Simple Algorithm
Quantum Computing · Rydberg State · 26 min

Optimal Approximate Cloning

The best imperfect copier and the 5/6 fidelity limit.

No-cloning says perfect copies are impossible. It does not say how close you can get. For qubits the answer is a crisp number, 5/65/6, and three independent routes reach it: an explicit machine, symmetry and representation theory, and the impossibility of signalling.

What you’ll learn

  • Formulate cloning as an optimization over channels, with single-copy and global fidelities, and justify restricting to covariant cloners.
  • Derive the Bužek–Hillery cloner's output, its shrinking factor η=2/3\eta=2/3 and its fidelity 5/65/6.
  • Prove that 5/65/6 is optimal for symmetric universal 1→21\to2 qubit cloning from positivity and no-signalling.
  • Apply the general N→MN\to M formula in dimension dd, and connect its M→∞M\to\infty limit to optimal state estimation.
  • Derive the optimal asymmetric trade-off and relate cloning to eavesdropping in QKD.

Cloning as an optimization

A 1→21\to2 cloner is a channel (a CPTP map) T:L(Cd)→L(Cd⊗Cd)T:\mathcal L(\C^d)\to\mathcal L(\C^d\otimes\C^d). For a pure input ψ=∣ψ⟩ ⁣⟨ψ∣\psi=\dyad\psi write ρAB=T(ψ)\rho_{AB}=T(\psi), with marginals ρA=Tr⁡BρAB\rho_A=\Tr_B\rho_{AB} and ρB=Tr⁡AρAB\rho_B=\Tr_A\rho_{AB}. Two figures of merit:

FA(ψ)=⟨ψ∣ρA∣ψ⟩,Fglob(ψ)=⟨ψψ∣ρAB∣ψψ⟩.F_A(\psi)=\bra\psi\rho_A\ket\psi,\qquad F_{\mathrm{glob}}(\psi)=\bra{\psi\psi}\rho_{AB}\ket{\psi\psi}.

A cloner is universal if these do not depend on ψ\psi, and symmetric if FA=FBF_A=F_B.

For qubits, covariance pins the marginals down completely. A qubit channel that commutes with every rotation can only rescale the Bloch vector, so each copy is a depolarized version of the input:

ρA=ηA ψ+(1−ηA)I2,FA=1+ηA2.\rho_A=\eta_A\,\psi+(1-\eta_A)\tfrac{\id}{2},\qquad F_A=\tfrac{1+\eta_A}{2}.

The Bloch vector of each copy is the input's, multiplied by the shrinking factor ηA\eta_A.

The Bužek–Hillery machine

Bužek and Hillery (1996) wrote down a unitary acting on the input aa, a blank bb in ∣0⟩\ket0, and one machine qubit xx:

U∣0⟩a∣0⟩b∣0⟩x=23 ∣00⟩ab∣↑⟩x+13 ∣Ψ+⟩ab∣↓⟩xU∣1⟩a∣0⟩b∣0⟩x=23 ∣11⟩ab∣↓⟩x+13 ∣Ψ+⟩ab∣↑⟩x\begin{aligned} U\ket{0}_a\ket{0}_b\ket{0}_x &= \sqrt{\tfrac23}\,\ket{00}_{ab}\ket{\uparrow}_x+\sqrt{\tfrac13}\,\ket{\Psi^+}_{ab}\ket{\downarrow}_x\\ U\ket{1}_a\ket{0}_b\ket{0}_x &= \sqrt{\tfrac23}\,\ket{11}_{ab}\ket{\downarrow}_x+\sqrt{\tfrac13}\,\ket{\Psi^+}_{ab}\ket{\uparrow}_x \end{aligned}
Here ∣Ψ+⟩=(∣01⟩+∣10⟩)/2\ket{\Psi^+}=(\ket{01}+\ket{10})/\sqrt2, and ∣↑⟩,∣↓⟩\ket\uparrow,\ket\downarrow are orthonormal machine states. The two right-hand sides are orthonormal, so the map extends to a unitary on all eight basis states.
DerivationReading off the copies
  1. ρab(0)=23∣00⟩ ⁣⟨00∣+13∣Ψ+⟩ ⁣⟨Ψ+∣\rho_{ab}(0)=\tfrac23\dyad{00}+\tfrac13\dyad{\Psi^+}

    Trace out the machine. Its two states are orthogonal, so the branches add incoherently.

