Thermalisation is a cornerstone of thermodynamics: a system placed in contact with a much larger environment at temperature T T T and chemical potential μ \mu μ generically relaxes to a state characterised by these two parameters alone. In classical statistical mechanics this is the grand canonical Boltzmann distribution, quantum mechanically it is the grand canonical Gibbs state
τ S = e − ( H S − μ N S ) / k B T Z S , (1) \tau_S = \frac{e^{-(H_S-\mu N_S)/k_B T}}{Z_S}, \tag*{(1)} τ S = Z S e − ( H S − μ N S ) / k B T , (1)
with Z S ≡ Tr ( e − ( H S − μ N S ) / k B T ) Z_S \equiv\operatorname{Tr}(e^{-(H_S-\mu N_S)/k_B T}) Z S ≡ Tr ( e − ( H S − μ N S ) / k B T ) the grand partition function. This state is determined entirely by the system Hamiltonian H S H_S H S , the system particle number N S N_S N S , and the temperature and chemical potential of the environment [1 ] . Equation (1) underpins much of quantum thermodynamics, it fixes the free state in thermodynamic resource theories [2 , 3 ] and is the reference equilibrium against which the performance of quantum batteries [4 , 5 ] and nanoscale heat engines [6 , 7 ] are assessed.
That the equilibrium state should depend only on H S H_S H S and N S N_S N S is, however, an assumption. It rests on the interaction energy between system and environment being negligible compared with the system’s own energy scales, so that the interaction does not contribute to the form of the equilibrium state. This is an accurate assumption for macroscopic systems but fails for microscopic ones, precisely the regime quantum thermodynamics is concerned with. For example, in quantum dots and molecular junctions the system-environment coupling is routinely comparable to both the level spacing and the temperature [8 , 9 , 10 , 11 ] . This motivates the field of strong coupling thermodynamics [12 , 13 , 14 , 15 , 16 , 17 , 18 ] , which retains the system–environment interaction explicitly in thermodynamic quantities. A central object of the field is the mean force Gibbs (MFG) state [12 , 13 ] . Suppose system and environment have jointly equilibrated, so that their combined state is the total grand canonical Gibbs state τ t o t = e − ( H − μ N ) / k B T / Tr ( e − ( H − μ N ) / k B T ) \tau_{\mathrm{tot}} = e^{-(H-\mu N)/k_B T}/\operatorname{Tr}(e^{-(H-\mu N)/k_B T}) τ tot = e − ( H − μ N ) / k B T / Tr ( e − ( H − μ N ) / k B T ) with H = H S + H R + λ V H = H_S + H_R + \lambda V H = H S + H R + λV where λ V \lambda V λV is the system-reservoir interaction with coupling strength λ \lambda λ and N = N S + N R N = N_S + N_R N = N S + N R the total particle number. This is the appropriate description whenever the combined object is itself macroscopic, either because it has equilibrated with a weakly coupled “super-bath” [19 , 20 ] at temperature T T T and chemical potential μ \mu μ , or by appeal to the eigenstate thermalisation hypothesis [21 , 22 ] . The reduced state of the system is then the (grand canonical) MFG state
τ M F = Tr R ( τ t o t ) . (2) \tau_{\mathrm{MF}} = \operatorname{Tr}_R(\tau_{\mathrm{tot}}). \tag*{(2)} τ MF = Tr R ( τ tot ) . (2)
Note that in the following, Gibbs and MFG state refer to the grand canonical versions unless stated otherwise. In the limit of weak coupling, this reduces to the Gibbs state of Eq. (1) . Unlike τ S \tau_S τ S , the MFG state need not commute with H S H_S H S : it can retain coherences in the bare energy eigenbasis [23 , 24 , 25 , 26 ] , a feature with no classical counterpart and one that is absent from the steady state predicted by standard secular master equations. This is a particularly attractive feature, as coherence is often seen as a resource [27 ] , for example to improve the efficiency of heat engines [28 ] .
Despite its compact definition, Eq. (2) is hard to evaluate and analytic progress has been limited to particular models or to asymptotic coupling regimes. Of particular note is the work of Cresser and Anders Cresser and Anders [23 ] , who obtained closed-form expressions for an arbitrary system coupled to a bosonic reservoir in both the weak and ultrastrong coupling limits, subsequently extended to anharmonic environments [29 ] for the ultrastrong limit, multiple channels [30 ] and complemented by effective-Hamiltonian [31 ] and numerically exact [32 , 33 ] treatments. However, these works are limited to the bosonic case. For fermions, even the definition of the MFG state given in Eq. (2) is ambiguous as the partial trace operation is defined for a Hilbert space with a tensor product structure, which fermions do not possess due to their canonical anti-commutation relation (CAR) algebra. This necessitates the use of the fermionic partial trace, defined via the mode reduced state [34 , 35 ] . The usual coupling operator is also different to the standard bosonic case, where for a canonical fermionic open system the interaction transfers a particle between system and reservoir
V = X † E + E † X , (3) V = X^\dagger E + E^\dagger X, \tag*{(3)} V = X † E + E † X , (3)
where X X X is an odd-parity system operator, typically an annihilation operator, and E E E is a linear combination of reservoir annihilation operators. This differs from X ⊗ B X \otimes B X ⊗ B with Hermitian factors considered in [23 ] .
In this work we close this gap. Our main result, Eq. (11) , is a closed-form expression for the MFG state of a generic fermionic system coupled through Eq. (3) to a number-conserving fermionic reservoir up to second order in the coupling. We use the fermionic partial trace as defined in [34 , 35 ] ensuring its validity. The Kubo-Martin-Schwinger condition fixes all coefficient functions in terms of a single reservoir function. We derive both the low and high temperature limits where the Gibbs state is recovered at high temperature and coherences in the bare energy eigenbasis persist at low temperature. We also derive the corresponding Hamiltonian of mean force. Exchange coupling is also a legitimate choice for bosonic reservoirs and Eq. (11) covers both statistics, although its main content is the fermionic case.
We verify Eq. (11) against numerically exact calculations, showing that at weak and intermediate coupling the steady state agrees with the MFG state predicted by our formula. There is no general proof that the MFG state is the long-time state of an open system [13 ] , and the existing numerical and analytical evidence mostly concerns bosonic environments [26 , 32 , 33 , 36 ] . So for all the examples we consider, we compare our static formula with the steady state obtained from real-time dynamics. For quadratic Hamiltonians, which include the first two examples we consider, both the exact dynamics and the exact MFG state follow directly from the single-particle correlation matrix [37 ] (see also Sec. SVI A). For interacting systems no such result is available. For the single impurity Anderson model we obtain the steady state from real-time dynamics, using a matrix product state (MPS) description of the reservoir based on the thermofield chain mapping [38 , 39 , 40 , 41 ] . For all three examples we find agreement with our static prediction, providing evidence that the MFG state is indeed the relevant equilibrium state for fermionic open systems. The paper is organised as follows. Section I introduces the setup and the fermionic partial trace. Section
tion II presents the main result and its general properties. Section III outlines the numerical approach we use to benchmark our result, which is applied in Section IV to the resonant level model, two modes coupled to a shared reservoir, the single impurity Anderson model. The Supplemental Material contains the derivations for Eq. (11) , the high- and low-temperature limits, the Hamiltonian of mean force, the examples and additional details on the non-interacting case.
Setup
We consider the canonical open-system setting
H = H S + H R + λ V , (4) H = H_S + H_R + \lambda V, \tag*{(4)} H = H S + H R + λV , (4)
where H S H_S H S is the system Hamiltonian, H R H_R H R the reservoir Hamiltonian and λ V \lambda V λV their interaction. We use the standard continuum description for a Gaussian reservoir
H R = ∫ d ε ε a ε † a ε , (5) H_R = \int\mathrm{d}\varepsilon\, \varepsilon\, a_\varepsilon^\dagger a_\varepsilon, \tag*{(5)} H R = ∫ d ε ε a ε † a ε , (5)
V = ∫ d ε J ( ε ) ( a ε † X + X † a ε ) , (6) V = \int\mathrm{d}\varepsilon\sqrt{J(\varepsilon)} \left(a_\varepsilon^\dagger X + X^\dagger a_\varepsilon\right), \tag*{(6)} V = ∫ d ε J ( ε ) ( a ε † X + X † a ε ) , (6)
where a ε † , a ε a_\varepsilon^\dagger,a_\varepsilon a ε † , a ε create and annihilate a reservoir excitation at frequency ε \varepsilon ε , satisfying { a ε , a ε ′ † } = δ ( ε − ε ′ ) \{a_\varepsilon,a_{\varepsilon'}^\dagger\} = \delta(\varepsilon-\varepsilon') { a ε , a ε ′ † } = δ ( ε − ε ′ ) for a fermionic reservoir and [ a ε , a ε ′ † ] = δ ( ε − ε ′ ) [a_\varepsilon,a_{\varepsilon'}^\dagger] = \delta(\varepsilon-\varepsilon') [ a ε , a ε ′ † ] = δ ( ε − ε ′ ) for a bosonic one. We use natural units k B = ℏ = 1 k_B = \hbar= 1 k B = ℏ = 1 throughout and write β ≡ 1 / T \beta\equiv1/T β ≡ 1/ T for the inverse temperature. Here λ \lambda λ is the coupling strength, which for the normalisation ∫ d ε J ( ε ) = 1 \int\mathrm{d}\varepsilon J(\varepsilon) = 1 ∫ d ε J ( ε ) = 1 used below carries units of energy. X X X is a system operator and J ( ε ) J(\varepsilon) J ( ε ) is the reservoir spectral density. Note that X X X need not be Hermitian and in general will not be, rather it is usually a fermionic annihilation operator such that V V V describes hoppings between the system and the environment at each frequency ε \varepsilon ε with strength λ J ( ε ) \lambda\sqrt{J(\varepsilon)} λ J ( ε ) . We take the coupling to conserve the total particle number N = N S + N R N = N_S + N_R N = N S + N R , where
N R = ∫ d ε a ε † a ε N_R = \int\mathrm{d}\varepsilon\, a_\varepsilon^\dagger a_\varepsilon N R = ∫ d ε a ε † a ε
and N S N_S N S is the system number operator, with [ H S , N S ] = 0 [H_S,N_S] = 0 [ H S , N S ] = 0 . Since [ N R , a ε ] = − a ε [N_R,a_\varepsilon] = -a_\varepsilon [ N R , a ε ] = − a ε , this requires [ N S , X ] = − X [N_S,X] = -X [ N S , X ] = − X , i.e. X X X removes a single particle from the system, as is the case when X X X is a fermionic annihilation operator. Then [ H , N ] = 0 [H,N] = 0 [ H , N ] = 0 , and the global equilibrium state entering Eq. (2) is the grand canonical Gibbs state τ t o t ∝ e − β ( H − μ N ) \tau_{\mathrm{tot}} \propto e^{-\beta(H-\mu N)} τ tot ∝ e − β ( H − μ N ) . Its generator can be written as
H − μ N = K S + K R + λ V , (7) H - \mu N = K_S + K_R + \lambda V, \tag*{(7)} H − μ N = K S + K R + λV , (7)
K S ≡ H S − μ N S , K R ≡ ∫ d ε ( ε − μ ) a ε † a ε , K_S \equiv H_S - \mu N_S,\qquad K_R \equiv\int\mathrm{d}\varepsilon\,(\varepsilon-\mu)a_\varepsilon^\dagger a_\varepsilon, K S ≡ H S − μ N S , K R ≡ ∫ d ε ( ε − μ ) a ε † a ε ,
so that the system Gibbs state of Eq. (1) is τ S = e − β K S / Z S \tau_S = e^{-\beta K_S}/Z_S τ S = e − β K S / Z S .
As mentioned in the introduction, for fermions the partial trace in Eq. (2) needs care. This can be dealt with via the definition of the reduced state. For a state ρ \rho ρ of the system and reservoir, the reduced state ρ S ≡ Tr R ( ρ ) \rho_S \equiv\operatorname{Tr}_R(\rho) ρ S ≡ Tr R ( ρ ) is the unique operator on the system Fock space F S \mathcal{F}_S F S such that
Tr S ( ρ S O S ) = Tr ( ρ O S ) ∀ O S ∈ A S , (8) \operatorname{Tr}_S(\rho_S O_S)=\operatorname{Tr}(\rho O_S)\quad\forall O_S\in\mathcal{A}_S, \tag*{(8)} Tr S ( ρ S O S ) = Tr ( ρ O S ) ∀ O S ∈ A S , (8)
where A S \mathcal{A}_S A S is the algebra generated by the system operators s ^ i † s ^ i , s ^ i s ^ j † , s ^ i , s ^ i † \hat{s}_i^\dagger\hat{s}_i,\hat{s}_i\hat{s}_j^\dagger,\hat{s}_i,\hat{s}_i^\dagger s ^ i † s ^ i , s ^ i s ^ j † , s ^ i , s ^ i † for i = 1 , … , L S i=1,\ldots,L_S i = 1 , … , L S where L S L_S L S is the number of system modes and { s ^ i , s ^ j † } = δ i j \{\hat{s}_i,\hat{s}_j^\dagger\}=\delta_{ij} { s ^ i , s ^ j † } = δ ij . This defines what is meant by Tr R \operatorname{Tr}_R Tr R in Eq. (2) , which coincides with the partial trace for bosons due to the underlying tensor product structure, but for fermions this is not always the case [34 , 35 ] . To define states in a Fock space, a specific mode ordering must be chosen, for example in the full Fock space F \mathcal{F} F states can be defined as
∣ n S , m R ⟩ = ( s ^ 1 † ) n 1 … ( s ^ L S † ) n L S ( a ^ 1 † ) m 1 … ( a ^ L R † ) m L R ∣ v a c ⟩ , (9) \lvert n_S,m_R\rangle=(\hat{s}_1^\dagger)^{n_1}\ldots(\hat{s}_{L_S}^\dagger)^{n_{L_S}}(\hat{a}_1^\dagger)^{m_1}\ldots(\hat{a}_{L_R}^\dagger)^{m_{L_R}}\lvert\mathrm{vac}\rangle, \tag*{(9)} ∣ n S , m R ⟩ = ( s ^ 1 † ) n 1 … ( s ^ L S † ) n L S ( a ^ 1 † ) m 1 … ( a ^ L R † ) m L R ∣ vac ⟩ , (9)
where L S , L R L_S,L_R L S , L R are the number of system and reservoir modes respectively and we have put the system modes first, where for notational simplicity we have used a finite, discrete set of reservoir modes. In [34 ] , it is shown that if the modes being traced out form a contiguous set in the given mode ordering and the global state (τ t o t \tau_{\mathrm{tot}} τ tot in our case) is parity symmetric which is guaranteed from superselection rules [34 , 42 , 43 ] , then the reduced state ρ S \rho_S ρ S coincides with the operator obtained from the conventional partial trace under a Jordan Wigner transform (See Sec. SI A for more details). We thus fix the mode ordering as Eq. (9) from the start.
Second order MFG state and Hamiltonian of Mean Force
Here we present the main result, a closed form expression for the mean force Gibbs state for a total fermionic Hamiltonian given by Eqs. (4) -(6) up to second order in λ \lambda λ , with an error of O ( λ 4 ) O(\lambda^4) O ( λ 4 ) . We decompose the coupling operator in the eigenoperator basis of K S K_S K S ,
X = ∑ ω X ω , X ω = ∑ n , m : ε n − ε m = ω Π n X Π m , (10) X=\sum_\omega X_\omega,\qquad X_\omega=\sum_{n,m:\,\varepsilon_n-\varepsilon_m=\omega}\Pi_nX\Pi_m, \tag*{(10)} X = ω ∑ X ω , X ω = n , m : ε n − ε m = ω ∑ Π n X Π m , (10)
where [ K S , X ω ] = ω X ω [K_S,X_\omega]=\omega X_\omega [ K S , X ω ] = ω X ω and Π n \Pi_n Π n projects onto the (possibly degenerate) eigenspace of K S K_S K S with eigenvalue ε n \varepsilon_n ε n . Projecting onto distinct Bohr frequencies rather than distinct transitions allows for degenerate energy gaps. Because every matrix element of X X X lowers N S N_S N S by one, each ω \omega ω is a Bohr frequency of H S H_S H S shifted by μ \mu μ , and the operators X ω X_\omega X ω coincide with those obtained by decomposing X X X with respect to H S H_S H S . Combining Eq. (10) with the Kubo expansion of e − β ( H − μ N ) e^{-\beta(H-\mu N)} e − β ( H − μ N ) [23 , 44 , 45 , 46 ] and performing the reservoir trace (See Sec. SI for details) gives the MFG state
τ M F ( 2 ) = τ S + λ 2 ( Δ d i a g + Δ o f f ) , (11) \tau_{\mathrm{MF}}^{(2)}=\tau_S+\lambda^2(\Delta_{\mathrm{diag}}+\Delta_{\mathrm{off}}), \tag*{(11)} τ MF ( 2 ) = τ S + λ 2 ( Δ diag + Δ off ) , (11)
Δ d i a g = ∑ ω τ S ( X ω † X ω − Tr S [ τ S X ω † X ω ] ) D 1 ( ω ) + τ S ( X ω X ω † − Tr S [ τ S X ω X ω † ] ) D 2 ( ω ) , (12) \Delta_{\mathrm{diag}}=\sum_\omega\tau_S\left(X_\omega^\dagger X_\omega-\operatorname{Tr}_S[\tau_SX_\omega^\dagger X_\omega]\right)D_1(\omega)
+\tau_S\left(X_\omega X_\omega^\dagger-\operatorname{Tr}_S[\tau_SX_\omega X_\omega^\dagger]\right)D_2(\omega), \tag*{(12)} Δ diag = ω ∑ τ S ( X ω † X ω − Tr S [ τ S X ω † X ω ] ) D 1 ( ω ) + τ S ( X ω X ω † − Tr S [ τ S X ω X ω † ] ) D 2 ( ω ) , (12)
Δ o f f = ∑ ω ≠ ω ′ ( ω ′ − ω ) − 1 [ ( X ω ′ X ω † τ S + τ S X ω X ω ′ † ) A 2 ( ω ) − ( X ω ′ † X ω τ S + τ S X ω † X ω ′ ) A 1 ( ω ) ] . (13) \Delta_{\mathrm{off}}=\sum_{\omega\ne\omega'}(\omega'-\omega)^{-1}\left[(X_{\omega'}X_\omega^\dagger\tau_S+\tau_SX_\omega X_{\omega'}^\dagger)A_2(\omega)
-(X_{\omega'}^\dagger X_\omega\tau_S+\tau_SX_\omega^\dagger X_{\omega'})A_1(\omega)\right]. \tag*{(13)} Δ off = ω = ω ′ ∑ ( ω ′ − ω ) − 1 [ ( X ω ′ X ω † τ S + τ S X ω X ω ′ † ) A 2 ( ω ) − ( X ω ′ † X ω τ S + τ S X ω † X ω ′ ) A 1 ( ω ) ] . (13)
The four coefficient functions are imaginary-time transforms of the two reservoir correlation functions
G 1 ( β , − i τ ) = Tr R [ τ R E ( − i τ ) E † ] , (14) G_1(\beta,-i\tau)=\operatorname{Tr}_R[\tau_R E(-i\tau)E^\dagger], \tag*{(14)} G 1 ( β , − i τ ) = Tr R [ τ R E ( − i τ ) E † ] , (14)
G 2 ( β , − i τ ) = Tr R [ τ R E † ( − i τ ) E ] , (15) G_2(\beta,-i\tau)=\operatorname{Tr}_R[\tau_R E^\dagger(-i\tau)E], \tag*{(15)} G 2 ( β , − i τ ) = Tr R [ τ R E † ( − i τ ) E ] , (15)
where E = ∫ d ε J ( ε ) a ε E=\int d\varepsilon\sqrt{J(\varepsilon)}a_\varepsilon E = ∫ d ε J ( ε ) a ε , and the imaginary-time evolution is generated by K R K_R K R , E ( − i τ ) = e τ K R E e − τ K R E(-i\tau)=e^{\tau K_R}Ee^{-\tau K_R} E ( − i τ ) = e τ K R E e − τ K R . Namely,
A 1 ( ω ) = ∫ 0 β d τ e − τ ω G 1 ( β , − i τ ) , (16) A_1(\omega)=\int_0^\beta d\tau\,e^{-\tau\omega}G_1(\beta,-i\tau), \tag*{(16)} A 1 ( ω ) = ∫ 0 β d τ e − τ ω G 1 ( β , − i τ ) , (16)
A 2 ( ω ) = ∫ 0 β d τ e τ ω G 2 ( β , − i τ ) , (17) A_2(\omega)=\int_0^\beta d\tau\,e^{\tau\omega}G_2(\beta,-i\tau), \tag*{(17)} A 2 ( ω ) = ∫ 0 β d τ e τ ω G 2 ( β , − i τ ) , (17)
D 1 ( ω ) = β A 1 ( ω ) + A 1 ′ ( ω ) , D 2 ( ω ) = β A 2 ( ω ) − A 2 ′ ( ω ) . (18) D_1(\omega)=\beta A_1(\omega)+A_1'(\omega),\qquad D_2(\omega)=\beta A_2(\omega)-A_2'(\omega). \tag*{(18)} D 1 ( ω ) = β A 1 ( ω ) + A 1 ′ ( ω ) , D 2 ( ω ) = β A 2 ( ω ) − A 2 ′ ( ω ) . (18)
Closed forms for A 1 , 2 A_{1,2} A 1 , 2 and D 1 , 2 D_{1,2} D 1 , 2 in terms of J ( ε ) J(\varepsilon) J ( ε ) and the Fermi (or Bose) function are given in Sec. SI F, Eqs. (S51)–(S54). Two correlation functions appear, one for each term of the exchange coupling. G 1 G_1 G 1 virtually propagates a particle into the reservoir, and G 2 G_2 G 2 virtually propagates a hole [47 ] . They are not independent, however. Because the reservoir is thermal, the imaginary-time Kubo–Martin–Schwinger condition G 1 ( β , − i τ ) = G 2 ( β , − i ( β − τ ) ) G_1(\beta,-i\tau)=G_2(\beta,-i(\beta-\tau)) G 1 ( β , − i τ ) = G 2 ( β , − i ( β − τ )) holds, and (see Sec. SI E)
A 1 ( ω ) = e − β ω A 2 ( ω ) , D 1 ( ω ) = e − β ω A 2 ′ ( ω ) . (19) A_1(\omega)=e^{-\beta\omega}A_2(\omega),\qquad D_1(\omega)=e^{-\beta\omega}A_2'(\omega). \tag*{(19)} A 1 ( ω ) = e − β ω A 2 ( ω ) , D 1 ( ω ) = e − β ω A 2 ′ ( ω ) . (19)
All four coefficients are therefore fixed by A 2 A_2 A 2 and its derivative. Since every ω \omega ω is a Bohr frequency of H S H_S H S shifted by μ \mu μ , the denominators ω ′ − ω \omega'-\omega ω ′ − ω of Δ o f f \Delta_{\mathrm{off}} Δ off and those inside A 1 , 2 A_{1,2} A 1 , 2 and D 1 , 2 D_{1,2} D 1 , 2 do not depend on μ \mu μ ; the chemical potential enters only through τ S \tau_S τ S and through the reservoir occupation. An equivalent form, in which the coefficient functions remain bounded as T → 0 T\to0 T → 0 and which makes both the low-temperature limit and the comparison with the bosonic result of Ref. [23 ] direct, is given in Sec. SI I. It is also possible to write τ M F ( 2 ) \tau_{\mathrm{MF}}^{(2)} τ MF ( 2 ) in exponential form, in terms of a Hamiltonian of mean force. Writing τ M F ∝ e − β ( H M F − μ N S ) \tau_{\mathrm{MF}} \propto e^{-\beta(H_{\mathrm{MF}}-\mu N_S)} τ MF ∝ e − β ( H MF − μ N S ) , we have (see Sec. SIV)
H M F = H S + λ 2 K ( 2 ) + O ( λ 4 ) , H_{\mathrm{MF}}=H_S+\lambda^2K^{(2)}+O(\lambda^4), H MF = H S + λ 2 K ( 2 ) + O ( λ 4 ) ,
K ( 2 ) = − 1 β ∑ ω [ D 1 ( ω ) X ω † X ω + D 2 ( ω ) X ω X ω † ] − ∑ ω ≠ ω ′ [ A 2 ( ω ′ ) − A 2 ( ω ) e β ω ′ − e β ω X ω † X ω ′ + A 1 ( ω ′ ) − A 1 ( ω ) e − β ω ′ − e − β ω X ω X ω ′ † ] , (20) K^{(2)}=-\frac{1}{\beta}\sum_{\omega}\left[D_1(\omega)X_\omega^\dagger X_\omega+D_2(\omega)X_\omega X_\omega^\dagger\right]
-\sum_{\omega\ne\omega'}\left[\frac{A_2(\omega')-A_2(\omega)}{e^{\beta\omega'}-e^{\beta\omega}}X_\omega^\dagger X_{\omega'}
+\frac{A_1(\omega')-A_1(\omega)}{e^{-\beta\omega'}-e^{-\beta\omega}}X_\omega X_{\omega'}^\dagger\right], \tag*{(20)} K ( 2 ) = − β 1 ω ∑ [ D 1 ( ω ) X ω † X ω + D 2 ( ω ) X ω X ω † ] − ω = ω ′ ∑ [ e β ω ′ − e β ω A 2 ( ω ′ ) − A 2 ( ω ) X ω † X ω ′ + e − β ω ′ − e − β ω A 1 ( ω ′ ) − A 1 ( ω ) X ω X ω ′ † ] , (20)
where H M F = K M F + μ N S = H S + λ 2 K ( 2 ) H_{\mathrm{MF}}=K_{\mathrm{MF}}+\mu N_S=H_S+\lambda^2K^{(2)} H MF = K MF + μ N S = H S + λ 2 K ( 2 ) because K ( 2 ) K^{(2)} K ( 2 ) conserves N S N_S N S . Note that this form is more restrictive than Eq. (11) . Gibbs state calculated from Eq. (20) reproduces Eq. (11) only when the mean force corrections to the populations are small compared with the bare populations themselves. As T → 0 T\to0 T → 0 the bare populations of excited states vanish exponentially while their corrections remain finite, so the exponential form is restricted to sufficiently high temperatures, whereas Eq. (11) is not (see Sec. SIV).
Both X ω † X ω X_\omega^\dagger X_\omega X ω † X ω and X ω X ω † X_\omega X_\omega^\dagger X ω X ω † commute with H S H_S H S , so Δ d i a g \Delta_{\mathrm{diag}} Δ diag is block diagonal with respect to the eigenspaces of H S H_S H S , and diagonal if H S H_S H S is non-degenerate [48 ] . The off-diagonal sum is not,
[ τ M F ( 2 ) , H S ] = λ 2 ∑ ω ≠ ω ′ { A 2 ( ω ) [ τ S X ω X ω ′ † − X ω ′ X ω † τ S ] + A 1 ( ω ) [ τ S X ω † X ω ′ − X ω ′ † X ω τ S ] } , (21) [\tau_{\mathrm{MF}}^{(2)},H_S]=\lambda^2\sum_{\omega\ne\omega'}\left\{A_2(\omega)\left[\tau_SX_\omega X_{\omega'}^\dagger-X_{\omega'}X_\omega^\dagger\tau_S\right]+A_1(\omega)\left[\tau_SX_\omega^\dagger X_{\omega'}-X_{\omega'}^\dagger X_\omega\tau_S\right]\right\}, \tag*{(21)} [ τ MF ( 2 ) , H S ] = λ 2 ω = ω ′ ∑ { A 2 ( ω ) [ τ S X ω X ω ′ † − X ω ′ X ω † τ S ] + A 1 ( ω ) [ τ S X ω † X ω ′ − X ω ′ † X ω τ S ] } , (21)
which is generically non-zero. As in the bosonic case [23 ] , the equilibrium state retains coherences across energy sectors in the eigenbasis of H S H_S H S . The high- and low-temperature limits are derived in Secs. SII and SIII respectively. At high temperature the correction vanishes, τ M F ( 2 ) = τ S + O ( λ 2 β 2 ) \tau_{\mathrm{MF}}^{(2)}=\tau_S+O(\lambda^2\beta^2) τ MF ( 2 ) = τ S + O ( λ 2 β 2 ) , faster than the O ( β ) O(\beta) O ( β ) deviation of τ S \tau_S τ S itself from the maximally mixed state, and as O ( λ 2 β 3 ) O(\lambda^2\beta^3) O ( λ 2 β 3 ) when X X X is linear in system annihilation operators, as in all three examples below. At low temperature the coherences persist.
