New Phase Transitions in Optimal States for Memory Channels Vahid Karimipour 1, Zohreh Meghdadi2 , Laleh Memarzadeh 3, † † § 9 0 0 Department of Physics, Sharif University of Technology, 2 † 11155-9161, Tehran, Iran n a J Dipartimento di Fisica, Universit‘a di Camerino, I-62032 Camerino, Italy. § 7 1 ] h p - t n Abstract a u We investigate the question of optimal input ensembles for memory channels q and construct a rather large class of Pauli channels with correlated noise which [ canbestudiedanalyticallywithregardtotheentanglementoftheiroptimalinput 1 ensembles. Inamoredetailedstudyofasubclassofthesechannels,thecomplete v phase diagram of the two-qubit channel, which shows three distinct phases is 9 obtained. While increasing the correlation generally changes the optimal state 4 from separable to maximally entangled states, this is done via an intermediate 6 2 regionwherebothseparableandmaximallyentangledstatesareoptimal. Amore . concrete model, based on random rotations of the error operators which mimic 1 the behavior of this subclass of channels is also presented. 0 9 0 : v i X r a PACS Numbers: 03.67.HK, 05.40.Ca 1Corresponding author:[email protected] 2email:[email protected] 3email:[email protected] 1 Introduction A basic question in quantum information theory [1, 2, 3, 4, 5] is whether the use of entangledstatesforencodingclassicalinformationcanincreasetherateofinformation transmission though a channel or not. A proper calculation of the so called Holevo capacity [3] of a channel, representing by a Completely Positive Trace preserving (CPT)mapΦ,requirestheoptimization ofHolevoinformationover ensembleofinput stateswhenweencodeinformationintoarbitrarylongstringsofquantumstates(more precisely states in the tensor product of the Hilbert space of one state) and carrying out the limiting procedure C := lim C , where n n →∞ 1 C := Sup χ (ε) (1) n ε n n is the capacity of the channel, when we send strings of n quantum states into the channel. Here ε := p ,ρ is the ensemble of input states, i i { } χ (ε) := S( p Φ(ρ )) p S(Φ(ρ )) (2) n i i i i − i i X X is the Holevo information of the ensemble and S(ρ) tr(ρlogρ) is the von Neu- ≡ − mann entropy of a state ρ. To find the capacity of a given channel we should find the ensemble which maximizes this quantity and we call it the optimal ensemble of input states. Then the properties of this ensemble can be studied and ask whether this ensemble includes entangled states or not. Theimportanceofthisquestionstemsfromthefactthatentanglementisavitalquan- tum mechanical resource in many tasks in quantum information processing, however it is a held belief that entanglement is so fragile in the presence of noise. So it would be interesting if one can show by using this property higher rate of data transmission through noisy channels is achievable. The difficulty in answering this question is not only because of the optimization of Holevo information over multi-parameter space but also due to the fact that no concrete classification of multi-particle entangled states exist. A much simpler problem is to calculate C , rather than C , which is equivalent 2 n to refresh the channel after each two uses, and see whether entangled states can en- hance Holevo information or not. This problem has been tackled by many authors [6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16] and for a kind of correlated channel it has been shown that entanglement can enhance the Holevo information, if the correlation is above a certain threshold value. This correlated channel was first introduced in [9] as follows 3 Φ(ρ) = P σ σ ρσ σ , (3) ij i j i j ⊗ ⊗ i,j=0 X whereσ ,i = 0,1,2,3 representPaulioperatorsI,σ ,σ ,σ andP ,theprobabilities i x y z ij of errors σ σ are correlated in a special way, namely i j ⊗ P = (1 µ)p p +µδ p . (4) ij i j ij j − 1 The parameter µ [0,1], called memory factor signifies the amount of correlation ∈ in the noise of the channel. For µ = 0, the errors on the two consecutive qubits are completely un-correlated, while for µ = 1, the two errors