  2. ρa(0)=23∣0⟩ ⁣⟨0∣+13⋅I2=56∣0⟩ ⁣⟨0∣+16∣1⟩ ⁣⟨1∣\rho_a(0)=\tfrac23\dyad0+\tfrac13\cdot\tfrac{\id}{2}=\tfrac56\dyad0+\tfrac16\dyad1

    Use Tr⁡b∣Ψ+⟩ ⁣⟨Ψ+∣=I/2\Tr_b\dyad{\Psi^+}=\id/2. By symmetry ρb(0)\rho_b(0) is the same.

  3. ρa(ψ)=23 ψ+13⋅I2for every pure ψ\rho_a(\psi)=\tfrac23\,\psi+\tfrac13\cdot\tfrac{\id}{2}\qquad\text{for every pure }\psi

    The machine is covariant, so the same shrinking factor η=23\eta=\tfrac23 holds for every input. (Check one case in the reveal below.)

  4. F=1+η2=56,Fglob=⟨ψψ∣ρab∣ψψ⟩=23F=\tfrac{1+\eta}{2}=\tfrac56,\qquad F_{\mathrm{glob}}=\bra{\psi\psi}\rho_{ab}\ket{\psi\psi}=\tfrac23

    Each copy has fidelity 5/65/6, and a joint test of "both outputs equal ψ\psi" passes with probability 2/32/3. The copies are correlated: Fglob≠FAFBF_{\mathrm{glob}}\neq F_AF_B.

Think first Check universality on an input that is not a basis state: compute ⟨+∣ρa∣+⟩\bra+\rho_a\ket+ for the input ∣+⟩\ket+.

By linearity the output is 12[(23∣00⟩+13∣Ψ+⟩)∣↑⟩+(13∣Ψ+⟩+23∣11⟩)∣↓⟩]\tfrac{1}{\sqrt2}\big[(\sqrt{\tfrac23}\ket{00}+\sqrt{\tfrac13}\ket{\Psi^+})\ket\uparrow+(\sqrt{\tfrac13}\ket{\Psi^+}+\sqrt{\tfrac23}\ket{11})\ket\downarrow\big]. Project qubit aa onto ∣+⟩\ket+ in each machine branch. The ∣↑⟩\ket\uparrow branch leaves 12(36∣0⟩b+16∣1⟩b)\tfrac12\big(\tfrac{3}{\sqrt6}\ket0_b+\tfrac{1}{\sqrt6}\ket1_b\big), of squared norm 512\tfrac{5}{12}, and the ∣↓⟩\ket\downarrow branch contributes another 512\tfrac{5}{12}. Total: ⟨+∣ρa∣+⟩=56\bra+\rho_a\ket+=\tfrac56.

Why nothing beats 5/6

DerivationProof: positivity plus no-signalling
  1. ρab(m⃗)=14[I⊗I+η (m⃗⋅σ⃗⊗I+I⊗m⃗⋅σ⃗)+∑jktjk(m⃗) σj⊗σk]\rho_{ab}(\vec m)=\tfrac14\Big[\id\otimes\id+\eta\,(\vec m\cdot\vec\sigma\otimes\id+\id\otimes\vec m\cdot\vec\sigma)+\textstyle\sum_{jk}t_{jk}(\vec m)\,\sigma_j\otimes\sigma_k\Big]

    The general two-qubit output for an input with Bloch vector m⃗\vec m and equal depolarized marginals. Without loss of generality the cloner is covariant (twirling) and swap-symmetric (randomly swapping the outputs preserves both fidelities when they are equal).

  2. tjk(m⃗)=t δjk+t′ mjmkt_{jk}(\vec m)=t\,\delta_{jk}+t'\,m_jm_k

    Covariance makes tjkt_{jk} a rotation-covariant tensor built from m⃗\vec m alone; swap symmetry makes it symmetric, which excludes ϵjklml\epsilon_{jkl}m_l.

  3. 12[ρab(n⃗)+ρab(−n⃗)]=14[I⊗I+t∑jσj⊗σj+t′∑jknjnk σj⊗σk]\tfrac12\big[\rho_{ab}(\vec n)+\rho_{ab}(-\vec n)\big]=\tfrac14\Big[\id\otimes\id+t\textstyle\sum_j\sigma_j\otimes\sigma_j+t'\sum_{jk}n_jn_k\,\sigma_j\otimes\sigma_k\Big]

    No-signalling: let Bob's input be half of a singlet. By measuring along any axis n⃗\vec n she likes, Alice steers it to ±n⃗\pm\vec n with probability ½ each. Bob's averaged output must not depend on her choice, so t′=0t'=0. (For a map already known to be linear this is automatic; the point is that no-signalling alone forces it.)

