Efficient tools for quantum metrology with uncorrelated noise
aa r X i v : . [ qu a n t - ph ] J u l Efficient tools for quantum metrology withuncorrelated noise
Jan Kołodyński and Rafał Demkowicz-Dobrzański
Faculty of Physics, University of Warsaw, 00-681 Warszawa, PolandE-mail: [email protected] and [email protected]
Abstract.
Quantum metrology offers an enhanced performance in experiments suchas gravitational wave-detection, magnetometry or atomic clocks frequency calibration.The enhancement, however, requires a delicate tuning of relevant quantum featuressuch as entanglement or squeezing. For any practical application the inevitableimpact of decoherence needs to be taken into account in order to correctly quantifythe ultimate attainable gain in precision. We compare the applicability and theeffectiveness of various methods of calculating the ultimate precision bounds resultingfrom the presence of decoherence. This allows us to put a number of seeminglyunrelated concepts into a common framework and arrive at an explicit hierarchy ofquantum metrological methods in terms of the tightness of the bounds they provide. Inparticular, we show a way to extend the techniques originally proposed in Demkowicz-Dobrzański R, Kołodyński J and Guţă M 2012
Nat. Commun. ontents N independent quantum channels 13 N limit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144.1.1 Classical simulation (CS) method . . . . . . . . . . . . . . . . . . 154.1.2 Quantum simulation (QS) method . . . . . . . . . . . . . . . . . 154.1.3 Channel extension (CE) method . . . . . . . . . . . . . . . . . . . 174.2 Finite- N channel extension (CE) method . . . . . . . . . . . . . . . . . . 18 Appendix A.1 Dephasing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26Appendix A.2 Depolarization . . . . . . . . . . . . . . . . . . . . . . . . . . . 26Appendix A.3 Loss . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27Appendix A.4 Spontaneous emission (amplitude damping) . . . . . . . . . . 28
Appendix B Equivalence of RLD-based bound applicability and localclassical simulability of a channel 28Appendix C RLD-based bound as a special case of asymptotic CE bound 29 ppendix D Optimal local QS of a channel 30Appendix E Finite- N CE method as a semi-definite programming task 31 fficient tools for quantum metrology with uncorrelated noise
1. Introduction
Quantum enhanced metrology has recently enjoyed a great success at experimental levelleading to new results in atomic spectroscopy [1–4], magnetometry [5–7] and opticalinterferometry [8–11] with prominent achievements in gravitational waves sensing [12].As predicted in the incipient theoretical results of [13–17], a physical parameter unitarilyencoded into a quantum system—a probe—consisting of N entangled, non-interactingparticles (atoms or photons) can be extracted with a precision that is limited by thequantum-mechanical uncertainty relations and not the more restrictive central limittheorem of classical statistics. Hence, the uncertainty in reconstructing the encodedparameter, such as e.g. optical phase delay or frequency difference, can in principle beproportional to /N , the so-called Heisenberg Limit (HL), rather than / √ N , commonlyreferred to as the Standard Quantum Limit (SQL) or the shot (projection) noise.However, this dramatic scaling improvement can be illusive, as both the experimentalresults and theoretical toy-models have indicated that achieving the ideal HL is a hardtask owing to the strong destructive impact of imperfections, which should be alwaysaccounted for in realistic scenarios.An important question that has been considerably addressed by many researchers[18–32] reads: how and to what extent can the noise effects be compensated inquantum metrological setups? In the case of atomic spectroscopy, it has already beenindicated in [24, 25] that the effects of uncorrelated noise, independently affecting theatoms within a probe, have a dramatic impact on quantum protocols—most likelyrestricting the ultimate precision scaling to become SQL-like for high enough N , sothat the quantum enhancement is asymptotically limited to a multiplicative constantfactor . In optical interferometry, photonic loss is the main obstacle to practicalimplementations of quantum enhanced protocols [27–30] and the asymptotic SQL-likescaling is again inevitable, as proven in [31, 32]. Similarly to the atomic case, theasymptotic improvement constant becomes an essential feature which determines theachievable precision for high N . Following the above exemplary models, methods ofquantifying the asymptotic quantum enhancement have been proposed for arbitrarykinds of probes with decoherence present [33–39]. Recently, general procedures havebeen established that are capable of deriving practical bounds on ultimate achievableprecision in realistic quantum metrological setups [38, 39]. In particular, in the case ofuncorrelated noise it has been demonstrated that the sole analysis of the evolution of a single particle often leads to surprisingly informative bounds on the precision achievablewith arbitrarily entangled multi-particle inputs [39]—see Figure 1 for an outline ofa relevant metrological scheme and a summary of the single particle (single channel)methods investigated further on in this paper. For completeness, it should be notedthat there are some specific metrological models with noise, in which asymptotic scalingpower enhancement is nevertheless possible [22, 23, 40]. Still, applicability of thesemodels is limited, since in any practical implementation the noise types considered willalways be accompanied by some generic decoherence processes for which the constant fficient tools for quantum metrology with uncorrelated noise Figure 1. Quantum metrology and the single channel methods ( a ) General scheme for quantum-enhanced metrology with uncorrelated noise. The N particles within the probe in a quantum state ψ N in evolve and decohere independentlywhile sensing an unknown parameter ϕ (e.g. phase). An estimate ˜ ϕ of the parameteris inferred from a measurement result on the final state of the probe ρ Nϕ .( b ) Precision bounds from single channel analysis. The ultimate precision is boundedby the / √ N scaling (SQL), if for small variations δϕ around ϕ the channel Λ ϕ canbe expressed as a parameter-independent map Φ that is also fed a classical, diagonalstate p ϕ – CS, or a general, quantum state σ ϕ – QS, varying smoothly with ϕ . Still,for all such channels and more, the tightest bound on precision is obtained by the CEmethod, in which the map Λ ϕ is replaced by its extension Λ ϕ ⊗ I . factor bound on the maximal quantum enhancement will force the asymptotic precisionscaling to be SQL-like.In this paper, we provide new insight into relations between seemingly unrelatedmethods used for derivations of various quantum metrological bounds and order themwith respect to their predictive power. Firstly, focussing on the geometric intuitivemethod of channel classical simulation (CS) introduced in [36, 39], we prove that itsapplicability is equivalent to the approach proposed in [37]. Moreover, we show thatthe criterion for classical simulability of a channel can be generalized to a quantumsimulation (QS) condition [36], which coincides with the channel programmabilitypostulate of [34]. Although the idea of QS allows to prove the asymptotic SQL-likebehaviour for a wider range of decoherence models, we demonstrate that the channelextension (CE) method of [39] encompasses all CS, [37], QS and [34] approachesproviding more stringent bounds on precision. We also comment on the problem of fficient tools for quantum metrology with uncorrelated noise b .Most importantly, we go beyond the results of [39] and show that the CE methodmay be applied not only in the asymptotic regime but also when dealing with finiteprobes of N particles. Similarly to the asymptotic case, it corresponds then to anoptimisation procedure over Kraus representations of a given channel that can berecast into an efficiently solvable semi-definite programming task. We apply ourresults to phase/frequency estimation with various noise models including: dephasing,depolarization, loss and spontaneous emission, restricting ourselves to the cases of noisecommuting with the parameterized unitary part of the evolution. This assumptionmakes the analysis more transparent, but is not indispensable, since our methods maybe effectively employed for any single particle evolution described by a general Lindbladequation [54] reshaped into the corresponding Kraus representation [55]. What is more,as our finite- N CE method applies to models for which its asymptotic version fails, itcan be used to upper-bound the asymptotic scaling for channels surpassing the SQL.As an example, in [40], our new method has been already utilized to predict asymptoticsuper-classical scaling for a channel with a non-commutative noise. Additionally, inorder to stress the generality of the methods, in the final section of this paper weshow that they can be applied not only to noisy unitary parameter estimation tasksbut also to ones in which the decoherence strength itself is estimated. Finally, weshould also clarify that noise correlation and memory effects [26, 56–58] are beyondthe scope of this paper; while the non-Markovian effects, provided they affect each ofthe particles independently, might be analyzed using the tools presented, correlationsbetween the decoherence processes affecting the particles does not fit well into theframework advocated here.This paper is organised as follows. In Section 2 we introduce the mathematicaltools of estimation theory designed to quantify the achievable precision in metrologicalschemes and discuss their applicability in the quantum setting. In Section 3 we studyhow these concepts may be utilized when the estimated parameter is encoded duringthe evolution of a given quantum system. Section 4 is devoted to quantum systems thatconsist of N particles undergoing independent evolution and contains the main resultsof the paper—methods allowing to quantify the ultimate precision both in the finite- N and asymptotic N regimes as well as their direct application to phase estimationschemes. In Section 5 we show how these methods should be accommodated in order toencompass the frequency estimation tasks of atomic spectroscopy, whereas in Section 6when considering metrological scenarios in which the strength of noise or loss is theparameter to be estimated. Section 7 contains additional discussion on consequences ofthe results obtained as well as an outlook on future research. Section 8 summarizes the fficient tools for quantum metrology with uncorrelated noise