Numerical methods
To validate Eq. (11) , we compare to numerics. When the system Hamiltonian is quadratic in mode operators, H S = ∑ i j h i j s i † s j H_S=\sum_{ij}h_{ij}s_i^\dagger s_j H S = ∑ ij h ij s i † s j , the total Hamiltonian H = ∑ i j H i j d i † d j H=\sum_{ij}H_{ij}d_i^\dagger d_j H = ∑ ij H ij d i † d j is also quadratic, where { d i } \{d_i\} { d i } denotes both the system and reservoir modes in the ordering of Eq. (9) . For Gaussian initial states, both the real-time dynamics and the MFG state can then be obtained exactly from the single particle correlation matrix C i j ( t ) = Tr [ ρ ( t ) d i † d j ] C_{ij}(t)=\operatorname{Tr}[\rho(t)d_i^\dagger d_j] C ij ( t ) = Tr [ ρ ( t ) d i † d j ] , which evolves unitarily as (see Sec. SVI)
C ( t ) = e − i H t C ( 0 ) e i H t , (22) C(t)=e^{-iHt}C(0)e^{iHt}, \tag*{(22)} C ( t ) = e − i H t C ( 0 ) e i H t , (22)
where ρ ( t ) \rho(t) ρ ( t ) is the combined state of the system and reservoir. For the MFG state, we write the generator of Eq. (7) as H − μ N = ∑ i j K i j d i † d j H-\mu N=\sum_{ij}K_{ij}d_i^\dagger d_j H − μ N = ∑ ij K ij d i † d j , so that the correlation matrix of the global Gibbs state is C = f ( K ) C=f(K) C = f ( K ) , where f ( w ) = ( e β w + 1 ) − 1 f(w)=(e^{\beta w}+1)^{-1} f ( w ) = ( e β w + 1 ) − 1 . As the MFG state is Gaussian, it is determined by the system block C S = [ f ( K ) ] i , j ∈ S C_S=[f(K)]_{i,j\in S} C S = [ f ( K ) ] i , j ∈ S alone [37 , 49 ] (see also Sec. SVI A),
τ M F = exp ( − ∑ i j M i j s i † s j ) Tr S exp ( − ∑ i j M i j s i † s j ) , M = ln ( C S − 1 − 1 ) , (23) \tau_{\mathrm{MF}}=\frac{\exp\left(-\sum_{ij}M_{ij}s_i^\dagger s_j\right)}{\operatorname{Tr}_S\exp\left(-\sum_{ij}M_{ij}s_i^\dagger s_j\right)},\qquad M=\ln(C_S^{-1}-1), \tag*{(23)} τ MF = Tr S exp ( − ∑ ij M ij s i † s j ) exp ( − ∑ ij M ij s i † s j ) , M = ln ( C S − 1 − 1 ) , (23)
so that β − 1 M \beta^{-1}M β − 1 M is the single particle matrix of H M F − μ N S H_{\mathrm{MF}}-\mu N_S H MF − μ N S , and occupations and coherences can be read off directly as ⟨ s i † s j ⟩ M F = ( C S ) i j \langle s_i^\dagger s_j\rangle_{\mathrm{MF}}=(C_S)_{ij} ⟨ s i † s j ⟩ MF = ( C S ) ij . For interacting systems there is no exact solution in general, and a numerical method must be used. For this we use the thermofield chain mapping [38 , 39 , 40 , 41 ] , computing the real-time dynamics until convergence from an initial product state ρ ( 0 ) = ρ 0 T R \rho(0)=\rho_0T_R ρ ( 0 ) = ρ 0 T R . The thermal reservoir is purified with an auxiliary copy, and a Bogoliubov rotation of the doubled reservoir gives two independent reservoirs with spectral densities J ( ϵ ) f ( ϵ − μ ) J(\epsilon)f(\epsilon-\mu) J ( ϵ ) f ( ϵ − μ ) and J ( ϵ ) [ 1 − f ( ϵ − μ ) ] J(\epsilon)[1-f(\epsilon-\mu)] J ( ϵ ) [ 1 − f ( ϵ − μ )] , which are initially filled and empty respectively. Each reservoir is mapped onto a semi-infinite tight-binding chain using orthogonal polynomials [50 ] , with the system coupled only to the first site of each chain. We represent the system and both chains as an MPS and compute the dynamics using the time-dependent variational principle [51 ] as implemented in ITensor [52 ] . For more details on the approach see [40 , 41 ] . We use chains of O ( 100 ) O(100) O ( 100 ) sites, a time step δ t = 0.1 / D \delta t=0.1/D δ t = 0.1/ D , a maximum bond dimension of 400 and a truncation cutoff of 10 − 9 10^{-9} 1 0 − 9 , all of which we have checked for convergence. With the system placed first in the mode ordering, its reduced state coincides with the fermionic reduced state (see Sec. SIA). For the quadratic models we evolve Eq. (22) in the same chain geometry [40 ] . To evaluate Eq. (23) we instead apply the orthogonal polynomial mapping to the physical spectral density J ( ϵ ) J(\epsilon) J ( ϵ ) . We find C S C_S C S converged to 10 − 12 10^{-12} 1 0 − 12 when the chain is truncated at 200 sites.
Examples
We now consider some examples. For the two quadratic models we compute the exact MFG state via Eq. (23) , and have confirmed it agrees with the long-time state of the exact correlation-matrix dynamics as described by Eq. (22) for both the chain mapping discretisation discussed in Sec. III and linear discretisation [53 ] . We therefore refer to both as the exact result. For the interacting SIAM we compare Eq. (11) to the converged real-time dynamics using the thermofield approach as described in Sec. III. For all cases we choose a semi-elliptical spectral density J ( ϵ ) = 2 π D 1 − ( ϵ / D ) 2 J(\epsilon)=\frac{2}{\pi D}\sqrt{1-(\epsilon/D)^2} J ( ϵ ) = π D 2 1 − ( ϵ / D ) 2 , although it must be emphasised that this choice is relatively arbitrary and not the focus of the discussion. D D D is the half-bandwidth, which we will use as our unit of energy. This definition fixes the normalisation as ∫ − D D J ( ϵ ) d ϵ = 1 \int_{-D}^{D}J(\epsilon)d\epsilon=1 ∫ − D D J ( ϵ ) d ϵ = 1 . In each example below we compare three states: the bare Gibbs state τ S \tau_S τ S , our prediction τ M F ( 2 ) \tau_{\mathrm{MF}}^{(2)} τ MF ( 2 ) , and the (numerically) exact result.
Resonant level model - We begin with the resonant level model (RLM): a single fermionic mode with H S = ε s n H_S=\varepsilon_s n H S = ε s n , n ≡ s † s n\equiv s^\dagger s n ≡ s † s , coupled to the reservoir through X = s X=s X = s . Here X X X is already an eigenoperator with ω 0 = μ − ε s \omega_0=\mu-\varepsilon_s ω 0 = μ − ε s , so Δ o f f \Delta_{\mathrm{off}} Δ off is empty and Eq. (11) gives (see Sec. SV A)
⟨ n ⟩ M F ( 2 ) = f s + λ 2 f s ( 1 − f s ) [ D 1 ( ω 0 ) − D 2 ( ω 0 ) ] , (24) \langle n\rangle_{\mathrm{MF}}^{(2)}=f_s+\lambda^2 f_s(1-f_s)\left[D_1(\omega_0)-D_2(\omega_0)\right], \tag*{(24)} ⟨ n ⟩ MF ( 2 ) = f s + λ 2 f s ( 1 − f s ) [ D 1 ( ω 0 ) − D 2 ( ω 0 ) ] , (24)
where f s ≡ f ( ε s − μ ) f_s\equiv f(\varepsilon_s-\mu) f s ≡ f ( ε s − μ ) and f f f is the Fermi Dirac distribution. This admits a clear physical interpretation and relates to the standard Green’s function treatment of the RLM [54 , 55 ] . Defining ⟨ n ⟩ M F ( 2 ) = f s + δ ⟨ n ⟩ \langle n\rangle_{\mathrm{MF}}^{(2)}=f_s+\delta\langle n\rangle ⟨ n ⟩ MF ( 2 ) = f s + δ ⟨ n ⟩ , the deviation from the bare Gibbs value can be decomposed as (see Sec. SV A)
δ ⟨ n ⟩ = ∂ f s ∂ ε s Λ ( ε s ) − λ 2 P ∫ d ε J ( ε ) f s − f ( ε − μ ) ( ε − ε s ) 2 , (25) \delta\langle n\rangle=\frac{\partial f_s}{\partial\varepsilon_s}\Lambda(\varepsilon_s)-\lambda^2\mathcal{P}\int d\varepsilon\,J(\varepsilon)\frac{f_s-f(\varepsilon-\mu)}{(\varepsilon-\varepsilon_s)^2}, \tag*{(25)} δ ⟨ n ⟩ = ∂ ε s ∂ f s Λ ( ε s ) − λ 2 P ∫ d ε J ( ε ) ( ε − ε s ) 2 f s − f ( ε − μ ) , (25)
Λ ( ε s ) = λ 2 P ∫ d ε J ( ε ) ε s − ε . (26) \Lambda(\varepsilon_s)=\lambda^2\mathcal{P}\int d\varepsilon\,\frac{J(\varepsilon)}{\varepsilon_s-\varepsilon}. \tag*{(26)} Λ ( ε s ) = λ 2 P ∫ d ε ε s − ε J ( ε ) . (26)
The first term describes a level shift by Λ \Lambda Λ due to the hybridisation with the reservoir, which the occupation responds to with the Fermi-function susceptibility. The second term represents the broadening of the level, sampling f f f away from ε s \varepsilon_s ε s .
At low temperatures the coefficient functions diverge, leading to numerical precision issues, so we instead use the bounded form derived in Sec. SIII for the points at lowest T T T , although the numerical differences between forms are negligible above T ∼ 0.05 D T\sim0.05D T ∼ 0.05 D for all models considered. Figure 1 (a) shows ⟨ n ⟩ \langle n\rangle ⟨ n ⟩ at fixed T / D = 0.1 T/D=0.1 T / D = 0.1 against λ \lambda λ . The bare Gibbs state predicts the level to be essentially empty at this temperature. As the coupling grows, hybridisation with the reservoir shifts and broadens the level, giving a finite occupation that increases with λ \lambda λ . Our perturbative MFG prediction captures this, matching the exact result at low to intermediate coupling and
Figure 1. Steady state occupation of the resonant level model H S = ε s s † s H_S=\varepsilon_s s^\dagger s H S = ε s s † s across (a) coupling at fixed temperature and (b) temperature at fixed coupling, with the bare Gibbs, MFG and exact predictions shown. The energy of the level is fixed at ε s = 0.5 D \varepsilon_s=0.5D ε s = 0.5 D with μ = 0 \mu=0 μ = 0 .
departing from it at strong coupling as expected. Figure 1 (b) shows the same quantity against temperature at fixed λ / D = 0.25 \lambda/D=0.25 λ / D = 0.25 : the occupation remains finite as T → 0 T\to0 T → 0 , and both curves converge to the Gibbs state at high temperature.
Two independent modes coupled to a shared reservoir - We now consider a model that allows for coherence in the bare system energy eigenbasis. Specifically, we consider two modes that do not interact directly but share a reservoir. An interesting feature of this model is that any coherences that arise must be mediated by the reservoir rather than directly from interactions between the two modes. This is the fermionic analogue of [26 ] in which they study reservoir mediated coherences in the mean force Gibbs state for a multi-qubit spin boson problem. The Hamiltonian is given by
H S = ε 1 n 1 + ε 2 n 2 , (27) H_S=\varepsilon_1 n_1+\varepsilon_2 n_2, \tag*{(27)} H S = ε 1 n 1 + ε 2 n 2 , (27)
with coupling operator X = s 1 + s 2 X=s_1+s_2 X = s 1 + s 2 . Each mode is an eigenoperator, X ω 1 = s 1 X_{\omega_1}=s_1 X ω 1 = s 1 , X ω 2 = s 2 X_{\omega_2}=s_2 X ω 2 = s 2 with frequencies ω 1 = μ − ε 1 \omega_1=\mu-\varepsilon_1 ω 1 = μ − ε 1 , ω 2 = μ − ε 2 \omega_2=\mu-\varepsilon_2 ω 2 = μ − ε 2 , and X = X ω 1 + X ω 2 X=X_{\omega_1}+X_{\omega_2} X = X ω 1 + X ω 2 . As in [26 ] , there exists a decoherence-free state when ε 1 = ε 2 \varepsilon_1=\varepsilon_2 ε 1 = ε 2 . The antisymmetric combination d − = ( s 1 − s 2 ) / 2 d_-=(s_1-s_2)/\sqrt{2} d − = ( s 1 − s 2 ) / 2 does not appear in X = s 1 + s 2 = 2 d + X=s_1+s_2=\sqrt{2}d_+ X = s 1 + s 2 = 2 d + , and for ε 1 = ε 2 \varepsilon_1=\varepsilon_2 ε 1 = ε 2 there is no hopping between d + d_+ d + and d − d_- d − in H S H_S H S , so d − d_- d − would be entirely decoupled from the reservoir, known as a dark state [56 , 57 ] . In this case the MFG need not be the state the system relaxes to, as this will depend on the initial occupation of the dark state. To avoid this, we consider ε 1 ≠ ε 2 \varepsilon_1\ne\varepsilon_2 ε 1 = ε 2 .
At second order in the coupling, each mode’s population is blind to the other’s presence and the occupations are given by Eq. (24) applied to each mode respectively (see Sec. SV B). The modes are not totally blind to each other however, as they would be in the bare Gibbs state, they acquire a coherence, given by
⟨ s 1 † s 2 ⟩ τ M F ( 2 ) = λ 2 Φ ( ε 1 ) − Φ ( ε 2 ) ε 1 − ε 2 , (28) \langle s_1^\dagger s_2\rangle_{\tau_{\mathrm{MF}}^{(2)}}=\lambda^2\frac{\Phi(\varepsilon_1)-\Phi(\varepsilon_2)}{\varepsilon_1-\varepsilon_2}, \tag*{(28)} ⟨ s 1 † s 2 ⟩ τ MF ( 2 ) = λ 2 ε 1 − ε 2 Φ ( ε 1 ) − Φ ( ε 2 ) , (28)
where Φ ( ε ) \Phi(\varepsilon) Φ ( ε ) is defined in Sec. SV B along with additional details on τ M F ( 2 ) \tau_{\mathrm{MF}}^{(2)} τ MF ( 2 ) . This is reflected through the Hamiltonian of mean force. Evaluating Eq. (20) gives a reservoir-mediated hopping term,
H M F = ∑ i = 1 , 2 ε ~ i n i + g ( s 1 † s 2 + s 2 † s 1 ) , (29) H_{\mathrm{MF}}=\sum_{i=1,2}\widetilde{\varepsilon}_i n_i+g\left(s_1^\dagger s_2+s_2^\dagger s_1\right), \tag*{(29)} H MF = i = 1 , 2 ∑ ε i n i + g ( s 1 † s 2 + s 2 † s 1 ) , (29)
g = λ 2 Φ ( ε 1 ) − Φ ( ε 2 ) f 1 − f 2 , (30) g=\lambda^2\frac{\Phi(\varepsilon_1)-\Phi(\varepsilon_2)}{f_1-f_2}, \tag*{(30)} g = λ 2 f 1 − f 2 Φ ( ε 1 ) − Φ ( ε 2 ) , (30)
where ε ~ i \widetilde{\varepsilon}_i ε i are defined in the same way as the renormalised energy of the Hamiltonian of mean force for the RLM, and we have omitted an overall constant in H M F H_{\mathrm{MF}} H MF (see Sec. SV B). Note that Eqs. (29) –(30) are only valid at sufficiently high temperature (see Sec. SIV), but the coherence, Eq. (28) , holds more generally. Figure 2 (a), (c) show Eq. (28) and the exact result at fixed T against coupling and at fixed coupling against T respectively. We again see good agreement at weak coupling and high temperature as expected, although the deviation occurs at lower λ \lambda λ . This can easily be understood from the definition of X = s 1 + s 2 = 2 d + X = s_1 + s_2 = \sqrt{2}d_+ X = s 1 + s 2 = 2 d + , so the reservoir couples to the bright mode with strength 2 λ \sqrt{2}\lambda 2 λ , meaning deviations set in at smaller λ \lambda λ than for the RLM. The coherence also remains finite as T → 0 T \to0 T → 0 , consistent with the persistence of coherences at low temperatures found in Sec. SIII. Figure 2 (b), (Figure 2 d) show the associated comparison but now for the full state in terms of the trace distance T ( ρ , σ ) = 1 2 Tr ( ρ − σ ) † ( ρ − σ ) T(\rho,\sigma) = \frac{1}{2}\operatorname{Tr}\sqrt{(\rho-\sigma)^\dagger(\rho-\sigma)} T ( ρ , σ ) = 2 1 Tr ( ρ − σ ) † ( ρ − σ ) , with τ M F ( 2 ) \tau_{\mathrm{MF}}^{(2)} τ MF ( 2 ) against the exact result, in addition to the difference between the bare Gibbs and the exact result. The behaviour is qualitatively the same for both the trace distance and the coherence.
Figure 2. Comparison of the steady state as predicted by Eq. (11) , the exact result and the bare Gibbs state of the two mode model H S = ε 1 n 1 + ε 2 n 2 H_S = \varepsilon_1 n_1 + \varepsilon_2 n_2 H S = ε 1 n 1 + ε 2 n 2 . (a) Coherence against coupling at fixed temperature T / D = 0.1 T/D = 0.1 T / D = 0.1 . (b) Trace distance between predictions against coupling at T / D = 0.1 T/D = 0.1 T / D = 0.1 . (c) Coherence against temperature at fixed coupling λ / D = 0.3 \lambda/D = 0.3 λ / D = 0.3 (d) Trace distance between predictions against temperature at fixed coupling λ / D = 0.3 \lambda/D = 0.3 λ / D = 0.3 . The energies of the two levels are fixed at ε 1 , ε 2 = 0.3 D , 0.4 D \varepsilon_1,\varepsilon_2 = 0.3D, 0.4D ε 1 , ε 2 = 0.3 D , 0.4 D , with μ = 0 \mu= 0 μ = 0 .
Single impurity Anderson model - As a final example, we consider the paradigmatic single impurity Anderson model (SIAM) [58 ] , with Hamiltonian
H S = ε d ( n ↑ + n ↓ ) + U n ↑ n ↓ , n σ = d σ † d σ , (31) H_S = \varepsilon_d(n_\uparrow+n_\downarrow) + U n_\uparrow n_\downarrow,\qquad n_\sigma= d_\sigma^\dagger d_\sigma, \tag*{(31)} H S = ε d ( n ↑ + n ↓ ) + U n ↑ n ↓ , n σ = d σ † d σ , (31)
coupled to a spinful reservoir,
H = H S + ∑ σ ∫ d ε ε a ε σ † a ε σ + λ ∑ σ ∫ d ε J ( ε ) ( d σ † a ε σ + a ε σ † d σ ) . (32) H = H_S + \sum_\sigma\int d\varepsilon\, \varepsilon\, a_{\varepsilon\sigma}^\dagger a_{\varepsilon\sigma}
+ \lambda\sum_\sigma\int d\varepsilon\, \sqrt{J(\varepsilon)}(d_\sigma^\dagger a_{\varepsilon\sigma}+a_{\varepsilon\sigma}^\dagger d_\sigma). \tag*{(32)} H = H S + σ ∑ ∫ d ε ε a ε σ † a ε σ + λ σ ∑ ∫ d ε J ( ε ) ( d σ † a ε σ + a ε σ † d σ ) . (32)
We take σ ∈ { ↑ , ↓ } \sigma\in\{\uparrow,\downarrow\} σ ∈ { ↑ , ↓ } to label spin, where σ ˉ \bar{\sigma} σ ˉ denotes the opposite spin. Unlike the previous two examples, this model is interacting, so the exact state is not accessible via the correlation-matrix approach. It also involves two reservoirs, one for each spin, coupled through X σ = d σ X_\sigma= d_\sigma X σ = d σ and E σ = ∫ d ε J ( ε ) a ε σ E_\sigma= \int d\varepsilon\sqrt{J(\varepsilon)}a_{\varepsilon\sigma} E σ = ∫ d ε J ( ε ) a ε σ . However, as the two spin reservoirs are independent, Eq. (11) holds with the correction summed over σ \sigma σ (see Sec. SVC). Each spin operator decomposes into two eigenoperators X σ , ω 1 = d σ ( 1 − n σ ˉ ) X_{\sigma,\omega_1} = d_\sigma(1-n_{\bar{\sigma}}) X σ , ω 1 = d σ ( 1 − n σ ˉ ) , X σ , ω 2 = d σ n σ ˉ X_{\sigma,\omega_2} = d_\sigma n_{\bar{\sigma}} X σ , ω 2 = d σ n σ ˉ so that X σ = X σ , ω 1 + X σ , ω 2 X_\sigma= X_{\sigma,\omega_1}+X_{\sigma,\omega_2} X σ = X σ , ω 1 + X σ , ω 2 , where ω 1 = μ − ε ~ d \omega_1 = \mu-\widetilde{\varepsilon}_d ω 1 = μ − ε d , ω 2 = μ − ε ~ d − U ~ \omega_2 = \mu-\widetilde{\varepsilon}_d-\widetilde{U} ω 2 = μ − ε d − U . These two Bohr frequencies correspond to adding a particle to the empty and to the singly occupied level respectively. Defining the Fock states ∣ 0 ⟩ |0\rangle ∣0 ⟩ , ∣ σ ⟩ = d σ † ∣ 0 ⟩ |\sigma\rangle= d_\sigma^\dagger|0\rangle ∣ σ ⟩ = d σ † ∣0 ⟩ , ∣ 2 ⟩ = d ↑ † d ↓ † ∣ 0 ⟩ |2\rangle= d_\uparrow^\dagger d_\downarrow^\dagger|0\rangle ∣2 ⟩ = d ↑ † d ↓ † ∣0 ⟩ , Eq. (11) gives
τ M F ( 2 ) = ∑ k p k ( 0 ) [ 1 + λ 2 ( w k − w ˉ ) ] ∣ k ⟩ ⟨ k ∣ , w ˉ ≡ ∑ k p k ( 0 ) w k , (33) \tau_{\mathrm{MF}}^{(2)} = \sum_k p_k^{(0)}\left[1+\lambda^2(w_k-\bar{w})\right]|k\rangle\langle k|,\qquad\bar{w}\equiv\sum_k p_k^{(0)}w_k, \tag*{(33)} τ MF ( 2 ) = k ∑ p k ( 0 ) [ 1 + λ 2 ( w k − w ˉ ) ] ∣ k ⟩ ⟨ k ∣ , w ˉ ≡ k ∑ p k ( 0 ) w k , (33)
for k ∈ { 0 , ↑ , ↓ , 2 } k\in\{0,\uparrow,\downarrow,2\} k ∈ { 0 , ↑ , ↓ , 2 } , where p k ( 0 ) p_k^{(0)} p k ( 0 ) are the bare Gibbs state occupations. For the definitions of w k w_k w k in terms of D 1 , D 2 D_1,D_2 D 1 , D 2 see Sec. SVC. As required by charge and S z S_z S z conservation, the state τ M F ( 2 ) \tau_{\mathrm{MF}}^{(2)} τ MF ( 2 ) is diagonal. Figure 3 (b) shows the double occupation as predicted by Eq. (33) and the converged steady state of the MPS calculation against λ \lambda λ for the particle hole symmetric case ε d = − U / 2 \varepsilon_d=-U/2 ε d = − U /2 at strong interaction U U U . We see good agreement in the expected regime of weak to intermediate coupling. The Hamiltonian of mean force can be written analogously to the bare Hamiltonian H S H_S H S with effective energies and interaction ε ~ d , U ~ \widetilde{\varepsilon}_d,\widetilde{U} ε d , U (see Sec. SVC). Figure 3 (a) shows the time evolution of the four diagonal components of the impurity. Together, these results demonstrate the accuracy and validity of our result in the expected parameter regime.
Figure 3. Comparison between the MFG prediction Eq. (11) , the bare Gibbs state and the numerically exact value from MPS real time evolution. (a) Time evolution of the four diagonal components of the impurity state with initial state ρ 0 = ∣ 0 ⟩ ⟨ 0 ∣ \rho_0 = |0\rangle\langle0| ρ 0 = ∣0 ⟩ ⟨ 0∣ , at coupling λ = 0.15 D \lambda= 0.15D λ = 0.15 D . (b) Steady state double occupancy of the MFG prediction compared to the converged values from the MPS calculation and the bare Gibbs state. Parameters: β = 5 / D \beta= 5/D β = 5/ D , ε d = − 0.4 D \varepsilon_d = -0.4D ε d = − 0.4 D , U = 0.8 D U = 0.8D U = 0.8 D , μ = 0 \mu= 0 μ = 0 .
Conclusion
We have derived the grand-canonical MFG state to second order in the coupling for an arbitrary fermionic system exchanging particles with a fermionic reservoir, together with the corresponding Hamiltonian of mean force, for which the MFG state takes a Gibbs-like form. The correction separates into particle and hole channels weighted by the empty and filled reservoir states, and the KMS condition fixes all coefficients in terms of a single reservoir function. As in the bosonic case [23 ] , the MFG state generically carries coherences in the bare energy eigenbasis, where we found for two independent modes these are mediated entirely through the reservoir and persist at T = 0 T = 0 T = 0 , giving the fermionic analogue to [26 ] . We compared with exact results for the quadratic models and with thermofield MPS dynamics for the single impurity Anderson model, finding good agreement at weak to intermediate coupling as expected. Our results add to the literature on the MFG state which has thus far been mostly restricted to the bosonic regime. This provides a reference state needed to define strong-coupling thermodynamic quantities for fermionic systems such as quantum dots [8 , 9 ] , molecular junctions [10 , 11 ] and thermal machines [59 ] . A natural extension would be to explore the possibility of tractable expressions for the fermionic MFG beyond the perturbative coupling regime, although this is a challenging problem, especially as there exists no fermionic analogue of the polaron transform which is the usual route taken for bosonic systems [13 ] . One possible approach would be to use the reaction-coordinate mapping [60 ] to treat the strong system-reservoir coupling exactly and use Eq. (11) on the residual reservoir coupling, although we leave this for future work.
Acknowledgments
D.J.S acknowledges support from EPSRC Grant EP/Z534936/1.
