Introduction
Entanglement is a fundamental resource for a wide range of quantum information processing tasks, including quantum communication, distributed quantum computation, and quantum cryptography [1 , 2 , 3 , 4 , 5 , 6 ] . However, noise arising during entanglement generation, distribution, and storage typically degrades shared entangled states. Entanglement purification protocols (EPPs) [7 ] were therefore introduced to distill high-fidelity entangled states from noisy inputs using local operations and classical communication (LOCC) [8 , 9 , 10 ] .
Many EPPs, however, are tailored to specified input-state families and assume access to independently and identically distributed (i.i.d.) copies. In realistic scenarios, only finitely many incompletely characterized input states may be available, and they may not be identical. These limitations motivate two longstanding objectives in entanglement purification: reducing the prior information required about the inputs and dispensing with the assumption that i.i.d. copies are available [11 , 12 ] .
Progress toward the first objective includes the Schmidt projection method of entanglement concentration, which requires no prior knowledge of the Schmidt coefficients but assumes identically prepared pure-state inputs [13 ] , and EPPs for unknown or partially known states under an exchangeability assumption [14 ] . Universal protocols [15 ] include distortion-free entanglement concentration for unknown bipartite pure states [16 ] and restricted classes of multipartite pure states [17 ] , as well as asymptotic entanglement distillation from unknown i.i.d. copies under non-entangling operations [18 ] . Progress toward relaxing the assumption of i.i.d. copies includes optimal single-copy probabilistic transformations of bipartite pure states [19 , 20 , 21 ] and entanglement enhancement and local filtering for two-qubit states [22 , 23 , 24 ] . General one-shot formulations have further characterized nonasymptotic entanglement distillation under LOCC
and broader classes of operations [25 , 26 , 27 ] . The universal results above retain the i.i.d. assumption, whereas the finite-resource results apply only to particular input classes or under additional assumptions. Thus, the two objectives have largely been pursued separately. This raises a natural question: does there exist a universal LOCC EPP applicable to finitely many, possibly nonidentical input states when only limited prior information is available?
More recently, this question has been investigated within the framework of universal n n n -to-1 LOCC EPPs [28 , 29 , 30 , 31 , 32 ] . Given n n n input states ρ 1 , … , ρ n \rho_1,\ldots,\rho_n ρ 1 , … , ρ n , let Λ 0 \Lambda_0 Λ 0 denote the successful branch of an n n n -to-1 LOCC EPP acting on their total state ρ i n = ⨂ j = 1 n ρ j \rho^{\mathrm{in}} = \bigotimes_{j=1}^{n}\rho_j ρ in = ⨂ j = 1 n ρ j . Such an EPP is universal if, for every admissible choice of the input states, the success probability p ρ i n = Tr [ Λ 0 ( ρ i n ) ] p_{\rho^{\mathrm{in}}} = \operatorname{Tr}[\Lambda_0(\rho^{\mathrm{in}})] p ρ in = Tr [ Λ 0 ( ρ in )] is positive and the normalized successful output ρ ′ = Λ 0 ( ρ i n ) / p ρ i n \rho' = \Lambda_0(\rho^{\mathrm{in}})/p_{\rho^{\mathrm{in}}} ρ ′ = Λ 0 ( ρ in ) / p ρ in satisfies F ( ρ ′ ) ≥ max 1 ≤ j ≤ n F ( ρ j ) F(\rho') \ge\max_{1\le j\le n} F(\rho_j) F ( ρ ′ ) ≥ max 1 ≤ j ≤ n F ( ρ j ) , where all fidelities are evaluated with respect to the same prescribed target state [28 ] . The same work also proved that no such EPP exists for the full set of two-qubit entangled states. However, it has remained open whether this no-go result extends, at fixed finite local dimensions, to all higher-dimensional bipartite entangled states and all multipartite entangled states.
In this work, we resolve the above question. Specifically, we show that for every n ≥ 2 n \ge2 n ≥ 2 and any fixed finite local dimensions, no n n n -to-1 LOCC EPP is universal for either the full set of bipartite entangled states or the full set of multipartite entangled states. Moreover, we show that the bipartite and multipartite no-go results remain valid under separable operations, when the allowed inputs are restricted to states with fidelities arbitrarily close to unity, and with the assistance from a finite-dimensional pure catalyst. Beyond entanglement, we establish an analogous no-go theorem for coherence purification under strictly incoherent operations.
Preliminaries
Let us consider n ≥ 2 n \ge2 n ≥ 2 entangled states ρ 1 , … , ρ n \rho_1,\ldots,\rho_n ρ 1 , … , ρ n shared by m ≥ 2 m \ge2 m ≥ 2 spatially separated parties, denoted by P = \mathcal{P} = P = { A , B , C , … } \{A, B, C, \ldots\} { A , B , C , … } . Let H = ⨂ X ∈ P H X \mathcal{H} = \bigotimes_{X\in\mathcal{P}} \mathcal{H}_X H = ⨂ X ∈ P H X , where H X \mathcal{H}_X H X is the finite-dimensional local Hilbert space of party X X X . Each ρ j \rho_j ρ j acts on an isomorphic copy of H \mathcal{H} H , whose local subsystem corresponding to party X X X is denoted by X j X_j X j , with H X j ≅ H X \mathcal{H}_{X_j} \cong\mathcal{H}_X H X j ≅ H X . Their tensor product ρ i n : = ⨂ j = 1 n ρ j \rho^{\mathrm{in}} := \bigotimes_{j=1}^{n}\rho_j ρ in := ⨂ j = 1 n ρ j is the total input state. For the output system, the local subsystem of party X X X is denoted by X ′ X' X ′ , with H X ′ ≅ H X \mathcal{H}_{X'} \cong\mathcal{H}_X H X ′ ≅ H X . Let B ( K ) \mathcal{B}(\mathcal{K}) B ( K ) denote the space of linear operators on a Hilbert space K \mathcal{K} K . An n n n -to-1 EPP is a two-outcome LOCC quantum instrument Λ = { Λ 0 , Λ 1 } \Lambda= \{\Lambda_0,\Lambda_1\} Λ = { Λ 0 , Λ 1 } [9 , 33 ] , where Λ 0 , Λ 1 : B ( H ⊗ n ) → B ( H ) \Lambda_0,\Lambda_1 : \mathcal{B}(\mathcal{H}^{\otimes n})\to\mathcal{B}(\mathcal{H}) Λ 0 , Λ 1 : B ( H ⊗ n ) → B ( H ) are completely positive and trace-nonincreasing maps corresponding to the success and failure branches, respectively, and Λ 0 + Λ 1 \Lambda_0+\Lambda_1 Λ 0 + Λ 1 is trace preserving. For the total input state ρ i n \rho^{\mathrm{in}} ρ in , the success probability is p s = Tr [ Λ 0 ( ρ i n ) ] p_s = \operatorname{Tr}[\Lambda_0(\rho^{\mathrm{in}})] p s = Tr [ Λ 0 ( ρ in )] . Whenever p s > 0 p_s>0 p s > 0 , the normalized output conditioned on success is ρ ′ = Λ 0 ( ρ i n ) / p s \rho'=\Lambda_0(\rho^{\mathrm{in}})/p_s ρ ′ = Λ 0 ( ρ in ) / p s .
We first define universality for EPPs acting on fixed finite-dimensional bipartite states. For m = 2 m=2 m = 2 , let H = C d 1 ⊗ C d 2 \mathcal{H}=\mathbb{C}^{d_1}\otimes\mathbb{C}^{d_2} H = C d 1 ⊗ C d 2 , with the two factors associated with parties A and B, respectively. For the j j j th input state ρ j \rho_j ρ j , the corresponding local subsystems are denoted by A j A_j A j and B j B_j B j . Let S d 1 , d 2 \mathcal{S}_{d_1,d_2} S d 1 , d 2 denote the set of all entangled states on H \mathcal{H} H . Set d = min { d 1 , d 2 } d=\min\{d_1,d_2\} d = min { d 1 , d 2 } and choose a maximally entangled target state of Schmidt rank d d d , denoted by ∣ Φ d ⟩ |\Phi_d\rangle ∣ Φ d ⟩ . Maximally entangled states are natural targets for purification because they are key resources for various quantum information processing tasks, such as quantum teleportation, quantum networks, and quantum communication [1 , 5 ] . In suitable local bases, the target state can be written as ∣ Φ d ⟩ = d − 1 / 2 ∑ r = 0 d − 1 ∣ r ⟩ A ∣ r ⟩ B |\Phi_d\rangle=d^{-1/2}\sum_{r=0}^{d-1}|r\rangle_A|r\rangle_B ∣ Φ d ⟩ = d − 1/2 ∑ r = 0 d − 1 ∣ r ⟩ A ∣ r ⟩ B , and we denote Φ d : = ∣ Φ d ⟩ ⟨ Φ d ∣ \Phi_d:=|\Phi_d\rangle\langle\Phi_d| Φ d := ∣ Φ d ⟩ ⟨ Φ d ∣ . For any state σ \sigma σ on H \mathcal{H} H , the fidelity [34 ] with the target state is F ( σ ) : = ⟨ Φ d ∣ σ ∣ Φ d ⟩ F(\sigma):=\langle\Phi_d|\sigma|\Phi_d\rangle F ( σ ) := ⟨ Φ d ∣ σ ∣ Φ d ⟩ . Generalizing the definition introduced for 2 ⊗ 2 2\otimes2 2 ⊗ 2 entangled states in Ref. [28 ] , we call an n n n -to-1 EPP { Λ 0 , Λ 1 } \{\Lambda_0,\Lambda_1\} { Λ 0 , Λ 1 } universal for S d 1 , d 2 \mathcal{S}_{d_1,d_2} S d 1 , d 2 if, for every total input state ρ i n = ⨂ j = 1 n ρ j \rho^{\mathrm{in}}=\bigotimes_{j=1}^{n}\rho_j ρ in = ⨂ j = 1 n ρ j with ρ j ∈ S d 1 , d 2 \rho_j\in\mathcal{S}_{d_1,d_2} ρ j ∈ S d 1 , d 2 for all j = 1 , … , n j=1,\ldots,n j = 1 , … , n , the protocol succeeds with probability p s ( ρ i n ) : = Tr [ Λ 0 ( ρ i n ) ] > 0 p_s(\rho^{\mathrm{in}}):=\operatorname{Tr}[\Lambda_0(\rho^{\mathrm{in}})]>0 p s ( ρ in ) := Tr [ Λ 0 ( ρ in )] > 0 and, conditioned on success, produces the normalized output state ρ ′ = Λ 0 ( ρ i n ) / p s ( ρ i n ) \rho'=\Lambda_0(\rho^{\mathrm{in}})/p_s(\rho^{\mathrm{in}}) ρ ′ = Λ 0 ( ρ in ) / p s ( ρ in ) satisfying F ( ρ ′ ) ≥ max 1 ≤ j ≤ n F ( ρ j ) F(\rho')\geq\max_{1\leq j\leq n}F(\rho_j) F ( ρ ′ ) ≥ max 1 ≤ j ≤ n F ( ρ j ) .
We next define universality for EPPs acting on genuinely multipartite entangled states. For m ≥ 3 m\geq3 m ≥ 3 , let d X ≥ 2 d_X\geq2 d X ≥ 2 denote the local dimension of party X ∈ P X\in\mathcal{P} X ∈ P , and write d = ( d A , d B , d C , … ) \mathbf{d}=(d_A,d_B,d_C,\ldots) d = ( d A , d B , d C , … ) . We take H X = C d X \mathcal{H}_X=\mathbb{C}^{d_X} H X = C d X for every X ∈ P X\in\mathcal{P} X ∈ P and let G d \mathcal{G}_{\mathbf{d}} G d denote the set of genuinely multipartite entangled states on H = ⨂ X ∈ P C d X \mathcal{H}=\bigotimes_{X\in\mathcal{P}}\mathbb{C}^{d_X} H = ⨂ X ∈ P C d X . Set d = min X ∈ P d X d=\min_{X\in\mathcal{P}}d_X d = min X ∈ P d X and D = ∏ X ∈ P d X D=\prod_{X\in\mathcal{P}}d_X D = ∏ X ∈ P d X . For each party X ∈ P X\in\mathcal{P} X ∈ P , let { ∣ r ⟩ X } r = 0 d X − 1 \{|r\rangle_X\}_{r=0}^{d_X-1} { ∣ r ⟩ X } r = 0 d X − 1 denote a suitable local basis. As the target state, we choose the generalized GHZ state [35 ] ∣ Ψ d ⟩ = d − 1 / 2 ∑ r = 0 d − 1 ⨂ X ∈ P ∣ r ⟩ X |\Psi_d\rangle=d^{-1/2}\sum_{r=0}^{d-1}\bigotimes_{X\in\mathcal{P}}|r\rangle_X ∣ Ψ d ⟩ = d − 1/2 ∑ r = 0 d − 1 ⨂ X ∈ P ∣ r ⟩ X , with Ψ d : = ∣ Ψ d ⟩ ⟨ Ψ d ∣ \Psi_d:=|\Psi_d\rangle\langle\Psi_d| Ψ d := ∣ Ψ d ⟩ ⟨ Ψ d ∣ . For any state σ \sigma σ on H \mathcal{H} H , its fidelity with the target GHZ state is F ( σ ) : = ⟨ Ψ d ∣ σ ∣ Ψ d ⟩ F(\sigma):=\langle\Psi_d|\sigma|\Psi_d\rangle F ( σ ) := ⟨ Ψ d ∣ σ ∣ Ψ d ⟩ . An n n n -to-1 EPP is universal for G d \mathcal{G}_{\mathbf{d}} G d if, for every total input state ρ i n = ⨂ j = 1 n ρ j \rho^{\mathrm{in}}=\bigotimes_{j=1}^{n}\rho_j ρ in = ⨂ j = 1 n ρ j with ρ j ∈ G d \rho_j\in\mathcal{G}_{\mathbf{d}} ρ j ∈ G d , the protocol succeeds with probability p s ( ρ i n ) > 0 p_s(\rho^{\mathrm{in}})>0 p s ( ρ in ) > 0 and, conditioned on success, produces an output state ρ ′ \rho' ρ ′ satisfying F ( ρ ′ ) ≥ max 1 ≤ j ≤ n F ( ρ j ) F(\rho')\geq\max_{1\leq j\leq n}F(\rho_j) F ( ρ ′ ) ≥ max 1 ≤ j ≤ n F ( ρ j ) .
No universal LOCC entanglement purification.
With the above definitions in place, we first consider the bipartite case.
Theorem 1. For any finite local dimensions d 1 , d 2 ≥ 2 d_1,d_2\geq2 d 1 , d 2 ≥ 2 and any n ≥ 2 n\geq2 n ≥ 2 , no n n n -to-1 LOCC EPP is universal for S d 1 , d 2 \mathcal{S}_{d_1,d_2} S d 1 , d 2 .
Proof. We begin by considering the case d 1 = d 2 = d d_1=d_2=d d 1 = d 2 = d and write S d : = S d , d \mathcal{S}_d:=\mathcal{S}_{d,d} S d := S d , d . Suppose, for contradiction, that a universal n n n -to-1 LOCC EPP for S d \mathcal{S}_d S d exists.
First, we construct input states for which universality requires the normalized output conditioned on success to be maximally entangled.
For each pair A j B j A_jB_j A j B j , let ∣ Φ d ⟩ j = d − 1 / 2 ∑ r = 0 d − 1 ∣ r ⟩ A j ∣ r ⟩ B j |\Phi_d\rangle_j=d^{-1/2}\sum_{r=0}^{d-1}|r\rangle_{A_j}|r\rangle_{B_j} ∣ Φ d ⟩ j = d − 1/2 ∑ r = 0 d − 1 ∣ r ⟩ A j ∣ r ⟩ B j and Φ d ( j ) : = ∣ Φ d ⟩ j ⟨ Φ d ∣ \Phi_d^{(j)}:=|\Phi_d\rangle_j\langle\Phi_d| Φ d ( j ) := ∣ Φ d ⟩ j ⟨ Φ d ∣ . Define τ j = p Φ d ( j ) + ( 1 − p ) I A j B j / d 2 \tau_j=p\Phi_d^{(j)}+(1-p)I_{A_jB_j}/d^2 τ j = p Φ d ( j ) + ( 1 − p ) I A j B j / d 2 , where I A j B j I_{A_jB_j} I A j B j is the identity on H A j ⊗ H B j \mathcal{H}_{A_j}\otimes\mathcal{H}_{B_j} H A j ⊗ H B j and 1 d + 1 < p < 1 \frac{1}{d+1}<p<1 d + 1 1 < p < 1 . For this range of p p p , every τ j \tau_j τ j is full rank and entangled, and hence τ j ∈ S d \tau_j\in\mathcal{S}_d τ j ∈ S d . We consider the following n + 1 n+1 n + 1 total input states
Ω 0 = ⨂ j = 1 n Φ d ( j ) , Ω k = Φ d ( k ) ⊗ ⨂ j ≠ k τ j , k = 1 , … , n . (1) \Omega_0=\bigotimes_{j=1}^{n}\Phi_d^{(j)},\qquad
\Omega_k=\Phi_d^{(k)}\otimes\bigotimes_{j\ne k}\tau_j,\quad k=1,\ldots,n. \tag*{(1)} Ω 0 = j = 1 ⨂ n Φ d ( j ) , Ω k = Φ d ( k ) ⊗ j = k ⨂ τ j , k = 1 , … , n . (1) Let ∣ Φ d ⟩ o u t = d − 1 / 2 ∑ r = 0 d − 1 ∣ r ⟩ A ′ ∣ r ⟩ B ′ |\Phi_d\rangle_{\mathrm{out}}=d^{-1/2}\sum_{r=0}^{d-1}|r\rangle_{A'}|r\rangle_{B'} ∣ Φ d ⟩ out = d − 1/2 ∑ r = 0 d − 1 ∣ r ⟩ A ′ ∣ r ⟩ B ′ denote the corresponding maximally entangled target state on the output pair A ′ B ′ A'B' A ′ B ′ . Since each Ω k \Omega_k Ω k contains the target state Φ d \Phi_d Φ d among its n n n input states, the maximum input fidelity is 1. Universality therefore requires the normalized output conditioned on success to equal Φ d o u t \Phi_d^{\mathrm{out}} Φ d out . Equivalently,
Λ 0 ( Ω k ) = p k Φ d o u t , k = 0 , … , n , (2) \Lambda_0(\Omega_k)=p_k\Phi_d^{\mathrm{out}},\qquad k=0,\ldots,n, \tag*{(2)} Λ 0 ( Ω k ) = p k Φ d out , k = 0 , … , n , (2) where p k : = Tr [ Λ 0 ( Ω k ) ] > 0 p_k:=\operatorname{Tr}[\Lambda_0(\Omega_k)]>0 p k := Tr [ Λ 0 ( Ω k )] > 0 . Since Λ 0 \Lambda_0 Λ 0 is the successful branch of an LOCC protocol, it admits a product Kraus representation Λ 0 ( ρ ) = ∑ μ K μ ρ K μ † \Lambda_0(\rho)=\sum_\mu K_\mu\rho K_\mu^\dagger Λ 0 ( ρ ) = ∑ μ K μ ρ K μ † , with K μ = M μ ⊗ N μ K_\mu=M_\mu\otimes N_\mu K μ = M μ ⊗ N μ . Here, M μ : ⨂ j = 1 n H A j → H A ′ M_\mu:\bigotimes_{j=1}^{n}\mathcal{H}_{A_j}\to\mathcal{H}_{A'} M μ : ⨂ j = 1 n H A j → H A ′ and N μ : ⨂ j = 1 n H B j → H B ′ N_\mu:\bigotimes_{j=1}^{n}\mathcal{H}_{B_j}\to\mathcal{H}_{B'} N μ : ⨂ j = 1 n H B j → H B ′ .
Second, we derive support constraints for every product Kraus operator in the successful branch and then select one satisfying a nonzero condition.
Let L o u t \mathcal{L}_{\mathrm{out}} L out be the subspace spanned by ∣ Φ d ⟩ o u t |\Phi_d\rangle_{\mathrm{out}} ∣ Φ d ⟩ out . Then P ⊥ = I A ′ B ′ − Φ d o u t P_\perp=I_{A'B'}-\Phi_d^{\mathrm{out}} P ⊥ = I A ′ B ′ − Φ d out is the orthogonal projector onto L o u t ⊥ \mathcal{L}_{\mathrm{out}}^\perp L out ⊥ , where I A ′ B ′ I_{A'B'} I A ′ B ′ is the identity on the output Hilbert space. For k = 1 , … , n k=1,\ldots,n k = 1 , … , n , Eq. (2) implies
P ⊥ Λ 0 ( Ω k ) P ⊥ = ∑ μ ( P ⊥ K μ Ω k 1 / 2 ) ( P ⊥ K μ Ω k 1 / 2 ) † = 0. (3) P_\perp\Lambda_0(\Omega_k)P_\perp
=\sum_\mu(P_\perp K_\mu\Omega_k^{1/2})(P_\perp K_\mu\Omega_k^{1/2})^\dagger=0. \tag*{(3)} P ⊥ Λ 0 ( Ω k ) P ⊥ = μ ∑ ( P ⊥ K μ Ω k 1/2 ) ( P ⊥ K μ Ω k 1/2 ) † = 0. (3) Since all terms in Eq. (3) are positive semidefinite and their sum is zero, each term must vanish separately. Consequently, P ⊥ K μ Ω k 1 / 2 = 0 P_\perp K_\mu\Omega_k^{1/2}=0 P ⊥ K μ Ω k 1/2 = 0 for every μ \mu μ . As the range of Ω k 1 / 2 \Omega_k^{1/2} Ω k 1/2 equals supp Ω k \operatorname{supp}\Omega_k supp Ω k , it follows that K μ ( supp Ω k ) ⊆ L o u t K_\mu(\operatorname{supp}\Omega_k)\subseteq\mathcal{L}_{\mathrm{out}} K μ ( supp Ω k ) ⊆ L out for every μ \mu μ . Given that every τ j \tau_j τ j is full rank, we have supp Ω k = span { ∣ Φ d ⟩ k } ⊗ H k \operatorname{supp}\Omega_k=\operatorname{span}\{|\Phi_d\rangle_k\}\otimes\mathcal{H}_k supp Ω k = span { ∣ Φ d ⟩ k } ⊗ H k , where H k : = ⨂ j ≠ k ( H A j ⊗ H B j ) \mathcal{H}_k:=\bigotimes_{j\ne k}(\mathcal{H}_{A_j}\otimes\mathcal{H}_{B_j}) H k := ⨂ j = k ( H A j ⊗ H B j ) is the Hilbert space of all pairs except A k B k A_kB_k A k B k . Accordingly, we obtain
K μ ( ∣ Φ d ⟩ k ⊗ ∣ ψ ⟩ k ) ∈ L o u t , (4) K_\mu\bigl(|\Phi_d\rangle_k\otimes|\psi\rangle_k\bigr)\in\mathcal{L}_{\mathrm{out}}, \tag*{(4)} K μ ( ∣ Φ d ⟩ k ⊗ ∣ ψ ⟩ k ) ∈ L out , (4) for every μ \mu μ , every k = 1 , … , n k=1,\ldots,n k = 1 , … , n , and every ∣ ψ ⟩ k ∈ H k |\psi\rangle_k\in\mathcal{H}_k ∣ ψ ⟩ k ∈ H k . In particular, the vector ∣ Ω 0 ⟩ : = ⨂ j = 1 n ∣ Φ d ⟩ j |\Omega_0\rangle:=\bigotimes_{j=1}^{n}|\Phi_d\rangle_j ∣ Ω 0 ⟩ := ⨂ j = 1 n ∣ Φ d ⟩ j belongs to supp Ω k \operatorname{supp}\Omega_k supp Ω k for every k = 1 , … , n k=1,\ldots,n k = 1 , … , n . We use Ω 0 \Omega_0 Ω 0 to select a Kraus operator that acts nontrivially on this common vector. Since Ω 0 \Omega_0 Ω 0 is pure, Eq. (2) gives ∑ μ ∣ ψ μ ⟩ ⟨ ψ μ ∣ = p 0 Φ d o u t \sum_\mu|\psi_\mu\rangle\langle\psi_\mu|=p_0\Phi_d^{\mathrm{out}} ∑ μ ∣ ψ μ ⟩ ⟨ ψ μ ∣ = p 0 Φ d out , where ∣ ψ μ ⟩ : = K μ ∣ Ω 0 ⟩ |\psi_\mu\rangle:=K_\mu|\Omega_0\rangle ∣ ψ μ ⟩ := K μ ∣ Ω 0 ⟩ . The condition p 0 > 0 p_0>0 p 0 > 0 ensures that at least one ∣ ψ μ ⟩ |\psi_\mu\rangle ∣ ψ μ ⟩ is nonzero. Moreover, the one-dimensional support of Φ d o u t \Phi_d^{\mathrm{out}} Φ d out implies that
every ∣ v μ ⟩ |v_\mu\rangle ∣ v μ ⟩ belongs to L o u t \mathcal{L}_{\mathrm{out}} L out . We may therefore choose a product Kraus operator, denoted by K = M ⊗ N K=M\otimes N K = M ⊗ N , such that
K ∣ Ω 0 ⟩ = c ∣ Φ d ⟩ o u t , c ≠ 0. (5) K|\Omega_0\rangle=c|\Phi_d\rangle_{\mathrm{out}},\qquad c\ne0. \tag*{(5)} K ∣ Ω 0 ⟩ = c ∣ Φ d ⟩ out , c = 0. (5)
The selected K K K therefore satisfies Eq. (4) for every k = 1 , … , n k=1,\ldots,n k = 1 , … , n .
