How many copies are needed for approximate cloning of structured pure states?

What cloning sample complexity actually counts

The no-cloning theorem forbids a channel that maps N copies of an unknown pure state to N+K copies exactly. The quantitative version is more useful: a channel is an ε-approximate cloner from N to N+K copies if, for every state ψ in the promised family S, the trace distance between the channel output and ψ⊗(N+K) is at most ε. The paper's central quantity, cloning sample complexity N_min,S(ε), asks how many input copies N are needed to reach a target error ε.

This is a different question from how well a single fixed channel performs, and it is also different from tomography sample complexity, which asks how many copies are needed to learn the state. The paper treats N_min,S(ε) as the answer to a threshold question — for what N do ε-approximate cloners from N to N+1 copies exist — and it reports both constructive upper bounds and, where available, lower bounds that apply to arbitrary channels.

The baseline is the unstructured family of all pure states on a Hilbert space H, whose complexity is Θ(dim H/ε). For n qudits or n fermionic modes dim H grows exponentially in n, and for bosons it is infinite, so the unstructured answer is unusable in exactly the many-body settings where cloning would matter. The paper's question is whether restricting the promise to a structured family S ⊊ S_H buys a real reduction.

Learning-based cloning fails on continuous families

A natural shortcut is measure-and-prepare: use the N input copies to learn an estimate ψ̃ of the state, then output ψ̃⊗(N+K). For stabilizer states this approach was recently shown to be asymptotically optimal, with N_min,S = Θ(n) at constant ε. The paper finds that this shortcut does not generalize.

The deciding property is whether the family is discrete or continuous. For a continuous family S, the best estimate ψ̃ carries an error η ≃ N^(−1/2), and raising that estimate to the N-th tensor power amplifies the error to a constant ε. As a result, measure-and-prepare cloning cannot reach an arbitrarily small ε at any N and K. The obstruction already appears for the simplest continuous case, pure states on a single qubit.

The contrast the paper draws is instructive: the standard N→N+1 cloner for unstructured pure states is ε-approximate at N = Θ(1/ε), but it is not of measure-and-prepare form. So the failure is specific to the learn-then-prepare strategy, not to approximate cloning itself, and it motivates the paper's search for cloners that do not route through state estimation.

Resolving orbits yield a canonical cloner

The paper's main technical device is a class of families it calls resolving orbits. A family S is an orbit of a group G under a representation φ, and it is N-resolving when φ⊗N restricted to the N-copy span — the span of {|ψ⟩⊗N : ψ ∈ S} — is irreducible and square-integrable. Square-integrability is what keeps the construction meaningful in infinite-dimensional settings, where the copy span has infinite dimension but the formal degree d_N remains finite.

For such a family the integral of ψ⊗N over S resolves the identity on the N-copy span, giving ∫ ψ⊗N = (1/d_N) π_N, where π_N projects onto that span. This yields an explicit channel T(ρ_N) = (d_N/d_(N+K)) π_(N+K) (ρ_N ⊗ I_K) π_(N+K), which the paper calls the canonical cloner. Its accuracy is set by ε = 1 − f_(T,S), with cloning fidelity f_(T,S) = d_N/d_(N+K).

The construction generalizes earlier work: the family of all pure states is itself a resolving orbit whose N-copy span is the symmetric subspace, recovering Werner's cloner, and the framework also covers the generalized coherent states treated by Chiribella and Yang. For compact G with finite-dimensional H, the canonical cloner is also optimal. The paper notes that the tight relation ε = 1 − f between trace distance and fidelity is specific to this channel rather than a general identity.

Fermionic Gaussian and Slater states become polynomial

Fermionic Gaussian states form an orbit of the Pin(2n) group, and their q-particle subfamily — the Slater states — forms an orbit of U(n). In physics terms these are Bogoliubov transformations and single-particle transformations. To apply the canonical cloner, the paper shows that these groups act irreducibly on the N-copy spans. For Slater states the base of the orbit is the Fermi sea state, which is a highest-weight vector; the Cartan product theorem then implies that its N-th tensor power is also highest-weight and generates an irreducible representation, with dimensionality d_N given by the Weyl product formula.

The general Gaussian case needs extra steps because Pin(2n) is disconnected, so highest-weight theory does not apply directly. Once the resolving property is established, the sample complexities follow from the asymptotics of the fidelity ratio d_N/d_(N+1). The resulting optimal cloners give N_min,S_ff(ε) = n(n−1)/(2ε) · (1 + O(ε)) for fermionic Gaussian states and N_min,S_ff,q(ε) = q(n−q)/ε · (1 + O(ε)) for q-particle Slater states, as ε → 0, uniformly for n ≥ 2 and 1 ≤ q ≤ n−1.

The comparison that matters is against the unstructured baseline: for fermions this is an exponential-to-polynomial reduction in n. The Slater-state formula also carries a structural message, since q(n−q) vanishes at q = 0 and q = n and is largest at half filling — the number of copies needed tracks how far the particle number sits from the extremes of the mode range.

