Introduction
Bellomo, Lo Franco, and Compagno showed that two noninteracting qubits, each coupled to an independent non-Markovian reservoir, can recover entanglement [1 ] after a finite interval of complete disentanglement [2 , 3 ] . Their construction also applies to many noninteracting qubits with independent local reservoirs. Here we ask how coupling the local memory modes [4 , 5 ] changes the late-time pair concurrence [6 ] of N N N qubits, and whether this behavior can be understood from the spectral structure of the resulting reservoir network.
Collective dynamics can arise through coupling to a shared environment or through feedback, as demonstrated in studies of chemical and other coupled oscillator systems [7 , 8 , 9 ] . Here we examine how coupling the local environments of otherwise noninteracting qubits into a network changes their entanglement dynamics. We formulate the problem using coupled bosonic modes [10 , 11 , 12 , 13 ] , motivated by cavity and circuit QED systems [14 , 15 , 16 , 17 ] . We consider a finite, simple, connected, undirected graph with N ≥ 3 N \ge3 N ≥ 3 vertices, each hosting a qubit coupled to a bosonic memory mode, and the graph edges specify coherent hopping between the memory modes. Only the memory modes lose excitations, through independent local channels at zero temperature. This defines a lossy Jaynes-Cummings-Hubbard model [14 , 18 , 19 ] on the graph. We consider initially empty memory modes and arbitrary qubit states with at most two excitations.
Coupled cavities have been studied as auxiliary environments that prolong entanglement [20 , 21 ] , as lossy lattices that mediate interactions between emitters [22 , 23 ] , and as arrays supporting one-excitation transport [24 ] and atomic entanglement distribution [25 , 26 ] . Di Fidio, Ares, and Sperling obtained exact one-excitation propagation in general lossy cavity networks with one atom per cavity [27 ] . Beyond the one-excitation sector, earlier work includes two-polariton states in closed chains [28 , 29 ] , collective decay from one or two localized excitations [30 ] , approximately additive decay rates of multiexcitation subradiant states in emitter chains [31 , 32 , 33 ] , and two-excitation bound states that outlive all one-excitation states [34 ] .
With two excitations initially present, loss contributes to the one-excitation block of the density matrix [35 , 36 ] . Relating late-time concurrence to the graph spectrum therefore requires accounting for this contribution and controlling the decay of the two-excitation sector. We establish that the smallest amplitude decay rates in the two sectors satisfy γ 2 ≥ 2 γ 1 \gamma_2 \ge2\gamma_1 γ 2 ≥ 2 γ 1 for arbitrary uniform detuning and finite positive coupling, and determine the conditions for equality. At nonnegative detuning the inequality is strict, allowing the leading concurrence to be expressed in terms of the least damped one-excitation modes, with coefficients that retain contributions from two-excitation loss.
The physical mechanism follows from the collective modes of the reservoir network, reflecting the broader role of graph spectra in organizing spatial patterns in networked dynamical systems [37 , 38 , 39 ] . The adjacency eigenvalues shift the frequencies of the collective memory modes relative to the qubits. Channels with larger effective detuning have least damped eigenmodes with smaller amplitudes in the lossy memory degrees of freedom, and consequently longer lifetimes. At resonance, the spectral radius therefore determines the slowest decay rate, while the Perron vector fixes the asymptotic ratios of pair concurrence on nonbipartite graphs when the slowest mode is populated. On bipartite graphs, spectral symmetry preserves two equally damped modes with different frequencies. Their interference can produce arbitrarily late revivals when it overcomes the exponential decay envelope. Here a revival means renewed growth of concurrence after a decrease, not necessarily following a period of complete disentanglement [40 , 41 ] . Loss from two excitations can populate these long-lived modes and change the revival behavior even though the two-excitation sector decays faster. We also identify an exceptional detuned regime where γ 2 = 2 γ 1 \gamma_2 = 2\gamma_1 γ 2 = 2 γ 1 and both sectors contribute to the leading concurrence.
Model and spectral channels
We consider a finite simple connected undirected graph G G G with N ≥ 3 N \ge3 N ≥ 3 vertices and adjacency matrix A A A . Each vertex r ∈ V = { 1 , … , N } r \in V = \{1, \ldots, N\} r ∈ V = { 1 , … , N } hosts a qubit with transition frequency ω q \omega_q ω q and a bosonic memory mode with bare frequency ω m \omega_m ω m . The qubit has ground state ∣ g r ⟩ \lvert g_r\rangle ∣ g r ⟩ and excited state ∣ e r ⟩ \lvert e_r\rangle ∣ e r ⟩ with raising and lowering operators σ r + = ∣ e r ⟩ ⟨ g r ∣ \sigma_r^+ = \lvert e_r\rangle\langle g_r\rvert σ r + = ∣ e r ⟩ ⟨ g r ∣ and σ r − = ( σ r + ) † \sigma_r^- = (\sigma_r^+)^\dagger σ r − = ( σ r + ) † . The memory mode has annihilation operator m r m_r m r with [ m r , m s † ] = δ r s [m_r,m_s^\dagger] = \delta_{rs} [ m r , m s † ] = δ r s and [ m r , m s ] = 0 [m_r,m_s] = 0 [ m r , m s ] = 0 . In the frame rotating at ω q \omega_q ω q , with ℏ = 1 \hbar= 1 ℏ = 1 , the Jaynes–Cummings–Hubbard Hamiltonian is
H = Δ ∑ r m r † m r + g ∑ r ( σ r + m r + σ r − m r † ) + η ∑ r , s A r s m r † m s . (1) H = \Delta\sum_r m_r^\dagger m_r + g \sum_r \left(\sigma_r^+m_r + \sigma_r^-m_r^\dagger\right) + \eta\sum_{r,s} A_{rs}m_r^\dagger m_s. \tag*{(1)} H = Δ r ∑ m r † m r + g r ∑ ( σ r + m r + σ r − m r † ) + η r , s ∑ A r s m r † m s . (1)
where Δ = ω m − ω q \Delta= \omega_m - \omega_q Δ = ω m − ω q is the uniform detuning, with resonance at Δ = 0 \Delta= 0 Δ = 0 , and g , η > 0 g,\eta> 0 g , η > 0 are uniform coupling strengths. Here A r s = 1 A_{rs} = 1 A r s = 1 if vertices r r r and s s s are joined by an edge, and zero otherwise.
Only the memory modes lose excitations, through independent local channels at zero temperature Figure 1 (a). With S S S denoting all qubits and M M M all memory modes, the joint state ρ S M \rho_{SM} ρ S M obeys the Lindblad equation [42 , 43 ]
Figure 1. (a) Each qubit couples to a lossy memory mode. The memory modes hop on G 0 G_0 G 0 and lose population at rate 2 λ 2\lambda 2 λ . (b) Loss feeds lower excitation blocks. The source-free blocks in the top row receive no feeding from higher excitation sectors. Conjugate blocks are omitted. (c) Representative graphs and their spectral radii R R R . The graph G 0 G_0 G 0 has edges 12, 13, 23, 34, and G 0 + 24 G_0+24 G 0 + 24 also has the edge joining vertices 2 and 4. Red and blue vertices form the two parts of each bipartite graph. Nonbipartite graphs have yellow vertices. Vertices 1, 2, 3 form one triangle of the prism.
ρ ˙ S M = − i [ H , ρ S M ] + 2 λ ∑ r D [ m r ] ρ S M , (2) \dot{\rho}_{SM} = -i[H,\rho_{SM}] + 2\lambda\sum_r \mathcal{D}[m_r]\rho_{SM}, \tag*{(2)} ρ ˙ S M = − i [ H , ρ S M ] + 2 λ r ∑ D [ m r ] ρ S M , (2)
where λ > 0 \lambda> 0 λ > 0 and D [ L ] ρ = L ρ L † − 1 2 { L † L , ρ } \mathcal{D}[L]\rho= L\rho L^\dagger- \frac{1}{2}\{L^\dagger L,\rho\} D [ L ] ρ = L ρ L † − 2 1 { L † L , ρ } . The memory modes are initially in their joint vacuum ∣ Ω M ⟩ \lvert\Omega_M\rangle ∣ Ω M ⟩ , uncorrelated with the qubits,
ρ S M ( 0 ) = ρ S ( 0 ) ⊗ ∣ Ω M ⟩ ⟨ Ω M ∣ , (3) \rho_{SM}(0) = \rho_S(0) \otimes\lvert\Omega_M\rangle\langle\Omega_M\rvert, \tag*{(3)} ρ S M ( 0 ) = ρ S ( 0 ) ⊗ ∣ Ω M ⟩ ⟨ Ω M ∣ , (3)
where ρ S ( 0 ) \rho_S(0) ρ S ( 0 ) is an arbitrary density operator supported on Q 0 ⊕ Q 1 ⊕ Q 2 \mathcal{Q}_0 \oplus\mathcal{Q}_1 \oplus\mathcal{Q}_2 Q 0 ⊕ Q 1 ⊕ Q 2 , and Q n \mathcal{Q}_n Q n is the span of the qubit states with exactly n n n excitations. With ∣ Ω S ⟩ = ⨂ r ∣ g r ⟩ \lvert\Omega_S\rangle= \bigotimes_r \lvert g_r\rangle ∣ Ω S ⟩ = ⨂ r ∣ g r ⟩ the qubit vacuum, this subspace is spanned by ∣ Ω S ⟩ \lvert\Omega_S\rangle ∣ Ω S ⟩ , ∣ r ⟩ = σ r + ∣ Ω S ⟩ \lvert r\rangle= \sigma_r^+\lvert\Omega_S\rangle ∣ r ⟩ = σ r + ∣ Ω S ⟩ for r ∈ V r \in V r ∈ V , and ∣ r s ⟩ = σ r + σ s + ∣ Ω S ⟩ \lvert rs\rangle= \sigma_r^+\sigma_s^+\lvert\Omega_S\rangle ∣ r s ⟩ = σ r + σ s + ∣ Ω S ⟩ for r < s r < s r < s .
The total excitation operator is
N = ∑ r ( σ r + σ r − + m r † m r ) . (4) \mathcal{N} = \sum_r \left(\sigma_r^+\sigma_r^- + m_r^\dagger m_r\right). \tag*{(4)} N = r ∑ ( σ r + σ r − + m r † m r ) . (4)
Since [ H , N ] = 0 [H,\mathcal{N}] = 0 [ H , N ] = 0 and [ N , m r ] = − m r [\mathcal{N},m_r] = -m_r [ N , m r ] = − m r , the Hamiltonian preserves total excitation number and each loss lowers it by one.
Equation (2) is equivalent to
ρ ˙ S M = − i ( H e f f ρ S M − ρ S M H e f f † ) + 2 λ ∑ r m r ρ S M m r † , (5) \dot{\rho}_{SM} = -i\left(H_{\mathrm{eff}}\rho_{SM} - \rho_{SM}H_{\mathrm{eff}}^\dagger\right) + 2\lambda\sum_r m_r\rho_{SM}m_r^\dagger, \tag*{(5)} ρ ˙ S M = − i ( H eff ρ S M − ρ S M H eff † ) + 2 λ r ∑ m r ρ S M m r † , (5)
with
H e f f = H − i λ ∑ r m r † m r . (6) H_{\mathrm{eff}} = H - i\lambda\sum_r m_r^\dagger m_r. \tag*{(6)} H eff = H − iλ r ∑ m r † m r . (6)
The operator H e f f H_{\mathrm{eff}} H eff generates the unnormalized evolution conditioned on no quantum jumps.
Let Π n \Pi_n Π n project onto the sector with n n n total excitations. Since [ H e f f , N ] = 0 [H_{\mathrm{eff}},\mathcal{N}] = 0 [ H eff , N ] = 0 , the conditional evolution preserves each excitation sector, with generator
H n = Π n H e f f Π n ∣ ran Π n . (7) H_n = \Pi_n H_{\mathrm{eff}}\Pi_n\big|_{\operatorname{ran}\Pi_n}. \tag*{(7)} H n = Π n H eff Π n ran Π n . (7)
The smallest amplitude decay rate in each excited sector is
γ n = min z ∈ spec H n ( − Im z ) , n = 1 , 2. (8) \gamma_n = \min_{z\in\operatorname{spec}H_n}(-\operatorname{Im}z), \qquad n = 1,2. \tag*{(8)} γ n = z ∈ spec H n min ( − Im z ) , n = 1 , 2. (8)
An eigenvector of H n H_n H n with eigenvalue z z z defines an n n n -excitation mode with frequency Re z \operatorname{Re}z Re z . Modes with decay rate γ n \gamma_n γ n are called least damped.
Let ρ ( n , m ) = Π n ρ S M Π m \rho^{(n,m)} = \Pi_n\rho_{SM}\Pi_m ρ ( n , m ) = Π n ρ S M Π m denote the ( n , m ) (n,m) ( n , m ) block of the joint density matrix, and write ∣ 0 ⟩ = ∣ Ω S ⟩ ⊗ ∣ Ω M ⟩ \lvert0\rangle= \lvert\Omega_S\rangle\otimes\lvert\Omega_M\rangle ∣ 0 ⟩ = ∣ Ω S ⟩ ⊗ ∣ Ω M ⟩ . With H 0 = 0 H_0 = 0 H 0 = 0 and Γ r ( n ) = 2 λ Π n − 1 m r Π n \Gamma_r^{(n)} = \sqrt{2\lambda}\Pi_{n-1}m_r\Pi_n Γ r ( n ) = 2 λ Π n − 1 m r Π n ,
ρ ˙ ( n , m ) = − i ( H n ρ ( n , m ) − ρ ( n , m ) H m † ) + ∑ r Γ r ( n + 1 ) ρ ( n + 1 , m + 1 ) Γ r ( m + 1 ) † , (9) \dot{\rho}^{(n,m)} = -i\left(H_n\rho^{(n,m)} - \rho^{(n,m)}H_m^\dagger\right) + \sum_r \Gamma_r^{(n+1)}\rho^{(n+1,m+1)}\Gamma_r^{(m+1)\dagger}, \tag*{(9)} ρ ˙ ( n , m ) = − i ( H n ρ ( n , m ) − ρ ( n , m ) H m † ) + r ∑ Γ r ( n + 1 ) ρ ( n + 1 , m + 1 ) Γ r ( m + 1 ) † , (9)
for 0 ≤ n , m ≤ 2 0 \le n,m \le2 0 ≤ n , m ≤ 2 , with ρ ( n , m ) = 0 \rho^{(n,m)} = 0 ρ ( n , m ) = 0 for n > 2 n > 2 n > 2 or m > 2 m > 2 m > 2 . Such excitation-sector cascades are used in exact treatments of Lindblad equations without gain [44 ] . We refer to the sum over r r r as loss feeding: dissipation transfers contributions from higher to lower excitation blocks Figure 1 (b), while the conditional evolution within each sector is governed by H n H_n H n .
In the ordered basis
B 1 = ( σ 1 + ∣ 0 ⟩ , … , σ N + ∣ 0 ⟩ , m 1 † ∣ 0 ⟩ , … , m N † ∣ 0 ⟩ ) , (10) \mathcal{B}_1 = \left(\sigma_1^+\lvert0\rangle,\ldots,\sigma_N^+\lvert0\rangle,m_1^\dagger\lvert0\rangle,\ldots,m_N^\dagger\lvert0\rangle\right), \tag*{(10)} B 1 = ( σ 1 + ∣ 0 ⟩ , … , σ N + ∣ 0 ⟩ , m 1 † ∣ 0 ⟩ , … , m N † ∣ 0 ⟩ ) , (10)
the sector generator is
H 1 = ( 0 g I N g I N η A + ( Δ − i λ ) I N ) , (11) H_1 =
\begin{pmatrix}
0 & g\mathbb{I}_N \\
g\mathbb{I}_N & \eta A + (\Delta-i\lambda)\mathbb{I}_N
\end{pmatrix}, \tag*{(11)} H 1 = ( 0 g I N g I N η A + ( Δ − iλ ) I N ) , (11)
where I N \mathbb{I}_N I N is the N × N N \times N N × N identity matrix. The spectral decomposition of the adjacency matrix is
A = ∑ θ ∈ spec A θ E θ . (12) A = \sum_{\theta\in\operatorname{spec} A}\theta E_\theta. \tag*{(12)} A = θ ∈ spec A ∑ θ E θ . (12)
where E θ E_\theta E θ is the orthogonal projector onto the eigenspace with eigenvalue θ \theta θ .
For a normalized eigenvector u u u of A A A with eigenvalue θ \theta θ , define
∣ q u ⟩ = ∑ r u r ∣ r ⟩ ⊗ ∣ Ω M ⟩ , ∣ m u ⟩ = ∣ Ω S ⟩ ⊗ ∑ r u r m r † ∣ Ω M ⟩ . (13) \begin{aligned}
|q_u\rangle&= \sum_r u_r |r\rangle\otimes|\Omega_M\rangle,\\
|m_u\rangle&= |\Omega_S\rangle\otimes\sum_r u_r m_r^\dagger|\Omega_M\rangle. \tag*{(13)}
\end{aligned} ∣ q u ⟩ ∣ m u ⟩ = r ∑ u r ∣ r ⟩ ⊗ ∣ Ω M ⟩ , = ∣ Ω S ⟩ ⊗ r ∑ u r m r † ∣ Ω M ⟩ . (13)
Using A u = θ u Au=\theta u A u = θ u , we obtain
H 1 ∣ q u ⟩ = g ∣ m u ⟩ , H_1|q_u\rangle= g|m_u\rangle, H 1 ∣ q u ⟩ = g ∣ m u ⟩ ,
H 1 ∣ m u ⟩ = g ∣ q u ⟩ + ( Δ + η θ − i λ ) ∣ m u ⟩ . (14) H_1|m_u\rangle= g|q_u\rangle+ (\Delta+\eta\theta-i\lambda)|m_u\rangle. \tag*{(14)} H 1 ∣ m u ⟩ = g ∣ q u ⟩ + ( Δ + η θ − iλ ) ∣ m u ⟩ . (14)
Thus S u = span { ∣ q u ⟩ , ∣ m u ⟩ } \mathcal{S}_u=\operatorname{span}\{|q_u\rangle,|m_u\rangle\} S u = span { ∣ q u ⟩ , ∣ m u ⟩} is invariant under H 1 H_1 H 1 , and in the ordered orthonormal basis ( ∣ q u ⟩ , ∣ m u ⟩ ) (|q_u\rangle,|m_u\rangle) ( ∣ q u ⟩ , ∣ m u ⟩) the restriction H 1 ∣ S u H_1|_{\mathcal{S}_u} H 1 ∣ S u is represented by
h θ = ( 0 g g D θ − i λ ) , D θ = Δ + η θ . (15) h_\theta=
\begin{pmatrix}
0 & g\\
g & D_\theta-i\lambda
\end{pmatrix},
\qquad D_\theta=\Delta+\eta\theta. \tag*{(15)} h θ = ( 0 g g D θ − iλ ) , D θ = Δ + η θ . (15)
We call h θ h_\theta h θ a spectral channel, with effective detuning D θ = ( ω m + η θ ) − ω q D_\theta=(\omega_m+\eta\theta)-\omega_q D θ = ( ω m + η θ ) − ω q . Each adjacency eigenvalue thus shifts the frequency of a collective memory mode relative to the qubits. The channel occurs in H 1 H_1 H 1 with multiplicity rank E θ \operatorname{rank}E_\theta rank E θ , and its eigenvalues are the roots of z 2 − ( D θ − i λ ) z − g 2 = 0 z^2-(D_\theta-i\lambda)z-g^2=0 z 2 − ( D θ − iλ ) z − g 2 = 0 .
For either eigenvalue z z z of h θ h_\theta h θ , let χ = ( 1 , z / g ) T \chi=(1,z/g)^{\mathsf{T}} χ = ( 1 , z / g ) T , so that h θ χ = z χ h_\theta\chi=z\chi h θ χ = z χ . Multiplying this equation by χ † \chi^\dagger χ † from the left and taking imaginary parts gives
− Im z = λ ∣ z ∣ 2 g 2 + ∣ z ∣ 2 . (16) -\operatorname{Im}z=\lambda\frac{|z|^2}{g^2+|z|^2}. \tag*{(16)} − Im z = λ g 2 + ∣ z ∣ 2 ∣ z ∣ 2 . (16)
The decay rate is therefore proportional to the fraction of the mode residing in the lossy memory degrees of freedom. Since the roots have product − g 2 ≠ 0 -g^2\ne0 − g 2 = 0 , both decay rates are positive, so γ 1 > 0 \gamma_1>0 γ 1 > 0 . For D θ ≠ 0 D_\theta\ne0 D θ = 0 the two roots have different moduli (Appendix A), and by Eq. (16) the root of smaller modulus, denoted z θ z_\theta z θ , also has the smaller decay rate.
Let γ ( D ) \gamma(D) γ ( D ) denote the smaller amplitude decay rate of a spectral channel with effective detuning D D D . It is even in D D D and strictly decreases with ∣ D ∣ |D| ∣ D ∣ (Appendix A). Thus, increasing the magnitude of the effective detuning suppresses the memory-mode component of the channel’s least damped eigenmode and prolongs its lifetime. Since spec H 1 = ⋃ θ ∈ spec A spec h θ \operatorname{spec}H_1=\bigcup_{\theta\in\operatorname{spec}A}\operatorname{spec}h_\theta spec H 1 = ⋃ θ ∈ spec A spec h θ ,
γ 1 = min θ ∈ spec A γ ( Δ + η θ ) . (17) \gamma_1=\min_{\theta\in\operatorname{spec}A}\gamma(\Delta+\eta\theta). \tag*{(17)} γ 1 = θ ∈ spec A min γ ( Δ + η θ ) . (17)
The channel h θ h_\theta h θ has an eigenvalue z z z satisfying − Im z = γ 1 -\operatorname{Im}z=\gamma_1 − Im z = γ 1 if and only if ∣ Δ + η θ ∣ = max ϑ ∈ spec A ∣ Δ + η ϑ ∣ |\Delta+\eta\theta|=\max_{\vartheta\in\operatorname{spec}A}|\Delta+\eta\vartheta| ∣Δ + η θ ∣ = max ϑ ∈ spec A ∣Δ + η ϑ ∣ . At resonance, Δ = 0 \Delta=0 Δ = 0 , this condition is equivalent to ∣ θ ∣ = max ϑ ∈ spec A ∣ ϑ ∣ |\theta|=\max_{\vartheta\in\operatorname{spec}A}|\vartheta| ∣ θ ∣ = max ϑ ∈ spec A ∣ ϑ ∣ .
Let R R R denote the spectral radius of A A A ,
R = max θ ∈ spec A ∣ θ ∣ . (18) R=\max_{\theta\in\operatorname{spec}A}|\theta|. \tag*{(18)} R = θ ∈ spec A max ∣ θ ∣. (18)
Connectedness makes R R R a simple eigenvalue with a positive unit eigenvector v v v , called the Perron vector [45 ] . A graph is bipartite if its vertices split into two parts and every edge joins opposite parts. A connected graph is nonbipartite exactly when it contains a cycle of odd length [45 ] . For a connected graph, − R -R − R is an eigenvalue exactly when the graph is bipartite. In that case, let V ± V_\pm V ± denote the two parts and J J J the diagonal matrix with J r r = ± 1 J_{rr}=\pm1 J r r = ± 1 for r ∈ V ± r\in V_\pm r ∈ V ± , so that J A J = − A JAJ=-A J A J = − A and w = J v w=Jv w = J v is a unit eigenvector at − R -R − R .
Least damped channels and decay rates
Let θ − = min ϑ ∈ spec A ϑ \theta_-=\min_{\vartheta\in\operatorname{spec}A}\vartheta θ − = min ϑ ∈ spec A ϑ and define
M = { θ ∈ spec A : γ ( Δ + η θ ) = γ 1 } . \mathcal{M}=\{\theta\in\operatorname{spec}A:\gamma(\Delta+\eta\theta)=\gamma_1\}. M = { θ ∈ spec A : γ ( Δ + η θ ) = γ 1 } .
