Quantum superreplication of states and gates

Although the no-cloning theorem forbids perfect replication of quantum information, it is sometimes possible to produce large numbers of replicas with vanishingly small error. This phenomenon, known as quantum superreplication, can occur for both quantum states and quantum gates. The aim of this paper is to review the central features of quantum superreplication and provide a unified view of existing results. The paper also includes new results. In particular, we show that when quantum superreplication can be achieved, it can be achieved through estimation up to an error of size O(M/N2), where N and M are the number of input and output copies, respectively. Quantum strategies still offer an advantage for superreplication in that they allow for exponentially faster reduction of the error. Using the relation with estimation, we provide i) an alternative proof of the optimality of Heisenberg scaling in quantum metrology, ii) a strategy for estimating arbitrary unitary gates with a mean square error scaling as log N/N2, and iii) a protocol that generates O(N2) nearly perfect copies of a generic pure state U |0〉 while using the corresponding gate U only N times. Finally, we point out that superreplication can be achieved using interactions among k systems, provided that k is large compared to M2/N2.


I. INTRODUCTION
The no-cloning theorem [1,2] is one of the cornerstones of quantum information theory, with implications permeating the whole field [3,4].Most famously, the impossibility to copy non-orthogonal quantum states provides a working principle for quantum cryptography with applications to key distribution [5,6], unforgeable banknotes [7], and secret sharing [8].The no-cloning theorem forbids perfect cloning.A natural question, originally asked by Bužek and Hillery [9], is how well cloning can be approximated by the processes allowed by quantum mechanics.The question is relevant both to cryptographic applications and to the foundations of quantum theory, shedding light on relation between quantum and classical copy machines [10][11][12][13][14][15] and providing benchmarks that certify the advantages of quantum information processing over classical information processing [16][17][18][19][20].
Among the cloning machines allowed by quantum mechanics, one can distinguish two types: deterministic and probabilistic machines.Deterministic machines produce approximate copies with certainty, while probabilistic machines sometimes produce a failure message indicating that the copying process has gone wrong.In general, probabilistic machines can produce more accurate copies at the price of a reduced probability of success.For example, Duan and Guo [21] showed that a set of non-orthogonal states can be cloned without error as long as they are linearly independent.More recently, Fiurášek [22] showed that probabilistic cloners can offer an advantage even for linearly dependent states, including e. g. coherent states of harmonic oscillators with known amplitude.Ralph and Lund [23] proposed a concrete optical setup achieving noiseless probabilistic amplification and cloning of coherent states.The possibility of noiseless probabilistic amplification was later extended to the case where the coherent state amplitude is Gaussian-distributed [19].Although the probability of success vanishes as the accuracy increases, a nearly perfect amplification of coherent states has been observed experimentally for small values of the amplitude [24][25][26][27].
The recent developments on probabilistic amplification and cloning of coherent states motivate the search for new scenarios where cloning of non-orthogonal states can be achieved with vanishingly small error.In [28] we considered the task of cloning clock states, that is, quantum states generated by time evolution under a known Hamiltonian.Here, the cloner is given N identical copies of the same clock state and attempts to produce M = M (N ) copies.Inspired by the asymptotic framework of information theory, we considered the scenario where N is large.Here, the main question is how fast M (N ) can grow under the condition that the copying process is reliable, meaning that the global error on all the output copies vanishes in the large N limit.For deterministic cloners, we showed that the number of reliable copies has to scale as with lim In words, this means that the number of extra copies produced by the cloner must be negligible compared to the number of input copies.This result can be seen as an asymptotic no-cloning theorem, which extends the original no-cloning theorem from perfect cloners to cloners that become perfect in the asymptotic limit.The origin of the asymptotic no-cloning theorem is the requirement that the error vanishes globally on all the clones.In general, using tomography it is easy to produce copies that, individually, have an error vanishing order 1/N .However, this does not guarantee that the error vanishes at the global level.Indeed, the errors accumulate when multiple copies are examined jointly.When this fact is taken into account, it turns out that only a cloner satisfying Eq. ( 1) can have a vanishing error at the global level.For probabilistic cloners of clock states, the situation is different: the number of reliable copies produced by a probabilistic machine can scale as When δ is zero, this means that the number of extra copies scales as N , evading the limitation posed by the asymptotic no-cloning theorem.When δ is larger than zero, the number of extra copies grows faster than N and one can produce up to M = Θ(N 2 ) copies.In both cases, we refer to this phenomenon as superreplication, emphasizing the fact that the number of extra-copies grows beyond the limits imposed by the asymptotic no-cloning theorem.Note however, that the number of output copies allowed by Eq. ( 2) can grow at most at rate M = Θ N 2− , where is an arbitrarily small constant.It turns out that, in finite dimensions, the quadratic replication rate is the ultimate limit for the superreplication of clock states, in close connection with the Heisenberg limit of quantum metrology [28,29].
In addition to producing a large number of extra copies, probabilistic cloners exhibit a better scaling of the error, which vanishes as exp[−cN 2 /M (N )] for a suitable constant c > 0 depending on the particular set of clock states under consideration.In contrast, the error for deterministic machines can scale at most as 1/N 4 [28].The benefits of probabilistic cloners, however, do not come for free.The price to pay is a very small probability of success: precisely, the probability of superreplication has to be small compared to exp [−M/N ].This fact poses a severe limitation on the ability to observe superreplication, especially when M grows much faster than N .Nevertheless, the cloner can be devised in such a way that, when superreplication fails, the original input state is almost undisturbed.Precisely, Winter's gentle measurement lemma [30] implies that the error introduced by the failure of superreplication scales as O √ p succ , where p succ is the probability of success.For superreplication processes with δ > 0, this means that the error introduced by the failure of superreplication vanishes faster than exp −N δ/2 .As a consequence, the state resulting from a failed superreplication can still be used to achieve standard, deterministic cloning with asymptotically optimal performance: for large N , the error exp −N δ/2 is covered by the error introduced by deterministic cloning.In summary, superreplication is a rare event, but trying to observe it does not prevent the application of standard deterministic cloning techniques.Interestingly, the limitation on the probability of success is lifted if we consider the problem of cloning gates, instead of states.Cloning a quantum gate means simulating M uses of it while actually using it only N < M times.A no-cloning theorem for quantum gates was proven by D'Ariano, Perinotti, and one of the authors [31], who showed that no quantum protocol can perfectly clone a generic quantum gate.Recently, Dür, Sekatski, and Skotiniotis [32] analyzed the cloning of phase gates, i. e. unitary gates describing the time evolution with a known Hamiltonian.They devised a quantum network that simulates up to N 2 uses of an unknown phase gate by using it only N times.The network works deterministically and has vanishing error on average over all input states.We refer to this phenomenon as gate superreplication.The discovery by Dür et al opened the question whether superreplication can be achieved not only for phase gates, but also for arbitrary quantum gates.In Ref. [33] we answered the question in the affirmative, constructing a universal quantum network that replicates completely unknown unitary gates.The network has vanishing error on almost all input states, except for a vanishingly small fraction of the Hilbert space.These "bad states" can be characterized explicitly, making it easy to identify the applications where gate superreplication can be safely employed.