are exactly the same. The idea behind this, is that when the channel relaxation time is much less than the time interval between the passage of the two qubits, theerrors will bethe same. What was observed in [9] and in subsequent works [12, 14, 16] was that when µ passes a critical value µ , the optimal input state jumps suddenly from product states to maximally c entangled states. It is important to note that although the sharp transition of optimal ensemble of inputstates from productto maximally entangled states is interesting and cause non- analytical behavior of the channel, it is not obvious that the same happens for all kinds of correlated channels. the correlation (4) is one out of many different forms of correlations that one can envisage for a correlated noisy channel. Apart from the mathematical possibility of defining many other forms of correlations, one can also argue on physical grounds, in favor of other forms of correlation. A fully correlated channel which exerts a noise operator on the first qubit, need not exerts the same error operator on the second qubit, as in (4). In fact it is natural to expects that on exerting the first error, the state of the environment will change and depending on this new state, it will exert new errors with conditional probabilities on the second qubit even if the time interval between two consecutive uses of the channel be small. In this paper we try to shed light on these issues and to provide a basis for study- ing more examples and extend the study of correlated noisy channels by considering more general forms of correlations. Our study will be along the line of references [9, 12, 14, 16], that is we do not consider a specific model of environment, rather we take the abstract definition of the channel as a CPT linear map, defined by its Kraus decomposition[17]. Westartinsection 2withthemostgeneralformoftheaction ofa correlated Pauli channel on two qubits, where by two plausible requirements, we will restrict the parameters so that the members of this class can be studied by analytical means. Then we focus on one particular subclass and study in detail the optimal ensemble which maximizes its Holevo information. In this same section we propose a general definition of the correlation parameter, and show that when the correlation of the noise in this channel increases, the optimal ensemble changes from a product ensemble to a maximally entangled one. The phase diagram of the model also shows otherinterestingtransitionsnotalreadyobservedinotherworks[9,12,14,16]. Infact the phase diagram contains three distinct phases, a product one and two maximally entangled ones, which differ with each other by a relative phase. We also observe that when we consider the correlation parameter and move through this phase diagram, the optimal ensemble changes from product to maximally entangled, however this transition is mediated by a region of correlation, where both maximally entangled and separable ensembles are optimal. Finally we construct a rather concrete model for this particular type of correlation. The paper concludes with a discussion. 2 2 The correlated action of a Pauli channel on two qubits The general action of a Pauli channel on two qubits is defined by the following CPT map Φ(ρ) = P σ σ ρσ σ , (5) ij i j i j ⊗ ⊗ ij X where P is the probability of the errors σ and σ on the first and second qubits ij j i entering the channel respectively. The number of independent error probabilities are 15 due to normalization P = 1. i,j ij P In order to see how much correlated the noise on the two consecutive qubits are, we form the marginal error probabilities p(1) and p(2) and evaluate the following dis- (1) (2) tance between the two probability distributions, p and p p , denoted as ij i j C 3 1 (1) (2) := P p p (6) C 2 | ij − i j | i,j=0 X (1) (2) where p := P and p = P . For uncorrelated noise we will have = 0 i j ij i j ji C and for a fully correlated noise will attain its maximum value. For the probability P C P distribution (4), turns out