We call h θ h_\theta h θ a least damped channel when θ ∈ M \theta\in\mathcal{M} θ ∈ M . The ordering of γ ( D ) \gamma(D) γ ( D ) gives
M = { { R } , Δ > − η 2 ( R + θ − ) , { R , θ − } , Δ = − η 2 ( R + θ − ) , { θ − } , Δ < − η 2 ( R + θ − ) . (19) \mathcal{M}=
\begin{cases}
\{R\}, & \Delta>-\dfrac{\eta}{2}(R+\theta_-),\\
\{R,\theta_-\}, & \Delta=-\dfrac{\eta}{2}(R+\theta_-),\\
\{\theta_-\}, & \Delta<-\dfrac{\eta}{2}(R+\theta_-).
\end{cases} \tag*{(19)} M = ⎩ ⎨ ⎧ { R } , { R , θ − } , { θ − } , Δ > − 2 η ( R + θ − ) , Δ = − 2 η ( R + θ − ) , Δ < − 2 η ( R + θ − ) . (19)
At Δ = − η ( R + θ − ) / 2 \Delta=-\eta(R+\theta_-)/2 Δ = − η ( R + θ − ) /2 , the two extremal channels have equal decay rates, with D R = − D θ − = η ( R − θ − ) / 2 D_R=-D_{\theta_-}=\eta(R-\theta_-)/2 D R = − D θ − = η ( R − θ − ) /2 and z θ − = − z R ‾ z_{\theta_-}=-\overline{z_R} z θ − = − z R .
At resonance, Δ = 0 \Delta=0 Δ = 0 , we have M = { R } \mathcal{M}=\{R\} M = { R } for nonbipartite graphs and M = { R , − R } \mathcal{M}=\{R,-R\} M = { R , − R } for bipartite graphs. In both cases, every z θ z_\theta z θ with θ ∈ M \theta\in\mathcal{M} θ ∈ M has algebraic multiplicity one in H 1 H_1 H 1 (Appendix A).
For every θ ∈ M \theta\in\mathcal{M} θ ∈ M , we have D θ ≠ 0 D_\theta\ne0 D θ = 0 , and z θ z_\theta z θ is the unique eigenvalue of h θ h_\theta h θ with − Im z θ = γ 1 -\operatorname{Im}z_\theta=\gamma_1 − Im z θ = γ 1 . Let P θ P_\theta P θ denote the corresponding spectral projection of H 1 H_1 H 1 , given explicitly in Appendix A2. For fixed g , η , λ > 0 g,\eta,\lambda>0 g , η , λ > 0 and Δ ∈ R \Delta\in\mathbb{R} Δ ∈ R , there exists $\varepsilon>0 such that
e − i H 1 t = ∑ θ ∈ M e − i z θ t P θ + O ( e − ( γ 1 + ε ) t ) . (20) e^{-iH_1t}=\sum_{\theta\in\mathcal{M}}e^{-iz_\theta t}P_\theta+O\left(e^{-(\gamma_1+\varepsilon)t}\right). \tag*{(20)} e − i H 1 t = θ ∈ M ∑ e − i z θ t P θ + O ( e − ( γ 1 + ε ) t ) . (20)
where the remainder is measured in operator norm (Appendix A2). Thus, at long times, the one-excitation dynamics is governed by the least damped spectral channels.
Proposition 1 (Sector decay bound and its equality case). For the model of Sec. II, we have
γ 2 ≥ 2 γ 1 > 0 (21) \gamma_2\ge2\gamma_1>0 \tag*{(21)} γ 2 ≥ 2 γ 1 > 0 (21)
for all g , η , λ > 0 g,\eta,\lambda>0 g , η , λ > 0 and Δ ∈ R \Delta\in\mathbb{R} Δ ∈ R . Equality γ 2 = 2 γ 1 \gamma_2=2\gamma_1 γ 2 = 2 γ 1 holds if and only if Δ ≤ − η ( R + θ − ) / 2 \Delta\le-\eta(R+\theta_-)/2 Δ ≤ − η ( R + θ − ) /2 and there exists a nonzero symmetric matrix S S S with zero diagonal satisfying A S = θ − S AS=\theta_-S A S = θ − S . In particular, if Δ ≥ 0 \Delta\ge0 Δ ≥ 0 or if θ − \theta_- θ − is a simple eigenvalue of A A A , then
γ 2 > 2 γ 1 . (22) \gamma_2>2\gamma_1. \tag*{(22)} γ 2 > 2 γ 1 . (22)
The bound establishes a hierarchy between excitation sectors. When the inequality is strict, the leading concurrence at order e − 2 γ 1 t e^{-2\gamma_1t} e − 2 γ 1 t is determined by the slow one-excitation coherence, although its coefficient can retain contributions from loss out of the two-excitation sector. This distinction underlies the asymptotic concurrence analysis below.
The proof is given in Appendix B. For Δ ≥ 0 \Delta\ge0 Δ ≥ 0 , strict inequality follows from R + θ − ≥ 0 R+\theta_- \ge0 R + θ − ≥ 0 , with equality R + θ − = 0 R+\theta_-=0 R + θ − = 0 only for bipartite graphs, whose smallest adjacency eigenvalue θ − = − R \theta_-=-R θ − = − R is simple. The equality condition can nevertheless be realized at negative detuning. On the complete graph K N K_N K N , θ − = − 1 \theta_-=-1 θ − = − 1 , and every symmetric matrix with zero diagonal and zero row sums satisfies A S = − S AS=-S A S = − S . Nonzero matrices of this kind exist for N ≥ 4 N\ge4 N ≥ 4 , giving γ 2 = 2 γ 1 \gamma_2=2\gamma_1 γ 2 = 2 γ 1 whenever Δ ≤ − η ( N − 2 ) / 2 \Delta\le-\eta(N-2)/2 Δ ≤ − η ( N − 2 ) /2 . We examine the case K 4 K_4 K 4 in § 0.5 F.
Late-time concurrence
For vertices i < j i<j i < j , the pair state is
ρ i j ( t ) = Tr M , V ∖ { i , j } ρ S M ( t ) , (23) \rho_{ij}(t)=\operatorname{Tr}_{M,V\setminus\{i,j\}}\rho_{SM}(t), \tag*{(23)} ρ ij ( t ) = Tr M , V ∖ { i , j } ρ S M ( t ) , (23)
where the trace removes all memory modes and all qubits except i , j i,j i , j . We use the ordered basis ( ∣ 11 ⟩ , ∣ 10 ⟩ , ∣ 01 ⟩ , ∣ 00 ⟩ ) (|11\rangle,|10\rangle,|01\rangle,|00\rangle) ( ∣11 ⟩ , ∣10 ⟩ , ∣01 ⟩ , ∣00 ⟩) , with the first label referring to qubit i i i and 1 denoting an excitation.
Let Y = σ y ⊗ σ y Y=\sigma_y\otimes\sigma_y Y = σ y ⊗ σ y , where σ y \sigma_y σ y is the Pauli y y y matrix. For any positive semidefinite 4 × 4 4\times4 4 × 4 matrix X X X , define
C ( X ) = max { 0 , ξ 1 − ξ 2 − ξ 3 − ξ 4 } , (24) C(X)=\max\{0,\xi_1-\xi_2-\xi_3-\xi_4\}, \tag*{(24)} C ( X ) = max { 0 , ξ 1 − ξ 2 − ξ 3 − ξ 4 } , (24)
where ξ 1 ≥ ξ 2 ≥ ξ 3 ≥ ξ 4 ≥ 0 \xi_1\ge\xi_2\ge\xi_3\ge\xi_4\ge0 ξ 1 ≥ ξ 2 ≥ ξ 3 ≥ ξ 4 ≥ 0 are the square roots of the eigenvalues of X Y X ∗ Y XYX^*Y X Y X ∗ Y , and complex conjugation is taken in the stated basis. For a normalized state this is the two-qubit concurrence [6 ] , and the same formula defines its homogeneous extension, C ( a X ) = a C ( X ) C(aX)=aC(X) C ( a X ) = a C ( X ) for a ≥ 0 a\ge0 a ≥ 0 . We write C i j ( t ) = C ( ρ i j ( t ) ) C_{ij}(t)=C(\rho_{ij}(t)) C ij ( t ) = C ( ρ ij ( t )) .
For every such X X X ,
∣ C ( X ) − 2 ∣ X 23 ∣ ∣ ≤ 2 X 11 X 44 . (25) \left|C(X)-2|X_{23}|\right|\le2\sqrt{X_{11}X_{44}}. \tag*{(25)} ∣ C ( X ) − 2∣ X 23 ∣ ∣ ≤ 2 X 11 X 44 . (25)
Here X 23 = ⟨ 10 ∣ X ∣ 01 ⟩ X_{23}=\langle10|X|01\rangle X 23 = ⟨ 10∣ X ∣01 ⟩ . The lower estimate follows from the bound obtained from the diagonal and antidiagonal entries [46 ] , and Appendix C.1 proves both estimates.
The pair population ( ρ i j ) 11 (\rho_{ij})_{11} ( ρ ij ) 11 comes entirely from ρ ( 2 , 2 ) \rho^{(2,2)} ρ ( 2 , 2 ) . When γ 2 > 2 γ 1 \gamma_2>2\gamma_1 γ 2 > 2 γ 1 , its square root decays faster than e − 2 γ 1 t e^{-2\gamma_1t} e − 2 γ 1 t , so the doubly excited pair population cannot contribute at leading order. Equation (25) then reduces the concurrence at order e − 2 γ 1 t e^{-2\gamma_1t} e − 2 γ 1 t to 2 ∣ ( ρ i j ) 23 ∣ 2|(\rho_{ij})_{23}| 2∣ ( ρ ij ) 23 ∣ .
Write B n m = ρ ( n , m ) ( 0 ) B_{nm}=\rho^{(n,m)}(0) B nm = ρ ( n , m ) ( 0 ) and ∣ q i ⟩ = σ i + ∣ 0 ⟩ |q_i\rangle=\sigma_i^+|0\rangle ∣ q i ⟩ = σ i + ∣0 ⟩ . Let v a v_a v a run through orthonormal real bases of the adjacency eigenspaces with θ ∈ M \theta\in\mathcal{M} θ ∈ M , and let z a z_a z a be the associated root z θ z_\theta z θ . The corresponding right and left eigenvectors are
x a = ( v a z a v a / g ) , y a † = x a T 1 + z a 2 / g 2 . (26) x_a=\begin{pmatrix}v_a\\ z_av_a/g\end{pmatrix},\qquad y_a^\dagger=\frac{x_a^\mathsf{T}}{1+z_a^2/g^2}. \tag*{(26)} x a = ( v a z a v a / g ) , y a † = 1 + z a 2 / g 2 x a T . (26)
The bound γ 2 ≥ 2 γ 1 > γ 1 \gamma_{2} \ge2\gamma_{1} > \gamma_{1} γ 2 ≥ 2 γ 1 > γ 1 ensures convergence of the integral defining the matrix Q Q Q ,
Q a b = y a † B 11 y b + ∫ 0 ∞ e i ( z a − z ˉ b ) u y a † F 11 ( u ) y b d u , Q_{ab}=y_{a}^{\dagger}B_{11}y_{b}+\int_{0}^{\infty}e^{i(z_{a}-\bar{z}_{b})u}y_{a}^{\dagger}F_{11}(u)y_{b}\,\mathrm{d}u, Q ab = y a † B 11 y b + ∫ 0 ∞ e i ( z a − z ˉ b ) u y a † F 11 ( u ) y b d u ,
F 11 ( u ) = ∑ r Γ r ( 2 ) ρ ( 2 , 2 ) ( u ) Γ r ( 2 ) † . (27) F_{11}(u)=\sum_{r}\Gamma_{r}^{(2)}\rho^{(2,2)}(u)\Gamma_{r}^{(2)\dagger}. \tag*{(27)} F 11 ( u ) = r ∑ Γ r ( 2 ) ρ ( 2 , 2 ) ( u ) Γ r ( 2 ) † . (27)
The positive semidefinite matrix Q Q Q describes the surviving populations and coherences of the least damped one-excitation modes. Its first term projects the initial one-excitation state onto these modes, while the integral accounts for excitation transferred into them by loss from the two-excitation sector. Thus, even when two excitations decay faster, they can influence the coefficients governing late-time concurrence. Define
Φ i j ( t ) = ∑ a , b Q a b v a ( i ) v b ( j ) e − i ( ω a − ω b ) t , ω a = Re z a . (28) \Phi_{ij}(t)=\sum_{a,b}Q_{ab}v_{a}(i)v_{b}(j)e^{-i(\omega_{a}-\omega_{b})t},\quad\omega_{a}=\operatorname{Re}z_{a}. \tag*{(28)} Φ ij ( t ) = a , b ∑ Q ab v a ( i ) v b ( j ) e − i ( ω a − ω b ) t , ω a = Re z a . (28)
When γ 2 = 2 γ 1 \gamma_{2}=2\gamma_{1} γ 2 = 2 γ 1 , contributions from both excitation sectors can survive at the same leading order. To retain these terms, define the balanced pair matrix
Λ t = diag ( e γ 1 t , 1 , 1 , e − γ 1 t ) , \Lambda_{t}=\operatorname{diag}(e^{\gamma_{1}t},1,1,e^{-\gamma_{1}t}), Λ t = diag ( e γ 1 t , 1 , 1 , e − γ 1 t ) ,
X i j ( t ) = e 2 γ 1 t Λ t ρ i j ( t ) Λ t . (29) X_{ij}(t)=e^{2\gamma_{1}t}\Lambda_{t}\rho_{ij}(t)\Lambda_{t}. \tag*{(29)} X ij ( t ) = e 2 γ 1 t Λ t ρ ij ( t ) Λ t . (29)
Each entry of X i j ( t ) X_{ij}(t) X ij ( t ) is a finite sum ∑ s p s ( t ) e s t \sum_{s}p_{s}(t)e^{st} ∑ s p s ( t ) e s t with distinct complex exponents s s s and polynomial coefficients p s ( t ) p_{s}(t) p s ( t ) . Define X i j ( 0 ) ( t ) X_{ij}^{(0)}(t) X ij ( 0 ) ( t ) by retaining the terms with Re s = 0 \operatorname{Re}s=0 Re s = 0 .
Proposition 2 (Asymptotic laws for exact concurrence). For every initial state of the form Eq. (3) and every pair, there is β > 0 \beta>0 β > 0 such that
C i j ( t ) = e − 2 γ 1 t F i j ( t ) + O ( e − ( 2 γ 1 + β ) t ) , (30) C_{ij}(t)=e^{-2\gamma_{1}t}F_{ij}(t)+O(e^{-(2\gamma_{1}+\beta)t}), \tag*{(30)} C ij ( t ) = e − 2 γ 1 t F ij ( t ) + O ( e − ( 2 γ 1 + β ) t ) , (30) where
F i j ( t ) = { 2 ∣ Φ i j ( t ) ∣ , γ 2 > 2 γ 1 , C ( X i j ( 0 ) ( t ) ) , γ 2 = 2 γ 1 . (31) F_{ij}(t)=
\begin{cases}
2|\Phi_{ij}(t)|, & \gamma_{2}>2\gamma_{1},\\
\mathcal{C}(X_{ij}^{(0)}(t)), & \gamma_{2}=2\gamma_{1}.
\end{cases} \tag*{(31)} F ij ( t ) = { 2∣ Φ ij ( t ) ∣ , C ( X ij ( 0 ) ( t )) , γ 2 > 2 γ 1 , γ 2 = 2 γ 1 . (31)
Appendix C gives the entries of X i j ( 0 ) X_{ij}^{(0)} X ij ( 0 ) , proves its positivity, and establishes these asymptotic laws. We call F i j F_{ij} F ij the leading profile, e − 2 γ 1 t F i j ( t ) e^{-2\gamma_{1}t}F_{ij}(t) e − 2 γ 1 t F ij ( t ) the leading concurrence, and e 2 γ 1 t C i j ( t ) e^{2\gamma_{1}t}C_{ij}(t) e 2 γ 1 t C ij ( t ) the rescaled concurrence.
A revival of C i j C_{ij} C ij is a triple of times t 0 < t 1 < t 2 t_{0}<t_{1}<t_{2} t 0 < t 1 < t 2 with C i j ( t 1 ) < min { C i j ( t 0 ) , C i j ( t 2 ) } C_{ij}(t_{1})<\min\{C_{ij}(t_{0}),C_{ij}(t_{2})\} C ij ( t 1 ) < min { C ij ( t 0 ) , C ij ( t 2 )} . We say that C i j C_{ij} C ij has arbitrarily late revivals if it has a revival with t 0 > T t_{0}>T t 0 > T for every T T T .
Perron pair ratios on nonbipartite graphs
At resonance, Eq. (30) becomes
C i j ( t ) = 2 e − 2 γ 1 t ∣ Φ i j ( t ) ∣ + R i j ( t ) , C_{ij}(t)=2e^{-2\gamma_{1}t}|\Phi_{ij}(t)|+R_{ij}(t), C ij ( t ) = 2 e − 2 γ 1 t ∣ Φ ij ( t ) ∣ + R ij ( t ) ,
R i j ( t ) = O ( e − ( 2 γ 1 + β ) t ) . (32) R_{ij}(t)=O(e^{-(2\gamma_{1}+\beta)t}). \tag*{(32)} R ij ( t ) = O ( e − ( 2 γ 1 + β ) t ) . (32)
For a nonbipartite graph, the Perron mode is the unique least damped channel, giving Φ i j = Q R R v i v j \Phi_{ij}=Q_{RR}v_{i}v_{j} Φ ij = Q R R v i v j . When Q R R > 0 Q_{RR}>0 Q R R > 0 , the concurrence of every pair therefore inherits the spatial weights of the Perron eigenvector, and
lim t → ∞ C i j ( t ) C k l ( t ) = v i v j v k v l . (33) \lim_{t\to\infty}\frac{C_{ij}(t)}{C_{kl}(t)}=\frac{v_{i}v_{j}}{v_{k}v_{l}}. \tag*{(33)} t → ∞ lim C k l ( t ) C ij ( t ) = v k v l v i v j . (33)
Thus, although concurrence decays in magnitude, its relative distribution across pairs approaches a pattern determined solely by the graph. The condition Q R R > 0 Q_{RR}>0 Q R R > 0 is essential. The graph with edges 12 12 12 , 15 15 15 , 25 25 25 , 34 34 34 , 35 35 35 , 45 45 45 contains triangles, but the initial state ( ∣ 1 ⟩ − ∣ 3 ⟩ ) / 2 (|1\rangle-|3\rangle)/\sqrt{2} ( ∣1 ⟩ − ∣3 ⟩) / 2 has support only on the adjacency eigenvectors ( 1 , ± 1 , − 1 , ∓ 1 , 0 ) / 2 (1,\pm1,-1,\mp1,0)/2 ( 1 , ± 1 , − 1 , ∓ 1 , 0 ) /2 at eigenvalues ± 1 \pm1 ± 1 , so Q R R = 0 Q_{RR}=0 Q R R = 0 . Interference between the qubit amplitudes of these two channels causes C 13 C_{13} C 13 to repeatedly vanish and revive, even at arbitrarily late times (Appendix A4). Thus, odd cycles alone do not exclude arbitrarily late revivals.
Perron pair ratios on bipartite graphs
At resonance, a bipartite graph retains two equally damped channels, associated with R R R and − R -R − R . Their different frequencies produce persistent interference in the leading concurrence profile when Q R , − R ≠ 0 Q_{R,-R}\ne0 Q R , − R = 0 .
Define
Ω = 2 ∣ ω R ∣ , U = Q R R − Q − R , − R , \Omega=2|\omega_{R}|,\qquad U=Q_{RR}-Q_{-R,-R}, Ω = 2∣ ω R ∣ , U = Q R R − Q − R , − R ,
F = 2 ∣ Q R , − R ∣ , ν = F / Tr Q , (34) F=2|Q_{R,-R}|,\qquad\nu=F/\operatorname{Tr}Q, \tag*{(34)} F = 2∣ Q R , − R ∣ , ν = F / Tr Q , (34)
with Tr Q > 0 \operatorname{Tr}Q>0 Tr Q > 0 . Since Q Q Q is positive semidefinite, 0 ≤ ν ≤ 1 0\le\nu\le1 0 ≤ ν ≤ 1 . Taking v R = v v_{R}=v v R = v and v − R = w = J v v_{-R}=w=Jv v − R = w = J v gives
∣ Φ i j ( t ) ∣ = v i v j Tr Q [ 1 + ν cos ( Ω t + ϕ ) ] , (35) |\Phi_{ij}(t)|=v_{i}v_{j}\operatorname{Tr}Q[1+\nu\cos(\Omega t+\phi)], \tag*{(35)} ∣ Φ ij ( t ) ∣ = v i v j Tr Q [ 1 + ν cos ( Ω t + ϕ )] , (35)
∣ Φ i j ( t ) ∣ = v i v j U 2 + F 2 sin 2 ( Ω t + ϕ ′ ) , (36) |\Phi_{ij}(t)|=v_{i}v_{j}\sqrt{U^{2}+F^{2}\sin^{2}(\Omega t+\phi')}, \tag*{(36)} ∣ Φ ij ( t ) ∣ = v i v j U 2 + F 2 sin 2 ( Ω t + ϕ ′ ) , (36)
for vertices in the same and opposite parts, respectively. With ω R < 0 \omega_{R}<0 ω R < 0 from Eq. (72) , the phases are ϕ ′ = arg Q R , − R \phi'=\arg Q_{R,-R} ϕ ′ = arg Q R , − R and ϕ = ϕ ′ + π ( 1 − J i j ) / 2 \phi=\phi'+\pi(1-J_{ij})/2 ϕ = ϕ ′ + π ( 1 − J ij ) /2 , and they do not enter when Q R , − R = 0 Q_{R,-R}=0 Q R , − R = 0 .
The time dependence is identical for all pairs within each of three classes: those entirely in V + V_{+} V + , those entirely in V − V_{-} V − , and those connecting the two parts. Consequently, the Perron ratio in Eq. (33) holds between any two pairs within the same class, provided their common leading profile remains bounded away from zero. For same-part pairs this is guaranteed by Tr Q > F \operatorname{Tr}Q>F Tr Q > F , while for opposite-part pairs it follows from U ≠ 0 U\ne0 U = 0 .
If Q R R Q − R , − R = 0 Q_{RR}Q_{-R,-R}=0 Q R R Q − R , − R = 0 , positivity of Q Q Q forces F = 0 F=0 F = 0 , and the temporal factor reduces to Tr Q \operatorname{Tr}Q Tr Q for every pair. In this case, Eq. (33) holds between arbitrary pairs, including those belonging to different classes. When the common leading profile has recurring zeros, subleading contributions can become relevant near those times, and the asymptotic expansion alone does not establish the ratio limit.
Numerical graph comparisons
We propagate Eq. (2) exactly on the space with at most two total excitations and evaluate concurrence from the complete pair state. Unless stated otherwise, Δ = 0 \Delta= 0 Δ = 0 , g = 0.3 λ g = 0.3\lambda g = 0.3 λ , and η = 3 λ \eta= 3\lambda η = 3 λ . We use the graphs G 0 G_0 G 0 and G 0 + 24 G_0 + 24 G 0 + 24 of Figure 1 (c), the four-vertex path P 4 P_4 P 4 , the complete bipartite graph K 3 , 3 K_{3,3} K 3 , 3 , and the triangular prism. The comparisons between graphs use
∣ ψ i j ⟩ = 0.1 e 0.4 i ∣ Ω S ⟩ + 0.8 ∣ i ⟩ + ∣ j ⟩ 2 + 0.1 e 0.7 i ∣ i j ⟩ . (37) |\psi_{ij}\rangle= \sqrt{0.1}e^{0.4i}|\Omega_S\rangle+ \sqrt{0.8}\frac{|i\rangle+ |j\rangle}{\sqrt{2}} + \sqrt{0.1}e^{0.7i}|ij\rangle. \tag*{(37)} ∣ ψ ij ⟩ = 0.1 e 0.4 i ∣ Ω S ⟩ + 0.8 2 ∣ i ⟩ + ∣ j ⟩ + 0.1 e 0.7 i ∣ ij ⟩ . (37)
Appendix § 0.5.5 describes the numerical methods and independent checks.