2. Quantum Fisher Information
Let us assume that after measuring a physical system an outcome is obtained that canbe represented by a random variable X distributed with some probability distribution p ϕ ( X ) . If the system is classically described, all its properties can be simultaneouslydetermined, so that X can in principle be multidimensional and contain as muchinformation about the system as allowed by the available resources. The estimationtask corresponds then to determining with highest precision the quantity ϕ based on theobserved value of X . As stated by the Cramér-Rao bound [59] any unbiased strategy todetermine the unknown parameter after repeating the procedure k times, must providean estimate ˜ ϕ with uncertainty that is lower bounded by ∆ ˜ ϕ ≥ p k F cl [ p ϕ ] , where F cl [ p ϕ ] = Z dx ˙ p ϕ ( x ) p ϕ ( x ) (1)is the ( classical ) Fisher Information (FI) ‡ .The / √ k dependence in (1) is a consequence of the central limit theorem andthe fact that the k procedures are independent. This manifests itself by the additivityproperty of the FI, i.e. F cl (cid:2) p kϕ (cid:3) = k F cl [ p ϕ ] for X k . Equation (1) shows that the FIis a local quantity containing information about infinitesimal variations of ϕ . That iswhy, FI is designed to be used in the so called local estimation approach in which smallparameter fluctuations are to be sensed. This small deviations regime may always bereached after many procedure repetitions ( k → ∞ ) and in this limit the Cramér-Raobound is known to be saturable via e.g. max-likelihood estimation schemes [59]. In a quantum estimation scenario, the parameter ϕ is encoded in a quantum state ρ ϕ .A general measurement, mathematically represented by the elements of the positiveoperator valued measure (POVMs) M x that satisfy M x ≥ , R dx M x = I § [60, 61], isperformed yielding outcome statistics p ϕ ( X ) = Tr { ρ ϕ M X } . Establishing the optimalestimation strategy corresponds then not only to the correct interpretation of themeasurement results, but also to a non-trivial optimization over the class of all POVMsto find the measurement scheme maximizing the precision. In this case the quantumCramér-Rao bound can be derived [62–64], which is independent of the choice of thePOVMs and solely determined by the dependence of the output state on the estimated ‡ Throughout the paper, we depict derivatives w.r.t. the estimated parameter with an ‘overdot’, sothat e.g. ˙ p ϕ ( x ) ≡ ∂ ϕ p ϕ ( x ) , ˙ ρ ϕ ≡ ∂ ϕ ρ ϕ and ˙ K ( ϕ ) ≡ ∂ ϕ K ( ϕ ) . § We denote by I —the identity operator and by I —the identity superoperator. fficient tools for quantum metrology with uncorrelated noise ∆ ˜ ϕ ≥ p k F Q [ ρ ϕ ] with F Q [ ρ ϕ ] = Tr n ρ ϕ L S [ ρ ϕ ] o (2)being now the quantum Fisher information (QFI). The Hermitian operator L S [ ρ ϕ ] is theso called symmetric logarithmic derivative (SLD), which can be unambiguously definedfor any state ρ ϕ via the relation ˙ ρ ϕ = (cid:0) ρ ϕ L S [ ρ ϕ ]+ L S [ ρ ϕ ] ρ ϕ (cid:1) . Then, in the eigenbasis of ρ ϕ = P i λ i ( ϕ ) | e i ( ϕ ) ih e i ( ϕ ) | with {| e i ( ϕ ) i} i forming a complete basis ( ∀ i : 0 ≤ λ i ≤ ) L S [ ρ ϕ ] = X i,j h e i ( ϕ ) | ˙ ρ ϕ | e j ( ϕ ) i λ i ( ϕ ) + λ j ( ϕ ) | e i ( ϕ ) ih e j ( ϕ ) | , (3)where the sum is taken over the terms with non-vanishing denominator. QFI is anadditive quantity on product states and in particular F Q (cid:2) ρ ⊗ kϕ (cid:3) = k F Q [ ρ ϕ ] . Thus, the √ k term in the denominator of (2) may be equivalently interpreted as the number ofindependent repetitions of an experiment with a state ρ ϕ or a single shot experimentwith a multi-party state ρ ⊗ kϕ . Crucially, as proven in [64, 65], there always exist ameasurement strategy, e.g. projection measurement in the eigenbasis of the SLD, forwhich bounds (1) and (2) coincide. Hence, as in the classical case, the saturability of(2) is guaranteed, but again only in the k → ∞ limit. For pure states, ρ ϕ = | ψ ϕ ih ψ ϕ | , the QFI in (2) simplifies to F Q [ | ψ ϕ i ] =4 (cid:18)D ˙ ψ ϕ | ˙ ψ ϕ E − (cid:12)(cid:12)(cid:12)D ˙ ψ ϕ | ψ ϕ E(cid:12)(cid:12)(cid:12) (cid:19) k . Yet, as indicated by (3), for general mixed states thecomputation of QFI involves diagonalisation of ρ ϕ , which may be infeasible when dealingwith large systems. That is why it is often necessary to look for upper bounds onQFI that would be efficiently calculable even at the expense of saturability. For thispurpose, definitions of QFI were proposed that do not involve computing the SLD, butare specified at the level of state purifications: ρ ϕ = Tr E {| Ψ( ϕ ) ih Ψ( ϕ ) |} . In [38] the QFIof any ρ ϕ has been proven to be equal to the smallest QFI of its purifications | Ψ( ϕ ) i ¶ : F Q [ ρ ϕ ] = min Ψ( ϕ ) F Q [ | Ψ( ϕ ) i ] = 4 min Ψ( ϕ ) (cid:26)D ˙Ψ( ϕ ) | ˙Ψ( ϕ ) E − (cid:12)(cid:12)(cid:12)D ˙Ψ( ϕ ) | Ψ( ϕ ) E(cid:12)(cid:12)(cid:12) (cid:27) . (4)Independently, in [35] another purification-based QFI definition has been constructed: F Q [ ρ ϕ ] = 4 min Ψ( ϕ ) D ˙Ψ( ϕ ) | ˙Ψ( ϕ ) E . (5)Despite apparent difference, Eqs. (4) and (5) are equivalent and one can prove thatany purification minimizing one of them is likewise optimal for the other and satisfiesthe condition (cid:12)(cid:12)(cid:12) ˙Ψ( ϕ ) E = L S [ ρ ϕ ] ⊗ I E | Ψ( ϕ ) i causing the second term of (4) to vanish. k We shorten the notation of functions and superoperators of pure states, so that F [ | ψ i ] ≡ F [ | ψ ih ψ | ] and Λ[ | ψ i ] ≡ Λ[ | ψ ih ψ | ] . ¶ See also an alternative formulation based on the convex roof formula, which is valid for unitaryparameter estimation [66, 67]. fficient tools for quantum metrology with uncorrelated noise Ψ( ϕ ) (4) must provide a strictly tighter bound on QFIthan (5), the latter definition, owing to its elegant form, allows for more agility in theminimization procedure, so that it has been efficiently utilized in [35, 39] and is also thebase for this paper. On the other hand, a natural way to construct a bound on QFI and avoid the SLDcomputation is to relax the Hermiticity condition of the logarithmic derivative. If anon-Hermitian L [ ρ ϕ ] satisfying ∂ ϕ ρ ϕ = (cid:0) ρ ϕ L [ ρ ϕ ]+ L [ ρ ϕ ] † ρ ϕ (cid:1) can be found, as provenin [63, 65], an upper limit on QFI in (2) is obtained: F Q [ ρ ϕ ] ≤ Tr (cid:8) ρ ϕ L [ ρ ϕ ] L [ ρ ϕ ] † (cid:9) . Inparticular, if and only if ∂ ϕ ρ ϕ is contained within the support of ρ ϕ , one can construct the Right Logarithmic Derivative (RLD) by setting L [ ρ ϕ ] = L R [ ρ ϕ ] = ρ − ϕ ∂ ϕ ρ ϕ and formulatean upper bound on QFI of a simpler form: F Q [ ρ ϕ ] ≤ F RLDQ [ ρ ϕ ] = Tr (cid:8) ρ − ϕ ( ∂ ϕ ρ ϕ ) (cid:9) . (6)Although (6) is tight only when L R [ ρ ϕ ] = L S [ ρ ϕ ] , it still allows one to quantify precisionwell for channel estimation tasks [37], as described in the following section. Lastly, oneshould note that we are not considering here multi-parameter estimation schemes, forwhich the RLD may sometimes provide tighter bounds than the SLD [68, 69].
3. Estimation of a quantum channel
As in metrological setups the estimated parameter is encoded in the evolution of asystem, we identify ρ ϕ = Λ ϕ [ ρ in ] as the final state of a system that started from an input ρ in . The preparation of ρ in is controlled in order to achieve the most precise estimateof ϕ that parametrizes some general channel —a Completely Positive Trace Preserving (CPTP) map Λ ϕ [60, 61]. Although the form of Λ ϕ in general strongly depends on themodel considered, the Quantum Cramér-Rao bound always applies, so that precisionis upper bounded according to (2) with QFI F Q [Λ ϕ [ ρ in ]] . Furthermore, as the QFI is aconvex quantity [70], one should restrict oneself to pure input states when seeking theoptimal one. Hence, as shown in Figure 2 a , we define the channel QFI as the maximalQFI after performing the input optimisation, so that it has a concrete operational andapplication-like interpretation F [Λ ϕ ] = max ψ in F Q [ Λ ϕ [ | ψ in i ]] . (7)For instance, while estimating the duration of the evolution ( ϕ ≡ t ) in an ideal,decoherence-free setting, the CPTP map is unitary leading to a pure channel output.Thus, Λ ϕ [ | ψ in i ] = U t [ | ψ in i ] = e − i Ht | ψ in ih ψ in | e i Ht with H being the Hamiltonian of theevolution. Hence, the definition (7) corresponds to F [ U t ] = max ψ in F Q (cid:2) e − i Ht | ψ in i (cid:3) = 4 max ψ in (cid:8) h ψ in | H | ψ in i − h ψ in | H | ψ in i (cid:9) = 4 max ψ in ∆ H, (8) fficient tools for quantum metrology with uncorrelated noise Figure 2. Channel QFI based on the output state purification ( a ) Channel QFI as the QFI of the output state maximized over all pure input states.( b ) Channel QFI obtained from the output state purification generated by a local,fictitious, parameter-dependent environment rotation.