In this paper we review the key facts about superreplication of states and gates, unifying the ideas presented in the literature and emphasizing the connections between superreplication and other tasks, such as estimating quantum gates and generating states using gates as oracles.In addition to the review part, the paper contains a number of new results: 1. We show that, whenever quantum superreplication is achievable, it can also be achieved through estimation.
However, the error for estimation-based strategies will vanish with a power law, while the error for the genuine quantum strategies vanishes faster than any polynomial.
2. We establish an equivalence between optimality of the Heisenberg scaling in quantum metrology and the ultimate limit on the rate of superreplication, encapsulated in Eq. ( 2).On the one hand, the Heisenberg scaling implies the impossibility to produce more than M = O(N 2 ) copies of a clock state with vanishing error [28].On the other hand, here we prove the converse result, showing that the limit on the replication rate set by Eq. ( 2) implies the optimality of the Heisenberg scaling.
3. We explore the possibility to achieve superreplication without acting globally on all the N input copies and on M − N additional blank copies.Specifically, we consider strategies where the N input copies are divided into subgroups and each subgroup generates new copies, mimicking a scenario where groups of cells fuse together and then split into approximate clones.For replication strategies of this form, we show that interactions among groups of O(M 2 /N 2− ) particles are necessary and sufficient to achieve superreplication.With respect to the original superreplication protocol, the modified protocol offers a reduction of the size of the interactions by a factor M/N 2− , which is asymptotically large for all the replication rates allowed by Eq. (2).
4. We construct a simple protocol that estimates an unknown unitary gate with mean square error log N/N 2 .The protocol uses gate superreplication to produce M = Θ(N 2 / log N ) copies and gate tomography to estimate the gate within a mean square error of size 1/M .The resulting scaling of the error is close to the optimal scaling 1/N 2 [34][35][36][37].

5.
We analyze the task of generating M copies of a quantum state U |0 given N uses of a completely unknown unitary gate U [33].In this setting we show that M = Θ(N 2− ) copies of the state can be generated with fidelity going to 1 in the large N limit, for every > 0.
6.We examine the replication of quantum gates in the worst case over all possible input states.In the worst-case scenario, we establish an asymptotic no-cloning theorem for quantum gates, stating that the number reliable copies of the gate scales as , where f (N ) vanishes in the large N limit.
The paper is organised as follows.Section II introduces the task of asymptotic cloning and precisely analyses the phenomenon of quantum state superreplication.Section III analyses the relationship between state superreplication and quantum metrology.Section IV discusses the scale of the interactions needed for state superreplication.In Section V we review the existing results on the superreplication of quantum gates.Several applications of gate superreplication are presented in Section VI.Finally, the conclusions are drawn in Section VII.

II. QUANTUM STATE SUPERREPLICATION
Deterministic cloners.Although perfect cloning of non-orthogonal quantum states is forbidden by the no-cloning theorem [1,2], approximate cloning can be realized to various degrees of accuracy, depending on the set of states to be cloned [10][11][12][13][14][15].Consider a scenario where one is given N identical systems, each of them prepared in the same |ψ x , and the goal is to generate M systems in a state close to M perfect copies of the state |ψ x .For the moment, we assume the cloning process to be deterministic, meaning that it produces approximate clones with unit probability.Mathematically, a deterministic cloner is described by a completely positive trace-preserving linear map C, a.k.a. a quantum channel, sending states on the Hilbert space H ⊗N to states on the Hilbert space H ⊗M , where H is the Hilbert space of a single copy.The cloning accuracy can be quantified by the worst case fidelity between the ideal state of M perfect copies and the actual output state of the cloner.Explicitly, the average fidelity is given by where p x is the prior probability of the state |ψ x .The quantum channel that maximizes the fidelity is the optimal N -to-M cloner for the ensemble {|ψ x , p x }.In general, the form of the optimal channel varies for different ensembles, depending both on the states and on their prior probabilities.
Example: the equatorial qubit cloner.The general settings of optimal cloning can be nicely illustrated in the example of equatorial qubit states [38][39][40].Here, the ensemble consists of states of the form |ψ t = (|0 + e −it |1 )/ √ 2 where θ is drawn uniformly at random from the interval [0, 2π).The input state can be expanded as where {|N, n | n = 0, . . ., N } is the orthonormal basis of the Dicke states, defined as π being a permutation of the N qubits, U π being the unitary operator that implements the permutation π, and the sum running over the symmetric group S N .
For convenience of discussion, we focus on the case where both N and M are even.In this case, the optimal cloning channel has the simple form C(ρ) = V ρV † , where V is the isometry defined as [40] V Plugging the above relation into the definition of the average fidelity (3), one obtains the optimal value The asymptotic no-cloning theorem.Inspired by the asymptotic framework of information theory, we now focus on quantum cloners in the limit N → ∞.We model the cloning process as a sequence of cloners (C N ) N ∈N , where C N transforms N copies into M (N ) approximate copies for a given function M (N ).We call the cloning process reliable, if the cloning fidelity goes to one in the asymptotic limit, namely lim The key question here is how many extra copies can be produced reliably.Can we produce N extra copies or more?To provide a rigorous answer, we define the rate of a cloning process as In words: if the rate is α, then the number of output copies grows as M (N ) = N + Θ(N α ).We say that a cloning rate α is achievable iff there exists a reliable cloning process that has rate equal to α.For a given set of states, the task is to find the maximum achievable rate over all cloning processes.In the case of deterministic processes, the following theorem places a tight limit on the number of copies that can be produced reliably: Theorem 1 (Asymptotic no-cloning theorem for quantum states [28]).No deterministic process can reliably clone a continuous set of quantum states at a rate α ≥ 1.
The proof idea comes from the standard quantum limit of metrology [29] which limits the precision in the deterministic estimation of a parameter t encoded into N product copies of a state |ψ t .The standard quantum limit states that the mean square error vanishes as c/N , where c is a constant that depends on the encoding t → |ψ t .Intuitively, a deterministic cloner violating theorem 1 would contradict the standard quantum limit, because one could increase the precision by first cloning the probe states and then measuring them.
The argument based on the standard quantum limit of metrology is partly heuristic.A complete argument can be made for clock states, i. e. quantum states of the form The argument is conceptually independent of the standard quantum limit and allows one to prove a strong converse of the asymptotic no-cloning theorem: Theorem 2 (Strong converse of the asymptotic no-cloning theorem [28]).Every deterministic process that clones clock states at a rate α > 1 will have necessarily a vanishing fidelity in the large N limit.