Figure 2 (a,Figure 2 b) show that adding edge 24 lowers 2 γ 1 / λ 2\gamma_1/\lambda 2 γ 1 / λ from 0.004124 to 0.002984, but the concurrence curves nevertheless cross. Thus, a slower asymptotic decay does not guarantee greater concurrence at intermediate times. The pair ratios approach their respective Perron limits of 2.170 and 1.281. In Figure 2 (c,Figure 2 d), the 3-regular graphs K 3 , 3 K_{3,3} K 3 , 3 and the prism share R = 3 R = 3 R = 3 , a uniform Perron vector, and γ 1 = 0.001094 λ \gamma_1 = 0.001094\lambda γ 1 = 0.001094 λ . Nevertheless, their long-time behavior differs: the bipartite graph retains two least damped modes and exhibits an oscillatory concurrence profile, whereas the nonbipartite graph has a single least damped mode and approaches a constant rescaled profile.
Figure 2. Graph dependence of concurrence at Δ = 0 \Delta= 0 Δ = 0 , g = 0.3 λ g = 0.3\lambda g = 0.3 λ , and η = 3 λ \eta= 3\lambda η = 3 λ . Colored solid curves are numerical results, and dark dashed curves are asymptotic predictions, with the Perron limits in (b). (a,b) Preparation ∣ ψ 13 ⟩ |\psi_{13}\rangle ∣ ψ 13 ⟩ on G 0 G_0 G 0 and G 0 + 24 G_0 + 24 G 0 + 24 . Adding edge 24 reduces the asymptotic decay rate, but the concurrence curves cross near λ t = 446 \lambda t = 446 λ t = 446 . The ratios C 13 / C 14 C_{13}/C_{14} C 13 / C 14 approach different Perron limits. (c,d) Preparation ∣ ψ 12 ⟩ |\psi_{12}\rangle ∣ ψ 12 ⟩ on K 3 , 3 K_{3,3} K 3 , 3 and the prism. Both graphs have R = 3 R = 3 R = 3 and a uniform Perron vector, but their remaining adjacency spectra differ. Their rescaled concurrences approach periodic and constant profiles, respectively. On K 3 , 3 K_{3,3} K 3 , 3 , C 12 = 0 C_{12} = 0 C 12 = 0 for 795.39 ≤ λ t ≤ 796.42 795.39 \le\lambda t \le796.42 795.39 ≤ λ t ≤ 796.42 . The insets in (d) share one vertical scale and show this interval and a positive late minimum.
Late-time monotonicity and revival thresholds
Eventual strict decrease on nonbipartite graphs
At resonance, a nonbipartite graph has a unique least damped Perron mode. When Q R R > 0 Q_{RR} > 0 Q R R > 0 , the leading concurrence of every pair is a positive constant times e − 2 γ 1 t e^{-2\gamma_1 t} e − 2 γ 1 t . In fact, the exact concurrence is eventually strictly decreasing for every pair.
Establishing strict decrease requires control not only of the remainder but also of its variation in time. For fixed m > 0 m > 0 m > 0 , Appendix § 0.5.4 proves that there are constants K , T > 0 K, T > 0 K , T > 0 such that
∣ R i j ( t + h ) − R i j ( t ) ∣ ≤ K e − ( 2 γ 1 + β ) t ∣ h ∣ (38) |R_{ij}(t+h)-R_{ij}(t)| \leq K e^{-(2\gamma_1+\beta)t}|h| \tag*{(38)} ∣ R ij ( t + h ) − R ij ( t ) ∣ ≤ K e − ( 2 γ 1 + β ) t ∣ h ∣ (38)
whenever t ≥ T t \geq T t ≥ T , λ ∣ h ∣ ≤ 1 \lambda|h| \leq1 λ ∣ h ∣ ≤ 1 , and ∣ Φ i j ∣ ≥ m |\Phi_{ij}| \geq m ∣ Φ ij ∣ ≥ m on the interval between t t t and t + h t+h t + h . When, in addition, the derivative of the leading concurrence is at most − δ e − 2 γ 1 t -\delta e^{-2\gamma_1t} − δ e − 2 γ 1 t for some δ > 0 \delta> 0 δ > 0 , this bound ensures that the exact concurrence is eventually strictly decreasing. For a nonbipartite graph with Q R R > 0 Q_{RR} > 0 Q R R > 0 , Φ i j = Q R R v i v j \Phi_{ij} = Q_{RR}v_i v_j Φ ij = Q R R v i v j is a positive constant, so both conditions hold.
Revival dichotomy on bipartite graphs
For bipartite graphs at resonance, the two least damped modes produce an oscillatory concurrence profile when Q R , − R ≠ 0 Q_{R,-R} \ne0 Q R , − R = 0 . Whether this produces arbitrarily late revivals depends on whether the oscillations can overcome the exponential decay envelope. Define the critical visibility
ν c r i t = 2 γ 1 4 γ 1 2 + Ω 2 . (39) \nu_{\mathrm{crit}} = \frac{2\gamma_1}{\sqrt{4\gamma_1^2+\Omega^2}}. \tag*{(39)} ν crit = 4 γ 1 2 + Ω 2 2 γ 1 . (39)
Equations (B10) and (B2) give ν c r i t = γ 1 ( λ − γ 1 ) / g \nu_{\mathrm{crit}} = \sqrt{\gamma_1(\lambda-\gamma_1)}/g ν crit = γ 1 ( λ − γ 1 ) / g . Since γ 1 = γ ( η R ) < λ / 2 \gamma_1 = \gamma(\eta R) < \lambda/2 γ 1 = γ ( η R ) < λ /2 and χ ( λ − x ) \chi(\lambda-x) χ ( λ − x ) increases for x < λ / 2 x < \lambda/2 x < λ /2 , Eq. (51) shows that ν c r i t \nu_{\mathrm{crit}} ν crit decreases strictly as η R \eta R η R increases at fixed g g g and λ \lambda λ .
Theorem 1 (Revival dichotomy). Assume that G G G is bipartite with parts V + V_+ V + and V − V_- V − , Δ = 0 \Delta= 0 Δ = 0 , and Tr Q > 0 \operatorname{Tr} Q > 0 Tr Q > 0 . For a pair of vertices i < j i < j i < j , the following statements hold:
If i , j ∈ V + i,j \in V_+ i , j ∈ V + or i , j ∈ V − i,j \in V_- i , j ∈ V − , then ν > ν c r i t \nu> \nu_{\mathrm{crit}} ν > ν crit gives arbitrarily late revivals and ν < ν c r i t \nu< \nu_{\mathrm{crit}} ν < ν crit gives eventual strict decrease.
Suppose i i i and j j j lie in opposite parts. If U ≠ 0 U \ne0 U = 0 , arbitrarily late revivals occur when
F 2 U 2 > 2 ν c r i t 1 − ν c r i t . (40) \frac{F^2}{U^2} > \frac{2\nu_{\mathrm{crit}}}{1-\nu_{\mathrm{crit}}}. \tag*{(40)} U 2 F 2 > 1 − ν crit 2 ν crit . (40) The reverse strict inequality gives eventual strict decrease. If U = 0 < F U = 0 < F U = 0 < F , arbitrarily late revivals occur. If U = F = 0 U = F = 0 U = F = 0 , then F i j ( t ) ≡ 0 \mathcal{F}_{ij}(t) \equiv0 F ij ( t ) ≡ 0 .
These conditions show that bipartiteness alone does not guarantee revivals: their occurrence also depends on the coherence and relative populations of the two slow modes. For example, the normalized one-excitation state ( ∣ v ⟩ + ∣ w ⟩ ) / 2 (|v\rangle+ |w\rangle)/\sqrt{2} ( ∣ v ⟩ + ∣ w ⟩) / 2 , with ∣ v ⟩ = ∑ r v r ∣ r ⟩ |v\rangle= \sum_r v_r|r\rangle ∣ v ⟩ = ∑ r v r ∣ r ⟩ and ∣ w ⟩ = ∑ r w r ∣ r ⟩ |w\rangle= \sum_r w_r|r\rangle ∣ w ⟩ = ∑ r w r ∣ r ⟩ , satisfies ν = 1 \nu= 1 ν = 1 and U = 0 < F U = 0 < F U = 0 < F , so every pair has arbitrarily late revivals. The revival thresholds follow from comparing the maximum logarithmic growth rate of ∣ Φ i j ∣ |\Phi_{ij}| ∣ Φ ij ∣ with the exponential decay rate 2 γ 1 2\gamma_1 2 γ 1 . The dichotomy is established in Appendix § 0.5.4 , although the threshold equality cases remain unclassified. Similarly, a vanishing leading profile does not imply vanishing exact concurrence, since subleading modes may determine its late-time behavior.
Loss feeding
To isolate the effect of loss feeding, we fix the initial one-excitation block on P 4 P_4 P 4 and vary only the two-excitation and vacuum populations,
ρ S ( 0 ) = 0.15 ∣ v ⟩ ⟨ v ∣ + p 2 ∣ 13 ⟩ ⟨ 13 ∣ + ( 0.85 − p 2 ) ∣ Ω S ⟩ ⟨ Ω S ∣ , (41) \rho_S(0) = 0.15|v\rangle\langle v| + p_2|13\rangle\langle13| + (0.85-p_2)|\Omega_S\rangle\langle\Omega_S|, \tag*{(41)} ρ S ( 0 ) = 0.15∣ v ⟩ ⟨ v ∣ + p 2 ∣13 ⟩ ⟨ 13∣ + ( 0.85 − p 2 ) ∣ Ω S ⟩ ⟨ Ω S ∣ , (41)
where 0 ≤ p 2 ≤ 0.85 0 \leq p_2 \leq0.85 0 ≤ p 2 ≤ 0.85 and ∣ v ⟩ = ∑ r v r ∣ r ⟩ |v\rangle= \sum_r v_r|r\rangle ∣ v ⟩ = ∑ r v r ∣ r ⟩ . The initial one-excitation term contributes only to Q R R Q_{RR} Q R R , whereas loss feeding contributes to every entry of Q Q Q . Increasing p 2 p_2 p 2 can therefore change the interference between the least damped modes without changing the initial one-excitation state.
For p 2 = 0.1 p_2 = 0.1 p 2 = 0.1 and 0.8 0.8 0.8 , ν \nu ν takes the values 0.0959 0.0959 0.0959 and 0.3082 0.3082 0.3082 , respectively, on opposite sides of ν c r i t = 0.2004 \nu_{\mathrm{crit}} = 0.2004 ν crit = 0.2004 . For the pair 34 in opposite parts, U U U is independent of p 2 p_2 p 2 , while F 2 / U 2 F^2/U^2 F 2 / U 2 changes from 0.0148 0.0148 0.0148 to 0.9485 0.9485 0.9485 , crossing the threshold 2 ν c r i t / ( 1 − ν c r i t ) = 0.5012 2\nu_{\mathrm{crit}/(1-\nu_{\mathrm{crit}})} = 0.5012 2 ν crit / ( 1 − ν crit ) = 0.5012 . The two thresholds are crossed at p 2 = 0.2963 p_2 = 0.2963 p 2 = 0.2963 and 0.5815 0.5815 0.5815 , respectively. Figure 3 shows the corresponding changes in the exact concurrence. Without the loss-feeding integral in Eq. (27) , neither transition would be predicted.
Figure 3. Loss from two excitations changes both revival conditions for a fixed graph and a fixed initial one-excitation block. The state is Eq. (41) on P 4 P_4 P 4 , with Δ = 0 \Delta=0 Δ = 0 , g = 0.3 λ g=0.3\lambda g = 0.3 λ , and η = 3 λ \eta=3\lambda η = 3 λ . (a) ν \nu ν and its threshold ν c r i t \nu_{\mathrm{crit}} ν crit as the initial two-excitation population p 2 p_2 p 2 varies. (b,c) Numerical C 13 C_{13} C 13 and C 34 C_{34} C 34 (solid) and asymptotic predictions (dashed) for p 2 = 0.1 p_2=0.1 p 2 = 0.1 and 0.8 0.8 0.8 , below and above the thresholds of Theorem 1 , respectively. In (b), C 13 = 0 C_{13}=0 C 13 = 0 for λ t ≤ 35.93 \lambda t\leq35.93 λ t ≤ 35.93 and 58.18 ≤ λ t ≤ 84.61 58.18\leq\lambda t\leq84.61 58.18 ≤ λ t ≤ 84.61 at p 2 = 0.1 p_2=0.1 p 2 = 0.1 , and for λ t ≤ 37.42 \lambda t\leq37.42 λ t ≤ 37.42 and 55.68 ≤ λ t ≤ 98.78 55.68\leq\lambda t\leq98.78 55.68 ≤ λ t ≤ 98.78 at p 2 = 0.8 p_2=0.8 p 2 = 0.8 . In (c), C 34 = 0 C_{34}=0 C 34 = 0 for 7.73 ≤ λ t ≤ 56.95 7.73\leq\lambda t\leq56.95 7.73 ≤ λ t ≤ 56.95 at p 2 = 0.1 p_2=0.1 p 2 = 0.1 , and for 53.03 ≤ λ t ≤ 71.34 53.03\leq\lambda t\leq71.34 53.03 ≤ λ t ≤ 71.34 and 137.19 ≤ λ t ≤ 212.87 137.19\leq\lambda t\leq212.87 137.19 ≤ λ t ≤ 212.87 at p 2 = 0.8 p_2=0.8 p 2 = 0.8 .
Late revival times
For a bipartite graph at resonance, consider a pair in the same part with Tr Q > 0 \operatorname{Tr} Q > 0 Tr Q > 0 and ν c r i t < ν < 1 \nu_{\mathrm{crit}} < \nu< 1 ν crit < ν < 1 . The leading concurrence has local maxima at
t n = ϑ − χ − φ + 2 π n Ω , ϑ = arctan Ω 2 γ 1 , χ = arccos ( − ν c r i t ν ) . (42) t_n = \frac{\vartheta- \chi- \varphi+ 2\pi n}{\Omega},
\qquad
\vartheta= \arctan\frac{\Omega}{2\gamma_1},
\qquad
\chi= \arccos\left(-\frac{\nu_{\mathrm{crit}}}{\nu}\right). \tag*{(42)} t n = Ω ϑ − χ − φ + 2 π n , ϑ = arctan 2 γ 1 Ω , χ = arccos ( − ν ν crit ) . (42)
For a pair in opposite parts with F 2 ( 1 − ν c r i t ) > 2 ν c r i t U 2 F^2(1-\nu_{\mathrm{crit}}) > 2\nu_{\mathrm{crit}}U^2 F 2 ( 1 − ν crit ) > 2 ν crit U 2 , the leading concurrence has local maxima at
t n ′ = ϑ + χ ′ − 2 φ ′ + 2 π n 2 Ω , χ ′ = arccos [ ν c r i t ( 1 + 2 U 2 F 2 ) ] . (43) t'_n = \frac{\vartheta+ \chi' - 2\varphi' + 2\pi n}{2\Omega},
\qquad
\chi' = \arccos\left[\nu_{\mathrm{crit}}\left(1+\frac{2U^2}{F^2}\right)\right]. \tag*{(43)} t n ′ = 2Ω ϑ + χ ′ − 2 φ ′ + 2 π n , χ ′ = arccos [ ν crit ( 1 + F 2 2 U 2 ) ] . (43)
This condition is equivalent to Eq. (40) when U ≠ 0 U \ne0 U = 0 and holds whenever U = 0 < F U = 0 < F U = 0 < F . For every sufficiently large integer n n n , the exact concurrence has at least one local maximum at t n + O ( e − β t n ) t_n+O(e^{-\beta t_n}) t n + O ( e − β t n ) for a pair in the same part and at t n ′ + O ( e − β t n ′ ) t'_n+O(e^{-\beta t'_n}) t n ′ + O ( e − β t n ′ ) for a pair in opposite parts. Appendix D proves these location bounds using Eq. (38) .
For the preparation in Eq. (41) with p 2 = 0.8 p_2=0.8 p 2 = 0.8 , pair 13 satisfies ν c r i t < ν < 1 \nu_{\mathrm{crit}}<\nu<1 ν crit < ν < 1 , while pair 34 satisfies F 2 ( 1 − ν c r i t ) > 2 ν c r i t U F^2(1-\nu_{\mathrm{crit}})>2\nu_{\mathrm{crit}}U F 2 ( 1 − ν crit ) > 2 ν crit U . Figure 4 compares the predicted times from Eqs. (42) and (43) with the local maxima t n e x t_n^{\mathrm{ex}} t n ex and t n ′ e x t_n^{\prime\mathrm{ex}} t n ′ ex of the exact concurrence. The first predicted maximum of C 34 C_{34} C 34 shown in Figure 4 falls within the interval 137.19 ≤ λ t ≤ 212.87 137.19\leq\lambda t\leq212.87 137.19 ≤ λ t ≤ 212.87 , where the exact concurrence vanishes. For 150 ≤ λ t ≤ 3000 150\leq\lambda t\leq3000 150 ≤ λ t ≤ 3000 , the numerical maxima correspond one to one with the predicted times.
Figure 4. Late revival times for Eq. (41) on P 4 P_4 P 4 with p 2 = 0.8 p_2=0.8 p 2 = 0.8 , Δ = 0 \Delta=0 Δ = 0 , g = 0.3 λ g=0.3\lambda g = 0.3 λ , and η = 3 λ \eta=3\lambda η = 3 λ . (a,b) Numerical C 13 C_{13} C 13 and C 34 C_{34} C 34 (solid) and asymptotic predictions (dashed), as in Figure 3 (b,c). Filled markers show the exact local maxima at t n e x t_n^{\rm ex} t n ex and t n ′ e x t_n^{\prime\rm ex} t n ′ ex , and open markers show the leading concurrence at t n t_n t n and t n ′ t_n' t n ′ from Eqs. (42) and (43) . Markers are shown for λ t ≥ 150 \lambda t \ge150 λ t ≥ 150 . In (b), C 34 = 0 C_{34}=0 C 34 = 0 at the first open marker. (c) Shifts of the exact maxima from the predicted times. Dotted reference lines are proportional to e − ( γ 2 − 2 γ 1 ) t e^{-(\gamma_2-2\gamma_1)t} e − ( γ 2 − 2 γ 1 ) t and pass through the last marker of each set, so only their slope is fixed.
The computed shifts t n e x − t n t_n^{\mathrm{ex}}-t_n t n ex − t n and t n ′ e x − t n ′ t_n^{\prime\mathrm{ex}}-t'_n t n ′ ex − t n ′ are positive and exhibit approximately geometric decay. Their successive ratios are consistent with 0.3894 0.3894 0.3894 for C 13 C_{13} C 13 and 0.6240 0.6240 0.6240 for C 34 C_{34} C 34 , obtained from e − 2 π ( γ 2 − 2 γ 1 ) / Ω e^{-2\pi(\gamma_2-2\gamma_1)/\Omega} e − 2 π ( γ 2 − 2 γ 1 ) /Ω and e − π ( γ 2 − 2 γ 1 ) / Ω e^{-\pi(\gamma_2-2\gamma_1)/\Omega} e − π ( γ 2 − 2 γ 1 ) /Ω , respectively, where γ 2 − 2 γ 1 = 0.005324 λ \gamma_2-2\gamma_1=0.005324\lambda γ 2 − 2 γ 1 = 0.005324 λ .
Odd cycles and detuning
At resonance, adding an edge i j ij ij gives R ′ ≥ v T A ′ v = R + 2 v i v j > R R'\geq v^{\mathsf{T}}A'v=R+2v_iv_j>R R ′ ≥ v T A ′ v = R + 2 v i v j > R for the original unit Perron vector, where primes refer to the modified graph. Since γ ( D ) \gamma(D) γ ( D ) decreases with ∣ D ∣ \lvert D\rvert ∣ D ∣ , the slowest decay rate satisfies γ 1 ′ < γ 1 \gamma'_1<\gamma_1 γ 1 ′ < γ 1 at every finite coupling. Adding an edge within one part of a bipartite graph creates an odd cycle and leaves a single least damped channel, whereas an edge joining opposite parts preserves bipartiteness and the two least damped channels.
Let Q ′ Q' Q ′ and C i j ′ C'_{ij} C ij ′ denote the matrix in Eq. (27) and the concurrence for the modified graph with the same initial state. If the modified graph is nonbipartite and Q R ′ / R ′ ′ > 0 Q'_{R'/R'}>0 Q R ′ / R ′ ′ > 0 , its concurrence has a positive leading coefficient and decays more slowly than the concurrence on the original graph. Consequently, at sufficiently late times, the modified graph has greater concurrence than the original graph for every qubit pair, C i j ′ ( t ) > C i j ( t ) C'_{ij}(t)>C_{ij}(t) C ij ′ ( t ) > C ij ( t ) .
Uniform detuning provides another way to remove the degeneracy of the least damped channels without changing the graph. By Proposition 3 of Appendix B, a bipartite graph has γ 2 > 2 γ 1 \gamma_2>2\gamma_1 γ 2 > 2 γ 1 at every uniform detuning. At nonzero detuning, let γ \gamma γ label the unique least damped one-excitation mode. If Q a a > 0 Q_{aa}>0 Q aa > 0 , the leading profile is a positive constant, and the concurrence is eventually strictly decreasing (Appendix C).
For Δ > 0 \Delta>0 Δ > 0 , write γ + = γ ( η R + Δ ) \gamma_+=\gamma(\eta R+\Delta) γ + = γ ( η R + Δ ) and γ − = γ ( η R − Δ ) \gamma_-=\gamma(\eta R-\Delta) γ − = γ ( η R − Δ ) for the rates of the channels labeled by R R R and − R -R − R . Their difference satisfies
δ γ = γ − − γ + = 2 ∣ d γ d D ( η R ) ∣ Δ + O ( Δ 3 ) (44) \delta\gamma=\gamma_- - \gamma_+ = 2\left|\frac{d\gamma}{dD}(\eta R)\right|\Delta+O(\Delta^3) \tag*{(44)} δ γ = γ − − γ + = 2 d D d γ ( η R ) Δ + O ( Δ 3 ) (44)
as Δ → 0 + \Delta\to0^+ Δ → 0 + . For sufficiently small detuning, γ − < 2 γ + ≤ γ 2 \gamma_-<2\gamma_+\leq\gamma_2 γ − < 2 γ + ≤ γ 2 , so the integrals in Eq. (27) , evaluated at z R z_R z R and z − R z_{-R} z − R , converge. The cross term between the two channels decays as e − ( γ + + γ − ) t e^{-(\gamma_++\gamma_-)t} e − ( γ + + γ − ) t , compared with e − 2 γ + t e^{-2\gamma_+t} e − 2 γ + t for the population of the less damped channel. Its relative contribution therefore decays as e − δ γ t e^{-\delta\gamma t} e − δ γ t , suppressing the interference responsible for persistent oscillations at resonance.
Figure 5 (b,c) compares P 4 P_4 P 4 at Δ = 0 \Delta=0 Δ = 0 and Δ = λ \Delta=\lambda Δ = λ , with g = 0.3 λ g=0.3\lambda g = 0.3 λ , η = 3 λ \eta=3\lambda η = 3 λ , and preparation ∣ ψ 13 ⟩ \lvert\psi_{13}\rangle ∣ ψ 13 ⟩ . At Δ = λ \Delta=\lambda Δ = λ , M = { ( 1 + 5 ) / 2 } \mathcal{M}=\{(1+\sqrt{5})/2\} M = {( 1 + 5 ) /2 } and Q R R > 0 Q_{RR}>0 Q R R > 0 , so the concurrence is eventually strictly decreasing, and the rescaled concurrence approaches 0.17630054 0.17630054 0.17630054 . Revivals nevertheless occur at finite times; after the last local maximum at λ t = 822.11 \lambda t = 822.11 λ t = 822.11 , the computed curve decreases up to λ t = 1700 \lambda t = 1700 λ t = 1700 .