We next reduce K K K to a product operator acting only on the first two pairs while retaining the nonzero condition and the support constraints corresponding to k = 1 , 2 k=1,2 k = 1 , 2 . If n = 2 n=2 n = 2 , we simply set K ~ = K \widetilde{K}=K K = K , and no reduction is required. We therefore consider n > 2 n>2 n > 2 . Let R = { 3 , … , n } \mathcal{R}=\{3,\ldots,n\} R = { 3 , … , n } denote the set of indices of the remaining pairs. Using the local bases in which ∣ Φ d ⟩ |\Phi_d\rangle ∣ Φ d ⟩ has the standard form, for a multi-index z = ( z 3 , … , z n ) ∈ { 0 , … , d − 1 } n − 2 z=(z_3,\ldots,z_n)\in\{0,\ldots,d-1\}^{n-2} z = ( z 3 , … , z n ) ∈ { 0 , … , d − 1 } n − 2 , define ∣ z ⟩ A R = ⨂ j = 3 n ∣ z j ⟩ A j |z\rangle_{A_{\mathcal{R}}}=\bigotimes_{j=3}^{n}|z_j\rangle_{A_j} ∣ z ⟩ A R = ⨂ j = 3 n ∣ z j ⟩ A j and ∣ z ⟩ B R = ⨂ j = 3 n ∣ z j ⟩ B j |z\rangle_{B_{\mathcal{R}}}=\bigotimes_{j=3}^{n}|z_j\rangle_{B_j} ∣ z ⟩ B R = ⨂ j = 3 n ∣ z j ⟩ B j . The product of the maximally entangled states on the remaining pairs can be expanded as
⨂ j = 3 n ∣ Φ d ⟩ j = d − ( n − 2 ) / 2 ∑ z ∣ z ⟩ A R ∣ z ⟩ B R . \bigotimes_{j=3}^{n}|\Phi_d\rangle_j=d^{-(n-2)/2}\sum_z|z\rangle_{A_{\mathcal{R}}}|z\rangle_{B_{\mathcal{R}}}. j = 3 ⨂ n ∣ Φ d ⟩ j = d − ( n − 2 ) /2 z ∑ ∣ z ⟩ A R ∣ z ⟩ B R .
For each z z z , let us define the local operators M z : H A 1 ⊗ H A 2 → H A ′ M_z:\mathcal{H}_{A_1}\otimes\mathcal{H}_{A_2}\to\mathcal{H}_{A'} M z : H A 1 ⊗ H A 2 → H A ′ and N z : H B 1 ⊗ H B 2 → H B ′ N_z:\mathcal{H}_{B_1}\otimes\mathcal{H}_{B_2}\to\mathcal{H}_{B'} N z : H B 1 ⊗ H B 2 → H B ′ by M z ∣ ξ ⟩ = M ( ∣ ξ ⟩ ⊗ ∣ z ⟩ A R ) M_z|\xi\rangle=M(|\xi\rangle\otimes|z\rangle_{A_{\mathcal{R}}}) M z ∣ ξ ⟩ = M ( ∣ ξ ⟩ ⊗ ∣ z ⟩ A R ) and N z ∣ η ⟩ = N ( ∣ η ⟩ ⊗ ∣ z ⟩ B R ) N_z|\eta\rangle=N(|\eta\rangle\otimes|z\rangle_{B_{\mathcal{R}}}) N z ∣ η ⟩ = N ( ∣ η ⟩ ⊗ ∣ z ⟩ B R ) . Using Eq. (5) , we obtain K ∣ Ω 0 ⟩ = d − ( n − 2 ) / 2 ∑ z ( M z ⊗ N z ) ( ∣ Φ d ⟩ 1 ⊗ ∣ Φ d ⟩ 2 ) ≠ 0 K|\Omega_0\rangle=d^{-(n-2)/2}\sum_z(M_z\otimes N_z)(|\Phi_d\rangle_1\otimes|\Phi_d\rangle_2)\ne0 K ∣ Ω 0 ⟩ = d − ( n − 2 ) /2 ∑ z ( M z ⊗ N z ) ( ∣ Φ d ⟩ 1 ⊗ ∣ Φ d ⟩ 2 ) = 0 . Hence there exists at least one multi-index z 0 z_0 z 0 such that
( M z 0 ⊗ N z 0 ) ( ∣ Φ d ⟩ 1 ⊗ ∣ Φ d ⟩ 2 ) ≠ 0. (6) (M_{z_0}\otimes N_{z_0})(|\Phi_d\rangle_1\otimes|\Phi_d\rangle_2)\ne0. \tag*{(6)} ( M z 0 ⊗ N z 0 ) ( ∣ Φ d ⟩ 1 ⊗ ∣ Φ d ⟩ 2 ) = 0. (6)
Let us further define the corresponding product operator K ~ : = M ~ ⊗ N ~ \widetilde{K}:=\widetilde{M}\otimes\widetilde{N} K := M ⊗ N , with M ~ : = M z 0 \widetilde{M}:=M_{z_0} M := M z 0 and N ~ : = N z 0 \widetilde{N}:=N_{z_0} N := N z 0 . For any ∣ ϕ ⟩ 2 ∈ H A 2 ⊗ H B 2 |\phi\rangle_2\in\mathcal{H}_{A_2}\otimes\mathcal{H}_{B_2} ∣ ϕ ⟩ 2 ∈ H A 2 ⊗ H B 2 , the vector ∣ Φ d ⟩ 1 ⊗ ∣ ϕ ⟩ 2 ⊗ ∣ z 0 ⟩ A R ∣ z 0 ⟩ B R |\Phi_d\rangle_1\otimes|\phi\rangle_2\otimes|z_0\rangle_{A_{\mathcal{R}}}|z_0\rangle_{B_{\mathcal{R}}} ∣ Φ d ⟩ 1 ⊗ ∣ ϕ ⟩ 2 ⊗ ∣ z 0 ⟩ A R ∣ z 0 ⟩ B R belongs to supp Ω 1 \operatorname{supp}\Omega_1 supp Ω 1 , because every τ j \tau_j τ j with j ≠ 1 j\ne1 j = 1 is full rank. Hence Eq. (4) implies
K ~ ( ∣ Φ d ⟩ 1 ⊗ ∣ ϕ ⟩ 2 ) ∈ L o u t . (7) \widetilde{K}(|\Phi_d\rangle_1\otimes|\phi\rangle_2)\in\mathcal{L}_{\mathrm{out}}. \tag*{(7)} K ( ∣ Φ d ⟩ 1 ⊗ ∣ ϕ ⟩ 2 ) ∈ L out . (7)
Applying Eq. (4) with k = 2 k=2 k = 2 yields
K ~ ( ∣ ϕ ⟩ 1 ⊗ ∣ Φ d ⟩ 2 ) ∈ L o u t , (8) \widetilde{K}(|\phi\rangle_1\otimes|\Phi_d\rangle_2)\in\mathcal{L}_{\mathrm{out}}, \tag*{(8)} K ( ∣ ϕ ⟩ 1 ⊗ ∣ Φ d ⟩ 2 ) ∈ L out , (8)
for every ∣ ϕ ⟩ 1 ∈ H A 1 ⊗ H B 1 |\phi\rangle_1\in\mathcal{H}_{A_1}\otimes\mathcal{H}_{B_1} ∣ ϕ ⟩ 1 ∈ H A 1 ⊗ H B 1 .
Third, we show that the nonzero condition in Eq. (6) and the support constraints in Eqs. (7) and (8) are incompatible with the product form of K ~ \widetilde{K} K .
We decompose M ~ \widetilde{M} M and N ~ \widetilde{N} N with respect to the local subsystems A 1 A_1 A 1 and B 1 B_1 B 1 of the first pair as M ~ = ∑ i = 0 d − 1 M i ( ⟨ i ∣ A 1 ⊗ I A 2 ) \widetilde{M}=\sum_{i=0}^{d-1}M_i(\langle i|_{A_1}\otimes I_{A_2}) M = ∑ i = 0 d − 1 M i (⟨ i ∣ A 1 ⊗ I A 2 ) and N ~ = ∑ j = 0 d − 1 N j ( ⟨ j ∣ B 1 ⊗ I B 2 ) \widetilde{N}=\sum_{j=0}^{d-1}N_j(\langle j|_{B_1}\otimes I_{B_2}) N = ∑ j = 0 d − 1 N j (⟨ j ∣ B 1 ⊗ I B 2 ) , where M i : H A 2 → H A ′ M_i:\mathcal{H}_{A_2}\to\mathcal{H}_{A'} M i : H A 2 → H A ′ and N j : H B 2 → H B ′ N_j:\mathcal{H}_{B_2}\to\mathcal{H}_{B'} N j : H B 2 → H B ′ . Since these operators map between d d d -dimensional spaces, we identify them with elements of M d ( C ) M_d(\mathbb{C}) M d ( C ) using the local bases in which ∣ Φ d ⟩ |\Phi_d\rangle ∣ Φ d ⟩ has the standard form. For any X ∈ M d ( C ) X\in M_d(\mathbb{C}) X ∈ M d ( C ) , define ∣ ϕ X ⟩ j = ( X ⊗ I B j ) ∣ Φ d ⟩ j |\phi_X\rangle_j=(X\otimes I_{B_j})|\Phi_d\rangle_j ∣ ϕ X ⟩ j = ( X ⊗ I B j ) ∣ Φ d ⟩ j for j = 1 , 2 j=1,2 j = 1 , 2 . As X X X ranges over M d ( C ) M_d(\mathbb{C}) M d ( C ) , the vectors ∣ ϕ X ⟩ j |\phi_X\rangle_j ∣ ϕ X ⟩ j range over the entire space H A j ⊗ H B j \mathcal{H}_{A_j}\otimes\mathcal{H}_{B_j} H A j ⊗ H B j . Using the identity ( A ⊗ B ) ∣ Φ d ⟩ = ( A B T ⊗ I d ) ∣ Φ d ⟩ (A\otimes B)|\Phi_d\rangle=(AB^{T}\otimes I_d)|\Phi_d\rangle ( A ⊗ B ) ∣ Φ d ⟩ = ( A B T ⊗ I d ) ∣ Φ d ⟩ , where the superscript T T T denotes transposition, we obtain
K ~ ( ∣ Φ d ⟩ 1 ⊗ ∣ ϕ X ⟩ 2 ) = 1 d ( ∑ i M i X N i T ⊗ I B ′ ) ∣ Φ d ⟩ o u t . (9) \widetilde{K}(|\Phi_d\rangle_1\otimes|\phi_X\rangle_2)=\frac{1}{\sqrt{d}}\left(\sum_iM_iXN_i^{T}\otimes I_{B'}\right)|\Phi_d\rangle_{\mathrm{out}}. \tag*{(9)} K ( ∣ Φ d ⟩ 1 ⊗ ∣ ϕ X ⟩ 2 ) = d 1 ( i ∑ M i X N i T ⊗ I B ′ ) ∣ Φ d ⟩ out . (9)
By Eq. (7) , the vector in Eq. (9) belongs to L o u t \mathcal{L}_{\mathrm{out}} L out for every X X X . Since ∣ Φ d ⟩ o u t |\Phi_d\rangle_{\mathrm{out}} ∣ Φ d ⟩ out has full Schmidt rank, the map Y ↦ ( Y ⊗ I B ′ ) ∣ Φ d ⟩ o u t Y\mapsto(Y\otimes I_{B'})|\Phi_d\rangle_{\mathrm{out}} Y ↦ ( Y ⊗ I B ′ ) ∣ Φ d ⟩ out is injective. Hence there exists a linear functional α : M d ( C ) → C \alpha:M_d(\mathbb{C})\to\mathbb{C} α : M d ( C ) → C such that, for every X ∈ M d ( C ) X\in M_d(\mathbb{C}) X ∈ M d ( C ) ,
∑ i M i X N i T = α ( X ) I A ′ . (10) \sum_iM_iXN_i^{T}=\alpha(X)I_{A'}. \tag*{(10)} i ∑ M i X N i T = α ( X ) I A ′ . (10)
Similarly, expanding ∣ ϕ X ⟩ 1 = d − 1 / 2 ∑ r , s X r s ∣ r ⟩ A 1 ∣ s ⟩ B 1 |\phi_X\rangle_1=d^{-1/2}\sum_{r,s}X_{rs}|r\rangle_{A_1}|s\rangle_{B_1} ∣ ϕ X ⟩ 1 = d − 1/2 ∑ r , s X r s ∣ r ⟩ A 1 ∣ s ⟩ B 1 gives
K ~ ( ∣ ϕ X ⟩ 1 ⊗ ∣ Φ d ⟩ 2 ) = 1 d ( ∑ r , s X r s M r N s T ⊗ I B ′ ) ∣ Φ d ⟩ o u t . (11) \widetilde{K}(|\phi_X\rangle_1\otimes|\Phi_d\rangle_2)=\frac{1}{\sqrt{d}}\left(\sum_{r,s}X_{rs}M_rN_s^{T}\otimes I_{B'}\right)|\Phi_d\rangle_{\mathrm{out}}. \tag*{(11)} K ( ∣ ϕ X ⟩ 1 ⊗ ∣ Φ d ⟩ 2 ) = d 1 ( r , s ∑ X r s M r N s T ⊗ I B ′ ) ∣ Φ d ⟩ out . (11)
By Eq. (8) , the vector in Eq. (11) belongs to L o u t \mathcal{L}_{\mathrm{out}} L out for every X X X . Hence there exists a linear functional β : M d ( C ) → C \beta:M_d(\mathbb{C})\to\mathbb{C} β : M d ( C ) → C such that, for every X ∈ M d ( C ) X\in M_d(\mathbb{C}) X ∈ M d ( C ) ,
∑ r , s X r s M r N s T = β ( X ) I A ′ . (12) \sum_{r,s}X_{rs}M_rN_s^{T}=\beta(X)I_{A'}. \tag*{(12)} r , s ∑ X r s M r N s T = β ( X ) I A ′ . (12)
In particular, for X = I d X=I_d X = I d , Eqs. (9) and (11) both yield the same nonzero vector K ~ ( ∣ Φ d ⟩ 1 ⊗ ∣ Φ d ⟩ 2 ) \widetilde{K}(|\Phi_d\rangle_1\otimes|\Phi_d\rangle_2) K ( ∣ Φ d ⟩ 1 ⊗ ∣ Φ d ⟩ 2 ) . Hence, α ( I d ) = β ( I d ) ≠ 0 \alpha(I_d)=\beta(I_d)\ne0 α ( I d ) = β ( I d ) = 0 . Let E r s = ∣ r ⟩ ⟨ s ∣ E_{rs}=|r\rangle\langle s| E r s = ∣ r ⟩ ⟨ s ∣ denote the standard matrix units. Setting X = E r s X=E_{rs} X = E r s in Eq. (12) gives
M r N s T = c r s I A ′ , (13) M_rN_s^{T}=c_{rs}I_{A'}, \tag*{(13)} M r N s T = c r s I A ′ , (13)
where c r s : = β ( E r s ) c_{rs}:=\beta(E_{rs}) c r s := β ( E r s ) . Since I d = ∑ i E i i I_d=\sum_iE_{ii} I d = ∑ i E ii and β \beta β is linear, we have ∑ i c i i = β ( I d ) ≠ 0 \sum_i c_{ii}=\beta(I_d)\ne0 ∑ i c ii = β ( I d ) = 0 . Hence there exists an index t t t such that c t t ≠ 0 c_{tt}\ne0 c tt = 0 . For this index, Eq. (13) gives M t N t T = c t t I A ′ M_tN_t^{T}=c_{tt}I_{A'} M t N t T = c tt I A ′ . Since c t t ≠ 0 c_{tt}\ne0 c tt = 0 , both M t M_t M t and N t N_t N t are invertible. Setting s = t s=t s = t in Eq. (13) and using the invertibility of N t T N_t^{T} N t T gives M r = c r t c t t M t M_r=\frac{c_{rt}}{c_{tt}}M_t M r = c tt c r t M t . Likewise, setting r = t r=t r = t and using the invertibility of M t M_t M t gives N s = c t s c t t N t N_s=\frac{c_{ts}}{c_{tt}}N_t N s = c tt c t s N t . Thus, there exist λ i , μ i ∈ C \lambda_i,\mu_i\in\mathbb{C} λ i , μ i ∈ C such that
M i = λ i M t and N i = μ i N t , (14) M_i=\lambda_iM_t\quad\text{and}\quad N_i=\mu_iN_t, \tag*{(14)} M i = λ i M t and N i = μ i N t , (14)
for every i i i . Substituting Eq. (14) into Eq. (10) , we have
S M t X N t T = α ( X ) I A ′ , with S : = ∑ i λ i μ i . (15) S M_tXN_t^{T}=\alpha(X)I_{A'},\qquad\text{with }S:=\sum_i\lambda_i\mu_i. \tag*{(15)} S M t X N t T = α ( X ) I A ′ , with S := i ∑ λ i μ i . (15)
Taking X = I d X=I_d X = I d and using M t N t T = c t t I A ′ M_tN_t^{T}=c_{tt}I_{A'} M t N t T = c tt I A ′ , we obtain S c t t = α ( I d ) Sc_{tt}=\alpha(I_d) S c tt = α ( I d ) . Since α ( I d ) ≠ 0 \alpha(I_d)\ne0 α ( I d ) = 0 and c t t ≠ 0 c_{tt}\ne0 c tt = 0 , it follows that S ≠ 0 S\ne0 S = 0 . Moreover, since N t T = c t t M t − 1 N_t^{T}=c_{tt}M_t^{-1} N t T = c tt M t − 1 , Eq. (15) becomes S c t t M t X M t − 1 = α ( X ) I A ′ Sc_{tt}M_tXM_t^{-1}=\alpha(X)I_{A'} S c tt M t X M t − 1 = α ( X ) I A ′ . Multiplying this equation from the left by M t − 1 M_t^{-1} M t − 1 and from the right by M t M_t M t , we obtain X = α ( X ) S c t t I d X=\frac{\alpha(X)}{Sc_{tt}}I_d X = S c tt α ( X ) I d for every X ∈ M d ( C ) X\in M_d(\mathbb{C}) X ∈ M d ( C ) . However, choosing X = ∣ 0 ⟩ ⟨ 1 ∣ X=|0\rangle\langle1| X = ∣0 ⟩ ⟨ 1∣ would require ∣ 0 ⟩ ⟨ 1 ∣ |0\rangle\langle1| ∣0 ⟩ ⟨ 1∣ to be proportional to I d I_d I d , which is impossible for d ≥ 2 d\ge2 d ≥ 2 . Hence no universal n n n -to-1 LOCC EPP exists for S d \mathcal{S}_d S d .
It remains to consider the case d 1 ≠ d 2 d_1\ne d_2 d 1 = d 2 . Let d = min { d 1 , d 2 } d=\min\{d_1,d_2\} d = min { d 1 , d 2 } , and suppose, to the contrary, that a universal n n n -to-1 LOCC EPP for S d 1 , d 2 \mathcal{S}_{d_1,d_2} S d 1 , d 2 exists. Repeat the construction in Eq. (1) with τ j = p Φ d ( j ) + ( 1 − p ) I A j B j / ( d 1 d 2 ) \tau_j=p\Phi_{d}^{(j)}+(1-p)I_{A_jB_j}/(d_1d_2) τ j = p Φ d ( j ) + ( 1 − p ) I A j B j / ( d 1 d 2 ) , where 1 / ( d + 1 ) < p < 1 1/(d+1)<p<1 1/ ( d + 1 ) < p < 1 . Each τ j \tau_j τ j is full rank. It is also entangled for the chosen range of p p p : indeed, its partial transpose τ j T B j \tau_j^{T_{B_j}} τ j T B j has the eigenvalue ( 1 − p ) / ( d 1 d 2 ) − p / d < 0 (1-p)/(d_1d_2)-p/d<0 ( 1 − p ) / ( d 1 d 2 ) − p / d < 0 , where we have used d 1 d 2 ≥ d 2 d_1d_2\ge d^2 d 1 d 2 ≥ d 2 . Since Φ d ( j ) \Phi_d^{(j)} Φ d ( j ) is entangled as well, every Ω k \Omega_k Ω k is an (unusual total) input state. Since each Ω k \Omega_k Ω k contains a maximally entangled state among its n n n input states, universality again gives Eq. (2) .