( c ) Extended channel QFI independent of the maximization over the input states. Theenvironment rotation corresponds to a choice of Kraus representation of the channel. and the optimal states are the ones that maximize the Hamiltonian variance. Notealso that in this case the quantum Cramér-Rao bound takes the form of the time-energy uncertainty relation, ∆ H · ∆˜ t ≥ / [71, 72], with ∆˜ t being the uncertainty in theestimated duration. In order to employ the definition (5), we utilize the Stinespring theorem [61] and expressthe channel Λ ϕ as a unitary map U SE ϕ on the system combined with an environmentdisregarded after the evolution. In this way, the output state and its purificationrespectively read Λ ϕ [ | ψ in i ] = Tr E {| Ψ( ϕ ) ih Ψ( ϕ ) |} and | Ψ( ϕ ) i = U SE ϕ | ψ in i ⊗ | i , where | i is an arbitrary fixed state chosen to be the first vector in the basis {| i i} ri =1 ofthe environment Hilbert space H r E . By specifying the dimension r of H r E to be equalto the rank of Λ ϕ , we can generate all non-trivial purifications, ˜Ψ( ϕ ) , by applyinga fictitious, possibly ϕ -dependent unitary rotation, u E ϕ , to the environment, so that (cid:12)(cid:12)(cid:12) ˜Ψ( ϕ ) E = ˜ U SE ϕ | ψ in i⊗| i with ˜ U SE ϕ = u E ϕ U SE ϕ . Furthermore, writing the channel action inits Kraus representation form Λ ϕ [ | ψ in i ] = P ri =1 ˜ K i ( ϕ ) | ψ in ih ψ in | ˜ K i ( ϕ ) † , we can identifythe Kraus operators corresponding to ˜Ψ( ϕ ) as ˜ K i ( ϕ ) = h i | ˜ U SE ϕ | i = r X j =1 u E ϕ,ij K j ( ϕ ) , (9)where u E ϕ,ij = h i | u E ϕ | j i and K j ( ϕ ) = h j | U SE ϕ | i are the Kraus operators of theoriginal purification Ψ( ϕ ) . Hence, by picking an environment unitary rotation u E ϕ , fficient tools for quantum metrology with uncorrelated noise Λ ϕ . Moreover, as the QFI is a local quantity, we can restrict ourselves to infinitesimalrotations, u E ϕ = e − i h ( ϕ − ϕ ) , in the vicinity of the real value ϕ that are generated by someHermitian h . Taking without loss of generality ϕ = 0 , we obtain the channel version of(5) shown in Figure 2 b as F [Λ ϕ ] = 4 max ψ in min h h ˙˜Ψ( ϕ ) | ˙˜Ψ( ϕ ) i = 4 max ψ in min h h ψ in | r X i =1 ˙˜ K i ( ϕ ) † ˙˜ K i ( ϕ ) | ψ in i , (10)where | ˙˜Ψ( ϕ ) i = ( ˙ U SE ϕ − i hU SE ϕ ) | ψ in i ⊗ | i and similarly ˙˜ K i ( ϕ ) = ˙ K i ( ϕ ) − i P rj =1 h ij K j ( ϕ ) with h ij = h i | h | j i being the elements of the generator of Kraus representation rotations(9). Crucially, Figure 2 b and Equation (10) indicate that the optimal purification/Krausrepresentation corresponds to the choice of an artificial environment that rotates locallywith ϕ hindering as much as possible information about the estimated parameter.In order to make the reader familiar with the above formalism, we apply thedefinition (10) to the previously mentioned case of the evolution duration estimation U t [ | ψ in i ] . As the evolution is unitary and r = 1 , the environment/Kraus rotationscorrespond just to a phase variation u E t = e − i h t with h being a real scalar. Hence,(10) simplifies to (8) as expected: F [ U t ] = 4 max ψ in min h (cid:8) h ψ in | H | ψ in i − h h ψ in | H | ψ in i + h (cid:9) = 4 max ψ in ∆ H (11)with minimum occurring at h = h ψ in | H | ψ in i . Consistently, we would also arrive at (8),if we had used the other QFI purification-based definition (4) as shown in [38]. The channel QFI (7) is affected by any ϕ -variations in the form of Λ ϕ , quantifyingthe distinguishability between maps Λ ϕ and Λ ϕ + δϕ . However, any such disturbancemay be noticeable only for input states which lead to a measurable change of thechannel output that is at best in some “orthogonal direction”. As a consequence, thequantity min h { . . . } in (10) depends strongly on the pure input ψ in , as the minimumoccurs for Kraus operators (cid:8) K opt i ( ϕ ) (cid:9) ri =1 which fulfill the condition ˙ K opt i ( ϕ ) | ψ in i = L S [Λ ϕ [ | ψ in i ]] K opt i ( ϕ ) | ψ in i required for the purification of (4) and (5) to be optimal.Maximization of this quantity over | ψ in i is difficult in general, due to the impossibilityof exchanging the order of max and min in (10) [35].Yet, one may construct a natural upper bound on the channel QFI (7) by extending the input space, H S , by an equally-large auxiliary space, H A , which is unaffected bythe map and measured along with the channel output (see Figure 2 c ). This way, byemploying extended input states entangled between these two spaces, | ψ extin i ∈ H S ×H A ,one may acquire full available information about ϕ imprinted by the map Λ ϕ on theextended output state. The analogue of (10) defines then the extended channel QFI [35]: F [Λ ϕ ⊗ I ] = 4 max ρ Sin min h Tr S ( ρ Sin r X i =1 ˙˜ K i ( ϕ ) † ˙˜ K i ( ϕ ) ) = 4 min h (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) r X i =1 ˙˜ K i ( ϕ ) † ˙˜ K i ( ϕ ) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) , (12) fficient tools for quantum metrology with uncorrelated noise k . . . k represents the operator norm. The first expression is obtained by tracingover the auxiliary space H A , what leads to the maximisation over all mixed states ρ Sin = Tr A {| ψ extin ih ψ extin |} . However, this is exactly (10) with the pure input state replacedby a general mixed one, in which case the order of max and min can be swapped [35] + .Consistently, if the optimal input state of (12) is pure, (10) and (12) become equivalentmanifesting the uselessness of entanglement between H S and H A and the irrelevance ofthe auxiliary space.Importantly, the extended channel QFI (12) can always be efficiently evaluatednumerically by means of semi-definite programming [39] and we show that this is aspecial case of a more general task of bounding N -parallel channels QFI, as explainedin Appendix E. For phase estimation schemes with relevant noise models including: dephasing , depolarization , loss and spontaneous emission , we determine analyticallyboth (10) and (12) to verify if the use of extended, entangled input states may improveestimation at the single channel level. The corresponding unextended/extended channelQFIs are presented in Table 1, whereas the optimal purifications yielding the minimumof (12) can be found in Appendix A along with the details of the channels considered.The results justify that extension enhances the precision only for depolarization andspontaneous emission channels, for which the input states maximally entangled between H S and H A are optimal. In [37] the applicability of the RLD-based bound (6) has been addressed in the contextof channels. By defining the
Choi-Jamiołkowski (C-J) representation [61] of a particularmap Λ ϕ as Ω Λ ϕ = Λ ϕ ⊗ I [ | I i ] with | I i = P dim H S i =1 | i i ⊗ | i i , it has been proven that theextended channel QFI can be further upper-bounded by F [Λ ϕ ⊗ I ] ≤ F RLD [Λ ϕ ⊗ I ] = (cid:13)(cid:13)(cid:13) Tr A n ˙Ω Λ ϕ Ω − ϕ ˙Ω Λ ϕ o(cid:13)(cid:13)(cid:13) , (13)where k . . . k is again the operator norm and Ω − ϕ is the inverse of Ω Λ ϕ restricted only tothe support of the C-J matrix. Most importantly, the bound (13) is determined solely bythe form of Λ ϕ , i.e. its C-J representation, and does not contain any extra informationabout the space of input states accepted by the map. This contrasts the definitionsof purification-based unextended/extended channel QFIs (10)/(12) and facilitates theanalyticity of the results presented in Table 1. On the other hand, as indicated inSection 2.4, both applicability and tightness of the RLD-based bounds are limited. Thebound (13) is valid only when ˙Ω ϕ is fully supported by Ω Λ ϕ [37]. However, we givean intuitive reason for this restriction by proving that this condition is equivalent (seeAppendix B) to the notion of channel Λ ϕ being ϕ -non-extremal , as introduced in [39]and also revisited in the following section. This confirms that the exclusive dependenceof (13) on Ω Λ ϕ has indeed a strong geometrical meaning. Moreover, although the RLD-based bounds depicted in Table 1 for the relevant channels seem to be far above the + We should stress that (12) does not correspond to the situation of using mixed states as inputs for unextended channel, as mixed states never outperform pure state inputs due to convexity of the QFI. fficient tools for quantum metrology with uncorrelated noise Noisemodel F [Λ ϕ ] F [Λ ϕ ⊗I ] F CEas in [39] F QSas F RLD [Λ ϕ ⊗I ] F CSas in [39]
Dephasing η η η − η η − η η − η η − η Depolarization η η η η (1 − η )(1+2 η ) 2 η (1 − η )(1+2 η ) 2 η (1+ η )(1 − η )(1+3 η ) 4 η (1 − η )(1+3 η ) Loss η η η − η η − η n.a. n.a. Spontaneousemission η η ( √ η ) η − η n.a. n.a. n.a. Table 1. Channel phase estimation sensitivity quantified via QFIs and theirasymptotic bounds.
The noise models of metrological relevance discussed in thepaper are listed in the first column. Decoherence strength increases with a decreaseof the η parameter (0 ≤ η < (see Appendix A for details). From left to right—single channel QFI, extended channel QFI, upper bounds on asymptotic channel QFI(16) in ascending order: CE bound (see Section 4.1.3), QS bound (see Section 4.1.2),RLD-based bound (see Section 3.4), CS bound (see Section 4.1.1). [n.a.—not available] corresponding channel QFIs—(7) and (12), they are of great significance. The bound(13) is additive for any maps Λ (1) ϕ , Λ (2) ϕ to which it applies [37], thus F RLD (cid:2)(cid:0) Λ (1) ϕ ⊗I (cid:1) ⊗ (cid:0) Λ (2) ϕ ⊗ I (cid:1)(cid:3) = F RLD (cid:2) Λ (1) ϕ ⊗I (cid:3) + F RLD (cid:2) Λ (2) ϕ ⊗I (cid:3) ∴ F h (Λ ϕ ⊗I ) ⊗ N i ≤ F RLD h (Λ ϕ ⊗I ) ⊗ N i = N F RLD [Λ ϕ ⊗I ] . (14)Hence, it constrains not only the QFI of a single extended channel, but also restricts theQFI of N extended channels used in parallel to scale at most linearly with N . Crucially,as the extension can only improve the precision, (14) is also a valid upper-bound on theQFI of N uses of an unextended channel, which asymptotic precision scaling is thenlimited to a constant factor improvement over the SQL (see Section 4.1).
