Introduction
Introduction.— Maxwell’s demon is a thought experiment in which feedback control based on measurement outcomes appears to violate the second law of thermodynamics [1 , 2 , 3 , 4 ] . In modern information thermodynamics, the paradox is resolved by incorporating the information acquired through measurement and the cost of erasing memory [5 , 6 , 7 , 8 ] , and the last decades have witnessed remarkable success in this field especially for few-body systems amenable to precise control [9 , 10 , 11 , 12 , 13 , 14 , 15 ] . However, despite recent progress [16 , 17 ] , many questions remain as to how information thermodynamics should be formulated in many-body systems, the demon’s original setting. Such questions have especially regained importance in light of recent experimental advances in quantum computers and quantum simulators, which have made measurement-based feedback control of otherwise unitary many-body dynamics feasible [18 , 19 , 20 , 21 ] . It is therefore important to explore how information thermodynamics extends to closed quantum many-body systems.
In parallel, the extractable work in closed quantum dynamics (without measurement and feedback) has recently been studied extensively as a fundamental problem in quantum thermodynamics and quantum-battery research [12 , 22 , 23 , 24 , 25 , 26 , 27 , 28 , 29 , 30 , 31 , 32 , 33 , 34 , 35 , 36 , 37 , 38 ] . The maximum work extractable under arbitrary unitary operations is called ergotropy [22 ] , and a state whose energy expectation value cannot be lowered by any unitary operation is called a passive state [39 , 40 ] . This property is an (idealized) formulation of Planck’s principle, which provides one way to characterize quantum thermal equilibrium. The canonical ensemble ρ e n s = e − β H / Z \rho^{\mathrm{ens}} = e^{-\beta H}/Z ρ ens = e − β H / Z with β ≥ 0 \beta\ge0 β ≥ 0 is passive, and no work can be extracted from it by unitary operations alone.
In a closed quantum many-body system, however, not only a canonical ensemble but also pure states can represent thermal equilibrium with respect to local observables. One of such classes of states is microscopic thermal equilibrium (MITE), where the expectation value of every local observable agrees with that in the corresponding canonical ensemble [41 ] . Importantly, such thermal equilibrium with respect to local observables does not immediately imply passivity in the conventional sense. Indeed, if arbitrary unitary operations are allowed, we can transform the initial thermal pure state to the ground state, which leads to extensive erg otropy [22 ] . In realistic quantum many-body systems, however, arbitrary global unitaries cannot be implemented freely. It has recently been shown that no extensive work can be extracted from finite-time operations under local Hamiltonians, i.e., recovering the second law of thermodynamics for thermal pure states [33 , 42 ] .
Against this background, one crucial question is how the second law on work extraction can be modified under realistic (e.g., local) measurement and feedback operations in closed quantum many-body systems. We note that related notions, such as daemonic ergotropy, characterize the enhancement of ergotropy by measurement and feedback [43 ] . However, it remains unclear how such measurement-assisted work extraction should be formulated thermodynamically in closed many-body systems (using, e.g., temperature), especially for thermal pure states under locality constraints.
In this Letter, we develop an information-thermodynamic framework for closed quantum many-body systems initially in MITE under experimentally and physically motivated restrictions: local quantum measurements and finite-time unitary controls generated by local Hamiltonians. When we additionally assume each local term in the feedback Hamiltonian to depend only on measurement outcomes obtained in nearby regions, we rigorously prove that the work extractable from MITE (a pure state) coincides in the thermodynamic limit with that extractable from the canonical ensemble (a mixed state). The latter is expressed in terms of the QC mutual information [5 , 44 , 45 , 46 ] , the Holevo information [47 , 48 , 49 ] , and non-equilibrium free energy. Thus, even when the entire system is in a pure state from which macroscopic work could be extracted under unrestricted operations, the work extractable under these constraints is bounded in terms of measurement-related information quantities. In contrast, if we allow feedback based on distant measurement outcomes, we show that long-range correlations hidden in a pure MITE state can be exploited for work extraction exceeding that from the canonical ensemble despite finite-time quantum operations. Importantly, our results are relevant to quantum simulators and quantum computers, where pure many-body states are naturally prepared whereas preparing an exact canonical ensemble is generally challenging.
Measurement-and-feedback protocol in closed many-body systems.— We consider the work extractable by measurement and feedback control applied to a closed quantum many-body system of volume V = L d V = L^d V = L d . Here, a closed quantum many-body system exchanges no heat with a heat bath but is subject to external operations consisting of measurement and feedback. We focus on a cyclic protocol, and take the initial and final Hamiltonians to be H i = H f = H = ∑ X h X H_i = H_f = H = \sum_X h_X H i = H f = H = ∑ X h X , where h X h_X h X is a local operator supported on a subregion X X X , whose diameter diam ( X ) \operatorname{diam}(X) diam ( X ) does not depend on the system size L L L . Here, we define the diameter of the region X X X by diam ( X ) : = max { dist ( x , y ) : x , y ∈ X } \operatorname{diam}(X) := \max\{\operatorname{dist}(x,y) : x,y \in X\} diam ( X ) := max { dist ( x , y ) : x , y ∈ X } , where dist ( x , y ) \operatorname{dist}(x,y) dist ( x , y ) is the Euclidean distance.
We consider a protocol similar to that in [5 ] , where the measurement and feedback are applied only once. It consists of three steps. First, a unitary operation U 1 U_1 U 1 is applied before the measurement. Next, a measurement described by Kraus operators { M k } \{M_k\} { M k } is performed, where k k k is a set of measurement outcomes and { M k } \{M_k\} { M k } satisfies ∑ k M k † M k = 1 \sum_k M_k^\dagger M_k = 1 ∑ k M k † M k = 1 . Finally, conditioned on k k k , a feedback unitary U k U_k U k is applied. For an initial state ρ i \rho_i ρ i , the probability of obtaining outcome k k k and the postfeedback state conditioned on outcome k k k are P k = Tr [ M k U 1 ρ i U 1 † M k † ] P_k = \operatorname{Tr}\left[M_k U_1 \rho_i U_1^\dagger M_k^\dagger\right] P k = Tr [ M k U 1 ρ i U 1 † M k † ] and ρ k ′ = U k M k U 1 ρ i U 1 † M k † U k † / P k \rho'_k = U_k M_k U_1 \rho_i U_1^\dagger M_k^\dagger U_k^\dagger/P_k ρ k ′ = U k M k U 1 ρ i U 1 † M k † U k † / P k , respectively. Averaging over the outcomes { k } \{k\} { k } gives the final state ρ f = ∑ k P k ρ k ′ = ∑ k U k M k U 1 ρ i U 1 † M k † U k † \rho_f = \sum_k P_k \rho'_k = \sum_k U_k M_k U_1 \rho_i U_1^\dagger M_k^\dagger U_k^\dagger ρ f = ∑ k P k ρ k ′ = ∑ k U k M k U 1 ρ i U 1 † M k † U k † .