The asymptotic no-cloning theorem implies that a deterministic cloning process cannot produce more than a negligible number of extra replicas in the asymptotic limit.Let us check it a few examples.The first example is the cloning of equatorial qubit states, introduced earlier in the paper.In the large N limit, the cloning fidelity is given by F ≈ 2 √ M N /(M + N ) [cf.Eq. ( 4)].It is straightforward to see that any achievable cloning rate must be smaller than one.Another example is the universal cloning of pure states.For d-dimensional quantum systems, the optimal fidelity is F = d+N −1 N / d+M −1 M [41].For large N and M , the fidelity scales as (N/M ) d−1 and converges to one if and only if the cloning rate satisfies the condition α < 1.The example of universal cloning shows that one can always find a deterministic process that reliably produces M = N + Θ(N δ ) copies, for every desired exponent δ > 0.
It is important to stress that the asymptotic no-cloning theorem holds for continuous sets of states, but does not place any restriction on the cloning rate on discrete sets of states.For example, every finite set of quantum states can be cloned with vanishing error in the large N limit, no matter how large is M [15].Hence, for finite sets of states every rate is achievable.Probabilistic cloners.The asymptotic no-cloning theorem poses a stringent limit on the ability to replicate information.In the following we will see that the limit can be broken by allowing the cloner to be probabilistic.
To analyze probabilistic cloners it is useful to introduce the notion of quantum instrument [42,43].A quantum instrument consists of an indexed set of completely positive, trace non-increasing maps (a.k.a.quantum operations) {M 1 , M 2 , . . .}.When an input state ρ is fed into the quantum instrument, the quantum operation M i occurs with probability p i = Tr[M i (ρ)] and the instrument outputs a quantum system in the state ρ = M i (ρ)/p i .In the context of our cloning problem, we consider a quantum instrument consisting of two quantum operations M yes and M no , with M yes describing the realization of a successful cloning process.When acting on the N -copy input state |ψ x ⊗N , the probabilistic cloner succeeds with probability in which case it produces the M -copy output state Conditionally on the success of the cloning process, the average fidelity is equal to where p(x|yes) is the conditional probability given by Bayes' rule.Inserting Eqs. ( 5) and ( 6) in the above expression, we obtain the explicit formula The maximum fidelity over all possible quantum instruments represents the ultimate performance allowed by quantum mechanics, even if the probability of success is arbitrarily small.In the following we will focus on the maximum fidelity in the asymptotic limit of large N and M .
Superreplication of equatorial qubit states.Let us start from the simple example of equatorial qubit states.In this case, the optimal probabilistic cloner is given by a quantum instrument with successful operation M yes (ρ) = QρQ † given by (recall that we are assuming N and M even for simplicity of presentation).Inserting the expression of the optimal cloner in the the fidelity (8), we obtain the optimal value having used Hoeffding's inequality.One can immediately see that any cloning rate α < 2 is achievable: in this case, the cloning error vanishes with N faster than the inverse of every polynomial.Interestingly, such rapid decay of the error is impossible for deterministic cloners, whose error can vanish at most as 1/N 4 [28].In summary, allowing the cloner to claim failure results into two advantages: i) the cloning rate can be boosted beyond the limits of the asymptotic no-cloning theorem, and ii) the cloning error vanishes with N at a speed that could not be achieved by deterministic cloners.The contrast between deterministic and probabilistic cloners can be also seen in the nonasymptotic setting: for example, when N = 20 and M = 120, the probabilistic fidelity maintains the high value 94.52%, while the deterministic fidelity drops to 69.39%.
The optimal probabilistic cloner dramatically outperforms all deterministic cloners in terms of rate and accuracy.However, these advantages do not come for free: the price to pay is that the probability of success vanishes in the large N limit.For example, the probability of success for qubit cloner of Eq. ( 9) is given by Using Stirling's approximation one can see that the above probability satisfies the asymptotic equality This example indicates a trade-off between the cloning rate and the success probability.In the following we will see that such trade-off is a generic feature of state superreplication.
General case: superreplication of quantum clocks.The superreplication of equatorial qubit states can be generalised to the broad family of clock states, of the form |ψ t = e −itH |ψ with t ∈ R and H † = H.For every family of clock states, one can find a probabilistic N -to-M cloner that achieves fidelity [28] F Here, K is the number of distinct energy levels of the Hamiltonian H and p min is the smallest non-zero probability in the probability distribution of the energy for the state |ψ .It is immediate to see that every cloning rate α < 2 is achievable, and thus clock states can be superreplicated.Again, the cloning error vanishes with N faster than the inverse of every polynomial-a fast scaling that could not be achieved by deterministic cloners.In other words, the probabilistic cloner produces not only more but also better replicas.
The exponential decay of the success probability is a necessary condition for superreplication of clock states: indeed, every reliable superreplication process must have a success probability that is vanishingly small compared to exp [−M/N ].In other words, as soon as a cloning process violates the asymptotic no-cloning theorem, the probability of the process must vanish in the large N limit.A strong converse result can also be proven [28]: Theorem 3. Every cloning process with rate α ≥ 1 and success probability of order exp [−M (N )/N ] or larger will necessarily have vanishing fidelity.
On the other hand, it is possible to show that superreplication can be achieved with success probability p yes > exp −M/N 1− for every > 0.
Informally, we can think of superreplication as a lottery where one invests the initial copies of the state in the hope to obtain a much larger number of approximate copies.Despite the low probability of success, the superreplication lottery has almost no risk of loss: indeed, it is possible to guarantee that, when replication fails, the cloner returns the N input copies, up to an error that is vanishingly small compared to exp [−M/2N ].This result follows from the gentle measurement lemma [30].Since the error is small, the input copies can still be used for deterministic cloning.Recall that the cloning error of deterministic cloners vanishes at most as 1/N 4 .Hence, the error introduced by the failed attempt at superrplication is negligible compared to the cloning error.In summary, one can achieve the same asymptotic performances of the optimal deterministic cloning, while still allowing for the chance to observe the exotic superreplication phenomenon.
Building on this observation, we can construct a cloner that makes a series of attempts to copy clock states probabilistically, with every failed attempt resulting into a small deterioration of the input copies [44].This cloner increases the probability of success, at the price of a performance that decreases with the number of attempts.In Figure II we show a plot of the fidelity as a function of the total probability of success in the case of equatorial qubit states with N = 50 and M = 1000.From the plot one can see the high fidelity/low probability region corresponding to the first attempts and the rapid decrease of the fidelity with the number of attempts.
The Heisenberg limit for superreplication.We have seen that clock states can be superreplicated for every cloning rate smaller than two.Higher cloning rates are forbidden by the following Theorem 4 (Heisenberg limit for superreplication [28]).In finite dimensions, no physical process can achieve a cloning rate larger than two for a set of states containing clock states.shows the tradeoff between fidelity and success probability for N = 50 and M = 1000.The points on the blue line represent successive attempts to generate M copies through the optimal probabilistic process.The black dashed line represents the fidelity of the optimal deterministic cloner.In this example, the errors introduced by failed attempts cause the fidelity to fall below the optimal deterministic fidelity when the probability of success becomes larger than 20%.