Supplementary Material for: Second-order mean force Gibbs state of fermionic open systems
David J. Strachan
Derivation of the main result
Here we give a detailed derivation of Eq. (11) at arbitrary chemical potential μ \mu μ . We consider a system and environment with total Hamiltonian
H = H S + H R + λ V , (34) H = H_S + H_R + \lambda V, \tag*{(34)} H = H S + H R + λV , (34)
where H S H_S H S and H R H_R H R are the system and reservoir Hamiltonians and λ V \lambda V λV their interaction, with λ \lambda λ the coupling parameter. In the bosonic literature V V V is usually taken to be of the factorised form V = X ⊗ B V = X \otimes B V = X ⊗ B with X X X and B B B Hermitian system and reservoir operators [13 , 23 ] . Here we instead take
V = X † E + E † X , (35) V = X^\dagger E + E^\dagger X, \tag*{(35)} V = X † E + E † X , (35)
which describes an exchange, or hopping, interaction, with X X X and E E E being odd-parity system and reservoir operators respectively. Importantly, this does not assume a tensor product structure as fermions do not possess one. This is the generic coupling for fermionic reservoirs, and is also a legitimate choice for bosonic ones i.e. the Jaynes–Cummings interaction. Specifically we consider the canonical setup as outlined in section I in which we have
E = ∫ d ε J ( ε ) a ε . (36) E = \int\mathrm{d}\varepsilon\,\sqrt{J(\varepsilon)}\,a_\varepsilon. \tag*{(36)} E = ∫ d ε J ( ε ) a ε . (36)
We assume that the coupling conserves the total particle number N = N S + N R N = N_S + N_R N = N S + N R , with N R = ∫ d ε a ε † a ε N_R = \int\mathrm{d}\varepsilon\,a^\dagger_\varepsilon a_\varepsilon N R = ∫ d ε a ε † a ε , and that [ H S , N S ] = 0 [H_S,N_S] = 0 [ H S , N S ] = 0 . Since [ N R , E ] = − E [N_R,E] = -E [ N R , E ] = − E , number conservation of Eq. (35) requires
[ N S , X ] = − X , (37) [N_S,X] = -X, \tag*{(37)} [ N S , X ] = − X , (37)
i.e. X X X removes exactly one particle from the system, so that [ V , N ] = 0 [V,N] = 0 [ V , N ] = 0 and [ H , N ] = 0 [H,N] = 0 [ H , N ] = 0 . The global state is then the grand canonical Gibbs state τ t o t = e − β ( H − μ N ) / Tr ( e − β ( H − μ N ) ) \tau_{\mathrm{tot}} = e^{-\beta(H-\mu N)}/\operatorname{Tr}(e^{-\beta(H-\mu N)}) τ tot = e − β ( H − μ N ) / Tr ( e − β ( H − μ N ) ) , and its generator can be written as
H − μ N = K S + K R + λ V ≡ K 0 + λ V , K S ≡ H S − μ N S , K R ≡ ∫ d ε ξ a ε † a ε , ξ ≡ ε − μ . (38) H - \mu N = K_S + K_R + \lambda V \equiv K_0 + \lambda V,\qquad K_S \equiv H_S - \mu N_S,\qquad K_R \equiv\int\mathrm{d}\varepsilon\,\xi\,a^\dagger_\varepsilon a_\varepsilon,\qquad\xi\equiv\varepsilon- \mu. \tag*{(38)} H − μ N = K S + K R + λV ≡ K 0 + λV , K S ≡ H S − μ N S , K R ≡ ∫ d ε ξ a ε † a ε , ξ ≡ ε − μ . (38)
The system Gibbs state is τ S = e − β K S / Z S \tau_S = e^{-\beta K_S}/Z_S τ S = e − β K S / Z S with Z S = Tr S ( e − β K S ) Z_S = \operatorname{Tr}_S(e^{-\beta K_S}) Z S = Tr S ( e − β K S ) , and the reservoir Gibbs state is τ R = e − β K R / Z R \tau_R = e^{-\beta K_R}/Z_R τ R = e − β K R / Z R .
Fermionic partial trace and parity
For fermions, the partial trace is somewhat ambiguous due to the lack of a tensor product structure of the fermionic Fock space. Fermionic Fock spaces are defined by the algebra of canonical anticommutation relations (CAR) A \mathcal{A} A , defined by
{ b i † , b j } = δ i j , (39) \{b_i^\dagger,b_j\} = \delta_{ij}, \tag*{(39)} { b i † , b j } = δ ij , (39)
{ b i , b j } = { b i † , b j † } = 0 , (40) \{b_i,b_j\} = \{b_i^\dagger,b_j^\dagger\} = 0, \tag*{(40)} { b i , b j } = { b i † , b j † } = 0 , (40)
where b i † b_i^\dagger b i † (b i b_i b i ) are creation (annihilation) operators. Fock states are then defined through a specific ordering of mode operators,
∣ j 1 , … , j n ⟩ ≡ ( b 1 † ) j 1 ( b 2 † ) j 2 ⋯ ( b n † ) j n ∣ v a c ⟩ . (41) |j_1,\ldots,j_n\rangle\equiv(b_1^\dagger)^{j_1}(b_2^\dagger)^{j_2}\cdots(b_n^\dagger)^{j_n}|\mathrm{vac}\rangle. \tag*{(41)} ∣ j 1 , … , j n ⟩ ≡ ( b 1 † ) j 1 ( b 2 † ) j 2 ⋯ ( b n † ) j n ∣ vac ⟩ . (41)
For bosons this ordering is arbitrary as creation operators of different modes commute, but for fermions this is not the case. It is clear from this that ∣ j 1 , … , j n ⟩ |j_1,\ldots,j_n\rangle ∣ j 1 , … , j n ⟩ cannot be written as a tensor product ∣ j 1 ⟩ ⊗ ⋯ ⊗ ∣ j n ⟩ |j_1\rangle\otimes\cdots\otimes|j_n\rangle ∣ j 1 ⟩ ⊗ ⋯ ⊗ ∣ j n ⟩ due to this anticommutativity, and as such the standard partial trace definition in terms of such a tensor product does not directly apply. We can instead define a fermionic partial trace, or mode reduced state [34 , 35 ] ρ S = Tr R ( ρ ) \rho_S = \operatorname{Tr}_R(\rho) ρ S = Tr R ( ρ ) defined as the unique operator on F S \mathcal{F}_S F S with
Tr S ( ρ S O S ) = Tr ( ρ O S ) ∀ O S ∈ A S , (42) \operatorname{Tr}_S(\rho_S O_S) = \operatorname{Tr}(\rho O_S) \qquad\forall O_S \in\mathcal{A}_S, \tag*{(42)} Tr S ( ρ S O S ) = Tr ( ρ O S ) ∀ O S ∈ A S , (42)
where A S \mathcal{A}_S A S is the canonical anti-commutation relation (CAR) algebra of the system. Let P = ( − 1 ) N P = (-1)^N P = ( − 1 ) N denote the total fermion parity of the bipartite space, with N N N the total number operator. In [34 ] , it is shown that if the modes being traced out are a contiguous set in the mode ordering and the global state ρ \rho ρ is parity symmetric, i.e. [ P , ρ ] = 0 [P,\rho] = 0 [ P , ρ ] = 0 , then the mode reduced state ρ S \rho_S ρ S coincides with the operator obtained from conventional partial trace methods under a Jordan Wigner transform.
We will use this when tracing out the reservoir in the definition of τ M F \tau_{\mathrm{MF}} τ MF . We thus fix the mode ordering and therefore the definitions of the mode algebras A S \mathcal{A}_S A S and A R \mathcal{A}_R A R to be system modes first,
∣ n S , m R ⟩ = ( s 1 † ) n 1 ⋯ ( s L S † ) n L S ( a 1 † ) m 1 ⋯ ( a L R † ) m L R ∣ v a c ⟩ , (43) \lvert n_S,m_R\rangle= (s_1^\dagger)^{n_1}\cdots(s_{L_S}^\dagger)^{n_{L_S}}(a_1^\dagger)^{m_1}\cdots(a_{L_R}^\dagger)^{m_{L_R}}\lvert\mathrm{vac}\rangle, \tag*{(43)} ∣ n S , m R ⟩ = ( s 1 † ) n 1 ⋯ ( s L S † ) n L S ( a 1 † ) m 1 ⋯ ( a L R † ) m L R ∣ vac ⟩ , (43)
where L S , L R L_S,L_R L S , L R are the number of system and reservoir modes respectively. We will also need to use some properties related to parity in the derivation, which we lay out here. Write P = P S P R P = P_S P_R P = P S P R with P S , P R P_S,P_R P S , P R the system and reservoir parities. An operator O O O is even if P O P † = O P O P^\dagger= O P O P † = O and odd if P O P † = − O P O P^\dagger= -O P O P † = − O ; equivalently, even operators contain an even number of mode operators. In Sec. SI A, we use the following property. Let O S ∈ A S O_S \in\mathcal{A}_S O S ∈ A S and O R ∈ A R O_R \in\mathcal{A}_R O R ∈ A R . If either O S O_S O S or O R O_R O R is even, then [ O S , O R ] = 0 [O_S,O_R] = 0 [ O S , O R ] = 0 . If both are odd, then { O S , O R } = 0 \{O_S,O_R\} = 0 { O S , O R } = 0 . This is immediate from the canonical anticommutation relations: moving an odd operator past an odd operator on disjoint modes generates a single minus sign, whereas moving an even one past anything generates none. Fermionic systems follow superselection rules [34 , 42 , 43 ] , so any physical Hamiltonian is even; [ H S , O R ] = [ H R , O S ] = 0 [H_S,O_R] = [H_R,O_S] = 0 [ H S , O R ] = [ H R , O S ] = 0 for arbitrary O S , O R O_S,O_R O S , O R , and the same holds for K S K_S K S and K R K_R K R since number operators are even. This is what allows the β \beta β -interaction picture below to factorise. If X X X is odd, X † X X^\dagger X X † X , X X † X X^\dagger X X † , X 2 † X 1 X_2^\dagger X_1 X 2 † X 1 and X 2 X 1 † X_2 X_1^\dagger X 2 X 1 † are all even and commute with reservoir operators; this is what makes the two fermionic swaps in Eqs. (49) –(50) cancel.
Kubo expansion and the reservoir trace
We start from the Kubo expansion [23 , 44 , 45 , 46 ]
e − β ( K 0 + λ V ) = e − β K 0 ( 1 − λ ∫ 0 β d β 1 e β 1 K 0 V e − β 1 K 0 + λ 2 ∫ 0 β d β 2 ∫ 0 β 2 d β 1 e β 2 K 0 V e − ( β 2 − β 1 ) K 0 V e − β 1 K 0 + O ( λ 3 ) ) . (44) e^{-\beta(K_0+\lambda V)} = e^{-\beta K_0}\left(1-\lambda\int_0^\beta d\beta_1\,e^{\beta_1 K_0}V e^{-\beta_1 K_0}+\lambda^2\int_0^\beta d\beta_2\int_0^{\beta_2}d\beta_1\,e^{\beta_2 K_0}V e^{-(\beta_2-\beta_1)K_0}V e^{-\beta_1 K_0}+\mathcal{O}(\lambda^3)\right). \tag*{(44)} e − β ( K 0 + λV ) = e − β K 0 ( 1 − λ ∫ 0 β d β 1 e β 1 K 0 V e − β 1 K 0 + λ 2 ∫ 0 β d β 2 ∫ 0 β 2 d β 1 e β 2 K 0 V e − ( β 2 − β 1 ) K 0 V e − β 1 K 0 + O ( λ 3 ) ) . (44)
We define the “β \beta β -interaction” picture V ( − i β ′ ) = e β ′ K 0 V e − β ′ K 0 V(-i\beta') = e^{\beta'K_0}V e^{-\beta'K_0} V ( − i β ′ ) = e β ′ K 0 V e − β ′ K 0 . As K S K_S K S and K R K_R K R commute with reservoir and system operators respectively (Sec. SI A), we have
V ( − i β ′ ) = X † ( − i β ′ ) E ( − i β ′ ) + E † ( − i β ′ ) X ( − i β ′ ) , (45) V(-i\beta') = X^\dagger(-i\beta')E(-i\beta') + E^\dagger(-i\beta')X(-i\beta'), \tag*{(45)} V ( − i β ′ ) = X † ( − i β ′ ) E ( − i β ′ ) + E † ( − i β ′ ) X ( − i β ′ ) , (45)
where
X ( − i β ′ ) = e β ′ K S X e − β ′ K S , E ( − i β ′ ) = e β ′ K R E e − β ′ K R . (46) X(-i\beta') = e^{\beta'K_S}X e^{-\beta'K_S}, \qquad E(-i\beta') = e^{\beta'K_R}E e^{-\beta'K_R}. \tag*{(46)} X ( − i β ′ ) = e β ′ K S X e − β ′ K S , E ( − i β ′ ) = e β ′ K R E e − β ′ K R . (46)
With this, Eq. (44) becomes
e − β ( K 0 + λ V ) = e − β K 0 ( 1 − λ ∫ 0 β d β 1 V ( − i β 1 ) + λ 2 ∫ 0 β d β 2 ∫ 0 β 2 d β 1 V ( − i β 2 ) V ( − i β 1 ) + O ( λ 3 ) ) . (47) e^{-\beta(K_0+\lambda V)} = e^{-\beta K_0}\left(1-\lambda\int_0^\beta d\beta_1 V(-i\beta_1)+\lambda^2\int_0^\beta d\beta_2\int_0^{\beta_2}d\beta_1 V(-i\beta_2)V(-i\beta_1)+\mathcal{O}(\lambda^3)\right). \tag*{(47)} e − β ( K 0 + λV ) = e − β K 0 ( 1 − λ ∫ 0 β d β 1 V ( − i β 1 ) + λ 2 ∫ 0 β d β 2 ∫ 0 β 2 d β 1 V ( − i β 2 ) V ( − i β 1 ) + O ( λ 3 ) ) . (47)
Introducing the shorthand X j = X ( − i β j ) X_j=X(-i\beta_j) X j = X ( − i β j ) , E j = E ( − i β j ) E_j=E(-i\beta_j) E j = E ( − i β j ) , the O ( λ 2 ) \mathcal{O}(\lambda^2) O ( λ 2 ) term contains four contributions,
V ( − i β 2 ) V ( − i β 1 ) = X 2 † E 2 X 1 † E 1 + X 2 † E 2 E 1 † X 1 + E 2 † X 2 X 1 † E 1 + E 2 † X 2 E 1 † X 1 . (48) V(-i\beta_2)V(-i\beta_1)=X_2^\dagger E_2X_1^\dagger E_1+X_2^\dagger E_2E_1^\dagger X_1+E_2^\dagger X_2X_1^\dagger E_1+E_2^\dagger X_2E_1^\dagger X_1. \tag*{(48)} V ( − i β 2 ) V ( − i β 1 ) = X 2 † E 2 X 1 † E 1 + X 2 † E 2 E 1 † X 1 + E 2 † X 2 X 1 † E 1 + E 2 † X 2 E 1 † X 1 . (48)
The reservoir Gibbs state is block diagonal in excitation number, so the first and last terms, which contain E 2 E 1 E_2E_1 E 2 E 1 and E 2 † E 1 † E_2^\dagger E_1^\dagger E 2 † E 1 † respectively and therefore connect different number sectors, vanish on tracing over the reservoir. The remaining two terms factorise into system and reservoir parts. X 2 † X 1 X_2^\dagger X_1 X 2 † X 1 and X 2 X 1 † X_2 X_1^\dagger X 2 X 1 † are even-parity system operators, so no net sign is generated when the two system operators are brought together, the two fermionic swaps required in each case cancel,
X 2 † E 2 E 1 † X 1 = X 2 † X 1 E 2 E 1 † , (49) X_2^\dagger E_2 E_1^\dagger X_1 = X_2^\dagger X_1 E_2 E_1^\dagger, \tag*{(49)} X 2 † E 2 E 1 † X 1 = X 2 † X 1 E 2 E 1 † , (49)
E 2 † X 2 X 1 † E 1 = X 2 X 1 † E 2 † E 1 . (50) E_2^\dagger X_2 X_1^\dagger E_1 = X_2 X_1^\dagger E_2^\dagger E_1. \tag*{(50)} E 2 † X 2 X 1 † E 1 = X 2 X 1 † E 2 † E 1 . (50)
We can now trace over the reservoir. As the reservoir modes are contiguous and all operators involved have even parity, the fermionic partial trace as defined by Eq. (42) coincides with the usual partial trace here (Sec. SI A), leading to the following expression for the unnormalised mean force Gibbs state
ρ ~ S = e − β K S ( 1 + λ 2 ∑ ω , ω ′ X ω † X ω ′ ∫ 0 β d β 2 ∫ 0 β 2 d β 1 G 1 ( β , − i ( β 2 − β 1 ) ) + λ 2 ∑ ω , ω ′ X ω X ω ′ † ∫ 0 β d β 2 ∫ 0 β 2 d β 1 G 2 ( β , − i ( β 2 − β 1 ) ) + O ( λ 4 ) ) , (51) \widetilde{\rho}_S = e^{-\beta K_S}\left(1+\lambda^2\sum_{\omega,\omega'}X_\omega^\dagger X_{\omega'}\int_0^\beta d\beta_2\int_0^{\beta_2}d\beta_1\,G_1(\beta,-i(\beta_2-\beta_1))+\lambda^2\sum_{\omega,\omega'}X_\omega X_{\omega'}^\dagger\int_0^\beta d\beta_2\int_0^{\beta_2}d\beta_1\,G_2(\beta,-i(\beta_2-\beta_1))+\mathcal{O}(\lambda^4)\right), \tag*{(51)} ρ S = e − β K S 1 + λ 2 ω , ω ′ ∑ X ω † X ω ′ ∫ 0 β d β 2 ∫ 0 β 2 d β 1 G 1 ( β , − i ( β 2 − β 1 )) + λ 2 ω , ω ′ ∑ X ω X ω ′ † ∫ 0 β d β 2 ∫ 0 β 2 d β 1 G 2 ( β , − i ( β 2 − β 1 )) + O ( λ 4 ) , (51)
where
G 1 ( β , − i ( β 2 − β 1 ) ) = Tr R [ τ R E ( − i β 2 ) E † ( − i β 1 ) ] = Tr R [ τ R E ( − i ( β 2 − β 1 ) ) E † ] , (52) G_1(\beta,-i(\beta_2-\beta_1))=\operatorname{Tr}_R\left[\tau_R E(-i\beta_2)E^\dagger(-i\beta_1)\right]=\operatorname{Tr}_R\left[\tau_R E(-i(\beta_2-\beta_1))E^\dagger\right], \tag*{(52)} G 1 ( β , − i ( β 2 − β 1 )) = Tr R [ τ R E ( − i β 2 ) E † ( − i β 1 ) ] = Tr R [ τ R E ( − i ( β 2 − β 1 )) E † ] , (52)
G 2 ( β , − i ( β 2 − β 1 ) ) = Tr R [ τ R E † ( − i β 2 ) E ( − i β 1 ) ] = Tr R [ τ R E † ( − i ( β 2 − β 1 ) ) E ] . (53) G_2(\beta,-i(\beta_2-\beta_1))=\operatorname{Tr}_R\left[\tau_R E^\dagger(-i\beta_2)E(-i\beta_1)\right]=\operatorname{Tr}_R\left[\tau_R E^\dagger(-i(\beta_2-\beta_1))E\right]. \tag*{(53)} G 2 ( β , − i ( β 2 − β 1 )) = Tr R [ τ R E † ( − i β 2 ) E ( − i β 1 ) ] = Tr R [ τ R E † ( − i ( β 2 − β 1 )) E ] . (53)
Dependence on β 2 − β 1 \beta_2-\beta_1 β 2 − β 1 alone follows from imaginary-time translation invariance of the equilibrium reservoir state. The O ( λ ) \mathcal{O}(\lambda) O ( λ ) and O ( λ 3 ) \mathcal{O}(\lambda^3) O ( λ 3 ) contributions vanish because they contain odd-parity environment operators, whose reservoir trace against τ R \tau_R τ R is zero. Equations (S19)–(S20) are the imaginary-time analogues of the reservoir temporal correlation functions familiar from open-system dynamics [13 ] .
Eigenoperator decomposition
To arrive at a practical expression we decompose X X X into eigenoperators of K S K_S K S ,
X = ∑ ω X ω , X ω = ∑ n , m : ε n − ε m = ω Π n X Π m , [ K S , X ω ] = ω X ω , (54) X=\sum_\omega X_\omega,\qquad X_\omega=\sum_{n,m:\,\varepsilon_n-\varepsilon_m=\omega}\Pi_nX\Pi_m,\qquad[K_S,X_\omega]=\omega X_\omega, \tag*{(54)} X = ω ∑ X ω , X ω = n , m : ε n − ε m = ω ∑ Π n X Π m , [ K S , X ω ] = ω X ω , (54)
where Π n \Pi_n Π n projects onto the (possibly degenerate) ε n \varepsilon_n ε n eigenspace of K S K_S K S . The sum over ω \omega ω therefore runs over distinct values of ω \omega ω rather than over distinct transitions as several transitions may share the same Bohr frequency, so are collected into a single X ω X_\omega X ω . This grouping is essential, because separating such transitions would generate spurious ( ω ′ − ω ) − 1 (\omega'-\omega)^{-1} ( ω ′ − ω ) − 1 terms with vanishing denominators in Eq. (61) . It follows that
X ( − i β ′ ) = e β ′ K S X e − β ′ K S = ∑ ω e β ′ ω X ω , (55) X(-i\beta')=e^{\beta'K_S}Xe^{-\beta'K_S}=\sum_\omega e^{\beta'\omega}X_\omega, \tag*{(55)} X ( − i β ′ ) = e β ′ K S X e − β ′ K S = ω ∑ e β ′ ω X ω , (55)
and, taking the adjoint of [ K S , X ω ] = ω X ω [K_S,X_\omega]=\omega X_\omega [ K S , X ω ] = ω X ω to obtain [ K S , X ω † ] = − ω X ω † [K_S,X_\omega^\dagger]=-\omega X_\omega^\dagger [ K S , X ω † ] = − ω X ω † ,
X † ( − i β ′ ) = e β ′ K S X † e − β ′ K S = ∑ ω e − β ′ ω X ω † . (56) X^\dagger(-i\beta')=e^{\beta'K_S}X^\dagger e^{-\beta'K_S}=\sum_\omega e^{-\beta'\omega}X_\omega^\dagger. \tag*{(56)} X † ( − i β ′ ) = e β ′ K S X † e − β ′ K S = ω ∑ e − β ′ ω X ω † . (56)
Substituting into Eq. (51) ,
ρ ~ S = e − β K S ( 1 + λ 2 ∑ ω , ω ′ X ω † X ω ′ ∫ 0 β d β 2 ∫ 0 β 2 d β 1 e − β 2 ω e β 1 ω ′ G 1 ( β , − i ( β 2 − β 1 ) ) + λ 2 ∑ ω , ω ′ X ω X ω ′ † ∫ 0 β d β 2 ∫ 0 β 2 d β 1 e β 2 ω e − β 1 ω ′ G 2 ( β , − i ( β 2 − β 1 ) ) + O ( λ 4 ) ) . (57) \widetilde{\rho}_S=e^{-\beta K_S}\left(1+\lambda^2\sum_{\omega,\omega'}X_\omega^\dagger X_{\omega'}\int_0^\beta d\beta_2\int_0^{\beta_2}d\beta_1\,e^{-\beta_2\omega}e^{\beta_1\omega'}G_1(\beta,-i(\beta_2-\beta_1))+\lambda^2\sum_{\omega,\omega'}X_\omega X_{\omega'}^\dagger\int_0^\beta d\beta_2\int_0^{\beta_2}d\beta_1\,e^{\beta_2\omega}e^{-\beta_1\omega'}G_2(\beta,-i(\beta_2-\beta_1))+\mathcal{O}(\lambda^4)\right). \tag*{(57)} ρ S = e − β K S 1 + λ 2 ω , ω ′ ∑ X ω † X ω ′ ∫ 0 β d β 2 ∫ 0 β 2 d β 1 e − β 2 ω e β 1 ω ′ G 1 ( β , − i ( β 2 − β 1 )) + λ 2 ω , ω ′ ∑ X ω X ω ′ † ∫ 0 β d β 2 ∫ 0 β 2 d β 1 e β 2 ω e − β 1 ω ′ G 2 ( β , − i ( β 2 − β 1 )) + O ( λ 4 ) . (57)
Changing variables and swapping the order of integration gives