Because the closed quantum many-body system exchanges no heat with a heat bath, we define the extracted work as the net decrease in the system’s energy induced by the sequence of external operations [5 ] :
W = Tr [ H ρ i ] − Tr [ H ρ f ] . (1) W = \operatorname{Tr}[H\rho_i] - \operatorname{Tr}[H\rho_f]. \tag*{(1)} W = Tr [ H ρ i ] − Tr [ H ρ f ] . (1)
Although a physical implementation of the measurement may exchange energy with the system, the resulting change in the system energy is reflected in ρ f \rho_f ρ f ; the energetic costs associated with the measurement apparatus and feedback controller, including information processing and memory reset, are not included in W W W [5 , 8 ] . A positive value W > 0 W > 0 W > 0 means that energy has been extracted from the system.
Our main assumption is that U 1 U_1 U 1 and U k U_k U k are generated by local Hamiltonians in an L L L -independent time and that { M k } \{M_k\} { M k } describes local measurement as specified below. If we first assume additionally that each local term in the feedback Hamiltonian depends only on measurement outcomes obtained within an L L L -independent distance of its support, we obtain our main theorem: the work W W W extractable from general initial MITE is given by
β W ≃ β W e n s = I Q C − χ − D ( ρ f e n s ∥ ρ i e n s ) . (2) \beta W \simeq\beta W_{\mathrm{ens}} = I_{\mathrm{QC}} - \chi- D\left(\rho_f^{\mathrm{ens}} \middle\| \rho_i^{\mathrm{ens}}\right). \tag*{(2)} β W ≃ β W ens = I QC − χ − D ( ρ f ens ρ i ens ) . (2)
Here, W e n s W_{\mathrm{ens}} W ens denotes the work extracted by applying the same protocol to the initial canonical ensemble ρ i e n s = e − β H / Z \rho_i^{\mathrm{ens}} = e^{-\beta H}/Z ρ i ens = e − β H / Z , producing the final (generally nonequilibrium) state ρ f e n s \rho_f^{\mathrm{ens}} ρ f ens . The symbol ≃ \simeq ≃ means that subsequent corrections are neglected for large L L L , and the right-hand side consists of information-theoretical quantities defined for this protocol with the canonical initial state. In the following, we detail the definition of each quantity, the conditions for the equality, and their meanings.
Work extraction from MITE under constraints on locality and operational times.— We first discuss the first equality in Eq. (2) , which serves as a main achievement in our work. A microscopic thermal equilibrium (MITE) ρ i M I T E \rho_i^{\mathrm{MITE}} ρ i MITE [41 ] with inverse temperature β \beta β is defined as a (possibly pure) state where expectation values of local observables O O O are indistinguishable from those of the canonical ensemble ρ i e n s \rho_i^{\mathrm{ens}} ρ i ens with the same β \beta β :
Tr [ ρ i M I T E O ] ≃ Tr [ ρ i e n s O ] . (3) \operatorname{Tr}[\rho_i^{\mathrm{MITE}} O] \simeq\operatorname{Tr}[\rho_i^{\mathrm{ens}} O]. \tag*{(3)} Tr [ ρ i MITE O ] ≃ Tr [ ρ i ens O ] . (3)
Here, A ≃ B A \simeq B A ≃ B means ∣ A − B ∣ = o ( 1 ) |A-B| = o(1) ∣ A − B ∣ = o ( 1 ) and o ( V ) o(V) o ( V ) for intensive and extensive A A A , respectively. To keep the discussion general, we consider an extended notion of MITE and assume that the local observable O O O has support S l o c S_{\mathrm{loc}} S loc whose radius scales sublinearly with the system size L L L : diam [ S l o c ] = O ( L s ) \operatorname{diam}[S_{\mathrm{loc}}] = O(L^s) diam [ S loc ] = O ( L s ) . Note that s = 0 s = 0 s = 0 corresponds to the ordinary MITE [41 ] , but the MITE with 0 < s < 1 0 < s < 1 0 < s < 1 can also occur for typical states randomly taken from the microcanonical ensemble [50 , 51 ] . As shown below, our results hold for each fixed s s s (0 < s < 1 0 < s < 1 0 < s < 1 ).
Next, we write the unitary operation applied before the measurement as U 1 ( t ′ ) : = T exp [ − i ∫ 0 t ′ H 1 ( u ) d u ] U_1(t') := \mathcal{T} \exp\left[-i\int_0^{t'} H_1(u)\,\mathrm{d}u\right] U 1 ( t ′ ) := T exp [ − i ∫ 0 t ′ H 1 ( u ) d u ] and the feedback operation applied after the measurement as U k ( t ) : = T exp [ − i ∫ 0 t H k ( u ) d u ] U_k(t) := \mathcal{T} \exp\left[-i\int_0^t H_k(u)\,\mathrm{d}u\right] U k ( t ) := T exp [ − i ∫ 0 t H k ( u ) d u ] , where T \mathcal{T} T denotes the time ordering. Crucially, we assume that the Hamiltonians H 1 ( u ) H_1(u) H 1 ( u ) and H k ( u ) H_k(u) H k ( u ) are given by the sum of local operators (whose supports do not increase with L L L ), and the durations of the two unitary operations satisfy t , t ′ = O ( L s ) t,t' = O(L^s) t , t ′ = O ( L s ) .