Because every τ j \tau_j τ j is full rank, supp Ω k = span { ∣ Φ d ⟩ k } ⊗ H k \operatorname{supp}\Omega_k=\operatorname{span}\{|\Phi_d\rangle_k\}\otimes\mathcal{H}_k supp Ω k = span { ∣ Φ d ⟩ k } ⊗ H k . Repeating the positivity and Kraus-selection arguments above, we obtain a product Kraus operator K = M ⊗ N K=M\otimes N K = M ⊗ N that satisfies Eqs. (4) and (5) for every k k k . We next restrict K K K to the d d d -dimensional subspaces supporting ∣ Φ d ⟩ |\Phi_d\rangle ∣ Φ d ⟩ . In the local bases in which ∣ Φ d ⟩ |\Phi_d\rangle ∣ Φ d ⟩ has the standard form, define K A j = span { ∣ r ⟩ A j } r = 0 d − 1 \mathcal{K}_{A_j}=\operatorname{span}\{|r\rangle_{A_j}\}_{r=0}^{d-1} K A j = span { ∣ r ⟩ A j } r = 0 d − 1 and K B j = span { ∣ r ⟩ B j } r = 0 d − 1 \mathcal{K}_{B_j}=\operatorname{span}\{|r\rangle_{B_j}\}_{r=0}^{d-1} K B j = span { ∣ r ⟩ B j } r = 0 d − 1 for every j j j , together with the corresponding output subspaces K A ′ \mathcal{K}_{A'} K A ′ and K B ′ \mathcal{K}_{B'} K B ′ . Let P A j P_{A_j} P A j , P B j P_{B_j} P B j , P A ′ P_{A'} P A ′ , and P B ′ P_{B'} P B ′ denote the orthogonal projectors onto K A j \mathcal{K}_{A_j} K A j , K B j \mathcal{K}_{B_j} K B j , K A ′ \mathcal{K}_{A'} K A ′ , and K B ′ \mathcal{K}_{B'} K B ′ , respectively, and set P A i n = ⨂ j = 1 n P A j P_A^{\mathrm{in}}=\bigotimes_{j=1}^{n}P_{A_j} P A in = ⨂ j = 1 n P A j and P B i n = ⨂ j = 1 n P B j P_B^{\mathrm{in}}=\bigotimes_{j=1}^{n}P_{B_j} P B in = ⨂ j = 1 n P B j . Define M ~ = P A ′ M P A i n \widetilde{M}=P_{A'}MP_A^{\mathrm{in}} M = P A ′ M P A in and N ~ = P B ′ N P B i n \widetilde{N}=P_{B'}NP_B^{\mathrm{in}} N = P B ′ N P B in , and K ~ = M ~ ⊗ N ~ \widetilde{K}=\widetilde{M}\otimes\widetilde{N} K = M ⊗ N . The input projections P A i n P_A^{\mathrm{in}} P A in and P B i n P_B^{\mathrm{in}} P B in act as the identity on ⨂ j = 1 n K A j \bigotimes_{j=1}^{n}\mathcal{K}_{A_j} ⨂ j = 1 n K A j and ⨂ j = 1 n K B j \bigotimes_{j=1}^{n}\mathcal{K}_{B_j} ⨂ j = 1 n K B j , while the output projection P A ′ ⊗ P B ′ P_{A'}\otimes P_{B'} P A ′ ⊗ P B ′ acts as the identity on L o u t = span { ∣ Φ d ⟩ o u t } ⊆ K A ′ ⊗ K B ′ \mathcal{L}_{\mathrm{out}}=\operatorname{span}\{|\Phi_d\rangle_{\mathrm{out}}\}\subseteq\mathcal{K}_{A'}\otimes\mathcal{K}_{B'} L out = span { ∣ Φ d ⟩ out } ⊆ K A ′ ⊗ K B ′ . Hence K ~ \widetilde{K} K remains nonzero on ⨂ j = 1 n ∣ Φ d ⟩ j \bigotimes_{j=1}^{n}|\Phi_d\rangle_j ⨂ j = 1 n ∣ Φ d ⟩ j and satisfies the restrictions of Eq. (4) to the chosen input subspaces. Applying to K ~ \widetilde{K} K the same reduction to the first two pairs used above yields a nonzero product operator K ~ ~ = M ~ ~ ⊗ N ~ ~ \widetilde{\widetilde{K}}=\widetilde{\widetilde{M}}\otimes\widetilde{\widetilde{N}} K = M ⊗ N satisfying the restrictions of Eqs. (7) and (8) to the chosen d d d -dimensional input and output subspaces. Since all these subspaces are d d d -dimensional, M ~ ~ \widetilde{\widetilde{M}} M and N ~ ~ \widetilde{\widetilde{N}} N may be decomposed as in the d 1 = d 2 d_1=d_2 d 1 = d 2 case, yielding Eqs. (10) and (12) . Therefore, the matrix argument developed above for the case d 1 = d 2 d_1=d_2 d 1 = d 2 applies without modification and yields the same contradiction. ■
Having established that no universal EPP exists in the bipartite setting, we next turn to the multipartite case.
Theorem 2. For any number of parties m ≥ 3 m\geq3 m ≥ 3 , any fixed finite local dimensions d = ( d X ) X ∈ P \boldsymbol{d}=(d_X)_{X\in\mathcal{P}} d = ( d X ) X ∈ P with d X ≥ 2 d_X\geq2 d X ≥ 2 , and any n ≥ 2 n\geq2 n ≥ 2 , no n n n -to-1 LOCC EPP is universal for G d \mathcal{G}_{\boldsymbol{d}} G d .
Proof sketch. For each j = 1 , … , n j=1,\ldots,n j = 1 , … , n , let Ψ d ( j ) \Psi_d^{(j)} Ψ d ( j ) denote the generalized GHZ state on ⨂ X ∈ P H X j \bigotimes_{X\in\mathcal{P}}\mathcal{H}_{X_j} ⨂ X ∈ P H X j , and define τ j = p Ψ d ( j ) + ( 1 − p ) I j / D \tau_j=p\Psi_d^{(j)}+(1-p)I_j/D τ j = p Ψ d ( j ) + ( 1 − p ) I j / D , where ( D − d ) / [ d ( D − 1 ) ] < p < 1 (D-d)/[d(D-1)]<p<1 ( D − d ) / [ d ( D − 1 )] < p < 1 . Each τ j \tau_j τ j is full rank and satisfies F ( τ j ) > 1 / d F(\tau_j)>1/d F ( τ j ) > 1/ d . Since the fidelity of every biseparable state with Ψ d ( j ) \Psi_d^{(j)} Ψ d ( j ) is at most 1 / d 1/d 1/ d , each τ j \tau_j τ j is genuinely multipartite entangled for the chosen range of p p p . Consider the n + 1 n+1 n + 1 inputs Θ 0 : = ⨂ j = 1 n Ψ d ( j ) \Theta_0:=\bigotimes_{j=1}^{n}\Psi_d^{(j)} Θ 0 := ⨂ j = 1 n Ψ d ( j ) and Θ k : = Ψ d ( k ) ⊗ ⨂ j ≠ k τ j \Theta_k:=\Psi_d^{(k)}\otimes\bigotimes_{j\neq k}\tau_j Θ k := Ψ d ( k ) ⊗ ⨂ j = k τ j for k = 1 , … , n k=1,\ldots,n k = 1 , … , n . Each Θ k \Theta_k Θ k contains a Ψ d \Psi_d Ψ d , so universality requires its normalized successful output to be the target state Ψ d \Psi_d Ψ d . The positivity and Kraus-selection arguments of Theorem 1 then yield the corresponding support constraints and a product Kraus operator that is nonzero on Θ 0 \Theta_0 Θ 0 . Across any bipartition of the parties, Ψ d \Psi_d Ψ d is maximally entangled with Schmidt rank d d d , while every multipartite product Kraus operator remains product across that bipartition. Treating the two sides of the bipartition as two effective parties, the support constraints therefore reduce to those in the bipartite proof, giving the same contradiction. The complete proof is provided in the Supplemental Material [36 ] .
We note that the proofs of Theorems 1 and 2 rely only on the product form of the Kraus operators. Since separable operations [37 ] also admit such product Kraus representations, both no-go theorems remain valid under separable operations. We next consider two further extensions of these results.
First, we show that the no-go results in Theorems 1 and 2 persist after restricting the allowed inputs to states above any prescribed fidelity threshold. Fix 0 < ϵ < 1 0<\epsilon<1 0 < ϵ < 1 . In the bipartite case, define S d 1 , d 2 ( ϵ ) : = { ρ ∈ S d 1 , d 2 : F ( ρ ) > 1 − ϵ } \mathcal{S}_{d_1,d_2}^{(\epsilon)}:=\{\rho\in\mathcal{S}_{d_1,d_2}:F(\rho)>1-\epsilon\} S d 1 , d 2 ( ϵ ) := { ρ ∈ S d 1 , d 2 : F ( ρ ) > 1 − ϵ } , and let d = min { d 1 , d 2 } d=\min\{d_1,d_2\} d = min { d 1 , d 2 } . The states used in the proof of Theorem 1 satisfy F ( Φ d ) = 1 F(\Phi_d)=1 F ( Φ d ) = 1 and F ( τ j ) = p + ( 1 − p ) / ( d 1 d 2 ) F(\tau_j)=p+(1-p)/(d_1d_2) F ( τ j ) = p + ( 1 − p ) / ( d 1 d 2 ) . For 1 / ( d + 1 ) < p < 1 1/(d+1)<p<1 1/ ( d + 1 ) < p < 1 , every τ j \tau_j τ j is full rank and entangled, and F ( τ j ) → 1 F(\tau_j)\to1 F ( τ j ) → 1 as p → 1 p\to1 p → 1 . We may therefore choose p p p within this range so that F ( τ j ) > 1 − ϵ F(\tau_j)>1-\epsilon F ( τ j ) > 1 − ϵ . All states used in the construction then belong to S d 1 , d 2 ( ϵ ) \mathcal{S}_{d_1,d_2}^{(\epsilon)} S d 1 , d 2 ( ϵ ) , while the full-rank property of the states τ j \tau_j τ j , and hence the support constraints derived in Theorem 1, remain unchanged. The same contradiction therefore shows that no n n n -to-1 LOCC EPP is universal for S d 1 , d 2 ( ϵ ) \mathcal{S}_{d_1,d_2}^{(\epsilon)} S d 1 , d 2 ( ϵ ) for any n ≥ 2 n\geq2 n ≥ 2 .
The multipartite case is analogous. Define G d ( ϵ ) : = { ρ ∈ G d : F ( ρ ) > 1 − ϵ } \mathcal{G}_{\boldsymbol{d}}^{(\epsilon)}:=\{\rho\in\mathcal{G}_{\boldsymbol{d}}:F(\rho)>1-\epsilon\} G d ( ϵ ) := { ρ ∈ G d : F ( ρ ) > 1 − ϵ } , and let d = min X ∈ P d X \boldsymbol{d}=\min_{X\in\mathcal{P}}d_X d = min X ∈ P d X and D = ∏ X ∈ P d X D=\prod_{X\in\mathcal{P}}d_X D = ∏ X ∈ P d X . In the multipartite construction, F ( Ψ d ) = 1 F(\Psi_d)=1 F ( Ψ d ) = 1 and F ( τ j ) = p + ( 1 − p ) / D F(\tau_j)=p+(1-p)/D F ( τ j ) = p + ( 1 − p ) / D . For ( D − d ) / [ d ( D − 1 ) ] < p < 1 (D-d)/[d(D-1)]<p<1 ( D − d ) / [ d ( D − 1 )] < p < 1 , every τ j \tau_j τ j is full rank and genuinely multipartite entangled, and F ( τ j ) → 1 F(\tau_j)\to1 F ( τ j ) → 1 as p → 1 p\to1 p → 1 . Choosing p < 1 p<1 p < 1 sufficiently close to one within this range places all states used in the construction in G d ( ϵ ) \mathcal{G}_{\boldsymbol{d}}^{(\epsilon)} G d ( ϵ ) without changing the support constraints derived in Theorem 2. The same contradiction therefore applies. These observations give the following corollary.
Corollary 1. For every 0 < ϵ < 1 0<\epsilon<1 0 < ϵ < 1 and n ≥ 2 n\geq2 n ≥ 2 , no n n n -to-1 LOCC EPP is universal for any S d 1 , d 2 ( ϵ ) \mathcal{S}_{d_1,d_2}^{(\epsilon)} S d 1 , d 2 ( ϵ ) with d 1 , d 2 ≥ 2 d_1,d_2\geq2 d 1 , d 2 ≥ 2 or any G d ( ϵ ) \mathcal{G}_{\boldsymbol{d}}^{(\epsilon)} G d ( ϵ ) with d = ( d X ) X ∈ P \boldsymbol{d}=(d_X)_{X\in\mathcal{P}} d = ( d X ) X ∈ P finite, ∣ P ∣ ≥ 3 |\mathcal{P}|\geq3 ∣ P ∣ ≥ 3 , and d X ≥ 2 d_X\geq2 d X ≥ 2 .
Second, we show that the no-go results in Theorems 1 and 2 persist even when the protocol is assisted by a finite-dimensional pure catalyst chosen independently of the input states [38 , 39 , 40 ] . Let ∣ χ ⟩ C |\chi\rangle_C ∣ χ ⟩ C be a catalyst shared among the parties and write χ : = ∣ χ ⟩ ⟨ χ ∣ \chi:=|\chi\rangle\langle\chi| χ := ∣ χ ⟩ ⟨ χ ∣ . An n n n -to-1 catalytic EPP with successful branch Γ 0 \Gamma_0 Γ 0 is called universal if, for every allowed total input state ρ i n = ⨂ j = 1 n ρ j \rho^{\mathrm{in}}=\bigotimes_{j=1}^{n}\rho_j ρ in = ⨂ j = 1 n ρ j , there exist a success probability q ρ i n > 0 q_{\rho}^{\mathrm{in}}>0 q ρ in > 0 and a normalized output state ω ρ i n \omega_{\rho}^{\mathrm{in}} ω ρ in such that Γ 0 ( ρ i n ⊗ χ ) = q ρ i n ω ρ i n ⊗ χ \Gamma_0(\rho^{\mathrm{in}}\otimes\chi)=q_{\rho}^{\mathrm{in}}\omega_{\rho}^{\mathrm{in}}\otimes\chi Γ 0 ( ρ in ⊗ χ ) = q ρ in ω ρ in ⊗ χ and F ( ω ρ i n ) ≥ max 1 ≤ j ≤ n F ( ρ j ) F(\omega_{\rho}^{\mathrm{in}})\geq\max_{1\leq j\leq n}F(\rho_j) F ( ω ρ in ) ≥ max 1 ≤ j ≤ n F ( ρ j ) . The following theorem extends Theorems 1 and 2 to catalytic EPPs. Its proof is given in the Supplemental Material [36 ] .
Theorem 3. Universal EPPs remain impossible under LOCC on the full bipartite and multipartite entangled-state sets considered in Theorems 1 and 2, even with the assistance of pure catalyst.
Before concluding, we note that the no-go results have an analogue in coherence purification [41 , 42 , 43 , 44 ] . Under strictly incoherent operations [41 , 45 , 46 , 47 ] , we establish that no universal n n n -to-1 coherence-purification protocol exists for any n ≥ 2 n\geq2 n ≥ 2 , even for inputs restricted to fidelities above an arbitrarily high threshold below one. The definitions and proof are given in the Supplemental Material [36 ] .
Conclusions
We have established a general no-go theorem for finite-input universal EPPs: for every n ≥ 2 n\geq2 n ≥ 2 and every fixed choice of finite local dimensions, no n n n -to-1 LOCC protocol is universal for either the full set of bipartite entangled states or the full set of multipartite entangled states. This settles the higher-dimensional bipartite and multipartite cases left open in Ref. [28 ] . Taken together, the extensions established above and the analogous coherence no-go result show that the obstruction is neither specific to LOCC nor unique to EPP.
Importantly, these no-go results do not end the search for universal purification, but redirect it toward physically relevant input classes. In a given experiment, the exact input states may be unknown, while the preparation conditions and dominant error processes may still be partially characterized or constrained [48 ] . A central question is therefore how to use such information to identify the broadest admissible input class for which a single EPP can be universal. Addressing this question can guide the design of EPPs for quantum computation, quantum communication, and quantum cryptography, while sharpening our understanding of the limits of entanglement purification.
Acknowledgements
This work was supported by the National Natural Science Foundation of China (Grant No. 12405019) and the Natural Science Foundation of Shandong Province (Grant No. ZR2025LLZ002).
Supplemental Material
In this Supplemental Material, we first provide the complete proof of Theorem 2 in the main text. We then prove Theorem 3, which extends Theorems 1 and 2 to catalytic EPPs. Finally, we give the definitions and proof of the no-go theorem for universal coherence purification under strictly incoherent operations, including the case of input sets whose minimum fidelity with the maximally coherent state can be chosen arbitrarily close to one.
Throughout, an identity operator acting on a Hilbert space is labeled by the corresponding space or subsystem. In particular, I K I_{\mathcal{K}} I K denotes the identity operator on K \mathcal{K} K . In matrix representations, I r I_r I r denotes the r × r r \times r r × r identity matrix.
SI. Complete proof of Theorem 2
Theorem 2. For any number of parties m ≥ 3 m \ge3 m ≥ 3 , any fixed finite local dimensions d = ( d X ) X ∈ P \boldsymbol{d}=(d_X)_{X\in\mathcal{P}} d = ( d X ) X ∈ P with d X ≥ 2 d_X \ge2 d X ≥ 2 , and any n ≥ 2 n \ge2 n ≥ 2 , no n n n -to-1 1 1 LOCC EPP is universal for G d \mathcal{G}_{\boldsymbol{d}} G d .
Proof. We argue by contradiction and suppose that a universal n n n -to-1 1 1 LOCC EPP for G d \mathcal{G}_{\boldsymbol{d}} G d exists.
First, we construct total input states for which universality requires the normalized output conditioned on success to be the generalized GHZ state.
Recall that d = min X ∈ P d X d=\min_{X\in\mathcal{P}}d_X d = min X ∈ P d X and D = ∏ X ∈ P d X D=\prod_{X\in\mathcal{P}}d_X D = ∏ X ∈ P d X . For each j = 1 , … , n j=1,\ldots,n j = 1 , … , n , let Ψ d ( j ) : = ∣ Ψ d ⟩ j ⟨ Ψ d ∣ \Psi_d^{(j)}:=|\Psi_d\rangle_j\langle\Psi_d| Ψ d ( j ) := ∣ Ψ d ⟩ j ⟨ Ψ d ∣ denote the generalized GHZ state on ⨂ X ∈ P H X j \bigotimes_{X\in\mathcal{P}}\mathcal{H}_{X_j} ⨂ X ∈ P H X j and define ∣ Ψ d ⟩ j = d − 1 / 2 ∑ r = 0 d − 1 ⨂ X ∈ P ∣ r ⟩ X j |\Psi_d\rangle_j=d^{-1/2}\sum_{r=0}^{d-1}\bigotimes_{X\in\mathcal{P}}|r\rangle_{X_j} ∣ Ψ d ⟩ j = d − 1/2 ∑ r = 0 d − 1 ⨂ X ∈ P ∣ r ⟩ X j . Define τ j = p Ψ d ( j ) + ( 1 − p ) I j / D \tau_j=p\Psi_d^{(j)}+(1-p)I_j/D τ j = p Ψ d ( j ) + ( 1 − p ) I j / D , where I j I_j I j is the identity on ⨂ X ∈ P H X j \bigotimes_{X\in\mathcal{P}}\mathcal{H}_{X_j} ⨂ X ∈ P H X j and ( D − d ) / [ d ( D − 1 ) ] < p < 1 (D-d)/[d(D-1)]<p<1 ( D − d ) / [ d ( D − 1 )] < p < 1 . Each τ j \tau_j τ j is full rank and satisfies F ( τ j ) = p + ( 1 − p ) / D > 1 / d F(\tau_j)=p+(1-p)/D>1/d F ( τ j ) = p + ( 1 − p ) / D > 1/ d . Across every bipartition of the parties, ∣ Ψ d ⟩ j |\Psi_d\rangle_j ∣ Ψ d ⟩ j has d d d equal
Schmidt coefficients. Thus, the fidelity of any pure state that is product across that bipartition with Ψ d ( j ) \Psi_d^{(j)} Ψ d ( j ) is at most 1 / d 1/d 1/ d . Since every biseparable state is a convex combination of pure states that are product across possibly different bipartitions, its fidelity with Ψ d ( j ) \Psi_d^{(j)} Ψ d ( j ) is also at most 1 / d 1/d 1/ d . It follows that every τ j \tau_j τ j is genuinely multipartite entangled and hence belongs to G d \mathcal{G}_{\boldsymbol{d}} G d .
Let us consider the following n + 1 n+1 n + 1 total input states belong:
Θ 0 = ⨂ j = 1 n Ψ d ( j ) , Θ k = Ψ d ( k ) ⊗ ⨂ j ≠ k τ j , k = 1 , … , n . (16) \Theta_0=\bigotimes_{j=1}^{n}\Psi_d^{(j)},\qquad
\Theta_k=\Psi_d^{(k)}\otimes\bigotimes_{j\ne k}\tau_j,\quad k=1,\ldots,n. \tag*{(16)} Θ 0 = j = 1 ⨂ n Ψ d ( j ) , Θ k = Ψ d ( k ) ⊗ j = k ⨂ τ j , k = 1 , … , n . (16)
Every input state in each Θ k \Theta_k Θ k belongs to G d \mathcal{G}_{\boldsymbol{d}} G d , so every Θ k \Theta_k Θ k is an allowed total input state. Let ∣ Ψ d ⟩ o u t |\Psi_d\rangle_{\mathrm{out}} ∣ Ψ d ⟩ out denote the target state on ⨂ X ∈ P H X ′ \bigotimes_{X\in\mathcal{P}}\mathcal{H}_{X'} ⨂ X ∈ P H X ′ , with Ψ d o u t : = ∣ Ψ d ⟩ o u t ⟨ Ψ d ∣ o u t \Psi_d^{\mathrm{out}}:=|\Psi_d\rangle_{\mathrm{out}}\langle\Psi_d|_{\mathrm{out}} Ψ d out := ∣ Ψ d ⟩ out ⟨ Ψ d ∣ out . Since each of Θ 0 , … , Θ n \Theta_0,\ldots,\Theta_n Θ 0 , … , Θ n contains one input state equal to the corresponding target state, the quantity max 1 ≤ j ≤ n F ( ρ j ) \max_{1\le j\le n}F(\rho_j) max 1 ≤ j ≤ n F ( ρ j ) in the universality condition equals one. Therefore, the normalized successful output must have unit fidelity with Ψ d o u t \Psi_d^{\mathrm{out}} Ψ d out . Equivalently,
Λ 0 ( Θ k ) = p k Ψ d o u t , k = 0 , … , n , (17) \Lambda_0(\Theta_k)=p_k\Psi_d^{\mathrm{out}},\qquad k=0,\ldots,n, \tag*{(17)} Λ 0 ( Θ k ) = p k Ψ d out , k = 0 , … , n , (17)
where p k : = Tr [ Λ 0 ( Θ k ) ] > 0 p_k:=\operatorname{Tr}[\Lambda_0(\Theta_k)]>0 p k := Tr [ Λ 0 ( Θ k )] > 0 . Since Λ 0 \Lambda_0 Λ 0 is the successful branch of an LOCC protocol, it admits a product Kraus representation Λ 0 ( ρ ) = ∑ μ K μ ρ K μ † \Lambda_0(\rho)=\sum_\mu K_\mu\rho K_\mu^\dagger Λ 0 ( ρ ) = ∑ μ K μ ρ K μ † , where K μ = ⨂ X ∈ P M μ , X K_\mu=\bigotimes_{X\in\mathcal{P}}M_{\mu,X} K μ = ⨂ X ∈ P M μ , X , with M μ , X : ⨂ j = 1 n H X j → H X M_{\mu,X}:\bigotimes_{j=1}^{n}\mathcal{H}_{X_j}\to\mathcal{H}_X M μ , X : ⨂ j = 1 n H X j → H X .