ρ ~ S = e − β K S ( 1 + λ 2 ∑ ω , ω ′ X ω † X ω ′ ∫ 0 β d β 1 ∫ β 1 β d β 2 e − β 2 ( ω − ω ′ ) e − β 1 ω ′ G 1 ( β , − i β 1 ) + λ 2 ∑ ω , ω ′ X ω X ω ′ † ∫ 0 β d β 1 ∫ β 1 β d β 2 e β 2 ( ω − ω ′ ) e β 1 ω ′ G 2 ( β , − i β 1 ) + O ( λ 4 ) ) . (58) \widetilde{\rho}_S=e^{-\beta K_S}\left(1+\lambda^2\sum_{\omega,\omega'}X_\omega^\dagger X_{\omega'}\int_0^\beta d\beta_1\int_{\beta_1}^{\beta}d\beta_2\,e^{-\beta_2(\omega-\omega')}e^{-\beta_1\omega'}G_1(\beta,-i\beta_1)+\lambda^2\sum_{\omega,\omega'}X_\omega X_{\omega'}^\dagger\int_0^\beta d\beta_1\int_{\beta_1}^{\beta}d\beta_2\,e^{\beta_2(\omega-\omega')}e^{\beta_1\omega'}G_2(\beta,-i\beta_1)+\mathcal{O}(\lambda^4)\right). \tag*{(58)} ρ S = e − β K S 1 + λ 2 ω , ω ′ ∑ X ω † X ω ′ ∫ 0 β d β 1 ∫ β 1 β d β 2 e − β 2 ( ω − ω ′ ) e − β 1 ω ′ G 1 ( β , − i β 1 ) + λ 2 ω , ω ′ ∑ X ω X ω ′ † ∫ 0 β d β 1 ∫ β 1 β d β 2 e β 2 ( ω − ω ′ ) e β 1 ω ′ G 2 ( β , − i β 1 ) + O ( λ 4 ) . (58)
The β 2 \beta_2 β 2 integral can now be carried out,
ρ ~ S = e − β K S ( 1 + λ 2 ∑ ω , ω ′ [ X ω † X ω ′ ω ′ − ω ( e − β ( ω − ω ′ ) ∫ 0 β d β 1 e − β 1 ω ′ G 1 ( β , − i β 1 ) − ∫ 0 β d β 1 e − β 1 ω G 1 ( β , − i β 1 ) ) + X ω X ω ′ † ω − ω ′ ( e β ( ω − ω ′ ) ∫ 0 β d β 1 e β 1 ω ′ G 2 ( β , − i β 1 ) − ∫ 0 β d β 1 e β 1 ω G 2 ( β , − i β 1 ) ) ] ) . (59) \tilde{\rho}_S=e^{-\beta K_S}\left(1+\lambda^2\sum_{\omega,\omega'}\left[\frac{X_\omega^\dagger X_{\omega'}}{\omega'-\omega}\left(e^{-\beta(\omega-\omega')}\int_0^\beta d\beta_1 e^{-\beta_1\omega'}G_1(\beta,-i\beta_1)-\int_0^\beta d\beta_1 e^{-\beta_1\omega}G_1(\beta,-i\beta_1)\right)+\frac{X_\omega X_{\omega'}^\dagger}{\omega-\omega'}\left(e^{\beta(\omega-\omega')}\int_0^\beta d\beta_1 e^{\beta_1\omega'}G_2(\beta,-i\beta_1)-\int_0^\beta d\beta_1 e^{\beta_1\omega}G_2(\beta,-i\beta_1)\right)\right]\right). \tag*{(59)} ρ ~ S = e − β K S 1 + λ 2 ω , ω ′ ∑ [ ω ′ − ω X ω † X ω ′ ( e − β ( ω − ω ′ ) ∫ 0 β d β 1 e − β 1 ω ′ G 1 ( β , − i β 1 ) − ∫ 0 β d β 1 e − β 1 ω G 1 ( β , − i β 1 ) ) + ω − ω ′ X ω X ω ′ † ( e β ( ω − ω ′ ) ∫ 0 β d β 1 e β 1 ω ′ G 2 ( β , − i β 1 ) − ∫ 0 β d β 1 e β 1 ω G 2 ( β , − i β 1 ) ) ] . (59)
Defining
A 1 ( ω ) = ∫ 0 β d τ e − τ ω G 1 ( β , − i τ ) , A 2 ( ω ) = ∫ 0 β d τ e τ ω G 2 ( β , − i τ ) , (60) A_1(\omega)=\int_0^\beta d\tau\,e^{-\tau\omega}G_1(\beta,-i\tau),\qquad A_2(\omega)=\int_0^\beta d\tau\,e^{\tau\omega}G_2(\beta,-i\tau), \tag*{(60)} A 1 ( ω ) = ∫ 0 β d τ e − τ ω G 1 ( β , − i τ ) , A 2 ( ω ) = ∫ 0 β d τ e τ ω G 2 ( β , − i τ ) , (60)
this can be written compactly as
ρ ~ S = e − β K S ( 1 + λ 2 ∑ ω , ω ′ [ X ω † X ω ′ e − β ( ω − ω ′ ) A 1 ( ω ′ ) − A 1 ( ω ) ω ′ − ω + X ω X ω ′ † e β ( ω − ω ′ ) A 2 ( ω ′ ) − A 2 ( ω ) ω − ω ′ ] ) . (61) \tilde{\rho}_S=e^{-\beta K_S}\left(1+\lambda^2\sum_{\omega,\omega'}\left[X_\omega^\dagger X_{\omega'}\frac{e^{-\beta(\omega-\omega')}A_1(\omega')-A_1(\omega)}{\omega'-\omega}+X_\omega X_{\omega'}^\dagger\frac{e^{\beta(\omega-\omega')}A_2(\omega')-A_2(\omega)}{\omega-\omega'}\right]\right). \tag*{(61)} ρ ~ S = e − β K S 1 + λ 2 ω , ω ′ ∑ [ X ω † X ω ′ ω ′ − ω e − β ( ω − ω ′ ) A 1 ( ω ′ ) − A 1 ( ω ) + X ω X ω ′ † ω − ω ′ e β ( ω − ω ′ ) A 2 ( ω ′ ) − A 2 ( ω ) ] . (61)
We now treat the ω = ω ′ \omega=\omega' ω = ω ′ and ω ≠ ω ′ \omega\ne\omega' ω = ω ′ cases separately. For ω = ω ′ \omega=\omega' ω = ω ′ , writing δ = ω ′ − ω \delta=\omega'-\omega δ = ω ′ − ω and expanding,
lim δ → 0 e β δ A 1 ( ω + δ ) − A 1 ( ω ) δ = β A 1 ( ω ) + d A 1 ( ω ) d ω , (62) \lim_{\delta\to0}\frac{e^{\beta\delta}A_1(\omega+\delta)-A_1(\omega)}{\delta}=\beta A_1(\omega)+\frac{dA_1(\omega)}{d\omega}, \tag*{(62)} δ → 0 lim δ e β δ A 1 ( ω + δ ) − A 1 ( ω ) = β A 1 ( ω ) + d ω d A 1 ( ω ) , (62)
and
− lim δ → 0 e − β δ A 2 ( ω + δ ) − A 2 ( ω ) δ = β A 2 ( ω ) − d A 2 ( ω ) d ω . (63) -\lim_{\delta\to0}\frac{e^{-\beta\delta}A_2(\omega+\delta)-A_2(\omega)}{\delta}=\beta A_2(\omega)-\frac{dA_2(\omega)}{d\omega}. \tag*{(63)} − δ → 0 lim δ e − β δ A 2 ( ω + δ ) − A 2 ( ω ) = β A 2 ( ω ) − d ω d A 2 ( ω ) . (63)
For the ω ≠ ω ′ \omega\ne\omega' ω = ω ′ terms we remove the exponential factors using
e − β K S X ω = X ω e − β K S e − β ω , e − β K S X ω † = X ω † e − β K S e β ω , (64) e^{-\beta K_S}X_\omega=X_\omega e^{-\beta K_S}e^{-\beta\omega},\qquad e^{-\beta K_S}X_\omega^\dagger=X_\omega^\dagger e^{-\beta K_S}e^{\beta\omega}, \tag*{(64)} e − β K S X ω = X ω e − β K S e − β ω , e − β K S X ω † = X ω † e − β K S e β ω , (64)
which follow from [ K S , X ω ] = ω X ω [K_S,X_\omega]=\omega X_\omega [ K S , X ω ] = ω X ω , giving
e − β K S X ω X ω ′ † = X ω X ω ′ † e − β K S e − β ( ω − ω ′ ) , e − β K S X ω † X ω ′ = X ω † X ω ′ e − β K S e β ( ω − ω ′ ) . (65) e^{-\beta K_S}X_\omega X_{\omega'}^\dagger=X_\omega X_{\omega'}^\dagger e^{-\beta K_S}e^{-\beta(\omega-\omega')},\qquad e^{-\beta K_S}X_\omega^\dagger X_{\omega'}=X_\omega^\dagger X_{\omega'}e^{-\beta K_S}e^{\beta(\omega-\omega')}. \tag*{(65)} e − β K S X ω X ω ′ † = X ω X ω ′ † e − β K S e − β ( ω − ω ′ ) , e − β K S X ω † X ω ′ = X ω † X ω ′ e − β K S e β ( ω − ω ′ ) . (65)
Using Eqs. (S29)–(S32), relabelling ω ↔ ω ′ \omega\leftrightarrow\omega' ω ↔ ω ′ in some terms and dividing by Z S ≡ Tr ( e − β K S ) Z_S\equiv\operatorname{Tr}(e^{-\beta K_S}) Z S ≡ Tr ( e − β K S ) , the unnormalised state becomes
ρ ~ S = τ S + λ 2 Δ ~ + O ( λ 4 ) , \tilde{\rho}_S=\tau_S+\lambda^2\tilde{\Delta}+O(\lambda^4), ρ ~ S = τ S + λ 2 Δ ~ + O ( λ 4 ) ,
Δ ~ = ∑ ω [ τ S X ω † X ω ( β A 1 ( ω ) + d A 1 ( ω ) d ω ) + τ S X ω X ω † ( β A 2 ( ω ) − d A 2 ( ω ) d ω ) ] + ∑ ω ≠ ω ′ ( ω ′ − ω ) − 1 [ ( X ω ′ X ω † τ S + τ S X ω X ω ′ † ) A 2 ( ω ) − ( X ω ′ † X ω τ S + τ S X ω † X ω ′ ) A 1 ( ω ) ] . (66) \tilde{\Delta}=\sum_\omega\left[\tau_S X_\omega^\dagger X_\omega\left(\beta A_1(\omega)+\frac{dA_1(\omega)}{d\omega}\right)+\tau_S X_\omega X_\omega^\dagger\left(\beta A_2(\omega)-\frac{dA_2(\omega)}{d\omega}\right)\right]
+\sum_{\omega\ne\omega'}(\omega'-\omega)^{-1}\left[(X_{\omega'}X_\omega^\dagger\tau_S+\tau_SX_\omega X_{\omega'}^\dagger)A_2(\omega)-(X_{\omega'}^\dagger X_\omega\tau_S+\tau_SX_\omega^\dagger X_{\omega'})A_1(\omega)\right]. \tag*{(66)} Δ ~ = ω ∑ [ τ S X ω † X ω ( β A 1 ( ω ) + d ω d A 1 ( ω ) ) + τ S X ω X ω † ( β A 2 ( ω ) − d ω d A 2 ( ω ) ) ] + ω = ω ′ ∑ ( ω ′ − ω ) − 1 [ ( X ω ′ X ω † τ S + τ S X ω X ω ′ † ) A 2 ( ω ) − ( X ω ′ † X ω τ S + τ S X ω † X ω ′ ) A 1 ( ω ) ] . (66)
To normalise ρ ~ S \tilde{\rho}_S ρ ~ S we require its trace. Taking the trace of the first relation in Eq. (S32), using cyclicity and rearranging gives
( 1 − e − β ( ω − ω ′ ) ) Tr S [ e − β K S X ω X ω ′ † ] = 0. (67) \left(1-e^{-\beta(\omega-\omega')}\right)\operatorname{Tr}_S\left[e^{-\beta K_S}X_\omega X_{\omega'}^\dagger\right]=0. \tag*{(67)} ( 1 − e − β ( ω − ω ′ ) ) Tr S [ e − β K S X ω X ω ′ † ] = 0. (67)
Since ω ≠ ω ′ \omega\ne\omega' ω = ω ′ , this gives Tr S [ e − β K S X ω X ω ′ † ] = 0 \operatorname{Tr}_S[e^{-\beta K_S}X_\omega X_{\omega'}^\dagger]=0 Tr S [ e − β K S X ω X ω ′ † ] = 0 , and the same argument applies to the other three off-diagonal terms, so only the diagonal terms of Δ ~ \tilde{\Delta} Δ ~ contribute to the trace. Their coefficients simplify. Differentiating Eq. (S27),
A 1 ′ ( ω ) = − ∫ 0 β d τ τ e − τ ω G 1 ( β , − i τ ) , A 2 ′ ( ω ) = ∫ 0 β d τ τ e τ ω G 2 ( β , − i τ ) , (68) A_1'(\omega)=-\int_0^\beta d\tau\,\tau e^{-\tau\omega}G_1(\beta,-i\tau),\qquad A_2'(\omega)=\int_0^\beta d\tau\,\tau e^{\tau\omega}G_2(\beta,-i\tau), \tag*{(68)} A 1 ′ ( ω ) = − ∫ 0 β d τ τ e − τ ω G 1 ( β , − i τ ) , A 2 ′ ( ω ) = ∫ 0 β d τ τ e τ ω G 2 ( β , − i τ ) , (68)
so that
D 1 ( ω ) ≡ β A 1 ( ω ) + A 1 ′ ( ω ) = ∫ 0 β d τ ( β − τ ) e − τ ω G 1 ( β , − i τ ) , (69) D_1(\omega)\equiv\beta A_1(\omega)+A_1'(\omega)=\int_0^\beta d\tau\,(\beta-\tau)e^{-\tau\omega}G_1(\beta,-i\tau), \tag*{(69)} D 1 ( ω ) ≡ β A 1 ( ω ) + A 1 ′ ( ω ) = ∫ 0 β d τ ( β − τ ) e − τ ω G 1 ( β , − i τ ) , (69)
D 2 ( ω ) ≡ β A 2 ( ω ) − A 2 ′ ( ω ) = ∫ 0 β d τ ( β − τ ) e τ ω G 2 ( β , − i τ ) . (70) D_2(\omega)\equiv\beta A_2(\omega)-A_2'(\omega)=\int_0^\beta d\tau\,(\beta-\tau)e^{\tau\omega}G_2(\beta,-i\tau). \tag*{(70)} D 2 ( ω ) ≡ β A 2 ( ω ) − A 2 ′ ( ω ) = ∫ 0 β d τ ( β − τ ) e τ ω G 2 ( β , − i τ ) . (70)
and hence
Z ~ S = Tr S ρ ~ S = 1 + λ 2 z , z ≡ Tr S Δ ~ = ∑ ω [ ⟨ X ω † X ω ⟩ τ S D 1 ( ω ) + ⟨ X ω X ω † ⟩ τ S D 2 ( ω ) ] . (71) \widetilde{Z}_S=\operatorname{Tr}_S\widetilde{\rho}_S=1+\lambda^2z,\qquad z\equiv\operatorname{Tr}_S\widetilde{\Delta}=\sum_{\omega}\left[\langle X_\omega^\dagger X_\omega\rangle_{\tau_S}D_1(\omega)+\langle X_\omega X_\omega^\dagger\rangle_{\tau_S}D_2(\omega)\right]. \tag*{(71)} Z S = Tr S ρ S = 1 + λ 2 z , z ≡ Tr S Δ = ω ∑ [ ⟨ X ω † X ω ⟩ τ S D 1 ( ω ) + ⟨ X ω X ω † ⟩ τ S D 2 ( ω ) ] . (71)
where ⟨ O ⟩ τ S ≡ Tr S [ τ S O ] \langle O\rangle_{\tau_S}\equiv\operatorname{Tr}_S[\tau_S O] ⟨ O ⟩ τ S ≡ Tr S [ τ S O ] . We now consider τ M F ( λ ) = ρ ~ S ( λ ) / Z ~ S ( λ ) \tau_{\mathrm{MF}}(\lambda)=\widetilde{\rho}_S(\lambda)/\widetilde{Z}_S(\lambda) τ MF ( λ ) = ρ S ( λ ) / Z S ( λ ) as a Taylor expansion following the approach in [33 ] , noting that it is even in λ \lambda λ ,
τ M F ( λ ) = τ M F ( 0 ) + 1 2 λ 2 d 2 τ M F d λ 2 ( 0 ) + O ( λ 4 ) . (72) \tau_{\mathrm{MF}}(\lambda)=\tau_{\mathrm{MF}}(0)+\frac{1}{2}\lambda^2\frac{d^2\tau_{\mathrm{MF}}}{d\lambda^2}(0)+O(\lambda^4). \tag*{(72)} τ MF ( λ ) = τ MF ( 0 ) + 2 1 λ 2 d λ 2 d 2 τ MF ( 0 ) + O ( λ 4 ) . (72)
Using Z ~ S ( 0 ) = 1 \widetilde{Z}_S(0)=1 Z S ( 0 ) = 1 , τ M F ( 0 ) = τ S \tau_{\mathrm{MF}}(0)=\tau_S τ MF ( 0 ) = τ S , 1 2 d 2 ρ ~ S d λ 2 ( 0 ) = Δ ~ \frac{1}{2}\frac{d^2\widetilde{\rho}_S}{d\lambda^2}(0)=\widetilde{\Delta} 2 1 d λ 2 d 2 ρ S ( 0 ) = Δ and 1 2 d 2 Z ~ S d λ 2 ( 0 ) = z \frac{1}{2}\frac{d^2\widetilde{Z}_S}{d\lambda^2}(0)=z 2 1 d λ 2 d 2 Z S ( 0 ) = z we find
τ M F ( 2 ) = τ S + λ 2 ( Δ ~ − z τ S ) + O ( λ 4 ) . (73) \tau_{\mathrm{MF}}^{(2)}=\tau_S+\lambda^2(\widetilde{\Delta}-z\tau_S)+O(\lambda^4). \tag*{(73)} τ MF ( 2 ) = τ S + λ 2 ( Δ − z τ S ) + O ( λ 4 ) . (73)
This gives the final result,
τ M F ( 2 ) = τ S + λ 2 ( Δ d i a g + Δ o f f ) = τ S + λ 2 ∑ ω [ τ S ( X ω † X ω − ⟨ X ω † X ω ⟩ τ S ) D 1 ( ω ) + τ S ( X ω X ω † − ⟨ X ω X ω † ⟩ τ S ) D 2 ( ω ) ] + λ 2 ∑ ω ≠ ω ′ ( ω ′ − ω ) − 1 [ ( X ω ′ X ω † τ S + τ S X ω X ω ′ † ) A 2 ( ω ) − ( X ω ′ † X ω τ S + τ S X ω † X ω ′ ) A 1 ( ω ) ] . (74) \begin{aligned}
\tau_{\mathrm{MF}}^{(2)}=\tau_S+\lambda^2(\Delta_{\mathrm{diag}}+\Delta_{\mathrm{off}}) \\
&=\tau_S+\lambda^2\sum_{\omega}\left[\tau_S\left(X_\omega^\dagger X_\omega-\langle X_\omega^\dagger X_\omega\rangle_{\tau_S}\right)D_1(\omega)+\tau_S\left(X_\omega X_\omega^\dagger-\langle X_\omega X_\omega^\dagger\rangle_{\tau_S}\right)D_2(\omega)\right] \\
&+\lambda^2\sum_{\omega\ne\omega'}(\omega'-\omega)^{-1}\left[\left(X_{\omega'}X_\omega^\dagger\tau_S+\tau_S X_\omega X_{\omega'}^\dagger\right)A_2(\omega)-\left(X_{\omega'}^\dagger X_\omega\tau_S+\tau_S X_\omega^\dagger X_{\omega'}\right)A_1(\omega)\right]. \tag*{(74)}
\end{aligned} τ MF ( 2 ) = τ S + λ 2 ( Δ diag + Δ off ) = τ S + λ 2 ω ∑ [ τ S ( X ω † X ω − ⟨ X ω † X ω ⟩ τ S ) D 1 ( ω ) + τ S ( X ω X ω † − ⟨ X ω X ω † ⟩ τ S ) D 2 ( ω ) ] + λ 2 ω = ω ′ ∑ ( ω ′ − ω ) − 1 [ ( X ω ′ X ω † τ S + τ S X ω X ω ′ † ) A 2 ( ω ) − ( X ω ′ † X ω τ S + τ S X ω † X ω ′ ) A 1 ( ω ) ] . (74)
This is Eq. (11) of the main text.
Coherences
Here we show that τ M F ( 2 ) \tau_{\mathrm{MF}}^{(2)} τ MF ( 2 ) contains coherences in the eigenbasis of H S H_S H S . Every operator appearing in Eq. (74) conserves N S N_S N S , so its commutator with H S H_S H S equals its commutator with K S K_S K S and the shifts of the frequency labels by μ \mu μ cancel throughout. We therefore work with H S H_S H S and the unshifted Bohr frequencies. For any Bohr frequency ω \omega ω ,
[ H S , X ω † X ω ] = [ H S , X ω † ] X ω + X ω † [ H S , X ω ] = − ω X ω † X ω + ω X ω † X ω = 0 , (75) [H_S,X_\omega^\dagger X_\omega]=[H_S,X_\omega^\dagger]X_\omega+X_\omega^\dagger[H_S,X_\omega]=-\omega X_\omega^\dagger X_\omega+\omega X_\omega^\dagger X_\omega=0, \tag*{(75)} [ H S , X ω † X ω ] = [ H S , X ω † ] X ω + X ω † [ H S , X ω ] = − ω X ω † X ω + ω X ω † X ω = 0 , (75)
and likewise [ H S , X ω X ω † ] = 0 [H_S,X_\omega X_\omega^\dagger]=0 [ H S , X ω X ω † ] = 0 . Since [ τ S , H S ] = 0 [\tau_S,H_S]=0 [ τ S , H S ] = 0 it follows that [ H S , Δ d i a g ] = 0 [H_S,\Delta_{\mathrm{diag}}]=0 [ H S , Δ diag ] = 0 . Now consider Δ o f f \Delta_{\mathrm{off}} Δ off . Using [ H S , X ω ′ X ω † ] = ( ω ′ − ω ) X ω ′ X ω † [H_S,X_{\omega'}X_\omega^\dagger]=(\omega'-\omega)X_{\omega'}X_\omega^\dagger [ H S , X ω ′ X ω † ] = ( ω ′ − ω ) X ω ′ X ω † together with the three analogous identities for X ω X ω ′ † X_\omega X_{\omega'}^\dagger X ω X ω ′ † , X ω ′ † X ω X_{\omega'}^\dagger X_\omega X ω ′ † X ω and X ω † X ω ′ X_\omega^\dagger X_{\omega'} X ω † X ω ′ gives
[ X ω ′ X ω † τ S , H S ] = ( ω − ω ′ ) X ω ′ X ω † τ S , [ τ S X ω X ω ′ † , H S ] = ( ω ′ − ω ) τ S X ω X ω ′ † , (76) [X_{\omega'}X_\omega^\dagger\tau_S,H_S]=(\omega-\omega')X_{\omega'}X_\omega^\dagger\tau_S,\qquad[\tau_S X_\omega X_{\omega'}^\dagger,H_S]=(\omega'-\omega)\tau_S X_\omega X_{\omega'}^\dagger, \tag*{(76)} [ X ω ′ X ω † τ S , H S ] = ( ω − ω ′ ) X ω ′ X ω † τ S , [ τ S X ω X ω ′ † , H S ] = ( ω ′ − ω ) τ S X ω X ω ′ † , (76)
[ X ω ′ † X ω τ S , H S ] = ( ω ′ − ω ) X ω ′ † X ω τ S , [ τ S X ω † X ω ′ , H S ] = ( ω − ω ′ ) τ S X ω † X ω ′ . (77) [X_{\omega'}^\dagger X_\omega\tau_S,H_S]=(\omega'-\omega)X_{\omega'}^\dagger X_\omega\tau_S,\qquad[\tau_S X_\omega^\dagger X_{\omega'},H_S]=(\omega-\omega')\tau_S X_\omega^\dagger X_{\omega'}. \tag*{(77)} [ X ω ′ † X ω τ S , H S ] = ( ω ′ − ω ) X ω ′ † X ω τ S , [ τ S X ω † X ω ′ , H S ] = ( ω − ω ′ ) τ S X ω † X ω ′ . (77)
In each case the factor ( ω ′ − ω ) (\omega'-\omega) ( ω ′ − ω ) produced by the commutator cancels the prefactor ( ω ′ − ω ) − 1 (\omega'-\omega)^{-1} ( ω ′ − ω ) − 1 in Eq. (74) , leaving
[ τ M F ( 2 ) , H S ] = λ 2 ∑ ω ≠ ω ′ { A 2 ( ω ) [ τ S X ω X ω ′ † − X ω ′ X ω † τ S ] + A 1 ( ω ) [ τ S X ω † X ω ′ − X ω ′ † X ω τ S ] } . (78) [\tau_{\mathrm{MF}}^{(2)},H_S]=\lambda^2\sum_{\omega\ne\omega'}\left\{A_2(\omega)\left[\tau_S X_\omega X_{\omega'}^\dagger-X_{\omega'}X_\omega^\dagger\tau_S\right]+A_1(\omega)\left[\tau_S X_\omega^\dagger X_{\omega'}-X_{\omega'}^\dagger X_\omega\tau_S\right]\right\}. \tag*{(78)} [ τ MF ( 2 ) , H S ] = λ 2 ω = ω ′ ∑ { A 2 ( ω ) [ τ S X ω X ω ′ † − X ω ′ X ω † τ S ] + A 1 ( ω ) [ τ S X ω † X ω ′ − X ω ′ † X ω τ S ] } . (78)
As in the bosonic case [23 ] , this is generically non-zero.
KMS relation between the two channels
The two correlation functions G 1 G_1 G 1 and G 2 G_2 G 2 are not independent. Since τ R \tau_R τ R is thermal, for 0 ≤ τ ≤ β 0\leq\tau\leq\beta 0 ≤ τ ≤ β ,
G 1 ( β , − i τ ) = 1 Z R Tr R [ e − β K R e τ K R E e − τ K R E † ] = 1 Z R Tr R [ e − ( β − τ ) K R E e − τ K R E † ] = 1 Z R Tr R [ e − τ K R E † e − ( β − τ ) K R E ] = 1 Z R Tr R [ e − β K R e ( β − τ ) K R E † e − ( β − τ ) K R E ] = G 2 ( β , − i ( β − τ ) ) . (79) \begin{aligned}
G_1(\beta,-i\tau)=\frac{1}{Z_R}\operatorname{Tr}_R\left[e^{-\beta K_R}e^{\tau K_R}Ee^{-\tau K_R}E^\dagger\right]=\frac{1}{Z_R}\operatorname{Tr}_R\left[e^{-(\beta-\tau)K_R}Ee^{-\tau K_R}E^\dagger\right] \\
&=\frac{1}{Z_R}\operatorname{Tr}_R\left[e^{-\tau K_R}E^\dagger e^{-(\beta-\tau)K_R}E\right]=\frac{1}{Z_R}\operatorname{Tr}_R\left[e^{-\beta K_R}e^{(\beta-\tau)K_R}E^\dagger e^{-(\beta-\tau)K_R}E\right] \\
&=G_2(\beta,-i(\beta-\tau)). \tag*{(79)}
\end{aligned} G 1 ( β , − i τ ) = Z R 1 Tr R [ e − β K R e τ K R E e − τ K R E † ] = Z R 1 Tr R [ e − ( β − τ ) K R E e − τ K R E † ] = Z R 1 Tr R [ e − τ K R E † e − ( β − τ ) K R E ] = Z R 1 Tr R [ e − β K R e ( β − τ ) K R E † e − ( β − τ ) K R E ] = G 2 ( β , − i ( β − τ )) . (79)
where the third equality is cyclicity of the trace. This is the imaginary-time KMS condition and requires only that the reservoir be in equilibrium, it holds for fermionic and bosonic reservoirs alike. Substituting Eq. (79) into the definition of A 1 A_1 A 1 and changing variables to u = β − τ u=\beta-\tau u = β − τ ,
A 1 ( ω ) = ∫ 0 β d τ e − τ ω G 2 ( β , − i ( β − τ ) ) = e − β ω ∫ 0 β d u e u ω G 2 ( β , − i u ) = e − β ω A 2 ( ω ) . (80) A_1(\omega)=\int_0^\beta d\tau\,e^{-\tau\omega}G_2(\beta,-i(\beta-\tau))=e^{-\beta\omega}\int_0^\beta du\,e^{u\omega}G_2(\beta,-iu)=e^{-\beta\omega}A_2(\omega). \tag*{(80)} A 1 ( ω ) = ∫ 0 β d τ e − τ ω G 2 ( β , − i ( β − τ )) = e − β ω ∫ 0 β d u e u ω G 2 ( β , − i u ) = e − β ω A 2 ( ω ) . (80)
Differentiating and using Eq. (69) ,
D 1 ( ω ) = β A 1 ( ω ) + A 1 ′ ( ω ) = β e − β ω A 2 ( ω ) − β e − β ω A 2 ( ω ) + e − β ω A 2 ′ ( ω ) = e − β ω A 2 ′ ( ω ) , (81) D_1(\omega)=\beta A_1(\omega)+A_1'(\omega)=\beta e^{-\beta\omega}A_2(\omega)-\beta e^{-\beta\omega}A_2(\omega)+e^{-\beta\omega}A_2'(\omega)=e^{-\beta\omega}A_2'(\omega), \tag*{(81)} D 1 ( ω ) = β A 1 ( ω ) + A 1 ′ ( ω ) = β e − β ω A 2 ( ω ) − β e − β ω A 2 ( ω ) + e − β ω A 2 ′ ( ω ) = e − β ω A 2 ′ ( ω ) , (81)
Together with Eq. (70) , all four coefficient functions are determined by A 2 ( ω ) A_2(\omega) A 2 ( ω ) and A 2 ′ ( ω ) A_2'(\omega) A 2 ′ ( ω ) .