We also impose locality on the measurement. Specifically, we represent the measurement operator corresponding to outcome k k k by a depth-d M d_{\mathrm{M}} d M local measurement circuit:
M k = m k d M d M m k d M − 1 d M − 1 ⋯ m k 1 1 , (4) M_k = m^{d_{\mathrm{M}}}_{k_{d_{\mathrm{M}}}}m^{d_{\mathrm{M}}-1}_{k_{d_{\mathrm{M}}-1}}\cdots m^1_{k_1}, \tag*{(4)} M k = m k d M d M m k d M − 1 d M − 1 ⋯ m k 1 1 , (4)
m k n n = ⨂ Y ∈ F n m k n , Y n , Y , (5) m^{n}_{k_n} = \bigotimes_{Y\in\mathcal{F}_n}m^{n,Y}_{k_{n,Y}}, \tag*{(5)} m k n n = Y ∈ F n ⨂ m k n , Y n , Y , (5)
where k = ( k 1 , … , k d M ) \boldsymbol{k}=(k_1,\ldots,k_{d_{\mathrm{M}}}) k = ( k 1 , … , k d M ) denotes the full record of local measurement outcomes with k n = ( k n , Y ) Y ∈ F n k_n=(k_{n,Y})_{Y\in\mathcal{F}_n} k n = ( k n , Y ) Y ∈ F n collecting the outcomes in the n n n th measurement layer. Here, F n \mathcal{F}_n F n denotes a set of mutually disjoint local regions Y Y Y , on which m k n , Y n , Y m^{n,Y}_{k_{n,Y}} m k n , Y n , Y is supported. We assume that a measurement represented in this form is low-depth and local, i.e., ( D − 1 ) d M = O ( L s ) (\mathcal{D}-1)d_{\mathrm{M}}=O(L^s) ( D − 1 ) d M = O ( L s ) , where D = max n max Y ∈ F n diam ( Y ) + 1 \mathcal{D}=\max_n\max_{Y\in\mathcal{F}_n}\operatorname{diam}(Y)+1 D = max n max Y ∈ F n diam ( Y ) + 1 (see Supplemental Material for details [52 ] ).
We further impose locality on the dependence of the feedback Hamiltonian on measurement outcomes. Physically, this means that the classical communication is spatially local. We write H k ( u ) = ∑ Z h k ( Z ) , Z ( u ) H_k(u)=\sum_Z h_{k^{(Z)},Z}(u) H k ( u ) = ∑ Z h k ( Z ) , Z ( u ) , where each term is supported on Z Z Z and k ( Z ) k^{(Z)} k ( Z ) collects the measurement outcomes on which that term depends. Our requirement means that k ( Z ) k^{(Z)} k ( Z ) is a set of k n , Y k_{n,Y} k n , Y satisfying
dist ( Z , Y ) ≤ ℓ , ℓ = O ( L s ) . (6) \operatorname{dist}(Z,Y)\le\ell,\qquad\ell=O(L^s). \tag*{(6)} dist ( Z , Y ) ≤ ℓ , ℓ = O ( L s ) . (6)
Here, dist ( Z , Y ) : = min z ∈ Z , y ∈ Y dist ( z , y ) \operatorname{dist}(Z,Y):=\min_{z\in Z,y\in Y}\operatorname{dist}(z,y) dist ( Z , Y ) := min z ∈ Z , y ∈ Y dist ( z , y ) .
Under these conditions, apply the same measurement-and-feedback protocol to ρ i e n s \rho_i^{\mathrm{ens}} ρ i ens and ρ i M I T E \rho_i^{\mathrm{MITE}} ρ i MITE . Then we can prove that
W ≃ W e n s (7) W\simeq W_{\mathrm{ens}} \tag*{(7)} W ≃ W ens (7)
holds for arbitrary β \beta β (including β = 0 \beta=0 β = 0 ). The rigorous formulation of the theorem and its proof are given in Supplemental Material [52 ] , where we apply the Lieb-Robinson bound [53 , 54 , 55 ] to the Hamiltonian in the Heisenberg picture. We emphasize that extending the unitary-only setting [33 , 42 ] to measurement and feedback is nontrivial, since even a local measurement can induce a nonlocal state update at the level of individual quantum trajectories. Nevertheless, under the above constraints, the extracted work agrees with that obtained from the corresponding canonical ensemble.
Optimized extractable work.— We next consider feedback operations optimized to maximize the extractable work. For fixed U 1 U_1 U 1 and measurement operators { M k } \{M_k\} { M k } , we denote the optimal work extractable from the canonical initial state under arbitrary unitary feedback by W e n s o p t W_{\mathrm{ens}}^{\mathrm{opt}} W ens opt . By definition, W e n s ≤ W e n s o p t W_{\mathrm{ens}}\le W_{\mathrm{ens}}^{\mathrm{opt}} W ens ≤ W ens opt holds. Equation (7) holds for every protocol satisfying the locality and time constraints stated above, including Eq. (6) . Therefore,
W \simeq W_{\mathrm{ens}}\le W_{\mathrm{ens}}^{\mathrm{opt}}.
{#eq:8}
Thus, although MITE may be pure and nonpassive, the work extractable under these constraints does not exceed the optimal work extractable from the corresponding canonical ensemble with the same U 1 U_1 U 1 and measurement operators.
Two remarks are in order. First, this argument requires all of the above constraints. If longer operation times, nonlocal Hamiltonians, or dependence on distant measurement outcomes (i.e., nonlocal classical communication) are allowed, the equivalence in Eq. (7) is no longer guaranteed. As discussed later, we indeed demonstrate the last possibility leads to additional work gain from long-range correlations in an entangled pure MITE state. Second, the equality condition in (8) is achieved for a certain setup, as discussed later.