The origin of the limit on the cloning rate is the Heisenberg limit of quantum metrology [29], which can be extended to probabilistic strategies in the case of finite dimensional systems [28].The intuitive idea is the following: the Heisenberg limit implies that the probabilistic estimation of t from the state |ψ t ⊗N has a minimum mean square error scaling as N −2 .If one could produce M = Θ(N α ) clones with sufficiently small error, then these clones could be used to estimate t.Performing individual measurements on the clones and collecting the statistics, one could make the mean square error as small as as O(1/N α ).Clearly, the Heisenberg limit implies α ≤ 2.
The argument based on the Heisenberg limit is partially heuristic, because is relies on the assumption that the replication error vanishes sufficiently fast.However, it is possible to make a complete argument based on the direct optimization of the probabilistic cloner.This argument is conceptually independent of the Heisenberg limit and allows to prove a strong converse [28]: Theorem 5 (Strong converse of the Heisenberg limit for superreplication).Every physical process that replicates clock states with replication rate α > 2 will necessarily have vanishing fidelity, no matter how small is its probability of success.
Note however that the restriction to finite-dimensional systems is essential for the above conclusions.For infinite dimensional systems, the probabilistic Heisenberg bound can easily become trivial.For example, consider the set of all coherent states with fixed amplitude r > 0, which can be seen as an infinite-dimensional example of clock states with a |n = √ n |n − 1 , ∀n ∈ N. Now, coherent states with known amplitude can be probabilistically cloned with fidelity arbitrarily close to 1 [22], by using Ralph and Lund's noiseless probabilistic amplifier [23].This means that we can pick every function M (N ) and build a probabilistic cloner that transforms N input copies into M (N ) approximate copies with error smaller than, say, 1/N .The cloning process constructed in this way is reliable and can have every desired rate α, thus breaching the Heisenberg limit α < 2. The catch is, of course, that the probability of success will decay faster for processes with higher rate.Pandey et al provide a general bound [45], implying that a replication process with rate α must satisfy the relation State superreplication beyond clock states.The superreplication of quantum clocks implies the superreplication of other relevant families of states, including, the multiphase covariant states To see that these states can be superreplicated, it is enough to consider the family of clock states √ n j |j j|, where n j is the j-th prime number.With this choice, the closure of the set {|ψ t | t ∈ R} coincides with the full set of multiphase covariant states {|ψ θ | θ ∈ [0, 2π) ×d−1 }.On the other hand, an N -to-M cloner for the clock states will achieve fidelity independently of t.By taking limits, we then obtain that every multiphase covariant state can be cloned with the same fidelity.
Another multiparameter family of states that can be superreplicated is the family of all maximally entangled states.The superreplication of maximally entangled states will be discussed in detail in Section V.
It is important to note that not all quantum states can be superreplicated.First of all, the symmetry of the input ensemble can sometime inhibit superreplication.For instance, if one tries to clone an arbitrary unknown finite dimensional state, the probabilistic cloner will yield the same fidelity as the optimal deterministic cloner [28].The lack of probabilistic advantages for universal cloning is an appealing feature from the point of view of the many-worlds interpretation of quantum mechanics [46], and can be even viewed as an axiom: Axiom 1 (Many-Worlds Fairness).The maximum rate at which a completely unknown state can be cloned is the same in all possible worlds.
Interestingly, Many-Worlds Fairness rules out the variant of quantum mechanics on real Hilbert spaces, originally considered by Stueckelberg [47] and recently analyzed by Hardy and Wootters [48,49] .The argument to exclude quantum mechanics on real Hilbert spaces is the following: if the wavefunction had only real amplitudes, then one branch of the wavefunction would allow one to superreplicate every state of a quantum bit, while the other branches would not, in violation of the Many-Worlds Fairness.The fact that ordinary quantum mechanics allows for complex amplitudes results in the Many-World Fairness property and in the impossibility of superreplicating completely unknown pure states.
Another limitation to superreplication can be observed for coherent states of the harmonic oscillator, parametrized as When z is chosen at random from a Gaussian distribution p(z) = λ/π e −λ |z| 2 , λ > 0, the optimal probabilistic fidelity is given by [19] F From the above expression we see that the fidelity vanishes for all replication rates α > 1.This fact is in stark contrast for situation for coherent states with known amplitude [22,23], for which every rate is achievable.Eq. ( 11) implies that superreplication with rate α = 1 is still possible, provided that λ is sufficiently large.For example, for λ ≥ 1 one can probabilistically transform N copies into M = 2N copies, thus violating the asymptotic no-cloning theorem.Here it is important to stress the role of the prior information: in the lack of it (λ = 0), the fidelity of the optimal probabilistic cloner is equal to the fidelity of the optimal deterministic cloner, originally proposed by Braunstein et al [50].

III. THE RELATION BETWEEN STATE SUPERREPLICATION AND STATE ESTIMATION
Superreplication via state estimation.A fundamental fact about quantum cloning is its asymptotic equivalence with state estimation [10][11][12][13][14][15]: when the number of output copies becomes large, the optimal cloning performance can be achieved by a classical cloning strategy, which consists in measuring the input copies and using the outcome of the measurement as an instruction to prepare the output copies.In the standard setting, one fixes the number of input copies N and lets the number of ouput copies M go to infinity.For probabilistic cloning, this scenario has been analyzed by Gendra et al [51], who proved the asymptotic equivalence between probabilistic cloning and probabilistic state estimation.In the case of clock states that are dense in an f -dimensional manifold, Gendra et al showed that the probabilistic fidelity decays as . An open question is whether probabilistic estimation also allows to achieve superreplication.We now show that this is indeed the case.
Consider a classical cloning strategy for the states {|ψ x | x ∈ X}, based on probabilistic estimation of the parameter x and on the preparation of the state |ψ x ⊗M , where x is the estimate.This is not the most general classical strategy, but will be sufficient to establish the possibility of superreplication.The successful quantum operation is given by where {P x } x∈X are positive operators satisfying the normalization condition for some operator P ?≥ 0 associated to the unsuccessful outcome of the estimation strategy.The fidelity of the strategy can be computed using Eq. ( 7), which gives E denoting the expectation with respect to the probability distribution p(x, x|yes) := p(x|yes) p(x|x).Using the convexity of the function f (y) = y M , we then obtain the inequality or equivalently meaning that the M -copy fidelity of our measure-and-prepare strategy is lower bounded by the M -th power of the single-copy fidelity.Now, suppose that the probabilistic single-copy fidelity approaches 1 as 1/N β for some β > 0, namely Then, Bernoulli's inequality yields the bound From the bound it is clear that every replication rate α < β can be achieved: the fidelity converges to 1 as long as M (N )/N β vanishes in the large N limit.In summary, we have proven the following Theorem 6.Let S = {U x | x ∈ X} be a continuous set of gates.A probabilistic strategy that estimates the gates in S with fidelity larger than 1 − O(1/N β ) can be used to replicate the states in S at every rate α < β via a measure-andprepare strategy Theorem 6 applies to arbitrary sets of states and to probabilistic strategies with arbitrary constraints on the probability of success.In the case of clock states, assuming no constraint on the probability of success, the singlecopy fidelity of phase estimation has the Heisenberg scaling [52].Hence, the classical strategy described above achieves superreplication for every replication rate α < 2. Note, however, that the classical superreplication strategy has an error vanishing with the power law 1/N 2−α , while the quantum superreplication strategy has an error vanishing faster than every polynomial.