Figure 5. Distinct effects of uniform detuning. Solid curves are numerical results and dashed curves are asymptotic predictions. (a) On K 4 K_4 K 4 , g = η = λ g=\eta=\lambda g = η = λ and Δ = − 2 λ \Delta=-2\lambda Δ = − 2 λ give γ 2 = 2 γ 1 \gamma_2=2\gamma_1 γ 2 = 2 γ 1 . The rescaled C 12 C_{12} C 12 for Eq. (46) with p = 0.01 p=0.01 p = 0.01 approaches the coefficient in Eq. (47) , which includes coherence between the vacuum and two qubit excitations. (b,c) Original and rescaled C 13 C_{13} C 13 on P 4 P_4 P 4 for Δ = 0 \Delta=0 Δ = 0 and λ \lambda λ , with g = 0.3 λ g=0.3\lambda g = 0.3 λ , η = 3 λ \eta=3\lambda η = 3 λ , and preparation ∣ ψ 13 ⟩ |\psi_{13}\rangle ∣ ψ 13 ⟩ . Each curve in (c) uses its own γ 1 \gamma_1 γ 1 . The resonant curve approaches a periodic profile, while the detuned curve approaches a constant. At resonance, C 13 C_{13} C 13 remains positive, and its minimum C 13 = 4.9 × 10 − 8 C_{13}=4.9\times10^{-8} C 13 = 4.9 × 1 0 − 8 near λ t = 1683 \lambda t=1683 λ t = 1683 lies below the range of (b). At Δ = λ \Delta=\lambda Δ = λ , C 13 = 0 C_{13}=0 C 13 = 0 for 87.91 ≤ λ t ≤ 96.43 87.91\leq\lambda t\leq96.43 87.91 ≤ λ t ≤ 96.43 . Detuning separates the amplitude decay rates of the R R R and − R -R − R channels, while finite-time revivals remain in (b).
Equal decay rates with degenerate least damped modes
By Proposition 1, equality γ 2 = 2 γ 1 \gamma_2 = 2\gamma_1 γ 2 = 2 γ 1 requires Δ ≤ − η ( R + θ − ) / 2 \Delta\le-\eta(R+\theta_-)/2 Δ ≤ − η ( R + θ − ) /2 and a repeated eigenvalue θ − \theta_- θ − , excluding bipartite graphs. In this case, contributions from both excitation sectors can survive at the same leading order, and Φ i j \Phi_{ij} Φ ij alone need not determine concurrence.
For the complete graph K 4 K_4 K 4 in Figure 1 , setting g = η = λ = 1 g=\eta=\lambda=1 g = η = λ = 1 and Δ = − 2 \Delta=-2 Δ = − 2 yields
γ 2 = 2 γ 1 = 1 − 3 5 − 6 . (45) \gamma_2=2\gamma_1=1-\sqrt{3\sqrt{5}-6}. \tag*{(45)} γ 2 = 2 γ 1 = 1 − 3 5 − 6 . (45)
Here M = { − 1 } \mathcal{M}=\{-1\} M = { − 1 } and rank E − 1 = 3 \operatorname{rank}E_{-1}=3 rank E − 1 = 3 . The eigenvalue of the least damped two-excitation modes is 2 z 2z 2 z , where z z z is the root of z 2 + ( 3 + i ) z − 1 = 0 z^2+(3+i)z-1=0 z 2 + ( 3 + i ) z − 1 = 0 of smaller modulus. Its algebraic and geometric multiplicities are both two. Appendix A 3 proves these statements using explicit eigenvectors and Lemmas 1 and 2.
In this example, the three least damped one-excitation modes have frequency ω = Re z \omega=\operatorname{Re}z ω = Re z , and the two least damped two-excitation modes have frequency 2 ω 2\omega 2 ω . Consequently, X i j ( 0 ) ( t ) X_{ij}^{(0)}(t) X ij ( 0 ) ( t ) evolves by local phase rotations, leaving its concurrence constant for every initial state (Appendix C 3).
Consider the qubit preparation
∣ s ⟩ = 1 2 ( ∣ 12 ⟩ − ∣ 13 ⟩ − ∣ 24 ⟩ + ∣ 34 ⟩ ) , ∣ ψ p ⟩ = 1 − p ∣ Ω S ⟩ + p ∣ s ⟩ , 0 < p < 1. (46) \begin{aligned}
|s\rangle&= \frac{1}{2}(|12\rangle-|13\rangle-|24\rangle+|34\rangle),\\
|\psi_p\rangle&= \sqrt{1-p}|\Omega_S\rangle+\sqrt{p}|s\rangle,\qquad0<p<1. \tag*{(46)}
\end{aligned} ∣ s ⟩ ∣ ψ p ⟩ = 2 1 ( ∣12 ⟩ − ∣13 ⟩ − ∣24 ⟩ + ∣34 ⟩) , = 1 − p ∣ Ω S ⟩ + p ∣ s ⟩ , 0 < p < 1. (46)
Put α = ( 1 + z 2 ) − 1 \alpha=(1+z^2)^{-1} α = ( 1 + z 2 ) − 1 and a 0 = ∣ α ∣ 2 / 2 a_0=|\alpha|^2/2 a 0 = ∣ α ∣ 2 /2 . For the pair 12, the matrix X 12 ( 0 ) X_{12}^{(0)} X 12 ( 0 ) in Eq. (79) has only diagonal and corner entries. Appendix A 5 evaluates these entries and gives, for 0 < p < 1 / 2 0<p<1/2 0 < p < 1/2 ,
C 12 ( t ) = 2 a 0 [ p ( 1 − p ) − p ] e − 2 γ 1 t + O ( e − ( 2 γ 1 + β ) t ) . (47) C_{12}(t)=2a_0[\sqrt{p(1-p)}-p]e^{-2\gamma_1t}+O(e^{-(2\gamma_1+\beta)t}). \tag*{(47)} C 12 ( t ) = 2 a 0 [ p ( 1 − p ) − p ] e − 2 γ 1 t + O ( e − ( 2 γ 1 + β ) t ) . (47)
The coefficient is positive, and the exact concurrence is eventually strictly decreasing (Appendix A § 0.7.5 ). The leading contribution arises from coherence between the vacuum and two-qubit excitation sectors, together with populations generated by loss; the coherence between ∣ 10 ⟩ |10\rangle ∣10 ⟩ and ∣ 01 ⟩ |01\rangle ∣01 ⟩ vanishes.
The same preparation also illustrates why a vanishing leading profile does not determine whether exact concurrence survives. At p = 1 / 2 p=1/2 p = 1/2 , the leading profile vanishes, but the exact concurrence is ∣ f ( t ) ∣ 4 / 4 |f(t)|^4/4 ∣ f ( t ) ∣ 4 /4 , where f ( t ) f(t) f ( t ) denotes the qubit amplitude evolving from an initially excited qubit in the D = − 3 D=-3 D = − 3 channel. It remains positive at late times and is asymptotic to ∣ α ∣ 4 e − 4 γ 1 t / 4 |\alpha|^4e^{-4\gamma_1t}/4 ∣ α ∣ 4 e − 4 γ 1 t /4 . For p > 1 / 2 p>1/2 p > 1/2 , concurrence is instead exactly zero at all sufficiently late times. Appendix A 5 derives both results from the full time-dependent pair state.
Figure 5 (a) tests Eq. (47) at p = 0.01 p=0.01 p = 0.01 , where the rescaled concurrence approaches C ( X 12 ( 0 ) ) = 0.07752961 C(X_{12}^{(0)})=0.07752961 C ( X 12 ( 0 ) ) = 0.07752961 , computed from Eq. (79) without fitting the trajectory.
Discussion and conclusions
We have shown that coupling the local environments of noninteracting qubits into a network changes not only the rate at which entanglement decays, but also its spatial distribution and temporal behavior. The underlying mechanism is the formation of collective reservoir modes whose frequencies and decay rates depend on the graph spectrum. As the magnitude of the effective detuning increases, the least damped modes have less weight in the lossy memory degrees of freedom and consequently decay more slowly. Network structure therefore determines which collective excitations survive longest, while their populations and mutual coherence determine the entanglement that remains at late times.
At resonance, this mechanism leads to a distinction between bipartite and nonbipartite networks. On nonbipartite graphs, the Perron mode is the unique least damped channel, and when populated, its spatial structure fixes the asymptotic ratios of pair concurrence. The concurrence of every pair then becomes strictly decreasing at sufficiently late times. Bipartite graphs instead retain two equally damped channels with different frequencies. Their interference produces an oscillatory leading concurrence profile, but arbitrarily late revivals occur when these oscillations overcome the exponential decay envelope. The resulting thresholds distinguish arbitrarily late revivals from eventual monotonic decay. Thus, networks with identical spectral radii, and hence identical slowest decay rates, can exhibit qualitatively different entanglement dynamics.
The role of dissipation becomes particularly important when the initial state contains two excitations. Although the two-excitation sector decays faster than the long-lived one-excitation modes at nonnegative detuning, loss transfers population into these modes and changes their relative weights and coherence. Consequently, changing the initial two-excitation population can move the system between revival and monotonic-decay regimes without changing the network or the initial one-excitation state. This effect is missed if the dynamics is restricted to the one-excitation sector. More generally, our bound
on the sector decay rates establishes when the leading concurrence is determined by one-excitation coherence and identifies the exceptional negative-detuning regime in which both sectors contribute at the same asymptotic order.
The model also suggests possible experimental tests in coupled-cavity and circuit-QED networks, where local qubit–resonator interactions, coherent hopping, and dissipation can be engineered. Small networks with different connectivity could be used to compare the spatial distribution and temporal evolution of pair concurrence. A particularly direct test would be to vary the initial two-excitation population while keeping the network and initial one-excitation block fixed, and examine the resulting change in revival behavior. Observing these effects would require resolving concurrence before additional decoherence processes suppress the surviving collective modes. The robustness of the predicted signatures to qubit relaxation, dephasing, and parameter disorder remains to be established.
Our analysis is exact for finite connected graphs with uniform couplings, local zero-temperature losses, and initial states containing at most two excitations. These assumptions isolate the effects of collective reservoir structure from other sources of decoherence. In particular, arbitrarily late revivals do not imply sustained entanglement, since their amplitudes decrease exponentially. What persists is the interference between the least damped collective modes, whose existence and spatial structure are determined by the network. This distinction provides a basis for understanding how reservoir connectivity influences entanglement dynamics, independently of direct interactions between the qubits.
Spectral channels
Ordering of channel decay rates
The roots z z z , z ′ z' z ′ of z 2 − ( D − i λ ) z − g 2 = 0 z^2-(D-i\lambda)z-g^2=0 z 2 − ( D − iλ ) z − g 2 = 0 have product z z ′ = − g 2 zz'=-g^2 z z ′ = − g 2 . If their moduli are equal, then ∣ z ∣ = ∣ z ′ ∣ = g |z|=|z'|=g ∣ z ∣ = ∣ z ′ ∣ = g and z ′ = − z z'=-z z ′ = − z , which implies D = Re ( z + z ′ ) = 0 D=\operatorname{Re}(z+z')=0 D = Re ( z + z ′ ) = 0 . Thus, for D ≠ 0 D\ne0 D = 0 , there is a unique root of smaller modulus, denoted by z z z , and ∣ z ∣ < g |z|<g ∣ z ∣ < g . The root equation gives
D = Re z ( 1 − g 2 ∣ z ∣ 2 ) . (48) D=\operatorname{Re}z\left(1-\frac{g^2}{|z|^2}\right). \tag*{(48)} D = Re z ( 1 − ∣ z ∣ 2 g 2 ) . (48)
Hence Re z < 0 \operatorname{Re}z<0 Re z < 0 for D > 0 D>0 D > 0 .
For D ≠ 0 D\ne0 D = 0 the roots are simple, and differentiating the root equation gives
d z d D = z 2 z − D + i λ , (49) \frac{dz}{dD}=\frac{z}{2z-D+i\lambda}, \tag*{(49)} d D d z = 2 z − D + iλ z , (49)
and therefore
d ∣ z ∣ 2 d D = 2 ∣ z ∣ 2 ( 2 Re z − D ) ∣ 2 z − D + i λ ∣ 2 < 0 , D > 0. (50) \frac{d|z|^2}{dD}=\frac{2|z|^2(2\operatorname{Re}z-D)}{|2z-D+i\lambda|^2}<0,\qquad D>0. \tag*{(50)} d D d ∣ z ∣ 2 = ∣2 z − D + iλ ∣ 2 2∣ z ∣ 2 ( 2 Re z − D ) < 0 , D > 0. (50)
Since the right-hand side of Eq. (16) is strictly increasing in ∣ z ∣ 2 |z|^2 ∣ z ∣ 2 , the smaller decay rate γ ( D ) \gamma(D) γ ( D ) strictly decreases for D > 0 D>0 D > 0 .
The transformation D ↦ − D D\mapsto-D D ↦ − D maps each root to − z -z − z and preserves its modulus and decay rate, so γ ( − D ) = γ ( D ) \gamma(-D)=\gamma(D) γ ( − D ) = γ ( D ) . The minimum decay rate is continuous at D = 0 D=0 D = 0 . Thus
γ ( D 1 ) > γ ( D 2 ) , 0 ≤ D 1 < D 2 . (51) \gamma(D_1)>\gamma(D_2),\qquad0\le D_1<D_2. \tag*{(51)} γ ( D 1 ) > γ ( D 2 ) , 0 ≤ D 1 < D 2 . (51)
Least damped modes at arbitrary detuning
Since R = max ϑ ∈ spec A ϑ R=\max_{\vartheta\in\operatorname{spec}A}\vartheta R = max ϑ ∈ spec A ϑ and η > 0 \eta>0 η > 0 , every θ ∈ spec A \theta\in\operatorname{spec}A θ ∈ spec A satisfies D θ − ≤ D θ ≤ D R D_{\theta_-}\le D_\theta\le D_R D θ − ≤ D θ ≤ D R , and ∣ D θ ∣ < max { ∣ D θ − ∣ , ∣ D R ∣ } |D_\theta|<\max\{|D_{\theta_-}|,|D_R|\} ∣ D θ ∣ < max { ∣ D θ − ∣ , ∣ D R ∣ } for θ − < θ < R \theta_-<\theta<R θ − < θ < R . Since γ ( D ) \gamma(D) γ ( D ) strictly decreases with ∣ D ∣ |D| ∣ D ∣ , the eigenvalues in M \mathcal{M} M are determined by
D R 2 − D θ − 2 = η ( R − θ − ) [ 2 Δ + η ( R + θ − ) ] . (52) D_R^2-D_{\theta_-}^2=\eta(R-\theta_-)\left[2\Delta+\eta(R+\theta_-)\right]. \tag*{(52)} D R 2 − D θ − 2 = η ( R − θ − ) [ 2Δ + η ( R + θ − ) ] . (52)
The factor η ( R − θ − ) \eta(R-\theta_-) η ( R − θ − ) is positive, which yields Eq. (19) .
At Δ = − η ( R + θ − ) / 2 \Delta=-\eta(R+\theta_-)/2 Δ = − η ( R + θ − ) /2 , we have D R = − D θ − D_R=-D_{\theta_-} D R = − D θ − , so the root symmetry gives z θ − = − z R ‾ z_{\theta_-}=-\overline{z_R} z θ − = − z R .
For θ ∈ M \theta\in\mathcal{M} θ ∈ M ,
∣ D θ ∣ = max { ∣ D θ − ∣ , ∣ D R ∣ } ≥ η 2 ( R − θ − ) > 0. (53) |D_\theta|=\max\{|D_{\theta_-}|,|D_R|\}\ge\frac{\eta}{2}(R-\theta_-)>0. \tag*{(53)} ∣ D θ ∣ = max { ∣ D θ − ∣ , ∣ D R ∣ } ≥ 2 η ( R − θ − ) > 0. (53)
Thus h θ h_\theta h θ has distinct eigenvalues, ∣ z θ ∣ < g |z_\theta|<g ∣ z θ ∣ < g , and γ 1 < λ / 2 \gamma_1<\lambda/2 γ 1 < λ /2 . Every eigenvalue z z z of h θ h_\theta h θ satisfies
D θ = z − g 2 z + i λ , (54) D_\theta=z-\frac{g^2}{z}+i\lambda, \tag*{(54)} D θ = z − z g 2 + iλ , (54)
so channels with different θ \theta θ cannot share an eigenvalue. An orthonormal basis of ran E θ \operatorname{ran}E_\theta ran E θ gives rank E θ \operatorname{rank}E_\theta rank E θ mutually orthogonal invariant subspaces S u S_u S u , each with matrix h θ h_\theta h θ . Hence each z θ z_\theta z θ with θ ∈ M \theta\in\mathcal{M} θ ∈ M has algebraic and geometric multiplicities rank E θ \operatorname{rank}E_\theta rank E θ in H 1 H_1 H 1 .
For θ ∈ M \theta\in\mathcal{M} θ ∈ M , let χ = ( 1 , z θ / g ) T \chi=(1,z_\theta/g)^T χ = ( 1 , z θ / g ) T . Since h θ T = h θ h_\theta^T=h_\theta h θ T = h θ , its spectral projection at z θ z_\theta z θ is χ χ T / ( χ T χ ) \chi\chi^T/(\chi^T\chi) χ χ T / ( χ T χ ) , whose denominator is nonzero because ∣ z θ ∣ < g |z_\theta|<g ∣ z θ ∣ < g . Combining this projection with E θ E_\theta E θ gives
P θ = 1 g 2 + z θ 2 ( g 2 E θ g z θ E θ g z θ E θ z θ 2 E θ ) , (55) P_\theta=\frac{1}{g^2+z_\theta^2}
\begin{pmatrix}
g^2E_\theta& gz_\theta E_\theta\\
gz_\theta E_\theta& z_\theta^2E_\theta
\end{pmatrix}, \tag*{(55)} P θ = g 2 + z θ 2 1 ( g 2 E θ g z θ E θ g z θ E θ z θ 2 E θ ) , (55)
with rank P θ = rank E θ P_\theta=\operatorname{rank}E_\theta P θ = rank E θ .
A repeated channel root requires D θ = 0 D_\theta=0 D θ = 0 and 2 g = λ 2g=\lambda 2 g = λ . Then ( h θ + i λ I 2 / 2 ) 2 = 0 (h_\theta+i\lambda I_2/2)^2=0 ( h θ + iλ I 2 /2 ) 2 = 0 and
e − i h θ t = e − λ t / 2 [ I 2 − i t ( h θ + i λ 2 I 2 ) ] . (56) e^{-ih_\theta t}=e^{-\lambda t/2}\left[I_2-it\left(h_\theta+\frac{i\lambda}{2}I_2\right)\right]. \tag*{(56)} e − i h θ t = e − λ t /2 [ I 2 − i t ( h θ + 2 iλ I 2 ) ] . (56)
Its decay rate satisfies λ / 2 > γ 1 \lambda/2>\gamma_1 λ /2 > γ 1 . All eigenvalues omitted from Eq. (20) have decay rates greater than γ 1 \gamma_1 γ 1 , so for fixed parameters, choosing ε > 0 \varepsilon>0 ε > 0 smaller than every such rate difference absorbs the possible factor t t t and proves the expansion. The same bound holds for the first time derivative of the remainder.
At Δ = 0 \Delta=0 Δ = 0 we have M ⊂ { R , − R } \mathcal{M}\subset\{R,-R\} M ⊂ { R , − R } , and connectedness makes R R R simple. When − R ∈ spec A -R\in\operatorname{spec}A − R ∈ spec A , the graph is bipartite, and J A J = − A JAJ=-A J A J = − A makes − R -R − R simple [45 ] . The multiplicity result above therefore gives algebraic multiplicity one for every z θ z_\theta z θ with θ ∈ M \theta\in\mathcal{M} θ ∈ M .
An exact equal-rate case
For K 4 K_4 K 4 with g = η = λ = 1 g=\eta=\lambda=1 g = η = λ = 1 and Δ = − 2 \Delta=-2 Δ = − 2 , the adjacency eigenvalues are 3 3 3 and − 1 -1 − 1 , the latter with effective detuning − 3 -3 − 3 . Let S S S be a symmetric matrix with zero diagonal and S 1 = 0 S1=0 S 1 = 0 , where 1 1 1 has every entry equal to one. Such matrices form a two-dimensional space on K 4 K_4 K 4 . If z z z is the root of z 2 + ( 3 + i ) z − 1 = 0 z^2+(3+i)z-1=0 z 2 + ( 3 + i ) z − 1 = 0 of smaller modulus, then
∣ Ψ S ⟩ = ∑ r < s S r s ( σ r + + z m r t ) ( σ s + + z m s t ) ∣ 0 ⟩ (57) |\Psi_S\rangle=\sum_{r<s}S_{rs}(\sigma_r^++zm_r^t)(\sigma_s^++zm_s^t)|0\rangle \tag*{(57)} ∣ Ψ S ⟩ = r < s ∑ S r s ( σ r + + z m r t ) ( σ s + + z m s t ) ∣0 ⟩ (57)
is an eigenvector of H 2 H_2 H 2 with eigenvalue 2 z 2z 2 z . Together with Eq. (21) , this proves Eq. (45) . The map S ↦ ∣ Ψ S ⟩ S\mapsto|\Psi_S\rangle S ↦ ∣ Ψ S ⟩ is injective through its two-qubit component and therefore gives two independent eigenvectors.
Only h − 1 h_{-1} h − 1 is a least damped channel, so every least damped one-excitation mode has eigenvalue z z z . By the proof of Lemma 1, every eigenvector of H 2 H_2 H 2 with decay rate 2 γ 1 2\gamma_1 2 γ 1 is an eigenvector of the auxiliary generator H ~ 2 \widetilde{H}_2 H 2 used in that proof, with eigenvalue 2 z 2z 2 z and zero qubit double-occupation amplitudes. Both channels have D θ ≠ 0 D_\theta\ne0 D θ = 0 , so H 1 H_1 H 1 is diagonalizable, and such a vector is a symmetric combination of products of the least damped modes ( u , z u ) (u,zu) ( u , z u ) with u ⊥ 1 u\perp1 u ⊥ 1 . Its two-qubit component is therefore a symmetric matrix S S S with zero diagonal and S 1 = 0 S1=0 S 1 = 0 , and the vector equals ∣ Ψ S ⟩ \lvert\Psi_S\rangle ∣ Ψ S ⟩ . Hence 2 z 2z 2 z has geometric multiplicity two, and Lemma 2 gives algebraic multiplicity two.
A preparation with Q R R = 0 Q_{RR}=0 Q R R = 0
For the five-vertex graph and initial state given in Sec. IV, the state lies in the span of the two adjacency eigenvectors u ± 1 = ( 1 , ± 1 , − 1 , ∓ 1 , 0 ) / 2 u_{\pm1}=(1,\pm1,-1,\mp1,0)/2 u ± 1 = ( 1 , ± 1 , − 1 , ∓ 1 , 0 ) /2 . Let z , z ′ z,z' z , z ′ be the channel roots at D = η > 0 D=\eta>0 D = η > 0 , with ∣ z ∣ < ∣ z ′ ∣ \lvert z\rvert<\lvert z'\rvert ∣ z ∣ < ∣ z ′ ∣ . Starting from a qubit excitation, this channel has qubit amplitude
f ( t ) = z ′ e − i z t − z e − i z ′ t z ′ − z , (58) f(t)=\frac{z'e^{-izt}-ze^{-iz't}}{z'-z}, \tag*{(58)} f ( t ) = z ′ − z z ′ e − i z t − z e − i z ′ t , (58)
and the channel at D = − η D=-\eta D = − η has qubit amplitude f ( t ) ‾ \overline{f(t)} f ( t ) by the root symmetry in Appendix A. The qubit amplitudes therefore satisfy c 1 = − c 3 = Re f ( t ) / 2 c_1=-c_3=\operatorname{Re}f(t)/\sqrt{2} c 1 = − c 3 = Re f ( t ) / 2 , so C 13 ( t ) = [ Re f ( t ) ] 2 C_{13}(t)=[\operatorname{Re}f(t)]^2 C 13 ( t ) = [ Re f ( t ) ] 2 . Since the second term in f f f decays faster and Re z ≠ 0 \operatorname{Re}z\ne0 Re z = 0 , this real part changes sign infinitely often at isolated zeros, which gives arbitrarily late revivals on a nonbipartite graph.