Second, we derive support constraints for every product Kraus operator in the successful branch and then select one satisfying a nonzero condition.
Let L o u t \mathcal{L}_{\mathrm{out}} L out be the subspace spanned by ∣ Ψ d ⟩ o u t |\Psi_d\rangle_{\mathrm{out}} ∣ Ψ d ⟩ out and let P ⊥ P_\perp P ⊥ be the orthogonal projector onto L o u t ⊥ \mathcal{L}_{\mathrm{out}}^\perp L out ⊥ . For every k = 1 , … , n k=1,\ldots,n k = 1 , … , n , Eq. (17) implies
P ⊥ Λ 0 ( Θ k ) P ⊥ = ∑ μ ( P ⊥ K μ Θ k 1 / 2 ) ( P ⊥ K μ Θ k 1 / 2 ) † = 0. (18) P_\perp\Lambda_0(\Theta_k)P_\perp
=\sum_\mu\left(P_\perp K_\mu\Theta_k^{1/2}\right)\left(P_\perp K_\mu\Theta_k^{1/2}\right)^\dagger=0. \tag*{(18)} P ⊥ Λ 0 ( Θ k ) P ⊥ = μ ∑ ( P ⊥ K μ Θ k 1/2 ) ( P ⊥ K μ Θ k 1/2 ) † = 0. (18)
Since all terms in Eq. (18) are positive semidefinite and their sum is zero, each term must vanish separately. Consequently, P ⊥ K μ Θ k 1 / 2 = 0 P_\perp K_\mu\Theta_k^{1/2}=0 P ⊥ K μ Θ k 1/2 = 0 for every μ \mu μ . As the range of Θ k 1 / 2 \Theta_k^{1/2} Θ k 1/2 equals supp Θ k \operatorname{supp}\Theta_k supp Θ k , it follows that K μ ( supp Θ k ) ⊆ L o u t K_\mu(\operatorname{supp}\Theta_k)\subseteq\mathcal{L}_{\mathrm{out}} K μ ( supp Θ k ) ⊆ L out for every μ \mu μ . Given that every τ j \tau_j τ j is full rank, we have supp Θ k = span { ∣ Ψ d ⟩ k } ⊗ ⨂ j ≠ k H X j \operatorname{supp}\Theta_k=\operatorname{span}\{|\Psi_d\rangle_k\}\otimes\mathop{\bigotimes}_{j\ne k}\mathcal{H}_{X_j} supp Θ k = span { ∣ Ψ d ⟩ k } ⊗ ⨂ j = k H X j . Here H k ˉ = ⨂ j ≠ k ⨂ X ∈ P H X j \mathcal{H}_{\bar{k}}=\bigotimes_{j\ne k}\bigotimes_{X\in\mathcal{P}}\mathcal{H}_{X_j} H k ˉ = ⨂ j = k ⨂ X ∈ P H X j is the Hilbert space of all input factors except the k k k th. Accordingly, we obtain
K μ ( ∣ Ψ d ⟩ k ⊗ ∣ ψ ⟩ k ˉ ) ∈ L o u t , (19) K_\mu\left(|\Psi_d\rangle_k\otimes|\psi\rangle_{\bar{k}}\right)\in\mathcal{L}_{\mathrm{out}}, \tag*{(19)} K μ ( ∣ Ψ d ⟩ k ⊗ ∣ ψ ⟩ k ˉ ) ∈ L out , (19)
for every μ \mu μ , every k = 1 , … , n k=1,\ldots,n k = 1 , … , n , and every ∣ ψ ⟩ k ˉ ∈ H k ˉ |\psi\rangle_{\bar{k}}\in\mathcal{H}_{\bar{k}} ∣ ψ ⟩ k ˉ ∈ H k ˉ . In particular, ⨂ j = 1 n ∣ Ψ d ⟩ j \bigotimes_{j=1}^{n}|\Psi_d\rangle_j ⨂ j = 1 n ∣ Ψ d ⟩ j belongs to supp Θ k \operatorname{supp}\Theta_k supp Θ k for every k = 1 , … , n k=1,\ldots,n k = 1 , … , n . We use Θ 0 \Theta_0 Θ 0 to select a Kraus operator that acts nontrivially on this common state. Since Θ 0 \Theta_0 Θ 0 is pure, Eq. (17) gives ∑ μ ∣ v μ ⟩ ⟨ v μ ∣ = p 0 Ψ d o u t \sum_\mu|v_\mu\rangle\langle v_\mu|=p_0\Psi_d^{\mathrm{out}} ∑ μ ∣ v μ ⟩ ⟨ v μ ∣ = p 0 Ψ d out , where ∣ v μ ⟩ : = K μ ⨂ j = 1 n ∣ Ψ d ⟩ j |v_\mu\rangle:=K_\mu\bigotimes_{j=1}^{n}|\Psi_d\rangle_j ∣ v μ ⟩ := K μ ⨂ j = 1 n ∣ Ψ d ⟩ j . The condition p 0 > 0 p_0>0 p 0 > 0 ensures that at least one ∣ v μ ⟩ |v_\mu\rangle ∣ v μ ⟩ is nonzero. Moreover, the one-dimensional support of Ψ d o u t \Psi_d^{\mathrm{out}} Ψ d out implies that every ∣ v μ ⟩ |v_\mu\rangle ∣ v μ ⟩ belongs to L o u t \mathcal{L}_{\mathrm{out}} L out . We may therefore choose a product Kraus operator, denoted by K = ⨂ X ∈ P M X K=\bigotimes_{X\in\mathcal{P}}M_X K = ⨂ X ∈ P M X , such that
K ⨂ j = 1 n ∣ Ψ d ⟩ j = c ∣ Ψ d ⟩ o u t , c ≠ 0. (20) K\bigotimes_{j=1}^{n}|\Psi_d\rangle_j=c|\Psi_d\rangle_{\mathrm{out}},\qquad c\ne0. \tag*{(20)} K j = 1 ⨂ n ∣ Ψ d ⟩ j = c ∣ Ψ d ⟩ out , c = 0. (20)
The selected K K K therefore satisfies Eq. (19) for every k = 1 , … , n k=1,\ldots,n k = 1 , … , n .
We next reduce K K K to a product operator acting only on the first two input factors while retaining the nonzero condition and the support constraints corresponding to k = 1 , 2 k=1,2 k = 1 , 2 . If n = 2 n=2 n = 2 , we set K ~ = K \widetilde{K}=K K = K and M ~ X = M X \widetilde{M}_X=M_X M X = M X for every X ∈ P X\in\mathcal{P} X ∈ P , and no reduction is required. We therefore consider n > 2 n>2 n > 2 . Let R = { 3 , … , n } \mathcal{R}=\{3,\ldots,n\} R = { 3 , … , n } denote the index set of the remaining input factors. For a multi-index z = ( z 3 , … , z n ) ∈ { 0 , … , d − 1 } n − 2 z=(z_3,\ldots,z_n)\in\{0,\ldots,d-1\}^{n-2} z = ( z 3 , … , z n ) ∈ { 0 , … , d − 1 } n − 2 , define
∣ z ⟩ X R = ⨂ j = 3 n ∣ z j ⟩ X j |z\rangle_{X_{\mathcal{R}}}=\bigotimes_{j=3}^{n}|z_j\rangle_{X_j} ∣ z ⟩ X R = j = 3 ⨂ n ∣ z j ⟩ X j
for every X ∈ P X\in\mathcal{P} X ∈ P . Thus, ⨂ j = 3 n ∣ Ψ d ⟩ j \bigotimes_{j=3}^{n}|\Psi_d\rangle_j ⨂ j = 3 n ∣ Ψ d ⟩ j can be expanded as
⨂ j = 3 n ∣ Ψ d ⟩ j = 1 d n − 2 ∑ z ⨂ X ∈ P ∣ z ⟩ X R . (21) \bigotimes_{j=3}^{n}|\Psi_d\rangle_j=\frac{1}{\sqrt{d^{n-2}}}\sum_z\bigotimes_{X\in\mathcal{P}}|z\rangle_{X_{\mathcal{R}}}. \tag*{(21)} j = 3 ⨂ n ∣ Ψ d ⟩ j = d n − 2 1 z ∑ X ∈ P ⨂ ∣ z ⟩ X R . (21)
For each z z z , define the local operators M X , z : H X 1 ⊗ H X 2 → H X ′ M_{X,z}:\mathcal{H}_{X_1}\otimes\mathcal{H}_{X_2}\to\mathcal{H}_{X'} M X , z : H X 1 ⊗ H X 2 → H X ′ by M X , z ∣ ξ ⟩ = M X ( ∣ ξ ⟩ ⊗ ∣ z ⟩ X R ) M_{X,z}|\xi\rangle=M_X(|\xi\rangle\otimes|z\rangle_{X_{\mathcal{R}}}) M X , z ∣ ξ ⟩ = M X ( ∣ ξ ⟩ ⊗ ∣ z ⟩ X R ) . Using Eqs. (20) and (21) , we obtain
K ( ⨂ j = 1 n ∣ Ψ d ⟩ j ) = 1 d n − 2 ∑ z ( ⨂ X ∈ P M X , z ) ( ∣ Ψ d ⟩ 1 ⊗ ∣ Ψ d ⟩ 2 ) ≠ 0. (22) K\left(\bigotimes_{j=1}^{n}|\Psi_d\rangle_j\right)=\frac{1}{\sqrt{d^{n-2}}}\sum_z\left(\bigotimes_{X\in\mathcal{P}}M_{X,z}\right)(|\Psi_d\rangle_1\otimes|\Psi_d\rangle_2)\ne0. \tag*{(22)} K ( j = 1 ⨂ n ∣ Ψ d ⟩ j ) = d n − 2 1 z ∑ ( X ∈ P ⨂ M X , z ) ( ∣ Ψ d ⟩ 1 ⊗ ∣ Ψ d ⟩ 2 ) = 0. (22)
Hence there exists at least one multi-index z 0 z_0 z 0 such that the product operator K ~ : = ⨂ X ∈ P M ~ X \widetilde{K}:=\bigotimes_{X\in\mathcal{P}}\widetilde{M}_X K := ⨂ X ∈ P M X , with M ~ X : = M X , z 0 \widetilde{M}_X:=M_{X,z_0} M X := M X , z 0 , satisfies
K ~ ( ∣ Ψ d ⟩ 1 ⊗ ∣ Ψ d ⟩ 2 ) ≠ 0. (23) \widetilde{K}(|\Psi_d\rangle_1\otimes|\Psi_d\rangle_2)\ne0. \tag*{(23)} K ( ∣ Ψ d ⟩ 1 ⊗ ∣ Ψ d ⟩ 2 ) = 0. (23)
For n > 2 n>2 n > 2 and any ∣ ϕ ⟩ 2 ∈ ⨂ X ∈ P H X 2 |\phi\rangle_2\in\bigotimes_{X\in\mathcal{P}}\mathcal{H}_{X_2} ∣ ϕ ⟩ 2 ∈ ⨂ X ∈ P H X 2 , the vector ∣ Ψ d ⟩ 1 ⊗ ∣ ϕ ⟩ 2 ⊗ ⨂ X ∈ P ∣ z 0 ⟩ X R |\Psi_d\rangle_1\otimes|\phi\rangle_2\otimes\bigotimes_{X\in\mathcal{P}}|z_0\rangle_{X_{\mathcal{R}}} ∣ Ψ d ⟩ 1 ⊗ ∣ ϕ ⟩ 2 ⊗ ⨂ X ∈ P ∣ z 0 ⟩ X R belongs to supp Θ 1 \operatorname{supp}\Theta_1 supp Θ 1 , because every τ j \tau_j τ j with j ≠ 1 j\ne1 j = 1 is full rank. Hence Eq. (19) implies
K ~ ( ∣ Ψ d ⟩ 1 ⊗ ∣ ϕ ⟩ 2 ) ∈ L o u t . (24) \widetilde{K}(|\Psi_d\rangle_1\otimes|\phi\rangle_2)\in\mathcal{L}_{\mathrm{out}}. \tag*{(24)} K ( ∣ Ψ d ⟩ 1 ⊗ ∣ ϕ ⟩ 2 ) ∈ L out . (24)
Applying Eq. (19) with k = 2 k=2 k = 2 yields
K ~ ( ∣ ϕ ⟩ 1 ⊗ ∣ Ψ d ⟩ 2 ) ∈ L o u t , (25) \widetilde{K}(|\phi\rangle_1\otimes|\Psi_d\rangle_2)\in\mathcal{L}_{\mathrm{out}}, \tag*{(25)} K ( ∣ ϕ ⟩ 1 ⊗ ∣ Ψ d ⟩ 2 ) ∈ L out , (25)
for every ∣ ϕ ⟩ 1 ∈ ⨂ X ∈ P H X 1 |\phi\rangle_1\in\bigotimes_{X\in\mathcal{P}}\mathcal{H}_{X_1} ∣ ϕ ⟩ 1 ∈ ⨂ X ∈ P H X 1 .
Third, we show that the nonzero property of K ~ \widetilde{K} K established in Eq. (23) and the two support constraints in Eqs. (24) and (25) reduce to the incompatible bipartite conditions derived in the proof of Theorem 1.
Consider the bipartition A ∣ A ‾ A\mid\overline{A} A ∣ A , where A ‾ : = P ∖ { A } \overline{A}:=\mathcal{P}\setminus\{A\} A := P ∖ { A } . For j = 1 , 2 j=1,2 j = 1 , 2 , define U j = span { ∣ r ⟩ A j : 0 ≤ r ≤ d − 1 } \mathcal{U}_j=\operatorname{span}\{|r\rangle_{A_j}:0\le r\le d-1\} U j = span { ∣ r ⟩ A j : 0 ≤ r ≤ d − 1 } and V j = span { ⨂ X ∈ A ‾ ∣ r ⟩ X j : 0 ≤ r ≤ d − 1 } \mathcal{V}_j=\operatorname{span}\{\bigotimes_{X\in\overline{A}}|r\rangle_{X_j}:0\le r\le d-1\} V j = span { ⨂ X ∈ A ∣ r ⟩ X j : 0 ≤ r ≤ d − 1 } , and define the corresponding output subspaces U o u t \mathcal{U}_{\mathrm{out}} U out and V o u t \mathcal{V}_{\mathrm{out}} V out analogously. Let P U j P_{\mathcal{U}_j} P U j , P V j P_{\mathcal{V}_j} P V j , P U o u t P_{\mathcal{U}_{\mathrm{out}}} P U out , and P V o u t P_{\mathcal{V}_{\mathrm{out}}} P V out denote the orthogonal projectors onto these subspaces. After grouping the tensor factors according to the bipartition A ∣ A ‾ A\mid\overline{A} A ∣ A , write K ~ = M ~ A ⊗ N ~ A ‾ \widetilde{K}=\widetilde{M}_A\otimes\widetilde{N}_{\overline{A}} K = M A ⊗ N A , where N ~ A ‾ : = ⨂ X ∈ A ‾ M ~ X \widetilde{N}_{\overline{A}}:=\bigotimes_{X\in\overline{A}}\widetilde{M}_X N A := ⨂ X ∈ A M X . Define a product operator K ^ : = M ^ ⊗ N ^ \widehat{K}:=\widehat{M}\otimes\widehat{N} K := M ⊗ N , where
M ^ : = P U o u t M ~ A ( P U 1 ⊗ P U 2 ) , N ^ : = P V o u t N ~ A ‾ ( P V 1 ⊗ P V 2 ) . (26) \widehat{M}:=P_{\mathcal{U}_{\mathrm{out}}}\widetilde{M}_A(P_{\mathcal{U}_1}\otimes P_{\mathcal{U}_2}),\qquad
\widehat{N}:=P_{\mathcal{V}_{\mathrm{out}}}\widetilde{N}_{\overline{A}}(P_{\mathcal{V}_1}\otimes P_{\mathcal{V}_2}). \tag*{(26)} M := P U out M A ( P U 1 ⊗ P U 2 ) , N := P V out N A ( P V 1 ⊗ P V 2 ) . (26)
For each j = 1 , 2 j=1,2 j = 1 , 2 , identify ⨂ X ∈ A ‾ ∣ r ⟩ X j \bigotimes_{X\in\overline{A}}|r\rangle_{X_j} ⨂ X ∈ A ∣ r ⟩ X j with a basis vector ∣ r ⟩ A ‾ j |r\rangle_{\overline{A}_j} ∣ r ⟩ A j of V j \mathcal{V}_j V j . Under this identification,
∣ Ψ d ⟩ j = 1 d ∑ r = 0 d − 1 ∣ r ⟩ A j ∣ r ⟩ A ‾ j = : ∣ Φ d ⟩ j . (27) |\Psi_d\rangle_j=\frac{1}{\sqrt{d}}\sum_{r=0}^{d-1}|r\rangle_{A_j}|r\rangle_{\overline{A}_j}=:|\Phi_d\rangle_j. \tag*{(27)} ∣ Ψ d ⟩ j = d 1 r = 0 ∑ d − 1 ∣ r ⟩ A j ∣ r ⟩ A j =: ∣ Φ d ⟩ j . (27)
We use the same identification on the output Hilbert space and denote the target state by ∣ Φ d ⟩ o u t |\Phi_d\rangle_{\mathrm{out}} ∣ Φ d ⟩ out . All vectors of the forms appearing on the left-hand sides of Eqs. (24) and (25) belong to L o u t \mathcal{L}_{\mathrm{out}} L out and are therefore proportional to ∣ Ψ d ⟩ o u t |\Psi_d\rangle_{\mathrm{out}} ∣ Ψ d ⟩ out , which lies in U o u t ⊗ V o u t \mathcal{U}_{\mathrm{out}}\otimes\mathcal{V}_{\mathrm{out}} U out ⊗ V out . Consequently, the projection P U o u t ⊗ P V o u t P_{\mathcal{U}_{\mathrm{out}}}\otimes P_{\mathcal{V}_{\mathrm{out}}} P U out ⊗ P V out leaves these vectors unchanged. It follows that
K ^ ( ∣ Φ d ⟩ 1 ⊗ ∣ ϕ ⟩ 2 ) ∈ span { ∣ Φ d ⟩ o u t } , \widehat{K}(|\Phi_d\rangle_1\otimes|\phi\rangle_2)\in\operatorname{span}\{|\Phi_d\rangle_{\mathrm{out}}\}, K ( ∣ Φ d ⟩ 1 ⊗ ∣ ϕ ⟩ 2 ) ∈ span { ∣ Φ d ⟩ out } ,
K ^ ( ∣ ϕ ⟩ 1 ⊗ ∣ Φ d ⟩ 2 ) ∈ span { ∣ Φ d ⟩ o u t } , (28) \widehat{K}(|\phi\rangle_1\otimes|\Phi_d\rangle_2)\in\operatorname{span}\{|\Phi_d\rangle_{\mathrm{out}}\}, \tag*{(28)} K ( ∣ ϕ ⟩ 1 ⊗ ∣ Φ d ⟩ 2 ) ∈ span { ∣ Φ d ⟩ out } , (28)
for every ∣ ϕ ⟩ j ∈ U j ⊗ V j |\phi\rangle_j\in\mathcal{U}_j\otimes\mathcal{V}_j ∣ ϕ ⟩ j ∈ U j ⊗ V j , j = 1 , 2 j=1,2 j = 1 , 2 . Likewise, Eq. (23) gives
K ^ ( ∣ Φ d ⟩ 1 ⊗ ∣ Φ d ⟩ 2 ) ≠ 0. (29) \widehat{K}(|\Phi_d\rangle_1\otimes|\Phi_d\rangle_2)\ne0. \tag*{(29)} K ( ∣ Φ d ⟩ 1 ⊗ ∣ Φ d ⟩ 2 ) = 0. (29)
Since the spaces U 1 \mathcal{U}_1 U 1 , U 2 \mathcal{U}_2 U 2 , U o u t \mathcal{U}_{\mathrm{out}} U out , V 1 \mathcal{V}_1 V 1 , V 2 \mathcal{V}_2 V 2 , and V o u t \mathcal{V}_{\mathrm{out}} V out are all d d d -dimensional, Eqs. (28) and (29) are precisely the two support constraints and the nonzero condition used in the matrix argument for the case d 1 = d 2 d_1=d_2 d 1 = d 2 in the proof of Theorem 1. That argument shows that no nonzero bipartite product operator K ^ \widehat{K} K can satisfy these conditions for d ≥ 2 d\ge2 d ≥ 2 . This contradiction means that no universal n n n -to-1 LOCC EPP exists for G d \mathcal{G}_d G d for any fixed finite local dimensions d = ( d X ) X ∈ P d=(d_X)_{X\in\mathcal{P}} d = ( d X ) X ∈ P .
Exact pure-state catalytic assistance
Previous studies have shown that catalytic assistance can enable state transformations that would otherwise be impossible under LOCC [38 , 39 , 40 ] . In this section, we investigate whether universal entanglement purification can be achieved with the assistance of a finite-dimensional pure catalyst. Let ∣ χ ⟩ C |\chi\rangle_{\mathcal{C}} ∣ χ ⟩ C be the pure catalyst shared among the parties and chosen independently of the input states, where H C : = ⨂ X ∈ P H C X \mathcal{H}_{\mathcal{C}}:=\bigotimes_{X\in\mathcal{P}}\mathcal{H}_{\mathcal{C}_X} H C := ⨂ X ∈ P H C X and H C X \mathcal{H}_{\mathcal{C}_X} H C X is a finite-dimensional Hilbert space associated with party X X X . We write χ : = ∣ χ ⟩ ⟨ χ ∣ \chi:=|\chi\rangle\langle\chi| χ := ∣ χ ⟩ ⟨ χ ∣ .