Deriving precision limits from superreplication.We have seen that the replication rate for clock states is determined by the Heisenberg limit in the probabilistic case and by the standard quantum limit in the deterministic case.On the other hand, we also seen strong converse results (Theorems 2 and 5) that prove the optimality of the replication rates without invoking the precision limits of quantum metrology.In fact, the quantum metrology limits can be derived from our our bounds on the replication rates.The argument is by contradiction: 1. Suppose that one could violate the standard quantum limit, using N copies of the clock state |ψ t to estimate the parameter t deterministically with error scaling as 1/N 1+ , > 0.Then, Eq. ( 13) would imply that one can produce M clones with M = Θ(N 1+ /2 ), thus violating Theorem 2.
2. Suppose that one could violate the Heisenberg limit of probabilistic metrology, using N copies of the clock state |ψ t to estimate the parameter t probabilistically with error scaling as 1/N 2+ , > 0.Then, Eq. ( 13) would imply that one can produce M clones with M = Θ(N 2+ /2 ), thus violating Theorem 5.
In conclusion, we have shown a complete equivalence between the limits on the scaling of the error in quantum metrology and the limits on the replication rate set by Theorems 2 and 5.

IV. SUPERREPLICATION WITH REDUCED INTERACTION SIZE
The divide-and-clone approach.To realise the optimal N -to-M probabilistic cloner, a global interaction involving at least M systems is required.This can be seen in the example of the equatorial qubit cloner, which is defined as an evolution acting coherently on the basis of the Dicke states, cf.Eq. ( 9).Of course, the need for interactions among large numbers of systems makes the implementation of cloning a challenging issue.Here we investigate the possibility to reduce the scale of the interactions.
A simple strategy to reduce the interaction scale is a "divide-and-clone" strategy, where one divides the input copies into groups and performs optimal cloning on each group.Suppose that we divide the N input copies into groups of N := Θ(N β ) copies.Applying the optimal N -to-M probabilistic cloner to each individual group, we then achieve fidelity for each group.The value of M can be chosen in order to reach the desired replication rate: in order to obtain M copies overall, one needs M to satisfy the condition M = M N/N .For a replication process producing M = Θ(N 1+δ ) copies, the condition yields M = Θ(N δ+β ).Since there are N/N groups, the overall fidelity of this strategy is From the above bound it is immediate to see that the fidelity converges to 1 whenever β > δ.Hence, superreplication of quantum clocks can be achieved with interactions among M = O N 2δ+ = O M 2 /N 2− particles, for every desired > 0. Now, recall that the original superreplication strategy [28] requires interactions among O(M ) particles.This means that the modified strategy leads to a reduction of the interaction scale by a factor N 2− /M .On the other hand, the modified strategy does not allow to achieve superreplication with interactions among less than Θ(N δ ) systems.Indeed, superreplication can only be achieved if the fidelity of the individual processes approaches 1.By the Heisenberg limit, this condition is satisfied only if N and M satisfy the asymptotic relation M (N ) 2 .Recalling that M has to scale as N β+δ and that N scales as N β , we then obtain that β > δ is a necessary condition.In summary, the strategy of dividing the input copies into non-interacting groups achieves superreplication with M = Θ(N 1+δ ) output copies if and only if the size of each group grows faster than N δ = M/N .
A sequential cloning approach.Another approach to reduce the scale of interactions is to generate clones by local mechanisms.For example, we can imagine a cloning process that involves repeated interactions among O(N ) systems.Let us analyze this idea in the case of equatorial qubit states: here we consider a sequence of cloners that take N to 2N copies, arranged in the following circuit: Each wire in the circuit represents a composite system of N qubits.At each step, the quantum operation M (N ) performs the optimal N -to-2N probabilistic cloning, given by the map as one can see substituting M with 2N in Eq. ( 9).In the following it will be convenient to express the map as We now analyze the performance of the sequential cloner after K − 1 steps, which result into the generation of KN approximate copies.At the first step, the input state is the N -copy state where the subscript 1 indicates the first group of N qubits.The first cloner transforms |Ψ t,0 into a state on the first and second group, yielding 2N qubits in the state where N 1 is a normalization constant and the set S n,1 is defined via the relation Iterating the cloning process for K − 1 steps we finally obtain K groups of qubits in the joint state , where c n,x is the coefficient given by The overall fidelity of the cloner with the state of N K perfect copies does not depend on t and is given by For K > 2, the cloning fidelity ( 15) is strictly smaller than the fidelity (8) of the optimal probabilistic cloner.A comparison of the sequential cloner, the optimal probabilistic cloner and the optimal deterministic cloner is illustrated in Fig. 2 for N = 8.Note that the fidelity of the sequential cloner decays quickly with the number of cloning steps.
In our example, the fidelity of the sequential cloner approaches the fidelity of the deterministic cloner as the number of output copies approaches M = N 2 = 64.This numerical result suggests that it may not be possible to achieve superreplication sequentially.However, it is an open question whether or not superreplication can be achieved by the the sequential cloner or by similar mechanisms.For example, one could construct a quantum cellular automaton for cloning, as proposed by D'Ariano, Macchiavello and Rossi in [53].They considered a tree-shaped network of deterministic 1-to-2 cloners.In order to achieve super-replication, one would have to extend the model, allowing for probabilistic cloners.Moreover, symmetry arguments imply that probabilistic 1-to-2 cloners do not offer any advantage over their deterministic counterparts [28].Hence, one has to consider local cloners with larger number of input copies.An interesting possibility is to construct a probabilistic cellular automaton where the building blocks are the optimal k-to-2k probabilisitc cloners with k > 1.