4. Estimation of N independent quantum channels In order to describe general metrological schemes depicted in Figure 1 a , we model theevolution of all particles within the probe as N identical, independent channels actingon a possibly entangled, pure input state of the whole probe (cid:12)(cid:12) ψ N in (cid:11) . The final outputstate of the probe then reads ρ Nϕ = Λ ⊗ Nϕ (cid:2)(cid:12)(cid:12) ψ N in (cid:11)(cid:3) yielding the N - channel QFI : F (cid:2) Λ ⊗ Nϕ (cid:3) = max ψ N in F Q (cid:2) Λ ⊗ Nϕ (cid:2)(cid:12)(cid:12) ψ N in (cid:11)(cid:3)(cid:3) , (15)which linear or quadratic dependence on N dictates respectively the SQL or HL scalingof precision. For example, when considering classical schemes that employ unentangledprobes, (cid:12)(cid:12) ψ N in (cid:11) = ⊗ Nn =1 | ψ in i , we are effectively dealing with N independent subsystems,so that F (cid:2) Λ ⊗ Nϕ (cid:3) = N F [Λ ϕ ] and the uncertainty of the estimate ˜ ϕ decreases as / √ N . fficient tools for quantum metrology with uncorrelated noise N limit To investigate channels that incorporate the uncorrelated noise restricting theasymptotic precision scaling to SQL, we define the asymptotic channel QFI as F as [Λ ϕ ] = lim N →∞ F (cid:2) Λ ⊗ Nϕ (cid:3) N . (16)Thus, for such SQL-bound channels, (16) is finite and F as [Λ ϕ ] ≥ F [Λ ϕ ] with equalityindicating the optimality of classical estimation schemes. Hence, (16) quantifies the maximal quantum precision enhancement that reads χ [Λ ϕ ] = lim N →∞ ∆ ˜ ϕ cl ∆ ˜ ϕ Q = s F as [Λ ϕ ] F [Λ ϕ ] ≥ . (17)However, as the computation of (16) is generally infeasible owing to the complexity ofQFI rising exponentially with N , one normally needs to construct an upper limit onthe N -channel QFI (15), F bound [Λ ⊗ Nϕ ] , from which the asymptotic form, F boundas , maybe deduced using (16) that upper-bounds both the N -channel QFI and the maximalquantum precision enhancement: F (cid:2) Λ ⊗ Nϕ (cid:3) ≤ N F boundas , χ [Λ ϕ ] ≤ s F boundas [Λ ϕ ] F [Λ ϕ ] . (18)Methods of constructing F boundas were proposed in [39] basing on the concepts ofchannel classical simulation (CS) and channel extension (CE). As mentioned already,the CS method applies only to ϕ -non-extremal channels, for which also the RLD-based bound (14) provides a valid example of F boundas . Yet, the notion of CS may begeneralized to the idea of channel quantum simulation (QS) introduced in [36], in orderto obtain an asymptotic bound applicable to a wider class of quantum maps. All thesefour approaches ( F boundas = F CSas , F CEas , F RLD [Λ ϕ ⊗I ] , F QSas respectively) are presented inTable 1 on the right hand side of the double-line for the relevant channels. As thelossy and spontaneous emission interferometry cases are examples of ϕ -extremal maps,they do not allow for finite F CSas and F RLD [Λ ϕ ⊗I ] to be constructed. In the case ofdepolarization channel, which is full-rank [61] and hence ϕ -non-extremal, despite thelack of a simple geometric interpretation of its value, F RLD [Λ ϕ ⊗I ] proves to be tighterthan F CSas . The more general QS approach not only applies also to the ( ϕ -extremal)lossy interferometry case, but also provides as accurate bounds as the CE method.Nevertheless, as the CE approach is proven to provide at least as tight bounds forthe broadest class of quantum channels containing ones to which the other methodsapply, we use it to quantify the maximal quantum precision enhancements (17) for thechannels considered (see Table 2). Below, we describe alternately the CS, QS and CEmethods—ordered according to their power and generality. fficient tools for quantum metrology with uncorrelated noise As introduced in [36] and depicted in Figure 1 b , a channel admits a CS of itself, if forany ̺ it is expressible in the form Λ ϕ [ ̺ ] = Φ[ ̺ ⊗ p ϕ ] = X i p ϕ,i Π i [ ̺ ] , (19)where p ϕ = P i p ϕ,i | e i ih e i | is a classical, diagonal density matrix in some basis and Φ is a ϕ -independent CPTP map acting on a larger input space. By defining Π i [ ̺ ] =Φ[ ̺ ⊗| e i ih e i | ] one obtains the second equality in (19), so that it becomes evident thatthe estimated ϕ parametrizes only the mixing probabilities of some ϕ -independentquantum maps. Then, the N -channel QFI (15) can be simply upper-bounded via F (cid:2) Λ ⊗ Nϕ (cid:3) ≤ N F cl [ p ϕ,i ] , where F cl [ p ϕ,i ] is the discrete version of classical FI in (1) andplays the role of F boundas in (18) [36, 39]. Moreover, as QFI is a local quantity, in order toconstruct a CS-based F boundas valid for small deviations δϕ around a given ϕ , the classicalsimulation must be feasible only locally, i.e. Λ ϕ [ ̺ ] = P i p ϕ,i Π i [ ̺ ] + O ( δϕ ) . Therefore,as proven in [36], if the C-J representation Ω Λ ϕ of a channel Λ ϕ allows for parameters ǫ ± > such that the matrices Ω Π ± = Ω Λ ϕ ± ǫ ± ˙Ω Λ ϕ are positive semi-definite at a given ϕ ,the channel is ϕ -non-extremal there and the necessary p ϕ,i can always be found. This isbecause one can construct Ω ˜Λ ϕ = p ϕ, + Ω Π + + p ϕ, − Ω Π − that up to O ( δϕ ) coincides with Ω Λ ϕ by choosing p ϕ, ± such that Ω ˜Λ ϕ = Ω Λ ϕ and ˙Ω ˜Λ ϕ = ˙Ω Λ ϕ . Hence, Λ ϕ [ ̺ ] = ˜Λ ϕ [ ̺ ] + O ( δϕ ) with ˜Λ ϕ [ ̺ ] = p ϕ, + Π + + p ϕ, − Π − , so that F cl [ p ϕ, ± ] = 1 / ( ǫ + ǫ − ) is a legal example of therequired finite bound valid at ϕ . Furthermore, in [39], it has been proven that forchannels of the form Λ ϕ [ ̺ ] = Λ[ U ϕ [ ̺ ]] this two-point construction is always optimal atany ϕ when maximal possible ǫ ± are chosen ∗ . Geometrically, imagining the convex setof all CPTP maps in their C-J representation that share input and output spaces of Λ ϕ ,the channels Ω Π ± lie at the intersection points of the tangent generated by ˙Ω Λ ϕ at Ω Λ ϕ and the boundary of the set. The F boundas of (18), which we refer to as the asymptoticCS bound — F CSas = 1 / ( ǫ max + ǫ max − ) , is dictated then by the “distances” ǫ max ± of the channelto the boundary measured along this tangent. Although the CS approach providesweaker bounds than the CE method [39], it gives an intuitive geometric explanation ofthe inevitable asymptotic SQL-like scaling of all ϕ -non-extremal maps. These naturallyinclude the full-rank channels [61], which lie inside the set of CPTP maps away from itsboundary. In [36], a natural generalization of the channel CS has been proposed which isschematically presented in Figure 1 b . This, so called, QS of a channel correspondsto expressing its action in a form similar to (19) that reads Λ ϕ [ ̺ ] = Φ[ ̺ ⊗ σ ϕ ] = Tr E Φ E σ n U (cid:16) ̺ ⊗ | ψ ϕ i E Φ E σ h ψ ϕ | (cid:17) U † o , (20)where now σ ϕ is a quantum, non-diagonal, ϕ -dependent density matrix, and the purifiedform on the right hand side involves both channel Φ environment space E Φ as well as ∗ Yet, it may prove optimal for channels of other type, as shown for transversal dephasing in [40]. fficient tools for quantum metrology with uncorrelated noise σ ϕ purification space E σ , such that σ ϕ = Tr E σ {| ψ ϕ ih ψ ϕ |} . Note that the notion of quantum simulability is equivalent to the channel programmability concept introducedin [34]. Following the same argumentation as in [39] for CS, the N -channel QFI (15) ofa quantum simulable channel —one admitting a QS of the form (20) with finite F Q [ σ ϕ ] —can be linearly bounded as F (cid:2) Λ ⊗ Nϕ (cid:3) ≤ N F Q [ σ ϕ ] , and therefore the asymptotic boundreads F boundas = F Q [ σ ϕ ] . Similarly to CS, a channel may admit many decompositions(20) and the optimal one must yield the lowest F Q [ σ ϕ ] . Therefore, without loss ofgenerality, in the search for the optimal QS, we may take U in (20) to act on the fullpurified system, i.e. also in E Φ and E σ spaces. This enlarges the set of all possibleQSs beyond the original ones U = U SE Φ ⊗ I E σ and yields F boundas = F Q [ | ψ ϕ i ] , which via(4) cannot be smaller than F Q [ σ ϕ ] . In fact, (4) assures that for any QS employing σ ϕ , there exists an “enlarged” decomposition (20) leading to the same F boundas with | ψ ϕ i being the minimal purification in (4). In conclusion, we may seek the optimal QSby analysing all possible decompositions of the form (20) that, owing to the localityof the QFI, must be feasible only for small deviations δϕ from a given ϕ , so that Λ ϕ [ ̺ ] = Tr E Φ E σ (cid:8) U ( ̺ ⊗| ψ ϕ ih ψ ϕ | ) U † (cid:9) + O ( δϕ ) . In Appendix D, we prove that, in orderfor the QS (20) to be possible locally at ϕ and lead to a finite asymptotic bound, Λ ϕ ofrank r must admit Kraus operators { K i ( ϕ ) } ri =1 that satisfy the two conditions:i r X i =1 ˙ K i ( ϕ ) † K i ( ϕ ) = 0 and r X i =1 ˙ K i ( ϕ ) † ˙ K i ( ϕ ) = 14 F Q [ | ψ ϕ i ] I . (21)Hence, by optimizing over all locally equivalent Kraus representations of Λ ϕ —the onesrelated to one another by rotations (9) generated by any Hermitian h —that satisfyconstraints (21), we may determine the asymptotic bound given by the optimal localQS, which we refer to as the asymptotic QS bound — F QSas , as follows F QSas = min h λ s.t. α ˜ K = λ I , β ˜ K = 0 , (22)where α ˜ K = P ri =1 ˙˜ K i ( ϕ ) † ˙˜ K i ( ϕ ) , β ˜ K = i P ri =1 ˙˜ K i ( ϕ ) † ˜ K i ( ϕ ) and λ has the interpretationof F boundas = F Q [ | ψ ϕ i ] for a local QS of the form (20). Before revisiting the CE methodexplicitly below, we should note that (22) resembles exactly the asymptotic CE boundof [39] with an extra constraint forcing the operator α ˜ K to be proportional to identity.This proves that indeed the QS method can never outperform the CE approach.Investigating the relevant quantum maps considered in Table 1, the QS methodmust naturally apply to dephasing and depolarization channels. These are ϕ -non-extremal maps, hence their locally constructible CSs (19) serve as examples of the moregeneral QSs (20). Consistently, the Kraus representations utilized in [39] to obtain theasymptotic CE bounds for these two channels fulfil the α ˜ K ∝ I constraint. Thus, QSis not only feasible in their case but also its asymptotic bound coincides with the oneof the superior CE method. Significantly, also in the case of the lossy interferometrythe optimal Kraus operators used in [39] to minimize the asymptotic CE bound satisfythe extra QS’s constraint. This fact indicates that also for ϕ -extremal channels QSmay prove to be as good as CE. However, in the case of spontaneous emission, the QS fficient tools for quantum metrology with uncorrelated noise Noisemodel
Dephasing Depolarization Loss Spontaneousemission χ [Λ ϕ ] = q − η ≤ q − η )(1+2 η ) = q − η ≤ q − η χ [Λ ϕ ⊗ I ] = q − η = q η (1 − η )(1+2 η ) = q − η = q √ η −√ η Table 2. Quantum phase estimation precision enhancement from the CEmethod.