Information-thermodynamic identity.— We now show the second equality in Eq. (2) :
β W e n s = I Q C − χ − D ( ρ f e n s ∥ ρ f e n s ) . (8) \beta W_{\mathrm{ens}}=I_{\mathrm{QC}}-\chi-D(\rho_f^{\mathrm{ens}}\Vert\rho_f^{\mathrm{ens}}). \tag*{(8)} β W ens = I QC − χ − D ( ρ f ens ∥ ρ f ens ) . (8)
This identity does not require any of the locality or operational-time constraints, including the locality of feedback dependence on measurement outcomes. Here, we have introduced the quantum relative entropy D ( ρ ∥ σ ) = Tr [ ρ ( log ρ − log σ ) ] D(\rho\Vert\sigma)=\operatorname{Tr}[\rho(\log\rho-\log\sigma)] D ( ρ ∥ σ ) = Tr [ ρ ( log ρ − log σ )] . In particular, β − 1 D ( ρ f e n s ∥ ρ f e n s ) \beta^{-1}D(\rho_f^{\mathrm{ens}}\Vert\rho_f^{\mathrm{ens}}) β − 1 D ( ρ f ens ∥ ρ f ens ) serves as the nonequilibrium free energy of ρ f e n s \rho_f^{\mathrm{ens}} ρ f ens [56 ] . We have also defined the quantum-classical (QC) mutual information [5 , 44 , 45 , 46 ] and the Holevo information [47 , 48 , 49 ] by using the von Neumann entropies S ( ρ ) = − Tr [ ρ log ρ ] S(\rho)=-\operatorname{Tr}[\rho\log\rho] S ( ρ ) = − Tr [ ρ log ρ ] of the branch states ρ k ′ e n s : = U k M k U 1 ρ i e n s U 1 † M k † U k † / P k e n s \rho_k'{}^{\mathrm{ens}}:=U_kM_kU_1\rho_i^{\mathrm{ens}}U_1^\dagger M_k^\dagger U_k^\dagger/P_k^{\mathrm{ens}} ρ k ′ ens := U k M k U 1 ρ i ens U 1 † M k † U k † / P k ens :
I Q C : = S ( ρ i e n s ) − ∑ k P k e n s S ( ρ k ′ e n s ) , (9) I_{\mathrm{QC}}:=S(\rho_i^{\mathrm{ens}})-\sum_k P_k^{\mathrm{ens}}S(\rho_k'{}^{\mathrm{ens}}), \tag*{(9)} I QC := S ( ρ i ens ) − k ∑ P k ens S ( ρ k ′ ens ) , (9)
χ : = S ( ρ f e n s ) − ∑ k P k e n s S ( ρ k e n s ) . (10) \chi:=S(\rho_f^{\mathrm{ens}})-\sum_k P_k^{\mathrm{ens}}S(\rho_k^{\mathrm{ens}}). \tag*{(10)} χ := S ( ρ f ens ) − k ∑ P k ens S ( ρ k ens ) . (10)
Because the feedback unitaries preserve the entropy of each conditional branch, I Q C I_{\mathrm{QC}} I QC is independent of { U k } \{U_k\} { U k } and quantifies the average entropy reduction achieved by conditioning on the measurement outcome, namely, the information acquired by the measurement. By contrast, χ \chi χ quantifies how much information about the measurement outcome remains encoded in the final state after feedback. In particular, χ = 0 \chi=0 χ = 0 if all postfeedback branches are identical, whereas χ \chi χ becomes large when different outcomes lead to more distinguishable final states. Also, I Q C − χ I_{\mathrm{QC}}-\chi I QC − χ represents the net entropy reduction.
Equation (8) is therefore a thermodynamic identity for a closed quantum many-body system, in which no heat bath is involved. It shows that the information I Q C I_{\mathrm{QC}} I QC acquired by the measurement is not converted directly into work. The extractable work is necessarily reduced by the Holevo information χ \chi χ and the relative entropy D ( ρ f e n s ∥ ρ i e n s ) D(\rho_f^{\mathrm{ens}}\Vert\rho_i^{\mathrm{ens}}) D ( ρ f ens ∥ ρ i ens ) (i.e., β W e n s < I Q C \beta W_{\mathrm{ens}}<I_{\mathrm{QC}} β W ens < I QC if W e n s > 0 W_{\mathrm{ens}}>0 W ens > 0 ). This differs from the setting of Ref. [5 ] , which considers contact with a heat bath. A detailed derivation of Eq. (8) , as well as the information-thermodynamic meaning of the optimal extractable work W e n s o p t W_{\mathrm{ens}}^{\mathrm{opt}} W ens opt based on Eq. (8) , is given in End Matter.
Example.— We now explicitly calculate the extractable work and illustrate the physical meaning of Eq. (7) for a simple setup. We consider a noninteracting V V V -qubit system H = g ∑ j = 1 V σ j 3 H=g\sum_{j=1}^{V}\sigma_j^3 H = g ∑ j = 1 V σ j 3 (g > 0 g>0 g > 0 ), where σ j 3 \sigma_j^3 σ j 3 is the Pauli operator acting only on site j j j : σ j 3 = σ 0 ⊗ σ 0 ⊗ ⋯ ⊗ σ 0 ⊗ σ 3 ⊗ σ 0 ⊗ ⋯ ⊗ σ 0 \sigma_j^3=\sigma^0\otimes\sigma^0\otimes\cdots\otimes\sigma^0\otimes\sigma^3\otimes\sigma^0\otimes\cdots\otimes\sigma^0 σ j 3 = σ 0 ⊗ σ 0 ⊗ ⋯ ⊗ σ 0 ⊗ σ 3 ⊗ σ 0 ⊗ ⋯ ⊗ σ 0 . The corresponding canonical ensemble is given by ρ i e n s = [ e − β g σ 3 2 cosh ( β g ) ] ⊗ V \rho_i^{\mathrm{ens}}=\left[\frac{e^{-\beta g\sigma^3}}{2\cosh(\beta g)}\right]^{\otimes V} ρ i ens = [ 2 c o s h ( β g ) e − β g σ 3 ] ⊗ V . We set the premeasurement unitary to U 1 = 1 U_1=1 U 1 = 1 . We perform the noisy local measurement on all qubits, whose strength is parametrized by η \eta η (0 ≤ η ≤ 1 0\le\eta\le1 0 ≤ η ≤ 1 ) [57 ] :
M k = ⨂ j = 1 V m k j ( η ) , m k ( η ) = σ 0 + k η σ 3 2 ( 1 + η 2 ) , k = ± 1. (11) M_{\boldsymbol{k}}=\bigotimes_{j=1}^{V}m_{k_j}(\eta),\qquad m_k(\eta)=\frac{\sigma^0+k\eta\sigma^3}{\sqrt{2(1+\eta^2)}},\qquad k=\pm1. \tag*{(11)} M k = j = 1 ⨂ V m k j ( η ) , m k ( η ) = 2 ( 1 + η 2 ) σ 0 + k η σ 3 , k = ± 1. (11)
The variable k j k_j k j is the outcome at site j j j of the σ 3 \sigma^3 σ 3 measurement and takes the value + + + or − - − . The case η = 1 \eta=1 η = 1 corresponds to a projective measurement, whereas η = 0 \eta=0 η = 0 corresponds to a measurement that yields no information. Note that, in this case, the depth is d M = 1 d_M=1 d M = 1 , Y Y Y is a single site, and F n ( = 1 ) \mathcal{F}_{n(=1)} F n ( = 1 ) is the collection of all single-site regions.
We also restrict the feedback to product unitaries as U k = ⨂ j = 1 V u k j U_{\boldsymbol{k}}=\bigotimes_{j=1}^{V}u_{k_j} U k = ⨂ j = 1 V u k j , where u k j u_{k_j} u k j acts only on site j j j and depends only on the measurement outcome obtained at that site (Z Z Z is a single site, and ℓ = 0 \ell=0 ℓ = 0 ). We then define the maximally extractable work under these single-qubit operations as W l o c − o p t = max { u k j } W W^{\mathrm{loc-opt}}=\max_{\{u_{k_j}\}}W W loc − opt = max { u k j } W for MITE and likewise W e n s l o c − o p t W_{\mathrm{ens}}^{\mathrm{loc-opt}} W ens loc − opt for the canonical ensemble.
