Simple Algorithm
Quantum Computing · Rydberg State · 30 min

Linear Optical QC & Fusion

KLM, measurement-induced nonlinearity and fusion-based designs.

Photons do not interact, yet Knill, Laflamme and Milburn proved in 2001 that single photons, beam splitters, detectors and fast feed-forward suffice for scalable quantum computing: the nonlinearity comes from measurement. Fusion-based architectures reorganize the same insight around small entangled resource states and heralded, erasure-like failures.

What you’ll learn

  • Describe passive linear optics as a representation of U(m)U(m) on Fock space, and compute multiphoton amplitudes as permanents.
  • Derive the KLM controlled-ZZ from nonlinear sign gates, its success probability, and the boosting of teleported gates to n/(n+1)n/(n+1).
  • Explain the 50% limit of linear-optics Bell measurements, fusion gates, and fusion-based quantum computation.
  • Use the Gottesman–Knill theorem and its continuous-variable analog to identify which photonic resources are classically simulable.
  • Describe GKP qubits and the role of squeezing in continuous-variable architectures.

Linear optics as a group representation

A passive linear-optical network on mm modes (beam splitters, phase shifters, mirrors) conserves photon number and acts linearly on creation operators,

aj†  ↦  ∑k=1mUkj ak†,U∈U(m).a_j^\dagger\;\mapsto\;\sum_{k=1}^m U_{kj}\,a_k^\dagger,\qquad U\in U(m).

On the nn-photon Fock space this induces a unitary φ(U)\varphi(U): a representation of U(m)U(m) of dimension (m+n−1n)\binom{m+n-1}{n}, exponentially larger than mm.

⟨T∣φ(U)∣S⟩=Perm⁡(UT,S)∏isi! ∏jtj!\bra{T}\varphi(U)\ket{S} = \frac{\operatorname{Perm}\big(U_{T,S}\big)}{\sqrt{\prod_i s_i!\,\prod_j t_j!}}
The amplitude from input occupations S=(s1,…,sm)S = (s_1,\dots,s_m) to output occupations TT, where UT,SU_{T,S} is the n×nn\times n matrix built from the rows and columns of UU repeated according to TT and SS. Permanents are #P-hard to compute in general, the root of boson-sampling arguments (Aaronson & Arkhipov, 2011).

Hong–Ou–Mandel interference is the smallest example. For a balanced beam splitter U=12(111−1)U = \tfrac{1}{\sqrt2}\begin{pmatrix}1&1\\1&-1\end{pmatrix}, the coincidence amplitude is ⟨1,1∣φ(U)∣1,1⟩=Perm⁡(U)=12(−1+1)=0\bra{1,1}\varphi(U)\ket{1,1} = \operatorname{Perm}(U) = \tfrac12(-1+1) = 0.

Measurement-induced nonlinearity: KLM

DerivationA controlled-Z from two nonlinear sign gates
  1. NS: α0∣0⟩+α1∣1⟩+α2∣2⟩ ⟼ α0∣0⟩+α1∣1⟩−α2∣2⟩\mathrm{NS}:\ \alpha_0\ket0 + \alpha_1\ket1 + \alpha_2\ket2\ \longmapsto\ \alpha_0\ket0 + \alpha_1\ket1 - \alpha_2\ket2

    The nonlinear sign gate flips the sign of the two-photon component of one mode. No linear-optical network does this deterministically; KLM realize it with one ancilla photon in two ancilla modes, a three-mode network and postselection on a particular detection pattern, succeeding with probability 14\tfrac14.

  2. ∣1⟩a∣1⟩b → BS  12(∣2,0⟩−∣0,2⟩) → NS⊗NS  −12(∣2,0⟩−∣0,2⟩) → BS−1  −∣1⟩a∣1⟩b\ket{1}_a\ket{1}_b\ \xrightarrow{\ \mathrm{BS}\ }\ \tfrac{1}{\sqrt2}\big(\ket{2,0}-\ket{0,2}\big)\ \xrightarrow{\ \mathrm{NS}\otimes\mathrm{NS}\ }\ -\tfrac{1}{\sqrt2}\big(\ket{2,0}-\ket{0,2}\big)\ \xrightarrow{\ \mathrm{BS}^{-1}\ }\ -\ket{1}_a\ket{1}_b

    Let aa and bb be the "1" rails of two dual-rail qubits. Only the logical input ∣11⟩\ket{11} puts a photon in both; HOM interference bunches them, and the NS gates flip the sign.