For all noise models specified in Appendix A, the asymptotic CE boundson the maximal quantum precision enhancement factors, χ [ • ] = p F as [ • ] / F [ • ] , arepresented. For a general quantum map Λ ϕ , the CE bound only upper-limits χ [Λ ϕ ] as F as [Λ ϕ ] ≤ F CEas . Yet, for dephasing and lossy interferometry, as indicated by “ = ”,the corresponding values of χ [Λ ϕ ] have been shown to be attainable [73, 74]. For anextended channel, χ [Λ ϕ ⊗ I ] is determined by the CE bound as F as [Λ ϕ ⊗ I ] = F CEas . method seizes to work, as the β ˜ K = 0 condition fixes α ˜ K to be disproportional to identity[39]. The CE method of [39] not only applies to the widest class of quantum maps containingall ϕ -non-extremal ones, but also provides more stringent bounds than the CS, RLD andQS equivalents, as respectively proven in [39], Appendix C and above. In this method,see Figure 1 b , each single channel is extended by an auxiliary ancilla as introducedin Section 3.3. In [35], it has been proven that one can then effectively bound the N -channel QFI (15) via the N -extended-channel QFI , so that F (cid:2) Λ ⊗ Nϕ (cid:3) ≤ F h (Λ ϕ ⊗ I ) ⊗ N i ≤ h (cid:8) N k α ˜ K k + N ( N − k β ˜ K k (cid:9) , (23)where again α ˜ K = P ri =1 ˙˜ K i ( ϕ ) † ˙˜ K i ( ϕ ) , β ˜ K = i P ri =1 ˙˜ K i ( ϕ ) † ˜ K i ( ϕ ) and h is the generator oflocal Kraus representation rotations (9). Crucially, if there exists a Kraus representationfor which the second term in (23) vanishes, F (cid:2) Λ ⊗ Nϕ (cid:3) must asymptotically scale linearlyin N . This requirement corresponds to the constraint β ˜ K = 0 already accounted in theQS method, which for any linearly independent Kraus operators { K i } ri =1 is equivalentto the existence of h such that [35] r X i,j =1 h ij K † i K j = i r X i =1 ˙ K i ( ϕ ) † K i ( ϕ ) . (24)What is more, for any channel that admits an h fulfilling (24), one can show basing onthe results of [35] that the second inequality in (23) is saturated in the N → ∞ limit, sothat the asymptotic extended channel QFI then reads F as [Λ ϕ ⊗ I ] = lim N →∞ F h (Λ ϕ ⊗ I ) ⊗ N i N = 4 min h β ˜ K =0 (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) r X i =1 ˙˜ K i ( ϕ ) † ˙˜ K i ( ϕ ) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) . (25) fficient tools for quantum metrology with uncorrelated noise F boundas of (18), which werefer to as the asymptotic CE bound — F CEas . We explicitly wrote the full form of (25)to emphasize its similarity to the extended single channel QFI (12). The essentialdifference in (25) is the constraint (24) yielding F as [Λ ϕ ⊗ I ] ≥ F [Λ ϕ ⊗ I ] and dictatingthe maximal quantum precision enhancement for an extended channel: χ [Λ ϕ ⊗ I ] = lim N →∞ ∆ ˜ ϕ extcl ∆ ˜ ϕ extQ = s F as [Λ ϕ ⊗ I ] F [Λ ϕ ⊗ I ] ≥ . (26)Similarly to (12), (25) is computable by means of semi-definite programming [39], sothat one can efficiently determine both (18) and (26). The CE-based bounds on χ [Λ ϕ ] and the exact values of χ [Λ ϕ ⊗ I ] for the relevant noise models are presented in Table 2.Although generally the CE method only upper-limits the maximal quantum precisionenhancement (17), it has been proven to quantify χ [Λ ϕ ] exactly in the case of dephasing[73] and lossy interferometer channels [74]. This has been achieved by showing that inputstates utilizing spin- and light- squeezing respectively yield a quantum enhancement thatasymptotically attains the corresponding CE-based bounds presented in Table 2. Onthe other hand, as indicated in Table 1, these channels are also examples of ones forwhich the extension does not improve the precision at the single channel level, so that χ [Λ ϕ ] = χ [Λ ϕ ⊗ I ] in Table 2. The question—when the lack of precision improvementdue to extension at the single channel level translates to the asymptotic regime, i.e F [Λ ϕ ] = F [Λ ϕ ⊗ I ] ⇐⇒ ? F as [Λ ϕ ] = F as [Λ ϕ ⊗ I ] , we leave open for future research. N channel extension (CE) method In Section 4.1, we have presented the CE method as the most effective one out of alldiscussed that provides the tightest upper limits on the maximal possible asymptoticquantum precision enhancement. On the other hand, in the case of experiments suchas optical interferometry with moderate numbers of photons [8, 9], the asymptotic CEbounds, despite still being valid, are far too weak to be useful. For very low values of N , the precision can be quantified numerically, for instance by brute-force type methodscomputing explicitly the QFI. However, in the intermediate N regime—being beyondthe reach of computational power, yet with N too low for the asymptotic methods tobe effective—more accurate bounds should play an important role.We propose the finite-N CE method which, despite being based on the properties ofa single channel, still provides bounds on precision that are relevant in the intermediate N regime. We utilize the upper-limit (23) on the N -extended-channel QFI and constructthe finite-N CE bound , F CE N , that reads F h (Λ ϕ ⊗ I ) ⊗ N i N ≤ F CE N = 4 min h (cid:8) k α ˜ K k + ( N − k β ˜ K k (cid:9) . (27)Following the suggestion of [39], in contrast to the asymptotic CE bound F CEas definedin (25), we do not impose the SQL-bounding condition β ˜ K = 0 (24), but we seek at each N for the minimal Kraus representation that is generated by some optimal h = h opt ( N ) fficient tools for quantum metrology with uncorrelated noise N -dependent. Still, as shown in Appendix E, F CE N can always be efficiently evaluated numerically by recasting the minimization over h in(27) into a semi-definite programming task. Moreover, as the finite- N CE bound variessmoothly between N = 1 and N = ∞ , at which it is tight, it provides more accuratebounds than its asymptotic version. For N = 1 , F CE N coincides with the extended channelQFI (12)— F CE N =1 = F [Λ ϕ ⊗ I ] , whereas in the N → ∞ limit it attains the asymptotic CEbound (25)— F CE N →∞ = F CEas .What is more, when considering channels for which the asymptotic CE method fails,as it is not possible to set β ˜ K = 0 in (23) for any Kraus representation, (27) still applies;it is the finite- N CE method that provides the correct CE-based bound in the N → ∞ limit that in principle may then surpass the SQL-like scaling. On the other hand, whendealing with estimation schemes in which one can moderate the effective amount ofloss (i.e. the form of Λ ϕ ) depending on the number of particles, the asymptotic bound F CEas may not actually be the tightest within the CE method. The β ˜ K of (23) and (27)becomes then a function of N and it may not be asymptotically optimal to set it equalto zero by imposing condition (24). Yet the finite- N CE method, being not constrainedwith β ˜ K = 0 , still yields the correct CE-based bound on precision as N → ∞ . Thisfact has been utilized in [40], where, owing to the N -dependence of β ˜ K , the finite- N CE method provided an asymptotic bound indeed tighter than the naively calculated F CEas . What is more, the F CE N →∞ has been numerically shown there to be saturable, whatproves the power of the more agile finite- N CE method. For phase estimation withvarious decoherence models including dephasing, depolarization, loss and spontaneousemission described in detail in Appendix A, we observe that the finite- N CE bound issimply related to its asymptotic form as F CE N = N F CEas N + F CEas , (28)where one should substitute for F CEas the corresponding asymptotic CE bounds presentedin Table 1 ♯ . For dephasing and loss decoherence models, we show explicitly in Figure 3both the asymptotic and the finite- N bounds accompanied by the plots of actualprecision achievable with explicit estimation strategies optimal either in the small orlarge N regime.In the first case, depicted in Figure 3 a , we consider a Ramsey spectroscopy setupof [13, 14] in which the probe consists of atoms prepared in a spin-squeezed state [73].The parameter is then encoded in the phase of a unitary rotation generated by the totalangular momentum of the atoms that simultaneously experience uncorrelated dephasing.After measuring probe’s total angular momentum perpendicular to the one generatingthe estimated phase change, the parameter is reconstructed with uncertainty plottedin Figure 3 a . For comparison, the maximal precision theoretically achievable withGreenberger-Horne-Zeilinger (GHZ) [75] input states is also shown. The QFI for a ♯ In the case of spontanous emission noise the formula is valid only for N ≥ , what we suspect to bea consequence of the spontaneous emission channel being an extremal map [61]. fficient tools for quantum metrology with uncorrelated noise Figure 3. Phase estimation CE method-based bounds on precision ( a ) Dephasing : Finite- N ( dashed ) and the asymptotic CE bounds ( dotted ) onestimation uncertainty as compared with the precision achieved by utilizing spin-squeezed states in a Ramsey spectroscopy setup ( solid black ) and GHZ states withoptimal measurement ( solid grey ) for a probe consisting of N atoms experiencinguncorrelated dephasing with η = 0 . .( b ) Loss : Lossy interferometry with particle survival probability η = 0 . , e.g. Mach-Zehnder interferometer experiencing photonic loss in both of its arms, with effectivepower transmission η . The smallest uncertainty in a phase estimation scheme isquantified by calculating the QFI for numerically optimized N -particle input states( solid black ) which only at low N can be approximated by N00N states ( solid grey ).Again, finite- N ( dashed ) and asymptotic CE bounds ( dotted ) on precision are shownfor comparison. GHZ-based strategy is F GHZ N = η N N which for low N attains the finite- N CE bound.This fact proves that in experiments with only few particles involved, such as [1], it isoptimal to use the GHZ states as inputs despite the uncorrelated dephasing present.In the second lossy interferometry case, shown in Figure 3 b , each particle is subjectto an independent loss process with survival probability η , as in e.g. Mach-Zehnderinterferometer with effective power transmittance η in both arms [29, 30] that representspreparation, transmission and detection loss [31]. Here the solid black line representsthe maximal QFI achieved with numerically optimised N -particle or equivalently N -photon states †† . As expected, it coincides for low N with the QFI attained by the socalled N00N states [76], F N00N N = η N N , which are the optical equivalents of the GHZstates previously considered. The plot indicates that the finite- N CE bound may beconsidered in this case to be tight only for moderate and very large N . Although in thecase of lossy optical interferometry, the maximal quantum enhancement of Table 2 canalso be achieved via an estimation strategy that employs squeezed-light as the inputwith mean number of photons constrained to N = ¯ N [74, 77], we cannot compare itsprecision with the one bounded through F CE N = ¯ N . As N ·F CE N is a convex quantity in N ,one cannot use it naively to constrain precision after replacing N by the mean numberof photons ¯ N . This contrasts the case of the (constant) asymptotic CE bound, whichyields N ·F CEas being linear in N , so that it also applies to estimation strategies employing †† As it is optimal to consider indistinguishable, bosonic particles [30]. fficient tools for quantum metrology with uncorrelated noise
5. Frequency estimation in atomic models
We apply the methods discussed above to the case of frequency estimation problems inatomic spectroscopy. The general Ramsey spectroscopy setup considered in [13, 14, 22–26, 40] corresponds to N identical two-level atoms—spin- / systems—with their statesseparated, where typically the detuning ω of an external oscillator frequency from theatoms transition frequency is to be estimated. We assume that the full experiment takesan overall time T , during which the estimation procedure is repeated k = T /t times,where t is the evolution duration of each experimental shot. The quantum Cramér-Raobound (2) on precision of the estimate ˜ ω can be then conveniently rewritten as ∆˜ ω ≥ p T f t [ ρ Nω ( t )] , (29)where f t [ ̺ ( t )] = F Q [ ̺ ( t )] /t is now the effective QFI per shot duration and ρ Nω ( t ) denotesthe final state of the whole probe containing N particles being measured in each shot.The total time T plays then the role of k in (2) and, after fixing t , the bound (29) canalways be saturated as T → ∞ . The evolution of the probe can be modelled by themaster equation of the Lindblad [54] form ∂ρ Nω ( t ) ∂t = N X n =1 i ω h σ ( n )3 , ρ Nω ( t ) i + L ( n ) (cid:2) ρ Nω ( t ) (cid:3) , (30)where σ ( n )3 is the Pauli operator generating a unitary rotation of the n ’th atom aroundthe z axis in its Bloch ball representation. The uncorrelated noise is representedby the Liouvillian part L ( n ) acting independently on each particle, here the n ’th, sothat effectively ρ Nω ( t ) = Λ ⊗ Nω ; t (cid:2)(cid:12)(cid:12) ψ N in (cid:11)(cid:3) with channel Λ ω ; t representing the overall singleparticle evolution over time t . To model the decoherence we consider the dephasing,depolarization, loss and spontaneous emission maps, which corresponding Liouvillianscan be found in Appendix A. As the estimated parameter corresponds now to ω = ϕ/t ,where ϕ is the phase of the unitary rotation, the QFI via a parameter change justrescales, so that f t [ ̺ ω ] = F Q [ ̺ ω ] /t = F Q [ ̺ ϕ ] t . Moreover, due to the commutativity of theunitary and the considered decoherence maps, we can without loss of generality utilizethe results presented for them in the previous sections. Defining the channel QFI forfrequency estimation tasks similarly to (7) as f [Λ ω ] = max ≤ t ≤ T max ψ in f t [Λ ω ; t [ | ψ in i ]] (31)we can compute all the appropriate expressions for QFIs and the asymptotic boundsof Table 1 as well as the finite- N bounds of (28) by substituting for the effective timedependence of the decoherence strength η ( t ) , which is determined by the master equation(30) (see Appendix A). Then, any quantity F listed in Table 1 transforms to its t -optimized equivalent as f = max ≤ t ≤ T F t . In Table 3 we present the channel QFIs relevant for fficient tools for quantum metrology with uncorrelated noise Noisemodel f [Λ ω ] f [Λ ω ⊗I ] f CE N ( N ≥ ) f CEas χ [Λ ω ] χ [Λ ω ⊗ I ] Dephasing e γ e γ N γ w [ N ]1+ ( e w N ] − ) N γ = √ e = √ e Depolarization e γ ≈ .