First, we find (see the Supplemental Material for details [52 ] )
W e n s l o c − o p t = V g max [ 0 , 2 η 1 + η 2 − tanh ( β g ) ] . (12) W_{\mathrm{ens}}^{\mathrm{loc-opt}}=Vg\max\left[0,\frac{2\eta}{1+\eta^2}-\tanh(\beta g)\right]. \tag*{(12)} W ens loc − opt = V g max [ 0 , 1 + η 2 2 η − tanh ( β g ) ] . (12)
Equation (12) shows that the maximum extractable work has a nonanalytic onset at a critical inverse temperature β c = g − 1 arctanh ( 2 η 1 + η 2 ) \beta_c=g^{-1}\operatorname{arctanh}\left(\frac{2\eta}{1+\eta^2}\right) β c = g − 1 arctanh ( 1 + η 2 2 η ) . This nonanalytic behavior arises because the optimal feedback operation switches at that boundary, which is regarded as a control-landscape transition [58 ] . We can also verify that Eq. (8) holds in this setting (see the Supplemental Material for details [52 ] ).
Next, consider work extraction from MITE. Since the measurements act on individual sites and the feedback at each site depends only on the measurement outcome obtained there, the protocol satisfies the locality conditions above. We can therefore apply Eq. (7) :
W l o c − o p t ≃ W e n s l o c − o p t . (13) W^{\mathrm{loc-opt}}\simeq W_{\mathrm{ens}}^{\mathrm{loc-opt}}. \tag*{(13)} W loc − opt ≃ W ens loc − opt . (13)
Thus, the work W l o c − o p t W^{\mathrm{loc-opt}} W loc − opt extractable from MITE also exhibits a nonanalytic onset at a parameter boundary, as in the canonical case, in the thermodynamic limit. Because the feedback Hamiltonian in this example contains only single-site terms, the corresponding Lieb-Robinson velocity is zero. Therefore, no condition need be imposed on the feedback time [52 ] .
Finally, since W e n s l o c − o p t ≤ W e n s o p t W_{\mathrm{ens}}^{\mathrm{loc-opt}}\le W_{\mathrm{ens}}^{\mathrm{opt}} W ens loc − opt ≤ W ens opt by definition, where W e n s o p t W_{\mathrm{ens}}^{\mathrm{opt}} W ens opt allows arbitrary feedback unitaries, we obtain
W l o c − o p t ≤ W e n s o p t (14) W^{\mathrm{loc-opt}}\le W_{\mathrm{ens}}^{\mathrm{opt}} \tag*{(14)} W loc − opt ≤ W ens opt (14)
from Eq. (13) . Here, we neglect the subextensive correction. This is a manifestation of Eq. (8) in the current setup. Interestingly, the equality condition for the inequality (14) is achieved with β = 0 \beta=0 β = 0 in our setting, as shown in End Matter. In that sense, inequality (8), which bounds the extractable work from MITE under these locality constraints by the optimal extractable work for the canonical ensemble, is tight.
We next relax the condition in Eq. (6) and demonstrate that feedback based on distant measurement outcomes can exploit long-range correlations hidden in MITE to extract an extensive amount of additional work relative to the corresponding canonical ensemble. We again consider the noninteracting Hamiltonian introduced above, assuming that the system is arranged in a one-dimensional chain V = L V=L V = L and L L L satisfies L = 3 n L=3n L = 3 n with n ∈ N n\in\mathbb{N} n ∈ N . As an initial state, we consider a product of GHZ triplets between distant sites:
∣ Ψ 0 ⟩ = ⨂ j = 1 n ∣ ↑ ↑ ↑ ⟩ j , j + n , j + 2 n + ∣ ↓ ↓ ↓ ⟩ j , j + n , j + 2 n 2 . (15) |\Psi_0\rangle=\bigotimes_{j=1}^{n}\frac{|\uparrow\uparrow\uparrow\rangle_{j,j+n,j+2n}+|\downarrow\downarrow\downarrow\rangle_{j,j+n,j+2n}}{\sqrt{2}}. \tag*{(15)} ∣ Ψ 0 ⟩ = j = 1 ⨂ n 2 ∣ ↑↑↑ ⟩ j , j + n , j + 2 n + ∣ ↓↓↓ ⟩ j , j + n , j + 2 n . (15)
This state satisfies the MITE condition with respect to the canonical ensemble ρ i e n s = 1 / 2 L \rho_i^{\mathrm{ens}}=1/2^L ρ i ens = 1/ 2 L at β = 0 \beta=0 β = 0 , while it has a long-range entanglement. Indeed, any region A A A with diam ( A ) < L / 3 \operatorname{diam}(A)<L/3 diam ( A ) < L /3 contains at most one site from each GHZ triplet, so its reduced density matrix is exactly 1 A / 2 ∣ A ∣ 1_A/2^{|A|} 1 A / 2 ∣ A ∣ .
We set U 1 = 1 U_1=1 U 1 = 1 and perform the noisy local measurement (11) on all sites. For each j j j , we look at the outcomes ( k j , k j + n , k j + 2 n ) (k_j,k_{j+n},k_{j+2n}) ( k j , k j + n , k j + 2 n ) , and if the number of + 1 +1 + 1 outcomes is larger than that of − 1 -1 − 1 outcomes among the three, we apply the unitary generated by σ j 3 + σ j + n 3 + σ j + 2 n 3 \sigma_j^3+\sigma_{j+n}^3+\sigma_{j+2n}^3 σ j 3 + σ j + n 3 + σ j + 2 n 3 . At t = π / 2 t=\pi/2 t = π /2 , the straightforward calculation leads to
W = g L ( 3 q − q 3 ) 2 , W e n s = g L q 2 , W e n s o p t = g L q , (16) W=\frac{gL(3q-q^3)}{2},\qquad W_{\mathrm{ens}}=\frac{gLq}{2},\qquad W_{\mathrm{ens}}^{\mathrm{opt}}=gLq, \tag*{(16)} W = 2 g L ( 3 q − q 3 ) , W ens = 2 g L q , W ens opt = g L q , (16)
where q = 2 η 1 + η 2 q = \frac{2\eta}{1+\eta^2} q = 1 + η 2 2 η . Note that while W e n s W_{\mathrm{ens}} W ens follows the same feedback protocol as the MITE, W e n s o p t W_{\mathrm{ens}}^{\mathrm{opt}} W ens opt allows arbitrary feedback protocols from the canonical ensemble. We have W > W e n s o p t > W e n s W > W_{\mathrm{ens}}^{\mathrm{opt}} > W_{\mathrm{ens}} W > W ens opt > W ens for 0 < η < 1 0 < \eta< 1 0 < η < 1 . Therefore, the measurement-and-feedback protocol extracts an extensive amount of additional work from ∣ Ψ 0 ⟩ \lvert\Psi_0\rangle ∣ Ψ 0 ⟩ compared with the corresponding canonical ensemble, exploiting the hidden long-range correlations associated with the entanglement structure of this pure MITE state. This example shows that Eq. (7) can fail even for constant-time feedback generated by on-site terms if the feedback depends on distant measurement outcomes. In the End Matter, we summarize the relation between the work extraction and the operational time.