Explicit coefficient functions
We now evaluate G 1 G_1 G 1 , G 2 G_2 G 2 and hence A 1 , 2 A_{1,2} A 1 , 2 , D 1 , 2 D_{1,2} D 1 , 2 for the canonical setup outlined in Sec. I. Eqs. (S3,S5) give a ε ( − i τ ) = e − τ ξ a ε a_\varepsilon(-i\tau)=e^{-\tau\xi}a_\varepsilon a ε ( − i τ ) = e − τ ξ a ε . Writing g g g for the thermal occupation, g = f g=f g = f with f ( ξ ) = ( e β ξ + 1 ) − 1 f(\xi)=(e^{\beta\xi}+1)^{-1} f ( ξ ) = ( e β ξ + 1 ) − 1 for fermions and g = n g=n g = n with n ( ξ ) = ( e β ξ − 1 ) − 1 n(\xi)=(e^{\beta\xi}-1)^{-1} n ( ξ ) = ( e β ξ − 1 ) − 1 for bosons, and σ = + 1 \sigma=+1 σ = + 1 for fermions, σ = − 1 \sigma=-1 σ = − 1 for bosons,
G 1 ( β , − i τ ) = ∫ d ε J ( ε ) e − τ ξ [ 1 − σ g ( ξ ) ] , (82) G_1(\beta,-i\tau)=\int d\varepsilon\,J(\varepsilon)e^{-\tau\xi}\left[1-\sigma g(\xi)\right], \tag*{(82)} G 1 ( β , − i τ ) = ∫ d ε J ( ε ) e − τ ξ [ 1 − σ g ( ξ ) ] , (82)
G 2 ( β , − i τ ) = ∫ d ε J ( ε ) e τ ξ g ( ξ ) , (83) G_2(\beta,-i\tau)=\int d\varepsilon\,J(\varepsilon)e^{\tau\xi}g(\xi), \tag*{(83)} G 2 ( β , − i τ ) = ∫ d ε J ( ε ) e τ ξ g ( ξ ) , (83)
with ξ = ε − μ \xi=\varepsilon-\mu ξ = ε − μ . Performing the τ \tau τ integrals in Eqs. (S27) and (S36)–(S37) with the correlation functions of Eqs. (S49)–(S50), and writing x ≡ ξ + ω = ε − μ + ω x\equiv\xi+\omega=\varepsilon-\mu+\omega x ≡ ξ + ω = ε − μ + ω ,
A 1 ( ω ) = ∫ d ε J ( ε ) [ 1 − σ g ( ξ ) ] 1 − e − β x x , (84) A_1(\omega)=\int d\varepsilon\,J(\varepsilon)\left[1-\sigma g(\xi)\right]\frac{1-e^{-\beta x}}{x}, \tag*{(84)} A 1 ( ω ) = ∫ d ε J ( ε ) [ 1 − σ g ( ξ ) ] x 1 − e − β x , (84)
A 2 ( ω ) = ∫ d ε J ( ε ) g ( ξ ) e β x − 1 x , (85) A_2(\omega)=\int d\varepsilon\,J(\varepsilon)g(\xi)\frac{e^{\beta x}-1}{x}, \tag*{(85)} A 2 ( ω ) = ∫ d ε J ( ε ) g ( ξ ) x e β x − 1 , (85)
D 1 ( ω ) = ∫ d ε J ( ε ) [ 1 − σ g ( ξ ) ] [ β x − 1 − e − β x x 2 ] , (86) D_1(\omega)=\int d\varepsilon\,J(\varepsilon)\left[1-\sigma g(\xi)\right]\left[\frac{\beta}{x}-\frac{1-e^{-\beta x}}{x^2}\right], \tag*{(86)} D 1 ( ω ) = ∫ d ε J ( ε ) [ 1 − σ g ( ξ ) ] [ x β − x 2 1 − e − β x ] , (86)
D 2 ( ω ) = ∫ d ε J ( ε ) g ( ξ ) [ − β x + e β x − 1 x 2 ] , (87) D_2(\omega)=\int d\varepsilon\,J(\varepsilon)g(\xi)\left[-\frac{\beta}{x}+\frac{e^{\beta x}-1}{x^2}\right], \tag*{(87)} D 2 ( ω ) = ∫ d ε J ( ε ) g ( ξ ) [ − x β + x 2 e β x − 1 ] , (87)
where we used
∫ 0 β d τ ( β − τ ) e − τ x = e − β x ∫ 0 β d s s e s x = β x − 1 − e − β x x 2 . (88) \int_0^\beta d\tau\,(\beta-\tau)e^{-\tau x}=e^{-\beta x}\int_0^\beta ds\,s e^{sx}=\frac{\beta}{x}-\frac{1-e^{-\beta x}}{x^2}. \tag*{(88)} ∫ 0 β d τ ( β − τ ) e − τ x = e − β x ∫ 0 β d s s e s x = x β − x 2 1 − e − β x . (88)
Each bracket is regular at x = 0 x=0 x = 0 : expanding 1 − e − β x = β x − 1 2 β 2 x 2 + 1 6 β 3 x 3 + … 1-e^{-\beta x}=\beta x-\frac{1}{2}\beta^2x^2+\frac{1}{6}\beta^3x^3+\ldots 1 − e − β x = β x − 2 1 β 2 x 2 + 6 1 β 3 x 3 + … gives
β x − 1 − e − β x x 2 = β 2 2 − β 3 x 6 + O ( x 2 ) . (89) \frac{\beta}{x}-\frac{1-e^{-\beta x}}{x^2}=\frac{\beta^2}{2}-\frac{\beta^3x}{6}+O(x^2). \tag*{(89)} x β − x 2 1 − e − β x = 2 β 2 − 6 β 3 x + O ( x 2 ) . (89)
High-temperature limit
In this section we derive the high-temperature limit, β → 0 \beta\to0 β → 0 , of the second-order fermionic MFG state. Following [23 ] , we use the product approximation from the Zassenhaus formula [61 ] ,
e − β ( K 0 + λ V ) = 1 2 [ e − β K 0 e − β λ V + e − β λ V e − β K 0 ] + β 3 12 ( λ [ K 0 , [ K 0 , V ] ] + λ 2 [ V , [ V , K 0 ] ] ) + O ( β 4 ) . (90) e^{-\beta(K_0+\lambda V)}=\frac{1}{2}\left[e^{-\beta K_0}e^{-\beta\lambda V}+e^{-\beta\lambda V}e^{-\beta K_0}\right]+\frac{\beta^3}{12}\left(\lambda[K_0,[K_0,V]]+\lambda^2[V,[V,K_0]]\right)+O(\beta^4). \tag*{(90)} e − β ( K 0 + λV ) = 2 1 [ e − β K 0 e − β λV + e − β λV e − β K 0 ] + 12 β 3 ( λ [ K 0 , [ K 0 , V ]] + λ 2 [ V , [ V , K 0 ]] ) + O ( β 4 ) . (90)
To arrive at Eq. (90) , take expansions of e − β ( K 0 + λ V ) e^{-\beta(K_0+\lambda V)} e − β ( K 0 + λV ) , e − β K 0 e − β λ V e^{-\beta K_0}e^{-\beta\lambda V} e − β K 0 e − β λV , e − β λ V e − β K 0 e^{-\beta\lambda V}e^{-\beta K_0} e − β λV e − β K 0 and combine them. Taking the trace and noting that Tr R [ e − β λ V e − β K R ] = Tr R [ e − β K R e − β λ V ] \operatorname{Tr}_{R}[e^{-\beta\lambda V}e^{-\beta K_R}]=\operatorname{Tr}_{R}[e^{-\beta K_R}e^{-\beta\lambda V}] Tr R [ e − β λV e − β K R ] = Tr R [ e − β K R e − β λV ] gives
Tr R [ e − β ( K 0 + λ V ) ] = Z R 2 { e − β K S Tr R [ τ R e − β λ V ] } + O ( λ 2 β 3 ) . (91) \operatorname{Tr}_{R}\left[e^{-\beta(K_0+\lambda V)}\right]=\frac{Z_R}{2}\left\{e^{-\beta K_S}\operatorname{Tr}_{R}\left[\tau_R e^{-\beta\lambda V}\right]\right\}+O(\lambda^2\beta^3). \tag*{(91)} Tr R [ e − β ( K 0 + λV ) ] = 2 Z R { e − β K S Tr R [ τ R e − β λV ] } + O ( λ 2 β 3 ) . (91)
The O ( λ 2 β 3 ) O(\lambda^{2}\beta^{3}) O ( λ 2 β 3 ) error comes from the λ 2 [ V , [ V , K 0 ] ] \lambda^{2}[V,[V,K_{0}]] λ 2 [ V , [ V , K 0 ]] term and λ [ K 0 , [ K 0 , V ] ] \lambda[K_{0},[K_{0},V]] λ [ K 0 , [ K 0 , V ]] vanishes as it is linear in V V V so its trace with τ R \tau_{R} τ R gives zero by number conservation. We now expand the reservoir average as
Tr R [ τ R e − β λ V ] = 1 + β 2 λ 2 2 Tr R [ τ R V 2 ] + O ( λ 3 β 3 ) , = 1 + β 2 λ 2 2 ( X X † ⟨ E † E ⟩ τ R + X † X ⟨ E E † ⟩ τ R ) . (92) \begin{aligned}
\operatorname{Tr}_{R}\left[\tau_{R}e^{-\beta\lambda V}\right] = 1 + \frac{\beta^{2}\lambda^{2}}{2}\operatorname{Tr}_{R}\left[\tau_{R}V^{2}\right] + O(\lambda^{3}\beta^{3}), \\
&= 1 + \frac{\beta^{2}\lambda^{2}}{2}\left(XX^{\dagger}\langle E^{\dagger}E\rangle_{\tau_{R}} + X^{\dagger}X\langle EE^{\dagger}\rangle_{\tau_{R}}\right). \tag*{(92)}
\end{aligned} Tr R [ τ R e − β λV ] = 1 + 2 β 2 λ 2 Tr R [ τ R V 2 ] + O ( λ 3 β 3 ) , = 1 + 2 β 2 λ 2 ( X X † ⟨ E † E ⟩ τ R + X † X ⟨ E E † ⟩ τ R ) . (92)
Here we use the fact that odd terms and E E EE E E , E † E † E^{\dagger}E^{\dagger} E † E † terms vanish by number conservation. We have also rearranged V 2 V^{2} V 2 in the same way as done in Eqs. (48) –(50) . Using E = ∫ d ε J ( ε ) a ε E = \int d\varepsilon\sqrt{J(\varepsilon)}a_{\varepsilon} E = ∫ d ε J ( ε ) a ε , ⟨ a ε † a ε ′ ⟩ τ R = δ ( ε − ε ′ ) f ( ξ ) \langle a_{\varepsilon}^{\dagger}a_{\varepsilon'}\rangle_{\tau_{R}} = \delta(\varepsilon-\varepsilon')f(\xi) ⟨ a ε † a ε ′ ⟩ τ R = δ ( ε − ε ′ ) f ( ξ ) , and expanding f ( ξ ) = 1 2 − 1 4 β ξ + O ( β 3 ξ 3 ) f(\xi) = \frac{1}{2} - \frac{1}{4}\beta\xi+ O(\beta^{3}\xi^{3}) f ( ξ ) = 2 1 − 4 1 β ξ + O ( β 3 ξ 3 ) ,
⟨ E E † ⟩ τ R = ∫ d ε J ( ε ) ( 1 − f ( ξ ) ) = 1 2 m 0 + 1 4 β m 1 + O ( β 3 ) , (93) \langle EE^{\dagger}\rangle_{\tau_{R}} = \int d\varepsilon J(\varepsilon)(1-f(\xi)) = \frac{1}{2}m_{0}+\frac{1}{4}\beta m_{1}+O(\beta^{3}), \tag*{(93)} ⟨ E E † ⟩ τ R = ∫ d ε J ( ε ) ( 1 − f ( ξ )) = 2 1 m 0 + 4 1 β m 1 + O ( β 3 ) , (93)
⟨ E † E ⟩ τ R = ∫ d ε J ( ε ) f ( ξ ) = 1 2 m 0 − 1 4 β m 1 + O ( β 3 ) , (94) \langle E^{\dagger}E\rangle_{\tau_{R}} = \int d\varepsilon J(\varepsilon)f(\xi) = \frac{1}{2}m_{0}-\frac{1}{4}\beta m_{1}+O(\beta^{3}), \tag*{(94)} ⟨ E † E ⟩ τ R = ∫ d ε J ( ε ) f ( ξ ) = 2 1 m 0 − 4 1 β m 1 + O ( β 3 ) , (94)
m k ≡ ∫ d ε J ( ε ) ( ε − μ ) k . (95) m_{k}\equiv\int d\varepsilon J(\varepsilon)(\varepsilon-\mu)^{k}. \tag*{(95)} m k ≡ ∫ d ε J ( ε ) ( ε − μ ) k . (95)
The m 1 m_{1} m 1 terms multiply β 2 \beta^{2} β 2 and so will be O ( β 3 ) O(\beta^{3}) O ( β 3 ) and can be dropped, giving
Tr R [ τ R e − β λ V ] = 1 + λ 2 β 2 m 0 4 C + O ( β 3 ) . (96) \operatorname{Tr}_{R}\left[\tau_{R}e^{-\beta\lambda V}\right] = 1 + \frac{\lambda^{2}\beta^{2}m_{0}}{4}C + O(\beta^{3}). \tag*{(96)} Tr R [ τ R e − β λV ] = 1 + 4 λ 2 β 2 m 0 C + O ( β 3 ) . (96)
with C = { X , X † } C=\{X,X^{\dagger}\} C = { X , X † } . Substituting this into Eq. (91) then gives for the un-normalised state
ρ ~ S = e − β K S + λ 2 β 2 m 0 8 { e − β K S , C } + O ( β 3 ) . (97) \widetilde{\rho}_{S}=e^{-\beta K_{S}}+\frac{\lambda^{2}\beta^{2}m_{0}}{8}\{e^{-\beta K_{S}},C\}+O(\beta^{3}). \tag*{(97)} ρ S = e − β K S + 8 λ 2 β 2 m 0 { e − β K S , C } + O ( β 3 ) . (97)
Using e − β K S = 1 + O ( β ) e^{-\beta K_{S}}=1+O(\beta) e − β K S = 1 + O ( β ) which gives { e − β K S , C } = 2 C + O ( β ) \{e^{-\beta K_{S}},C\}=2C+O(\beta) { e − β K S , C } = 2 C + O ( β ) and dividing by the trace,
τ M F = e − β K S + 1 4 λ 2 β 2 m 0 C Z S + 1 4 λ 2 β 2 m 0 Tr ( C ) + O ( β 3 ) = τ S + 1 4 λ 2 β 2 m 0 ( C Z S − τ S Tr ( C ) Z S ) + O ( β 3 ) , (98) \tau_{\mathrm{MF}}=\frac{e^{-\beta K_{S}}+\frac{1}{4}\lambda^{2}\beta^{2}m_{0}C}{Z_{S}+\frac{1}{4}\lambda^{2}\beta^{2}m_{0}\operatorname{Tr}(C)}+O(\beta^{3})=\tau_{S}+\frac{1}{4}\lambda^{2}\beta^{2}m_{0}\left(\frac{C}{Z_{S}}-\tau_{S}\frac{\operatorname{Tr}(C)}{Z_{S}}\right)+O(\beta^{3}), \tag*{(98)} τ MF = Z S + 4 1 λ 2 β 2 m 0 Tr ( C ) e − β K S + 4 1 λ 2 β 2 m 0 C + O ( β 3 ) = τ S + 4 1 λ 2 β 2 m 0 ( Z S C − τ S Z S Tr ( C ) ) + O ( β 3 ) , (98)
where the denominator has been expanded to first order, with the next term being O ( β 4 ) O(\beta^{4}) O ( β 4 ) . Using Z S = d S + O ( β ) Z_{S}=d_{S}+O(\beta) Z S = d S + O ( β ) , τ S = 1 / d S + O ( β ) \tau_{S}=1/d_{S}+O(\beta) τ S = 1/ d S + O ( β ) , we have
τ M F = τ S + λ 2 β 2 m 0 4 d S δ C + O ( β 3 ) , δ C ≡ C − ⟨ C ⟩ 0 , (99) \tau_{\mathrm{MF}}=\tau_{S}+\frac{\lambda^{2}\beta^{2}m_{0}}{4d_{S}}\delta C+O(\beta^{3}),\qquad\delta C\equiv C-\langle C\rangle_{0}, \tag*{(99)} τ MF = τ S + 4 d S λ 2 β 2 m 0 δ C + O ( β 3 ) , δ C ≡ C − ⟨ C ⟩ 0 , (99)
where ⟨ C ⟩ 0 = 1 Tr ( C ) / d S \langle C\rangle_{0}=1\operatorname{Tr}(C)/d_{S} ⟨ C ⟩ 0 = 1 Tr ( C ) / d S . The correction is traceless as required. This also holds for all coupling strengths, provided the temperature well exceeds the coupling energy scale. If X X X is linear in system annihilation operators, X = ∑ i u i s i X=\sum_{i}u_{i}s_{i} X = ∑ i u i s i , then C = ∑ i ∣ u i ∣ 2 1 C=\sum_{i}|u_{i}|^{2}1 C = ∑ i ∣ u i ∣ 2 1 and δ C = 0 \delta C=0 δ C = 0 , meaning τ M F = τ S + O ( β 3 ) \tau_{\mathrm{MF}}=\tau_{S}+O(\beta^{3}) τ MF = τ S + O ( β 3 ) . This is the case for all three models considered in the main text.
Low-temperature limit
In this section we derive the low-temperature limit, β → ∞ \beta\to\infty β → ∞ , of the second-order MFG state for a fermionic reservoir. Physically, one expects to replace the system Gibbs state by the ground state of the system and the reservoir Fermi function by a step function. However, care must be taken. The coefficient functions in Eq. (74) contain factors that grow exponentially with β \beta β . These are compensated by exponentially small factors in τ S \tau_{S} τ S and f f f . Finite contributions therefore arise from products of exponentially large and exponentially small factors, which are lost if the small factors are set to zero first. To remedy this, we rewrite the result in terms of coefficient functions that remain bounded as β → ∞ \beta\to\infty β → ∞ .
We introduce the particle and hole functions
B − ( ω ) ≡ P ∫ d ε J ( ε ) [ 1 − σ g ( ε − μ ) ] ε − μ + ω , B + ( ω ) ≡ P ∫ d ε J ( ε ) g ( ε − μ ) ε − μ + ω , (100) B_{-}(\omega) \equiv\mathcal{P}\int d\varepsilon\frac{J(\varepsilon)[1-\sigma g(\varepsilon-\mu)]}{\varepsilon-\mu+\omega}, \qquad B_{+}(\omega) \equiv\mathcal{P}\int d\varepsilon\frac{J(\varepsilon)g(\varepsilon-\mu)}{\varepsilon-\mu+\omega}, \tag*{(100)} B − ( ω ) ≡ P ∫ d ε ε − μ + ω J ( ε ) [ 1 − σ g ( ε − μ )] , B + ( ω ) ≡ P ∫ d ε ε − μ + ω J ( ε ) g ( ε − μ ) , (100)
where P \mathcal{P} P denotes the principal value and, as in Sec. SI F, g = f g=f g = f , σ = + 1 \sigma=+1 σ = + 1 for a fermionic reservoir and g = n g=n g = n , σ = − 1 \sigma=-1 σ = − 1 for a bosonic one. Since 0 ≤ f ≤ 1 0\leq f\leq1 0 ≤ f ≤ 1 , both remain bounded for fermions as β → ∞ \beta\to\infty β → ∞ at fixed ω \omega ω . We proceed to rewrite A 1 , 2 , D 1 , 2 A_{1,2},D_{1,2} A 1 , 2 , D 1 , 2 in terms of B − , B + B_{-},B_{+} B − , B + . In Eq. (85) , write g ( ξ ) e β x = g ( ξ ) e β ξ e β ω = [ 1 − σ g ( ξ ) ] e β ω g(\xi)e^{\beta x}=g(\xi)e^{\beta\xi}e^{\beta\omega}=[1-\sigma g(\xi)]e^{\beta\omega} g ( ξ ) e β x = g ( ξ ) e β ξ e β ω = [ 1 − σ g ( ξ )] e β ω using the detailed-balance relation g ( ξ ) e β ξ = 1 − σ g ( ξ ) g(\xi)e^{\beta\xi}=1-\sigma g(\xi) g ( ξ ) e β ξ = 1 − σ g ( ξ ) . The integrand of A 2 A_2 A 2 then splits into two terms, each with a simple pole at ξ = − ω \xi=-\omega ξ = − ω (x = 0 x=0 x = 0 ). The sum is regular, so each term may be taken as a principal value, giving
A 2 ( ω ) = e β ω B − ( ω ) − B + ( ω ) , A 1 ( ω ) = e − β ω A 2 ( ω ) = B − ( ω ) − e − β ω B + ( ω ) , (101) A_2(\omega)=e^{\beta\omega}B_{-}(\omega)-B_{+}(\omega), \qquad A_1(\omega)=e^{-\beta\omega}A_2(\omega)=B_{-}(\omega)-e^{-\beta\omega}B_{+}(\omega), \tag*{(101)} A 2 ( ω ) = e β ω B − ( ω ) − B + ( ω ) , A 1 ( ω ) = e − β ω A 2 ( ω ) = B − ( ω ) − e − β ω B + ( ω ) , (101)
where the second relation uses Eq. (80) . Differentiating Eq. (101) , A 2 ′ = β e β ω B − + e β ω B − ′ − B + ′ A'_2=\beta e^{\beta\omega}B_{-}+e^{\beta\omega}B'_{-}-B'_{+} A 2 ′ = β e β ω B − + e β ω B − ′ − B + ′ . From Eq. (81) and D 2 = β A 2 − A 2 ′ D_2=\beta A_2-A'_2 D 2 = β A 2 − A 2 ′ ,
D 1 ( ω ) = β B − ( ω ) + B − ′ ( ω ) − e − β ω B + ′ ( ω ) , D 2 ( ω ) = − β B + ( ω ) − e β ω B − ′ ( ω ) + B + ′ ( ω ) . (102) D_1(\omega)=\beta B_{-}(\omega)+B'_{-}(\omega)-e^{-\beta\omega}B'_{+}(\omega), \qquad D_2(\omega)=-\beta B_{+}(\omega)-e^{\beta\omega}B'_{-}(\omega)+B'_{+}(\omega). \tag*{(102)} D 1 ( ω ) = β B − ( ω ) + B − ′ ( ω ) − e − β ω B + ′ ( ω ) , D 2 ( ω ) = − β B + ( ω ) − e β ω B − ′ ( ω ) + B + ′ ( ω ) . (102)
The reason for doing this is that now all exponential dependence on β \beta β is in explicit factors e ± β ω e^{\pm\beta\omega} e ± β ω , multiplying bounded functions. Inserting Eq. (102) into τ S X ω † X ω D 1 + τ S X ω X ω † D 2 \tau_SX^\dagger_\omega X_\omega D_1+\tau_SX_\omega X^\dagger_\omega D_2 τ S X ω † X ω D 1 + τ S X ω X ω † D 2 and using Eq. (64) , we get e − β ω τ S X ω † X ω = X ω † τ S X ω e^{-\beta\omega}\tau_SX^\dagger_\omega X_\omega=X^\dagger_\omega\tau_SX_\omega e − β ω τ S X ω † X ω = X ω † τ S X ω and e β ω τ S X ω X ω † = X ω τ S X ω † e^{\beta\omega}\tau_SX_\omega X^\dagger_\omega=X_\omega\tau_SX^\dagger_\omega e β ω τ S X ω X ω † = X ω τ S X ω † . The diagonal part of Δ ~ \tilde{\Delta} Δ ~ is therefore
∑ ω { β [ B − ( ω ) τ S X ω † X ω − B + ( ω ) τ S X ω X ω † ] + B − ′ ( ω ) [ τ S X ω † , X ω ] + B + ′ ( ω ) [ τ S X ω , X ω † ] } . (103) \sum_\omega\left\{\beta\left[B_{-}(\omega)\tau_SX^\dagger_\omega X_\omega-B_{+}(\omega)\tau_SX_\omega X^\dagger_\omega\right]+B'_{-}(\omega)[\tau_SX^\dagger_\omega,X_\omega]+B'_{+}(\omega)[\tau_SX_\omega,X^\dagger_\omega]\right\}. \tag*{(103)} ω ∑ { β [ B − ( ω ) τ S X ω † X ω − B + ( ω ) τ S X ω X ω † ] + B − ′ ( ω ) [ τ S X ω † , X ω ] + B + ′ ( ω ) [ τ S X ω , X ω † ] } . (103)
Inserting Eq. (101) into Δ o f f \Delta_{\mathrm{off}} Δ off and using Eq. (64) gives four replacements:
e β ω X ω ′ X ω † τ S = X ω ′ τ S X ω † , e β ω τ S X ω X ω ′ † = X ω τ S X ω ′ † , e − β ω X ω ′ † X ω τ S = X ω ′ † τ S X ω , e − β ω τ S X ω † X ω ′ = X ω † τ S X ω ′ . (104) e^{\beta\omega}X_{\omega'}X^\dagger_\omega\tau_S=X_{\omega'}\tau_SX^\dagger_\omega,\qquad e^{\beta\omega}\tau_SX_\omega X^\dagger_{\omega'}=X_\omega\tau_SX^\dagger_{\omega'},\qquad e^{-\beta\omega}X^\dagger_{\omega'}X_\omega\tau_S=X^\dagger_{\omega'}\tau_SX_\omega,\qquad e^{-\beta\omega}\tau_SX^\dagger_\omega X_{\omega'}=X^\dagger_\omega\tau_SX_{\omega'}. \tag*{(104)} e β ω X ω ′ X ω † τ S = X ω ′ τ S X ω † , e β ω τ S X ω X ω ′ † = X ω τ S X ω ′ † , e − β ω X ω ′ † X ω τ S = X ω ′ † τ S X ω , e − β ω τ S X ω † X ω ′ = X ω † τ S X ω ′ . (104)
Collecting the coefficients of B − ( ω ) B_{-}(\omega) B − ( ω ) and B + ( ω ) B_{+}(\omega) B + ( ω ) ,
Δ o f f = ∑ ω ≠ ω ′ 1 ω ′ − ω { B − ( ω ) ( [ X ω ′ , τ S X ω † ] + H . c . ) + B + ( ω ) ( [ X ω ′ † , τ S X ω ] + H . c . ) } . (105) \Delta_{\mathrm{off}}=\sum_{\omega\neq\omega'}\frac{1}{\omega'-\omega}\left\{B_{-}(\omega)\left([X_{\omega'},\tau_SX^\dagger_\omega]+\mathrm{H.c.}\right)+B_{+}(\omega)\left([X^\dagger_{\omega'},\tau_SX_\omega]+\mathrm{H.c.}\right)\right\}. \tag*{(105)} Δ off = ω = ω ′ ∑ ω ′ − ω 1 { B − ( ω ) ( [ X ω ′ , τ S X ω † ] + H.c. ) + B + ( ω ) ( [ X ω ′ † , τ S X ω ] + H.c. ) } . (105)
The commutators in Eq. (103) and all of Δ o f f \Delta_{\mathrm{off}} Δ off are traceless, so
z = Tr Δ ~ = β ∑ ω { B − ( ω ) ⟨ X ω † X ω ⟩ τ S − B + ( ω ) ⟨ X ω X ω † ⟩ τ S } . (106) z=\operatorname{Tr}\tilde{\Delta}=\beta\sum_\omega\left\{B_{-}(\omega)\langle X^\dagger_\omega X_\omega\rangle_{\tau_S}-B_{+}(\omega)\langle X_\omega X^\dagger_\omega\rangle_{\tau_S}\right\}. \tag*{(106)} z = Tr Δ ~ = β ω ∑ { B − ( ω ) ⟨ X ω † X ω ⟩ τ S − B + ( ω ) ⟨ X ω X ω † ⟩ τ S } . (106)
Combining Eqs. (103) –(106) and using Eq. (73) , the second-order state in bounded form is
τ M F ( 2 ) = τ S + λ 2 { β ∑ ω [ B − ( ω ) τ S ( X ω † X ω − ⟨ X ω † X ω ⟩ τ S ) − B + ( ω ) τ S ( X ω X ω † − ⟨ X ω X ω † ⟩ τ S ) ] + ∑ ω [ B − ′ ( ω ) [ τ S X ω † , X ω ] + B + ′ ( ω ) [ τ S X ω , X ω † ] ] + ∑ ω ≠ ω ′ 1 ω ′ − ω [ B − ( ω ) ( [ X ω ′ , τ S X ω † ] + H . c . ) + B + ( ω ) ( [ X ω ′ † , τ S X ω ] + H . c . ) ] } . (107) \tau_{\mathrm{MF}}^{(2)}=\tau_S+\lambda^2\left\{\beta\sum_\omega\left[B_{-}(\omega)\tau_S\left(X^\dagger_\omega X_\omega-\langle X^\dagger_\omega X_\omega\rangle_{\tau_S}\right)-B_{+}(\omega)\tau_S\left(X_\omega X^\dagger_\omega-\langle X_\omega X^\dagger_\omega\rangle_{\tau_S}\right)\right]
+\sum_\omega\left[B'_{-}(\omega)[\tau_SX^\dagger_\omega,X_\omega]+B'_{+}(\omega)[\tau_SX_\omega,X^\dagger_\omega]\right]
+\sum_{\omega\neq\omega'}\frac{1}{\omega'-\omega}\left[B_{-}(\omega)\left([X_{\omega'},\tau_SX^\dagger_\omega]+\mathrm{H.c.}\right)+B_{+}(\omega)\left([X^\dagger_{\omega'},\tau_SX_\omega]+\mathrm{H.c.}\right)\right]\right\}. \tag*{(107)} τ MF ( 2 ) = τ S + λ 2 ⎩ ⎨ ⎧ β ω ∑ [ B − ( ω ) τ S ( X ω † X ω − ⟨ X ω † X ω ⟩ τ S ) − B + ( ω ) τ S ( X ω X ω † − ⟨ X ω X ω † ⟩ τ S ) ] + ω ∑ [ B − ′ ( ω ) [ τ S X ω † , X ω ] + B + ′ ( ω ) [ τ S X ω , X ω † ] ] + ω = ω ′ ∑ ω ′ − ω 1 [ B − ( ω ) ( [ X ω ′ , τ S X ω † ] + H.c. ) + B + ( ω ) ( [ X ω ′ † , τ S X ω ] + H.c. ) ] ⎭ ⎬ ⎫ . (107)
Equation (S74) is an exact rewriting of Eq. (74) , valid at all temperatures. It has the same structure as Eq. (3) of Ref. [23], with the single function D β ( ω ) D_\beta(\omega) D β ( ω ) of the Hermitian case replaced by two functions B ∓ ( ω ) B_{\mp}(\omega) B ∓ ( ω ) for the particle and hole channels.
The low-temperature limit
Now we have rewritten Eq. (74) in a bounded form, we can go ahead and take the low temperature limit, where in the following we assume that K S K_S K S has a non-degenerate ground state ∣ g ⟩ |g\rangle ∣ g ⟩ . We also restrict to the fermionic case here, although we note that the bounded form Eq. (107) applies to both fermions and bosons. We write K S ∣ n ⟩ = ε n ∣ n ⟩ K_S|n\rangle=\varepsilon_n|n\rangle K S ∣ n ⟩ = ε n ∣ n ⟩ and Ω n ≡ ϵ n − ϵ g \Omega_n \equiv\epsilon_n-\epsilon_g Ω n ≡ ϵ n − ϵ g , so that Ω n > 0 \Omega_n>0 Ω n > 0 for n ≠ g n\ne g n = g . We also assume J ( ϵ ) J(\epsilon) J ( ϵ ) is finite and smooth near ϵ = μ \epsilon=\mu ϵ = μ . The limit is taken in the O ( λ 2 ) O(\lambda^2) O ( λ 2 ) coefficient, i.e. at λ → 0 \lambda\to0 λ → 0 first.