Bosonic Gaussian states and the optimality gap

For bosons the paper separates two families. Centered pure Gaussian states S_sq have vanishing first moments but arbitrary squeezing and inter-mode correlations; general pure Gaussian states S_bg additionally carry unknown displacement. The centered family is described as the vacuum orbit of the metaplectic group, and adding displacements gives the Jacobi-group orbit. Irreducibility on the copy spans follows from the vacuum being the unique vector at zero total photon number, and the formal degrees and square-integrability threshold come from explicit integrals.

For centered states, explicit cloners together with a matching lower bound give N_min,S_sq(ε) = Θ_ε(n) for every fixed 0 < ε < 1 — linear in the number of modes at fixed error. For general bosonic Gaussian states the paper obtains a sandwich: n/(−log(1−ε)) − 1 ≤ N_min,S_bg(ε) ≤ n(n+3)/(2ε) + 2n for 0 < ε < 1. The upper bounds are constructive and the lower bounds apply to arbitrary channels.

The asymmetry between these two results is the point readers should not blur. The centered-state bound is optimal in its linear dependence on n at each fixed error, but the sharp joint dependence on n and ε remains open. For the full Gaussian family, the quadratic upper bound is not known to be optimal, and whether it can be improved to linear in n is also open. The paper states plainly that its bosonic Gaussian cloner certifies polynomial complexity without any constraint on the state's energy, but is not optimal in general. Optimality for these groups is also limited in scope: because the groups are not amenable, the construction is guaranteed optimal only among covariant channels, whereas for coherent states the displacement group is amenable and the canonical cloner is globally optimal with complexity Θ(n/ε).

Manifold dimension controls the scaling

The paper offers a unifying explanation for why so many of these families land on linear-in-manifold-dimension behavior. Through a controlled saddle-point evaluation of the frame potential at high N, it argues that for many families the cloning complexity is linear in the dimension of the family manifold. The same calculation accounts for the analogous scaling of tomography complexity, which is why the two notions of sample complexity move together here even though they answer different questions.

This dimension law is the paper's answer to the 'why' behind the individual formulas, and it is presented as an explanation rather than as a universal theorem covering every family considered. The phase-state family is explicitly outside the framework, and the bosonic Gaussian results show that the law does not by itself pin down the sharp joint dependence on n and ε.

Phase states need a separate construction

Qudit phase states are not covered by the resolving-orbit framework, so the paper builds an independent channel for them with optimal cloning fidelity. The analysis works in the occupation basis and uses a covariance reduction to a block semidefinite program, a balanced channel with a moment representation, and a performance operator whose fidelity is evaluated directly.

The optimality argument rests on balanced shifts, and the paper proves uniform fidelity bounds for the case of adding one copy. The technical bounds are explicit: for N ≥ 2d the fidelity deviation satisfies (1 − 1/d)^(N+1) + 10^(−9)(d−1)/N² ≤ δ_(N,d) ≤ (1 − 1/d)^(N+1) + 30(d−1)/N², and inverting these gives a sufficient copy count N = ⌈8(d log(1/δ) + √(d/δ))⌉. The numerical constants are stated as unoptimized.

A useful cross-check appears in the appendix on non-resolving orbit families: applying the general orbit bound to phase states reproduces the same worst-case lower bound as the general state cloner, namely the ratio of binomial-type dimensions (N+d−1 choose N)/(N+K+d−1 choose N+K). That places the phase-state result in the same family of bounds as the unstructured case, while the independent construction supplies the matching optimal fidelity.

Where these complexities do and do not transfer

The practical reading of this paper depends on which family a reader's problem actually involves, because the reported complexities are not interchangeable. Fermionic Gaussian and Slater states come with optimal cloners and closed-form copy counts in n and ε. Centered bosonic Gaussian states come with a linear-in-n count that is optimal at fixed error but whose joint n–ε dependence is unresolved. General bosonic Gaussian states come with a constructive quadratic upper bound and a weaker lower bound, with optimality unknown. Qudit phase states come with an optimal-fidelity channel built by a separate route.

Two boundaries deserve to travel with those numbers. First, the measure-and-prepare obstruction means that for continuous families, no amount of learning-then-preparing reaches small ε at any N or K, so a reader cannot substitute a tomography-based workflow and expect the same guarantees. Second, the optimality claims are not uniform: the canonical cloner is optimal for compact G with finite-dimensional H, but for the non-amenable bosonic Gaussian groups optimality is established only among covariant channels.

The paper also states that its results extend to N→N+K cloning for K > 1 across all considered families, and it positions itself as complementary to recent work on cloning stabilizer states, where learning-based cloning is asymptotically optimal. The sharpest decision threshold the paper offers is therefore not a single number but a family-by-family split: if the promised family is fermionic Gaussian or Slater, the copy count is known and optimal; if it is centered bosonic Gaussian, the linear scaling in n is the reliable part; if it is general bosonic Gaussian, treat the quadratic bound as constructive rather than tight.

Sources used in this answer