The successful branch Γ 0 \Gamma_0 Γ 0 of an n n n -to-1 EPP is universal with the assistance of the catalyst χ \chi χ if, for every allowed total input state ρ i n = ⨂ j = 1 n ρ j \rho^{\mathrm{in}}=\bigotimes_{j=1}^{n}\rho_j ρ in = ⨂ j = 1 n ρ j , it satisfies
Γ 0 ( ρ i n ⊗ χ ) = φ ρ i n ω ρ i n ⊗ χ (30) \Gamma_0(\rho^{\mathrm{in}}\otimes\chi)=\varphi_{\rho^{\mathrm{in}}}\omega_{\rho^{\mathrm{in}}}\otimes\chi \tag*{(30)} Γ 0 ( ρ in ⊗ χ ) = φ ρ in ω ρ in ⊗ χ (30)
and F ( ω ρ i n ) ≥ max 1 ≤ j ≤ n F ( ρ j ) F(\omega_{\rho^{\mathrm{in}}})\ge\max_{1\le j\le n}F(\rho_j) F ( ω ρ in ) ≥ max 1 ≤ j ≤ n F ( ρ j ) , where q φ ρ i n > 0 q_{\varphi_{\rho^{\mathrm{in}}}}>0 q φ ρ in > 0 is the success probability, ω ρ i n \omega_{\rho^{\mathrm{in}}} ω ρ in is the normalized output state, and the tensor-product form on the right-hand side expresses that the catalyst is returned exactly and without correlations with the output.
With the above definitions in place, we begin the proof of Theorem 3 in the main text by establishing the following lemma.
Lemma. Suppose that the local Hilbert spaces associated with A 1 , A 2 , A ′ , B 1 , B 2 A_1,A_2,A',B_1,B_2 A 1 , A 2 , A ′ , B 1 , B 2 , and B ′ B' B ′ are all d ( ≥ 2 ) d(\ge2) d ( ≥ 2 ) -dimensional. Let ∣ Φ d ⟩ j |\Phi_d\rangle_j ∣ Φ d ⟩ j , j = 1 , 2 j=1,2 j = 1 , 2 , and ∣ Φ d ⟩ o u t |\Phi_d\rangle_{\mathrm{out}} ∣ Φ d ⟩ out denote the corresponding maximally entangled states on the two input pairs and the output pair, respectively. Let ∣ χ ⟩ C A C B ∈ H C A ⊗ H C B |\chi\rangle_{\mathcal{C}_A\mathcal{C}_B}\in\mathcal{H}_{\mathcal{C}_A}\otimes\mathcal{H}_{\mathcal{C}_B} ∣ χ ⟩ C A C B ∈ H C A ⊗ H C B be any pure state, where H C A \mathcal{H}_{\mathcal{C}_A} H C A and H C B \mathcal{H}_{\mathcal{C}_B} H C B are finite-dimensional, and define L c a t : = span { ∣ Φ d ⟩ o u t ⊗ ∣ χ ⟩ } \mathcal{L}_{\mathrm{cat}}:=\operatorname{span}\{|\Phi_d\rangle_{\mathrm{out}}\otimes|\chi\rangle\} L cat := span { ∣ Φ d ⟩ out ⊗ ∣ χ ⟩} . There is no product operator K ^ = M ~ ⊗ N ~ \widehat{K}=\widetilde{M}\otimes\widetilde{N} K = M ⊗ N , with M ~ : H A 1 ⊗ H A 2 ⊗ H C A → H A ′ ⊗ H C A \widetilde{M}:\mathcal{H}_{A_1}\otimes\mathcal{H}_{A_2}\otimes\mathcal{H}_{\mathcal{C}_A}\to\mathcal{H}_{A'}\otimes\mathcal{H}_{\mathcal{C}_A} M : H A 1 ⊗ H A 2 ⊗ H C A → H A ′ ⊗ H C A , and N ~ : H B 1 ⊗ H B 2 ⊗ H C B → H B ′ ⊗ H C B \widetilde{N}:\mathcal{H}_{B_1}\otimes\mathcal{H}_{B_2}\otimes\mathcal{H}_{\mathcal{C}_B}\to\mathcal{H}_{B'}\otimes\mathcal{H}_{\mathcal{C}_B} N : H B 1 ⊗ H B 2 ⊗ H C B → H B ′ ⊗ H C B , that simultaneously satisfies the nonzero condition
K ^ ( ∣ Φ d ⟩ 1 ⊗ ∣ Φ d ⟩ 2 ⊗ ∣ χ ⟩ ) ≠ 0 (31) \widehat{K}(|\Phi_d\rangle_1\otimes|\Phi_d\rangle_2\otimes|\chi\rangle)\ne0 \tag*{(31)} K ( ∣ Φ d ⟩ 1 ⊗ ∣ Φ d ⟩ 2 ⊗ ∣ χ ⟩) = 0 (31)
and the two support constraints
K ~ ( ∣ Φ d ⟩ 1 ⊗ ∣ ϕ ⟩ 2 ⊗ ∣ χ ⟩ ) ∈ L c a t , \widetilde{K}(|\Phi_d\rangle_1 \otimes|\phi\rangle_2 \otimes|\chi\rangle) \in\mathcal{L}_{\mathrm{cat}}, K ( ∣ Φ d ⟩ 1 ⊗ ∣ ϕ ⟩ 2 ⊗ ∣ χ ⟩) ∈ L cat ,
K ~ ( ∣ ϕ ⟩ 1 ⊗ ∣ Φ d ⟩ 2 ⊗ ∣ χ ⟩ ) ∈ L c a t (32) \widetilde{K}(|\phi\rangle_1 \otimes|\Phi_d\rangle_2 \otimes|\chi\rangle) \in\mathcal{L}_{\mathrm{cat}} \tag*{(32)} K ( ∣ ϕ ⟩ 1 ⊗ ∣ Φ d ⟩ 2 ⊗ ∣ χ ⟩) ∈ L cat (32)
for every ∣ ϕ ⟩ j ∈ H A j ⊗ H B j |\phi\rangle_j \in\mathcal{H}_{A_j} \otimes\mathcal{H}_{B_j} ∣ ϕ ⟩ j ∈ H A j ⊗ H B j , j = 1 , 2 j=1,2 j = 1 , 2 .
Proof. First, we restrict the catalyst state to its Schmidt-support subspaces and translate the two support constraints into matrix equations.
Let the Schmidt decomposition of the catalyst be ∣ χ ⟩ = ∑ a = 1 R χ λ a ∣ a ⟩ C A ∣ a ⟩ C B |\chi\rangle= \sum_{a=1}^{R_\chi} \sqrt{\lambda_a}|a\rangle_{C_A}|a\rangle_{C_B} ∣ χ ⟩ = ∑ a = 1 R χ λ a ∣ a ⟩ C A ∣ a ⟩ C B , where R χ R_\chi R χ is its Schmidt rank. Define S A χ : = span { ∣ a ⟩ C A : 1 ≤ a ≤ R χ } \mathcal{S}_{A}^{\chi} := \operatorname{span}\{|a\rangle_{C_A}: 1 \le a \le R_\chi\} S A χ := span { ∣ a ⟩ C A : 1 ≤ a ≤ R χ } and S B χ : = span { ∣ a ⟩ C B : 1 ≤ a ≤ R χ } \mathcal{S}_{B}^{\chi} := \operatorname{span}\{|a\rangle_{C_B}: 1 \le a \le R_\chi\} S B χ := span { ∣ a ⟩ C B : 1 ≤ a ≤ R χ } , and let P A χ P_A^\chi P A χ and P B χ P_B^\chi P B χ denote the orthogonal projectors onto S A χ \mathcal{S}_{A}^{\chi} S A χ and S B χ \mathcal{S}_{B}^{\chi} S B χ , respectively. Define
M ‾ : = ( I A ′ ⊗ P A χ ) M ~ ( I A 1 ⊗ I A 2 ⊗ P A χ ) , \overline{M} := (I_{A'} \otimes P_A^\chi)\widetilde{M}(I_{A_1} \otimes I_{A_2} \otimes P_A^\chi), M := ( I A ′ ⊗ P A χ ) M ( I A 1 ⊗ I A 2 ⊗ P A χ ) ,
N ‾ : = ( I B ′ ⊗ P B χ ) N ~ ( I B 1 ⊗ I B 2 ⊗ P B χ ) (33) \overline{N} := (I_{B'} \otimes P_B^\chi)\widetilde{N}(I_{B_1} \otimes I_{B_2} \otimes P_B^\chi) \tag*{(33)} N := ( I B ′ ⊗ P B χ ) N ( I B 1 ⊗ I B 2 ⊗ P B χ ) (33)
and set K ‾ : = M ‾ ⊗ N ‾ \overline{K} := \overline{M} \otimes\overline{N} K := M ⊗ N . The input projections I A 1 ⊗ I A 2 ⊗ P A χ I_{A_1} \otimes I_{A_2} \otimes P_A^\chi I A 1 ⊗ I A 2 ⊗ P A χ and I B 1 ⊗ I B 2 ⊗ P B χ I_{B_1} \otimes I_{B_2} \otimes P_B^\chi I B 1 ⊗ I B 2 ⊗ P B χ leave the catalyst unchanged because ( P A χ ⊗ P B χ ) ∣ χ ⟩ = ∣ χ ⟩ (P_A^\chi\otimes P_B^\chi)|\chi\rangle= |\chi\rangle ( P A χ ⊗ P B χ ) ∣ χ ⟩ = ∣ χ ⟩ . Moreover, taking ∣ ϕ ⟩ 2 = ∣ Φ d ⟩ 2 |\phi\rangle_2 = |\Phi_d\rangle_2 ∣ ϕ ⟩ 2 = ∣ Φ d ⟩ 2 in the first support condition of Eq. (32) and using Eq. (31), we find that the vector in Eq. (31) is a nonzero element of L c a t \mathcal{L}_{\mathrm{cat}} L cat . The output projections I A ′ ⊗ P A χ I_{A'} \otimes P_A^\chi I A ′ ⊗ P A χ and I B ′ ⊗ P B χ I_{B'} \otimes P_B^\chi I B ′ ⊗ P B χ therefore leave unchanged all vectors on the left-hand sides of Eqs. (31) and (32). Consequently, Eqs. (31) and (32) remain valid with K ~ \widetilde{K} K replaced by K K K . We may therefore replace H C A \mathcal{H}_{C_A} H C A by S A χ \mathcal{S}_{A}^{\chi} S A χ and H C B \mathcal{H}_{C_B} H C B by S B χ \mathcal{S}_{B}^{\chi} S B χ . For notational simplicity, we relabel M ‾ \overline{M} M and N ‾ \overline{N} N as M ~ \widetilde{M} M and N ~ \widetilde{N} N , respectively. Define D χ : = diag ( λ 1 , … , λ R χ ) D_\chi:= \operatorname{diag}(\sqrt{\lambda_1},\ldots,\sqrt{\lambda_{R_\chi}}) D χ := diag ( λ 1 , … , λ R χ ) and Q χ : = I d ⊗ D χ Q_\chi:= I_d \otimes D_\chi Q χ := I d ⊗ D χ . Since λ a > 0 \lambda_a > 0 λ a > 0 for all a = 1 , … , R χ a=1,\ldots,R_\chi a = 1 , … , R χ , both D χ D_\chi D χ and Q χ Q_\chi Q χ are invertible. Decompose M ~ \widetilde{M} M and N ~ \widetilde{N} N with respect to A 1 A_1 A 1 and B 1 B_1 B 1 as M ~ = ∑ i = 0 d − 1 M i ( ⟨ i ∣ A 1 ⊗ I A 2 ⊗ I C A ) \widetilde{M} = \sum_{i=0}^{d-1} M_i(\langle i|_{A_1} \otimes I_{A_2} \otimes I_{C_A}) M = ∑ i = 0 d − 1 M i (⟨ i ∣ A 1 ⊗ I A 2 ⊗ I C A ) and N ~ = ∑ j = 0 d − 1 N j ( ⟨ j ∣ B 1 ⊗ I B 2 ⊗ I C B ) \widetilde{N} = \sum_{j=0}^{d-1} N_j(\langle j|_{B_1} \otimes I_{B_2} \otimes I_{C_B}) N = ∑ j = 0 d − 1 N j (⟨ j ∣ B 1 ⊗ I B 2 ⊗ I C B ) . Here, I C A I_{C_A} I C A and I C B I_{C_B} I C B denote the identity operators on S A χ \mathcal{S}_{A}^{\chi} S A χ and S B χ \mathcal{S}_{B}^{\chi} S B χ , respectively. Hence every M i M_i M i and N j N_j N j is represented by a d R χ × d R χ dR_\chi\times dR_\chi d R χ × d R χ matrix. For a bipartite vector ∣ v ⟩ = ∑ a , b = 1 d R χ V a b ∣ a ⟩ L ∣ b ⟩ R |v\rangle= \sum_{a,b=1}^{dR_\chi} V_{ab}|a\rangle_L|b\rangle_R ∣ v ⟩ = ∑ a , b = 1 d R χ V ab ∣ a ⟩ L ∣ b ⟩ R , define its coefficient matrix as V : = ( V a b ) ∈ M d R χ ( C ) V := (V_{ab}) \in M_{dR_\chi}(\mathbb{C}) V := ( V ab ) ∈ M d R χ ( C ) . If local operators A A A and B B B act on the two subsystems, respectively, then the coefficient matrix of ( A ⊗ B ) ∣ v ⟩ (A \otimes B)|v\rangle ( A ⊗ B ) ∣ v ⟩ is A V B T AVB^T A V B T . For X ∈ M d ( C ) X \in M_d(\mathbb{C}) X ∈ M d ( C ) , define ∣ ϕ X ⟩ j : = ( X ⊗ I B j ) ∣ Φ d ⟩ j |\phi_X\rangle_j := (X \otimes I_{B_j})|\Phi_d\rangle_j ∣ ϕ X ⟩ j := ( X ⊗ I B j ) ∣ Φ d ⟩ j , j = 1 , 2 j=1,2 j = 1 , 2 . The coefficient matrix of ∣ Φ d ⟩ o u t ⊗ ∣ χ ⟩ |\Phi_d\rangle_{\mathrm{out}} \otimes|\chi\rangle ∣ Φ d ⟩ out ⊗ ∣ χ ⟩ across A ′ C A ∣ B ′ C B A'C_A|B'C_B A ′ C A ∣ B ′ C B is Q χ / d Q_\chi/\sqrt{d} Q χ / d . The two support constraints in Eq. (32) therefore imply the existence of linear functionals α , β : M d ( C ) → C \alpha,\beta: M_d(\mathbb{C}) \to\mathbb{C} α , β : M d ( C ) → C such that
∑ i M i ( X ⊗ D χ ) N i T = α ( X ) Q χ (34) \sum_i M_i(X \otimes D_\chi)N_i^T = \alpha(X)Q_\chi \tag*{(34)} i ∑ M i ( X ⊗ D χ ) N i T = α ( X ) Q χ (34)
∑ r , s X r s M r Q χ N s T = β ( X ) Q χ (35) \sum_{r,s} X_{rs}M_rQ_\chi N_s^T = \beta(X)Q_\chi \tag*{(35)} r , s ∑ X r s M r Q χ N s T = β ( X ) Q χ (35)
for every X ∈ M d ( C ) X \in M_d(\mathbb{C}) X ∈ M d ( C ) , where X r s : = ⟨ r ∣ X ∣ s ⟩ X_{rs} := \langle r|X|s\rangle X r s := ⟨ r ∣ X ∣ s ⟩ and T T T denotes transposition in the fixed local bases. For X = I d X=I_d X = I d , the left-hand sides of (44) and (35) coincide and are the coefficient matrix of the nonzero vector in (31) . Hence α ( I d ) = β ( I d ) ≠ 0 \alpha(I_d)=\beta(I_d)\ne0 α ( I d ) = β ( I d ) = 0 .
Second, we show that the matrix constraints obtained above are incompatible with the nonzero condition.
Let E r s : = ∣ r ⟩ ⟨ s ∣ E_{rs}:=|r\rangle\langle s| E r s := ∣ r ⟩ ⟨ s ∣ denote the standard matrix units and set c r s : = β ( E r s ) c_{rs}:=\beta(E_{rs}) c r s := β ( E r s ) . Taking X = E r s X=E_{rs} X = E r s in (45) gives
M r Q χ N s T = c r s Q χ (36) M_rQ_\chi N_s^T = c_{rs}Q_\chi \tag*{(36)} M r Q χ N s T = c r s Q χ (36)
Since ∑ i c i i = β ( I d ) ≠ 0 \sum_i c_{ii}=\beta(I_d)\ne0 ∑ i c ii = β ( I d ) = 0 , there exists an index t t t such that c t t ≠ 0 c_{tt}\ne0 c tt = 0 . Equation (36) then gives M t Q χ N t T = c t t Q χ M_tQ_\chi N_t^T=c_{tt}Q_\chi M t Q χ N t T = c tt Q χ . Because Q χ Q_\chi Q χ is invertible, both M t M_t M t and N t N_t N t are invertible. Setting s = t s=t s = t and r = t r=t r = t , respectively, in Eq. (36) yields
M i = a i M t , N i = b i N t (37) M_i=a_iM_t,\qquad N_i=b_iN_t \tag*{(37)} M i = a i M t , N i = b i N t (37)
for some scalars a i , b i ∈ C a_i,b_i\in\mathbb{C} a i , b i ∈ C . Substituting (47) into (44) gives
γ M t ( X ⊗ D χ ) N t T = α ( X ) Q χ , γ : = ∑ i a i b i . (38) \gamma M_t(X\otimes D_\chi)N_t^T=\alpha(X)Q_\chi,\qquad\gamma:=\sum_i a_i b_i. \tag*{(38)} γ M t ( X ⊗ D χ ) N t T = α ( X ) Q χ , γ := i ∑ a i b i . (38)
Taking X = I d X=I_d X = I d gives γ c t t = α ( I d ) ≠ 0 \gamma c_{tt}=\alpha(I_d)\ne0 γ c tt = α ( I d ) = 0 . Moreover, N t T = c t t Q χ − 1 M t − 1 Q χ N_t^T=c_{tt}Q_\chi^{-1}M_t^{-1}Q_\chi N t T = c tt Q χ − 1 M t − 1 Q χ . Using (48) and Q χ − 1 ( X ⊗ I R χ ) Q χ = X ⊗ I R χ Q_\chi^{-1}(X\otimes I_{R_\chi})Q_\chi=X\otimes I_{R_\chi} Q χ − 1 ( X ⊗ I R χ ) Q χ = X ⊗ I R χ gives γ c t t M t ( X ⊗ I R χ ) M t − 1 Q χ = α ( X ) Q χ \gamma c_{tt}M_t(X\otimes I_{R_\chi})M_t^{-1}Q_\chi=\alpha(X)Q_\chi γ c tt M t ( X ⊗ I R χ ) M t − 1 Q χ = α ( X ) Q χ . Right-multiplying by Q χ − 1 Q_\chi^{-1} Q χ − 1 , then left-multiplying by M t − 1 M_t^{-1} M t − 1 and right-multiplying by M t M_t M t , yields
X ⊗ I R χ = α ( X ) γ c t t I d R χ (39) X\otimes I_{R_\chi}=\frac{\alpha(X)}{\gamma c_{tt}}I_{dR_\chi} \tag*{(39)} X ⊗ I R χ = γ c tt α ( X ) I d R χ (39)
for every X ∈ M d ( C ) X\in M_d(\mathbb{C}) X ∈ M d ( C ) . Choosing X = ∣ 0 ⟩ ⟨ 1 ∣ X=|0\rangle\langle1| X = ∣0 ⟩ ⟨ 1∣ contradicts (49) for d ≥ 2 d\ge2 d ≥ 2 . ▪
Using the lemma and the constructions in the proofs of Theorems 1 and 2, we obtain the following theorem.
Theorem 3. Universal EPPs remain impossible under LOCC on the full bipartite and multipartite entangled-state sets considered in Theorems 1 and 2, even with the assistance of pure catalyst.