As β → ∞ \beta\to\infty β → ∞ , τ S = ∣ g ⟩ ⟨ g ∣ + O ( e − β Ω 1 ) \tau_S=|g\rangle\langle g|+O(e^{-\beta\Omega_1}) τ S = ∣ g ⟩ ⟨ g ∣ + O ( e − β Ω 1 ) , where Ω 1 \Omega_1 Ω 1 is the smallest excitation energy, and f ( ϵ − μ ) → Θ ( μ − ϵ ) f(\epsilon-\mu)\to\Theta(\mu-\epsilon) f ( ϵ − μ ) → Θ ( μ − ϵ ) . Hence
B − ( ω ) → B − 0 ( ω ) ≡ P ∫ d ϵ J ( ϵ ) Θ ( ϵ − μ ) ϵ − μ + ω , B + ( ω ) → B + 0 ( ω ) ≡ P ∫ d ϵ J ( ϵ ) Θ ( μ − ϵ ) ϵ − μ + ω . (108) B_-(\omega)\to B_-^0(\omega)\equiv\mathcal{P}\int d\epsilon\,\frac{J(\epsilon)\Theta(\epsilon-\mu)}{\epsilon-\mu+\omega},\qquad B_+(\omega)\to B_+^0(\omega)\equiv\mathcal{P}\int d\epsilon\,\frac{J(\epsilon)\Theta(\mu-\epsilon)}{\epsilon-\mu+\omega}. \tag*{(108)} B − ( ω ) → B − 0 ( ω ) ≡ P ∫ d ϵ ϵ − μ + ω J ( ϵ ) Θ ( ϵ − μ ) , B + ( ω ) → B + 0 ( ω ) ≡ P ∫ d ϵ ϵ − μ + ω J ( ϵ ) Θ ( μ − ϵ ) . (108)
Because B ± B_\pm B ± and B ± ′ B'_\pm B ± ′ stay bounded, every term of Eq. (107) without an explicit factor of β \beta β may be evaluated by these substitutions. Because [ K S , X ω † X ω ] = 0 [K_S,X^\dagger_\omega X_\omega]=0 [ K S , X ω † X ω ] = 0 and ∣ g ⟩ |g\rangle ∣ g ⟩ is non-degenerate, X ω † X ω ∣ g ⟩ ∝ ∣ g ⟩ X^\dagger_\omega X_\omega|g\rangle\propto|g\rangle X ω † X ω ∣ g ⟩ ∝ ∣ g ⟩ . This gives
∣ g ⟩ ⟨ g ∣ ( X ω † X ω − ⟨ g ∣ X ω † X ω ∣ g ⟩ ) = 0 , (109) |g\rangle\langle g|\left(X^\dagger_\omega X_\omega-\langle g|X^\dagger_\omega X_\omega|g\rangle\right)=0, \tag*{(109)} ∣ g ⟩ ⟨ g ∣ ( X ω † X ω − ⟨ g ∣ X ω † X ω ∣ g ⟩ ) = 0 , (109)
and likewise for X ω X ω † X_\omega X^\dagger_\omega X ω X ω † . The first line of Eq. (107) therefore vanishes at T = 0 T=0 T = 0 . With τ S → ∣ g ⟩ ⟨ g ∣ \tau_S\to|g\rangle\langle g| τ S → ∣ g ⟩ ⟨ g ∣ , each remaining term is non-zero only for particular signs of ω \omega ω . The B − B_- B − and B − ′ B'_- B − ′ terms contain ⟨ g ∣ X ω † \langle g|X^\dagger_\omega ⟨ g ∣ X ω † or X ω ∣ g ⟩ X_\omega|g\rangle X ω ∣ g ⟩ , so they are non-zero only if X ω ∣ g ⟩ ≠ 0 X_\omega|g\rangle\ne0 X ω ∣ g ⟩ = 0 . That requires ω = Ω m > 0 \omega=\Omega_m>0 ω = Ω m > 0 for some state m m m . The pole of B − 0 ( Ω m ) B_-^0(\Omega_m) B − 0 ( Ω m ) at ϵ = μ − Ω m < μ \epsilon=\mu-\Omega_m<\mu ϵ = μ − Ω m < μ then lies outside the support of Θ ( ϵ − μ ) \Theta(\epsilon-\mu) Θ ( ϵ − μ ) . Similarly, the B + B_+ B + and B + ′ B'_+ B + ′ terms require X ω † ∣ g ⟩ ≠ 0 X^\dagger_\omega|g\rangle\ne0 X ω † ∣ g ⟩ = 0 , i.e. ω = − Ω m < 0 \omega=-\Omega_m<0 ω = − Ω m < 0 . The pole at ϵ = μ + Ω m > μ \epsilon=\mu+\Omega_m>\mu ϵ = μ + Ω m > μ lies outside the support of Θ ( μ − ϵ ) \Theta(\mu-\epsilon) Θ ( μ − ϵ ) . Every integral that survives is therefore an ordinary, non-singular integral, and no principal value is needed. With τ S → ∣ g ⟩ ⟨ g ∣ \tau_S\to|g\rangle\langle g| τ S → ∣ g ⟩ ⟨ g ∣ and B ± → B ± 0 B_\pm\to B_\pm^0 B ± → B ± 0 ,
lim β → ∞ τ M F ( 2 ) = ∣ g ⟩ ⟨ g ∣ + λ 2 { ∑ ω [ B − 0 ( ω ) [ ∣ g ⟩ ⟨ g ∣ X ω † , X ω ] + B + 0 ( ω ) [ ∣ g ⟩ ⟨ g ∣ X ω , X ω † ] ] + ∑ ω ≠ ω ′ 1 ω ′ − ω [ B − 0 ( ω ) ( [ X ω ′ , ∣ g ⟩ ⟨ g ∣ X ω † ] + H . c . ) + B + 0 ( ω ) ( [ X ω ′ † , ∣ g ⟩ ⟨ g ∣ X ω ] + H . c . ) ] } . (110) \begin{aligned}
\lim_{\beta\to\infty}\tau_{\mathrm{MF}}^{(2)}
={}&|g\rangle\langle g|+\lambda^2\Bigg\{\sum_\omega\left[B_-^0(\omega)\left[|g\rangle\langle g|X^\dagger_\omega,X_\omega\right]+B_+^0(\omega)\left[|g\rangle\langle g|X_\omega,X^\dagger_\omega\right]\right]\\
&+\sum_{\omega\ne\omega'}\frac{1}{\omega'-\omega}\Big[B_-^0(\omega)\left([X_{\omega'},|g\rangle\langle g|X^\dagger_\omega]+\mathrm{H.c.}\right)+B_+^0(\omega)\left([X^\dagger_{\omega'},|g\rangle\langle g|X_\omega]+\mathrm{H.c.}\right)\Big]\Bigg\}. \tag*{(110)}
\end{aligned} β → ∞ lim τ MF ( 2 ) = ∣ g ⟩ ⟨ g ∣ + λ 2 { ω ∑ [ B − 0 ( ω ) [ ∣ g ⟩ ⟨ g ∣ X ω † , X ω ] + B + 0 ( ω ) [ ∣ g ⟩ ⟨ g ∣ X ω , X ω † ] ] + ω = ω ′ ∑ ω ′ − ω 1 [ B − 0 ( ω ) ( [ X ω ′ , ∣ g ⟩ ⟨ g ∣ X ω † ] + H.c. ) + B + 0 ( ω ) ( [ X ω ′ † , ∣ g ⟩ ⟨ g ∣ X ω ] + H.c. ) ] } . (110)
At T = 0 T=0 T = 0 the reservoir is a filled Fermi sea, and Eq. (110) is the reduced state of the ground state of H − μ N H-\mu N H − μ N to second order, namely ∣ g ⟩ |g\rangle ∣ g ⟩ dressed by virtual transfer of a particle into the empty states above μ \mu μ (B − 0 B_-^0 B − 0 ) or of a reservoir particle from below μ \mu μ into the system (B + 0 B_+^0 B + 0 ). One key feature of Eq. (110) is that coherences persist in this limit. Taking the commutator with H S H_S H S gives
lim β → ∞ [ τ M F ( 2 ) , H S ] = λ 2 ∑ ω ≠ ω ′ { B − 0 ( ω ) ( X ω ∣ g ⟩ ⟨ g ∣ X ω ′ † + ∣ g ⟩ ⟨ g ∣ X ω † X ω ′ ) + B + 0 ( ω ) ( X ω † ∣ g ⟩ ⟨ g ∣ X ω ′ + X ω ′ X ω † ∣ g ⟩ ⟨ g ∣ ) } − H . c . , (111) \begin{aligned}
\lim_{\beta\to\infty}[\tau_{\mathrm{MF}}^{(2)},H_S]
={}&\lambda^2\sum_{\omega\ne\omega'}\Big\{B_-^0(\omega)\left(X_\omega|g\rangle\langle g|X^\dagger_{\omega'}+|g\rangle\langle g|X^\dagger_\omega X_{\omega'}\right)\\
&+B_+^0(\omega)\left(X^\dagger_\omega|g\rangle\langle g|X_{\omega'}+X_{\omega'}X^\dagger_\omega|g\rangle\langle g|\right)\Big\}-\mathrm{H.c.}, \tag*{(111)}
\end{aligned} β → ∞ lim [ τ MF ( 2 ) , H S ] = λ 2 ω = ω ′ ∑ { B − 0 ( ω ) ( X ω ∣ g ⟩ ⟨ g ∣ X ω ′ † + ∣ g ⟩ ⟨ g ∣ X ω † X ω ′ ) + B + 0 ( ω ) ( X ω † ∣ g ⟩ ⟨ g ∣ X ω ′ + X ω ′ X ω † ∣ g ⟩ ⟨ g ∣ ) } − H.c. , (111)
which is generically non-zero.
Hamiltonian of mean force
The mean force Gibbs state is often expressed through the Hamiltonian of mean force (HMF) [12 , 17 ] , the temperature-dependent operator for which τ M F \tau_{\mathrm{MF}} τ MF takes a Gibbs form. Here we derive the HMF to second order in λ \lambda λ for the exchange coupling of Eq. (35) , following the strategy of Ref. [30 ] . We define the grand canonical HMF through
e − β K M F ≡ 1 Z R Tr R e − β ( H − μ N ) . (112) e^{-\beta K_{\mathrm{MF}}}\equiv\frac{1}{Z_R}\operatorname{Tr}_R e^{-\beta(H-\mu N)}. \tag*{(112)} e − β K MF ≡ Z R 1 Tr R e − β ( H − μ N ) . (112)
Dividing by the bare reservoir partition function means K M F → K S K_{\mathrm{MF}}\to K_S K MF → K S for V = 0 V=0 V = 0 [12 ] . Because [ H , N ] = 0 [H,N]=0 [ H , N ] = 0 , and the partial trace is cyclic for the even reservoir operator N R N_R N R , we have [ N S , Tr R e − β ( H − μ N ) ] = Tr R [ N , e − β ( H − μ N ) ] = 0 [N_S,\operatorname{Tr}_R e^{-\beta(H-\mu N)}]=\operatorname{Tr}_R[N,e^{-\beta(H-\mu N)}]=0 [ N S , Tr R e − β ( H − μ N ) ] = Tr R [ N , e − β ( H − μ N ) ] = 0 . Hence [ K M F , N S ] = 0 [K_{\mathrm{MF}},N_S]=0 [ K MF , N S ] = 0 , and H M F ≡ K M F + μ N S H_{\mathrm{MF}}\equiv K_{\mathrm{MF}}+\mu N_S H MF ≡ K MF + μ N S is well defined, with τ M F = e − β ( H M F − μ N S ) / Z M F \tau_{\mathrm{MF}}=e^{-\beta(H_{\mathrm{MF}}-\mu N_S)}/Z_{\mathrm{MF}} τ MF = e − β ( H MF − μ N S ) / Z MF . We seek K M F ( 2 ) = K S + λ 2 K ( 2 ) + O ( λ 4 ) K_{\mathrm{MF}}^{(2)}=K_S+\lambda^2K^{(2)}+O(\lambda^4) K MF ( 2 ) = K S + λ 2 K ( 2 ) + O ( λ 4 ) . To do this, we will compare the Kubo expansion of Eq. (112) in terms of this expansion in λ \lambda λ with the corresponding Kubo expansion of the un-normalised MFG state, Eq. (61) .
The Kubo expansion with K S K_S K S as reference gives
e − β ( K S + λ 2 K ( 2 ) ) = e − β K S [ 1 − λ 2 ∫ 0 β d β ′ e β ′ K S K ( 2 ) e − β ′ K S + O ( λ 4 ) ] . (113) e^{-\beta(K_S+\lambda^2K^{(2)})}=e^{-\beta K_S}\left[1-\lambda^2\int_0^\beta d\beta'\,e^{\beta'K_S}K^{(2)}e^{-\beta'K_S}+O(\lambda^4)\right]. \tag*{(113)} e − β ( K S + λ 2 K ( 2 ) ) = e − β K S [ 1 − λ 2 ∫ 0 β d β ′ e β ′ K S K ( 2 ) e − β ′ K S + O ( λ 4 ) ] . (113)
Decomposing K ( 2 ) K^{(2)} K ( 2 ) into frequency components K ( 2 ) = ∑ ν K ν ( 2 ) K^{(2)}=\sum_{\nu}K^{(2)}_{\nu} K ( 2 ) = ∑ ν K ν ( 2 ) where [ K S , K ν ( 2 ) ] = ν K ν ( 2 ) [K_{S},K^{(2)}_{\nu}]=\nu K^{(2)}_{\nu} [ K S , K ν ( 2 ) ] = ν K ν ( 2 ) and using Eq. (55) we can evaluate this integral, giving
e − β ( K S + λ 2 K ( 2 ) ) = e − β K S [ 1 − λ 2 [ β K ν = 0 ( 2 ) + ∑ ν ≠ 0 K ν ( 2 ) e β ν − 1 ν ] + O ( λ 4 ) ] (114) e^{-\beta(K_{S}+\lambda^{2}K^{(2)})}=e^{-\beta K_{S}}\left[1-\lambda^{2}\left[\beta K^{(2)}_{\nu=0}+\sum_{\nu\ne0}K^{(2)}_{\nu}\frac{e^{\beta\nu}-1}{\nu}\right]+\mathcal{O}(\lambda^{4})\right] \tag*{(114)} e − β ( K S + λ 2 K ( 2 ) ) = e − β K S 1 − λ 2 β K ν = 0 ( 2 ) + ν = 0 ∑ K ν ( 2 ) ν e β ν − 1 + O ( λ 4 ) (114)
where the ν = 0 \nu=0 ν = 0 case has been separated for ease of comparison with Eq. (61) , which we will now consider. Note that the operators in Eq. (61) are already eigenoperators of K S K_{S} K S : X ω † X ω ′ X^{\dagger}_{\omega}X_{\omega'} X ω † X ω ′ has eigenvalue ν = ω ′ − ω \nu=\omega'-\omega ν = ω ′ − ω and X ω X ω ′ † X_{\omega}X^{\dagger}_{\omega'} X ω X ω ′ † has eigenvalue ν = ω − ω ′ \nu=\omega-\omega' ν = ω − ω ′ . Equating the off-diagonal terms with ν = ω − ω ′ \nu=\omega-\omega' ν = ω − ω ′
K ν ( 2 ) = − X ω ′ † X ω ω − ω ′ e β ( ω − ω ′ ) − 1 e − β ( ω − ω ′ ) A 1 ( ω ) − A 1 ( ω ′ ) ω − ω ′ − X ω X ω ′ † ω − ω ′ e β ( ω − ω ′ ) − 1 e β ( ω − ω ′ ) A 2 ( ω ′ ) − A 2 ( ω ) ω − ω ′ = − X ω ′ † X ω e − β ( ω − ω ′ ) A 1 ( ω ) − A 1 ( ω ′ ) e β ( ω − ω ′ ) − 1 − X ω X ω ′ † e β ( ω − ω ′ ) A 2 ( ω ′ ) − A 2 ( ω ) e β ( ω − ω ′ ) − 1 = − X ω ′ † X ω A 2 ( ω ) − A 2 ( ω ′ ) e β ω − e β ω ′ − X ω X ω ′ † A 1 ( ω ′ ) − A 1 ( ω ) e − β ω ′ − e − β ω . (115) \begin{aligned}
K^{(2)}_{\nu}&=-X^{\dagger}_{\omega'}X_{\omega}\frac{\omega-\omega'}{e^{\beta(\omega-\omega')}-1}\frac{e^{-\beta(\omega-\omega')}A_{1}(\omega)-A_{1}(\omega')}{\omega-\omega'}-X_{\omega}X^{\dagger}_{\omega'}\frac{\omega-\omega'}{e^{\beta(\omega-\omega')}-1}\frac{e^{\beta(\omega-\omega')}A_{2}(\omega')-A_{2}(\omega)}{\omega-\omega'}\\
&=-X^{\dagger}_{\omega'}X_{\omega}\frac{e^{-\beta(\omega-\omega')}A_{1}(\omega)-A_{1}(\omega')}{e^{\beta(\omega-\omega')}-1}-X_{\omega}X^{\dagger}_{\omega'}\frac{e^{\beta(\omega-\omega')}A_{2}(\omega')-A_{2}(\omega)}{e^{\beta(\omega-\omega')}-1}\\
&=-X^{\dagger}_{\omega'}X_{\omega}\frac{A_{2}(\omega)-A_{2}(\omega')}{e^{\beta\omega}-e^{\beta\omega'}}-X_{\omega}X^{\dagger}_{\omega'}\frac{A_{1}(\omega')-A_{1}(\omega)}{e^{-\beta\omega'}-e^{-\beta\omega}}. \tag*{(115)}
\end{aligned} K ν ( 2 ) = − X ω ′ † X ω e β ( ω − ω ′ ) − 1 ω − ω ′ ω − ω ′ e − β ( ω − ω ′ ) A 1 ( ω ) − A 1 ( ω ′ ) − X ω X ω ′ † e β ( ω − ω ′ ) − 1 ω − ω ′ ω − ω ′ e β ( ω − ω ′ ) A 2 ( ω ′ ) − A 2 ( ω ) = − X ω ′ † X ω e β ( ω − ω ′ ) − 1 e − β ( ω − ω ′ ) A 1 ( ω ) − A 1 ( ω ′ ) − X ω X ω ′ † e β ( ω − ω ′ ) − 1 e β ( ω − ω ′ ) A 2 ( ω ′ ) − A 2 ( ω ) = − X ω ′ † X ω e β ω − e β ω ′ A 2 ( ω ) − A 2 ( ω ′ ) − X ω X ω ′ † e − β ω ′ − e − β ω A 1 ( ω ′ ) − A 1 ( ω ) . (115)
where in the last line the KMS relation A 1 ( ω ) = e − β ω A 2 ( ω ) A_{1}(\omega)=e^{-\beta\omega}A_{2}(\omega) A 1 ( ω ) = e − β ω A 2 ( ω ) was used and both numerators and denominators were divided by e − β ω ′ e^{-\beta\omega'} e − β ω ′ and e β ω e^{\beta\omega} e β ω for the first and second terms respectively. For the diagonal component, we have
K ν = 0 ( 2 ) = − 1 β ∑ ω [ D 1 ( ω ) X ω † X ω + D 2 ( ω ) X ω X ω † ] . (116) K^{(2)}_{\nu=0}=-\frac{1}{\beta}\sum_{\omega}\left[D_{1}(\omega)X^{\dagger}_{\omega}X_{\omega}+D_{2}(\omega)X_{\omega}X^{\dagger}_{\omega}\right]. \tag*{(116)} K ν = 0 ( 2 ) = − β 1 ω ∑ [ D 1 ( ω ) X ω † X ω + D 2 ( ω ) X ω X ω † ] . (116)
Putting this together, the HMF to second order is given by
H M F ( 2 ) = H S + λ 2 K ( 2 ) + O ( λ 4 ) , H_{\mathrm{MF}}^{(2)}=H_{S}+\lambda^{2}K^{(2)}+\mathcal{O}(\lambda^{4}), H MF ( 2 ) = H S + λ 2 K ( 2 ) + O ( λ 4 ) ,
K ( 2 ) = − 1 β ∑ ω [ D 1 ( ω ) X ω † X ω + D 2 ( ω ) X ω X ω † ] − ∑ ω ≠ ω ′ [ A 2 ( ω ′ ) − A 2 ( ω ) e β ω ′ − e β ω X ω † X ω ′ + A 1 ( ω ′ ) − A 1 ( ω ) e − β ω ′ − e − β ω X ω X ω ′ † ] (117) K^{(2)}=-\frac{1}{\beta}\sum_{\omega}\left[D_{1}(\omega)X^{\dagger}_{\omega}X_{\omega}+D_{2}(\omega)X_{\omega}X^{\dagger}_{\omega}\right]-\sum_{\omega\ne\omega'}\left[\frac{A_{2}(\omega')-A_{2}(\omega)}{e^{\beta\omega'}-e^{\beta\omega}}X^{\dagger}_{\omega}X_{\omega'}+\frac{A_{1}(\omega')-A_{1}(\omega)}{e^{-\beta\omega'}-e^{-\beta\omega}}X_{\omega}X^{\dagger}_{\omega'}\right] \tag*{(117)} K ( 2 ) = − β 1 ω ∑ [ D 1 ( ω ) X ω † X ω + D 2 ( ω ) X ω X ω † ] − ω = ω ′ ∑ [ e β ω ′ − e β ω A 2 ( ω ′ ) − A 2 ( ω ) X ω † X ω ′ + e − β ω ′ − e − β ω A 1 ( ω ′ ) − A 1 ( ω ) X ω X ω ′ † ] (117)
where H M F ( 2 ) = K M F ( 2 ) + μ N S = H S + λ 2 K ( 2 ) H_{\mathrm{MF}}^{(2)}=K_{\mathrm{MF}}^{(2)}+\mu N_{S}=H_{S}+\lambda^{2}K^{(2)} H MF ( 2 ) = K MF ( 2 ) + μ N S = H S + λ 2 K ( 2 ) because K ( 2 ) K^{(2)} K ( 2 ) conserves N S N_{S} N S . However, this expression is more restrictive than Eq. (74) . Consider a non-degenerate eigenstate ∣ n ⟩ \lvert n\rangle ∣ n ⟩ of H S H_{S} H S with energy E n E_{n} E n , particle number N n N_{n} N n and bare population p n = e − β ( E n − μ N n ) / Z S p_{n}=e^{-\beta(E_{n}-\mu N_{n})}/Z_{S} p n = e − β ( E n − μ N n ) / Z S , and suppose for simplicity that τ M F ( 2 ) \tau_{\mathrm{MF}}^{(2)} τ MF ( 2 ) is diagonal in this basis, as for the RLM and the SIAM. The population of ∣ n ⟩ \lvert n\rangle ∣ n ⟩ in τ M F ( 2 ) \tau_{\mathrm{MF}}^{(2)} τ MF ( 2 ) , Eq. (74) , is then
⟨ n ∣ τ M F ( 2 ) ∣ n ⟩ = p n + λ 2 Δ n n = p n ( 1 + x n ) , x n = λ 2 Δ n n p n , (118) \langle n\lvert\tau_{\mathrm{MF}}^{(2)}\rvert n\rangle=p_{n}+\lambda^{2}\Delta_{nn}=p_{n}(1+x_{n}),\qquad x_{n}=\frac{\lambda^{2}\Delta_{nn}}{p_{n}}, \tag*{(118)} ⟨ n ∣ τ MF ( 2 ) ∣ n ⟩ = p n + λ 2 Δ nn = p n ( 1 + x n ) , x n = p n λ 2 Δ nn , (118)
where Δ n n = ⟨ n ∣ Δ d i a g ∣ n ⟩ \Delta_{nn}=\langle n\lvert\Delta_{\mathrm{diag}}\rvert n\rangle Δ nn = ⟨ n ∣ Δ diag ∣ n ⟩ . By definition, H M F − μ N S = − T ln τ M F + c o n s t H_{\mathrm{MF}}-\mu N_{S}=-T\ln\tau_{\mathrm{MF}}+\mathrm{const} H MF − μ N S = − T ln τ MF + const . Applying this to τ M F ( 2 ) \tau_{\mathrm{MF}}^{(2)} τ MF ( 2 ) , and using ⟨ n ∣ ln τ M F ( 2 ) ∣ n ⟩ = ln ⟨ n ∣ τ M F ( 2 ) ∣ n ⟩ \langle n\lvert\ln\tau_{\mathrm{MF}}^{(2)}\rvert n\rangle=\ln\langle n\lvert\tau_{\mathrm{MF}}^{(2)}\rvert n\rangle ⟨ n ∣ ln τ MF ( 2 ) ∣ n ⟩ = ln ⟨ n ∣ τ MF ( 2 ) ∣ n ⟩ since τ M F ( 2 ) \tau_{\mathrm{MF}}^{(2)} τ MF ( 2 ) is diagonal, the Hamiltonian of mean force that reproduces τ M F ( 2 ) \tau_{\mathrm{MF}}^{(2)} τ MF ( 2 ) exactly has diagonal elements
− T ln ⟨ n ∣ τ M F ( 2 ) ∣ n ⟩ = E n − μ N n − T ln ( 1 + x n ) + c o n s t . (119) -T\ln\langle n\lvert\tau_{\mathrm{MF}}^{(2)}\rvert n\rangle=E_{n}-\mu N_{n}-T\ln(1+x_{n})+\mathrm{const}. \tag*{(119)} − T ln ⟨ n ∣ τ MF ( 2 ) ∣ n ⟩ = E n − μ N n − T ln ( 1 + x n ) + const . (119)
H M F ( 2 ) = H S + λ 2 K ( 2 ) H_{\mathrm{MF}}^{(2)}=H_{S}+\lambda^{2}K^{(2)} H MF ( 2 ) = H S + λ 2 K ( 2 ) of Eq. (117) has diagonal elements
⟨ n ∣ H M F ( 2 ) − μ N S ∣ n ⟩ = E n − μ N n − T x n + c o n s t . (120) \langle n\lvert H_{\mathrm{MF}}^{(2)}-\mu N_{S}\rvert n\rangle=E_{n}-\mu N_{n}-Tx_{n}+\mathrm{const}. \tag*{(120)} ⟨ n ∣ H MF ( 2 ) − μ N S ∣ n ⟩ = E n − μ N n − T x n + const . (120)
Here const denotes terms independent of n n n , which only affect the normalisation. The grand canonical Gibbs state of H M F ( 2 ) H_{\mathrm{MF}}^{(2)} H MF ( 2 ) therefore reproduces τ M F ( 2 ) \tau_{\mathrm{MF}}^{(2)} τ MF ( 2 ) only if ln ( 1 + x n ) ≃ x n \ln(1+x_{n})\simeq x_{n} ln ( 1 + x n ) ≃ x n , i.e. ∣ x n ∣ ≪ 1 \lvert x_{n}\rvert\ll1 ∣ x n ∣ ≪ 1 . At low temperature this fails for excited states, their bare population p n p_{n} p n vanishes exponentially while the mean force correction Δ n n \Delta_{nn} Δ nn remains finite as T → 0 T\to0 T → 0 (See Sec. SIII), so x n x_{n} x n grows exponentially.
