \hat{x}} — a momentum-displacement — multiplies the even-n terms by +1 and odd-n terms by -1, flipping the relative sign just as Pauli Z should.

These are Gaussian operations on the mode. The logical Clifford group on GKP qubits is implemented entirely by Gaussian operations (displacements, squeezing, beam splitters). Only the logical T gate and related non-Clifford pieces require genuine non-Gaussian operations.

The error-correction payoff

Suppose a small displacement error — a noise process that shifts the mode by (\Delta x, \Delta p) with |\Delta x|, |\Delta p| < \sqrt{\pi}/2. The GKP codewords are periodic with period 2\sqrt{\pi} in x and in p separately. A homodyne measurement of \hat{x} modulo \sqrt{\pi} tells you \Delta x \bmod \sqrt{\pi} (which, if small, equals \Delta x). Applying \hat{D}(-\Delta x) corrects it. The same trick handles \Delta p errors via homodyne of \hat{p}.

So small displacement errors — the dominant noise mode in photonic and microwave cavities — are directly correctable by measurement and feedback, using only Gaussian operations. The GKP code turns a continuous noise model into a discrete logical qubit with protection.

Experimental status

GKP state preparation has progressed rapidly.

• 2019 (Campagne-Ibarcq et al., Yale). First realisation of a logical GKP qubit in a microwave cavity coupled to a transmon, with a coherence time exceeding the underlying cavity lifetime by an error-correction-protected factor.
• 2020 (Flühmann et al., ETH). First GKP qubit in a trapped ion's motional mode (one ion, one oscillator).
• 2023–2024. Multiple groups (Yale, Google, AWS, ETH) demonstrate GKP-state preparation with better fidelity and longer protected coherence times, approaching the threshold regime for fault-tolerant operation.

Photonic GKP has been harder: photonic modes do not have a built-in qubit (like a transmon) to mediate the non-Gaussian state preparation. Xanadu and several academic groups have published schemes using photon-number-resolved measurement on one arm of an entangled two-mode squeezed state as a preparation method, but the resulting GKP states have been restricted to few-peak (low-n) approximations so far.

Xanadu and the photonic CV programme

Xanadu Quantum Technologies, based in Toronto, is the company most closely identified with CV quantum computing. Their architecture:

• Squeezed-light sources. Integrated silicon-nitride waveguides with on-chip parametric oscillators generate squeezed vacuum states on demand, with roughly 10 dB of squeezing at gigahertz rates.
• Time-domain multiplexing. Instead of running many parallel waveguides in space, Xanadu multiplexes qumodes in time — different temporal windows of the same waveguide carry different modes. This allows hundreds to thousands of modes to be processed with a fixed hardware footprint.
• Programmable beam splitter networks. Thermo-optic phase shifters and tunable Mach–Zehnder networks implement any Gaussian unitary on the multi-mode state.
• Photon-number-resolving detection. Superconducting transition-edge sensors at ~100 mK provide the non-Gaussian measurement ingredient.

Borealis — 2022

Borealis was Xanadu's first cloud-accessible CV machine. It ran Gaussian Boson Sampling (GBS) — a sampling task (not a general computation) that involves Gaussian squeezing, a programmable linear network, and PNR detection. In the 2022 Nature paper, Borealis was demonstrated with 216 squeezed modes and argued to sample from a distribution that would take a classical supercomputer more than 9,000 years to reproduce. This is a supremacy-type claim, not a general-purpose computation.

GBS is instructive: it is not universal — the output is just a sample from a specific distribution — but it uses exactly the ingredients you need for general CV QC (squeezing + Gaussian network + non-Gaussian measurement). Borealis is a proof that the ingredients work at scale.

The roadmap

Xanadu's announced roadmap (as of 2025) targets:

• Photonic GKP state preparation at useful fidelity (ongoing research; published protocols; few-photon demonstrations).
• GKP-encoded logical qubits in a time-multiplexed photonic architecture.
• Fault-tolerant photonic quantum computing via GKP codes combined with outer concatenated codes, targeting a million-qubit system in the long run.

The target is ambitious. No group has yet demonstrated fault-tolerant CV computation; GKP qubits are single-mode and have their own error syndromes; multi-mode protocols require extensive time-multiplexed concatenation. But the ingredients are there, and the community regards CV as one of the two or three credible roads to fault tolerance alongside surface-code qubit approaches and topological approaches.

Two worked examples

Common confusions

• "CV means continuous operations, like rotating by any angle." No — every QC platform allows continuous rotation angles (at the cost of calibration). "Continuous variable" refers to the spectrum of the observable, not the precision of the gate. \hat{x} has continuous spectrum (-\infty, \infty) whereas \hat{Z} on a qubit has spectrum \{+1, -1\}. The distinction is in what you measure, not how you rotate.
• "Gaussian operations are quantum." They are quantum in that they act on quantum states, but their computational power on Gaussian inputs is classically simulable. The quantum mechanics is genuine; the computational speedup is not.
• "The vacuum has zero energy." The vacuum is the ground state of the oscillator, but its energy is \hbar\omega/2, not zero. Its quadrature variance 1/2 reflects genuine quantum fluctuations that ultimately show up as shot noise in interference experiments.
• "Squeezing violates Heisenberg." No — the uncertainty product \Delta x \cdot \Delta p stays \geq 1/2 for every state. Squeezing reduces one variance and increases the other while conserving the product. The famous "below vacuum noise" measurements beat the vacuum in one quadrature at the cost of worse noise in the orthogonal one.
• "CV is automatically more efficient than qubits because each mode is infinite-dimensional." The naive Hilbert-space-dimension argument is misleading. Computational resources are what matter, and those are bounded by what gates you can realistically implement. A Gaussian CV circuit on one mode gives no more power than a classical computer simulating a 2D point in phase space.
• "GKP states are easy because they are 'just combs.'" Ideal GKP states are unnormalisable and unphysical. Approximate GKP states require preparation protocols that inject significant non-Gaussian content, typically via photon-number-resolving measurements or transmon ancilla coupling. They are the non-Gaussian resource, not a Gaussian byproduct.
• "Xanadu has demonstrated fault-tolerant quantum computing." Xanadu's Borealis demonstrated Gaussian Boson Sampling — a non-universal supremacy-style task. Fault-tolerant CV QC on photonic hardware is a roadmap target, not a 2025 achievement.