to be = µ 3 p (1 p ), hence for fixed error proba- C C i=0 i − i bilities, the correlations indeed increase with µ. P There are two difficulties in studying such a channel for obtaining its optimal in- put ensemble and understanding how it depends on the correlation of the noise. The first one is that the probability distribution P can be correlated in many different ij ways and picking out a single parameter and designate it as the memory of the chan- nel, is just one possibility. The second problem is that the manifold of input states has itself 6 real parameters which means that the optimization task should be done over a 6 parameter space and thus analytical treatment of such a channel is almost impossible. To overcome these problems, we impose the following symmetry on the channel, Φ(σ σ ρσ σ ) = Φ(ρ). (7) 3 3 3 3 ⊗ ⊗ Thisisthesymmetrywhichhasbeenconsideredin[12]formakingthemodelamenable toanalytical treatment. Demandingthissymmetryreducesthenumberofparameters in P to 7 parameters: ij p t u s v q r w P =  , (8) w r q v    s u t p      with normalization relation between the parameters. However it is more plausible to assume that the marginal error probabilities on the first and the second qubits be equal, that is (1) (2) p = p i, (9) i i ∀ 3 Here we are assuming that the errors on a sequence of qubits should be the same, regardless of how we enumerate the qubits of the sequence. What really matters is that any two consecutive uses of the channel are correlated. This assumptions in addition to the constrained composed by the symmetry in (7 reduce the number of parameters from to 6 and the final form of the matrix of prob- abilities P with elements can be parameterized as follows: p η+ξ η ξ s − 4 4 η+γ q r η γ P =  4 −4 , (10) η γ r q η+γ − 4 4  s η−ξ η+ξ p   4 4    where p+q+r+η+s = 1. 2 The advantage of demanding the symmetry in (7) in not only in reducing the param- eters of P but also in reducing the parameters of the general input states. Following [12], we form the optimal ensemble by finding a state ρ which minimizes ∗ the output entropy and hence minimizing the second term in the right hand side of (2). The input ensemble is then formed as a uniform distribution of the states = ρ := (σ σ )ρ (σ σ ) . The reason is that the first term of the Holevo ij i j ∗ i j E { ⊗ ⊗ } quantity is maximized by this choise, i.e. 1 S( p Φ(ρ )) = S( σ σ Φ(ρ )σ σ ) ij ij i j ∗ i j 16 ⊗ ⊗ ij ij X X 1 = S( I) = 2, (11) 4 where in the first line we have used the covariance property of the channel: Φ(σ σ ρσ σ )= (σ σ )Φ(ρ)(σ σ ). (12) i j i j i j i j ⊗ ⊗ ⊗ ⊗ and in the second line the Shur’s first lemma for irreducible representations of the Pauli group. Therefore finding the optimum ensemble of input states reduces to finding a sin- gle input state which minimizes the output entropy and we call it optimum input state. Regarding the convexity of entropy we deduce that we should search for the optimal input state among pure states which in general has 6 parameters. However, we are able to restrict the form of the input states to the simple states which are invariant under the above symmetry in (7) ψ = cosθ 00 +sinθeiφ 11 , (13) | i | i | i Itiseasytofindtheoutputstateofthischannelwiththeaboveformoferrorprobabil- ities. Astraightforward calculation showsthat theoutputstate, inthecomputational 4 basis 00 , 01 , 10 , 11 will be {| i | i | i | i} ε 0 0 ε 00 03 0 ε ε 0 Φ(ρ) =  11 12  (14) 0 ε ε 0 ∗12 22    ε∗03 0 0 ε33    where ε = 2(p+s)cos2(θ)+2(q+r)sin2(θ) 00 ε = 2(p+s)sin2(θ)+2(q+r)cos2(θ) 33 ε = sin(2θ)[(p s)e iφ+(q r)eiφ] 03 − 1 − − ε = sin(2θ)[ξe iφ +γeiφ] 12 − 2 ε = ε = η. (15) 11 22 Due to the block diagonal structure of this matrix, its eigenvalues can be calculated in closed form and hence a complete specification of the optimal input states can be made for this general class of correlated Pauli channels. 