V. QUANTUM GATE SUPERREPLICATION
We have seen that state superreplication has necessarily a vanishing success probability in the asymptotic limit.Here we analyze the task of replicating quantum gates, where superreplication can be achieved with unit probability of success on average over all input states.Gate superreplication.In many applications, including quantum metrology and quantum algorithms, one is given access to a black box implementing an unknown quantum gate.In these applications the uses of the gate are a resource: indeed, a no-cloning theorem for gates asserts that it is impossible to perfectly simulate two uses of an unknown gate by using it only once [31].Still, a natural question is: how well can we simulate M uses of the unknown gate with N < M uses?Here the problem is to engineer a quantum computational network that uses the unknown gate as a subroutine, as in figure 3.In analogy with the superreplication of quantum states, we say that superreplication of quantum gates is possible iff one can find a sequence of networks with the property that the simulation error vanishes in the asymptotic limit and the number of extra copies of the input gate grows as N or faster.In the following we will see that gate superreplication can be achieved deterministically for almost all input states.The first result of this type was discovered by Dür, Sekatski, and Skotiniotis [32] in the case of phase gates, that is, gates of the form The performance of the simulation was quantified by the fidelity between the output state of the simulation and the output state of M perfect uses of the gate U t , averaged over t and over all possible input states.Using a deterministic network, the authors showed how to simulate M N 2 uses with asymptotically unit fidelity.This result can be extended from phase gates to arbitrary gates [33], as discussed in the following paragraphs.network.With this notation, the average fidelity reads where dU is the normalized Haar measure over the group of all unitary gates dΨ is the normalized Haar measure over the manifold of pure M -partite states.By Markov' inequality, a simulation with gate fidelity F gate [N → M ] ≥ 1 − is a simulation that works with fidelity F ≥ 1 − δ on all pure states except for a small fraction, whose probability is smaller than /δ.A replication process is described as a sequence of networks, with the N -th network transforming N copies of the unknown gate into M (N ) approximate copies.We say that the replication process is reliable iff lim In other words, for a reliable replication process one has F [N → M (N )] ≥ 1 − N for some N converging to zero.Choosing δ N = √ N , we then have that the replication process simulates the desired gate with fidelity larger than 1 − N on all pure states except a small fraction, whose probability is smaller than N /δ N = √ N .As in the case of state replication, we say that a process has replication rate and we say that the rate α is achievable iff there exists a reliable replication process with that rate.
In the following, we will see that every rate smaller than 2 is achievable.For simplicity, we discuss the case of qubit gates, although the protocol can be extended in a straightforward way to unitary gates on aribtrary finite-dimensional quantum systems.
The key to the construction of the universal gate cloning network is the decomposition of the Hilbert space of K qubits into orthogonal subspaces corresponding to different values of the total angular momentum.In formula, one has where j is the quantum number of the total angular momentum, R j is a representation space and M j is a multiplicity space [54].With respect to this decomposition, the action of K parallel uses of a generic qubit unitary U has the block-diagonal form where I jK is the identity on the multiplicity space M jK .
Note that U ⊗K acts non-trivially only in the representation spaces.This means that, essentially, the multiplicity spaces can be eliminated without losing any information about U .Likewise, they can be inflated by adding ancillas.The working principle for the simulation of U ⊗M given U ⊗N will be to conveniently adjust the size of the multiplicity spaces, as in the following Protocol 1 (Gate superreplication network [33]).To construct the replication network, one has to 1. Choose an ancilla A such that the rank of I jN ⊗ I A is no smaller than the rank of I jM for every j ∈ [0, N/2]; 2. Compose the gate U ⊗N with the identity on A, thus obtaining the gate 3. Project the resulting gate U ⊗N ⊗ I A into the subspace where the gate acts as 4. Embed the subspace H N into the Hilbert space H ⊗M .With a slight abuse of notation we use H N both for the subspace of H ⊗N ⊗ H A and for the corresponding subspace of H ⊗M .
5. Given an input state of M qubits, perform a projective measurement that projects the qubits either in the subspace H N or in its orthogonal complement H ⊥ N .If the qubits are projected into H N , then apply the gate U .If the qubits are projected into H ⊥ N , then perform the identity.The network described by protocol 1 perfectly imitates U ⊗M for any state in the subspace H N ⊂ H ⊗M .Now, it is possible to show that a random pure state in H ⊗M has a large overlap with the subspace H N whenever M N 2 .Indeed, we have the following Lemma 1.Let F N be the fidelity between the pure state |Ψ ∈ H ⊗M and its projection |Ψ N = P n |Ψ / P N |Ψ .The expectation value of the fidelity over the uniform measure is given by In the above inequality d j and m jM are the dimensions of the representation space R j and the multiplicity space M jM , respectively.The detailed proof of the last inequality can be found in the supplemental material of Ref. [33].
Using the above lemma, it is easy to show that the fidelity is close to 1 whenever M is small compared to N 2 .Indeed, the cloning fidelity ( 16) is lower bounded as In conclusion, the gate fidelity converges to 1 for all replication processes with replication rate α < 2, establishing the possibility of gate superreplication.Similar bounds can be found for arbitrary d-dimensional systems, in which case we have The inequality follows from the concentration of the Schur-Weyl measure [55,56], which has applications in quantum estimation [57] and information compression [58,59].The bound (17) allows one to produce M = Θ(N 2− ) copies with a super-polynomially fast vanishing error.In addition, one can produce M = Θ[N 2 / log N ] copies with an error vanishing faster than 1/N k for every desired k > 0. For this purpose, it is enough to choose so that Eq. ( 17) becomes As a final remark, we note that the size of the interactions used in the superreplication network can be reduced from O(M ) qubits to O(M/N ) qubits, by dividing the N input gates into groups of O(N 2 /M ) gates.This fact can be proven with the same argument used for quantum states.

VI. APPLICATIONS OF GATE SUPERREPLICATION
Superreplication of maximally entangled states.Inspired by gate superreplication, one can construct a protocol for superreplicating maximally entangled states, which achieves any cloning rate α < 2 [33].For qubits, we can cast an arbitrary maximally entangled state in the form The replication protocol works as follows: First, the unitary gate U can be extracted from the state via a gate teleportation protocol [60].The success probability of a single gate extraction is 1/4, which implies that U ⊗N can be extracted from N copies of the maximally entangled state with probability (1/4) N .The extracted gates are then used in the gate superreplication network, which simulates the gate U ⊗M with vanishing error whenever M grows as N α , α < 2. Finally, the approximation of U ⊗M is applied locally on M copies of the Bell state |Φ + , thus providing M approximate copies of the state |Φ U .The fidelity of the replicas can be bounded as having used the relationship between the entanglement fidelity and the gate fidelity derived by Horodecki et al [61].
The above equality implies that the fidelity of the state cloning protocol goes to one if and only if the fidelity of gate superreplication goes to one.Therefore, the above protocol is able to replicate maximally entangled states at every rate α < 2.
The idea of gate teleportation followed by gate superreplication can also be applied to achieve superreplication of other families of states.For instance, gate superreplication yields is an alternative superreplication protocol for equatorial qubit states.Given N copies of an equatorial qubit state (|0 + e −it |1 )/ √ 2, one can first generate N copies of the maximally entangled qubit state (|00 + e −it |11 )/ √ 2 by applying a CNOT gate to each input copy.Then, the maximally entangled qubit state can be superreplicated.Finally, applying a CNOT gate to each output copy allows to transform the (approximate) copies of the maximally entangled qubit state into (approximate) copies of the equatorial qubit state.The net result of this protocol is the superreplication of equatorial qubit states.