  4. spec⁡ρab(z^)=14{ 1+2η+t,  1−2η+t,  1+t,  1−3t }\operatorname{spec}\rho_{ab}(\hat z)=\tfrac14\big\{\,1+2\eta+t,\ \ 1-2\eta+t,\ \ 1+t,\ \ 1-3t\,\big\}

    With t′=0t'=0 and m⃗=z^\vec m=\hat z the operator is diagonal in {∣00⟩,∣11⟩,∣Ψ+⟩,∣Ψ−⟩}\{\ket{00},\ket{11},\ket{\Psi^+},\ket{\Psi^-}\}. Use ∑jσj⊗σj=2F−I\sum_j\sigma_j\otimes\sigma_j=2\mathbb F-\id, which is +1+1 on the triplet and −3-3 on the singlet.

  5. 1−3t≥0  and  1−2η+t≥0 ⟹ η≤1+t2≤23,F=1+η2≤561-3t\ge0\ \text{ and }\ 1-2\eta+t\ge0\ \Longrightarrow\ \eta\le\tfrac{1+t}{2}\le\tfrac23,\qquad F=\tfrac{1+\eta}{2}\le\tfrac56

    Positivity of ρab\rho_{ab}. Equality forces t=13t=\tfrac13 and η=23\eta=\tfrac23; the spectrum becomes {23,0,13,0}\{\tfrac23,0,\tfrac13,0\}, exactly the Bužek–Hillery output 23∣00⟩ ⁣⟨00∣+13∣Ψ+⟩ ⁣⟨Ψ+∣\tfrac23\dyad{00}+\tfrac13\dyad{\Psi^+}. ■\blacksquare

Check your understanding

A symmetric universal 1→21\to2 qubit cloner has correlation parameter t=0.2t=0.2. What is the largest shrinking factor that positivity allows?

  1. η=0.6\eta=0.6
  2. η=2/3\eta=2/3
  3. η=0.4\eta=0.4
  4. η=0.8\eta=0.8
Show the answer

A. η=0.6\eta=0.6

The binding eigenvalue is 1−2η+t≥01-2\eta+t\ge0, so η≤(1+t)/2=0.6\eta\le(1+t)/2=0.6; the other constraint, t≤13t\le\tfrac13, is satisfied. Only t=13t=\tfrac13 allows the optimum η=23\eta=\tfrac23.

Representation theory gives the same answer and much more. Werner (1998) showed that, for global fidelity, the optimal universal N→MN\to M cloner is unique:

T(ρ⊗N)=d[N]d[M] SM(ρ⊗N⊗I⊗(M−N))SM,d[K]=(d+K−1K),T(\rho^{\otimes N})=\frac{d[N]}{d[M]}\,S_M\big(\rho^{\otimes N}\otimes\id^{\otimes(M-N)}\big)S_M,\qquad d[K]=\binom{d+K-1}{K},

where SMS_M projects onto the symmetric subspace of (Cd)⊗M(\C^d)^{\otimes M} and d[K]d[K] is its dimension. Its global fidelity is d[N]/d[M]d[N]/d[M]; for d=2d=2, N=1N=1, M=2M=2 that is 2/32/3, as found above. Keyl and Werner (1999) proved that the same machine also maximizes the single-copy fidelity.

N → M, any dimension, and the estimation limit

For NN identical inputs and M≥NM\ge N outputs in dimension dd, the optimal universal single-copy fidelity is (Gisin–Massar 1997 for qubits; Werner 1998 and Keyl–Werner 1999 in general):

FN→M(d)=NM+(1−NM)N+1N+dF^{(d)}_{N\to M}=\frac{N}{M}+\Big(1-\frac{N}{M}\Big)\frac{N+1}{N+d}
For qubits this becomes F=M(N+1)+NM(N+2)F=\dfrac{M(N+1)+N}{M(N+2)}, i.e. shrinking factor ηN→M=N(M+2)M(N+2)\eta_{N\to M}=\dfrac{N(M+2)}{M(N+2)}.
N→MN\to MddOptimal FF
1→21\to225/6≈0.8335/6\approx0.833
1→31\to327/9≈0.7787/9\approx0.778
2→32\to3211/12≈0.91711/12\approx0.917
1→∞1\to\infty22/32/3 (state estimation)
1→21\to233/43/4
1→21\to2ddd+32(d+1)\dfrac{d+3}{2(d+1)}
Predict before you look

What is the optimal universal single-copy fidelity for 1→31\to3 qubit cloning?