2.1 Detailed study of a subclass and its phase diagram As an interesting and simple example, we consider a channel where only the param- eters p, q and r are non-vanishing, that is matrix of probabilities has the following form p 0 0 0 0 q r 0 P =  , (16) 0 r q 0    0 0 0 p      where due to normalization p+q +r = 1. The action of the Pauli channel on two 2 consecutive qubits will then be: Φ(ρ) = pρ+pσ σ ρσ σ +qσ σ ρσ σ +qσ σ ρσ σ 3 3 3 3 1 1 1 1 2 2 2 2 ⊗ ⊗ ⊗ ⊗ ⊗ ⊗ + rσ σ ρσ σ +rσ σ ρσ σ . (17) 1 2 1 2 2 1 2 1 ⊗ ⊗ ⊗ ⊗ The correlation parameter of this channel is found to be := 3p 4p2+ (q+r)2 q + (q+r)2 r . (18) C − | − | | − | Using (15), the eigenvalues of the output density matrix are 1 λ = 0, λ = 1 1 16[p(q+r)+Y sin2(2θ)] , (19) 1,2 3,4 2 ± − (cid:18) q (cid:19) where Y := q(r p)cos2φ+r(q p)sin2φ. (20) − − For minimization of output entropy which is S(Φ(ρ)) = λ logλ λ logλ , we 1 1 2 2 − − should maximize the difference between λ and λ . This is achieved by minimizing 1 2 5 Figure1: ColorOnline. Thephasediagramofmodel(14). Ineachphasetheminimum output entropy state (un-normalized) is specified. The triple point is (q,p) = (1,1). 6 6 Y sin22θ. Therefore if Y 0, we should choose θ = π and if Y 0, we should ≤ 4 ≥ choose θ = 0. Therefore the line Y = 0 determines the boundary of the maximally entangled (θ = π) and the product (θ = 0) optimal states. This line is determined 4 by the equations q(r p)= 0 and r(q p) = 0. In the (q,r) plane these two lines are − − specified by the equations 2q+r = 1 and 2r+q = 1. 2 2 When θ = 0, the value of Y is immaterial, however when θ = π, then the value of 4 Y should be minimized again. In the maximally entangled region, where θ = π, we 4 shouldstillminimizeY. Forq(r p) r(q p),Y takes itsminimumatφ = 0,andfor − ≥ − q(r p)< r(q p)it willtake its minimumat φ = π. Thusthe lineq(r p)= r(q p) − − 2 − − (q = r)willseparatetwotypesofoptimalmaximallyentangledstatesfromeachother. The phase diagram is shown in figure (1). It is seen from the phase diagram (2) that as the correlation parameter increases, the optimal input ensemble changes from product to maximally entangled. There are however two remarkable features in this diagram not encountered in previous studies. First we see that dependingon the values of the channel parameters, two different types of maximally entangled states, namely 1 (00 + 11 ) and 1 (00 +i11 ) are √2 | i | i √2 | i | i optimal. Although these two types of states, are transformed to each other by a local 1 0 operator I , since the channel is not covariant under this local operator, ⊗ 0 i ! they should be considered different as far as optimality of the encoding is concerned, althoughtheyareequivalentasfarastheirentanglementpropertiesareconcerned[20]. Second, whenwedrawthecontours of constantcorrelations inthisphasediagram, 6 Figure 2: Color Online. The phase diagram of model (14). In each phase the min- imum output entropy state (un-normalized) is specified. The contours of constant correlations are also shown. As the correlation increases, the optimal ensemble C changes from separable to maximally entangled one. For 0.43 < < 0.5, both ≈ C ≈ types of states are optimal, the grey region shown in the inset. we observed that there is a region of correlation, for which both separable and maxi- mally entangled states are optimal, depending on the values of the parameters q and r, figure (2). We say that the two phases coexist, a property which is reminiscent of first-order transitions. We should stress that if one draws the phase diagram of the model in [12], not in terms of the parameter µ, but in terms of the correlation parameter C, one will again see such coexistence region. Therefore and specially with regard to recent studies in relating transitions in channel capacity to the critical transitions in their environment [18, 19], it is an interesting issue to see if transitions in the channel should be characterized as first or second order. 