Applications
Resonant level model
The simplest application is the resonant level model (RLM), a single fermionic mode coupled to a continuous non-interacting reservoir,
H = ϵ s s † s + ∫ d ϵ ϵ a ϵ † a ϵ + λ ∫ d ϵ J ( ϵ ) ( s † a ϵ + a ϵ † s ) , (121) H=\epsilon_s s^\dagger s+\int d\epsilon\,\epsilon a_\epsilon^\dagger a_\epsilon+\lambda\int d\epsilon\,\sqrt{J(\epsilon)}(s^\dagger a_\epsilon+a_\epsilon^\dagger s), \tag*{(121)} H = ϵ s s † s + ∫ d ϵ ϵ a ϵ † a ϵ + λ ∫ d ϵ J ( ϵ ) ( s † a ϵ + a ϵ † s ) , (121)
so that X = s X=s X = s and E = ∫ J ( ϵ ) a ϵ d ϵ E=\int\sqrt{J(\epsilon)}a_\epsilon\,d\epsilon E = ∫ J ( ϵ ) a ϵ d ϵ . Here X X X is already an eigenoperator,
[ K S , s ] = − ( ϵ s − μ ) s ⟹ X = X ω 0 = s , ω 0 = μ − ϵ s , (122) [K_S,s]=-(\epsilon_s-\mu)s\quad\Longrightarrow\quad X=X_{\omega_0}=s,\qquad\omega_0=\mu-\epsilon_s, \tag*{(122)} [ K S , s ] = − ( ϵ s − μ ) s ⟹ X = X ω 0 = s , ω 0 = μ − ϵ s , (122)
this being the only non-zero Bohr-frequency component, so there are no off-diagonal contributions. With X ω 0 † X ω 0 = s † s ≡ n X_{\omega_0}^\dagger X_{\omega_0}=s^\dagger s\equiv n X ω 0 † X ω 0 = s † s ≡ n , X ω 0 X ω 0 † = s s † = 1 − n X_{\omega_0}X_{\omega_0}^\dagger=ss^\dagger=1-n X ω 0 X ω 0 † = s s † = 1 − n and τ S = diag ( 1 − f s , f s ) \tau_S=\operatorname{diag}(1-f_s,f_s) τ S = diag ( 1 − f s , f s ) with f s ≡ f ( ϵ s − μ ) f_s\equiv f(\epsilon_s-\mu) f s ≡ f ( ϵ s − μ ) , Eq. (S41) gives
τ M F ( 2 ) = τ S + λ 2 [ τ S ( n − Tr [ τ S n ] ) D 1 ( ω 0 ) + τ S ( ( 1 − n ) − Tr [ τ S ( 1 − n ) ] ) D 2 ( ω 0 ) ] . (123) \tau_{\mathrm{MF}}^{(2)}=\tau_S+\lambda^2\left[\tau_S\left(n-\operatorname{Tr}[\tau_S n]\right)D_1(\omega_0)+\tau_S\left((1-n)-\operatorname{Tr}[\tau_S(1-n)]\right)D_2(\omega_0)\right]. \tag*{(123)} τ MF ( 2 ) = τ S + λ 2 [ τ S ( n − Tr [ τ S n ] ) D 1 ( ω 0 ) + τ S ( ( 1 − n ) − Tr [ τ S ( 1 − n )] ) D 2 ( ω 0 ) ] . (123)
Since n − ⟨ n ⟩ τ S n-\langle n\rangle_{\tau_S} n − ⟨ n ⟩ τ S and ( 1 − n ) − ⟨ 1 − n ⟩ τ S (1-n)-\langle1-n\rangle_{\tau_S} ( 1 − n ) − ⟨ 1 − n ⟩ τ S are negatives of one another, there is a single correction to the occupation,
⟨ n ⟩ M F ( 2 ) = f s + λ 2 f s ( 1 − f s ) [ D 1 ( ω 0 ) − D 2 ( ω 0 ) ] . (124) \langle n\rangle_{\mathrm{MF}}^{(2)}=f_s+\lambda^2 f_s(1-f_s)\left[D_1(\omega_0)-D_2(\omega_0)\right]. \tag*{(124)} ⟨ n ⟩ MF ( 2 ) = f s + λ 2 f s ( 1 − f s ) [ D 1 ( ω 0 ) − D 2 ( ω 0 ) ] . (124)
To evaluate D 1 − D 2 D_1-D_2 D 1 − D 2 , it is easier to work with the bounded form Eq. (S69). With ω 0 = μ − ϵ s \omega_0=\mu-\epsilon_s ω 0 = μ − ϵ s the denominators in Eq. (S67) become ϵ − ϵ s \epsilon-\epsilon_s ϵ − ϵ s , so
B − ( ω 0 ) + B + ( ω 0 ) = P ∫ d ϵ J ( ϵ ) ϵ − ϵ s , f s B − ′ ( ω 0 ) − ( 1 − f s ) B + ′ ( ω 0 ) = − P ∫ d ϵ J ( ϵ ) f s − f ( ϵ − μ ) ( ϵ − ϵ s ) 2 , (125) B_-(\omega_0)+B_+(\omega_0)=\mathcal{P}\int d\epsilon\,\frac{J(\epsilon)}{\epsilon-\epsilon_s},\qquad f_s B_-'(\omega_0)-(1-f_s)B_+'(\omega_0)=-\mathcal{P}\int d\epsilon\,J(\epsilon)\frac{f_s-f(\epsilon-\mu)}{(\epsilon-\epsilon_s)^2}, \tag*{(125)} B − ( ω 0 ) + B + ( ω 0 ) = P ∫ d ϵ ϵ − ϵ s J ( ϵ ) , f s B − ′ ( ω 0 ) − ( 1 − f s ) B + ′ ( ω 0 ) = − P ∫ d ϵ J ( ϵ ) ( ϵ − ϵ s ) 2 f s − f ( ϵ − μ ) , (125)
where we used f s [ 1 − f ( ϵ − μ ) ] − ( 1 − f s ) f ( ϵ − μ ) = f s − f ( ϵ − μ ) f_s[1-f(\epsilon-\mu)]-(1-f_s)f(\epsilon-\mu)=f_s-f(\epsilon-\mu) f s [ 1 − f ( ϵ − μ )] − ( 1 − f s ) f ( ϵ − μ ) = f s − f ( ϵ − μ ) . The correction to the occupation is therefore
δ ⟨ n ⟩ = λ 2 ∫ d ϵ J ( ϵ ) [ β f s ( 1 − f s ) ϵ − ϵ s − f s − f ( ϵ − μ ) ( ϵ − ϵ s ) 2 ] (126) \delta\langle n\rangle=\lambda^2\int d\epsilon\,J(\epsilon)\left[\frac{\beta f_s(1-f_s)}{\epsilon-\epsilon_s}-\frac{f_s-f(\epsilon-\mu)}{(\epsilon-\epsilon_s)^2}\right] \tag*{(126)} δ ⟨ n ⟩ = λ 2 ∫ d ϵ J ( ϵ ) [ ϵ − ϵ s β f s ( 1 − f s ) − ( ϵ − ϵ s ) 2 f s − f ( ϵ − μ ) ] (126)
The integrand is regular at ϵ = ϵ s \epsilon=\epsilon_s ϵ = ϵ s : expanding f ( ϵ − μ ) = f s + f ′ ( ϵ s − μ ) ( ϵ − ϵ s ) + 1 2 f ′ ′ ( ϵ s − μ ) ( ϵ − ϵ s ) 2 + … f(\epsilon-\mu)=f_s+f'(\epsilon_s-\mu)(\epsilon-\epsilon_s)+\frac{1}{2}f''(\epsilon_s-\mu)(\epsilon-\epsilon_s)^2+\ldots f ( ϵ − μ ) = f s + f ′ ( ϵ s − μ ) ( ϵ − ϵ s ) + 2 1 f ′′ ( ϵ s − μ ) ( ϵ − ϵ s ) 2 + … and using β f s ( 1 − f s ) = − f ′ ( ϵ s − μ ) \beta f_s(1-f_s)=-f'(\epsilon_s-\mu) β f s ( 1 − f s ) = − f ′ ( ϵ s − μ ) , the two divergent pieces cancel and the integrand tends to 1 2 J ( ϵ s ) f ′ ′ ( ϵ s − μ ) \frac{1}{2}J(\epsilon_s)f''(\epsilon_s-\mu) 2 1 J ( ϵ s ) f ′′ ( ϵ s − μ ) . We define the level shift
Λ ( ϵ s ) ≡ λ 2 P ∫ d ϵ J ( ϵ ) ϵ s − ϵ , (127) \Lambda(\epsilon_s)\equiv\lambda^2\mathcal{P}\int d\epsilon\,\frac{J(\epsilon)}{\epsilon_s-\epsilon}, \tag*{(127)} Λ ( ϵ s ) ≡ λ 2 P ∫ d ϵ ϵ s − ϵ J ( ϵ ) , (127)
which is the real part of the retarded self-energy of the level evaluated at the bare pole, familiar from the Green’s function treatment of the RLM [54 , 55 ] . Then
δ ⟨ n ⟩ = ∂ f s ∂ ϵ s Λ ( ϵ s ) − λ 2 P ∫ d ϵ J ( ϵ ) f s − f ( ϵ − μ ) ( ϵ − ϵ s ) 2 . (128) \delta\langle n\rangle=\frac{\partial f_s}{\partial\epsilon_s}\Lambda(\epsilon_s)-\lambda^2\mathcal{P}\int d\epsilon\,J(\epsilon)\frac{f_s-f(\epsilon-\mu)}{(\epsilon-\epsilon_s)^2}. \tag*{(128)} δ ⟨ n ⟩ = ∂ ϵ s ∂ f s Λ ( ϵ s ) − λ 2 P ∫ d ϵ J ( ϵ ) ( ϵ − ϵ s ) 2 f s − f ( ϵ − μ ) . (128)
The first term is the naive expectation, the hybridisation renormalises the level by Λ \Lambda Λ , and the occupation responds with the Fermi-function susceptibility. The second term is not captured by any level shift, it represents the broadening of the level, sampling f f f away from ϵ s \epsilon_s ϵ s . We can recast this in terms of the Hamiltonian of mean force, where Eq. (S84) reduces to
H M F = ϵ ~ s n − λ 2 β D 2 ( ω 0 ) , ϵ ~ s = ϵ s − λ 2 β − 1 [ D 1 ( ω 0 ) − D 2 ( ω 0 ) ] . (129) H_{\mathrm{MF}}=\widetilde{\epsilon}_s n-\frac{\lambda^2}{\beta}D_2(\omega_0),\qquad\widetilde{\epsilon}_s=\epsilon_s-\lambda^2\beta^{-1}\left[D_1(\omega_0)-D_2(\omega_0)\right]. \tag*{(129)} H MF = ϵ s n − β λ 2 D 2 ( ω 0 ) , ϵ s = ϵ s − λ 2 β − 1 [ D 1 ( ω 0 ) − D 2 ( ω 0 ) ] . (129)
Two independent modes coupled to a shared reservoir
We now consider two modes that do not interact directly but share a reservoir. This is the minimal setting in which the reservoir can mediate correlations, and it is the exchange-coupled analogue of the multi-qubit spin–boson problem studied in Refs. [26 , 32 ] . Take
H S = ϵ 1 n 1 + ϵ 2 n 2 , n i = s i † s i , (130) H_S=\epsilon_1 n_1+\epsilon_2 n_2,\qquad n_i=s_i^\dagger s_i, \tag*{(130)} H S = ϵ 1 n 1 + ϵ 2 n 2 , n i = s i † s i , (130)
with no direct hopping between the modes, coupled to a single reservoir through the symmetric combination
X = s 1 + s 2 . (131) X=s_1+s_2. \tag*{(131)} X = s 1 + s 2 . (131)
The reservoir Hamiltonian and E E E are as in Sec. SV A. Since [ K S , s i ] = − ( ε i − μ ) s i [K_S,s_i]=-(\varepsilon_i-\mu)s_i [ K S , s i ] = − ( ε i − μ ) s i , each mode is already a Bohr-frequency eigenoperator:
X ω 1 = s 1 , ω 1 = μ − ε 1 , X ω 2 = s 2 , ω 2 = μ − ε 2 . (132) X_{\omega_1}=s_1,\quad\omega_1=\mu-\varepsilon_1,\qquad X_{\omega_2}=s_2,\quad\omega_2=\mu-\varepsilon_2. \tag*{(132)} X ω 1 = s 1 , ω 1 = μ − ε 1 , X ω 2 = s 2 , ω 2 = μ − ε 2 . (132)
No basis rotation is required, so X = X ω 1 + X ω 2 X=X_{\omega_1}+X_{\omega_2} X = X ω 1 + X ω 2 with exactly two Bohr frequencies. The diagonal part is
Δ d i a g = τ S ( n 1 − ⟨ n 1 ⟩ τ S ) D 1 ( ω 1 ) + τ S ( ( 1 − n 1 ) − ⟨ 1 − n 1 ⟩ τ S ) D 2 ( ω 1 ) + τ S ( n 2 − ⟨ n 2 ⟩ τ S ) D 1 ( ω 2 ) + τ S ( ( 1 − n 2 ) − ⟨ 1 − n 2 ⟩ τ S ) D 2 ( ω 2 ) , (133) \Delta_{\mathrm{diag}}=\tau_S(n_1-\langle n_1\rangle_{\tau_S})D_1(\omega_1)+\tau_S((1-n_1)-\langle1-n_1\rangle_{\tau_S})D_2(\omega_1)
+\tau_S(n_2-\langle n_2\rangle_{\tau_S})D_1(\omega_2)+\tau_S((1-n_2)-\langle1-n_2\rangle_{\tau_S})D_2(\omega_2), \tag*{(133)} Δ diag = τ S ( n 1 − ⟨ n 1 ⟩ τ S ) D 1 ( ω 1 ) + τ S (( 1 − n 1 ) − ⟨ 1 − n 1 ⟩ τ S ) D 2 ( ω 1 ) + τ S ( n 2 − ⟨ n 2 ⟩ τ S ) D 1 ( ω 2 ) + τ S (( 1 − n 2 ) − ⟨ 1 − n 2 ⟩ τ S ) D 2 ( ω 2 ) , (133)
and the off-diagonal part is
Δ o f f = 1 ε 1 − ε 2 [ ( s 2 s 1 † τ S + τ S s 1 s 2 † ) A 2 ( ω 1 ) − ( s 2 † s 1 τ S + τ S s 1 † s 2 ) A 1 ( ω 1 ) − ( s 1 s 2 † τ S + τ S s 2 s 1 † ) A 2 ( ω 2 ) + ( s 1 † s 2 τ S + τ S s 2 † s 1 ) A 1 ( ω 2 ) ] . (134) \Delta_{\mathrm{off}}=\frac{1}{\varepsilon_1-\varepsilon_2}\left[(s_2s_1^\dagger\tau_S+\tau_Ss_1s_2^\dagger)A_2(\omega_1)-(s_2^\dagger s_1\tau_S+\tau_Ss_1^\dagger s_2)A_1(\omega_1)
-(s_1s_2^\dagger\tau_S+\tau_Ss_2s_1^\dagger)A_2(\omega_2)+(s_1^\dagger s_2\tau_S+\tau_Ss_2^\dagger s_1)A_1(\omega_2)\right]. \tag*{(134)} Δ off = ε 1 − ε 2 1 [ ( s 2 s 1 † τ S + τ S s 1 s 2 † ) A 2 ( ω 1 ) − ( s 2 † s 1 τ S + τ S s 1 † s 2 ) A 1 ( ω 1 ) − ( s 1 s 2 † τ S + τ S s 2 s 1 † ) A 2 ( ω 2 ) + ( s 1 † s 2 τ S + τ S s 2 † s 1 ) A 1 ( ω 2 ) ] . (134)
The two modes are independent under τ S \tau_S τ S , which factorises into single-mode fermionic factors,
τ S = ∏ i = 1 , 2 [ ( 1 − f i ) ( 1 − n i ) + f i n i ] , f i ≡ f ( ε i − μ ) . (135) \tau_S=\prod_{i=1,2}\left[(1-f_i)(1-n_i)+f_i n_i\right],\qquad f_i\equiv f(\varepsilon_i-\mu). \tag*{(135)} τ S = i = 1 , 2 ∏ [ ( 1 − f i ) ( 1 − n i ) + f i n i ] , f i ≡ f ( ε i − μ ) . (135)
Labelling the Fock states by their occupations ν i ∈ { 0 , 1 } \nu_i\in\{0,1\} ν i ∈ { 0 , 1 } , the projector onto ∣ ν 1 ν 2 ⟩ |\nu_1\nu_2\rangle ∣ ν 1 ν 2 ⟩ is P ν = ∏ i [ ν i n i + ( 1 − ν i ) ( 1 − n i ) ] P_\nu=\prod_i[\nu_i n_i+(1-\nu_i)(1-n_i)] P ν = ∏ i [ ν i n i + ( 1 − ν i ) ( 1 − n i )] , and the bare populations are p ν ( 0 ) = ⟨ P ν ⟩ τ S = ∏ i f i ν i ( 1 − f i ) 1 − ν i p_\nu^{(0)}=\langle P_\nu\rangle_{\tau_S}=\prod_i f_i^{\nu_i}(1-f_i)^{1-\nu_i} p ν ( 0 ) = ⟨ P ν ⟩ τ S = ∏ i f i ν i ( 1 − f i ) 1 − ν i . Using ( 1 − n i ) − ⟨ 1 − n i ⟩ τ S = − ( n i − ⟨ n i ⟩ τ S ) (1-n_i)-\langle1-n_i\rangle_{\tau_S}=-(n_i-\langle n_i\rangle_{\tau_S}) ( 1 − n i ) − ⟨ 1 − n i ⟩ τ S = − ( n i − ⟨ n i ⟩ τ S ) and defining D i ≡ D 2 ( ω i ) − D 1 ( ω i ) \mathcal{D}_i\equiv D_2(\omega_i)-D_1(\omega_i) D i ≡ D 2 ( ω i ) − D 1 ( ω i ) we have Δ d i a g = − ∑ i τ S ( n i − f i ) D i \Delta_{\mathrm{diag}}=-\sum_i\tau_S(n_i-f_i)\mathcal{D}_i Δ diag = − ∑ i τ S ( n i − f i ) D i . Since P ν P_\nu P ν , τ S \tau_S τ S and n i n_i n i are all functions of the number operators they commute, so that Tr S [ P ν τ S ( n i − f i ) ] = p ν ( 0 ) ( ν i − f i ) \operatorname{Tr}_S[P_\nu\tau_S(n_i-f_i)]=p_\nu^{(0)}(\nu_i-f_i) Tr S [ P ν τ S ( n i − f i )] = p ν ( 0 ) ( ν i − f i ) , and the second-order population corrections δ p ν \delta p_\nu δ p ν , defined by p ν = p ν ( 0 ) + δ p ν p_\nu=p_\nu^{(0)}+\delta p_\nu p ν = p ν ( 0 ) + δ p ν , are
δ p 00 = λ 2 ( 1 − f 1 ) ( 1 − f 2 ) [ f 1 D 1 + f 2 D 2 ] , (136) \delta p_{00}=\lambda^2(1-f_1)(1-f_2)[f_1D_1+f_2D_2], \tag*{(136)} δ p 00 = λ 2 ( 1 − f 1 ) ( 1 − f 2 ) [ f 1 D 1 + f 2 D 2 ] , (136)
δ p 01 = λ 2 f 2 ( 1 − f 1 ) [ f 1 D 1 − ( 1 − f 2 ) D 2 ] , (137) \delta p_{01}=\lambda^2f_2(1-f_1)[f_1D_1-(1-f_2)D_2], \tag*{(137)} δ p 01 = λ 2 f 2 ( 1 − f 1 ) [ f 1 D 1 − ( 1 − f 2 ) D 2 ] , (137)
δ p 10 = λ 2 f 1 ( 1 − f 2 ) [ f 2 D 2 − ( 1 − f 1 ) D 1 ] , (138) \delta p_{10}=\lambda^2f_1(1-f_2)[f_2D_2-(1-f_1)D_1], \tag*{(138)} δ p 10 = λ 2 f 1 ( 1 − f 2 ) [ f 2 D 2 − ( 1 − f 1 ) D 1 ] , (138)
δ p 11 = − λ 2 f 1 f 2 [ ( 1 − f 1 ) D 1 + ( 1 − f 2 ) D 2 ] , (139) \delta p_{11}=-\lambda^2f_1f_2[(1-f_1)D_1+(1-f_2)D_2], \tag*{(139)} δ p 11 = − λ 2 f 1 f 2 [( 1 − f 1 ) D 1 + ( 1 − f 2 ) D 2 ] , (139)
which sum to zero as required by trace preservation. Summing δ p 10 + δ p 11 \delta p_{10}+\delta p_{11} δ p 10 + δ p 11 to obtain the correction to ⟨ n 1 ⟩ \langle n_1\rangle ⟨ n 1 ⟩ , the terms proportional to D 2 D_2 D 2 cancel, leaving
δ ⟨ n 1 ⟩ = − λ 2 f 1 ( 1 − f 1 ) D 1 = λ 2 f 1 ( 1 − f 1 ) [ D 1 ( ω 1 ) − D 2 ( ω 1 ) ] , (140) \delta\langle n_1\rangle=-\lambda^2f_1(1-f_1)D_1=\lambda^2f_1(1-f_1)[D_1(\omega_1)-D_2(\omega_1)], \tag*{(140)} δ ⟨ n 1 ⟩ = − λ 2 f 1 ( 1 − f 1 ) D 1 = λ 2 f 1 ( 1 − f 1 ) [ D 1 ( ω 1 ) − D 2 ( ω 1 )] , (140)
with no dependence on D 2 D_2 D 2 , and by the 1 ↔ 2 1\leftrightarrow2 1 ↔ 2 symmetry of the construction
⟨ n 1 ⟩ = f 1 + λ 2 f 1 ( 1 − f 1 ) [ D 1 ( ω 1 ) − D 2 ( ω 1 ) ] , ⟨ n 2 ⟩ = f 2 + λ 2 f 2 ( 1 − f 2 ) [ D 1 ( ω 2 ) − D 2 ( ω 2 ) ] . (141) \langle n_1\rangle=f_1+\lambda^2f_1(1-f_1)[D_1(\omega_1)-D_2(\omega_1)],\qquad\langle n_2\rangle=f_2+\lambda^2f_2(1-f_2)[D_1(\omega_2)-D_2(\omega_2)]. \tag*{(141)} ⟨ n 1 ⟩ = f 1 + λ 2 f 1 ( 1 − f 1 ) [ D 1 ( ω 1 ) − D 2 ( ω 1 )] , ⟨ n 2 ⟩ = f 2 + λ 2 f 2 ( 1 − f 2 ) [ D 1 ( ω 2 ) − D 2 ( ω 2 )] . (141)
This is exactly the single-mode RLM formula, Eq. (124) , applied independently to each mode. At this order each mode’s population is completely blind to the other mode’s presence, and to the fact that the reservoir is shared. The modes are not, however, blind to one another in every respect, as they would be in the bare Gibbs state, they acquire a coherence, ⟨ s 2 † s 1 ⟩ τ M F ( 2 ) \langle s_2^\dagger s_1\rangle_{\tau_{\mathrm{MF}}^{(2)}} ⟨ s 2 † s 1 ⟩ τ MF ( 2 ) . Now only Δ o f f \Delta_{\mathrm{off}} Δ off contributes and only the terms of Eq. (134) containing s 2 s 1 † s_2s_1^\dagger s 2 s 1 † or s 1 † s 2 s_1^\dagger s_2 s 1 † s 2 survive, the others contain s i s i s_is_i s i s i or s i † s i † s_i^\dagger s_i^\dagger s i † s i † and vanish. The surviving terms reduce by anticommutation to
s 2 † s 1 s 2 s 1 † = − n 2 ( 1 − n 1 ) , s 2 † s 1 s 1 † s 2 = ( 1 − n 1 ) n 2 , s_2^\dagger s_1s_2s_1^\dagger=-n_2(1-n_1),\qquad s_2^\dagger s_1s_1^\dagger s_2=(1-n_1)n_2, s 2 † s 1 s 2 s 1 † = − n 2 ( 1 − n 1 ) , s 2 † s 1 s 1 † s 2 = ( 1 − n 1 ) n 2 ,
s 1 † s 2 s 2 † s 1 = ( 1 − n 2 ) n 1 , s 2 s 1 † s 2 † s 1 = − ( 1 − n 2 ) n 1 , (142) s_1^\dagger s_2s_2^\dagger s_1=(1-n_2)n_1,\qquad s_2s_1^\dagger s_2^\dagger s_1=-(1-n_2)n_1, \tag*{(142)} s 1 † s 2 s 2 † s 1 = ( 1 − n 2 ) n 1 , s 2 s 1 † s 2 † s 1 = − ( 1 − n 2 ) n 1 , (142)
with τ S \tau_S τ S averages − f 2 ( 1 − f 1 ) -f_2(1-f_1) − f 2 ( 1 − f 1 ) , f 2 ( 1 − f 1 ) f_2(1-f_1) f 2 ( 1 − f 1 ) , f 1 ( 1 − f 2 ) f_1(1-f_2) f 1 ( 1 − f 2 ) and − f 1 ( 1 − f 2 ) -f_1(1-f_2) − f 1 ( 1 − f 2 ) . Collecting them,
⟨ s 2 † s 1 ⟩ τ M F ( 2 ) = λ 2 f 2 ( 1 − f 1 ) [ A 1 ( ω 2 ) − A 2 ( ω 1 ) ] + f 1 ( 1 − f 2 ) [ A 2 ( ω 2 ) − A 1 ( ω 1 ) ] ε 1 − ε 2 . (143) \langle s_2^\dagger s_1\rangle_{\tau_{\mathrm{MF}}^{(2)}}=\lambda^2\frac{f_2(1-f_1)[A_1(\omega_2)-A_2(\omega_1)]+f_1(1-f_2)[A_2(\omega_2)-A_1(\omega_1)]}{\varepsilon_1-\varepsilon_2}. \tag*{(143)} ⟨ s 2 † s 1 ⟩ τ MF ( 2 ) = λ 2 ε 1 − ε 2 f 2 ( 1 − f 1 ) [ A 1 ( ω 2 ) − A 2 ( ω 1 )] + f 1 ( 1 − f 2 ) [ A 2 ( ω 2 ) − A 1 ( ω 1 )] . (143)
This can be simplified using the KMS relation, which gives f i A 1 ( ω i ) = ( 1 − f i ) A 2 ( ω i ) f_i A_1(\omega_i) = (1-f_i) A_2(\omega_i) f i A 1 ( ω i ) = ( 1 − f i ) A 2 ( ω i ) . Grouping the four terms of Eq. (143) by frequency gives,
f 2 ( 1 − f 1 ) A 1 ( ω 2 ) + f 1 ( 1 − f 2 ) A 2 ( ω 2 ) = [ ( 1 − f 1 ) + f 1 ] ( 1 − f 2 ) A 2 ( ω 2 ) = ( 1 − f 2 ) A 2 ( ω 2 ) , f_2(1-f_1) A_1(\omega_2) + f_1(1-f_2) A_2(\omega_2) = [(1-f_1)+f_1](1-f_2) A_2(\omega_2) = (1-f_2) A_2(\omega_2), f 2 ( 1 − f 1 ) A 1 ( ω 2 ) + f 1 ( 1 − f 2 ) A 2 ( ω 2 ) = [( 1 − f 1 ) + f 1 ] ( 1 − f 2 ) A 2 ( ω 2 ) = ( 1 − f 2 ) A 2 ( ω 2 ) ,
f 2 ( 1 − f 1 ) A 2 ( ω 1 ) + f 1 ( 1 − f 2 ) A 1 ( ω 1 ) = [ f 2 + ( 1 − f 2 ) ] ( 1 − f 1 ) A 2 ( ω 1 ) = ( 1 − f 1 ) A 2 ( ω 1 ) , (144) f_2(1-f_1) A_2(\omega_1) + f_1(1-f_2) A_1(\omega_1) = [f_2+(1-f_2)](1-f_1) A_2(\omega_1) = (1-f_1) A_2(\omega_1), \tag*{(144)} f 2 ( 1 − f 1 ) A 2 ( ω 1 ) + f 1 ( 1 − f 2 ) A 1 ( ω 1 ) = [ f 2 + ( 1 − f 2 )] ( 1 − f 1 ) A 2 ( ω 1 ) = ( 1 − f 1 ) A 2 ( ω 1 ) , (144)
Using the definition of A 2 ( ω ) A_2(\omega) A 2 ( ω ) and rearranging, the coherence can be written as
⟨ s 2 † s 1 ⟩ M F ( 2 ) = λ 2 Φ ( ε 1 ) − Φ ( ε 2 ) ε 1 − ε 2 , Φ ( ε ) ≡ ∫ d ε ′ J ( ε ′ ) f ( ε ′ − μ ) − f ( ε − μ ) ε ′ − ε . (145) \langle s_2^\dagger s_1\rangle_{\mathrm{MF}}^{(2)} = \lambda^2 \frac{\Phi(\varepsilon_1)-\Phi(\varepsilon_2)}{\varepsilon_1-\varepsilon_2}, \qquad\Phi(\varepsilon) \equiv\int d\varepsilon' J(\varepsilon') \frac{f(\varepsilon'-\mu)-f(\varepsilon-\mu)}{\varepsilon'-\varepsilon}. \tag*{(145)} ⟨ s 2 † s 1 ⟩ MF ( 2 ) = λ 2 ε 1 − ε 2 Φ ( ε 1 ) − Φ ( ε 2 ) , Φ ( ε ) ≡ ∫ d ε ′ J ( ε ′ ) ε ′ − ε f ( ε ′ − μ ) − f ( ε − μ ) . (145)
The integrand of Φ \Phi Φ is regular at ε ′ = ε \varepsilon'=\varepsilon ε ′ = ε , tending to J ( ε ) f ′ ( ε − μ ) J(\varepsilon)f'(\varepsilon-\mu) J ( ε ) f ′ ( ε − μ ) at ε ′ → ε \varepsilon'\to\varepsilon ε ′ → ε . The presence of coherence can be understood through the Hamiltonian of mean force. Evaluating Eq. (117) gives
H M F = ∑ i = 1 , 2 ε ~ i n i + g ( s 1 † s 2 + s 2 † s 1 ) − λ 2 β [ D 2 ( ω 1 ) + D 2 ( ω 2 ) ] , g = λ 2 Φ ( ε 1 ) − Φ ( ε 2 ) f 1 − f 2 . (146) H_{\mathrm{MF}} = \sum_{i=1,2}\widetilde{\varepsilon}_i n_i + g(s_1^\dagger s_2+s_2^\dagger s_1)-\frac{\lambda^2}{\beta}[D_2(\omega_1)+D_2(\omega_2)], \qquad g=\lambda^2\frac{\Phi(\varepsilon_1)-\Phi(\varepsilon_2)}{f_1-f_2}. \tag*{(146)} H MF = i = 1 , 2 ∑ ε i n i + g ( s 1 † s 2 + s 2 † s 1 ) − β λ 2 [ D 2 ( ω 1 ) + D 2 ( ω 2 )] , g = λ 2 f 1 − f 2 Φ ( ε 1 ) − Φ ( ε 2 ) . (146)
with ε ~ i \widetilde{\varepsilon}_i ε i the renormalised RLM level of mode i i i . The MFG state is thus the Gibbs state of two modes coupled by a reservoir-mediated hopping, and the coherence of Eq. (145) reflects the rotation of their normal modes away from s 1 , s 2 s_1,s_2 s 1 , s 2 . Diagonalising Eq. (146) gives normal modes b ± b_\pm b ± defined as s 1 = cos θ b + − sin θ b − s_1=\cos\theta\,b_+-\sin\theta\,b_- s 1 = cos θ b + − sin θ b − , s 2 = sin θ b + + cos θ b − s_2=\sin\theta\,b_++\cos\theta\,b_- s 2 = sin θ b + + cos θ b − with eigenvalues E ± E_\pm E ± and angle defined by tan 2 θ = 2 g / ( ε ~ 1 − ε ~ 2 ) \tan2\theta=2g/(\widetilde{\varepsilon}_1-\widetilde{\varepsilon}_2) tan 2 θ = 2 g / ( ε 1 − ε 2 ) . The Gibbs state for H M F H_{\mathrm{MF}} H MF has thermal occupations ⟨ b k † b l ⟩ = δ k l f ( E k − μ ) \langle b_k^\dagger b_l\rangle=\delta_{kl}f(E_k-\mu) ⟨ b k † b l ⟩ = δ k l f ( E k − μ ) . Using sin 2 θ = 2 g / ( E + − E − ) \sin2\theta=2g/(E_+-E_-) sin 2 θ = 2 g / ( E + − E − ) this gives
⟨ s 2 † s 1 ⟩ = g f ( E + − μ ) − f ( E − − μ ) E + − E − ≈ g f 1 − f 2 ε 1 − ε 2 , (147) \langle s_2^\dagger s_1\rangle= g\frac{f(E_+-\mu)-f(E_--\mu)}{E_+-E_-} \approx g\frac{f_1-f_2}{\varepsilon_1-\varepsilon_2}, \tag*{(147)} ⟨ s 2 † s 1 ⟩ = g E + − E − f ( E + − μ ) − f ( E − − μ ) ≈ g ε 1 − ε 2 f 1 − f 2 , (147)
where the approximate equality follows from the fact that g = O ( λ 2 ) g=O(\lambda^2) g = O ( λ 2 ) , so the divided difference is only needed at zeroth order, where E ± = ε 1 , 2 E_\pm=\varepsilon_{1,2} E ± = ε 1 , 2 . Using the definition of g g g then gives Eq. (145) . However, as explained in section SIV, this is only valid at sufficiently high temperature.