Similar arguments apply to the superreplication of maximally entangled states in arbitrary finite dimensions.Every maximally entangled state can be parametrized as |Φ U = (U ⊗ I)|Φ + , where n=0 |n |n .Again, the unitary gate U can be probabilistically extracted from the state |Φ U , with probability p = 1/d 2 .Hence, superreplication of maximally entangled states can be achieved by i) simulating M copies of the gate U and ii) applying the simulated gates locally on M copies of the state |Φ + .The fidelity is then given by again, having used the relationship between the entanglement fidelity and the gate fidelity [61].In conclusion, the cloning fidelity for maximally entangled states approaches 1 if and only if the cloning fidelity for unitary gates approaches 1.Hence, the superreplication of quantum gates implies the (probabilistic) superreplication of maximally entangled states.In turn, the supereplication of maximally entangled states implies the superreplication of the uniform-weight multiphase states which are unitarily equivalent to the maximally entangled states Upper bound on the rate of gate replication.The relation between gate replication and replication of maximally entangled states can be used to derive the ultimate limit on the rates of gate replication in finite dimensions.The argument proceeds by contradition: Suppose that it is possible to devise a network that simulates Θ(N α ) uses of U for α ≥ 2 with vanishing error.Then we could use it to achieve superreplication of maximally entangled states at rate α ≥ 2. Now, the family of maximally entangled states contains the family of clock states Replication of these states at rate α ≥ 2 would contradict Theorem 5, which implies that every cloner with rate α ≥ 2 must have vanishing fidelity.Hence, the possibility of achieving gate superreplication at rates larger than quadratic is excluded.Note that the above argument applies not only to deterministic gate replication networks, but also to probabilistic networks.Using the correspondence between entanglement fidelity and gate fidelity, we obtain the following: Theorem 7. Every physical process that replicates phase gates with replication rate α > 2 will necessarily have vanishing gate fidelity (no matter how small the probability of success).
An alternative optimality proof for the quadratic replication rate was presented by Skotiniotis, Sekatski, and Dür [62] who devised an argument to bound the replication fidelity based on the no-signalling principle.Supergeneration of maximally entangled states.Gate superreplication can be achieved deterministically, while state superreplication can only be achieved with a vanishing probability.The origin of this sharp difference is in the fact that states and gates are inequivalent resources: while gates can be used to deterministically generate states as |ψ U = U |0 , the converse process is forbidden by the no-programming theorem [63].It is then useful to distinguish between the task of state cloning, where the input consists of N copies of the state |ψ U = U |0 , and the task of state generation, where the input consists of N copies of the gate U .Deterministic gate superreplication cannot be used to achieve deterministic state superreplication, but can be used to achieve deterministic state supergeneration, that is, the generation of up to N 2 almost perfect copies of the quantum state |ψ U from N copies of the gate U .A general supergeneration protocol works as follows: 1. Use a gate superreplication protocol to simulate M N 2 uses of the gate U 2. Applying the simulated gates on the state |0 ⊗N .
In general, the protocol can be tailored to the specific set of states that one wants to generate.For example, one could have i) clock states, where U is of the form U = e −itH , ii) maximally entangled states, where U is of the product form U = V A ⊗ I B with respect to some bipartition of the Hilbert space, and iii) arbitrary pure states, where U is a generic unitary.
The fidelity of supergeneration depends on the protocol used to replicate the gates.For example, the phase gate replication by Dür et al allows to supergenerate clock states [32], while our universal gate replication allows to supergenerate maximally entangled states of bipartite systems [33].Interestingly, the universal gate replication does not work for the set of all pure states, parametrized as In this case, the approach of simulating M uses of the gate U and applying it to the state |0 ⊗N does not work, because the state |0 ⊗N lies in the subspace H ⊥ N where our gate superreplication network fails.However, we will see in a couple of paragraphs that supergeneration of arbitrary pure states can be achieved via a suitable protocol based on gate estimation.
Gate estimation with quasi-Heisenberg scaling.The supergeneration of maximally entangled states has an elegant application to quantum metrology.Specifically, it allows for an easy proof of the fact that an unknown quantum gate can be estimated an error scaling as log N/N 2 .Here the error is defined as e N = 1 − F est,gate [N ], where F est,gate [N ] is the fidelity of gate estimation with N copies, namely p N ( Û |U ) being the probability distribution resulting from the estimate and F gate ( Û , U ) being the gate fidelity, defined as We now show that gate superreplication can be used to achieve estimation fidelity scaling as The proof goes as follows: given N copies of the gate U , we can produce M = Θ(N 2 / log N ) copies of the maximally entangled state |Φ U , with an error vanishing faster than 1/N k , for every desired k > 0-cf.Eq. ( 18).In particular, we set k = 2.With this choice, the error vanishes faster than 1/M and the M approximate copies can be used for state estimation, resulting into an estimate of the maximally entangled state with fidelity according to the central limit theorem.Here, the estimation fidelity is defined as where p M (Φ Û |Φ U ) is the probability distribution resulting from the estimation of a maximally entangled state with M nearly perfect copies.Now, we regard the estimation of the maximally entangled state |Φ U with M copies as a particular strategy for the estimation of the unitary gate U with N copies, meaning that we have Then, the estimation fidelity for the maximally entangled state can be easily converted into the gate fidelity for the corresponding gate, using the relation [61] F Since the state estimation fidelity converges to 1 as 1/M , the gate estimation fidelity will also converge to 1 as 1/M .Recalling that M scales as N 2 / log N , this proves Eq. ( 21).The error scaling e = log N/N 2 beats the central limit scaling of classical statistics and is close to the optimal quantum scaling 1/N 2 , which was derived in Refs.[34][35][36] for qubits and in Ref. [37] for general d-dimensional systems.The usefulness of our new derivation is that the proof is much simpler than the full optimization of the estimation strategy.
It is interesting to mention that, for d = 2, an alternative estimation strategy achieving scaling log N 2 /N 2 was proposed by Rudolph and Grover [64].Their protocol is sequential and uses unentangled states to reach a quasi-Heisenberg scaling.It is an open question whether a similar protocol exists for d > 2.