Exact evolution of the equal-rate preparation
For the preparation in Eq. (46) , let S S S be the symmetric matrix with nonzero upper-triangular entries ( S 12 , S 13 , S 24 , S 34 ) = ( 1 , − 1 , − 1 , 1 ) (S_{12},S_{13},S_{24},S_{34})=(1,-1,-1,1) ( S 12 , S 13 , S 24 , S 34 ) = ( 1 , − 1 , − 1 , 1 ) . Then ∣ s ⟩ = ∑ k < l S k l ∣ k l ⟩ / 2 \lvert s\rangle=\sum_{k<l}S_{kl}\lvert kl\rangle/2 ∣ s ⟩ = ∑ k < l S k l ∣ k l ⟩ /2 is normalized. Define f ( t ) , h ( t ) f(t),h(t) f ( t ) , h ( t ) by
\begin{pmatrix}f(t)\\h(t)\end{pmatrix}=e^{-ih_{-1}t}\begin{pmatrix}1\\0\end{pmatrix}. \tag*{(59)}\end{pmatrix}
The zero diagonal and zero row sums of S S S give the exact conditional evolution
e − i H 2 t ∣ s ⟩ = 1 2 ∑ k < l S k l [ f ( t ) σ k + + h ( t ) m k + ] [ f ( t ) σ l + + h ( t ) m l + ] ∣ 0 ⟩ . (60) e^{-iH_2t}\lvert s\rangle=\frac{1}{2}\sum_{k<l}S_{kl}\bigl[f(t)\sigma_k^+ +h(t)m_k^+\bigr]\bigl[f(t)\sigma_l^+ +h(t)m_l^+\bigr]\lvert0\rangle. \tag*{(60)} e − i H 2 t ∣ s ⟩ = 2 1 k < l ∑ S k l [ f ( t ) σ k + + h ( t ) m k + ] [ f ( t ) σ l + + h ( t ) m l + ] ∣ 0 ⟩ . (60)
A loss at vertex r r r and time u u u leaves the one-excitation vector whose spatial amplitudes form row r r r of S S S and whose channel amplitudes are f ( u ) , h ( u ) f(u),h(u) f ( u ) , h ( u ) , multiplied by 2 λ h ( u ) / 2 \sqrt{2}\lambda h(u)/2 2 λh ( u ) /2 . Since every row of S S S belongs to the − 1 -1 − 1 adjacency eigenspace, subsequent propagation gives channel amplitudes f ( t ) , h ( t ) f(t),h(t) f ( t ) , h ( t ) . The single-channel norm satisfies
d d t ( ∣ f ∣ 2 + ∣ h ∣ 2 ) = − 2 λ ∣ h ∣ 2 , 2 λ ∫ 0 ∞ ∣ h ( u ) ∣ 2 d u = 1. (61) \frac{d}{dt}\left(\lvert f\rvert^2+\lvert h\rvert^2\right)=-2\lambda\lvert h\rvert^2,\qquad2\lambda\int_0^\infty\lvert h(u)\rvert^2\,du=1. \tag*{(61)} d t d ( ∣ f ∣ 2 + ∣ h ∣ 2 ) = − 2 λ ∣ h ∣ 2 , 2 λ ∫ 0 ∞ ∣ h ( u ) ∣ 2 d u = 1. (61)
The contribution to f f f from the eigenvalue z z z is α e − i z t \alpha e^{-izt} α e − i z t , with α = ( 1 + z 2 ) − 1 \alpha=(1+z^2)^{-1} α = ( 1 + z 2 ) − 1 . The qubit block of e 2 γ 1 t ρ ( 1 , 1 ) ( t ) e^{2\gamma_1t}\rho_{(1,1)}(t) e 2 γ 1 t ρ ( 1 , 1 ) ( t ) therefore converges to p ∣ α ∣ 2 S T S / 4 p\lvert\alpha\rvert^2S^TS/4 p ∣ α ∣ 2 S T S /4 . For the pair 12, the identities ( S T S ) 11 = ( S T S ) 22 = 2 (S^TS)_{11}=(S^TS)_{22}=2 ( S T S ) 11 = ( S T S ) 22 = 2 and ( S T S ) 12 = 0 (S^TS)_{12}=0 ( S T S ) 12 = 0 give m 11 = m 22 = p a 0 m_{11}=m_{22}=pa_0 m 11 = m 22 = p a 0 and m 12 = 0 m_{12}=0 m 12 = 0 . Equation (A13) yields d = p ∣ α ∣ 4 / 4 = p a 0 2 d=p\lvert\alpha\rvert^4/4=pa_0^2 d = p ∣ α ∣ 4 /4 = p a 0 2 and b = p ( 1 − p ) α 2 e − 2 i ω t / 2 b=\sqrt{p(1-p)}\alpha^2e^{-2i\omega t}/2 b = p ( 1 − p ) α 2 e − 2 iω t /2 .
If the only nonzero off-diagonal entries of a positive semidefinite X X X are X 14 X_{14} X 14 and X 41 X_{41} X 41 , Eq. (24) gives C ( X ) = 2 max { 0 , ∣ X 14 ∣ − X 22 X 33 } C(X)=2\max\{0,\lvert X_{14}\rvert-\sqrt{X_{22}X_{33}}\} C ( X ) = 2 max { 0 , ∣ X 14 ∣ − X 22 X 33 } . Hence C ( X 12 ( 0 ) ) = 2 max { 0 , ∣ b ∣ − p a 0 } C(X_{12}^{(0)})=2\max\{0,\lvert b\rvert-pa_0\} C ( X 12 ( 0 ) ) = 2 max { 0 , ∣ b ∣ − p a 0 } , which proves Eq. (47) .
The same construction gives the pair state at every time. Write I ( t ) = 2 λ ∫ 0 t ∣ h ( u ) ∣ 2 d u = 1 − ∣ f ( t ) ∣ 2 − ∣ h ( t ) ∣ 2 I(t)=2\lambda\int_0^t\lvert h(u)\rvert^2du=1-\lvert f(t)\rvert^2-\lvert h(t)\rvert^2 I ( t ) = 2 λ ∫ 0 t ∣ h ( u ) ∣ 2 d u = 1 − ∣ f ( t ) ∣ 2 − ∣ h ( t ) ∣ 2 and suppress the time argument of f f f and h h h . The pair population ( ρ 12 ) 22 (\rho_{12})_{22} ( ρ 12 ) 22 receives the contributions p ( ∣ f ∣ 4 / 4 + ∣ f ∣ 2 ∣ h ∣ 2 / 2 ) p(\lvert f\rvert^4/4+\lvert f\rvert^2\lvert h\rvert^2/2) p (∣ f ∣ 4 /4 + ∣ f ∣ 2 ∣ h ∣ 2 /2 ) without loss and p ∣ f ∣ 2 I / 2 p\lvert f\rvert^2I/2 p ∣ f ∣ 2 I /2 after one loss. The same holds for entry 33. Orthogonality of the first two rows of S S S makes entry 23 zero.
Table A15.
Since the preparation has no one-excitation component, Eq. (9) keeps the (2,1) and (1,0) blocks zero, so only the diagonal and corner entries are nonzero. The exact concurrence is
C 12 ( t ) = [ ∣ f ∣ 2 ( p ( 1 − p ) − p ) + p 2 ∣ f ∣ 4 ] + , (62) C_{12}(t)=\left[\lvert f\rvert^2\left(\sqrt{p(1-p)}-p\right)+\frac{p}{2}\lvert f\rvert^4\right]_+, \tag*{(62)} C 12 ( t ) = [ ∣ f ∣ 2 ( p ( 1 − p ) − p ) + 2 p ∣ f ∣ 4 ] + , (62)
where [ x ] + = max ( 0 , x ) [x]_+=\max(0,x) [ x ] + = max ( 0 , x ) . For 0 < p < 1 / 2 0<p<1/2 0 < p < 1/2 , the coefficient of ∣ f ∣ 2 \lvert f\rvert^2 ∣ f ∣ 2 is positive. Since f − α e − i z t = ( 1 − α ) e − i z ′ t f-\alpha e^{-izt}=(1-\alpha)e^{-iz't} f − α e − i z t = ( 1 − α ) e − i z ′ t , where z ′ z' z ′ is the other root of the channel equation and Im z ′ = − ( λ − γ 1 ) \operatorname{Im}z' =-(\lambda-\gamma_1) Im z ′ = − ( λ − γ 1 ) , we have d ∣ f ∣ 2 / d t = − 2 γ 1 ∣ α ∣ 2 e − 2 γ 1 t + O ( e − λ t ) d\lvert f\rvert^2/dt=-2\gamma_1\lvert\alpha\rvert^2e^{-2\gamma_1t}+O(e^{-\lambda t}) d ∣ f ∣ 2 / d t = − 2 γ 1 ∣ α ∣ 2 e − 2 γ 1 t + O ( e − λ t ) . As 2 γ 1 < λ 2\gamma_1<\lambda 2 γ 1 < λ , ∣ f ∣ 2 \lvert f\rvert^2 ∣ f ∣ 2 and hence C 12 ( t ) C_{12}(t) C 12 ( t ) are eventually strictly decreasing. At p = 1 / 2 p=1/2 p = 1/2 , this is ∣ f ∣ 4 / 4 − ∣ α ∣ 2 e − 4 γ 1 t / 4 \lvert f\rvert^4/4-\lvert\alpha\rvert^2e^{-4\gamma_1t}/4 ∣ f ∣ 4 /4 − ∣ α ∣ 2 e − 4 γ 1 t /4 . For p > 1 / 2 p>1/2 p > 1/2 , the coefficient of ∣ f ∣ 2 \lvert f\rvert^2 ∣ f ∣ 2 is negative and f → 0 f\to0 f → 0 , so the expression is zero for every sufficiently late time.
Sector decay bound and its equality case
We prove Proposition 1 for g , η , λ > 0 g,\eta,\lambda>0 g , η , λ > 0 and Δ ∈ R \Delta\in\mathbb{R} Δ ∈ R . For a Hermitian matrix X X X , we write X ⪰ 0 X\succeq0 X ⪰ 0 if X X X is positive semidefinite and X ≻ 0 X\succ0 X ≻ 0 if X X X is positive definite. Section II and Appendix A2 show that 0 < γ 1 < λ / 2 0<\gamma_1<\lambda/2 0 < γ 1 < λ /2 and ∣ D θ ∣ = max ϑ ∈ spec A ∣ D ϑ ∣ \lvert D_\theta\rvert=\max_{\vartheta\in\operatorname{spec}A}\lvert D_\vartheta\rvert ∣ D θ ∣ = max ϑ ∈ spec A ∣ D ϑ ∣ for every θ ∈ M \theta\in\mathcal{M} θ ∈ M . Define
τ = λ − 2 γ 1 , d = max θ ∈ spec A D θ 2 τ 2 . (63) \tau=\lambda-2\gamma_1,\qquad d=\frac{\max_{\theta\in\operatorname{spec}A}D_\theta^2}{\tau^2}. \tag*{(63)} τ = λ − 2 γ 1 , d = τ 2 max θ ∈ spec A D θ 2 . (63)
Then τ , d > 0 \tau,d>0 τ , d > 0 , and for θ ∈ M \theta\in\mathcal{M} θ ∈ M , substituting z θ z_\theta z θ into its channel equation gives
g 2 = γ 1 ( λ − γ 1 ) ( 1 + d ) . (64) g^2=\gamma_1(\lambda-\gamma_1)(1+d). \tag*{(64)} g 2 = γ 1 ( λ − γ 1 ) ( 1 + d ) . (64)
Proof of the sector decay bound
Introduce auxiliary bosonic modes b r b_r b r by replacing σ r − \sigma_r^- σ r − and σ r + \sigma_r^+ σ r + with b r b_r b r and b r † b_r^\dagger b r † , where [ b r , b s † ] = δ r s [b_r,b_s^\dagger]=\delta_{rs} [ b r , b s † ] = δ r s , and leave the memory modes unchanged. The auxiliary one-excitation generator is still H 1 H_1 H 1 . Since the auxiliary model has no interactions between excitations, its two-excitation generator is
H ~ 2 = ( H 1 ⊗ I + I ⊗ H 1 ) ∣ Sym 2 H 1 , H 1 = C 2 N , \widetilde{\mathcal{H}}_2=\left.(H_1\otimes I+I\otimes H_1)\right|_{\operatorname{Sym}^2\mathcal{H}_1},\qquad\mathcal{H}_1=\mathbb{C}^{2N}, H 2 = ( H 1 ⊗ I + I ⊗ H 1 ) ∣ Sym 2 H 1 , H 1 = C 2 N ,
where Sym 2 H 1 \operatorname{Sym}^2\mathcal{H}_1 Sym 2 H 1 is the subspace spanned by u ⊗ v + v ⊗ u u\otimes v+v\otimes u u ⊗ v + v ⊗ u with u , v ∈ H 1 u,v\in\mathcal{H}_1 u , v ∈ H 1 . Let E : ran Π 2 → Sym 2 H 1 E:\operatorname{ran}\Pi_2\to\operatorname{Sym}^2\mathcal{H}_1 E : ran Π 2 → Sym 2 H 1 be the linear map that replaces σ r + \sigma_r^+ σ r + with b r † b_r^\dagger b r † . Then E † E = I E^\dagger E=I E † E = I , and ran E \operatorname{ran}E ran E is the orthogonal complement of the N N N states ( b r † ) 2 ∣ 0 ⟩ / 2 (b_r^\dagger)^2|0\rangle/\sqrt{2} ( b r † ) 2 ∣0 ⟩ / 2 . Let U = E ∣ Q 2 ⊗ ∣ Ω M ⟩ U=E|_{\mathcal{Q}_2\otimes|\Omega_M\rangle} U = E ∣ Q 2 ⊗ ∣ Ω M ⟩ . For simplicity, we denote E ψ E\psi E ψ by ψ \psi ψ . In the decomposition by zero or positive memory occupation,
H ~ 2 = ( 0 V † V K ) , H 2 = ( 0 U † V † V U K ) . \widetilde{\mathcal{H}}_2=\begin{pmatrix}0&V^\dagger\\V&K\end{pmatrix},\qquad
H_2=\begin{pmatrix}0&U^\dagger V^\dagger\\VU&K\end{pmatrix}. H 2 = ( 0 V V † K ) , H 2 = ( 0 V U U † V † K ) .
Thus H 2 = E † H ~ 2 E H_2=E^\dagger\widetilde{\mathcal{H}}_2E H 2 = E † H 2 E .
Let e r e_r e r and f r f_r f r be the vectors σ r + ∣ 0 ⟩ \sigma_r^+|0\rangle σ r + ∣0 ⟩ and m r † ∣ 0 ⟩ m_r^\dagger|0\rangle m r † ∣0 ⟩ of B 1 \mathcal{B}_1 B 1 , and put e r r = e r ⊗ e r e_{rr}=e_r\otimes e_r e r r = e r ⊗ e r . Then e r r = ( b r † ) 2 ∣ 0 ⟩ / 2 e_{rr}=(b_r^\dagger)^2|0\rangle/\sqrt{2} e r r = ( b r † ) 2 ∣0 ⟩ / 2 , so e r r † ψ = 0 e_{rr}^\dagger\psi=0 e r r † ψ = 0 for ψ ∈ ran E \psi\in\operatorname{ran}E ψ ∈ ran E . We call e r r † ψ e_{rr}^\dagger\psi e r r † ψ the qubit double-occupation amplitude of ψ ∈ Sym 2 H 1 \psi\in\operatorname{Sym}^2\mathcal{H}_1 ψ ∈ Sym 2 H 1 at vertex r r r . By the zero block of H ~ 2 \widetilde{\mathcal{H}}_2 H 2 , we have e r r † H ~ 2 e s s = 0 e_{rr}^\dagger\widetilde{\mathcal{H}}_2e_{ss}=0 e r r † H 2 e ss = 0 . Since H 1 e r = H 1 † e r = g f r H_1e_r=H_1^\dagger e_r=gf_r H 1 e r = H 1 † e r = g f r ,
H ~ 2 e r r = H ~ 2 † e r r = g ( e r ⊗ f r + f r ⊗ e r ) . (65) \widetilde{\mathcal{H}}_2e_{rr}=\widetilde{\mathcal{H}}_2^\dagger e_{rr}=g(e_r\otimes f_r+f_r\otimes e_r). \tag*{(65)} H 2 e r r = H 2 † e r r = g ( e r ⊗ f r + f r ⊗ e r ) . (65)
Define
ℓ = λ ( 1 + d ) , κ = 2 g ℓ , W = ( 1 − i κ i κ 1 ) . (66) \ell=\lambda(1+d),\qquad\kappa=\frac{2g}{\ell},\qquad W=\begin{pmatrix}1&-i\kappa\\i\kappa&1\end{pmatrix}. \tag*{(66)} ℓ = λ ( 1 + d ) , κ = ℓ 2 g , W = ( 1 iκ − iκ 1 ) . (66)
Equation (B2) gives g κ = 2 γ 1 ( λ − γ 1 ) / λ g\kappa=2\gamma_1(\lambda-\gamma_1)/\lambda g κ = 2 γ 1 ( λ − γ 1 ) / λ and κ 2 = 4 γ 1 ( λ − γ 1 ) / [ λ 2 ( 1 + d ) ] \kappa^2=4\gamma_1(\lambda-\gamma_1)/[\lambda^2(1+d)] κ 2 = 4 γ 1 ( λ − γ 1 ) / [ λ 2 ( 1 + d )] . Since 4 γ 1 ( λ − γ 1 ) ≤ λ 2 4\gamma_1(\lambda-\gamma_1)\le\lambda^2 4 γ 1 ( λ − γ 1 ) ≤ λ 2 and d > 0 d>0 d > 0 , we have κ < 1 \kappa<1 κ < 1 , which implies W ≻ 0 W\succ0 W ≻ 0 . For real D D D , let h D h_D h D be the matrix in Eq. (15) with D θ D_\theta D θ replaced by D D D . Direct multiplication gives
− i ( h D † W − W h D ) − 2 γ 1 W = Q D , Q D = ( 2 γ 1 τ / λ κ ( D − i τ ) κ ( D + i τ ) 2 ( λ − γ 1 ) τ / λ ) . (67) -i(h_D^\dagger W-Wh_D)-2\gamma_1W=Q_D,\qquad
Q_D=\begin{pmatrix}
2\gamma_1\tau/\lambda&\kappa(D-i\tau)\\
\kappa(D+i\tau)&2(\lambda-\gamma_1)\tau/\lambda
\end{pmatrix}. \tag*{(67)} − i ( h D † W − W h D ) − 2 γ 1 W = Q D , Q D = ( 2 γ 1 τ / λ κ ( D + i τ ) κ ( D − i τ ) 2 ( λ − γ 1 ) τ / λ ) . (67)
The diagonal entries of Q D Q_D Q D are positive. By the expression for κ 2 \kappa^2 κ 2 , we have det Q D = κ 2 ( d τ 2 − D 2 ) \det Q_D=\kappa^2(d\tau^2-D^2) det Q D = κ 2 ( d τ 2 − D 2 ) . Since D θ 2 ≤ d τ 2 D_\theta^2\le d\tau^2 D θ 2 ≤ d τ 2 , every Q D θ ⪰ 0 Q_{D_\theta}\succeq0 Q D θ ⪰ 0 . In the basis B 1 \mathcal{B}_1 B 1 , Eqs. (11) and (12) give H 1 = ∑ θ h D θ ⊗ E θ H_1=\sum_\theta h_{D_\theta}\otimes E_\theta H 1 = ∑ θ h D θ ⊗ E θ , where the first factor acts on the qubit and memory blocks. With W 1 = W ⊗ I N W_1=W\otimes I_N W 1 = W ⊗ I N and Q 1 = ∑ θ Q D θ ⊗ E θ ⪰ 0 Q_1=\sum_\theta Q_{D_\theta}\otimes E_\theta\succeq0 Q 1 = ∑ θ Q D θ ⊗ E θ ⪰ 0 ,
− i ( H 1 † W 1 − W 1 H 1 ) − 2 γ 1 W 1 = Q 1 . -i(H_1^\dagger W_1-W_1H_1)-2\gamma_1W_1=Q_1. − i ( H 1 † W 1 − W 1 H 1 ) − 2 γ 1 W 1 = Q 1 .
Restricted to Sym 2 H 1 \operatorname{Sym}^2\mathcal{H}_1 Sym 2 H 1 , the operators W 2 = W 1 ⊗ W 1 ≻ 0 W_2=W_1\otimes W_1\succ0 W 2 = W 1 ⊗ W 1 ≻ 0 and Q 2 = Q 1 ⊗ W 1 + W 1 ⊗ Q 1 ⪰ 0 Q_2=Q_1\otimes W_1+W_1\otimes Q_1\succeq0 Q 2 = Q 1 ⊗ W 1 + W 1 ⊗ Q 1 ⪰ 0 satisfy
− i ( H ~ 2 † W 2 − W 2 H ~ 2 ) − 4 γ 1 W 2 = Q 2 . (68) -i(\widetilde{\mathcal{H}}_2^\dagger W_2-W_2\widetilde{\mathcal{H}}_2)-4\gamma_1W_2=Q_2. \tag*{(68)} − i ( H 2 † W 2 − W 2 H 2 ) − 4 γ 1 W 2 = Q 2 . (68)
Using W 1 e r = e r + i κ f r W_1e_r=e_r+i\kappa f_r W 1 e r = e r + iκ f r , W 1 f r = f r − i κ e r W_1f_r=f_r-i\kappa e_r W 1 f r = f r − iκ e r , κ ℓ = 2 g \kappa\ell=2g κ ℓ = 2 g , and Eq. (B3), both sides of
W 2 ( H ~ 2 + i ℓ ) e r r = − ( 1 − κ 2 ) ( H ~ 2 † − i ℓ ) e r r (69) W_2(\widetilde{\mathcal{H}}_2+i\ell)e_{rr}=-(1-\kappa^2)(\widetilde{\mathcal{H}}_2^\dagger-i\ell)e_{rr} \tag*{(69)} W 2 ( H 2 + i ℓ ) e r r = − ( 1 − κ 2 ) ( H 2 † − i ℓ ) e r r (69)
equal ( 1 − κ 2 ) [ i ℓ e r r − g ( e r ⊗ f r + f r ⊗ e r ) ] (1-\kappa^2)[i\ell e_{rr}-g(e_r\otimes f_r+f_r\otimes e_r)] ( 1 − κ 2 ) [ i ℓ e r r − g ( e r ⊗ f r + f r ⊗ e r )] .
Let H 2 ψ = z ψ H_2\psi=z\psi H 2 ψ = z ψ with ψ ≠ 0 \psi\ne0 ψ = 0 . Then e r r † ψ = 0 e_{rr}^\dagger\psi=0 e r r † ψ = 0 , and the block forms give
H ~ 2 ψ = z ψ + ∑ r c r e r r , c r = e r r † H ~ 2 ψ . \widetilde{\mathcal{H}}_2\psi=z\psi+\sum_r c_re_{rr},\qquad c_r=e_{rr}^\dagger\widetilde{\mathcal{H}}_2\psi. H 2 ψ = z ψ + r ∑ c r e r r , c r = e r r † H 2 ψ .