Single impurity Anderson model
We now consider an interacting system, the single impurity Anderson model (SIAM), with Hamiltonian
H S = ε d ( n ↑ + n ↓ ) + U n ↑ n ↓ , n σ = d σ † d σ . (148) H_S=\varepsilon_d(n_\uparrow+n_\downarrow)+U n_\uparrow n_\downarrow,\qquad n_\sigma=d_\sigma^\dagger d_\sigma. \tag*{(148)} H S = ε d ( n ↑ + n ↓ ) + U n ↑ n ↓ , n σ = d σ † d σ . (148)
coupled to a spinful reservoir,
H = H S + ∑ σ ∫ d ε ε a ε σ † a ε σ + λ ∑ σ ∫ d ε J ( ε ) ( d σ † a ε σ + a ε σ † d σ ) . (149) H=H_S+\sum_\sigma\int d\varepsilon\,\varepsilon\,a_{\varepsilon\sigma}^\dagger a_{\varepsilon\sigma}+\lambda\sum_\sigma\int d\varepsilon\,\sqrt{J(\varepsilon)}(d_\sigma^\dagger a_{\varepsilon\sigma}+a_{\varepsilon\sigma}^\dagger d_\sigma). \tag*{(149)} H = H S + σ ∑ ∫ d ε ε a ε σ † a ε σ + λ σ ∑ ∫ d ε J ( ε ) ( d σ † a ε σ + a ε σ † d σ ) . (149)
The system has four Fock states which are all eigenstates of the grand-canonical Hamiltonian
∣ 0 ⟩ , ∣ σ ⟩ = d σ † ∣ 0 ⟩ , ∣ 2 ⟩ = d ↑ † d ↓ † ∣ 0 ⟩ , |0\rangle,\qquad|\sigma\rangle=d_\sigma^\dagger|0\rangle,\qquad|2\rangle=d_\uparrow^\dagger d_\downarrow^\dagger|0\rangle, ∣0 ⟩ , ∣ σ ⟩ = d σ † ∣0 ⟩ , ∣2 ⟩ = d ↑ † d ↓ † ∣0 ⟩ ,
K S ∣ 0 ⟩ = 0 , K S ∣ σ ⟩ = ( ε d − μ ) ∣ σ ⟩ , K S ∣ 2 ⟩ = [ 2 ( ε d − μ ) + U ] ∣ 2 ⟩ . (150) K_S|0\rangle=0,\qquad K_S|\sigma\rangle=(\varepsilon_d-\mu)|\sigma\rangle,\qquad K_S|2\rangle=[2(\varepsilon_d-\mu)+U]|2\rangle. \tag*{(150)} K S ∣0 ⟩ = 0 , K S ∣ σ ⟩ = ( ε d − μ ) ∣ σ ⟩ , K S ∣2 ⟩ = [ 2 ( ε d − μ ) + U ] ∣2 ⟩ . (150)
Unlike in the previous two examples, d σ d_\sigma d σ is not an eigenoperator. Removing a particle of spin σ \sigma σ changes the energy by an amount that depends on whether σ ˉ \bar{\sigma} σ ˉ is occupied. The two Bohr frequencies, corresponding to adding a particle to the empty and to the singly occupied level, are
ω 1 ≡ μ − ε d , ω 2 ≡ μ − ε d − U . (151) \omega_1\equiv\mu-\varepsilon_d,\qquad\omega_2\equiv\mu-\varepsilon_d-U. \tag*{(151)} ω 1 ≡ μ − ε d , ω 2 ≡ μ − ε d − U . (151)
Since [ K S , d σ ( 1 − n σ ˉ ) ] = ω 1 d σ ( 1 − n σ ˉ ) [K_S,d_\sigma(1-n_{\bar{\sigma}})]=\omega_1d_\sigma(1-n_{\bar{\sigma}}) [ K S , d σ ( 1 − n σ ˉ )] = ω 1 d σ ( 1 − n σ ˉ ) and [ K S , d σ n σ ˉ ] = ω 2 d σ n σ ˉ [K_S,d_\sigma n_{\bar{\sigma}}]=\omega_2d_\sigma n_{\bar{\sigma}} [ K S , d σ n σ ˉ ] = ω 2 d σ n σ ˉ , each channel splits into two Bohr-frequency components, X σ , ω 1 = d σ ( 1 − n σ ˉ ) X_{\sigma,\omega_1}=d_\sigma(1-n_{\bar{\sigma}}) X σ , ω 1 = d σ ( 1 − n σ ˉ ) , X σ , ω 2 = d σ n σ ˉ X_{\sigma,\omega_2}=d_\sigma n_{\bar{\sigma}} X σ , ω 2 = d σ n σ ˉ , so that X σ = X σ , ω 1 + X σ , ω 2 X_\sigma=X_{\sigma,\omega_1}+X_{\sigma,\omega_2} X σ = X σ , ω 1 + X σ , ω 2 . The products entering the diagonal part are projectors onto Fock states,
X σ , ω 1 † X σ , ω 1 = ∣ σ ⟩ ⟨ σ ∣ , X σ , ω 1 X σ , ω 1 † = ∣ 0 ⟩ ⟨ 0 ∣ , X_{\sigma,\omega_1}^\dagger X_{\sigma,\omega_1}=|\sigma\rangle\langle\sigma|,\qquad X_{\sigma,\omega_1}X_{\sigma,\omega_1}^\dagger=|0\rangle\langle0|, X σ , ω 1 † X σ , ω 1 = ∣ σ ⟩ ⟨ σ ∣ , X σ , ω 1 X σ , ω 1 † = ∣0 ⟩ ⟨ 0∣ ,
X σ , ω 2 † X σ , ω 2 = ∣ 2 ⟩ ⟨ 2 ∣ , X σ , ω 2 X σ , ω 2 † = ∣ σ ˉ ⟩ ⟨ σ ˉ ∣ . (152) X_{\sigma,\omega_2}^\dagger X_{\sigma,\omega_2}=|2\rangle\langle2|,\qquad X_{\sigma,\omega_2}X_{\sigma,\omega_2}^\dagger=|\bar{\sigma}\rangle\langle\bar{\sigma}|. \tag*{(152)} X σ , ω 2 † X σ , ω 2 = ∣2 ⟩ ⟨ 2∣ , X σ , ω 2 X σ , ω 2 † = ∣ σ ˉ ⟩ ⟨ σ ˉ ∣. (152)
Writing P k = ∣ k ⟩ ⟨ k ∣ P_k = |k\rangle\langle k| P k = ∣ k ⟩ ⟨ k ∣ for k ∈ { 0 , ↑ , ↓ , 2 } k \in\{0,\uparrow,\downarrow,2\} k ∈ { 0 , ↑ , ↓ , 2 } and p k ( 0 ) = ⟨ P k ⟩ τ S p_k^{(0)} = \langle P_k\rangle_{\tau_S} p k ( 0 ) = ⟨ P k ⟩ τ S , the diagonal part is
Δ d i a g = ∑ σ [ τ S ( P σ − p σ ( 0 ) ) D 1 ( ω 1 ) + τ S ( P 0 − p 0 ( 0 ) ) D 2 ( ω 1 ) + τ S ( P 2 − p 2 ( 0 ) ) D 1 ( ω 2 ) + τ S ( P σ ˉ − p σ ˉ ( 0 ) ) D 2 ( ω 2 ) ] . (153) \Delta_{\mathrm{diag}} = \sum_{\sigma}\left[\tau_S(P_\sigma-p_\sigma^{(0)})D_1(\omega_1)+\tau_S(P_0-p_0^{(0)})D_2(\omega_1)
+\tau_S(P_2-p_2^{(0)})D_1(\omega_2)+\tau_S(P_{\bar{\sigma}}-p_{\bar{\sigma}}^{(0)})D_2(\omega_2)\right]. \tag*{(153)} Δ diag = σ ∑ [ τ S ( P σ − p σ ( 0 ) ) D 1 ( ω 1 ) + τ S ( P 0 − p 0 ( 0 ) ) D 2 ( ω 1 ) + τ S ( P 2 − p 2 ( 0 ) ) D 1 ( ω 2 ) + τ S ( P σ ˉ − p σ ˉ ( 0 ) ) D 2 ( ω 2 ) ] . (153)
The off-diagonal part pairs X σ , ω 1 X_{\sigma,\omega_1} X σ , ω 1 with X σ , ω 2 X_{\sigma,\omega_2} X σ , ω 2 . Every such product contains the factor n σ ˉ ( 1 − n σ ˉ ) = 0 n_{\bar{\sigma}}(1-n_{\bar{\sigma}})=0 n σ ˉ ( 1 − n σ ˉ ) = 0 , meaning Δ o f f = 0 \Delta_{\mathrm{off}}=0 Δ off = 0 . The state τ M F ( 2 ) \tau_{\mathrm{MF}}^{(2)} τ MF ( 2 ) is therefore diagonal in the Fock basis as required by conservation of charge and S z S_z S z .
The bare Gibbs state is τ S = ∑ k p k ( 0 ) P k \tau_S=\sum_k p_k^{(0)}P_k τ S = ∑ k p k ( 0 ) P k with
p 0 ( 0 ) = 1 Z S , p σ ( 0 ) = e β ω 1 Z S , p 2 ( 0 ) = e β ( ω 1 + ω 2 ) Z S , Z S = 1 + 2 e β ω 1 + e β ( ω 1 + ω 2 ) . (154) p_0^{(0)}=\frac{1}{Z_S},\qquad p_\sigma^{(0)}=\frac{e^{\beta\omega_1}}{Z_S},\qquad p_2^{(0)}=\frac{e^{\beta(\omega_1+\omega_2)}}{Z_S},\qquad Z_S=1+2e^{\beta\omega_1}+e^{\beta(\omega_1+\omega_2)}. \tag*{(154)} p 0 ( 0 ) = Z S 1 , p σ ( 0 ) = Z S e β ω 1 , p 2 ( 0 ) = Z S e β ( ω 1 + ω 2 ) , Z S = 1 + 2 e β ω 1 + e β ( ω 1 + ω 2 ) . (154)
Since P k P_k P k and τ S \tau_S τ S commute, Tr [ P k τ S ( P k ′ − p k ′ ( 0 ) ) ] = p k ( 0 ) ( δ k k ′ − p k ′ ( 0 ) ) \operatorname{Tr}[P_k\tau_S(P_{k'}-p_{k'}^{(0)})]=p_k^{(0)}(\delta_{kk'}-p_{k'}^{(0)}) Tr [ P k τ S ( P k ′ − p k ′ ( 0 ) )] = p k ( 0 ) ( δ k k ′ − p k ′ ( 0 ) ) . Each state therefore collects D 1 D_1 D 1 from every transition in which it is in the state with one more particle, and D 2 D_2 D 2 from every transition in which it is in the state with one less particle. Defining
w 0 = 2 D 2 ( ω 1 ) , w σ = D 1 ( ω 1 ) + D 2 ( ω 2 ) , w 2 = 2 D 1 ( ω 2 ) , w ˉ ≡ ∑ k p k ( 0 ) w k , (155) w_0=2D_2(\omega_1),\qquad w_\sigma=D_1(\omega_1)+D_2(\omega_2),\qquad w_2=2D_1(\omega_2),\qquad\bar{w}\equiv\sum_k p_k^{(0)}w_k, \tag*{(155)} w 0 = 2 D 2 ( ω 1 ) , w σ = D 1 ( ω 1 ) + D 2 ( ω 2 ) , w 2 = 2 D 1 ( ω 2 ) , w ˉ ≡ k ∑ p k ( 0 ) w k , (155)
we obtain
τ M F ( 2 ) = ∑ k p k ( 0 ) [ 1 + λ 2 ( w k − w ˉ ) ] ∣ k ⟩ ⟨ k ∣ . (156) \tau_{\mathrm{MF}}^{(2)}=\sum_k p_k^{(0)}\left[1+\lambda^2(w_k-\bar{w})\right]|k\rangle\langle k|. \tag*{(156)} τ MF ( 2 ) = k ∑ p k ( 0 ) [ 1 + λ 2 ( w k − w ˉ ) ] ∣ k ⟩ ⟨ k ∣. (156)
The corrections to the double occupancy and to the spin-resolved occupation follow directly,
δ ⟨ n ↑ n ↓ ⟩ = λ 2 p 2 ( 0 ) ( w 2 − w ˉ ) , δ ⟨ n σ ⟩ = λ 2 [ p σ ( 0 ) ( w σ − w ˉ ) + p 2 ( 0 ) ( w 2 − w ˉ ) ] . (157) \delta\langle n_\uparrow n_\downarrow\rangle=\lambda^2p_2^{(0)}(w_2-\bar{w}),\qquad
\delta\langle n_\sigma\rangle=\lambda^2\left[p_\sigma^{(0)}(w_\sigma-\bar{w})+p_2^{(0)}(w_2-\bar{w})\right]. \tag*{(157)} δ ⟨ n ↑ n ↓ ⟩ = λ 2 p 2 ( 0 ) ( w 2 − w ˉ ) , δ ⟨ n σ ⟩ = λ 2 [ p σ ( 0 ) ( w σ − w ˉ ) + p 2 ( 0 ) ( w 2 − w ˉ ) ] . (157)
The Hamiltonian of mean force is a renormalised SIAM,
H M F = ε ~ d ( n ↑ + n ↓ ) + U ~ n ↑ n ↓ − λ 2 β w 0 , (158) H_{\mathrm{MF}}=\widetilde{\varepsilon}_d(n_\uparrow+n_\downarrow)+\widetilde{U}n_\uparrow n_\downarrow-\frac{\lambda^2}{\beta}w_0, \tag*{(158)} H MF = ε d ( n ↑ + n ↓ ) + U n ↑ n ↓ − β λ 2 w 0 , (158)
where
ε ~ d − ε d = − λ 2 β ( w σ − w 0 ) = − λ 2 β [ D 1 ( ω 1 ) + D 2 ( ω 2 ) − 2 D 2 ( ω 1 ) ] , \widetilde{\varepsilon}_d-\varepsilon_d=-\frac{\lambda^2}{\beta}(w_\sigma-w_0)=-\frac{\lambda^2}{\beta}\left[D_1(\omega_1)+D_2(\omega_2)-2D_2(\omega_1)\right], ε d − ε d = − β λ 2 ( w σ − w 0 ) = − β λ 2 [ D 1 ( ω 1 ) + D 2 ( ω 2 ) − 2 D 2 ( ω 1 ) ] ,
U ~ − U = − λ 2 β ( w 0 − 2 w σ + w 2 ) = − 2 λ 2 β { [ D 1 ( ω 2 ) − D 2 ( ω 2 ) ] − [ D 1 ( ω 1 ) − D 2 ( ω 1 ) ] } . (159) \widetilde{U}-U=-\frac{\lambda^2}{\beta}(w_0-2w_\sigma+w_2)=-\frac{2\lambda^2}{\beta}\left\{[D_1(\omega_2)-D_2(\omega_2)]-[D_1(\omega_1)-D_2(\omega_1)]\right\}. \tag*{(159)} U − U = − β λ 2 ( w 0 − 2 w σ + w 2 ) = − β 2 λ 2 { [ D 1 ( ω 2 ) − D 2 ( ω 2 )] − [ D 1 ( ω 1 ) − D 2 ( ω 1 )] } . (159)
Non-interacting fermions
If H H H is quadratic in mode operators and particle number conserving, then the exact dynamics can be obtained via unitary evolution of the single-particle correlation matrix C i j ( t ) = Tr ( ρ ( t ) d j † d i ) C_{ij}(t)=\operatorname{Tr}(\rho(t)d_j^\dagger d_i) C ij ( t ) = Tr ( ρ ( t ) d j † d i ) over all the system and reservoir modes { d i } = ( s 1 , … , s L S , a 1 , … , a L R ) \{d_i\}=(s_1,\ldots,s_{L_S},a_1,\ldots,a_{L_R}) { d i } = ( s 1 , … , s L S , a 1 , … , a L R ) . The ordering of the indices here is chosen such that the matrix C C C evolves unitarily, as we will now show. Defining U ( t ) = e − i H t U(t)=e^{-iHt} U ( t ) = e − i H t , we have
C i j ( t ) = ⟨ ψ ( t ) ∣ d j † d i ∣ ψ ( t ) ⟩ = ⟨ ψ ( 0 ) ∣ U † ( t ) d j † U ( t ) U † ( t ) d i U ( t ) ∣ ψ ( 0 ) ⟩ = ⟨ ψ ( 0 ) ∣ d j † ( t ) d i ( t ) ∣ ψ ( 0 ) ⟩ . (160) C_{ij}(t)=\langle\psi(t)|d_j^\dagger d_i|\psi(t)\rangle=\langle\psi(0)|U^\dagger(t)d_j^\dagger U(t)U^\dagger(t)d_iU(t)|\psi(0)\rangle
=\langle\psi(0)|d_j^\dagger(t)d_i(t)|\psi(0)\rangle. \tag*{(160)} C ij ( t ) = ⟨ ψ ( t ) ∣ d j † d i ∣ ψ ( t )⟩ = ⟨ ψ ( 0 ) ∣ U † ( t ) d j † U ( t ) U † ( t ) d i U ( t ) ∣ ψ ( 0 )⟩ = ⟨ ψ ( 0 ) ∣ d j † ( t ) d i ( t ) ∣ ψ ( 0 )⟩ . (160)
Differentiating d j † ( t ) d_j^\dagger(t) d j † ( t ) gives
d d j † ( t ) d t = i U † ( t ) [ H , d j † ] U ( t ) = i ∑ k H k j d k † ( t ) (161) \frac{d d_j^\dagger(t)}{dt}=iU^\dagger(t)[H,d_j^\dagger]U(t)=i\sum_k H_{kj}d_k^\dagger(t) \tag*{(161)} d t d d j † ( t ) = i U † ( t ) [ H , d j † ] U ( t ) = i k ∑ H k j d k † ( t ) (161)
where we have used [ H , d j † ] = ∑ k H k j d k † [H,d_j^\dagger]=\sum_k H_{kj}d_k^\dagger [ H , d j † ] = ∑ k H k j d k † for H = ∑ k l H k l d k † d l H=\sum_{kl}H_{kl}d_k^\dagger d_l H = ∑ k l H k l d k † d l . Defining the vector of operators d ⃗ † ( t ) = ( d 1 † ( t ) , … , d L † ( t ) ) T \vec{d}^{\dagger}(t)=(d_1^\dagger(t),\ldots,d_L^\dagger(t))^T d † ( t ) = ( d 1 † ( t ) , … , d L † ( t ) ) T , where L = L S + L R L=L_S+L_R L = L S + L R , we have
d d ⃗ † ( t ) d t = i H T d ⃗ † ( t ) . (162) \frac{d\vec{d}^{\dagger}(t)}{dt}=iH^T\vec{d}^{\dagger}(t). \tag*{(162)} d t d d † ( t ) = i H T d † ( t ) . (162)
which integrates to d ⃗ † ( t ) = U ∗ ( t ) d ⃗ † ( 0 ) \vec{d}^{\dagger}(t)=U^*(t)\vec{d}^{\dagger}(0) d † ( t ) = U ∗ ( t ) d † ( 0 ) with U ( t ) ≡ e − i H t U(t)\equiv e^{-iHt} U ( t ) ≡ e − i H t . Substituting this into Eq. (160) gives the result
C i j ( t ) = ⟨ ψ ( 0 ) ∣ ∑ k l U k j ∗ d k † ( U i l ∗ d l † ) † ∣ ψ ( 0 ) ⟩ = ∑ k l U k j ∗ U i l ⟨ ψ ( 0 ) ∣ d k † d l ∣ ψ ( 0 ) ⟩ = ∑ k l U k j ∗ U i l C l k ( 0 ) . (163) \begin{aligned}
C_{ij}(t)&=\langle\psi(0)|\sum_{kl}U^*_{kj}d_k^\dagger(U^*_{il}d_l^\dagger)^\dagger|\psi(0)\rangle\\
&=\sum_{kl}U^*_{kj}U_{il}\langle\psi(0)|d_k^\dagger d_l|\psi(0)\rangle\\
&=\sum_{kl}U^*_{kj}U_{il}C_{lk}(0). \tag*{(163)}
\end{aligned} C ij ( t ) = ⟨ ψ ( 0 ) ∣ k l ∑ U k j ∗ d k † ( U i l ∗ d l † ) † ∣ ψ ( 0 )⟩ = k l ∑ U k j ∗ U i l ⟨ ψ ( 0 ) ∣ d k † d l ∣ ψ ( 0 )⟩ = k l ∑ U k j ∗ U i l C l k ( 0 ) . (163)
Thus we have
C ( t ) = U C ( 0 ) U † . (164) C(t)=UC(0)U^\dagger. \tag*{(164)} C ( t ) = U C ( 0 ) U † . (164)
Exact mean force Gibbs state from the correlation matrix
For quadratic models the correlation matrix also gives the equilibrium state directly, without any time evolution. For simplicity, we consider a finite number of modes, but the continuum can be taken straightforwardly. We write the generator K K K of Eq. (38) in terms of a single particle matrix K i j K_{ij} K ij and diagonalise it in its eigenmodes,
K = ∑ i j K i j d i † d j = ∑ n ω n b n † b n . (165) K=\sum_{ij}K_{ij}d_i^\dagger d_j=\sum_n\omega_n b_n^\dagger b_n. \tag*{(165)} K = ij ∑ K ij d i † d j = n ∑ ω n b n † b n . (165)
The global Gibbs state factorises over the normal modes { b n } \{b_n\} { b n } , with ⟨ b n † b m ⟩ = δ n m f ( ω n ) \langle b_n^\dagger b_m\rangle=\delta_{nm}f(\omega_n) ⟨ b n † b m ⟩ = δ nm f ( ω n ) , where f ( ω ) f(\omega) f ( ω ) is the Fermi function. In the original basis, the single particle correlation matrix of the global Gibbs state is
C i j ≡ Tr ( τ t o t d i † d j ) = [ f ( K ) ] i j . (166) C_{ij}\equiv\operatorname{Tr}(\tau_{\mathrm{tot}}d_i^\dagger d_j)=[f(K)]_{ij}. \tag*{(166)} C ij ≡ Tr ( τ tot d i † d j ) = [ f ( K ) ] ij . (166)
Within the system block, we have C S = [ f ( K ) ] i , j ∈ S C_S=[f(K)]_{i,j\in S} C S = [ f ( K ) ] i , j ∈ S . For quadratic fermions, the reduced density matrix of a subsystem (provided the state is parity symmetric) is given in terms of the correlation matrix via [37 , 40 , 49 ]
τ M F = det ( 1 − C S ) exp { ∑ i j ∈ S [ ln ( C S ( 1 − C S ) − 1 ) ] i j s i † s j } . (167) \tau_{\mathrm{MF}}=\det(1-C_S)\exp\left\{\sum_{ij\in S}\left[\ln\left(C_S(1-C_S)^{-1}\right)\right]_{ij}s_i^\dagger s_j\right\}. \tag*{(167)} τ MF = det ( 1 − C S ) exp ⎩ ⎨ ⎧ ij ∈ S ∑ [ ln ( C S ( 1 − C S ) − 1 ) ] ij s i † s j ⎭ ⎬ ⎫ . (167)
or equivalently
τ M F = exp ( − ∑ i j M i j s i † s j ) Tr S exp ( − ∑ i j M i j s i † s j ) , M = ln ( C S − 1 − 1 ) . (168) \tau_{\mathrm{MF}}=\frac{\exp\left(-\sum_{ij}M_{ij}s_i^\dagger s_j\right)}{\operatorname{Tr}_S\exp\left(-\sum_{ij}M_{ij}s_i^\dagger s_j\right)},\qquad M=\ln\left(C_S^{-1}-1\right). \tag*{(168)} τ MF = Tr S exp ( − ∑ ij M ij s i † s j ) exp ( − ∑ ij M ij s i † s j ) , M = ln ( C S − 1 − 1 ) . (168)
β − 1 M \beta^{-1}M β − 1 M is thus the single-particle matrix of K M F = H M F − μ N S K_{\mathrm{MF}}=H_{\mathrm{MF}}-\mu N_S K MF = H MF − μ N S with temperature dependence through C S C_S C S .