2.2 A concrete and intuitive model of correlation In this section we construct a particular model for correlated noise in Pauli channels, which will reproduce the above example of correlation in a natural way. Consider a noisy Pauli channel acting on a qubit, defined as 3 Φ(ρ) = p σ ρσ , (21) i i i i=0 X 7 where p is the probability of error σ (σ = I) and 3 p = 1. When the first qubit i i 0 i=0 i passes through the channel, and an error operator σ acts on it, we assume that the Pi state of the channel changes randomly and therefore on the second qubit, it exerts not the same error or a fixed error for that matter, but a random rotation of the σ i operator, in the form σ˜i := Un,θσiUn†,θ, (22) whereUn,θ is arandomrotation aroundtheaxis nwith angle θ. Thusσ˜ has theeffect of the first error operator and also the random change in the environment. This ran- domness in contrast to a deterministic change in the environment state is physically plausible in view of the macroscopic nature of the environment. Therefore the action of the channel on two consecutive qubits may be written as follows 3 Φ(ρ)= p ρ+ p σ˜ σ ρ σ˜ σ . (23) 0 i i i i i ⊗ ⊗ i=1 X Sincetherotations arerandom,thecompletedefinitionofthechannelwillbegiven by integrating over theabove action withasuitableprobability distributionover random rotations. Thus the final definition will be Φ (ρ) = dnˆdθP(n,θ)Φ(ρ). (24) σ Z Clearly one can add more parameters to the above model, for example by taking different rotations to be along different axes. For simplicity, let us restrict ourself to a simple example in which the direction of all rotations are fixed in the z- axis and only the angle of rotation is random. Also to ensure the symmetry (7) we take p = p , and p = p , where p +p = 1. We take the probability distribution to be 0 3 1 2 0 1 2 a Gaussian with mean value θ and variance σ. Hence the channel will be defined by 0 dθ (θ−θ0)2 Φσ(ρ) = e− 2σ2 Φ(ρ), (25) σ√2π Z wherein this case σ˜i = e−2iθσzσie2iθσz. One can say that parameter σ is related to the memory of the channel. When σ = 0, the channel has full memory and it will exert a definite error operator (exactly the same error in the case θ = 0) on the second qubit 0 depending on the operator which it has exerted on the first. However for a non-zero small value of σ, the channel exerts errors on the second qubit which are close to the errors on the first qubit. As σ increases further the memory is lost further and thechannelwillexerterrorsfromalargerneighborhoodoftheerrorsonthefirstqubit. A remark is in order about the Gaussian distribution. The rotation operators are periodic which restrict the range of integration of θ to [0,2π]. However this makes the subsequent formulas unduly cumbersome without adding much to the physics. Instead we can assume the variance σ to be sufficiently less than 2π so that we can safely extend the range of integration of θ to ( , ) and use the simple results of −∞ ∞ 8 Figure3: ColorOnline. Thephasediagramofmodel(24). Ineachphasetheminimum output entropy state is specified. Gaussian integration. After rearranging and doingthe integrals, one findsthat this channel has theform (17) with the parameters as given below p = p , (26) 0 1+e 2σ2 − q = p ( ), (27) 1 2 1 e 2σ2 − r = p ( − ), (28) 1 2 The two independent parameters of this channel can be taken to be p and σ. In 1 terms of these parameters the phase diagram is shown in figure (3). Since we have always r < q, the region with 1 (00 + i11 ) optimal state is not covered in this √2 | i | i new phase diagram. The line which separates the product phase from the maximally entangled phase is now given by (29). Inserting the values of q and r from (27) and (28) in the relation 2r+q = 1 and simplifying we obtain 2 1 e 2σ2 = 3 . (29) − − p 1 The phase diagram in figure(2) is re-drawn in terms of the new parameters in figure (3). It is seen that depending on the value of p , the optimal ensemble changes from 1 separable to maximally entangled phase, whenthe memory passes a certain threshold (note that here a lower value of σ means a larger value of memory). Also there are values of p , where the optimal ensemble is always a maximally entangled one, no 1 matter how weak the memory is. This is related to the fact that for no value of the parameter σ, this channel is a product channel. 9