Put ϕ = ( H ~ 2 + i ℓ ) ψ \phi=(\widetilde{\mathcal{H}}_2+i\ell)\psi ϕ = ( H 2 + i ℓ ) ψ . Using e r r † ψ = 0 e_{rr}^\dagger\psi=0 e r r † ψ = 0 and e r r † H ~ 2 e s s = 0 e_{rr}^\dagger\widetilde{\mathcal{H}}_2e_{ss}=0 e r r † H 2 e ss = 0 ,
( z − H ~ 2 ) ϕ = − ∑ r c r ( H ~ 2 + i ℓ ) e r r , (z-\widetilde{\mathcal{H}}_2)\phi=-\sum_r c_r(\widetilde{\mathcal{H}}_2+i\ell)e_{rr}, ( z − H 2 ) ϕ = − r ∑ c r ( H 2 + i ℓ ) e r r ,
e r r † ( H ~ 2 + i ℓ ) ϕ = ( z + 2 i ℓ ) c r . e_{rr}^\dagger(\widetilde{\mathcal{H}}_2+i\ell)\phi=(z+2i\ell)c_r. e r r † ( H 2 + i ℓ ) ϕ = ( z + 2 i ℓ ) c r .
Equation (B7) then yields
ϕ † W 2 ( z − H ~ 2 ) ϕ = ( 1 − κ 2 ) ( z ‾ − 2 i ℓ ) ∑ r ∣ c r ∣ 2 . \phi^\dagger W_2(z-\widetilde{\mathcal{H}}_2)\phi=(1-\kappa^2)(\overline{z}-2i\ell)\sum_r|c_r|^2. ϕ † W 2 ( z − H 2 ) ϕ = ( 1 − κ 2 ) ( z − 2 i ℓ ) r ∑ ∣ c r ∣ 2 .
By Eq. (B6), the imaginary part of the left side is ( Im z + 2 γ 1 ) ϕ † W 2 ϕ + 1 2 ϕ † Q 2 ϕ (\operatorname{Im}z+2\gamma_1)\phi^\dagger W_2\phi+\frac{1}{2}\phi^\dagger Q_2\phi ( Im z + 2 γ 1 ) ϕ † W 2 ϕ + 2 1 ϕ † Q 2 ϕ . Therefore
0 = ( Im z + 2 γ 1 ) ϕ † W 2 ϕ + 1 2 ϕ † Q 2 ϕ + ( 1 − κ 2 ) ( Im z + 2 ℓ ) ∑ r ∣ c r ∣ 2 . (70) 0=(\operatorname{Im}z+2\gamma_1)\phi^\dagger W_2\phi+\frac{1}{2}\phi^\dagger Q_2\phi+(1-\kappa^2)(\operatorname{Im}z+2\ell)\sum_r|c_r|^2. \tag*{(70)} 0 = ( Im z + 2 γ 1 ) ϕ † W 2 ϕ + 2 1 ϕ † Q 2 ϕ + ( 1 − κ 2 ) ( Im z + 2 ℓ ) r ∑ ∣ c r ∣ 2 . (70)
Suppose Im z ≥ − 2 γ 1 \operatorname{Im}z\ge-2\gamma_1 Im z ≥ − 2 γ 1 . Since ℓ > λ > 2 γ 1 \ell>\lambda>2\gamma_1 ℓ > λ > 2 γ 1 and κ < 1 \kappa<1 κ < 1 , all three terms are nonnegative, so each vanishes. Thus every c r c_r c r vanishes and H ~ 2 ψ = z ψ \widetilde{\mathcal{H}}_2\psi=z\psi H 2 ψ = z ψ . If Im z > − 2 γ 1 \operatorname{Im}z>-2\gamma_1 Im z > − 2 γ 1 , the first term also gives ϕ = 0 \phi=0 ϕ = 0 . Since ϕ = ( z + i ℓ ) ψ \phi=(z+i\ell)\psi ϕ = ( z + i ℓ ) ψ and Im ( z + i ℓ ) > 0 \operatorname{Im}(z+i\ell)>0 Im ( z + i ℓ ) > 0 , this implies ψ = 0 \psi=0 ψ = 0 , a contradiction. No eigenvalue of H 2 H_2 H 2 has imaginary part greater than − 2 γ 1 -2\gamma_1 − 2 γ 1 , so γ 2 ≥ 2 γ 1 \gamma_2\ge2\gamma_1 γ 2 ≥ 2 γ 1 .
The equality case
Lemma 1 (Equality criterion). Equality γ 2 = 2 γ 1 \gamma_2=2\gamma_1 γ 2 = 2 γ 1 holds if and only if there exist ψ ≠ 0 \psi\ne0 ψ = 0 and z z z such that H 2 ψ = z ψ H_2\psi=z\psi H 2 ψ = z ψ , Im z = − 2 γ 1 \operatorname{Im}z=-2\gamma_1 Im z = − 2 γ 1 , and e r r † ψ = 0 e_{rr}^\dagger\psi=0 e r r † ψ = 0 for every r ∈ V r\in V r ∈ V .
For necessity, let H 2 ψ = z ψ H_2\psi=z\psi H 2 ψ = z ψ with ψ ≠ 0 \psi\ne0 ψ = 0 and Im z = − 2 γ 1 \operatorname{Im}z=-2\gamma_1 Im z = − 2 γ 1 . Equation (B8) gives c r = 0 c_r=0 c r = 0 for every r r r , so ψ \psi ψ is an eigenvector of H ~ 2 \widetilde{\mathcal{H}}_2 H 2 with zero qubit double-occupation amplitudes. Conversely, any auxiliary eigenvector satisfying these conditions lies in ran E \operatorname{ran}E ran E and is an eigenvector of H 2 H_2 H 2 . The sector decay bound then gives equality.
Lemma 2. If z ∈ spec H 2 z\in\operatorname{spec}H_2 z ∈ spec H 2 and Im z = − 2 γ 1 \operatorname{Im}z=-2\gamma_1 Im z = − 2 γ 1 , then z z z is semisimple , that is, every Jordan block of H 2 H_2 H 2 at z z z has size one.
Let Im z = − 2 γ 1 + ε \operatorname{Im}z=-2\gamma_1+\varepsilon Im z = − 2 γ 1 + ε and ξ ∈ ran Π 2 \xi\in\operatorname{ran}\Pi_2 ξ ∈ ran Π 2 , where 0 < ε ≤ λ 0<\varepsilon\le\lambda 0 < ε ≤ λ . By the sector decay bound, ψ = ( z − H 2 ) − 1 ξ \psi=(z-H_2)^{-1}\xi ψ = ( z − H 2 ) − 1 ξ exists. Put c r = e r r † H ~ 2 ψ c_r=e_{rr}^\dagger\widetilde{\mathcal{H}}_2\psi c r = e r r † H 2 ψ and ϕ = ( H ~ 2 + i ℓ ) ψ \phi=(\widetilde{\mathcal{H}}_2+i\ell)\psi ϕ = ( H 2 + i ℓ ) ψ . The block forms give H ~ 2 ψ = z ψ − ξ + ∑ r c r e r r \widetilde{\mathcal{H}}_2\psi=z\psi-\xi+\sum_r c_re_{rr} H 2 ψ = z ψ − ξ + ∑ r c r e r r . Using e r r † ξ = e r r † ψ = 0 e_{rr}^\dagger\xi=e_{rr}^\dagger\psi=0 e r r † ξ = e r r † ψ = 0 and e r r † H ~ 2 e s s = 0 e_{rr}^\dagger\widetilde{\mathcal{H}}_2e_{ss}=0 e r r † H 2 e ss = 0 ,
( z − H ~ 2 ) ϕ = ( H ~ 2 + i ℓ ) ( ξ − ∑ r c r e r r ) , (z-\widetilde{\mathcal{H}}_2)\phi=(\widetilde{\mathcal{H}}_2+i\ell)\left(\xi-\sum_r c_re_{rr}\right), ( z − H 2 ) ϕ = ( H 2 + i ℓ ) ( ξ − r ∑ c r e r r ) ,
e r r † ( H ~ 2 + i ℓ ) ϕ = ( z + 2 i ℓ ) c r − e r r † H ~ 2 ξ . e_{rr}^\dagger(\widetilde{\mathcal{H}}_2+i\ell)\phi=(z+2i\ell)c_r-e_{rr}^\dagger\widetilde{\mathcal{H}}_2\xi. e r r † ( H 2 + i ℓ ) ϕ = ( z + 2 i ℓ ) c r − e r r † H 2 ξ .
As in the derivation of Eq. (70) , Eqs. (68) and (69) give
ε ϕ † W 2 ϕ + 1 2 ϕ † Q 2 ϕ + ( 1 − κ 2 ) ( Im z + 2 ℓ ) ∥ c ∥ 2 = Im [ ϕ † W 2 ( H ~ 2 + i ℓ ) ξ ] − ( 1 − κ 2 ) Im ∑ r c r e r ‾ T H ~ 2 ξ (71) \varepsilon\phi^\dagger W_2\phi+\frac{1}{2}\phi^\dagger Q_2\phi+(1-\kappa^2)(\operatorname{Im}z+2\ell)\lVert c\rVert^2
=
\operatorname{Im}\left[\phi^\dagger W_2(\widetilde{H}_2+i\ell)\xi\right]-(1-\kappa^2)\operatorname{Im}\sum_r c_r\overline{e_r}^{\mathsf T}\widetilde{H}_2\xi \tag*{(71)} ε ϕ † W 2 ϕ + 2 1 ϕ † Q 2 ϕ + ( 1 − κ 2 ) ( Im z + 2 ℓ ) ∥ c ∥ 2 = Im [ ϕ † W 2 ( H 2 + i ℓ ) ξ ] − ( 1 − κ 2 ) Im r ∑ c r e r T H 2 ξ (71)
where ∥ c ∥ 2 = ∑ r ∣ c r ∣ 2 \lVert c\rVert^2=\sum_r\lvert c_r\rvert^2 ∥ c ∥ 2 = ∑ r ∣ c r ∣ 2 . The right side is at most a constant times ∥ ξ ∥ ( ∥ ϕ ∥ + ∥ c ∥ ) \lVert\xi\rVert(\lVert\phi\rVert+\lVert c\rVert) ∥ ξ ∥ (∥ ϕ ∥ + ∥ c ∥) . Since W 2 ≻ 0 W_2\succ0 W 2 ≻ 0 , Q 2 ⪰ 0 Q_2\succeq0 Q 2 ⪰ 0 , κ < 1 \kappa<1 κ < 1 , Im z + 2 ℓ ≥ 2 ( ℓ − γ 1 ) > 0 \operatorname{Im}z+2\ell\ge2(\ell-\gamma_1)>0 Im z + 2 ℓ ≥ 2 ( ℓ − γ 1 ) > 0 , and ε ≤ λ \varepsilon\le\lambda ε ≤ λ , the left side is at least a positive constant times ε ( ∥ ϕ ∥ + ∥ c ∥ ) 2 \varepsilon(\lVert\phi\rVert+\lVert c\rVert)^2 ε (∥ ϕ ∥ + ∥ c ∥ ) 2 . Both constants are independent of z z z and ξ \xi ξ , so ∥ ϕ ∥ + ∥ c ∥ = O ( ∥ ξ ∥ / ε ) \lVert\phi\rVert+\lVert c\rVert=O(\lVert\xi\rVert/\varepsilon) ∥ ϕ ∥ + ∥ c ∥ = O (∥ ξ ∥ / ε ) . Since ϕ = ( z + i ℓ ) ψ − ξ + ∑ r c r e r \phi=(z+i\ell)\psi-\xi+\sum_r c_r e_r ϕ = ( z + i ℓ ) ψ − ξ + ∑ r c r e r and ∣ z + i ℓ ∣ ≥ ℓ − 2 γ 1 > 0 \lvert z+i\ell\rvert\ge\ell-2\gamma_1>0 ∣ z + i ℓ ∣ ≥ ℓ − 2 γ 1 > 0 , this gives ∥ ( z − H 2 ) − 1 ∥ = O ( ε − 1 ) \lVert(z-H_2)^{-1}\rVert=O(\varepsilon^{-1}) ∥( z − H 2 ) − 1 ∥ = O ( ε − 1 ) as ε → 0 \varepsilon\to0 ε → 0 . If H 2 H_2 H 2 had a Jordan block of size m ≥ 2 m\ge2 m ≥ 2 at z 0 z_0 z 0 with Im z 0 = − 2 γ 1 \operatorname{Im}z_0=-2\gamma_1 Im z 0 = − 2 γ 1 , then ∥ ( z 0 + i ε − H 2 ) − 1 ∥ \lVert(z_0+i\varepsilon-H_2)^{-1}\rVert ∥( z 0 + i ε − H 2 ) − 1 ∥ would grow like ε − m \varepsilon^{-m} ε − m , a contradiction.
For θ ∈ M \theta\in\mathcal{M} θ ∈ M , write z θ = ω θ − i γ 1 z_\theta=\omega_\theta-i\gamma_1 z θ = ω θ − i γ 1 . Taking the imaginary part of the channel equation, we have ω θ = − γ 1 D θ / τ \omega_\theta=-\gamma_1D_\theta/\tau ω θ = − γ 1 D θ / τ . Since D θ ≠ 0 D_\theta\ne0 D θ = 0 , we have ω θ ≠ 0 \omega_\theta\ne0 ω θ = 0 . At resonance,
z R = ω R − i γ 1 , ω R = − γ 1 η R λ − 2 γ 1 < 0 , z_R=\omega_R-i\gamma_1,\qquad\omega_R=-\frac{\gamma_1\eta_R}{\lambda-2\gamma_1}<0, z R = ω R − i γ 1 , ω R = − λ − 2 γ 1 γ 1 η R < 0 ,
z − R = − ω R − i γ 1 (72) z_{-R}=-\omega_R-i\gamma_1 \tag*{(72)} z − R = − ω R − i γ 1 (72)
where the second root is present only for a bipartite graph. For θ ∈ M \theta\in\mathcal{M} θ ∈ M , Appendix A 2 shows that z θ z_\theta z θ is a semisimple eigenvalue of H 1 H_1 H 1 with eigenvectors
χ u = ( u z θ u / g ) , A u = θ u , u ≠ 0. (73) \chi_u=\begin{pmatrix}u\\ z_\theta u/g\end{pmatrix},\qquad Au=\theta u,\quad u\ne0. \tag*{(73)} χ u = ( u z θ u / g ) , A u = θ u , u = 0. (73)
By triangularizing H 1 H_1 H 1 , we have spec H ~ 2 ⊆ { z + z ′ : z , z ′ ∈ spec H 1 } \operatorname{spec}\widetilde{H}_2\subseteq\{z+z':z,z'\in\operatorname{spec}H_1\} spec H 2 ⊆ { z + z ′ : z , z ′ ∈ spec H 1 } . If − Im ( z + z ′ ) = 2 γ 1 -\operatorname{Im}(z+z')=2\gamma_1 − Im ( z + z ′ ) = 2 γ 1 , then − Im z = − Im z ′ = γ 1 -\operatorname{Im}z=-\operatorname{Im}z'=\gamma_1 − Im z = − Im z ′ = γ 1 . Since these summands are semisimple, every eigenvector of H ~ 2 \widetilde{H}_2 H 2 with decay rate 2 γ 1 2\gamma_1 2 γ 1 is a sum of symmetric products χ u ⊗ χ u ′ + χ u ′ ⊗ χ u \chi_u\otimes\chi_{u'}+\chi_{u'}\otimes\chi_u χ u ⊗ χ u ′ + χ u ′ ⊗ χ u for which the eigenvalues of χ u \chi_u χ u and χ u ′ \chi_{u'} χ u ′ add to its eigenvalue. The qubit double-occupation amplitude of such a product at vertex r r r is 2 u r u r ′ 2u_ru'_r 2 u r u r ′ .
If M = { θ } \mathcal{M}=\{\theta\} M = { θ } , the only auxiliary eigenvalue with decay rate 2 γ 1 2\gamma_1 2 γ 1 is 2 z θ 2z_\theta 2 z θ . Let { u k } \{u_k\} { u k } be a basis of the eigenspace A A A at θ \theta θ . The eigenvectors at 2 z θ 2z_\theta 2 z θ are the vectors ψ = ∑ k , l a k l χ u k ⊗ χ u l \psi=\sum_{k,l}a_{kl}\chi_{u_k}\otimes\chi_{u_l} ψ = ∑ k , l a k l χ u k ⊗ χ u l with a k l = a l k a_{kl}=a_{lk} a k l = a l k not all zero, and the qubit double-occupation amplitudes of ψ \psi ψ are the diagonal entries of
S = ∑ k , l a k l u k u l T . (74) S=\sum_{k,l}a_{kl}u_ku_l^{\mathsf T}. \tag*{(74)} S = k , l ∑ a k l u k u l T . (74)
The map from the coefficients a k l a_{kl} a k l to S S S is a bijection onto the symmetric matrices with A S = θ S AS=\theta S A S = θ S . For θ = R \theta=R θ = R , we have S = c v v T S=cvv^{\mathsf T} S = c v v T with c ≠ 0 c\ne0 c = 0 , so S r r = c v r 2 ≠ 0 S_{rr}=cv_r^2\ne0 S r r = c v r 2 = 0 .
If M = { R , θ − } \mathcal{M}=\{R,\theta_-\} M = { R , θ − } , then z θ − = − z R z_{\theta_-}=-z_R z θ − = − z R , and the auxiliary eigenvalues with decay rate 2 γ 1 2\gamma_1 2 γ 1 are 2 z R 2z_R 2 z R , z R + z θ − z_R+z_{\theta_-} z R + z θ − , and 2 z θ − 2z_{\theta_-} 2 z θ − . Their real parts 2 ω R 2\omega_R 2 ω R , 0 0 0 , and − 2 ω R -2\omega_R − 2 ω R are distinct, so each eigenvector belongs to one of the three eigenspaces. The eigenspaces at 2 z R 2z_R 2 z R and 2 z θ − 2z_{\theta_-} 2 z θ − are described by Eq. (74) with θ = R \theta=R θ = R and θ = θ − \theta=\theta_- θ = θ − . The eigenspace at z R + z θ − z_R+z_{\theta_-} z R + z θ − consists of the vectors χ v ⊗ χ u + χ u ⊗ χ v \chi_v\otimes\chi_u+\chi_u\otimes\chi_v χ v ⊗ χ u + χ u ⊗ χ v with A u = θ − u Au=\theta_-u A u = θ − u . Their qubit double-occupation amplitudes 2 v r u r 2v_ru_r 2 v r u r vanish at every vertex only if u = 0 u=0 u = 0 . By Eq. (19) , θ − ∈ M \theta_-\in\mathcal{M} θ − ∈ M if and only if Δ ≤ − η ( R + θ − ) / 2 \Delta\le-\eta(R+\theta_-)/2 Δ ≤ − η ( R + θ − ) /2 . Combining this with Lemma 1, we have γ 2 = 2 γ 1 \gamma_2=2\gamma_1 γ 2 = 2 γ 1 if and only if, in addition, A S = θ − S AS=\theta_-S A S = θ − S for some nonzero symmetric S S S with S r r = 0 S_{rr}=0 S r r = 0 . If θ − = R \theta_-=R θ − = R , S S S is simple with eigenvector u u u , then S = c u u T S=cuu^{\mathsf T} S = c u u T with c ≠ 0 c\ne0 c = 0 , and the entries S r r = c u r 2 S_{rr}=cu_r^2 S r r = c u r 2 do not all vanish. Since R + θ − ≥ 0 R+\theta_-\ge0 R + θ − ≥ 0 , if Δ ≥ 0 \Delta\ge0 Δ ≥ 0 then either Δ ≥ − η ( R + θ − ) / 2 \Delta\ge-\eta(R+\theta_-)/2 Δ ≥ − η ( R + θ − ) /2 , or R + θ − = 0 R+\theta_-=0 R + θ − = 0 and the graph is bipartite with θ − = − R \theta_-=-R θ − = − R . This proves Proposition 1.
Proposition 3. For a connected bipartite graph, γ 2 > 2 γ 1 \gamma_2>2\gamma_1 γ 2 > 2 γ 1 for every Δ ∈ R \Delta\in\mathbb{R} Δ ∈ R and every finite g , η , λ > 0 g,\eta,\lambda>0 g , η , λ > 0 .
This follows from Proposition 1, because θ − = − R \theta_-=-R θ − = − R is simple.
Proofs for late-time concurrence
We prove the asymptotic laws for γ 2 > 2 γ 1 \gamma_2>2\gamma_1 γ 2 > 2 γ 1 and γ 2 = 2 γ 1 \gamma_2=2\gamma_1 γ 2 = 2 γ 1 .
We use ∥ B ∥ 1 = Tr B † B \lVert B\rVert_1=\operatorname{Tr}\sqrt{B^\dagger B} ∥ B ∥ 1 = Tr B † B for the trace norm and ∥ B ∥ H S 2 = Tr ( B † B ) \lVert B\rVert_{\mathrm{HS}}^2=\operatorname{Tr}(B^\dagger B) ∥ B ∥ HS 2 = Tr ( B † B ) for the Hilbert-Schmidt norm, and ∥ B ∥ \lVert B\rVert ∥ B ∥ is the operator norm.
Bounds for a general pair matrix
Factor X = L L † X=LL^\dagger X = L L † with L L L a 4 × 4 4\times4 4 × 4 matrix, and denote its rows by u 1 , … , u 4 u_1,\ldots,u_4 u 1 , … , u 4 . The Wootters roots ξ k \xi_k ξ k in Eq. (24) are the singular values of
W = L T Y L = u 2 T u 3 + u 3 T u 2 − u 1 T u 4 − u 4 T u 1 . W=L^{\mathsf T}YL=u_2^{\mathsf T}u_3+u_3^{\mathsf T}u_2-u_1^{\mathsf T}u_4-u_4^{\mathsf T}u_1. W = L T Y L = u 2 T u 3 + u 3 T u 2 − u 1 T u 4 − u 4 T u 1 .
Let W 0 = u 2 T u 3 + u 3 T u 2 W_0=u_2^{\mathsf T}u_3+u_3^{\mathsf T}u_2 W 0 = u 2 T u 3 + u 3 T u 2 and denote its ordered singular values by s k s_k s k . Its two possibly nonzero singular values are ∥ u 2 ∥ ∥ u 3 ∥ ± ∣ u 2 u 3 T ∣ \lVert u_2\rVert\lVert u_3\rVert\pm\lvert u_2u_3^{\mathsf T}\rvert ∥ u 2 ∥ ∥ u 3 ∥ ± ∣ u 2 u 3 T ∣ , so s 1 − s 2 = 2 ∣ X 23 ∣ s_1-s_2=2\lvert X_{23}\rvert s 1 − s 2 = 2 ∣ X 23 ∣ and s 3 = s 4 = 0 s_3=s_4=0 s 3 = s 4 = 0 . The singular value perturbation inequality [47 ] gives
∣ C ( X ) − 2 ∣ X 23 ∣ ∣ ≤ ∑ k = 1 4 ∣ ξ k − s k ∣ ≤ ∥ W − W 0 ∥ 1 ≤ 2 ∥ u 1 ∥ ∥ u 4 ∥ = 2 X 11 X 44 . \lvert C(X)-2\lvert X_{23}\rvert\rvert\le\sum_{k=1}^{4}\lvert\xi_k-s_k\rvert
\le\lVert W-W_0\rVert_1
\le2\lVert u_1\rVert\lVert u_4\rVert
=2\sqrt{X_{11}X_{44}}. ∣ C ( X ) − 2 ∣ X 23 ∣∣ ≤ k = 1 ∑ 4 ∣ ξ k − s k ∣ ≤ ∥ W − W 0 ∥ 1 ≤ 2 ∥ u 1 ∥ ∥ u 4 ∥ = 2 X 11 X 44 .
where the row norms are Euclidean. This proves Eq. (25) .