Proof. (a) We first consider the bipartite case with d 1 = d 2 d_1=d_2 d 1 = d 2 . Assume, for contradiction, that there exist a d d d -dimensional pure catalyst ∣ χ ⟩ C A C B |\chi\rangle_{C_A C_B} ∣ χ ⟩ C A C B and an n n n -to-1 LOCC EPP assisted by this catalyst that is universal for S d \mathcal{S}_d S d . For each input pair A j B j A_jB_j A j B j , define τ j = p Φ d ( j ) + ( 1 − p ) I A j B j / d 2 \tau_j=p\Phi_d^{(j)}+(1-p)I_{A_jB_j}/d^2 τ j = p Φ d ( j ) + ( 1 − p ) I A j B j / d 2 , where 1 / ( d + 1 ) < p < 1 1/(d+1)<p<1 1/ ( d + 1 ) < p < 1 . As in the proof of Theorem 1, every τ j \tau_j τ j is full rank and entangled. Consider Ω 0 : = ⨂ j = 1 n Φ d ( j ) \Omega_0:=\bigotimes_{j=1}^n\Phi_d^{(j)} Ω 0 := ⨂ j = 1 n Φ d ( j ) and Ω k : = Φ d ( k ) ⊗ ⨂ j ≠ k τ j \Omega_k:=\Phi_d^{(k)}\otimes\bigotimes_{j\ne k}\tau_j Ω k := Φ d ( k ) ⊗ ⨂ j = k τ j for k = 1 , … , n k=1,\ldots,n k = 1 , … , n . Since each Ω k \Omega_k Ω k contains a factor Φ d \Phi_d Φ d , universality and exact catalyst return require
Γ 0 ( Ω k ⊗ χ ) = q k Φ d o u t ⊗ χ , q k > 0 , k = 0 , … , n (40) \Gamma_0(\Omega_k\otimes\chi)=q_k\Phi_d^{\mathrm{out}}\otimes\chi,\qquad q_k>0,\quad k=0,\ldots,n \tag*{(40)} Γ 0 ( Ω k ⊗ χ ) = q k Φ d out ⊗ χ , q k > 0 , k = 0 , … , n (40) Write the successful branch in a product Kraus representation as Γ 0 ( ρ ) = ∑ μ K μ ρ K μ † \Gamma_0(\rho)=\sum_\mu K_\mu\rho K_\mu^\dagger Γ 0 ( ρ ) = ∑ μ K μ ρ K μ † , where K μ = M μ ⊗ N μ K_\mu=M_\mu\otimes N_\mu K μ = M μ ⊗ N μ , and define P c a t ⊥ : = I A ′ B ′ C A C B − Φ d o u t ⊗ χ P_{\mathrm{cat}}^\perp:=I_{A'B'C_A C_B}-\Phi_d^{\mathrm{out}}\otimes\chi P cat ⊥ := I A ′ B ′ C A C B − Φ d out ⊗ χ . For k = 1 , … , n k=1,\ldots,n k = 1 , … , n , Eq. (40) gives
P c a t ⊥ Γ 0 ( Ω k ⊗ χ ) P c a t ⊥ = ∑ μ P c a t ⊥ K μ ( Ω k ⊗ χ ) K μ † P c a t ⊥ = 0. (41) P_{\mathrm{cat}}^\perp\Gamma_0(\Omega_k\otimes\chi)P_{\mathrm{cat}}^\perp
=\sum_\mu P_{\mathrm{cat}}^\perp K_\mu(\Omega_k\otimes\chi)K_\mu^\dagger P_{\mathrm{cat}}^\perp=0. \tag*{(41)} P cat ⊥ Γ 0 ( Ω k ⊗ χ ) P cat ⊥ = μ ∑ P cat ⊥ K μ ( Ω k ⊗ χ ) K μ † P cat ⊥ = 0. (41) Since every summand in Eq. (41) is positive semidefinite and their sum is zero, each summand must vanish. As χ \chi χ is a pure state, it follows that P c a t ⊥ K μ ( Ω k 1 / 2 ⊗ χ ) = 0 P_{\mathrm{cat}}^\perp K_\mu(\Omega_k^{1/2}\otimes\chi)=0 P cat ⊥ K μ ( Ω k 1/2 ⊗ χ ) = 0 for every μ \mu μ and k = 1 , … , n k=1,\ldots,n k = 1 , … , n . Because every τ j \tau_j τ j is full rank, we obtain supp Ω k = \operatorname{supp}\Omega_k= supp Ω k =
@page 7/11 starts_mid=1 ends_mid=1 and the two support constraints
K ~ ( ∣ Φ d ⟩ 1 ⊗ ∣ ϕ ⟩ 2 ⊗ ∣ χ ⟩ ) ∈ L c a t , \widetilde{K}(|\Phi_d\rangle_1 \otimes|\phi\rangle_2 \otimes|\chi\rangle) \in\mathcal{L}_{\mathrm{cat}}, K ( ∣ Φ d ⟩ 1 ⊗ ∣ ϕ ⟩ 2 ⊗ ∣ χ ⟩) ∈ L cat , K ~ ( ∣ ϕ ⟩ 1 ⊗ ∣ Φ d ⟩ 2 ⊗ ∣ χ ⟩ ) ∈ L c a t (42) \widetilde{K}(|\phi\rangle_1 \otimes|\Phi_d\rangle_2 \otimes|\chi\rangle) \in\mathcal{L}_{\mathrm{cat}} \tag*{(42)} K ( ∣ ϕ ⟩ 1 ⊗ ∣ Φ d ⟩ 2 ⊗ ∣ χ ⟩) ∈ L cat (42) for every ∣ ϕ ⟩ j ∈ H A j ⊗ H B j |\phi\rangle_j \in\mathcal{H}_{A_j} \otimes\mathcal{H}_{B_j} ∣ ϕ ⟩ j ∈ H A j ⊗ H B j , j = 1 , 2 j=1,2 j = 1 , 2 .
Proof. First, we restrict the catalyst state to its Schmidt-support subspaces and translate the two support constraints into matrix equations.
Let the Schmidt decomposition of the catalyst be ∣ χ ⟩ = ∑ a = 1 R χ λ a ∣ a ⟩ C A ∣ a ⟩ C B |\chi\rangle= \sum_{a=1}^{R_\chi} \sqrt{\lambda_a}|a\rangle_{C_A}|a\rangle_{C_B} ∣ χ ⟩ = ∑ a = 1 R χ λ a ∣ a ⟩ C A ∣ a ⟩ C B , where R χ R_\chi R χ is its Schmidt rank. Define S A χ : = span { ∣ a ⟩ C A : 1 ≤ a ≤ R χ } \mathcal{S}_{A}^{\chi} := \operatorname{span}\{|a\rangle_{C_A}: 1 \le a \le R_\chi\} S A χ := span { ∣ a ⟩ C A : 1 ≤ a ≤ R χ } and S B χ : = span { ∣ a ⟩ C B : 1 ≤ a ≤ R χ } \mathcal{S}_{B}^{\chi} := \operatorname{span}\{|a\rangle_{C_B}: 1 \le a \le R_\chi\} S B χ := span { ∣ a ⟩ C B : 1 ≤ a ≤ R χ } , and let P A χ P_A^\chi P A χ and P B χ P_B^\chi P B χ denote the orthogonal projectors onto S A χ \mathcal{S}_{A}^{\chi} S A χ and S B χ \mathcal{S}_{B}^{\chi} S B χ , respectively. Define
M ‾ : = ( I A ′ ⊗ P A χ ) M ~ ( I A 1 ⊗ I A 2 ⊗ P A χ ) , \overline{M} := (I_{A'} \otimes P_A^\chi)\widetilde{M}(I_{A_1} \otimes I_{A_2} \otimes P_A^\chi), M := ( I A ′ ⊗ P A χ ) M ( I A 1 ⊗ I A 2 ⊗ P A χ ) , N ‾ : = ( I B ′ ⊗ P B χ ) N ~ ( I B 1 ⊗ I B 2 ⊗ P B χ ) (43) \overline{N} := (I_{B'} \otimes P_B^\chi)\widetilde{N}(I_{B_1} \otimes I_{B_2} \otimes P_B^\chi) \tag*{(43)} N := ( I B ′ ⊗ P B χ ) N ( I B 1 ⊗ I B 2 ⊗ P B χ ) (43) and set K ‾ : = M ‾ ⊗ N ‾ \overline{K} := \overline{M} \otimes\overline{N} K := M ⊗ N . The input projections I A 1 ⊗ I A 2 ⊗ P A χ I_{A_1} \otimes I_{A_2} \otimes P_A^\chi I A 1 ⊗ I A 2 ⊗ P A χ and I B 1 ⊗ I B 2 ⊗ P B χ I_{B_1} \otimes I_{B_2} \otimes P_B^\chi I B 1 ⊗ I B 2 ⊗ P B χ leave the catalyst unchanged because ( P A χ ⊗ P B χ ) ∣ χ ⟩ = ∣ χ ⟩ (P_A^\chi\otimes P_B^\chi)|\chi\rangle= |\chi\rangle ( P A χ ⊗ P B χ ) ∣ χ ⟩ = ∣ χ ⟩ . Moreover, taking ∣ ϕ ⟩ 2 = ∣ Φ d ⟩ 2 |\phi\rangle_2 = |\Phi_d\rangle_2 ∣ ϕ ⟩ 2 = ∣ Φ d ⟩ 2 in the first support condition of Eq. (32) and using Eq. (31), we find that the vector in Eq. (31) is a nonzero element of L c a t \mathcal{L}_{\mathrm{cat}} L cat . The output projections I A ′ ⊗ P A χ I_{A'} \otimes P_A^\chi I A ′ ⊗ P A χ and I B ′ ⊗ P B χ I_{B'} \otimes P_B^\chi I B ′ ⊗ P B χ therefore leave unchanged all vectors on the left-hand sides of Eqs. (31) and (32). Consequently, Eqs. (31) and (32) remain valid with K ~ \widetilde{K} K replaced by K K K . We may therefore replace H C A \mathcal{H}_{C_A} H C A by S A χ \mathcal{S}_{A}^{\chi} S A χ and H C B \mathcal{H}_{C_B} H C B by S B χ \mathcal{S}_{B}^{\chi} S B χ . For notational simplicity, we relabel M ‾ \overline{M} M and N ‾ \overline{N} N as M ~ \widetilde{M} M and N ~ \widetilde{N} N , respectively. Define D χ : = diag ( λ 1 , … , λ R χ ) D_\chi:= \operatorname{diag}(\sqrt{\lambda_1},\ldots,\sqrt{\lambda_{R_\chi}}) D χ := diag ( λ 1 , … , λ R χ ) and Q χ : = I d ⊗ D χ Q_\chi:= I_d \otimes D_\chi Q χ := I d ⊗ D χ . Since λ a > 0 \lambda_a > 0 λ a > 0 for all a = 1 , … , R χ a=1,\ldots,R_\chi a = 1 , … , R χ , both D χ D_\chi D χ and Q χ Q_\chi Q χ are invertible. Decompose M ~ \widetilde{M} M and N ~ \widetilde{N} N with respect to A 1 A_1 A 1 and B 1 B_1 B 1 as M ~ = ∑ i = 0 d − 1 M i ( ⟨ i ∣ A 1 ⊗ I A 2 ⊗ I C A ) \widetilde{M} = \sum_{i=0}^{d-1} M_i(\langle i|_{A_1} \otimes I_{A_2} \otimes I_{C_A}) M = ∑ i = 0 d − 1 M i (⟨ i ∣ A 1 ⊗ I A 2 ⊗ I C A ) and N ~ = ∑ j = 0 d − 1 N j ( ⟨ j ∣ B 1 ⊗ I B 2 ⊗ I C B ) \widetilde{N} = \sum_{j=0}^{d-1} N_j(\langle j|_{B_1} \otimes I_{B_2} \otimes I_{C_B}) N = ∑ j = 0 d − 1 N j (⟨ j ∣ B 1 ⊗ I B 2 ⊗ I C B ) . Here, I C A I_{C_A} I C A and I C B I_{C_B} I C B denote the identity operators on S A χ \mathcal{S}_{A}^{\chi} S A χ and S B χ \mathcal{S}_{B}^{\chi} S B χ , respectively. Hence every M i M_i M i and N j N_j N j is represented by a d R χ × d R χ dR_\chi\times dR_\chi d R χ × d R χ matrix. For a bipartite vector ∣ v ⟩ = ∑ a , b = 1 d R χ V a b ∣ a ⟩ L ∣ b ⟩ R |v\rangle= \sum_{a,b=1}^{dR_\chi} V_{ab}|a\rangle_L|b\rangle_R ∣ v ⟩ = ∑ a , b = 1 d R χ V ab ∣ a ⟩ L ∣ b ⟩ R , define its coefficient matrix as V : = ( V a b ) ∈ M d R χ ( C ) V := (V_{ab}) \in M_{dR_\chi}(\mathbb{C}) V := ( V ab ) ∈ M d R χ ( C ) . If local operators A A A and B B B act on the two subsystems, respectively, then the coefficient matrix of ( A ⊗ B ) ∣ v ⟩ (A \otimes B)|v\rangle ( A ⊗ B ) ∣ v ⟩ is A V B T AVB^T A V B T . For X ∈ M d ( C ) X \in M_d(\mathbb{C}) X ∈ M d ( C ) , define ∣ ϕ X ⟩ j : = ( X ⊗ I B j ) ∣ Φ d ⟩ j |\phi_X\rangle_j := (X \otimes I_{B_j})|\Phi_d\rangle_j ∣ ϕ X ⟩ j := ( X ⊗ I B j ) ∣ Φ d ⟩ j , j = 1 , 2 j=1,2 j = 1 , 2 . The coefficient matrix of ∣ Φ d ⟩ o u t ⊗ ∣ χ ⟩ |\Phi_d\rangle_{\mathrm{out}} \otimes|\chi\rangle ∣ Φ d ⟩ out ⊗ ∣ χ ⟩ across A ′ C A ∣ B ′ C B A'C_A|B'C_B A ′ C A ∣ B ′ C B is Q χ / d Q_\chi/\sqrt{d} Q χ / d . The two support constraints in Eq. (32) therefore imply the existence of linear functionals α , β : M d ( C ) → C \alpha,\beta: M_d(\mathbb{C}) \to\mathbb{C} α , β : M d ( C ) → C such that
∑ i M i ( X ⊗ D χ ) N i T = α ( X ) Q χ (44) \sum_i M_i(X \otimes D_\chi)N_i^T = \alpha(X)Q_\chi \tag*{(44)} i ∑ M i ( X ⊗ D χ ) N i T = α ( X ) Q χ (44) ∑ r , s X r s M r Q χ N s T = β ( X ) Q χ (45) \sum_{r,s} X_{rs}M_rQ_\chi N_s^T = \beta(X)Q_\chi \tag*{(45)} r , s ∑ X r s M r Q χ N s T = β ( X ) Q χ (45) for every X ∈ M d ( C ) X \in M_d(\mathbb{C}) X ∈ M d ( C ) , where X r s : = ⟨ r ∣ X ∣ s ⟩ X_{rs} := \langle r|X|s\rangle X r s := ⟨ r ∣ X ∣ s ⟩ and T T T denotes transposition in the fixed local bases. For X = I d X=I_d X = I d , the left-hand sides of Eqs. (34) and (35) coincide and are the coefficient matrix of the nonzero vector in Eq. (31). Hence α ( I d ) = β ( I d ) ≠ 0 \alpha(I_d)=\beta(I_d)\ne0 α ( I d ) = β ( I d ) = 0 .
Second, we show that the matrix constraints obtained above are incompatible with the nonzero condition.
Let E r s : = ∣ r ⟩ ⟨ s ∣ E_{rs}:=|r\rangle\langle s| E r s := ∣ r ⟩ ⟨ s ∣ denote the standard matrix units and set c r s : = β ( E r s ) c_{rs}:=\beta(E_{rs}) c r s := β ( E r s ) . Taking X = E r s X=E_{rs} X = E r s in Eq. (35) gives
M r Q χ N s T = c r s Q χ (46) M_rQ_\chi N_s^T = c_{rs}Q_\chi \tag*{(46)} M r Q χ N s T = c r s Q χ (46) Since ∑ i c i i = β ( I d ) ≠ 0 \sum_i c_{ii}=\beta(I_d)\ne0 ∑ i c ii = β ( I d ) = 0 , there exists an index t t t such that c t t ≠ 0 c_{tt}\ne0 c tt = 0 . Equation (36) then gives M t Q χ N t T = c t t Q χ M_tQ_\chi N_t^T=c_{tt}Q_\chi M t Q χ N t T = c tt Q χ . Because Q χ Q_\chi Q χ is invertible, both M t M_t M t and N t N_t N t are invertible. Setting s = t s=t s = t and r = t r=t r = t , respectively, in Eq. (36) yields
M i = a i M t , N i = b i N t (47) M_i=a_iM_t,\qquad N_i=b_iN_t \tag*{(47)} M i = a i M t , N i = b i N t (47) for some scalars a i , b i ∈ C a_i,b_i\in\mathbb{C} a i , b i ∈ C . Substituting Eq. (37) into Eq. (34) gives
γ M t ( X ⊗ D χ ) N t T = α ( X ) Q χ , γ : = ∑ i a i b i . (48) \gamma M_t(X\otimes D_\chi)N_t^T=\alpha(X)Q_\chi,\qquad\gamma:=\sum_i a_i b_i. \tag*{(48)} γ M t ( X ⊗ D χ ) N t T = α ( X ) Q χ , γ := i ∑ a i b i . (48) Taking X = I d X=I_d X = I d gives γ c t t = α ( I d ) ≠ 0 \gamma c_{tt}=\alpha(I_d)\ne0 γ c tt = α ( I d ) = 0 . Moreover, N t T = c t t Q χ − 1 M t − 1 Q χ N_t^T=c_{tt}Q_\chi^{-1}M_t^{-1}Q_\chi N t T = c tt Q χ − 1 M t − 1 Q χ . Using Eq. (38) and Q χ − 1 ( X ⊗ I R χ ) Q χ = X ⊗ I R χ Q_\chi^{-1}(X\otimes I_{R_\chi})Q_\chi=X\otimes I_{R_\chi} Q χ − 1 ( X ⊗ I R χ ) Q χ = X ⊗ I R χ gives γ c t t M t ( X ⊗ I R χ ) M t − 1 Q χ = α ( X ) Q χ \gamma c_{tt}M_t(X\otimes I_{R_\chi})M_t^{-1}Q_\chi=\alpha(X)Q_\chi γ c tt M t ( X ⊗ I R χ ) M t − 1 Q χ = α ( X ) Q χ . Right-multiplying by Q χ − 1 Q_\chi^{-1} Q χ − 1 , then left-multiplying by M t − 1 M_t^{-1} M t − 1 and right-multiplying by M t M_t M t , yields
X ⊗ I R χ = α ( X ) γ c t t I d R χ (49) X\otimes I_{R_\chi}=\frac{\alpha(X)}{\gamma c_{tt}}I_{dR_\chi} \tag*{(49)} X ⊗ I R χ = γ c tt α ( X ) I d R χ (49) for every X ∈ M d ( C ) X\in M_d(\mathbb{C}) X ∈ M d ( C ) . Choosing X = ∣ 0 ⟩ ⟨ 1 ∣ X=|0\rangle\langle1| X = ∣0 ⟩ ⟨ 1∣ contradicts Eq. (39) for d ≥ 2 d\ge2 d ≥ 2 . ▪
Using the lemma and the constructions in the proofs of Theorems 1 and 2, we obtain the following theorem.
Theorem 3. Universal EPPs remain impossible under LOCC on the full bipartite and multipartite entangled-state sets considered in Theorems 1 and 2, even with the assistance of pure catalyst.
Proof. (a) We first consider the bipartite case with d 1 = d 2 d_1=d_2 d 1 = d 2 . Assume, for contradiction, that there exist a d d d -dimensional pure catalyst ∣ χ ⟩ C A C B |\chi\rangle_{C_A C_B} ∣ χ ⟩ C A C B and an n n n -to-1 LOCC EPP assisted by this catalyst that is universal for S d \mathcal{S}_d S d . For each input pair A j B j A_jB_j A j B j , define τ j = p Φ d ( j ) + ( 1 − p ) I A j B j / d 2 \tau_j=p\Phi_d^{(j)}+(1-p)I_{A_jB_j}/d^2 τ j = p Φ d ( j ) + ( 1 − p ) I A j B j / d 2 , where 1 / ( d + 1 ) < p < 1 1/(d+1)<p<1 1/ ( d + 1 ) < p < 1 . As in the proof of Theorem 1, every τ j \tau_j τ j is full rank and entangled. Consider Ω 0 : = ⨂ j = 1 n Φ d ( j ) \Omega_0:=\bigotimes_{j=1}^n\Phi_d^{(j)} Ω 0 := ⨂ j = 1 n Φ d ( j ) and Ω k : = Φ d ( k ) ⊗ ⨂ j ≠ k τ j \Omega_k:=\Phi_d^{(k)}\otimes\bigotimes_{j\ne k}\tau_j Ω k := Φ d ( k ) ⊗ ⨂ j = k τ j for k = 1 , … , n k=1,\ldots,n k = 1 , … , n . Since each Ω k \Omega_k Ω k contains a factor Φ d \Phi_d Φ d , universality and exact catalyst return require
Γ 0 ( Ω k ⊗ χ ) = q k Φ d o u t ⊗ χ , q k > 0 , k = 0 , … , n (50) \Gamma_0(\Omega_k\otimes\chi)=q_k\Phi_d^{\mathrm{out}}\otimes\chi,\qquad q_k>0,\quad k=0,\ldots,n \tag*{(50)} Γ 0 ( Ω k ⊗ χ ) = q k Φ d out ⊗ χ , q k > 0 , k = 0 , … , n (50) Write the successful branch in a product Kraus representation as Γ 0 ( ρ ) = ∑ μ K μ ρ K μ † \Gamma_0(\rho)=\sum_\mu K_\mu\rho K_\mu^\dagger Γ 0 ( ρ ) = ∑ μ K μ ρ K μ † , where K μ = M μ ⊗ N μ K_\mu=M_\mu\otimes N_\mu K μ = M μ ⊗ N μ , and define P c a t ⊥ : = I A ′ B ′ C A C B − Φ d o u t ⊗ χ P_{\mathrm{cat}}^\perp:=I_{A'B'C_A C_B}-\Phi_d^{\mathrm{out}}\otimes\chi P cat ⊥ := I A ′ B ′ C A C B − Φ d out ⊗ χ . For k = 1 , … , n k=1,\ldots,n k = 1 , … , n , Eq. (40) gives
P c a t ⊥ Γ 0 ( Ω k ⊗ χ ) P c a t ⊥ = ∑ μ P c a t ⊥ K μ ( Ω k ⊗ χ ) K μ † P c a t ⊥ = 0. (51) P_{\mathrm{cat}}^\perp\Gamma_0(\Omega_k\otimes\chi)P_{\mathrm{cat}}^\perp
=\sum_\mu P_{\mathrm{cat}}^\perp K_\mu(\Omega_k\otimes\chi)K_\mu^\dagger P_{\mathrm{cat}}^\perp=0. \tag*{(51)} P cat ⊥ Γ 0 ( Ω k ⊗ χ ) P cat ⊥ = μ ∑ P cat ⊥ K μ ( Ω k ⊗ χ ) K μ † P cat ⊥ = 0. (51) Since every summand in Eq. (41) is positive semidefinite and their sum is zero, each summand must vanish. As χ \chi χ is a pure state, it follows that P c a t ⊥ K μ ( Ω k 1 / 2 ⊗ χ ) = 0 P_{\mathrm{cat}}^\perp K_\mu(\Omega_k^{1/2}\otimes\chi)=0 P cat ⊥ K μ ( Ω k 1/2 ⊗ χ ) = 0 for every μ \mu μ and k = 1 , … , n k=1,\ldots,n k = 1 , … , n . Because every τ j \tau_j τ j is full rank, we obtain supp Ω k = \operatorname{supp}\Omega_k= supp Ω k =
span{ ∣ Φ d ⟩ k } ⊗ H k i n \{|\Phi_d\rangle_k\}\otimes\mathcal{H}^{\mathrm{in}}_k { ∣ Φ d ⟩ k } ⊗ H k in , where H k i n : = ⨂ j ≠ k ( H A j ⊗ H B j ) \mathcal{H}^{\mathrm{in}}_k:=\bigotimes_{j\ne k}(\mathcal{H}_{A_j}\otimes\mathcal{H}_{B_j}) H k in := ⨂ j = k ( H A j ⊗ H B j ) . It follows that
K μ ( ∣ Φ d ⟩ k ⊗ ∣ ψ ⟩ k ˉ ⊗ ∣ χ ⟩ ) ∈ L c a t (52) K_\mu\left(|\Phi_d\rangle_k\otimes|\psi\rangle_{\bar{k}}\otimes|\chi\rangle\right)\in\mathcal{L}_{\mathrm{cat}} \tag*{(52)} K μ ( ∣ Φ d ⟩ k ⊗ ∣ ψ ⟩ k ˉ ⊗ ∣ χ ⟩ ) ∈ L cat (52)
for every μ \mu μ , every k = 1 , … , n k=1,\ldots,n k = 1 , … , n , and every ∣ ψ ⟩ k ˉ ∈ H k i n |\psi\rangle_{\bar{k}}\in\mathcal{H}^{\mathrm{in}}_k ∣ ψ ⟩ k ˉ ∈ H k in . The condition q 0 > 0 q_0>0 q 0 > 0 ensures that at least one product Kraus operator, denoted by K = M ⊗ N K=M\otimes N K = M ⊗ N , acts nontrivially on ⨂ j = 1 n ∣ Φ d ⟩ j ⊗ ∣ χ ⟩ \bigotimes_{j=1}^{n}|\Phi_d\rangle_j\otimes|\chi\rangle ⨂ j = 1 n ∣ Φ d ⟩ j ⊗ ∣ χ ⟩ . We next reduce K K K to a product operator acting only on the first two input pairs and the catalyst, as required to apply the lemma. For n = 2 n=2 n = 2 , we simply set K ~ = K \widetilde{K}=K K = K , and no reduction is required. For n > 2 n>2 n > 2 , let R = { 3 , … , n } \mathcal{R}=\{3,\ldots,n\} R = { 3 , … , n } denote the set of indices of the remaining input pairs. For a multi-index z = ( z 3 , … , z n ) ∈ { 0 , … , d − 1 } n − 2 \boldsymbol{z}=(z_3,\ldots,z_n)\in\{0,\ldots,d-1\}^{n-2} z = ( z 3 , … , z n ) ∈ { 0 , … , d − 1 } n − 2 , define ∣ z ⟩ A R : = ⨂ j = 3 n ∣ z j ⟩ A j |\boldsymbol{z}\rangle_{A_{\mathcal{R}}}:=\bigotimes_{j=3}^{n}|z_j\rangle_{A_j} ∣ z ⟩ A R := ⨂ j = 3 n ∣ z j ⟩ A j and ∣ z ⟩ B R : = ⨂ j = 3 n ∣ z j ⟩ B j |\boldsymbol{z}\rangle_{B_{\mathcal{R}}}:=\bigotimes_{j=3}^{n}|z_j\rangle_{B_j} ∣ z ⟩ B R := ⨂ j = 3 n ∣ z j ⟩ B j . The product of the target states on the remaining input pairs can be expanded as ⨂ j = 3 n ∣ Φ d ⟩ j = d − ( n − 2 ) / 2 ∑ z ∣ z ⟩ A R ∣ z ⟩ B R \bigotimes_{j=3}^{n}|\Phi_d\rangle_j=d^{-(n-2)/2}\sum_{\boldsymbol{z}}|\boldsymbol{z}\rangle_{A_{\mathcal{R}}}|\boldsymbol{z}\rangle_{B_{\mathcal{R}}} ⨂ j = 3 n ∣ Φ d ⟩ j = d − ( n − 2 ) /2 ∑ z ∣ z ⟩ A R ∣ z ⟩ B R . Expanding the remaining target states in the nonzero vector K ( ⨂ j = 1 n ∣ Φ d ⟩ j ⊗ ∣ χ ⟩ ) K(\bigotimes_{j=1}^{n}|\Phi_d\rangle_j\otimes|\chi\rangle) K ( ⨂ j = 1 n ∣ Φ d ⟩ j ⊗ ∣ χ ⟩) gives a nonzero sum. Hence there exists at least one multi-index z 0 \boldsymbol{z}_0 z 0 for which the corresponding term is nonzero. Define M ~ \widetilde{M} M and N ~ \widetilde{N} N by fixing the inputs of M M M and N N N on the remaining input pairs to ∣ z 0 ⟩ A R |\boldsymbol{z}_0\rangle_{A_{\mathcal{R}}} ∣ z 0 ⟩ A R and ∣ z 0 ⟩ B R |\boldsymbol{z}_0\rangle_{B_{\mathcal{R}}} ∣ z 0 ⟩ B R , respectively, and set K ~ : = M ~ ⊗ N ~ \widetilde{K}:=\widetilde{M}\otimes\widetilde{N} K := M ⊗ N . The resulting product operator acts on the first two input pairs and the catalyst. In either case, K ~ \widetilde{K} K satisfies the nonzero condition in Eq. (31) , while Eq. (52) implies the two support constraints in Eq. (32). This contradicts the lemma. Hence no universal catalytic LOCC EPP exists for S d \mathcal{S}_d S d .