Conclusion and outlook
We have discussed information thermodynamics from possibly pure thermal states by measurement and feedback operations in closed quantum many-body systems, when the measurements are local and the unitary controls are generated by local Hamiltonians over finite times (for s = 0 s = 0 s = 0 ). We have first shown that, when each local feedback term depends only on nearby measurement outcomes, the work extractable from MITE is approximated by that obtained by applying the same protocol to a corresponding canonical ensemble. We have then expressed the extractable work in terms of the information-theoretical quantities.
For a noninteracting V V V -qubit system, we have demonstrated that the maximum work W l o c − o p t W_{\mathrm{loc-opt}} W loc − opt extractable from MITE under these locality constraints is bounded by the optimal extractable work W e n s o p t W_{\mathrm{ens}}^{\mathrm{opt}} W ens opt for a canonical initial state, and that the equality condition is achieved at the infinite temperature. We have also shown that long-range correlations in a pure MITE state can be exploited through feedback based on distant measurement outcomes, yielding an extensive amount of additional work relative to the corresponding canonical ensemble.
There are several future directions. First, it would be interesting to investigate how to relate the additional work gain from the quantum correlation to the known entanglement measures. It would also be interesting to investigate how recent LOCC-based approaches to characterizing many-body quantum correlations through work extraction [16 , 17 ] relate to our framework in interacting closed quantum many-body systems. Another intriguing direction is to extend our results to repeated and continuous measurement and feedback [59 , 60 , 61 ] . Finally, it is of importance to investigate the nature of control-landscape transitions induced by measurement and feedback, found in noninteracting systems in this work, in more complicated interacting systems.
Acknowledgments. We would like to thank Ken Funo, Yuuya Chiba, and Kohei Yoshimura for fruitful discussions. OpenAI’s ChatGPT (GPT-6) was used to find the example of the work extraction from GHZ states and to assist with the development and checking of parts of the derivation of the MITE–canonical equivalence. The AI-assisted mathematical arguments included in the manuscript were critically reviewed and independently verified by the authors. We are supported by JST ERATO Grant No. JPMJER2302, Japan. S.S. was supported by KAKENHI Grant No. JP26K17047. R.H. was supported by KAKENHI Grant No. JP24K16982.
References
End Matter
Derivation of the information-thermodynamic identity from the canonical state.— We give the details of the derivation of Eq. (8) . First, the relative entropy between the final state ρ f e n s \rho_f^{\mathrm{ens}} ρ f ens and the initial state ρ i e n s \rho_i^{\mathrm{ens}} ρ i ens is rewritten as
D ( ρ f e n s ∥ ρ i e n s ) = Tr [ ρ f e n s log ρ f e n s ] − Tr [ ρ f e n s log ρ i e n s ] = − S ( ρ f e n s ) − Tr [ ρ f e n s ( − β H − log Z ) ] = − S ( ρ f e n s ) + β Tr [ H ρ f e n s ] + log Z . (17) D(\rho_f^{\mathrm{ens}}\Vert\rho_i^{\mathrm{ens}})=\operatorname{Tr}[\rho_f^{\mathrm{ens}}\log\rho_f^{\mathrm{ens}}]-\operatorname{Tr}[\rho_f^{\mathrm{ens}}\log\rho_i^{\mathrm{ens}}]
=-S(\rho_f^{\mathrm{ens}})-\operatorname{Tr}[\rho_f^{\mathrm{ens}}(-\beta H-\log Z)]
=-S(\rho_f^{\mathrm{ens}})+\beta\operatorname{Tr}[H\rho_f^{\mathrm{ens}}]+\log Z. \tag*{(17)} D ( ρ f ens ∥ ρ i ens ) = Tr [ ρ f ens log ρ f ens ] − Tr [ ρ f ens log ρ i ens ] = − S ( ρ f ens ) − Tr [ ρ f ens ( − β H − log Z )] = − S ( ρ f ens ) + β Tr [ H ρ f ens ] + log Z . (17)
Thus,
β Tr [ H ρ f e n s ] = D ( ρ f e n s ∥ ρ i e n s ) + S ( ρ f e n s ) − log Z . (18) \beta\operatorname{Tr}[H\rho_f^{\mathrm{ens}}]=D(\rho_f^{\mathrm{ens}}\Vert\rho_i^{\mathrm{ens}})+S(\rho_f^{\mathrm{ens}})-\log Z. \tag*{(18)} β Tr [ H ρ f ens ] = D ( ρ f ens ∥ ρ i ens ) + S ( ρ f ens ) − log Z . (18)
Moreover,
β Tr [ H ρ i e n s ] = − log Z + S ( ρ i e n s ) . (19) \beta\operatorname{Tr}[H\rho_i^{\mathrm{ens}}]=-\log Z+S(\rho_i^{\mathrm{ens}}). \tag*{(19)} β Tr [ H ρ i ens ] = − log Z + S ( ρ i ens ) . (19)
From Eqs. (18) and (19) , the extractable work (1) from the canonical state is
β W e n s = S ( ρ i e n s ) − S ( ρ f e n s ) − D ( ρ f e n s ∥ ρ i e n s ) = S ( ρ i e n s ) − ∑ k P k e n s S ( ρ k ′ e n s ) − [ S ( ρ f e n s ) − ∑ k P k e n s S ( ρ k ′ e n s ) ] − D ( ρ f e n s ∥ ρ i e n s ) = I Q C − χ − D ( ρ f e n s ∥ ρ i e n s ) . (20) \beta W_{\mathrm{ens}}=S(\rho_i^{\mathrm{ens}})-S(\rho_f^{\mathrm{ens}})-D(\rho_f^{\mathrm{ens}}\Vert\rho_i^{\mathrm{ens}})
=S(\rho_i^{\mathrm{ens}})-\sum_k P_k^{\mathrm{ens}}S(\rho_k^{\prime\mathrm{ens}})
-\left[S(\rho_f^{\mathrm{ens}})-\sum_k P_k^{\mathrm{ens}}S(\rho_k^{\prime\mathrm{ens}})\right]-D(\rho_f^{\mathrm{ens}}\Vert\rho_i^{\mathrm{ens}})
=I_{\mathrm{QC}}-\chi-D(\rho_f^{\mathrm{ens}}\Vert\rho_i^{\mathrm{ens}}). \tag*{(20)} β W ens = S ( ρ i ens ) − S ( ρ f ens ) − D ( ρ f ens ∥ ρ i ens ) = S ( ρ i ens ) − k ∑ P k ens S ( ρ k ′ ens ) − [ S ( ρ f ens ) − k ∑ P k ens S ( ρ k ′ ens ) ] − D ( ρ f ens ∥ ρ i ens ) = I QC − χ − D ( ρ f ens ∥ ρ i ens ) . (20)
Thus, it can be expressed in terms of the quantum–classical (QC) mutual information (9) , the Holevo information (10) , and the relative entropy. Note that this equality becomes trivial (right- and left-hand sides become zero) for β = 0 \beta=0 β = 0 , although Eq. (7) nontrivially holds even for this case.