Leading pair matrix at equal rates
Let P n P_n P n be the spectral projection onto the generalized eigenspaces of H n H_n H n with decay rate γ n \gamma_n γ n . Its kernel is the sum of the remaining generalized eigenspaces. In particular, P 1 = ∑ θ ∈ M P θ P_1=\sum_{\theta\in\mathcal{M}}P_\theta P 1 = ∑ θ ∈ M P θ . Define
T n ( t ) = e γ n t e − i H n t P n . (75) T_n(t)=e^{\gamma_nt}e^{-iH_nt}P_n. \tag*{(75)} T n ( t ) = e γ n t e − i H n t P n . (75)
For θ ∈ M \theta\in\mathcal{M} θ ∈ M , z θ z_\theta z θ is a semisimple eigenvalue of H 1 H_1 H 1 . At γ 2 = 2 γ 1 \gamma_2 = 2\gamma_1 γ 2 = 2 γ 1 , Lemma 2 gives the same for the eigenvalues of H ~ 2 \widetilde{H}_2 H 2 with decay rate 2 γ 1 2\gamma_1 2 γ 1 . By Eq. (19) , and since z θ − = − z R ‾ z_{\theta_-} = -\overline{z_R} z θ − = − z R when both channels are least damped, the roots z θ z_\theta z θ with θ ∈ M \theta\in\mathcal{M} θ ∈ M have real parts in { ω , − ω } \{\omega,-\omega\} { ω , − ω } , where ω = Re z θ \omega= \operatorname{Re} z_\theta ω = Re z θ for any θ ∈ M \theta\in\mathcal{M} θ ∈ M . By the proof of Lemma 1, every eigenvector of H 2 H_2 H 2 with decay rate 2 γ 1 2\gamma_1 2 γ 1 is an eigenvector of H ~ 2 \widetilde{H}_2 H 2 , so its eigenvalue is a sum of two such roots. Thus T 1 T_1 T 1 , and T 2 T_2 T 2 at γ 2 = 2 γ 1 \gamma_2 = 2\gamma_1 γ 2 = 2 γ 1 , are finite sums of constant matrices times e i k ω t e^{ik\omega t} e ik ω t with integer k k k . Consequently Φ i j \Phi_{ij} Φ ij , the entries of X i j ( 0 ) X_{ij}^{(0)} X ij ( 0 ) given below, and hence every leading profile F i j \mathcal{F}_{ij} F ij are periodic with period 2 π / ∣ ω ∣ 2\pi/|\omega| 2 π /∣ ω ∣ . At γ 2 = 2 γ 1 \gamma_2 = 2\gamma_1 γ 2 = 2 γ 1 , which requires Δ ≠ 0 \Delta\ne0 Δ = 0 , set
K 11 = ∑ a , b x a Q a b x b † . (76) K_{11} = \sum_{a,b} x_a Q_{ab} x_b^\dagger. \tag*{(76)} K 11 = a , b ∑ x a Q ab x b † . (76)
and define
K 10 = P 1 B 10 + ∫ 0 ∞ e i H 1 u P 1 F 10 ( u ) d u , (77) K_{10} = P_1 B_{10} + \int_0^\infty e^{iH_1u}P_1F_{10}(u)\,du, \tag*{(77)} K 10 = P 1 B 10 + ∫ 0 ∞ e i H 1 u P 1 F 10 ( u ) d u , (77)
F 10 ( u ) = ∑ r Γ r ( 2 ) ρ ( 2 , 1 ) ( u ) Γ r ( 1 ) † , F_{10}(u) = \sum_r \Gamma_r^{(2)}\rho^{(2,1)}(u)\Gamma_r^{(1)\dagger}, F 10 ( u ) = r ∑ Γ r ( 2 ) ρ ( 2 , 1 ) ( u ) Γ r ( 1 ) † ,
whose convergence is proved below. In the matrix elements below, ∣ i j ⟩ = σ i + σ j + ∣ 0 ⟩ |ij\rangle= \sigma_i^+\sigma_j^+|0\rangle ∣ ij ⟩ = σ i + σ j + ∣0 ⟩ . Suppress the time arguments of T 1 , T 2 T_1,T_2 T 1 , T 2 and put
d = ⟨ i j ∣ T 2 B 22 T 2 † ∣ i j ⟩ , d = \langle ij|T_2B_{22}T_2^\dagger|ij\rangle, d = ⟨ ij ∣ T 2 B 22 T 2 † ∣ ij ⟩ ,
b = ⟨ i j ∣ T 2 B 20 ∣ 0 ⟩ , b = \langle ij|T_2B_{20}|0\rangle, b = ⟨ ij ∣ T 2 B 20 ∣0 ⟩ ,
t k = ⟨ i j ∣ T 2 B 21 T 1 † ∣ q k ⟩ , t_k = \langle ij|T_2B_{21}T_1^\dagger|q_k\rangle, t k = ⟨ ij ∣ T 2 B 21 T 1 † ∣ q k ⟩ ,
m k l = ⟨ q k ∣ T 1 K 11 T 1 † ∣ q l ⟩ , m_{kl} = \langle q_k|T_1K_{11}T_1^\dagger|q_l\rangle, m k l = ⟨ q k ∣ T 1 K 11 T 1 † ∣ q l ⟩ ,
a k = ⟨ q k ∣ T 1 K 10 ∣ 0 ⟩ , k , l ∈ { i , j } . (78) a_k = \langle q_k|T_1K_{10}|0\rangle,\qquad k,l \in\{i,j\}. \tag*{(78)} a k = ⟨ q k ∣ T 1 K 10 ∣0 ⟩ , k , l ∈ { i , j } . (78)
The matrix X i j ( 0 ) X_{ij}^{(0)} X ij ( 0 ) defined in Sec. IV has the form
X i j ( 0 ) ( t ) = ( d t i t j b t i ‾ m i i m i j a i t j ‾ m i j ‾ m j j a j b ‾ a i ‾ a j ‾ 1 ) . (79) X_{ij}^{(0)}(t) =
\begin{pmatrix}
d & t_i & t_j & b \\
\overline{t_i} & m_{ii} & m_{ij} & a_i \\
\overline{t_j} & \overline{m_{ij}} & m_{jj} & a_j \\
\overline{b} & \overline{a_i} & \overline{a_j} & 1
\end{pmatrix}. \tag*{(79)} X ij ( 0 ) ( t ) = d t i t j b t i m ii m ij a i t j m ij m j j a j b a i a j 1 . (79)
The proof below shows that this matrix is positive semidefinite for every t t t . Feeding from ρ ( 2 , 1 ) \rho^{(2,1)} ρ ( 2 , 1 ) contributes to a i , a j a_i,a_j a i , a j , and these entries cannot be discarded for general initial states.
Asymptotic laws
a. Convergence of the integrals. The operator in Eq. (76) equals P 1 B 11 P 1 † P_1B_{11}P_1^\dagger P 1 B 11 P 1 † plus the integral of e i H 1 u P 1 F 11 ( u ) P 1 † e − i H 1 † u e^{iH_1u}P_1F_{11}(u)P_1^\dagger e^{-iH_1^\dagger u} e i H 1 u P 1 F 11 ( u ) P 1 † e − i H 1 † u . The integrand is positive semidefinite and bounded by a polynomial times e − 2 ( γ 2 − γ 1 ) u e^{-2(\gamma_2-\gamma_1)u} e − 2 ( γ 2 − γ 1 ) u , which is integrable by Eq. (21) . Thus Q ⪰ 0 Q \succeq0 Q ⪰ 0 and the integral in Eq. (27) converges. Since ρ ( 2 , 1 ) ( u ) = e − i H 2 u B 21 e i H 1 † u \rho^{(2,1)}(u) = e^{-iH_2u}B_{21}e^{iH_1^\dagger u} ρ ( 2 , 1 ) ( u ) = e − i H 2 u B 21 e i H 1 † u , the integrand defining K 10 K_{10} K 10 is bounded by a polynomial times e − γ 2 u e^{-\gamma_2u} e − γ 2 u , and its integral also converges.
b. Positivity and the asymptotic law at equal rates. Factor ρ S M ( 0 ) = L L † \rho_{\mathrm{SM}}(0) = LL^\dagger ρ SM ( 0 ) = L L † and put L n = Π n L L_n = \Pi_nL L n = Π n L . With U n ( u ) = e − i H n u U_n(u) = e^{-iH_nu} U n ( u ) = e − i H n u , form the columns
V 0 = ( P 2 L 2 P 1 L 1 L 0 ) , V r ( u ) = ( 0 e i H 1 u P 1 Γ r ( 2 ) U 2 ( u ) L 2 Γ r ( 1 ) U 1 ( u ) L 1 ) . (80) V_0 =
\begin{pmatrix}
P_2L_2 \\
P_1L_1 \\
L_0
\end{pmatrix},
\qquad
V_r(u) =
\begin{pmatrix}
0 \\
e^{iH_1u}P_1\Gamma_r^{(2)}U_2(u)L_2 \\
\Gamma_r^{(1)}U_1(u)L_1
\end{pmatrix}. \tag*{(80)} V 0 = P 2 L 2 P 1 L 1 L 0 , V r ( u ) = 0 e i H 1 u P 1 Γ r ( 2 ) U 2 ( u ) L 2 Γ r ( 1 ) U 1 ( u ) L 1 . (80)
The Gram sum
Z = V 0 V 0 † + ∑ r ∫ 0 ∞ V r ( u ) V r ( u ) † d u + diag ( 0 , 0 , Tr B 22 ) (81) Z = V_0V_0^\dagger+ \sum_r \int_0^\infty V_r(u)V_r(u)^\dagger\,du + \operatorname{diag}(0,0,\operatorname{Tr}B_{22}) \tag*{(81)} Z = V 0 V 0 † + r ∑ ∫ 0 ∞ V r ( u ) V r ( u ) † d u + diag ( 0 , 0 , Tr B 22 ) (81)
is positive semidefinite. Its middle row contains K 11 K_{11} K 11 on the diagonal and K 10 K_{10} K 10 in the vacuum column. Its vacuum entry is Tr B 00 + Tr B 11 + Tr B 22 = 1 \operatorname{Tr}B_{00}+\operatorname{Tr}B_{11}+\operatorname{Tr}B_{22}=1 Tr B 00 + Tr B 11 + Tr B 22 = 1 , where the integral of the vacuum entry from V r V_r V r equals Tr B 11 \operatorname{Tr}B_{11} Tr B 11 because the initial one-excitation population is eventually lost. The other blocks are the corresponding projected initial blocks. Conjugating Z Z Z by diag ( T 2 , T 1 , 1 ) \operatorname{diag}(T_2,T_1,1) diag ( T 2 , T 1 , 1 ) and restricting to ( ∣ i j ⟩ , ∣ q i ⟩ , ∣ q j ⟩ , ∣ 0 ⟩ ) (|ij\rangle,|q_i\rangle,|q_j\rangle,|0\rangle) ( ∣ ij ⟩ , ∣ q i ⟩ , ∣ q j ⟩ , ∣0 ⟩) gives exactly the matrix in Eq. (79) , which is therefore positive semidefinite. Write γ = γ 1 \gamma=\gamma_1 γ = γ 1 and use Λ t \Lambda_t Λ t and X i j X_{ij} X ij from Eq. (29) . Since Λ t γ Λ t = Y \Lambda_t\gamma\Lambda_t=Y Λ t γ Λ t = Y , the local filtering identity [48 ] gives C i j = e − 2 γ t C ( X i j ) C_{ij}=e^{-2\gamma t}\mathcal{C}(X_{ij}) C ij = e − 2 γ t C ( X ij ) exactly. The blocks ( 2 , 2 ) (2,2) ( 2 , 2 ) , ( 2 , 1 ) (2,1) ( 2 , 1 ) , and ( 2 , 0 ) (2,0) ( 2 , 0 ) have no feeding term in Eq. (9) , and their projections by P 1 P_1 P 1 and P 2 P_2 P 2 , scaled as in Eq. (29) , give d d d , t k t_k t k , and b b b . Variation of constants for the ( 1 , 1 ) (1,1) ( 1 , 1 ) and ( 1 , 0 ) (1,0) ( 1 , 0 ) blocks yields m k l m_{kl} m k l and a k a_k a k . The pair ground population tends to one. On this scale, every traced external excitation adds a strictly decaying factor, and the remaining spectral terms and the tails of the convergent integrals defining Q Q Q and K 10 K_{10} K 10 also decay exponentially. Thus, for some δ > 0 \delta>0 δ > 0 ,
∥ X i j − X i j ( 0 ) ∥ 1 + ∥ X ˙ i j − X ˙ i j ( 0 ) ∥ 1 ≤ K e − δ t . (82) \lVert X_{ij}-X_{ij}^{(0)}\rVert_1+\lVert\dot{X}_{ij}-\dot{X}_{ij}^{(0)}\rVert_1\leq Ke^{-\delta t}. \tag*{(82)} ∥ X ij − X ij ( 0 ) ∥ 1 + ∥ X ˙ ij − X ˙ ij ( 0 ) ∥ 1 ≤ K e − δ t . (82)
Polynomial factors from Jordan blocks or from resonant feeding integrals are absorbed by a strict reduction of δ \delta δ . For positive semidefinite matrices A , B A,B A , B , square-root factors and the singular value trace norm inequality give
∣ C ( A ) − C ( B ) ∣ ≤ ( Tr A + Tr B ) ∥ A − B ∥ 1 . (83) |\mathcal{C}(A)-\mathcal{C}(B)|\leq\left(\sqrt{\operatorname{Tr}A}+\sqrt{\operatorname{Tr}B}\right)\sqrt{\lVert A-B\rVert_1}. \tag*{(83)} ∣ C ( A ) − C ( B ) ∣ ≤ ( Tr A + Tr B ) ∥ A − B ∥ 1 . (83)
We used ∥ A − B ∥ H S 2 ≤ ∥ A − B ∥ 1 \lVert\sqrt{A}-\sqrt{B}\rVert_{\mathrm{HS}}^2\leq\lVert A-B\rVert_1 ∥ A − B ∥ HS 2 ≤ ∥ A − B ∥ 1 [49 ] . Since the traces of X i j X_{ij} X ij and X i j ( 0 ) X_{ij}^{(0)} X ij ( 0 ) are bounded, Eqs. (82) and (83) prove Eq. (30) at γ 2 = 2 γ 1 \gamma_2=2\gamma_1 γ 2 = 2 γ 1 .
c. The strict inequality. For γ 2 > 2 γ 1 \gamma_2>2\gamma_1 γ 2 > 2 γ 1 , choose $2\gamma_1<\gamma_2'<\gamma_2$. Projecting with P 1 P_1 P 1 in the variation of constants formula for the ( 1 , 1 ) (1,1) ( 1 , 1 ) block gives ( ρ i j ) 23 = e − 2 γ 1 t [ Φ i j + O ( e − δ t ) ] (\rho_{ij})_{23}=e^{-2\gamma_1t}[\Phi_{ij}+O(e^{-\delta t})] ( ρ ij ) 23 = e − 2 γ 1 t [ Φ ij + O ( e − δ t )] . The pair entry 11 11 11 belongs entirely to ρ ( 2 , 2 ) a n d i s \rho^{(2,2)} and is ρ ( 2 , 2 ) an d i s O(e^{-2\gamma_2t})$, so Eq. (25) proves the first branch.
d. Revival criteria with a controlled remainder. Let s 0 < s 1 < s 2 s_0<s_1<s_2 s 0 < s 1 < s 2 be fixed nonnegative times. Suppose τ n → ∞ \tau_n\to\infty τ n → ∞ and, for a = 0 , 2 a=0,2 a = 0 , 2 ,
e − 2 γ 1 s a F i j ( τ n + s a ) − e − 2 γ 1 s 1 F i j ( τ n + s 1 ) ≥ δ n > 0 , (84) e^{-2\gamma_1s_a}\mathcal{F}_{ij}(\tau_n+s_a)-e^{-2\gamma_1s_1}\mathcal{F}_{ij}(\tau_n+s_1)\geq\delta_n>0, \tag*{(84)} e − 2 γ 1 s a F ij ( τ n + s a ) − e − 2 γ 1 s 1 F ij ( τ n + s 1 ) ≥ δ n > 0 , (84)
where e β τ n δ n → ∞ e^{\beta\tau_n}\delta_n\to\infty e β τ n δ n → ∞ . Equation (30) then gives arbitrarily late revivals of the exact concurrence. Since every F i j \mathcal{F}_{ij} F ij is periodic, it suffices that both differences in Eq. (84) are positive at τ n = 0 \tau_n=0 τ n = 0 , because taking τ n \tau_n τ n as multiples of the period then yields a constant δ n \delta_n δ n .
e. Strict decrease with one least damped mode. Assume γ 2 > 2 γ 1 \gamma_2>2\gamma_1 γ 2 > 2 γ 1 and Q b b = 0 Q_{bb}=0 Q bb = 0 for all b ≠ a b\ne a b = a . The rows of L ⋆ L_\star L ⋆ indexed by ∣ 10 ⟩ \lvert10\rangle ∣ 10 ⟩ and ∣ 01 ⟩ \lvert01\rangle ∣ 01 ⟩ are linearly dependent. The argument in Appendix D proves Eq. (38) for any Δ ∈ R \Delta\in\mathbb{R} Δ ∈ R . Here Φ i j = Q a a v a ( i ) v a ( j ) \Phi_{ij}=Q_{aa}v_a(i)v_a(j) Φ ij = Q aa v a ( i ) v a ( j ) is constant. If Φ i j ≠ 0 \Phi_{ij}\ne0 Φ ij = 0 , then C i j ( t ) C_{ij}(t) C ij ( t ) is eventually strictly decreasing.
f. A constant profile with multiple least damped modes at equal rates. Suppose all least damped one-excitation modes have frequency ω \omega ω and all least damped two-excitation modes have frequency 2 ω 2\omega 2 ω , with arbitrary multiplicities. Then
X i j ( 0 ) ( t ) = U ω ( t ) X i j ( 0 ) ( 0 ) U ω ( t ) † (85) X_{ij}^{(0)}(t)=U_\omega(t)X_{ij}^{(0)}(0)U_\omega(t)^\dagger \tag*{(85)} X ij ( 0 ) ( t ) = U ω ( t ) X ij ( 0 ) ( 0 ) U ω ( t ) † (85)
where U ω = diag ( e − 2 i ω t , e − i ω t , e − i ω t , 1 ) U_\omega=\operatorname{diag}(e^{-2i\omega t},e^{-i\omega t},e^{-i\omega t},1) U ω = diag ( e − 2 iω t , e − iω t , e − iω t , 1 ) is a product of local phase rotations. Thus F i j F_{ij} F ij is constant for every initial state.
Variation bounds at resonance
Assume Δ = 0 \Delta=0 Δ = 0 and write γ = γ 1 \gamma=\gamma_1 γ = γ 1 . All constants below refer to fixed parameters and initial state, and K K K may differ between occurrences. By Eq. (22) , we can choose 2 γ < γ 2 ′ < γ 2 2\gamma<\gamma_2'<\gamma_2 2 γ < γ 2 ′ < γ 2 . For some β 1 > 0 \beta_1>0 β 1 > 0 ,
e − i H 1 t = ∑ a ∈ M e − i z a t x a y a † + E ( t ) , e^{-iH_1t}=\sum_{a\in M}e^{-iz_at}x_ay_a^\dagger+E(t), e − i H 1 t = a ∈ M ∑ e − i z a t x a y a † + E ( t ) ,
∥ E ∥ + ∥ E ˙ ∥ ≤ K e − ( γ + β 1 ) t . (86) \lVert E\rVert+\lVert\dot{E}\rVert\le Ke^{-(\gamma+\beta_1)t}. \tag*{(86)} ∥ E ∥ + ∥ E ˙ ∥ ≤ K e − ( γ + β 1 ) t . (86)
Also ∥ e − i H 2 t ∥ + ∥ ∂ t e − i H 2 t ∥ ≤ K e − γ 2 ′ t \lVert e^{-iH_2t}\rVert+\lVert\partial_t e^{-iH_2t}\rVert\le Ke^{-\gamma_2't} ∥ e − i H 2 t ∥ + ∥ ∂ t e − i H 2 t ∥ ≤ K e − γ 2 ′ t . Choose 0 < β < min { β 1 , γ 2 ′ − 2 γ } 0<\beta<\min\{\beta_1,\gamma_2'-2\gamma\} 0 < β < min { β 1 , γ 2 ′ − 2 γ } . Let Q ( t ) Q(t) Q ( t ) be the matrix in Eq. (27) with its integral truncated at t t t . Positivity gives 0 ⪯ Q ( t ) ⪯ Q 0\preceq Q(t)\preceq Q 0 ⪯ Q ( t ) ⪯ Q , and Q − Q ( t ) Q-Q(t) Q − Q ( t ) decays exponentially through one derivative.
a. A differentiable factor on a fixed space. To control variation, define Λ t \Lambda_t Λ t and X = X i j X=X_{ij} X = X ij as in Eq. (29) . As in Appendix C, C ( X ) = e 2 γ t C i j C(X)=e^{2\gamma t}C_{ij} C ( X ) = e 2 γ t C ij .
Choose L i n L_{\mathrm{in}} L in with ρ S M ( 0 ) = L i n L i n † \rho_{\mathrm{SM}}(0)=L_{\mathrm{in}}L_{\mathrm{in}}^\dagger ρ SM ( 0 ) = L in L in † . Write U ( t ) = ⨁ n = 0 2 e − i H n t U(t)=\bigoplus_{n=0}^{2}e^{-iH_nt} U ( t ) = ⨁ n = 0 2 e − i H n t and Γ r = ∑ n = 1 2 Γ r ( n ) \Gamma_r=\sum_{n=1}^{2}\Gamma_r^{(n)} Γ r = ∑ n = 1 2 Γ r ( n ) . The amplitude for k k k losses is
Ψ k ( t , u ) = U ( t − u k ) Γ r k U ( u k − u k − 1 ) ⋯ Γ r 1 U ( u 1 ) L i n , (87) \Psi_k(t,u)=U(t-u_k)\Gamma_{r_k}U(u_k-u_{k-1})\cdots\Gamma_{r_1}U(u_1)L_{\mathrm{in}}, \tag*{(87)} Ψ k ( t , u ) = U ( t − u k ) Γ r k U ( u k − u k − 1 ) ⋯ Γ r 1 U ( u 1 ) L in , (87)
with 0 < u 1 < ⋯ < u k < t 0<u_1<\cdots<u_k<t 0 < u 1 < ⋯ < u k < t and k ≤ 2 k\le2 k ≤ 2 . For k = 0 k=0 k = 0 it is U ( t ) L i n U(t)L_{\mathrm{in}} U ( t ) L in . Summing Ψ k Ψ k † \Psi_k\Psi_k^\dagger Ψ k Ψ k † over labels and integrating over loss times gives the exact state. After the partial trace, retain the initial factor and traced configuration as column indices. Multiply the four pair rows by e 2 γ t , e γ t , e γ t , 1 e^{2\gamma t},e^{\gamma t},e^{\gamma t},1 e 2 γ t , e γ t , e γ t , 1 . For each k k k , change variables to u l = t s l u_l=ts_l u l = t s l and multiply the amplitudes by t k / 2 t^{k/2} t k /2 . The resulting rows define a factor L ( t ) L(t) L ( t ) on the fixed direct sum of the spaces L 2 ( 0 < s 1 < ⋯ < s k < 1 ) L^2(0<s_1<\cdots<s_k<1) L 2 ( 0 < s 1 < ⋯ < s k < 1 ) with their finite column indices, where the k = 0 k=0 k = 0 summand is finite dimensional. The factor satisfies X = L L † X=LL^\dagger X = L L † .