27 34 e γ N γ α w β [ N ]2+ (cid:16) e α wβ [ N ] − (cid:17)(cid:16) e α wβ [ N ] +2 (cid:17) N γ ≤ q e ≈ . q e Loss e γ e γ Nγ w [ N ]1+ ( e w N ] − ) N γ = √ e = √ e Spontaneousemission e γ wγ ( e ˜ w/ ) Nγ w [ N ]4+ ( e w N ] − ) N γ ≤ √ e = 1+ e ˜ w/ √ ˜ w Table 3. QFIs, CE bounds and quantum enhancements in frequencyestimation.
In frequency estimation tasks, the precision is maximized by adjustingthe single experimental shot duration t . The t -optimised (extended) channel QFIsas well as their finite- N and asymptotic CE bounds are presented, where w x [ N ] =1 + W (cid:2) x − N e N (cid:3) , ˜ w = 1 + 2 W h √ e i and W [ x ] is the Lambert W function. As in thecase of depolarizing channel not all solutions possess an analytical form, only theirnumerical approximations are shown with α ≈ . and β ≈ . . Right of the double-line the maximal quantum precision enhancements are listed for the maps considered.In the case of dephasing noise the ultimate √ e factor has already been reported in[24, 38]. For unextended depolarization and spontaneous emission maps, the derivedenhancement factors may possibly be not achievable. frequency estimation, their asymptotic and finite- N CE bounds, as well as the maximalquantum precision enhancements for each model considered. In the case of dephasing,we recover the results of [24, 38, 40], whereas for depolarizing, loss and spontaneousemission maps we obtain the QFIs and their bounds, which to our knowledge havenot been reported in the literature before. However, similarly to the case of quantumphase estimation summarized in Table 2, the obtained quantum enhancement factorsfor depolarization and spontaneous emission channels serve only as bounds, as they arenot guaranteed to be saturable.
6. Estimation of decoherence strength
Lastly, we would like to emphasize that the CS, QS and CE methods described inSection 4 also apply to estimation tasks in which the estimated parameter is not encodedin the unitary, noiseless part of the system evolution. Examples of such schemes are theexperimentally motivated ones, in which one tries to quantify the effective strength ofnoise or loss present in the apparatus. That is why we consider again the channelsdescribed in Appendix A, but this time with the parameter to be estimated beingthe decoherence strength η . This kind of problems has been widely considered notonly in the estimation theory [44, 78–80], but also when examining issues of channeldiscrimination [45, 46] with particular application in quantum reading [50–52]. Ascompared to unitary rotations, the nature of the estimated parameter is dramatically fficient tools for quantum metrology with uncorrelated noise Channel considered F [Λ η ] F [Λ η ⊗I ] F CSas = F QSas = F CEas
Dephasing − η − η − η [82] Depolarization − η [70] − η )(1+3 η ) [70] − η )(1+3 η ) [35, 82] Loss η (1 − η ) 1 η (1 − η ) 1 η (1 − η ) Spontaneous emission η (1 − η ) [83] η (1 − η ) [83] η (1 − η ) Table 4. Decoherence strength estimation quantified via channel QFIs andtheir asymptotic bounds.
Definitions of channels listed in the first column can befound in Appendix A. In contrast to phase estimation examples given in Table 1, thevariable to be estimated here is the decoherence parameter η , (0 ≤ η < . Due tothe different nature of the estimated parameter, the geometrical CS method providesbounds that not just most tightly limit the asymptotic extended channel QFIs, butactually coincide with its value. The results prove that only in the case of depolarizationchannel the precision can be enhanced with the use of quantum estimation strategies,as for all other cases F [Λ η ] = F [Λ η ⊗ I ] = F as . different. In unitary parameter case, the use of entangled input state of N particlesresults in an effective N -times higher “angular speed” of rotation leading to the HL in theabsence of noise. In decoherence strength estimation tasks, a change in the parametervalue can be geometrically interpreted in the space of all valid quantum channels asa “movement” in the direction away from the boundary of the space of the relevantCPTP maps, for which “speed” cannot be naively amplified N -times when employing N parallel channels. Hence, as in the case of lossy unitary rotation estimation, the optimalentangled inputs must lead not to scaling but constant factor quantum enhancements,which again can be quantified by the methods of Section 4. This also explains that for allthe four noise models considered, the purely geometrical notion of classical simulability isenough to bound most tightly the maximal asymptotic precision of estimation. However,as for them also F CSas = F QSas = F CEas = F [Λ η ⊗ I ] , the CS-based asymptotic quantumenhancement corresponds to the classical estimation strategy that employs independentbut extended channels. The fact that factorizable inputs—uncorrelated in between theextended N channels but possibly requiring entanglement between each single particleand its ancilla—are optimal for noise estimation with extended channels, has alreadybeen noticed for the low-noise [81] and generalized Pauli [82] channels, of which thelatter contain the dephasing and depolarization maps studied here.In the case of dephasing channel, we further realize that the extension at thesingle channel level is also unnecessary, as F [Λ η ] = F [Λ η ⊗ I ] = 1 / (1 − η ) , andthe geometrically dictated bound of CS is attainable classically just by employingunentangled qubits in any pure state lying on the Bloch sphere equatorial plane. fficient tools for quantum metrology with uncorrelated noise F CSas = F [Λ η ] = F [Λ η ⊗ I ] = 1 / ( η (1 − η )) , which at the single channel level is achieved by a photon inany mixed state with the ancilla being redundant. On the one hand, this emphasizesthat entanglement between the photons entering the interferometer is unnecessary andagrees with the results of [78, 79] confirming that the total photon number fluctuationsare really the ones that limit the precision. These can be reduced by employing Gaussianstates [78] or in principle fully eliminated by the use of Fock states [79] that attain theCS-bound.The case of depolarization map is different, as it is known that for qubits [70, 84]the precision of estimation may be improved by extending the channel, i.e. F [Λ η ] < F [Λ η ⊗ I ] . This leaves the space for possible enhancement thanks to the use ofentangled probes between unextended channels and indeed this fact has been observedalready when considering two depolarization channels used in parallel [70]. The resultsare summarized in Table 4.
7. Further discussion
We would also like to point that the SQL-like bounds, universally valid in practicalmetrological scenarios, allow one to avoid some of the controversies characteristic foridealized decoherence-free scenarios. When decoherence is not present and the probestates with indefinite number of particles are considered, such as e.g. squeezed states oflight, the exact form of HL needs to be reconsidered [85–87] since the direct replacementof N with mean number of particles ¯ N may make the HL invalid. Moreover, thefinal claims on the achievable precision scaling may strongly depend on the form of apriori parameter knowledge assumed, and lead to some apparent contradictions [88, 89].These difficulties do not arise in the realistic metrological schemes, as the asymptoticSQL-like bounds are valid also when N is replaced by ¯ N for indefinite particle numberstate [39, 77]. The bounds derived in the local approach (small parameter fluctuation)based on the calculation of the QFI are saturable in a single-shot scenario unlikethe decoherence-free case when only after some number of independently repeatedexperiments one may expect to approach the theoretical limits [90, 91]. This is due to thefact that by employing input states of grouped particles, which possess no correlationsin between the groupings, and by letting the groups to be of finite but sufficiently largesize, one can attain the ultimate asymptotic SQL bound up to any precision. Sincesaturability of the QFI bounds for independently prepared probes is well established[62–64, 92], the operational meaning of the QFI is also clear in the single shot scenario.The above argument also suggests the asymptotically optimal form of the input states,which should include ones that do not possess long-range correlations in between theparticles. This observation has already been made in [93] and indicates that in methodsdesigned to search for the optimal inputs in scenarios with uncorrelated noise one may fficient tools for quantum metrology with uncorrelated noise a priori knowledge should not play an important role. Inparticular, we expect that various methods such as Bayesian [95–98] or informationtheoretic [99, 100] should recover the bounds compatible with the ones obtained via the local approach considered in this paper. This statement is known to hold in the case ofoptical interferometry with losses [31, 32]; it is an intriguing question whether analogousclaims can be made in more general scenarios.
8. Summary and outlook
We have constructed explicit methods capable of determining fundamental bounds onquantum enhancement in metrological setups in presence of uncorrelated noise. Themethods are based on the study of the structure of a single-particle quantum channelthat represents the decoherence process. The methods do not require any kind ofeducated guess, nor an involved numerical optimization—given a set of Kraus operatorsrepresenting the channel, bounds on precision can be derived immediately without theneed to search e.g. for the optimal input states. We have discussed the efficiency ofCS, QS, RLD and CE methods in providing asymptotic bounds on precision for phaseestimation under a number of different decoherence models. We have also generalizedthe CE method in order to provide tighter bounds in the regime of finite number ofparticles and we have showed that this generalization can be again cast in the form ofa semi-definte program. The methods have also been applied to a related problem offrequency estimation. Moreover, it has been shown that when thinking of estimationof the decoherence parameter itself already the simplest approach based on the CSmethod typically provides the tightest bounds. While the methods are efficient asthey avoid the search for the optimal many particle input states, formulation of anexplicit optimal estimation strategy may in general require performing such a search.Hopefully, the optimal states are expected to have a relatively simple structure and canbe searched within a restricted class of states such as e.g. squeezed or matrix productstates [73, 77, 93, 94]. Once the precision calculated for a given input state hits thefundamental bound one is guaranteed that the optimal strategy has been identified. Thenatural future work on our methods is to generalize them and study their applicabilityin the multi-parameter estimation schemes where it is a priori not clear which of thedifferent approaches will be the most fruitful and whether some nontrivial new boundsmay be derived.