Universal supergeneration of pure states.We now show that all pure states can be supergenerated.To this purpose, we parametrize the manifold of pure states as {|ψ U = U |0 | U ∈ SU(d)}, where |0 is a fixed pure state.To achieve supergeneration, we use a classical strategy based on the estimation of the unknown gate U and on the preparation of the state |ψ Û ⊗M conditional on the estimate Û .The fidelity of this strategy is given by where p N ( Û |U ) is the probability distribution resulting from gate estimation.The above choice of strategy implies that we have the bound Now, note that the single-copy fidelity satisfies the relation where F est,gate [N ] is the gate estimation fidelity defined in Eq. ( 20) and in the second equality we used the fact that the optimal gate estimation strategy is covariant [65], i. e. it satisfies the condition p Combining Eq. ( 23) with the bounds ( 22) and ( 21) we finally obtain meaning that arbitrary pure states can be supergenerated at every rate smaller than quadratic.Since the above protocol is based on estimation, the error vanishes only with a power law.It is an open question whether there exists a universal quantum supergeneration protocol whose error vanishes faster than any inverse polynomial.
Equivalence between gate superreplication and gate estimation.As in the case of states, superreplication can be achieved by via gate estimation.The argument is the same used in the derivation of Eq. ( 13): in short, one can show that every estimation strategy with fidelity scaling as F est,gate [N ] = 1 − O(1/N β ) can be used to achieve gate replication with fidelity meaning that every replication rate smaller than β can be achieved.In particular, we know from Eq. ( 21) that we can choose β = 2 − .Hence, every replication rate smaller than quadratic can be achieved via state estimation.Note that, however, the replication error goes to zero only with a power law, while the quantum replication network has a much better scaling of the error.The connection between gate superreplication and estimation can be used to prove the Heisenberg limit [29].The proof is by contradiction: Suppose that one could estimate the parameters of a unitary gate with fidelity scaling as F est,gate [N ] = 1 − O(1/N 2+ ).Then, Eq. ( 24) would imply that one can simulate M = O N 2+ /2 copies of the gate, in contradiction with Theorem 7. In summary, the quadratic scaling of the Heisenberg limit can be derived solely from considerations about the optimal rates of gate superreplication.
Asymptotic no-cloning theorem for quantum gates in the worst case scenario.We have seen that gate superreplication can be achieved on all input states except for a vanishingly small fraction.A natural question is whether one can find a superreplication protocol that works on all input states.The problem can be addressed by evaluating the worst-case fidelity where C (N ) U is the channel implemented by the replication network.In this setting, we say that a gate replication process is reliable in the worst case iff lim N →∞ F gate,worst [N → M (N )] = 1 and we say that the replication rate α is achievable in the worst case iff there exists a replication protocol that has that rate and is reliable in the worst case.The supremum over all achievable rates is determined by the following Theorem 8 (Asymptotic no-cloning theorem for quantum gates).For finite dimensional quantum systems, no physical process can reliably replicate phase gates at rate α ≥ 1 in the worst case scenario.
Here we give a heuristic proof, based on optimal phase estimation.When an unknown phase gate is used N times, the optimal estimation strategy has mean square error c/N 2 , for some suitable constant c > 0 [66,67].The bound holds both for deterministic and probabilistic strategies [68] and the fact that the optimal estimation strategy can be achieved deterministically plays a crucial role in our argument.Now, suppose that one can produce M reliable replicas, up to an error scaling as 1/M β , for some β > 2. In this case, one could use the replicas for phase estimation, reducing the error to this result following the fact that the error of a deterministic estimation strategy is a continuous function of the input state.Since the above strategy cannot be better than the optimal strategy, we obtain the inequality which implies that M can grow at most as M = N + Θ(N α ) with α < 1.In addition, by Taylor-expanding the r.h.s. of Eq. ( 26) we obtain the inequality β < 3 − α, meaning that the error can vanish at most as 1/N 3 .In summary, superreplication and exponentially vanishing errors are forbidden in the worst case scenario.

VII. CONCLUSION AND OUTLOOK
In this paper we reviewed the phenomenon of superreplication of quantum states and gates, emphasizing the applications and connections with other tasks in quantum information.In particular, we clarified the relation between superreplication and the precision limits of quantum metrology, showing that i) estimation can be used to achieve superreplication with error vanishing as O(M/N 2 ), probabilistically in the case of states and deterministically in the case of gates, ii) gate superreplication allows for a simpler proof of the quadratic precision enhancement in the estimation of an unknown gate, iii) the optimality of the Heisenberg scaling for the estimation of unitary gates can be derived from the ultimate limit to the rate of superreplication.In addition, we showed that N uses of a completely unknown gate are sufficient to generate O(N 2 ) approximate copies of the corresponding pure state, with an error that vanishes in the large N limit.
Among the future research directions, an important one is the study of replication processes for mixed states and noisy channels.We expect that also in this case there is an equivalence between replication and estimation.In the presence of noise, it is well known that the Heisenberg scaling is often inhibited, leading to different sub-Heisenberg scalings [69][70][71][72][73].We then expect that the replication rates will also have intermediate values, ranging between α = 1 and α = 2 depending on the type of noise.The study of superreplication in the noisy case is also expected to shed light on the optimal strategies for the estimation of noisy channels, whose performances are sometimes hard to characterize analytically.Another interesting research direction is the study of replication processes that are subject to constraints, e. g. on the ability to perform joint operations on composite systems.Recently, Kumagai and Hayashi [74] investigated the problem of cloning bipartite quantum states using only local operations and classical communication.When the state to be cloned is perfectly known, they showed that the number of extra copies scales as √ N .Note the contrast with the situation where the LOCC restriction is not imposed, in which case one can produce Θ(N δ ) extra copies for every δ < 1.Besides the constraints due to the locality of the operations, other physical constraints can arise from conservation laws, such as energy and angular momentum conservation [75][76][77][78][79].The search for energy-preserving cloners has been considered in our earlier work [44], where we identified the optimal operations for the transformation of pure states.In this scenario, it is interesting to examine how the replication rates are affected by the presence of limited resources, or, conversely, what are the resources needed to achieve a desired replication rate.

9 FFIG. 1 :
FIG.1: Replication of equatorial qubit states through successive attempts of probabilistic cloning.The figure shows the tradeoff between fidelity and success probability for N = 50 and M = 1000.The points on the blue line represent successive attempts to generate M copies through the optimal probabilistic process.The black dashed line represents the fidelity of the optimal deterministic cloner.In this example, the errors introduced by failed attempts cause the fidelity to fall below the optimal deterministic fidelity when the probability of success becomes larger than 20%.

FFIG. 2 :
FIG. 2:Comparison between the three cloners.In this figure we compare the performance of the network cloner to the one of the probabilistic cloner as well as the one of the deterministic cloner.The fidelity is plotted as a function of the output number, fixing the input number to N = 8.The green line (with numerics represented by red dots) represents the fidelity-output curve for the network cloner.The blue line (with numerics represented by red dots) stands for the fidelity-output curve for the probabilistic cloner, and the black line (with numerics represented by black dots) stands for the fidelity-output curve for the deterministic cloner.

FIG. 3 :
FIG. 3: Quantum gate cloning.A quantum network for gate cloning.Given N uses of an unknown unitary U (green boxes on the left), the network (blue boxes on the left) simulates M > N uses of U (green boxes on the right).