We next consider the d 1 , d 2 d_1,d_2 d 1 , d 2 case. Set d = min { d 1 , d 2 } d=\min\{d_1,d_2\} d = min { d 1 , d 2 } and, for each j = 1 , … , n j=1,\ldots,n j = 1 , … , n , let τ j = p Φ d ( j ) + ( 1 − p ) I A j B j / ( d 1 d 2 ) \tau_j=p\Phi_{d}^{(j)}+(1-p)I_{A_jB_j}/(d_1d_2) τ j = p Φ d ( j ) + ( 1 − p ) I A j B j / ( d 1 d 2 ) , where $1/(d+1)<p<1$. Let Ω 0 : = ⨂ j = 1 n Φ d ( j ) \Omega_0:=\bigotimes_{j=1}^{n}\Phi_d^{(j)} Ω 0 := ⨂ j = 1 n Φ d ( j ) and Ω k : = Φ d ( k ) ⊗ ⨂ j ≠ k τ j \Omega_k:=\Phi_d^{(k)}\otimes\bigotimes_{j\ne k}\tau_j Ω k := Φ d ( k ) ⊗ ⨂ j = k τ j for k = 1 , … , n k=1,\ldots,n k = 1 , … , n . Since 1 − p > 0 1-p>0 1 − p > 0 and τ j B j \tau_j^{B_j} τ j B j has an eigenvalue ( 1 − p ) / ( d 1 d 2 ) − p / d < 0 (1-p)/(d_1d_2)-p/d<0 ( 1 − p ) / ( d 1 d 2 ) − p / d < 0 , each τ j \tau_j τ j is full rank and entangled. Consequently, Ω 0 , … , Ω n \Omega_0,\ldots,\Omega_n Ω 0 , … , Ω n are allowed total input states. The positivity, Kraus-selection, and reduction arguments used above for the case d 1 = d 2 d_1=d_2 d 1 = d 2 yield a product operator K ~ = M ~ ⊗ N ~ \widetilde{K}=\widetilde{M}\otimes\widetilde{N} K = M ⊗ N acting on the first two input pairs and the catalyst such that K ~ ( ∣ Φ d ⟩ 1 ⊗ ∣ Φ d ⟩ 2 ⊗ ∣ χ ⟩ ) = c ∣ Φ d ⟩ o u t ⊗ ∣ χ ⟩ \widetilde{K}(|\Phi_d\rangle_1\otimes|\Phi_d\rangle_2\otimes|\chi\rangle)=c|\Phi_d\rangle_{\mathrm{out}}\otimes|\chi\rangle K ( ∣ Φ d ⟩ 1 ⊗ ∣ Φ d ⟩ 2 ⊗ ∣ χ ⟩) = c ∣ Φ d ⟩ out ⊗ ∣ χ ⟩ for some c ≠ 0 c\ne0 c = 0 . Moreover, K ~ ( ∣ Φ d ⟩ 1 ⊗ ∣ ψ ⟩ 2 ⊗ ∣ χ ⟩ ) ∈ L c a t \widetilde{K}(|\Phi_d\rangle_1\otimes|\psi\rangle_2\otimes|\chi\rangle)\in\mathcal{L}_{\mathrm{cat}} K ( ∣ Φ d ⟩ 1 ⊗ ∣ ψ ⟩ 2 ⊗ ∣ χ ⟩) ∈ L cat and K ~ ( ∣ ψ ⟩ 1 ⊗ ∣ Φ d ⟩ 2 ⊗ ∣ χ ⟩ ) ∈ L c a t \widetilde{K}(|\psi\rangle_1\otimes|\Phi_d\rangle_2\otimes|\chi\rangle)\in\mathcal{L}_{\mathrm{cat}} K ( ∣ ψ ⟩ 1 ⊗ ∣ Φ d ⟩ 2 ⊗ ∣ χ ⟩) ∈ L cat for every ∣ ψ ⟩ j ∈ H A j ⊗ H B j |\psi\rangle_j\in\mathcal{H}_{A_j}\otimes\mathcal{H}_{B_j} ∣ ψ ⟩ j ∈ H A j ⊗ H B j , j = 1 , 2 j=1,2 j = 1 , 2 . For j = 1 , 2 j=1,2 j = 1 , 2 , let K A j : = span { ∣ r ⟩ A j } r = 0 d − 1 \mathcal{K}_{A_j}:=\operatorname{span}\{|r\rangle_{A_j}\}_{r=0}^{d-1} K A j := span { ∣ r ⟩ A j } r = 0 d − 1 and K B j : = span { ∣ r ⟩ B j } r = 0 d − 1 \mathcal{K}_{B_j}:=\operatorname{span}\{|r\rangle_{B_j}\}_{r=0}^{d-1} K B j := span { ∣ r ⟩ B j } r = 0 d − 1 , and let P A j P_{A_j} P A j and P B j P_{B_j} P B j denote the projectors onto these subspaces, respectively. For the output subsystems A ′ A' A ′ and B ′ B' B ′ , let K A ′ : = span { ∣ r ⟩ A ′ } r = 0 d − 1 \mathcal{K}_{A'}:=\operatorname{span}\{|r\rangle_{A'}\}_{r=0}^{d-1} K A ′ := span { ∣ r ⟩ A ′ } r = 0 d − 1 and K B ′ : = span { ∣ r ⟩ B ′ } r = 0 d − 1 \mathcal{K}_{B'}:=\operatorname{span}\{|r\rangle_{B'}\}_{r=0}^{d-1} K B ′ := span { ∣ r ⟩ B ′ } r = 0 d − 1 and let P A ′ P_{A'} P A ′ and P B ′ P_{B'} P B ′ denote the projectors onto these subspaces, respectively. Define M ~ : = ( P A ′ ⊗ I C A ) M ~ ( P A 1 ⊗ P A 2 ⊗ I C A ) \widetilde{M}:=(P_{A'}\otimes I_{C_A})\widetilde{M}(P_{A_1}\otimes P_{A_2}\otimes I_{C_A}) M := ( P A ′ ⊗ I C A ) M ( P A 1 ⊗ P A 2 ⊗ I C A ) and N ~ : = ( P B ′ ⊗ I C B ) N ~ ( P B 1 ⊗ P B 2 ⊗ I C B ) \widetilde{N}:=(P_{B'}\otimes I_{C_B})\widetilde{N}(P_{B_1}\otimes P_{B_2}\otimes I_{C_B}) N := ( P B ′ ⊗ I C B ) N ( P B 1 ⊗ P B 2 ⊗ I C B ) . By construction, the input projections act as the identity on K A 1 ⊗ K A 2 ⊗ H C A \mathcal{K}_{A_1}\otimes\mathcal{K}_{A_2}\otimes\mathcal{H}_{C_A} K A 1 ⊗ K A 2 ⊗ H C A and K B 1 ⊗ K B 2 ⊗ H C B \mathcal{K}_{B_1}\otimes\mathcal{K}_{B_2}\otimes\mathcal{H}_{C_B} K B 1 ⊗ K B 2 ⊗ H C B , respectively. Since ∣ Φ d ⟩ o u t ∈ K A ′ ⊗ K B ′ |\Phi_d\rangle_{\mathrm{out}}\in\mathcal{K}_{A'}\otimes\mathcal{K}_{B'} ∣ Φ d ⟩ out ∈ K A ′ ⊗ K B ′ , the output projections ( P A ′ ⊗ I C A ) ( P B ′ ⊗ I C B ) (P_{A'}\otimes I_{C_A})(P_{B'}\otimes I_{C_B}) ( P A ′ ⊗ I C A ) ( P B ′ ⊗ I C B ) leave every vector in L c a t \mathcal{L}_{\mathrm{cat}} L cat unchanged. Consequently, K ~ : = M ~ ⊗ N ~ \widetilde{K}:=\widetilde{M}\otimes\widetilde{N} K := M ⊗ N is nonzero on ∣ Φ d ⟩ 1 ⊗ ∣ Φ d ⟩ 2 ⊗ ∣ χ ⟩ |\Phi_d\rangle_1\otimes|\Phi_d\rangle_2\otimes|\chi\rangle ∣ Φ d ⟩ 1 ⊗ ∣ Φ d ⟩ 2 ⊗ ∣ χ ⟩ and satisfies two support conditions of the form in Eq. (52) , with the arbitrary input vector restricted to K A j ⊗ K B j \mathcal{K}_{A_j}\otimes\mathcal{K}_{B_j} K A j ⊗ K B j for $j=1,2. Each of the subspaces K A 1 , K A 2 , K B 1 , K B 2 , K A ′ , K B ′ \mathcal{K}_{A_1},\mathcal{K}_{A_2},\mathcal{K}_{B_1},\mathcal{K}_{B_2},\mathcal{K}_{A'},\mathcal{K}_{B'} K A 1 , K A 2 , K B 1 , K B 2 , K A ′ , K B ′ is d d d -dimensional and may therefore be identified with C d \mathbb{C}^d C d . Under these identifications, K ~ \widetilde{K} K satisfies Eqs. (31) and (32). This contradicts the lemma and proves the bipartite statement for all finite d 1 , d 2 ≥ 2 d_1,d_2\ge2 d 1 , d 2 ≥ 2 .
(b) We then consider the multipartite case. Suppose, for contradiction, that a catalytic LOCC EPP is universal for G d \mathcal{G}_d G d . Let us consider the states τ j , Ω 0 \tau_j,\Omega_0 τ j , Ω 0 , and Ω k \Omega_k Ω k introduced in the proof of Theorem 2. Every τ j \tau_j τ j is full rank and genuinely multipartite entangled, so all Ω k \Omega_k Ω k , with k = 0 , … , n k=0,\ldots,n k = 0 , … , n , are allowed input states. Since each Ω k \Omega_k Ω k contains a factor Ψ d \Psi_d Ψ d , universality and exact catalyst return require
Γ 0 ( Ω k ⊗ χ ) = q k Ψ d o u t ⊗ χ , q k > 0 , k = 0 , … , n . (53) \Gamma_0(\Omega_k\otimes\chi)=q_k\Psi_d^{\mathrm{out}}\otimes\chi,\qquad q_k>0,\qquad k=0,\ldots,n. \tag*{(53)} Γ 0 ( Ω k ⊗ χ ) = q k Ψ d out ⊗ χ , q k > 0 , k = 0 , … , n . (53)
Repeating the positivity, Kraus-selection, and reduction arguments above gives a multipartite product operator K ~ = ⨂ X ∈ P M ~ X \widetilde{K}=\bigotimes_{X\in\mathcal{P}}\widetilde{M}_X K = ⨂ X ∈ P M X acting on the first two input factors and the catalyst. This operator acts nontrivially on ∣ Ψ d ⟩ 1 ⊗ ∣ Ψ d ⟩ 2 ⊗ ∣ χ ⟩ |\Psi_d\rangle_1\otimes|\Psi_d\rangle_2\otimes|\chi\rangle ∣ Ψ d ⟩ 1 ⊗ ∣ Ψ d ⟩ 2 ⊗ ∣ χ ⟩ and maps both ∣ Ψ d ⟩ 1 ⊗ ∣ ϕ ⟩ 2 ⊗ ∣ χ ⟩ |\Psi_d\rangle_1\otimes|\phi\rangle_2\otimes|\chi\rangle ∣ Ψ d ⟩ 1 ⊗ ∣ ϕ ⟩ 2 ⊗ ∣ χ ⟩ and ∣ ϕ ⟩ 1 ⊗ ∣ Ψ d ⟩ 2 ⊗ ∣ χ ⟩ |\phi\rangle_1\otimes|\Psi_d\rangle_2\otimes|\chi\rangle ∣ ϕ ⟩ 1 ⊗ ∣ Ψ d ⟩ 2 ⊗ ∣ χ ⟩ into L c a t m u l t i \mathcal{L}_{\mathrm{cat}}^{\mathrm{multi}} L cat multi for every ∣ ϕ ⟩ ∈ H o u t |\phi\rangle\in\mathcal{H}_{\mathrm{out}} ∣ ϕ ⟩ ∈ H out . Let A ∈ P \mathcal{A}\in\mathcal{P} A ∈ P and write A ˉ : = P ∖ { A } \bar{\mathcal{A}}:=\mathcal{P}\setminus\{\mathcal{A}\} A ˉ := P ∖ { A } . Let H C A : = ⨂ X ∈ A ˉ H C X \mathcal{H}_{\mathcal{C}_{\mathcal{A}}}:=\bigotimes_{X\in\bar{\mathcal{A}}}\mathcal{H}_{\mathcal{C}_X} H C A := ⨂ X ∈ A ˉ H C X , and let I C A ˉ I_{C_{\bar{\mathcal{A}}}} I C A ˉ denote the identity operator on H C A ˉ \mathcal{H}_{\mathcal{C}_{\bar{\mathcal{A}}}} H C A ˉ . Use the d d d -dimensional input and output subspaces U 1 , U 2 , U o u t \mathcal{U}_1,\mathcal{U}_2,\mathcal{U}_{\mathrm{out}} U 1 , U 2 , U out and V 1 , V 2 , V o u t \mathcal{V}_1,\mathcal{V}_2,\mathcal{V}_{\mathrm{out}} V 1 , V 2 , V out introduced in the proof of Theorem 2. Let P U i P_{\mathcal{U}_i} P U i and P V i P_{\mathcal{V}_i} P V i denote the projectors onto the corresponding input subspaces, and let P U o u t P_{\mathcal{U}_{\mathrm{out}}} P U out and P V o u t P_{\mathcal{V}_{\mathrm{out}}} P V out denote the output projectors. Across the bipartition A ∣ A ˉ \mathcal{A}|\bar{\mathcal{A}} A ∣ A ˉ , write K ~ = M ~ A ⊗ N ~ A ˉ \widetilde{K}=\widetilde{M}_{\mathcal{A}}\otimes\widetilde{N}_{\bar{\mathcal{A}}} K = M A ⊗ N A ˉ , where N ~ A ˉ : = ⨂ X ∈ A ˉ M ~ X \widetilde{N}_{\bar{\mathcal{A}}}:=\bigotimes_{X\in\bar{\mathcal{A}}}\widetilde{M}_X N A ˉ := ⨂ X ∈ A ˉ M X . Define M ~ : = ( P U o u t ⊗ I C A ˉ ) M ~ A ( P U 1 ⊗ P U 2 ⊗ I C A ˉ ) \widetilde{M}:=(P_{\mathcal{U}_{\mathrm{out}}}\otimes I_{C_{\bar{\mathcal{A}}}})\widetilde{M}_{\mathcal{A}}(P_{\mathcal{U}_1}\otimes P_{\mathcal{U}_2}\otimes I_{C_{\bar{\mathcal{A}}}}) M := ( P U out ⊗ I C A ˉ ) M A ( P U 1 ⊗ P U 2 ⊗ I C A ˉ ) and N ~ : = ( P V o u t ⊗ I C A ) N ~ A ˉ ( P V 1 ⊗ P V 2 ⊗ I C A ) \widetilde{N}:=(P_{\mathcal{V}_{\mathrm{out}}}\otimes I_{C_{\mathcal{A}}})\widetilde{N}_{\bar{\mathcal{A}}}(P_{\mathcal{V}_1}\otimes P_{\mathcal{V}_2}\otimes I_{C_{\mathcal{A}}}) N := ( P V out ⊗ I C A ) N A ˉ ( P V 1 ⊗ P V 2 ⊗ I C A ) . The input projections P U 1 ⊗ P U 2 ⊗ I C A ˉ P_{\mathcal{U}_1}\otimes P_{\mathcal{U}_2}\otimes I_{C_{\bar{\mathcal{A}}}} P U 1 ⊗ P U 2 ⊗ I C A ˉ and P V 1 ⊗ P V 2 ⊗ I C A P_{\mathcal{V}_1}\otimes P_{\mathcal{V}_2}\otimes I_{C_{\mathcal{A}}} P V 1 ⊗ P V 2 ⊗ I C A restrict the two arbitrary input states to the chosen subspaces. Since every resulting output is proportional to ∣ Ψ d ⟩ o u t ⊗ ∣ χ ⟩ |\Psi_d\rangle_{\mathrm{out}}\otimes|\chi\rangle ∣ Ψ d ⟩ out ⊗ ∣ χ ⟩ , and ∣ Ψ d ⟩ o u t |\Psi_d\rangle_{\mathrm{out}} ∣ Ψ d ⟩ out belongs to U o u t ⊗ V o u t \mathcal{U}_{\mathrm{out}}\otimes\mathcal{V}_{\mathrm{out}} U out ⊗ V out , the output projections P U o u t ⊗ I C A ˉ P_{\mathcal{U}_{\mathrm{out}}}\otimes I_{C_{\bar{\mathcal{A}}}} P U out ⊗ I C A ˉ and P V o u t ⊗ I C A P_{\mathcal{V}_{\mathrm{out}}}\otimes I_{C_{\mathcal{A}}} P V out ⊗ I C A leave these states unchanged. Under the identifications used in the proof of Theorem 2, each target state ∣ Ψ d ⟩ o u t |\Psi_d\rangle_{\mathrm{out}} ∣ Ψ d ⟩ out becomes a maximally entangled state of Schmidt rank d d d across A ∣ A ˉ \mathcal{A}|\bar{\mathcal{A}} A ∣ A ˉ . Moreover, K ~ : = M ~ ⊗ N ~ \widetilde{K}:=\widetilde{M}\otimes\widetilde{N} K := M ⊗ N is a product operator across A ∣ A ˉ \mathcal{A}|\bar{\mathcal{A}} A ∣ A ˉ and remains nonzero on ∣ Ψ d ⟩ 1 ⊗ ∣ Ψ d ⟩ 2 ⊗ ∣ χ ⟩ |\Psi_d\rangle_1\otimes|\Psi_d\rangle_2\otimes|\chi\rangle ∣ Ψ d ⟩ 1 ⊗ ∣ Ψ d ⟩ 2 ⊗ ∣ χ ⟩ . Therefore satisfies Eqs. (31) and (32), with the two sides of the bipartition treated as effective parties. This contradicts the lemma. Hence no universal catalytic LOCC EPP exists for either S d 1 , d 2 \mathcal{S}_{d_1,d_2} S d 1 , d 2 or G d \mathcal{G}_d G d .
SIII. Universal coherence purification under strictly incoherent operations
We first recall the basic notions of coherence purification under strictly incoherent operations [41 , 45 , 46 ] . Fix an orthonormal basis { ∣ r ⟩ } r = 0 d − 1 \{|r\rangle\}_{r=0}^{d-1} { ∣ r ⟩ } r = 0 d − 1 of C d \mathbb{C}^d C d , called the incoherent basis. A state δ \delta δ on C d \mathbb{C}^d C d is incoherent if and only if it can be written as δ = ∑ r = 0 d − 1 δ r ∣ r ⟩ ⟨ r ∣ \delta=\sum_{r=0}^{d-1}\delta_r|r\rangle\langle r| δ = ∑ r = 0 d − 1 δ r ∣ r ⟩ ⟨ r ∣ , where δ r ≥ 0 \delta_r\ge0 δ r ≥ 0 and ∑ r = 0 d − 1 δ r = 1 \sum_{r=0}^{d-1}\delta_r=1 ∑ r = 0 d − 1 δ r = 1 . A state that is not diagonal in the incoherent basis is called coherent. We denote the sets of incoherent and coherent states on C d \mathbb{C}^d C d by I d \mathcal{I}_d I d and C d \mathcal{C}_d C d , respectively. Relative to this basis, we define the maximally coherent state by ∣ φ d ⟩ = d − 1 / 2 ∑ r = 0 d − 1 ∣ r ⟩ |\varphi_d\rangle=d^{-1/2}\sum_{r=0}^{d-1}|r\rangle ∣ φ d ⟩ = d − 1/2 ∑ r = 0 d − 1 ∣ r ⟩ and write φ d : = ∣ φ d ⟩ ⟨ φ d ∣ \varphi_d:=|\varphi_d\rangle\langle\varphi_d| φ d := ∣ φ d ⟩ ⟨ φ d ∣ .