Acknowledgments
We would like to thank Mădălin Guţă and Lorenzo Maccone for valuable feedback aswell as Konrad Banaszek for constant support. J.K. also acknowledges Michal Sedlák fficient tools for quantum metrology with uncorrelated noise
Appendix A. Channels considered
We adopt the standard notation in which I d represents a d × d identity matrix and { σ i } i =1 are the Pauli operators. In Section 4.2 parameter ϕ to be estimated is the rotation anglearound the z axis of the Bloch ball generated by the unitary operator U ϕ = exp[ i σ ϕ/ .We consider maps D η with decoherence parameter η that commute with such rotation,whence Λ ϕ [ ̺ ] = D η (cid:2) U ϕ ̺U † ϕ (cid:3) = U ϕ D η [ ̺ ] U † ϕ , and are defined accordingly by the Krausoperators presented below. For each case, we also specify the purification determiningthe extended channel QFI (12) ( F [Λ ϕ ⊗ I ] in Table 1) in the form of the optimalgenerator of Kraus representation rotation h , as introduced in (10). Dealing withfrequency estimation tasks discussed in Section 5 we construct the effective one-particleKraus operators by substituting ϕ → ωt and η → η ( t ) in the nominal ones, where ω is theestimated frequency detuning. For all models, we explicitly write the Liouvillian L ( n ) determining the noise affecting each particle, see (30), and the effective form of η ( t ) .When discussing decoherence strength estimation in Section 6, we consider solely eachof the following noise maps with η being now the parameter to be estimated: Λ ϕ = η = D η . Appendix A.1. Dephasing • Decoherence parameter , η , represents the final equatorial radius of the Bloch ballshrank uniformly in the xy plane by the channel. • Kraus operators : K = r η I , K = r − η σ . (A.1) • Optimal purification determining the extended channel QFI (12): h = p − η σ . (A.2) • One-particle Liouvillian for frequency estimation tasks: L ( n ) [ ̺ ] = γ (cid:16) σ ( n )3 ̺ σ ( n )3 − ̺ (cid:17) ∴ η ( t ) = e − γt . (A.3) Appendix A.2. Depolarization • Decoherence parameter , η , represents the final radius of the Bloch ball shrunkisotropically by the channel. fficient tools for quantum metrology with uncorrelated noise • Kraus operators : K = r η I , ( K i = r − η σ i − ) i =2 ... . (A.4) • Optimal purification determining the extended channel QFI (12): h = 12 ξ (cid:20) σ (cid:21) ξ with ξ = p (1 + 3 η ) (1 − η )1 + η . (A.5) • One-particle Liouvillian for frequency estimation tasks: L ( n ) [ ̺ ] = γ X i =1 σ ( n ) i ̺ σ ( n ) i − ̺ ! ∴ η ( t ) = e − γ t . (A.6) Appendix A.3. Loss • Decoherence parameter , η , represents survival probability of each of the particlesthat are subject to independent loss processes. The channel on a single particleis formally a map from a two- to a three-dimensional system with the output’sthird dimension corresponding to the vacuum mode responsible for the particleloss. Although in this case one should strictly write Λ ϕ = D η (cid:2) U ϕ ̺U † ϕ (cid:3) = ˜ U ϕ D η [ ̺ ] ˜ U † ϕ with ˜ U ϕ acting on a different Hilbert space, the effects of U ϕ and ˜ U ϕ are physicallyindistinguishable, as the particle losses commute with the acquired phase (forinstance see [30]). In the case of optical interferometry, η represents the effectivepower transmittance assumed to be equal in both arms and accounts for preparationand transmission loss as well as detector inefficiencies in a Mach-Zehnder setup [31]. • Kraus operators : K = √ − η , K = √ − η , K = √ η √ η . (A.7) • Optimal purification determining the extended channel QFI (12): h = − (cid:20) σ (cid:21)
000 0 0 . (A.8) • One-particle Liouvillian for the frequency estimation tasks: L ( n ) [ ̺ ] = γ X m =0 (cid:18) σ ( n ) m, + ̺ σ ( n ) m, − − n σ ( n ) m, − σ ( n ) m, + , ̺ o(cid:19) ∴ η ( t ) = e − γt , (A.9)where σ ( n ) m, + = | vac ih m | are the generators of transition to the vacuum mode fromqubit basis states | i and | i , such that σ ( n ) m, − = σ ( n ) † m, + . fficient tools for quantum metrology with uncorrelated noise Appendix A.4. Spontaneous emission (amplitude damping) • Decoherence parameter , η , represents the radius of the disk obtained by projectingthe deformed Bloch ball outputted by the channel onto the xy plane. • Kraus operators : K = √ η ! , K = √ − η ! . (A.10) • Optimal purification determining the extended channel QFI (12): h = 12 ξ − ! with ξ = 1 − √ η √ η . (A.11) • One-particle Liouvillian for the frequency estimation tasks ( σ ± = ( σ ± i σ ) ): L ( n ) [ ̺ ] = γ (cid:18) σ ( n )+ ̺ σ ( n ) − − n σ ( n ) − σ ( n )+ , ̺ o(cid:19) ∴ η ( t ) = e − γt . (A.12) Appendix B. Equivalence of RLD-based bound applicability and localclassical simulability of a channel
Given a channel—a CPTP map Λ ϕ : H in → H out —we define its C-J matrix representation[61] as Ω Λ ϕ = Λ ϕ ⊗ I [ | I ih I | ] = P i | K i ( ϕ ) ih K i ( ϕ ) | , where { K i ( ϕ ) } ri =1 are r linearlyindependent Kraus operators of Λ ϕ ; we adopt a concise notation for bipartite states, inwhich | φ i = P dim H in i,j =1 h i | φ | j i | i i | j i = φ ⊗ I | I i = I ⊗ φ T | I i with | I i = P dim H in i =1 | i i | i i . Forsimplicity, from now onwards we drop the explicit ϕ dependence of operators, assumingthat the estimation is performed for small variations δϕ around a given, fixed ϕ . InSup. Mat. of [39] (Equation (S9)) it has been proven that the condition for any channelto be ϕ -non-extremal at ϕ is equivalent to the statement that there exists a non-zeroHermitian matrix µ ij such that ˙Ω Λ ϕ = X ij µ ij | K i ih K j | . (B.1)On the other hand, the RLD-based bound exists there if and only if [37] P Ω ⊥ ˙Ω ϕ P Ω ⊥ = 0 (B.2)where P Ω ⊥ is the projection onto the null-space Ω ⊥ , i.e. the subspace orthogonal to Ω Λ ϕ ,so that ∀ i : P Ω ⊥ | K i i = 0 . The (B.1) implies (B.2), as by substitution P Ω ⊥ X ij µ ij | K i ih K j | ! P Ω ⊥ = X ij X p µ ip h K p | K p i µ pj ! P Ω ⊥ | K i ih K j | P Ω ⊥ = 0 , (B.3) fficient tools for quantum metrology with uncorrelated noise ϕ -non-extremal channel must admit an RLD-based bound on its extended QFI.In order to prove the other direction, we split the derivatives of each C-J eigenvectorinto components supported by Ω Λ ϕ and in the null-space Ω ⊥ : (cid:12)(cid:12)(cid:12) ˙ K i E = P j ν ij | K j i + (cid:12)(cid:12) K ⊥ i (cid:11) .Hence, after substituting for ˙Ω Λ ϕ the (B.2) then simplifies to X i (cid:12)(cid:12) K ⊥ i (cid:11) h K i | ! X j | K j i (cid:10) K ⊥ j (cid:12)(cid:12)! = 0 , (B.4)and since A † A = 0 implies A † = A = 0 and {| K i i} i are orthogonal, we conclude that all (cid:12)(cid:12) K ⊥ i (cid:11) = 0 . Thus, (B.2) implies that (cid:12)(cid:12)(cid:12) ˙ K i E = P j ν ij | K j i , which due to the local ambiguityof Kraus representations (9) is equivalent to (cid:12)(cid:12)(cid:12) ˙˜ K i E = P j ( ν ij − i h ij ) (cid:12)(cid:12)(cid:12) ˜ K j E for any Hermitian h . Therefore, without loss of generality, we may set h = − i ν AH after splitting ν intoits Hermitian and anti-Hermitian parts ν = ν H + ν AH , so that (cid:12)(cid:12)(cid:12) ˙˜ K i E = P j ν H ij (cid:12)(cid:12)(cid:12) ˜ K j E with ν H = 0 for any non-trivial channel. Finally, we can write ˙Ω ϕ = X i (cid:12)(cid:12)(cid:12) ˙˜ K i ED ˜ K i (cid:12)(cid:12)(cid:12) + (cid:12)(cid:12)(cid:12) ˜ K i ED ˙˜ K i (cid:12)(cid:12)(cid:12) = 2 X ij ν H ji (cid:12)(cid:12)(cid:12) ˜ K i ED ˜ K j (cid:12)(cid:12)(cid:12) (B.5)and satisfy the condition (B.1). (cid:4) Appendix C. RLD-based bound as a special case of asymptotic CE bound
For a channel that admits an RLD-based bound, in order to obtain the CS condition(B.5) in Appendix B, we chose h = − i ν AH that actually satisfies the β ˜ K = 0 constraint(24) of the CE method. This can be verified by taking the Tr H out { . . . } of the both sidesof the identity X ij h ij | K j ih K i | = X ij i (cid:16) ν ij − ν † ij (cid:17) | K j ih K i | = i X i (cid:12)(cid:12)(cid:12) ˙ K i E h K i | − | K i i D ˙ K i (cid:12)(cid:12)(cid:12) , (C.1)which results in (24). This is consistent, as the CE method must apply to any ϕ -non-extremal channel [39] admitting an RLD-based bound. Furthermore, the asymptotic CEbound (25) is at least as tight as the RLD-based bound (13) on the extended channelQFI (12). We prove this by substituting (B.5) into the definition of F RLD [Λ ϕ ⊗ I ] in(13), so that F RLD [Λ ϕ ⊗ I ] = 4 (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Tr H out (X ij ν H ji (cid:12)(cid:12)(cid:12) ˜ K i E X pq ν H pq D ˜ K q (cid:12)(cid:12)(cid:12))(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) = 4 (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)X i ˙˜ K † i ˙˜ K i (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) , (C.2)where we have used the fact that D ˜ K j (cid:12)(cid:12)(cid:12) Ω − ϕ (cid:12)(cid:12)(cid:12) ˜ K p E = δ jp . Hence, F RLD [Λ ϕ ⊗ I ] isan example of an asymptotic CE-based bound with a possibly sub-optimal Krausrepresentation chosen such that ∀ i : (cid:12)(cid:12)(cid:12) ˙˜ K i E = P j ν H ij (cid:12)(cid:12)(cid:12) ˜ K j E and β ˜ K = 0 . (cid:4) fficient tools for quantum metrology with uncorrelated noise Appendix D. Optimal local QS of a channel