  3. ∣1,0⟩, ∣0,1⟩, ∣0,0⟩ ⟼ unchanged\ket{1,0},\ \ket{0,1},\ \ket{0,0}\ \longmapsto\ \text{unchanged}

    With at most one photon in modes a,ba, b, the NS gates act trivially (they only touch two-photon components), and the second beam splitter undoes the first.

  4. CZwithpsucc=(14)2=116\mathrm{CZ}\quad\text{with}\quad p_{\rm succ} = \big(\tfrac14\big)^2 = \tfrac{1}{16}

    Both NS gates must succeed; failures are heralded by the ancilla detectors.

A 1/161/16 gate cannot be used directly inside a circuit. KLM's second idea is gate teleportation: apply the probabilistic gate to an entangled ancilla state offline, retrying until it succeeds, then teleport the data qubits through it. Teleportation needs a Bell measurement, which linear optics performs only probabilistically, so KLM built generalized teleportation circuits with nn-photon entangled ancillas that succeed with probability n/(n+1)n/(n+1). A failure is heralded and amounts to a known computational-basis measurement of the qubit, which a simple code can absorb.

Predict before you look

A teleported CZ uses two KLM teleportations, each succeeding with probability n/(n+1)n/(n+1). With n=9n = 9, what is the success probability of the CZ, in percent?

Show the answer

81 %

(9/10)2=0.81(9/10)^2 = 0.81. The failure probability falls only as 1/n1/n per teleportation while the ancilla states grow with nn: KLM is scalable in principle, but its overhead drove the search for leaner schemes built on cluster states and fusion.

Fusion and resource states

Measurement-based computing moves all entangling into the preparation of a resource state, so photonics only has to grow such states. Browne and Rudolph (2005) introduced fusion gates, two-photon measurements that merge two cluster states. A Type-II fusion is a linear-optics Bell measurement on one photon from each cluster. On success (probability 12\tfrac12 in its basic form) it measures the parities X1X2X_1X_2 and Z1Z2Z_1Z_2, joining the clusters. On failure it measures the photons separately, which yields only one of the two parities: the structure is damaged at a known place rather than corrupted invisibly.

Fusion-based quantum computation (FBQC; Bartolucci et al., 2023) makes fusion the whole architecture. Many identical resource-state generators produce constant-size entangled states (for example 4- or 6-photon graph states); each photon undergoes only a few operations and is consumed in a fusion. The fusions are arranged so that their outcomes form the syndrome of a topological code. A failed fusion is a heralded erasure of one parity, and photon loss is a detectable erasure too, so the decoder knows where the damage is. The paper develops a stabilizer formalism for such fusion networks and analyzes thresholds under erasure and Pauli errors and under a linear-optics loss model.

Check your understanding

Why does FBQC treat a failed fusion as an erasure rather than as a Pauli error?

  1. Failed fusions do not affect the computation
  2. A failed fusion is always repaired by repeating it on the same photons
  3. Erasures and Pauli errors are mathematically identical
  4. The failure is heralded by the detection pattern, so the decoder knows which parity outcome is missing, and located errors are much easier to correct
Show the answer

D. The failure is heralded by the detection pattern, so the decoder knows which parity outcome is missing, and located errors are much easier to correct

The detectors announce the failure, and the missing parity is a known gap in the syndrome. Codes tolerate far higher rates of located errors than of unlocated ones.

Continuous variables and GKP

Light also carries continuous-variable (CV) quantum information in the quadratures q^,p^\hat q, \hat p of each mode, with [q^,p^]=i\comm{\hat q}{\hat p} = i in units where ℏ=1\hbar = 1. Gaussian states (vacuum, coherent, squeezed), Gaussian operations (beam splitters, phase shifts, squeezers, displacements) and homodyne detection form the cheap, deterministic part of the toolkit. Squeezed light can be woven deterministically into very large cluster states multiplexed in time, demonstrated in 2019 by Asavanant et al. and by Larsen et al.

Universality therefore needs a non-Gaussian ingredient: photon counting, a cubic-phase gate or non-Gaussian states. The most developed choice is the GKP code (Gottesman, Kitaev & Preskill, 2001), which encodes a qubit in a single oscillator.