A strictly incoherent operation (SIO) from H i n \mathcal{H}_{\mathrm{in}} H in to H o u t \mathcal{H}_{\mathrm{out}} H out is a completely positive and trace-preserving map E \mathcal{E} E that admits a Kraus representation E ( ρ ) = ∑ μ K μ ρ K μ † \mathcal{E}(\rho)=\sum_\mu K_\mu\rho K_\mu^\dagger E ( ρ ) = ∑ μ K μ ρ K μ † in which every K μ K_\mu K μ is strictly incoherent. Here, H i n \mathcal{H}_{\mathrm{in}} H in and H o u t \mathcal{H}_{\mathrm{out}} H out are equipped with incoherent bases { ∣ r ⟩ i n } r = 0 d i n − 1 \{\lvert r\rangle_{\mathrm{in}}\}_{r=0}^{d_{\mathrm{in}}-1} {∣ r ⟩ in } r = 0 d in − 1 and { ∣ s ⟩ o u t } s = 0 d o u t − 1 \{\lvert s\rangle_{\mathrm{out}}\}_{s=0}^{d_{\mathrm{out}}-1} {∣ s ⟩ out } s = 0 d out − 1 , respectively. A linear operator K : H i n → H o u t K : \mathcal{H}_{\mathrm{in}} \to\mathcal{H}_{\mathrm{out}} K : H in → H out is strictly incoherent if K ∣ r ⟩ i n = α r ∣ f ( r ) ⟩ o u t K\lvert r\rangle_{\mathrm{in}} = \alpha_r\lvert f(r)\rangle_{\mathrm{out}} K ∣ r ⟩ in = α r ∣ f ( r ) ⟩ out and K † ∣ s ⟩ o u t = β s ∣ g ( s ) ⟩ i n K^\dagger\lvert s\rangle_{\mathrm{out}} = \beta_s\lvert g(s)\rangle_{\mathrm{in}} K † ∣ s ⟩ out = β s ∣ g ( s ) ⟩ in for every r r r and s s s , where α r , β s ∈ C \alpha_r,\beta_s \in\mathbb{C} α r , β s ∈ C may be zero, and f ( r ) f(r) f ( r ) and g ( s ) g(s) g ( s ) are output and input incoherent basis labels, respectively. Equivalently, the matrix of K K K with respect to these bases has at most one nonzero entry in each row and each column [41 , 46 ] .
An n n n -to-1 SIO coherence purification protocol is a two-outcome quantum instrument Λ = { Λ 0 , Λ 1 } \Lambda= \{\Lambda_0,\Lambda_1\} Λ = { Λ 0 , Λ 1 } consisting of a success branch Λ 0 \Lambda_0 Λ 0 and a failure branch Λ 1 \Lambda_1 Λ 1 , each of which is a completely positive and trace-nonincreasing map from B ( ( C d ) ⊗ n ) \mathcal{B}((\mathbb{C}^{d})^{\otimes n}) B (( C d ) ⊗ n ) to B ( C d ) \mathcal{B}(\mathbb{C}^{d}) B ( C d ) and admits a Kraus representation consisting of strictly incoherent operators, with Λ 0 + Λ 1 \Lambda_0+\Lambda_1 Λ 0 + Λ 1 being trace preserving. Here, the input incoherent basis is the tensor product of the incoherent bases of the n n n input factors, while the output incoherent basis is { ∣ r ⟩ } r = 0 d − 1 \{\lvert r\rangle\}_{r=0}^{d-1} {∣ r ⟩ } r = 0 d − 1 .
An n n n -to-1 SIO coherence purification protocol Λ = { Λ 0 , Λ 1 } \Lambda= \{\Lambda_0,\Lambda_1\} Λ = { Λ 0 , Λ 1 } is universal for C d \mathbb{C}^{d} C d if, for every total input state ρ i n = ρ 1 ⊗ ⋯ ⊗ ρ n \rho^{\mathrm{in}} = \rho_1 \otimes\cdots\otimes\rho_n ρ in = ρ 1 ⊗ ⋯ ⊗ ρ n with ρ j ∈ C d \rho_j \in\mathbb{C}^{d} ρ j ∈ C d for all j = 1 , … , n j=1,\ldots,n j = 1 , … , n , the success probability p s : = Tr [ Λ 0 ( ρ i n ) ] p_s := \operatorname{Tr}[\Lambda_0(\rho^{\mathrm{in}})] p s := Tr [ Λ 0 ( ρ in )] is positive and the normalized output conditioned on success, ρ ′ : = Λ 0 ( ρ i n ) / p s \rho' := \Lambda_0(\rho^{\mathrm{in}})/p_s ρ ′ := Λ 0 ( ρ in ) / p s , satisfies
F c ( ρ ′ ) ≥ max 1 ≤ j ≤ n F c ( ρ j ) . (54) F_c(\rho') \ge\max_{1\le j\le n} F_c(\rho_j). \tag*{(54)} F c ( ρ ′ ) ≥ 1 ≤ j ≤ n max F c ( ρ j ) . (54)
Here, F c ( ρ ) : = ⟨ ϕ d ∣ ρ ∣ ϕ d ⟩ F_c(\rho) := \langle\phi_d\lvert\rho\lvert\phi_d\rangle F c ( ρ ) := ⟨ ϕ d ∣ ρ ∣ ϕ d ⟩ denotes the fidelity of a state ρ \rho ρ on C d \mathbb{C}^{d} C d with the maximally coherent target state ∣ ϕ d ⟩ \lvert\phi_d\rangle ∣ ϕ d ⟩ .
For ϵ > 0 \epsilon>0 ϵ > 0 , define the restricted set
C d ( ϵ ) : = { ρ ∈ C d : F c ( ρ ) > 1 − ϵ } . (55) \mathbb{C}^{(\epsilon)}_d := \{\rho\in\mathbb{C}^{d}:F_c(\rho)>1-\epsilon\}. \tag*{(55)} C d ( ϵ ) := { ρ ∈ C d : F c ( ρ ) > 1 − ϵ } . (55)
Theorem 4. For every d ≥ 2 d\ge2 d ≥ 2 and any n ≥ 2 n\ge2 n ≥ 2 , no n n n -to-1 SIO coherence-purification protocol is universal for C d \mathbb{C}^{d} C d . Moreover, for every ϵ > 0 \epsilon>0 ϵ > 0 , no such protocol is universal for C d ( ϵ ) \mathbb{C}^{(\epsilon)}_d C d ( ϵ ) .
Proof. (a) We prove the statement for C d \mathbb{C}^{d} C d . Suppose, for contradiction, that a universal n n n -to-1 SIO coherence-purification protocol exists.
First, we construct total input states for which universality requires the normalized output conditioned on success to be the maximally coherent state.
Consider the n + 1 n+1 n + 1 total input states Σ 0 : = ⨂ j = 1 n φ d ( j ) \Sigma_0 := \bigotimes_{j=1}^{n}\varphi_d^{(j)} Σ 0 := ⨂ j = 1 n φ d ( j ) and Σ k : = φ d ( k ) ⊗ ⨂ j ≠ k σ j \Sigma_k := \varphi_d^{(k)}\otimes\bigotimes_{j\ne k}\sigma_j Σ k := φ d ( k ) ⊗ ⨂ j = k σ j for k = 1 , … , n k=1,\ldots,n k = 1 , … , n . Here, for each j = 1 , … , n j=1,\ldots,n j = 1 , … , n , let ∣ ϕ d ⟩ j : = d − 1 / 2 ∑ r = 0 d − 1 ∣ r ⟩ j \lvert\phi_d\rangle_j := d^{-1/2}\sum_{r=0}^{d-1}\lvert r\rangle_j ∣ ϕ d ⟩ j := d − 1/2 ∑ r = 0 d − 1 ∣ r ⟩ j denote the maximally coherent state on the j j j th input Hilbert space and set φ d ( j ) : = ∣ ϕ d ⟩ j ⟨ ϕ d ∣ j \varphi_d^{(j)} := \lvert\phi_d\rangle_j\langle\phi_d\rvert_j φ d ( j ) := ∣ ϕ d ⟩ j ⟨ ϕ d ∣ j . For each j = 1 , … , n j=1,\ldots,n j = 1 , … , n , let
σ j : = p φ d ( j ) + ( 1 − p ) I j d , 0 < p < 1. (56) \sigma_j := p\varphi_d^{(j)}+(1-p)\frac{I_j}{d},\qquad0<p<1. \tag*{(56)} σ j := p φ d ( j ) + ( 1 − p ) d I j , 0 < p < 1. (56)
where I j I_j I j is the identity on the j j j th input Hilbert space. Since 0 < p < 1 0<p<1 0 < p < 1 , every σ j \sigma_j σ j is full rank and coherent. Thus, Σ 0 , … , Σ n \Sigma_0,\ldots,\Sigma_n Σ 0 , … , Σ n are allowed total input states. Let ∣ ϕ d ⟩ o u t \lvert\phi_d\rangle_{\mathrm{out}} ∣ ϕ d ⟩ out denote the maximally coherent target state on the output Hilbert space and set φ d o u t : = ∣ ϕ d ⟩ o u t ⟨ ϕ d ∣ \varphi_d^{\mathrm{out}}:=\lvert\phi_d\rangle_{\mathrm{out}}\langle\phi_d\rvert φ d out := ∣ ϕ d ⟩ out ⟨ ϕ d ∣ . Moreover, each of Σ k \Sigma_k Σ k contains at least one maximally coherent target state, so the maximum input fidelity in Eq. (54) is one. Universality therefore requires
Λ 0 ( Σ k ) = p k φ d o u t , k = 0 , … , n , (57) \Lambda_0(\Sigma_k)=p_k\varphi_d^{\mathrm{out}},\qquad k=0,\ldots,n, \tag*{(57)} Λ 0 ( Σ k ) = p k φ d out , k = 0 , … , n , (57)
where p k : = Tr [ Λ 0 ( Σ k ) ] > 0 p_k:=\operatorname{Tr}[\Lambda_0(\Sigma_k)]>0 p k := Tr [ Λ 0 ( Σ k )] > 0 .
Second, we derive a support constraint and select a Kraus operator that acts nontrivially on Σ 0 \Sigma_0 Σ 0 .
By definition of an SIO coherence-purification protocol, the successful branch Λ 0 \Lambda_0 Λ 0 admits a Kraus representation Λ 0 ( ρ ) = ∑ μ K μ ρ K μ † \Lambda_0(\rho)=\sum_\mu K_\mu\rho K_\mu^\dagger Λ 0 ( ρ ) = ∑ μ K μ ρ K μ † in which every K μ K_\mu K μ is strictly incoherent. Let L c : = span { ∣ ϕ d ⟩ o u t } \mathcal{L}_c:=\operatorname{span}\{\lvert\phi_d\rangle_{\mathrm{out}}\} L c := span {∣ ϕ d ⟩ out } , and let P ⊥ : = I o u t − φ d o u t P_\perp:=I_{\mathrm{out}}-\varphi_d^{\mathrm{out}} P ⊥ := I out − φ d out denote the orthogonal projector onto L c ⊥ \mathcal{L}_c^\perp L c ⊥ , where I o u t I_{\mathrm{out}} I out is the identity on the output Hilbert space. For k = 1 , … , n k=1,\ldots,n k = 1 , … , n , Eq. (57) gives
P ⊥ Λ 0 ( Σ k ) P ⊥ = ∑ μ P ⊥ K μ Σ k K μ † P ⊥ = 0. (58) P_\perp\Lambda_0(\Sigma_k)P_\perp=\sum_\mu P_\perp K_\mu\Sigma_k K_\mu^\dagger P_\perp=0. \tag*{(58)} P ⊥ Λ 0 ( Σ k ) P ⊥ = μ ∑ P ⊥ K μ Σ k K μ † P ⊥ = 0. (58)
All summands in Eq. (58) are positive semidefinite, and their sum is zero; hence each summand must vanish. It follows that P ⊥ K μ Σ k 1 / 2 = 0 P_\perp K_\mu\Sigma_k^{1/2}=0 P ⊥ K μ Σ k 1/2 = 0 for every μ \mu μ and k = 1 , … , n k=1,\ldots,n k = 1 , … , n . Since every σ j \sigma_j σ j is full rank, the range of Σ k 1 / 2 = span { ∣ ϕ d ⟩ k } ⊗ H k c \Sigma_k^{1/2}=\operatorname{span}\{\lvert\phi_d\rangle_k\}\otimes\mathcal{H}^{c}_k Σ k 1/2 = span {∣ ϕ d ⟩ k } ⊗ H k c , where H k c : = ⨂ j ≠ k C d \mathcal{H}^{c}_k:=\bigotimes_{j\ne k}\mathbb{C}^{d} H k c := ⨂ j = k C d is the Hilbert space of all input factors except the k k k th; consequently, P ⊥ K μ Σ k 1 / 2 = 0 P_\perp K_\mu\Sigma_k^{1/2}=0 P ⊥ K μ Σ k 1/2 = 0 implies
K μ ( ∣ ϕ d ⟩ k ⊗ ∣ ϕ ⟩ k ˉ ) ∈ L c (59) K_\mu(\lvert\phi_d\rangle_k\otimes\lvert\phi\rangle_{\bar{k}})\in\mathcal{L}_c \tag*{(59)} K μ (∣ ϕ d ⟩ k ⊗ ∣ ϕ ⟩ k ˉ ) ∈ L c (59)
for every μ \mu μ , every k = 1 , … , n k=1,\ldots,n k = 1 , … , n , and every ∣ ϕ ⟩ k ˉ ∈ H k c \lvert\phi\rangle_{\bar{k}}\in\mathcal{H}^{c}_k ∣ ϕ ⟩ k ˉ ∈ H k c . For the pure input state Σ 0 \Sigma_0 Σ 0 , Eq. (57) becomes ∑ μ ∣ v μ ⟩ ⟨ v μ ∣ = p 0 φ d o u t \sum_\mu\lvert v_\mu\rangle\langle v_\mu\rvert=p_0\varphi_d^{\mathrm{out}} ∑ μ ∣ v μ ⟩ ⟨ v μ ∣ = p 0 φ d out , where ∣ v μ ⟩ = K μ ∣ ϕ d ⟩ ⊗ n \lvert v_\mu\rangle=K_\mu\lvert\phi_d\rangle^{\otimes n} ∣ v μ ⟩ = K μ ∣ ϕ d ⟩ ⊗ n . The condition p 0 > 0 p_0>0 p 0 > 0 ensures that at least one ∣ v μ ⟩ \lvert v_\mu\rangle ∣ v μ ⟩ is nonzero. We may therefore choose an SIO Kraus operator, denoted by K K K , such that
K ∣ ϕ d ⟩ ⊗ n ≠ 0. (60) K\lvert\phi_d\rangle^{\otimes n}\ne0. \tag*{(60)} K ∣ ϕ d ⟩ ⊗ n = 0. (60)
The selected K K K also satisfies Eq. (59) for every k = 1 , … , n k=1,\ldots,n k = 1 , … , n .
Third, we show that the nonzero condition and the support constraints are incompatible with the row-and-column structure of an SIO Kraus operator.
Label the incoherent basis of the total input Hilbert space by strings x = ( x 1 , … , x n ) ∈ { 0 , … , d − 1 } n \mathbf{x}=(x_1,\ldots,x_n)\in\{0,\ldots,d-1\}^{n} x = ( x 1 , … , x n ) ∈ { 0 , … , d − 1 } n , with ∣ x ⟩ : = ∣ x 1 ⟩ 1 ⊗ ⋯ ⊗ ∣ x n ⟩ n \lvert\mathbf{x}\rangle:=\lvert x_1\rangle_1\otimes\cdots\otimes\lvert x_n\rangle_n ∣ x ⟩ := ∣ x 1 ⟩ 1 ⊗ ⋯ ⊗ ∣ x n ⟩ n . The selected Kraus operator can be written as K = ∑ x ∈ S K c x ∣ f ( x ) ⟩ o u t ⟨ x ∣ K=\sum_{\mathbf{x}\in S_K}c_{\mathbf{x}}\lvert f(\mathbf{x})\rangle_{\mathrm{out}}\langle\mathbf{x}\rvert K = ∑ x ∈ S K c x ∣ f ( x ) ⟩ out ⟨ x ∣ , where S K S_K S K is the set of labels of the nonzero columns of K K K , c x ≠ 0 c_{\mathbf{x}}\ne0 c x = 0 for every x ∈ S K \mathbf{x}\in S_K x ∈ S K , and f : S K → { 0 , … , d − 1 } f:S_K\to\{0,\ldots,d-1\} f : S K → { 0 , … , d − 1 } assigns the output row occupied by each nonzero column. Because every output row contains at most one nonzero entry, f f f is injective. Consequently, the cardinality of S K S_K S K , denoted by ∣ S K ∣ \lvert S_K\rvert ∣ S K ∣ , satisfies
∣ S K ∣ ≤ d . (61) \lvert S_K\rvert\le d. \tag*{(61)} ∣ S K ∣ ≤ d . (61)
Equation Eq. (60) implies that S K S_K S K is nonempty. Fix a string x = ( x 1 , … , x n ) ∈ S K \mathbf{x}=(x_1,\ldots,x_n)\in S_K x = ( x 1 , … , x n ) ∈ S K , which we call the selected string. Keeping its last n − 1 n-1 n − 1 components fixed, consider the vector ∣ ϕ d ⟩ 1 ⊗ ∣ x 2 , … , x n ⟩ \lvert\phi_d\rangle_1\otimes\lvert x_2,\ldots,x_n\rangle ∣ ϕ d ⟩ 1 ⊗ ∣ x 2 , … , x n ⟩ , where ∣ x 2 , … , x n ⟩ : = ⨂ j = 2 n ∣ x j ⟩ j \lvert x_2,\ldots,x_n\rangle:=\bigotimes_{j=2}^{n}\lvert x_j\rangle_j ∣ x 2 , … , x n ⟩ := ⨂ j = 2 n ∣ x j ⟩ j . This vector has support only on the d d d basis strings ( r , x 2 , … , x n ) (r,x_2,\ldots,x_n) ( r , x 2 , … , x n ) , with r = 0 , … , d − 1 r=0,\ldots,d-1 r = 0 , … , d − 1 , and contains the selected string x \mathbf{x} x . Its image under K K K is nonzero because the column labeled by x \mathbf{x} x contributes a nonzero component, whereas distinct nonzero columns occupy distinct output rows and therefore cannot cancel one another. By Eq. (59) , this nonzero output must be proportional to ∣ ϕ d ⟩ o u t \lvert\phi_d\rangle_{\mathrm{out}} ∣ ϕ d ⟩ out and hence has the form ( c / d ) ∑ r = 0 d − 1 ∣ r ⟩ o u t (c/\sqrt{d})\sum_{r=0}^{d-1}\lvert r\rangle_{\mathrm{out}} ( c / d ) ∑ r = 0 d − 1 ∣ r ⟩ out for some c ≠ 0 c\ne0 c = 0 . Since each output row of K K K contains at most one nonzero entry, all d d d strings ( r , x 2 , … , x n ) (r,x_2,\ldots,x_n) ( r , x 2 , … , x n ) must belong to S K S_K S K . Together with Eq. (61) , this gives
S K = { ( r , x 2 , … , x n ) : 0 ≤ r ≤ d − 1 } . (62) S_K=\{(r,x_2,\ldots,x_n):0\le r\le d-1\}. \tag*{(62)} S K = {( r , x 2 , … , x n ) : 0 ≤ r ≤ d − 1 } . (62)
We next apply Eq. (59) with k = 2 k=2 k = 2 to the vector ∣ x 1 ⟩ 1 ⊗ ∣ ϕ d ⟩ 2 ⊗ ∣ x 3 , … , x n ⟩ \lvert x_1\rangle_1 \otimes\lvert\phi_d\rangle_2 \otimes\lvert x_3,\ldots,x_n\rangle ∣ x 1 ⟩ 1 ⊗ ∣ ϕ d ⟩ 2 ⊗ ∣ x 3 , … , x n ⟩ , where, for n = 2 n=2 n = 2 , the last tensor factor is omitted. This vector contains the same selected basis string x ∈ S K \mathbf{x}\in S_K x ∈ S K , so its image under K K K is again nonzero. Repeating the preceding argument gives
S K = { ( x 1 , r , x 3 , … , x n ) : 0 ≤ r ≤ d − 1 } . (63) S_K=\{(x_1,r,x_3,\ldots,x_n):0\le r\le d-1\}. \tag*{(63)} S K = {( x 1 , r , x 3 , … , x n ) : 0 ≤ r ≤ d − 1 } . (63)
For d ≥ 2 d\ge2 d ≥ 2 , Eqs. (62) and (63) are incompatible: the set in Eq. (62) contains a string whose first coordinate r r r differs from x 1 x_1 x 1 , whereas every string in Eq. (63) has first coordinate x 1 x_1 x 1 . This contradiction proves that no n n n -to-1 1 1 SIO coherence-purification protocol is universal for C d C_d C d .
(b) It remains to prove the statement for C d ( ϵ ) C_d^{(\epsilon)} C d ( ϵ ) . The states in Eq. (56) satisfy F c ( σ j ) = p + ( 1 − p ) / d F_c(\sigma_j)=p+(1-p)/d F c ( σ j ) = p + ( 1 − p ) / d , which approaches one as p → 1 p\to1 p → 1 . For every fixed ϵ > 0 \epsilon>0 ϵ > 0 , we may therefore choose p < 1 p<1 p < 1 sufficiently close to one so that every σ j \sigma_j σ j belongs to C d ( ϵ ) C_d^{(\epsilon)} C d ( ϵ ) . The state ϕ d \phi_d ϕ d also belong to C d ( ϵ ) C_d^{(\epsilon)} C d ( ϵ ) and have unit fidelity. Hence all total input states used in the preceding contradiction can be constructed from states in C d ( ϵ ) C_d^{(\epsilon)} C d ( ϵ ) . The same contradiction proves that no n n n -to-1 1 1 SIO coherence-purification protocol is universal for C d ( ϵ ) C_d^{(\epsilon)} C d ( ϵ ) .