Show the answer

0.778

F=M(N+1)+NM(N+2)=3⋅2+13⋅3=79≈0.778F=\frac{M(N+1)+N}{M(N+2)}=\frac{3\cdot2+1}{3\cdot3}=\frac79\approx0.778, between the 1→21\to2 value 5/65/6 and the estimation limit 2/32/3.

Asymmetric cloners and the frontier

Drop the symmetry requirement: how good can copy BB be if copy AA must have shrinking factor ηA\eta_A? An elegant construction superposes the two trivial cloners.

DerivationThe optimal asymmetric trade-off
  1. ∣Ψ⟩RABC=a ∣Φ+⟩RA∣Φ+⟩BC+b ∣Φ+⟩RB∣Φ+⟩AC,a,b≥0\ket\Psi_{RABC}=a\,\ket{\Phi^+}_{RA}\ket{\Phi^+}_{BC}+b\,\ket{\Phi^+}_{RB}\ket{\Phi^+}_{AC},\qquad a,b\ge0

    Work with the Choi state: a reference RR is maximally entangled with the input. In the first branch copy AA is the input and BB is half of a fresh pair (junk); in the second the roles swap. CC is the machine.

  2. ⟨ΦRA+ΦBC+∣ΦRB+ΦAC+⟩=12 ⟹ a2+b2+ab=1\braket{\Phi^+_{RA}\Phi^+_{BC}|\Phi^+_{RB}\Phi^+_{AC}}=\tfrac12\ \Longrightarrow\ a^2+b^2+ab=1

    Normalization. One checks Tr⁡ABC∣Ψ⟩ ⁣⟨Ψ∣=I/2\Tr_{ABC}\dyad\Psi=\id/2, so ∣Ψ⟩\ket\Psi is the Choi state of a valid isometry; it is covariant by construction, since each Φ+\Phi^+ is invariant under Uˉ⊗U\bar U\otimes U.

  3. FeA=∥(⟨Φ+∣RA⊗IBC)∣Ψ⟩∥2=(a+b2)2F_e^{A}=\Big\lVert\big(\bra{\Phi^+}_{RA}\otimes\id_{BC}\big)\ket\Psi\Big\rVert^2=\big(a+\tfrac b2\big)^2

    The entanglement fidelity of the channel to AA, using (⟨Φ+∣RA⊗I)∣Φ+⟩RB∣Φ+⟩AC=12∣Φ+⟩BC\big(\bra{\Phi^+}_{RA}\otimes\id\big)\ket{\Phi^+}_{RB}\ket{\Phi^+}_{AC}=\tfrac12\ket{\Phi^+}_{BC}.

  4. ηA=4FeA−13=a(a+b),ηB=b(a+b)\eta_A=\frac{4F_e^A-1}{3}=a(a+b),\qquad\eta_B=b(a+b)

    For a depolarizing channel Fe=(1+3η)/4F_e=(1+3\eta)/4; then use a2+b2+ab=1a^2+b^2+ab=1.

  5. ηA2+ηB2+ηAηB=ηA+ηB\eta_A^2+\eta_B^2+\eta_A\eta_B=\eta_A+\eta_B

    Eliminating aa and bb gives an ellipse through (1,0)(1,0), (0,1)(0,1) and the symmetric optimum (23,23)(\tfrac23,\tfrac23). Equivalently FA=1−b22F_A=1-\tfrac{b^2}{2} and FB=1−a22F_B=1-\tfrac{a^2}{2}. This family is optimal (Niu–Griffiths 1998; Cerf 2000).