∣μˉ⟩∝∑s∈Z∣q^=(2s+μ)π⟩,Sq=e−2iπ p^,Sp=e2iπ q^\ket{\bar\mu}\propto\sum_{s\in\Z}\ket{\hat q = (2s+\mu)\sqrt\pi},\qquad S_q = e^{-2i\sqrt\pi\,\hat p},\qquad S_p = e^{2i\sqrt\pi\,\hat q}
Ideal square-lattice GKP codewords (μ∈{0,1}\mu\in\{0,1\}) and the two commuting stabilizers, displacements by 2π2\sqrt\pi. The logical Xˉ\bar X and Zˉ\bar Z are displacements by π\sqrt\pi.

Key takeaways

  • Passive linear optics represents U(m)U(m) on Fock space; any UU needs m(m−1)/2m(m-1)/2 beam splitters, and multiphoton amplitudes are permanents.
  • KLM: measurement-induced nonlinearity gives a heralded CZ with probability 1/161/16, and gate teleportation boosts success to n/(n+1)n/(n+1) per teleportation.
  • Linear-optics Bell measurements without ancillas succeed at most half the time; fusion-based architectures are built around such heralded, erasure-like failures.
  • Gaussian circuits (Bartlett et al.) and Pauli-measured stabilizer resources (Gottesman–Knill) are classically simulable; GKP states supply the non-Gaussian resource and turn small shifts into correctable errors.

Check yourself

In the interactive lesson these are scored with your confidence, and what you miss comes back for review.

Question 1

Three photons enter a three-mode interferometer UU, one per input mode. The amplitude for one photon in each output mode is

  1. det⁡U\det U
  2. Perm⁡(U)\operatorname{Perm}(U)
  3. ∏jUjj\prod_j U_{jj}
  4. Tr⁡U\Tr U
Show the answer

B. Perm⁡(U)\operatorname{Perm}(U)

Sum over all ways of routing the three photons to the three outputs, every term with a plus sign because photons are bosons: the permanent. Fermions would give the determinant.

Question 2

The KLM controlled-ZZ built from two nonlinear sign gates succeeds with probability

  1. 1/21/2
  2. 1/41/4
  3. 1/91/9
  4. 1/161/16
Show the answer

D. 1/161/16

Each NS gate succeeds with probability 14\tfrac14, and both must succeed.

Question 3

Why can't a photonic computer that uses only squeezed states, beam splitters and homodyne detection outperform classical computers?

  1. Such circuits are efficiently simulable by tracking mean vectors and covariance matrices (Bartlett et al., 2002); a non-Gaussian element is required
  2. Squeezed states decohere too quickly
  3. Homodyne detection is too noisy
  4. Beam splitters cannot entangle squeezed states
Show the answer

A. Such circuits are efficiently simulable by tracking mean vectors and covariance matrices (Bartlett et al., 2002); a non-Gaussian element is required

Gaussian in, Gaussian operations, Gaussian measurements: the state stays Gaussian and is described by polynomially many numbers. Photon counting, cubic-phase gates or GKP states break the spell.

Question 4

A GKP qubit suffers a position shift uu. Ideal GKP error correction recovers the logical state whenever

  1. ∣u∣<π\abs u < \sqrt\pi
  2. ∣u∣<2π\abs u < 2\sqrt\pi
  3. ∣u∣<π/2\abs u < \sqrt\pi/2
  4. always, for any uu
Show the answer

C. ∣u∣<π/2\abs u < \sqrt\pi/2

Logical codewords are offset by π\sqrt\pi. Measuring q^ mod π\hat q \bmod \sqrt\pi identifies the shift correctly unless it exceeds half that spacing, where it is mistaken for a logical Xˉ\bar X.

Further reading

  • E. Knill, R. Laflamme & G. J. Milburn, "A scheme for efficient quantum computation with linear optics", Nature 409, 46 (2001).
  • D. E. Browne & T. Rudolph, "Resource-efficient linear optical quantum computation", Physical Review Letters 95, 010501 (2005).
  • S. Bartolucci et al., "Fusion-based quantum computation", Nature Communications (2023).
  • D. Gottesman, A. Kitaev & J. Preskill, "Encoding a qubit in an oscillator", Physical Review A 64, 012310 (2001).
  • P. Kok, W. J. Munro, K. Nemoto, T. C. Ralph, J. P. Dowling & G. J. Milburn, "Linear optical quantum computing with photonic qubits", Reviews of Modern Physics 79, 135 (2007).