This factor is continuously differentiable in Hilbert-Schmidt norm. Indeed, an amplitude A ( t , u ) A(t,u) A ( t , u ) of the scaled rows becomes G t ( s ) = t k / 2 A ( t , t s ) G_t(s)=t^{k/2}A(t,ts) G t ( s ) = t k /2 A ( t , t s ) , with
G ˙ t = t k / 2 [ k 2 t A + ∂ t A + ∑ l = 1 k s l ∂ u l A ] u = t s . (88) \dot{G}_t=t^{k/2}\left[\frac{k}{2t}A+\partial_tA+\sum_{l=1}^{k}s_l\partial_{u_l}A\right]_{u=ts}. \tag*{(88)} G ˙ t = t k /2 [ 2 t k A + ∂ t A + l = 1 ∑ k s l ∂ u l A ] u = t s . (88)
For λ t ≥ 1 \lambda t\ge1 λ t ≥ 1 , its squared Hilbert-Schmidt norm is bounded by a constant times the integral of ∣ A ∣ 2 + ∣ ∂ t A ∣ 2 + ∑ l ∣ ∂ u l A ∣ 2 \lvert A\rvert^2+\lvert\partial_tA\rvert^2+\sum_l\lvert\partial_{u_l}A\rvert^2 ∣ A ∣ 2 + ∣ ∂ t A ∣ 2 + ∑ l ∣ ∂ u l A ∣ 2 over the original loss times. Each partial derivative inserts a finite sector generator beside a propagator or jump, preserving the decay bounds. For one loss, the middle rows are bounded by K e − ( γ 2 ′ − γ ) t Ke^{-(\gamma_2'-\gamma)t} K e − ( γ 2 ′ − γ ) t and the fourth row by K e − γ t Ke^{-\gamma t} K e − γ t . For two losses, only the fourth row survives, with bound K e − γ 2 ′ ( u 2 − u 1 ) Ke^{-\gamma_2'(u_2-u_1)} K e − γ 2 ′ ( u 2 − u 1 ) . These bounds are square integrable on the infinite ordered domains. The rows without loss are bounded, with first row O ( e − ( γ 2 ′ − 2 γ ) t ) O(e^{-(\gamma_2'-2\gamma)t}) O ( e − ( γ 2 ′ − 2 γ ) t ) . Thus L L L and L ˙ \dot L L ˙ are bounded.
Set the first row of L L L to zero, retain in the middle rows only a final one-excitation state with its final propagator multiplied by P 1 P_1 P 1 , keep the fourth row unchanged, and call the result L ⋆ L_\star L ⋆ . For a discarded middle row with one loss, its squared norm and that of its derivative obey
K e − 2 β 1 t ∫ 0 t e − 2 ( γ 2 ′ − γ − β 1 ) u d u ≤ K e − 2 β t . (89) Ke^{-2\beta_1t}\int_0^t e^{-2(\gamma_2'-\gamma-\beta_1)u}\,du\le Ke^{-2\beta t}. \tag*{(89)} K e − 2 β 1 t ∫ 0 t e − 2 ( γ 2 ′ − γ − β 1 ) u d u ≤ K e − 2 β t . (89)
Equation (D3) gives the derivative bound, and the strict choice of β \beta β absorbs a possible factor of t t t . The other discarded terms decay at least as fast, so
∥ L − L ⋆ ∥ H S + ∥ L ˙ − L ˙ ⋆ ∥ H S ≤ K e − β t . (90) \lVert L-L_\star\rVert_{\mathrm{HS}}+\lVert\dot L-\dot L_\star\rVert_{\mathrm{HS}}\le Ke^{-\beta t}. \tag*{(90)} ∥ L − L ⋆ ∥ HS + ∥ L ˙ − L ˙ ⋆ ∥ HS ≤ K e − β t . (90)
The retained middle rows have the form ℓ k ⋆ = ∑ a v a ( k ) e − i ω a t ζ a ( t ) \ell_k^\star=\sum_a v_a(k)e^{-i\omega_at}\zeta_a(t) ℓ k ⋆ = ∑ a v a ( k ) e − i ω a t ζ a ( t ) for k = i , j k=i,j k = i , j , where ζ a ζ b † = Q a b ( t ) \zeta_a\zeta_b^\dagger=Q_{ab}(t) ζ a ζ b † = Q ab ( t ) . Different loss histories occupy orthogonal summands, so this Gram matrix adds the initial and feeding terms without interference between them. The inner product of the retained rows is therefore b ⋆ = Φ i j + O ( e − β t ) b_\star=\Phi_{ij}+O(e^{-\beta t}) b ⋆ = Φ ij + O ( e − β t ) , also through one derivative.
b. Small singular values and the increment estimate. Put W = L † L W=L^\dagger L W = L † L and W ⋆ = L ⋆ † L ⋆ W_\star=L_\star^\dagger L_\star W ⋆ = L ⋆ † L ⋆ . Equation (D5) makes their difference exponentially small through one derivative. The two possibly nonzero singular values of W ⋆ W_\star W ⋆ are
ξ 1 , 2 ⋆ = a ⋆ c ⋆ ± ∣ b ⋆ ∣ , a ⋆ = ∥ ℓ i ⋆ ∥ 2 , c ⋆ = ∥ ℓ j ⋆ ∥ 2 . (91) \xi_{1,2}^\star=\sqrt{a_\star c_\star}\pm\lvert b_\star\rvert,\qquad a_\star=\lVert\ell_i^\star\rVert^2,\quad c_\star=\lVert\ell_j^\star\rVert^2. \tag*{(91)} ξ 1 , 2 ⋆ = a ⋆ c ⋆ ± ∣ b ⋆ ∣ , a ⋆ = ∥ ℓ i ⋆ ∥ 2 , c ⋆ = ∥ ℓ j ⋆ ∥ 2 . (91)
Work on late intervals where ∣ Φ i j ∣ ≥ m > 0 \lvert\Phi_{ij}\rvert\ge m>0 ∣ Φ ij ∣ ≥ m > 0 . The first root and its gap from the other are then bounded below. For a nonbipartite graph or a pair in the same part, v a ( i ) / v i = v a ( j ) / v j v_a(i)/v_i=v_a(j)/v_j v a ( i ) / v i = v a ( j ) / v j for each retained channel, so the two rows are parallel for every Q ( t ) Q(t) Q ( t ) and the rank is one. For a pair in opposite parts, a ⋆ c ⋆ − ∣ b ⋆ ∣ 2 = 4 v i 2 v j 2 det Q ( t ) a_\star c_\star-\lvert b_\star\rvert^2=4v_i^2v_j^2\det Q(t) a ⋆ c ⋆ − ∣ b ⋆ ∣ 2 = 4 v i 2 v j 2 det Q ( t ) . If det Q = 0 \det Q=0 det Q = 0 , the inequalities 0 ⪯ Q ( t ) ⪯ Q 0\preceq Q(t)\preceq Q 0 ⪯ Q ( t ) ⪯ Q again give rank one. If det Q > 0 \det Q>0 det Q > 0 , the second root stays positive at late times. Thus the leading rank r ∈ { 1 , 2 } r\in\{1,2\} r ∈ { 1 , 2 } is fixed and its nonzero roots have uniform spectral gaps.
Let P P P and P ⋆ P_{\star} P ⋆ project onto these r r r right singular spaces. The resolvent identity for W † W W^{\dagger}W W † W and W ⋆ † W ⋆ W_{\star}^{\dagger}W_{\star} W ⋆ † W ⋆ , differentiated once and integrated around their separated roots, gives ξ k − ξ k ⋆ = O ( e − β t ) \xi_k-\xi_k^{\star}=O(e^{-\beta t}) ξ k − ξ k ⋆ = O ( e − β t ) and P − P ⋆ = O ( e − β t ) P-P_{\star}=O(e^{-\beta t}) P − P ⋆ = O ( e − β t ) , also through one derivative. The remaining singular values have sum ∥ B 0 ∥ 1 \lVert B_0\rVert_1 ∥ B 0 ∥ 1 , where
B 0 = W ( I − P ) = ( W − W ⋆ ) ( I − P ) + W ⋆ ( P ⋆ − P ) . (92) \begin{aligned}
B_0 &= W(\mathbb{I}-P)\\
&=(W-W_{\star})(\mathbb{I}-P)+W_{\star}(P_{\star}-P). \tag*{(92)}
\end{aligned} B 0 = W ( I − P ) = ( W − W ⋆ ) ( I − P ) + W ⋆ ( P ⋆ − P ) . (92)
This identity bounds B 0 B_0 B 0 and B ˙ 0 \dot{B}_0 B ˙ 0 by K e − β t Ke^{-\beta t} K e − β t in operator norm. Their ranks are at most four and eight, respectively, the latter because the derivative is a limit of differences of two rank-four operators. The same bounds therefore hold in trace norm.
The concurrence is positive on these intervals at late times. Using ξ 1 ⋆ − ξ 2 ⋆ = 2 ∣ b ⋆ ∣ \xi_1^{\star}-\xi_2^{\star}=2\lvert b_{\star}\rvert ξ 1 ⋆ − ξ 2 ⋆ = 2 ∣ b ⋆ ∣ gives
C i j ( t ) = 2 e − 2 γ t ∣ Φ i j ( t ) ∣ + f ( t ) − ∥ B ( t ) ∥ 1 , (93) C_{ij}(t)=2e^{-2\gamma t}\lvert\Phi_{ij}(t)\rvert+f(t)-\lVert B(t)\rVert_1, \tag*{(93)} C ij ( t ) = 2 e − 2 γ t ∣ Φ ij ( t )∣ + f ( t ) − ∥ B ( t ) ∥ 1 , (93)
∣ f ∣ + ∣ f ˙ ∣ + ∥ B ∥ 1 + ∥ B ˙ ∥ 1 ≤ K e − ( 2 γ + β ) t , (94) \lvert f\rvert+\lvert\dot{f}\rvert+\lVert B\rVert_1+\lVert\dot{B}\rVert_1\leq Ke^{-(2\gamma+\beta)t}, \tag*{(94)} ∣ f ∣ + ∣ f ˙ ∣ + ∥ B ∥ 1 + ∥ B ˙ ∥ 1 ≤ K e − ( 2 γ + β ) t , (94)
where B = e − 2 γ t B 0 B=e^{-2\gamma t}B_0 B = e − 2 γ t B 0 and f f f collects the smooth large root differences and 2 e − 2 γ t ( ∣ b ⋆ ∣ − ∣ Φ i j ∣ ) 2e^{-2\gamma t}(\lvert b_{\star}\rvert-\lvert\Phi_{ij}\rvert) 2 e − 2 γ t (∣ b ⋆ ∣ − ∣ Φ ij ∣) . The inequality ∣ ∥ B ( t + h ) ∥ 1 − ∥ B ( t ) ∥ 1 ∣ ≤ ∥ B ( t + h ) − B ( t ) ∥ 1 \lvert\lVert B(t+h)\rVert_1-\lVert B(t)\rVert_1\rvert\leq\lVert B(t+h)-B(t)\rVert_1 ∣∥ B ( t + h ) ∥ 1 − ∥ B ( t ) ∥ 1 ∣ ≤ ∥ B ( t + h ) − B ( t ) ∥ 1 now proves Eq. (38) .
c. Strict decrease and revivals. If the derivative of the leading concurrence is at most − δ e − 2 γ t -\delta e^{-2\gamma t} − δ e − 2 γ t , with δ > 0 \delta>0 δ > 0 , Eq. (38) preserves this sign in every sufficiently late increment of the exact concurrence. This proves strict decrease for a nonzero constant Φ i j \Phi_{ij} Φ ij . For pairs in the same part, the maximal derivative coefficient after dividing by 2 v i v j Tr Q e − 2 γ t 2v_iv_j\operatorname{Tr}Qe^{-2\gamma t} 2 v i v j Tr Q e − 2 γ t is
max x { − Ω ν sin x − 2 γ ( 1 + ν cos x ) } = ν Ω 2 + 4 γ 2 − 2 γ . (95) \max_x\{-\Omega\nu\sin x-2\gamma(1+\nu\cos x)\}=\nu\sqrt{\Omega^2+4\gamma^2}-2\gamma. \tag*{(95)} x max { − Ω ν sin x − 2 γ ( 1 + ν cos x )} = ν Ω 2 + 4 γ 2 − 2 γ . (95)
It is negative for ν < ν c r i t \nu<\nu_{\mathrm{crit}} ν < ν crit , and the profile is then bounded away from zero. For pairs in opposite parts with U ≠ 0 U\neq0 U = 0 , the maximal logarithmic derivative of the profile is Ω F 2 / ( 2 ∣ U ∣ U 2 + F 2 ) \Omega F^2/(2\lvert U\rvert\sqrt{U^2+F^2}) Ω F 2 / ( 2 ∣ U ∣ U 2 + F 2 ) . Comparison with 2 γ 2\gamma 2 γ gives Eq. (40) and strict decrease under the reverse inequality.
Above either threshold, each period contains separated intervals of decrease and increase for the leading concurrence. Hence both differences in Eq. (84) are positive at τ n = 0 \tau_n=0 τ n = 0 for suitable s 0 < s 1 < s 2 s_0<s_1<s_2 s 0 < s 1 < s 2 , and Appendix C 3 gives arbitrarily late revivals. Recurring zeros of the profile yield the same conclusion when ν = 1 \nu=1 ν = 1 or U = 0 < F U=0<F U = 0 < F .
With x = Ω t + ϕ x=\Omega t+\phi x = Ω t + ϕ , the coefficient maximized in Eq. (95) equals − 2 γ − ν Ω 2 + 4 γ 2 cos ( x − ϑ ) -2\gamma-\nu\sqrt{\Omega^2+4\gamma^2}\cos(x-\vartheta) − 2 γ − ν Ω 2 + 4 γ 2 cos ( x − ϑ ) . It changes sign from positive to negative at x − ϑ = − χ x-\vartheta=-\chi x − ϑ = − χ , which gives Eq. (42) . For pairs in opposite parts, put x = Ω t + ϕ ′ x=\Omega t+\phi' x = Ω t + ϕ ′ . The logarithmic derivative of the leading concurrence equals
F 2 Ω 2 + 4 γ 2 2 ( U 2 + F 2 sin 2 x ) [ cos ( 2 x − ϑ ) − ν c r i t ( 1 + 2 U 2 F 2 ) ] , \frac{F^2\sqrt{\Omega^2+4\gamma^2}}{2(U^2+F^2\sin^2x)}
\left[\cos(2x-\vartheta)-\nu_{\mathrm{crit}}\left(1+\frac{2U^2}{F^2}\right)\right], 2 ( U 2 + F 2 sin 2 x ) F 2 Ω 2 + 4 γ 2 [ cos ( 2 x − ϑ ) − ν crit ( 1 + F 2 2 U 2 ) ] ,
which changes sign from positive to negative at 2 x − ϑ = χ ′ 2x-\vartheta=\chi' 2 x − ϑ = χ ′ . This gives Eq. (43) . The conditions ν c r i t < ν \nu_{\mathrm{crit}}<\nu ν crit < ν and F 2 ( 1 − ν c r i t ) > 2 ν c r i t U 2 F^2(1-\nu_{\mathrm{crit}})>2\nu_{\mathrm{crit}}U^2 F 2 ( 1 − ν crit ) > 2 ν crit U 2 make both zeros simple, and ∣ Φ i j ∣ \lvert\Phi_{ij}\rvert ∣ Φ ij ∣ stays bounded away from zero near them. Write t n t_n t n for a maximum in either equation. At t n t_n t n , the leading concurrence has a strictly negative second derivative proportional to e − 2 γ t n e^{-2\gamma t_n} e − 2 γ t n . In a fixed small neighbourhood where this bound remains uniform, the derivative of the leading concurrence has magnitude at least c e − 2 γ t ∣ t − t n ∣ ce^{-2\gamma t}\lvert t-t_n\rvert c e − 2 γ t ∣ t − t n ∣ , with positive sign before t n t_n t n and negative sign after it. Equation (38) then makes the exact concurrence strictly increasing before t n − K e − β t n t_n-Ke^{-\beta t_n} t n − K e − β t n and strictly decreasing after t n + K e − β t n t_n+Ke^{-\beta t_n} t n + K e − β t n within that neighbourhood. By continuity, the intervening interval contains at least one local maximum, which proves the stated location bound.
Numerical methods
We construct the exact space with at most two total excitations, retaining bosonic square-root occupation factors and excluding qubit double occupation. We propagate the density matrix with the exponential of the sparse Lindblad generator using SciPy [50 ] and trace out all other qubits and memory modes to obtain the complete pair state. For the figures, the pair entries in the row and column of ∣ 11 ⟩ \lvert11\rangle ∣ 11 ⟩ come from the blocks ρ ( 2 , m ) \rho^{(2,m)} ρ ( 2 , m ) , which receive no loss feeding, and are evaluated directly from e − i H 2 t e^{-iH_2t} e − i H 2 t and e − i H 1 t e^{-iH_1t} e − i H 1 t . Local extrema are located by bounded minimization and zeros of ξ 1 − ξ 2 − ξ 3 − ξ 4 \xi_1-\xi_2-\xi_3-\xi_4 ξ 1 − ξ 2 − ξ 3 − ξ 4 by root finding, with additional samples near each zero.
Near the vacuum, direct diagonalization of the product in Eq. (24) can fail to resolve small populations accurately. We instead set d a = max ( X a a , 0 ) d_a=\sqrt{\max(X_{aa},0)} d a = max ( X aa , 0 ) and diagonalize the correlation matrix K a b = X a b / ( d a d b ) K_{ab}=X_{ab}/(d_ad_b) K ab = X ab / ( d a d b ) on its nonzero diagonal support. If K = U diag ( κ ) U † K=U\operatorname{diag}(\kappa)U^{\dagger} K = U diag ( κ ) U † , restore L = diag ( d ) U diag ( κ ) L=\operatorname{diag}(d)U\operatorname{diag}(\sqrt{\kappa}) L = diag ( d ) U diag ( κ ) . The Wootters roots are the singular values of L † L L^{\dagger}L L † L . Positivity sets a row and column to zero when its diagonal vanishes. Small negative eigenvalues of K K K , greater than − 10 − 8 -10^{-8} − 1 0 − 8 , are treated as numerical roundoff and set to zero, while more negative eigenvalues cause the calculation to terminate. We then verify the reconstructed matrix and check that it is positive semidefinite. All solid curves show the concurrence of the complete pair state.
The integral in Eq. (27) is evaluated without fitting. For each retained pair of roots, solve
− i ( H 2 − z a ) χ a b + i χ a b ( H 2 † − z b ‾ ) = − ρ ( 2 , 2 ) ( 0 ) . (96) -i(H_2-z_a)\chi_{ab}+i\chi_{ab}(H_2^{\dagger}-\overline{z_b})=-\rho^{(2,2)}(0). \tag*{(96)} − i ( H 2 − z a ) χ ab + i χ ab ( H 2 † − z b ) = − ρ ( 2 , 2 ) ( 0 ) . (96)
Its contribution to Q a b Q_{ab} Q ab is ∑ r y a † Γ r ( 2 ) χ a b Γ r ( 2 ) † y b \sum_r y_a^{\dagger}\Gamma_r^{(2)}\chi_{ab}\Gamma_r^{(2)\dagger}y_b ∑ r y a † Γ r ( 2 ) χ ab Γ r ( 2 ) † y b . Before solving, we check the sufficient convergence condition 2 γ 2 > γ a + γ b 2\gamma_2>\gamma_a+\gamma_b 2 γ 2 > γ a + γ b , where γ a = − Im z a \gamma_a=-\operatorname{Im}z_a γ a = − Im z a .
For Eq. (41) , Q ( 1 ) = 0.15 Q ( 1 ) + p 2 Q ( 2 ) Q^{(1)}=0.15Q^{(1)}+p_2Q^{(2)} Q ( 1 ) = 0.15 Q ( 1 ) + p 2 Q ( 2 ) , where Q ( 1 ) Q^{(1)} Q ( 1 ) has only Q R R ( 1 ) = 0.9933366153 Q_{RR}^{(1)}=0.9933366153 Q R R ( 1 ) = 0.9933366153 . With J 11 = 1 J_{11}=1 J 11 = 1 , unit population in ∣ 13 ⟩ \lvert13\rangle ∣ 13 ⟩ gives
Q R R ( 2 ) = Q − R , − R ( 2 ) = 0.2011667011 , Q_{RR}^{(2)}=Q_{-R,-R}^{(2)}=0.2011667011, Q R R ( 2 ) = Q − R , − R ( 2 ) = 0.2011667011 ,
Q R , − R ( 2 ) = 0.0906937803 − 0.0000495797 i . (97) Q_{R,-R}^{(2)}=0.0906937803-0.0000495797i. \tag*{(97)} Q R , − R ( 2 ) = 0.0906937803 − 0.0000495797 i . (97)
Hermiticity fixes the other entry. These values give the stated crossings p 2 = 0.29626396 p_2=0.29626396 p 2 = 0.29626396 and 0.58152236 0.58152236 0.58152236 .
For Figure 4 , the state has no coherence between excitation numbers. The only nonzero off-diagonal entries of each pair matrix are then X 23 X_{23} X 23 and X 32 X_{32} X 32 , so C = 2 max { 0 , ∣ X 23 ∣ − X 11 X 44 } C = 2\max\{0, |X_{23}| - \sqrt{X_{11}X_{44}}\} C = 2 max { 0 , ∣ X 23 ∣ − X 11 X 44 } . Each exact maximum after the last zero of C C C is a root of d C / d t dC/dt d C / d t , bracketed by a sign change on a grid of step 0.5 / λ 0.5/\lambda 0.5/ λ . The evolution of the other entries of the full state does not involve its vacuum population. We therefore exclude the vacuum population during propagation and reconstruct X 44 X_{44} X 44 from the unit trace, avoiding numerical errors dominated by the much larger vacuum contribution. A closed form built from the eigendecompositions of H 1 H_1 H 1 and H 2 H_2 H 2 , evaluated in 40-digit arithmetic, reproduces every computed maximum to within 10 − 11 / λ 10^{-11}/\lambda 1 0 − 11 / λ .
An independent implementation using tensor-product operators reproduces the matrix elements obtained from the primary construction. Density matrix integration with DOP853, at relative and absolute tolerances 2 × 10 − 12 2 \times10^{-12} 2 × 1 0 − 12 and 2 × 10 − 14 2 \times10^{-14} 2 × 1 0 − 14 , agrees with sparse exponential propagation to within 1.5 × 10 − 12 1.5 \times10^{-12} 1.5 × 1 0 − 12 in the tested pair matrices, including those for the feeding preparation and the detuned K 4 K_4 K 4 state. A general K 4 K_4 K 4 test mixes 98% of Eq. (46) at p = 0.01 p = 0.01 p = 0.01 with 2% of a random density operator on Q 0 ⊕ Q 1 ⊕ Q 2 Q_0 \oplus Q_1 \oplus Q_2 Q 0 ⊕ Q 1 ⊕ Q 2 , generated from an 11 × 11 11 \times11 11 × 11 complex Gaussian matrix with NumPy seed 20260923. Every cross-sector coherence block is nonzero. The calculation retains all least damped modes and includes the B 21 B_{21} B 21 contribution to K 10 K_{10} K 10 . Quadrature and Sylvester evaluations agree within 1.4 × 10 − 12 1.4 \times10^{-12} 1.4 × 1 0 − 12 . At λ t = 100 \lambda t = 100 λ t = 100 , X 12 ( t ) X_{12}(t) X 12 ( t ) differs from X 12 ( 0 ) ( t ) X_{12}^{(0)}(t) X 12 ( 0 ) ( t ) by less than 2.2 × 10 − 9 2.2 \times10^{-9} 2.2 × 1 0 − 9 entrywise.
Additional checks cover 200 random pair matrices, the exact equal-rate preparation at p = 0.01 p = 0.01 p = 0.01 , 1 / 2 1/2 1/2 , repeated channel roots, the example with Q R R = 0 Q_{RR} = 0 Q R R = 0 , and both strict sides of each bipartite threshold. The latter two tests use g / λ = 0.3 g/\lambda= 0.3 g / λ = 0.3 , 1 1 1 , 2 2 2 .
The spectral checks cover 56 resonant cases on seven graphs, the exact detuned K 4 K_4 K 4 equality point, and 276 additional parameter choices spanning resonance and nonzero detuning, up to g / λ = 30 g/\lambda= 30 g / λ = 30 . The smallest computed value of γ 2 − 2 γ 1 \gamma_2 - 2\gamma_1 γ 2 − 2 γ 1 at nonzero detuning is above − 10 − 14 λ -10^{-14}\lambda − 1 0 − 14 λ , within roundoff, and selected H 2 H_2 H 2 eigenpair residuals are below 10 − 14 10^{-14} 1 0 − 14 relative to the matrix norm.