Information theoretical meaning of W e n s o p t W_{\mathrm{ens}}^{\mathrm{opt}} W ens opt .— Equation (8) also clarifies the information-thermodynamic meaning of the optimal extractable work W e n s o p t W_{\mathrm{ens}}^{\mathrm{opt}} W ens opt . We denote the conditional state immediately after the measurement and before feedback by ρ ~ k e n s = M k U 1 ρ i e n s U 1 † M k † / P k e n s \widetilde{\rho}_k^{\mathrm{ens}}=M_kU_1\rho_i^{\mathrm{ens}}U_1^\dagger M_k^\dagger/P_k^{\mathrm{ens}} ρ k ens = M k U 1 ρ i ens U 1 † M k † / P k ens . Let H ∣ E α ⟩ = E α ∣ E α ⟩ H|E_\alpha\rangle=E_\alpha|E_\alpha\rangle H ∣ E α ⟩ = E α ∣ E α ⟩ with E 1 ≤ E 2 ≤ ⋯ E_1\le E_2\le\cdots E 1 ≤ E 2 ≤ ⋯ , and let r k , 1 ≥ r k , 2 ≥ ⋯ r_{k,1}\ge r_{k,2}\ge\cdots r k , 1 ≥ r k , 2 ≥ ⋯ be the eigenvalues of ρ ~ k e n s \widetilde{\rho}_k^{\mathrm{ens}} ρ k ens . Since a feedback unitary preserves these eigenvalues, the branch energy is minimized by assigning the largest eigenvalues to the lowest energy levels. Thus, if arbitrary feedback unitaries are allowed, the optimal postfeedback branch is ρ k ′ e n s , o p t = ∑ α r k , α ∣ E α ⟩ ⟨ E α ∣ \rho_k^{\prime\mathrm{ens,opt}}=\sum_\alpha r_{k,\alpha}|E_\alpha\rangle\langle E_\alpha| ρ k ′ ens , opt = ∑ α r k , α ∣ E α ⟩ ⟨ E α ∣ . This state is passive. Applying Eq. (8) to this optimal feedback gives
β W e n s ≤ β W e n s o p t = I Q C − χ o p t − D ( ρ f e n s , o p t ∥ ρ i e n s ) (21) \beta W_{\mathrm{ens}}\le\beta W_{\mathrm{ens}}^{\mathrm{opt}}=I_{\mathrm{QC}}-\chi^{\mathrm{opt}}-D(\rho_f^{\mathrm{ens,opt}}\Vert\rho_i^{\mathrm{ens}}) \tag*{(21)} β W ens ≤ β W ens opt = I QC − χ opt − D ( ρ f ens , opt ∥ ρ i ens ) (21)
with ρ f e n s , o p t = ∑ k P k e n s ρ k ′ e n s , o p t \rho_f^{\mathrm{ens,opt}}=\sum_kP_k^{\mathrm{ens}}\rho_k^{\prime\mathrm{ens,opt}} ρ f ens , opt = ∑ k P k ens ρ k ′ ens , opt . Here, I Q C I_{\mathrm{QC}} I QC is unchanged by the feedback optimization, whereas χ o p t \chi^{\mathrm{opt}} χ opt and ρ f e n s , o p t \rho_f^{\mathrm{ens,opt}} ρ f ens , opt are evaluated for the optimal feedback. Hence, Eq. (8) provides an information-thermodynamic decomposition of the optimal extractable work in a closed quantum system. Interestingly, I Q C I_{\mathrm{QC}} I QC , χ o p t \chi^{\mathrm{opt}} χ opt and D ( ρ f e n s , o p t ∥ ρ i e n s ) D(\rho_f^{\mathrm{ens,opt}}\Vert\rho_i^{\mathrm{ens}}) D ( ρ f ens , opt ∥ ρ i ens ) are expressed by classical populations (e.g., r k , n r_{k,n} r k , n ) alone, since ρ f e n s , o p t \rho_f^{\mathrm{ens,opt}} ρ f ens , opt and ρ i e n s \rho_i^{\mathrm{ens}} ρ i ens are both diagonalized in the energy basis.
Equality condition in Eq. (14) .— Here, we show that the equality condition in Eq. (14) is achieved for β = 0 \beta=0 β = 0 in the setup of noninteracting V V V qubits. Since W l o c − o p t ≃ W e n s l o c − o p t W_{\mathrm{loc-opt}}\simeq W_{\mathrm{ens}}^{\mathrm{loc-opt}} W loc − opt ≃ W ens loc − opt , it is enough to show W e n s l o c − o p t = W e n s o p t W_{\mathrm{ens}}^{\mathrm{loc-opt}}=W_{\mathrm{ens}}^{\mathrm{opt}} W ens loc − opt = W ens opt . For this purpose, we will show that the product feedback operation described by U k = ⨂ j = 1 V u k j U_k=\bigotimes_{j=1}^{V}u_{k j} U k = ⨂ j = 1 V u k j already makes every conditional branch passive at β = 0 \beta=0 β = 0 . Then, allowing feedback operations that create correlations between qubits cannot lower the energy any further.