Key takeaways

  • Optimal cloners can be taken covariant; qubit copies are then depolarized, ρ=ηψ+(1−η)I/2\rho=\eta\psi+(1-\eta)\id/2, with F=(1+η)/2F=(1+\eta)/2.
  • Bužek–Hillery: η=2/3\eta=2/3, F=5/6F=5/6, Fglob=2/3F_{\mathrm{glob}}=2/3, equivalently T(ρ)=23S(ρ⊗I)ST(\rho)=\tfrac23S(\rho\otimes\id)S. Positivity plus no-signalling proves it optimal.
  • General law: FN→M=NM+(1−NM)N+1N+dF_{N\to M}=\frac NM+\big(1-\frac NM\big)\frac{N+1}{N+d}, approaching the estimation limit N+1N+d\frac{N+1}{N+d} as M→∞M\to\infty.
  • Asymmetric cloners trace the ellipse ηA2+ηB2+ηAηB=ηA+ηB\eta_A^2+\eta_B^2+\eta_A\eta_B=\eta_A+\eta_B: the same trade-off an eavesdropper faces.

Check yourself

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

Question 1

The Bužek–Hillery output for input ∣0⟩\ket0 is ρab=23∣00⟩ ⁣⟨00∣+13∣Ψ+⟩ ⁣⟨Ψ+∣\rho_{ab}=\tfrac23\dyad{00}+\tfrac13\dyad{\Psi^+}. What is its global fidelity ⟨00∣ρab∣00⟩\bra{00}\rho_{ab}\ket{00}?

  1. 5/65/6
  2. 25/3625/36
  3. 2/32/3
  4. 1/31/3
Show the answer

C. 2/32/3

Only the first term overlaps ∣00⟩\ket{00}, so Fglob=23=d[1]/d[2]F_{\mathrm{glob}}=\tfrac23=d[1]/d[2]. The copies are correlated, so Fglob≠FAFBF_{\mathrm{glob}}\neq F_AF_B.

Question 2

For qutrits (d=3d=3), what does the optimal universal 1→M1\to M single-copy fidelity tend to as M→∞M\to\infty?

  1. 2/32/3
  2. 3/43/4
  3. 1/31/3
  4. 1/21/2
Show the answer

D. 1/21/2

The limit is the estimation fidelity N+1N+d=24=12\frac{N+1}{N+d}=\frac24=\frac12 for N=1N=1, d=3d=3.

Question 3

In the no-signalling proof of the 5/65/6 bound, what forces the coefficient t′t' of mjmk σj⊗σkm_jm_k\,\sigma_j\otimes\sigma_k to vanish?

  1. Positivity of ρab\rho_{ab}
  2. Bob's averaged output cannot depend on the axis along which Alice steers his input
  3. Swap symmetry of the two outputs
  4. Normalization, Tr⁡ρab=1\Tr\rho_{ab}=1
Show the answer

B. Bob's averaged output cannot depend on the axis along which Alice steers his input

Averaging over Alice's two outcomes leaves the term t′∑jknjnkσj⊗σkt'\sum_{jk}n_jn_k\sigma_j\otimes\sigma_k, which depends on her chosen axis n⃗\vec n. No-signalling demands independence from n⃗\vec n, so t′=0t'=0.

Question 4

On the optimal asymmetric boundary ηA2+ηB2+ηAηB=ηA+ηB\eta_A^2+\eta_B^2+\eta_A\eta_B=\eta_A+\eta_B, copy AA must have ηA=0.9\eta_A=0.9. What is the best possible ηB\eta_B, approximately?

  1. ηB≈0.35\eta_B\approx0.35
  2. ηB≈0.10\eta_B\approx0.10
  3. ηB≈0.67\eta_B\approx0.67
  4. ηB≈0.50\eta_B\approx0.50
Show the answer

A. ηB≈0.35\eta_B\approx0.35

Solve ηB2−0.1 ηB−0.09=0\eta_B^2-0.1\,\eta_B-0.09=0: ηB=0.1+0.372≈0.354\eta_B=\tfrac{0.1+\sqrt{0.37}}{2}\approx0.354, so FB≈0.68F_B\approx0.68. Making one copy nearly perfect leaves the other only slightly better than measure-and-prepare (η=13\eta=\tfrac13).

Further reading

  • V. Bužek & M. Hillery, "Quantum copying: Beyond the no-cloning theorem", Physical Review A 54, 1844 (1996).
  • N. Gisin & S. Massar, "Optimal quantum cloning machines", Physical Review Letters 79, 2153 (1997).
  • R. F. Werner, "Optimal cloning of pure states", Physical Review A 58, 1827 (1998).
  • N. Gisin, "Quantum cloning without signaling", Physics Letters A 242, 1 (1998).
  • V. Scarani, S. Iblisdir, N. Gisin & A. Acín, "Quantum cloning", Reviews of Modern Physics 77, 1225 (2005).