A channel Λ ϕ of rank r , in order to be locally quantum simulable within small deviations δϕ from a given ϕ , must fulfil the condition (see Section 4.1.2) Λ ϕ [ ̺ ] = Tr E Φ E σ (cid:8) U ( ̺ ⊗ | ψ ϕ ih ψ ϕ | ) U † (cid:9) + O ( δϕ ) = r ′ ≥ r X i =1 ¯ K i ( ϕ ) ̺ ¯ K i ( ϕ ) † + O ( δϕ ) , (D.1)where ¯ K i ( ϕ ) = h i | U | ψ ϕ i and {| i i} r ′ i =1 form any basis in the r ′ dimensional H E Φ × H E σ space containing ψ ϕ . Hence, Λ ϕ must admit a Kraus representation { ˜ K i } r ′ i =1 (withpossibly linearly dependent Kraus operators, as for generality we assume r ′ ≥ r ) thatcoincides with the one of (D.1) up to O ( δϕ ) , i.e. satisfies ˜ K i = ¯ K i and ˙˜ K i = ˙¯ K i for all i . We construct a valid decomposition of | ˙ ψ ϕ i into its (normalized) components paralleland perpendicular to ψ ϕ : (cid:12)(cid:12)(cid:12) ˙ ψ ϕ E = i a | ψ ϕ i− i b (cid:12)(cid:12) ψ ⊥ ϕ (cid:11) , where we can choose a, b ∈ R becauseof ∂ ϕ h ψ ϕ | ψ ϕ i = 0 and the irrelevance of the global phase. Then, the asymptotic bound F boundas of (18) determined by the local QS (D.1) at ϕ simply reads F Q [ | ψ ϕ i ] = 4 b andthe required Kraus operators { ˜ K i } r ′ i =1 of Λ ϕ must fulfil conditions ˜ K i = h i | U | ψ ϕ i and ˙˜ K i = h i | U (cid:12)(cid:12)(cid:12) ˙ ψ ϕ E = i a ˜ K i − i b h i | U (cid:12)(cid:12) ψ ⊥ ϕ (cid:11) . Hence, for the local QS of channel Λ ϕ to be valid b must be finite and we must be always able to construct U = ˜ K ab ˜ K + i b ˙˜ K • . . . • ˜ K ab ˜ K + i b ˙˜ K • . . . • ˜ K ab ˜ K + i b ˙˜ K ... . . . ...... ... • . . . • (D.2)with first two columns fixed to give for all i the correct h i | U | ψ ϕ i and h i | U (cid:12)(cid:12) ψ ⊥ ϕ (cid:11) respectively. Due to locality, all entries marked with • in (D.2) can be chosen freelyto satisfy the unitarity condition U † U = U U † = I . Yet, this constraint still requires theKraus operators to simultaneously fulfil i P r ′ i =1 ˙˜ K † i ˜ K i = a I and P r ′ i =1 ˙˜ K † i ˙˜ K i = ( b + a ) I .Without loss of generality, we may shift their phase at ϕ , so that ˜ K i → e − i aϕ ˜ K i andthe conditions become independent of a , i.e. i P r ′ i =1 ˙˜ K † i ˜ K i = 0 and P r ′ i =1 ˙˜ K † i ˙˜ K i = b I .Furthermore, these constraints do not require r ′ > r , as rewriting for example the firstone as i P r ′ i =1 D ˙ ψ ϕ (cid:12)(cid:12)(cid:12) U | i i h i | U | ψ ϕ i = 0 , one can always resolve the identity with somebasis vectors P r ′ i =1 | i i h i | = P ri =1 | e i i h e i | and define linearly independent Kraus operators { K i = h e i | U | ψ ϕ i} ri =1 also fulfilling the necessary requirements.Finally, we may conclude that Λ ϕ is locally quantum simulable at ϕ , if it admitsthere a Kraus representation satisfying conditions (21) stated in the main text, whichdue to locality can be generated via (9) by some Hermitian r × r matrix h . (cid:4) EFERENCES Appendix E. Finite- N CE method as a semi-definite programming task
The finite-N CE bound has been defined in (27) as F CE N = 4 min h (cid:8) k α ˜ K k + ( N − k β ˜ K k (cid:9) , (E.1)where k·k denotes the operator norm, α ˜ K = P i ˙˜ K † i ˙˜ K i and β ˜ K = i P i ˙˜ K † i ˜ K i . Given achannel Λ ϕ from a d in - to a d out -dimensional Hilbert space and the set of its linearlyindependent Kraus operators ( d out × d in matrices) { K i } ri =1 , in order to compute F CE N , weshould minimize (E.1) over locally equivalent Kraus representations (9) of Λ ϕ generatedby all Hermitian, r × r matrices h .Basing on the results of [39], where the β ˜ K = 0 constraint (24) is also imposed on(E.1), we show that F CE N can always be evaluated by means of semi-definite programming(SDP). Adopting a concise notation in which K is a column vector containing thestarting Kraus operators K i as its elements, we can associate all locally equivalent Krausrepresentations ˜ K in (E.1) with those generated by any h via ˜ K = K and ˙˜ K = ˙ K − i h K .By constructing matrices ( I d represents a d × d identity matrix) A = " √ λ a I d in ˙˜ K † ˙˜ K √ λ a I r · d out B = √ λ b I d in (cid:18) i ˙˜ K † ˜K (cid:19) † i ˙˜ K † ˜K √ λ b I d in , (E.2)which positive semi-definiteness conditions correspond respectively to α ˜ K = ˙˜ K † ˙˜ K ≤ λ a I d in β † ˜ K β ˜ K = ˜K † ˙˜ K ˙˜ K † ˜ K ≤ λ b I d in , (E.3)we rewrite (E.1) into the required SDP form F CE N = 4 min h { λ a + ( N − λ b } , (E.4)s.t. A ≥ , B ≥ . For the purpose of this paper we have implemented a semi-definite program usingthe CVX package for Matlab [102], which efficiently evaluates (E.4) given the set ofKraus operators and their derivatives of a generic channel Λ ϕ . The fact that only K and ˙ K are involved in (E.4) is a consequence of the QFI, and hence the bound F CE N ,being a local quantity.Lastly, one should note that by slightly modifying the program in (E.4) we are ableto also efficiently evaluate: the extended channel QFI (12), as F [Λ ϕ ⊗ I ] = F CE N =1 ; andthe asymptotic extended channel QFI (25), F as [Λ ϕ ⊗ I ] = F CEas , by setting N = 1 andimposing the β ˜ K = 0 constraint (24) as already pursued in [39]. References [1] Leibfried D, Barrett M D, Schaetz T, Britton J, Chiaverini J, Itano W M, JostJ D, Langer C and Wineland D J 2004
Science
EFERENCES
Science
Nature
Nature
Phys. Rev. Lett.
Phys. Rev. Lett.
Phys. Rev. Lett.
Nature
Science
Phys. Rev. Lett. New J. Phys. Nature Phys. Phys. Rev. A R6797–R6800[14] Bollinger J J, Itano W M, Wineland D J and Heinzen D J 1996
Phys. Rev. A R4649–R4652[15] Bužek V, Derka R and Massar S 1999
Phys. Rev. Lett. Phys. Rev. Lett. Phys. Rev. Lett. Nature Photon. Nature Photon. Nature Phys. Contemp. Phys. Phys. Rev. Lett. Phys. Rev. Lett. Phys. Rev. Lett. Phys. Rev. A EFERENCES
New J. Phys. Phys. Rev. A New J. Phys. Phys. Rev. Lett.
Phys. Rev. A Phys. Rev. A Phys. Rev. A J. Phys. A: Math. Gen. IEEE Trans. Inf. Theory J. Phys. A: Math. Theor. ArXiv e-prints
Comm. Math. Phys.
Nature Phys. Nat. Commun. ArXiv e-prints
Phys. Rev. A Phys. Rev. Lett. Phys. Rev. Lett. Phys. Rev. A Phys. Rev. A Phys. Rev. A Phys. Rev. Lett.
Phys. Rev. A Phys. Rev. A Phys. Rev. Lett.
New J.Phys. Phys. Rev. A Phys. Rev. Lett.
The Theory of Open Quantum Systems (Oxford University Press)[55] Andersson E, Cresser J D and Hall M J W 2007
J. Mod. Opt. EFERENCES
Phys. Rev. A Phys. Rev. Lett.
ArXiv e-prints
Fundamentals of Statistical Signal Processing: Estimation Theory (Prentice Hall)[60] Nielsen M A and Chuang I L 2000
Quantum Computing and Quantum Information (Cambridge University Press)[61] Bengtsson I and Życzkowski K 2006
Geometry of quantum states: an introductionto quantum entanglement (Cambridge Univeristy Press)[62] Helstrom C W 1976
Quantum detection and estimation theory (Academic Press)[63] Holevo A S 1982
Probabilistic and Statistical Aspects of Quantum Theory (NorthHolland)[64] Braunstein S L and Caves C M 1994
Phys. Rev. Lett. On Fisher Information of Quantum Statistical Models (World)chap 9, pp 113–125[66] Tóth G and Petz D 2013
Phys. Rev. A ArXiv e-prints
Math. Eng. Tech. Rep. ArXiv e-prints
Phys. Rev. A Ann. Phys.
135 – 173[72] Aharonov Y, Massar S and Popescu S 2002
Phys. Rev. A Phys. Rev. A Phys. Rev. D Bells Theorem, Quantum Theory,and Conceptions of the Universe (Dordrecht, The Netherlands: Kluwer AcademicPublishers)[76] Lee H, Kok P and Dowling J P 2002
J. Mod. Opt. ArXiv e-prints
Phys. Rev. Lett. Phys.Rev. A ArXiv e-prints
J. Phys. A: Math. Gen. EFERENCES
J. Phys. A: Math. Gen. Phys. Rev. A J. Phys. A: Math. Theor. Phys. Rev. Lett. (12) 120501[86] Hofmann H F 2009
Phys. Rev. A Phys. Rev. Lett.
Phys. Rev. Lett.
Phys. Rev. Lett.
Phys. Rev. Lett.
Phys. Rev. Lett.
Comm. Math. Phys.
Phys. Rev. Lett. ArXiv e-prints
Phys. Rev. A Phys. Rev. A J. Phys. B: At., Mol. Opt. Phys. Phys. Rev. A New J. Phys. ArXiv e-prints
Phys. Rev. A063836[102] Grant M and Boyd S 2012 Cvx: Matlab software for disciplined convexprogramming URL