Going deeper

The symplectic formalism

An n-mode Gaussian state is fully characterised by a displacement vector \mu \in \mathbb{R}^{2n} and a covariance matrix \sigma \in \mathbb{R}^{2n \times 2n}, where the phase-space coordinates are \mathbf{r} = (x_1, p_1, x_2, p_2, \ldots, x_n, p_n):

A Gaussian unitary acts on (\mu, \sigma) by \mu \to S\mu + d, \sigma \to S \sigma S^T, where S \in Sp(2n, \mathbb{R}) is a symplectic matrix (S \Omega S^T = \Omega with \Omega the standard symplectic form) and d \in \mathbb{R}^{2n} is a displacement. Homodyne measurements give Gaussian integrals with known formulas. Everything is classical algebra of size O(n^2).

The Lloyd–Braunstein universality theorem

Let \mathcal{G} denote the set of Gaussian unitaries plus the cubic phase gate e^{i\gamma \hat{x}^3}. Lloyd and Braunstein (1999) proved:

Theorem. Any polynomial P(\hat{x}_1, \hat{p}_1, \ldots, \hat{x}_n, \hat{p}_n)-Hamiltonian unitary e^{iP} can be approximated to any precision \epsilon by a product of unitaries from \mathcal{G}, with total length polynomial in the polynomial degree and 1/\epsilon.

The proof uses a Trotter-like decomposition: combine e^{iA} and e^{iB} using the approximation e^{i(A+B)\delta} \approx e^{iA\delta} e^{iB\delta}, and recursively generate higher-order polynomials via nested commutators of Gaussian + cubic terms. The algebra of polynomials on \hat{x}, \hat{p} is generated by the quadratic and cubic terms via commutators, and any polynomial-Hamiltonian unitary is reachable from these.

GKP codes combined with outer codes

GKP alone protects against small mode-displacement errors but not against large-displacement errors or against mode loss (photon loss). A full fault-tolerant architecture concatenates GKP with an outer qubit code — typically a surface code over many GKP-encoded modes:

1. Each physical mode is encoded as a GKP qubit, protected against small displacement errors by stabilizer measurements.
2. A lattice of GKP qubits is arranged in the 2D surface-code geometry.
3. Surface-code syndrome extraction runs over GKP qubits; the inner GKP layer suppresses small errors, the outer surface code handles the rare logical failures of the inner code.

This concatenation was proved to yield a threshold at around 10 dB of GKP squeezing (Fukui–Tomita–Noiri, 2018; Vuillot et al., 2019). Current experimental GKP states are approaching this threshold in microwave-cavity platforms.

The link to boson sampling and classical hardness

Gaussian Boson Sampling — the task Borealis ran — samples from the output distribution of a Gaussian input passed through a random linear network and measured with PNR detectors. The output distribution is governed by hafnians of sub-matrices of the scattering matrix, a permanent-like quantity that is #P-hard to compute classically (Aaronson–Arkhipov for boson sampling; Hamilton et al. for Gaussian variant). This is why 216 squeezed modes suffice for a supremacy-style argument.

GBS is not universal CV QC, but it shows that the ingredients (squeezing + linear optics + PNR) are individually sufficient for hard sampling problems — and that these ingredients can be scaled on integrated photonics.

How CV compares with qubit approaches

Neither approach is obviously dominant. CV's one-mode-per-logical-qubit potential is enormously attractive, but the non-Gaussian resource cost and the photonic loss channel keep it on the same order of magnitude as qubit surface codes in realistic estimates.

Where this leads next

• Photonic Qubits — the discrete-variable photonic approach. Read this for the qubit-encoding language; CV is the alternative.
• Measurement-Based Photonics — the cluster-state paradigm. CV has a direct analogue (CV cluster states) used in Xanadu's architecture.
• Xanadu — the photonic CV company; hardware detail on Borealis and the GKP roadmap.
• GKP Encoding — the dedicated article on GKP codes; covers approximate states, syndrome extraction, and thresholds.

References

1. Samuel L. Braunstein, Peter van Loock, Quantum information with continuous variables (2005), Rev. Mod. Phys. 77, 513 — arXiv:quant-ph/0410100.
2. Daniel Gottesman, Alexei Kitaev, John Preskill, Encoding a qubit in an oscillator (2001), Phys. Rev. A 64, 012310 — arXiv:quant-ph/0008040.
3. Seth Lloyd, Samuel L. Braunstein, Quantum computation over continuous variables (1999), Phys. Rev. Lett. 82, 1784 — arXiv:quant-ph/9810082.
4. Stephen D. Bartlett, Barry C. Sanders, Samuel L. Braunstein, Kae Nemoto, Efficient classical simulation of continuous variable quantum information processes (2002), Phys. Rev. Lett. 88, 097904 — arXiv:quant-ph/0109047.
5. Lars S. Madsen et al. (Xanadu), Quantum computational advantage with a programmable photonic processor (2022), Nature 606 — Nature 606, 75–81.
6. Wikipedia, Continuous-variable quantum information.