The initial canonical state is the infinite-temperature state, ρ i e n s = ( σ 0 / 2 ) ⊗ V \rho_i^{\mathrm{ens}}=(\sigma^0/2)^{\otimes V} ρ i ens = ( σ 0 /2 ) ⊗ V . After obtaining the single-qubit outcome k = ± 1 k=\pm1 k = ± 1 , the normalized postmeasurement state for each qubit reads
m k ( η ) ( σ 0 / 2 ) m k † ( η ) Tr [ m k ( η ) ( σ 0 / 2 ) m k † ( η ) ] = σ 0 + k q σ 3 2 , q = 2 η 1 + η 2 . (22) \frac{m_k(\eta)(\sigma^0/2)m_k^\dagger(\eta)}{\operatorname{Tr}[m_k(\eta)(\sigma^0/2)m_k^\dagger(\eta)]}=\frac{\sigma^0+kq\sigma^3}{2},\qquad q=\frac{2\eta}{1+\eta^2}. \tag*{(22)} Tr [ m k ( η ) ( σ 0 /2 ) m k † ( η )] m k ( η ) ( σ 0 /2 ) m k † ( η ) = 2 σ 0 + k q σ 3 , q = 1 + η 2 2 η . (22)
Equation (22) shows that, conditioned on the outcome, the measurement prepares a qubit biased along the σ 3 \sigma^3 σ 3 axis: the k = + 1 k=+1 k = + 1 branch is biased toward σ 3 = + 1 \sigma^3=+1 σ 3 = + 1 , whereas the k = − 1 k=-1 k = − 1 branch is biased toward σ 3 = − 1 \sigma^3=-1 σ 3 = − 1 . The locally optimal feedback at β = 0 \beta=0 β = 0 , u − = σ 0 u_-=\sigma^0 u − = σ 0 and u + = σ 1 u_+=\sigma^1 u + = σ 1 , then uses the outcome only to orient this bias toward the lower-energy direction:
u k ( σ 0 + k q σ 3 2 ) u k † = σ 0 − q σ 3 2 . (23) u_k\left(\frac{\sigma^0+kq\sigma^3}{2}\right)u_k^\dagger=\frac{\sigma^0-q\sigma^3}{2}. \tag*{(23)} u k ( 2 σ 0 + k q σ 3 ) u k † = 2 σ 0 − q σ 3 . (23)
Consequently, for every outcome k k k , the normalized state after measurement and feedback is
ρ k ′ e n s = ( σ 0 − q σ 3 2 ) ⊗ V = exp [ − a ∑ j = 1 V σ j 3 ] ( 2 cosh a ) V = exp [ − ( a / a ) H ] ( 2 cosh a ) V , (24) \rho_k^{\prime\mathrm{ens}}=\left(\frac{\sigma^0-q\sigma^3}{2}\right)^{\otimes V}
=\frac{\exp[-a\sum_{j=1}^{V}\sigma_j^3]}{(2\cosh a)^V}
=\frac{\exp[-(a/a)H]}{(2\cosh a)^V}, \tag*{(24)} ρ k ′ ens = ( 2 σ 0 − q σ 3 ) ⊗ V = ( 2 cosh a ) V exp [ − a ∑ j = 1 V σ j 3 ] = ( 2 cosh a ) V exp [ − ( a / a ) H ] , (24)
where a = arctanh q a=\operatorname{arctanh}q a = arctanh q . Equation (24) shows that each conditional branch is already a canonical state of the full Hamiltonian at a nonnegative inverse temperature a / q a/q a / q . Therefore it is passive. Any further unitary feedback, including an entangling one, cannot lower the energy once the branch has already been put in this passive state. Hence the equality in Eq. (14) holds when β = 0 \beta=0 β = 0 .
Summary of the operational time and extractable work.— In Table 1, we summarize the relation between the operational time and extractable work from MITE with s = 0 s=0 s = 0 for different protocols, i.e., (a) protocol with no measurement and feedback, (b) feedback-based protocol using nearby measurement outcomes, and (c) feedback-based protocol using distant measurement outcomes. We assume that the unitary operations are generated by local Hamiltonians, and, if any, measurement is composed of local gates. We consider two regimes, i.e., constant times (e.g., t = O ( L 0 ) t = O(L^{0}) t = O ( L 0 ) ) and longer times (e.g., t = O ( L ) t = O(L) t = O ( L ) ). In the case (b), we can obtain positive extensive work prohibited in the case (a) even for constant times, although it is still bounded by the optimal work predicted by the canonical ensemble. In case (c), we can use long-range correlation to obtain additional work extensively larger than that from the canonical ensemble, even for the constant times.
Table I. Summary of the operational time and extractable work from the MITE in closed quantum many-body systems.
Supplemental Material
Derivation of the MITE–canonical equivalence
We give the details of the derivation of Eq. (7) . Assume that the initial state is a MITE state ρ i M I T E \rho_i^{\mathrm{MITE}} ρ i MITE and that the Hamiltonian H = ∑ X h X H=\sum_X h_X H = ∑ X h X is a sum of local observables. Then, for the extractable work
W = Tr [ H ρ i M I T E ] − Tr [ H ρ f ] , (25) W=\operatorname{Tr}[H\rho_i^{\mathrm{MITE}}]-\operatorname{Tr}[H\rho_f], \tag*{(25)} W = Tr [ H ρ i MITE ] − Tr [ H ρ f ] , (25)
the first term satisfies
Tr [ H ρ i M I T E ] ≃ Tr [ H ρ i e n s ] (26) \operatorname{Tr}[H\rho_i^{\mathrm{MITE}}]\simeq\operatorname{Tr}[H\rho_i^{\mathrm{ens}}] \tag*{(26)} Tr [ H ρ i MITE ] ≃ Tr [ H ρ i ens ] (26)
by the definition of MITE. The nontrivial contribution is the second term, namely, the final-state energy. For the initial state ρ i M I T E \rho_i^{\mathrm{MITE}} ρ i MITE , the second term in Eq. (25) can be written as
Tr [ H ρ f ] = ∑ X , k Tr [ ρ i M I T E U 1 † M k † U k † ( t ) h X U k ( t ) M k U 1 ] . (27) \operatorname{Tr}[H\rho_f]=\sum_{X,k}\operatorname{Tr}\left[\rho_i^{\mathrm{MITE}}U_1^\dagger M_k^\dagger U_k^\dagger(t)h_XU_k(t)M_kU_1\right]. \tag*{(27)} Tr [ H ρ f ] = X , k ∑ Tr [ ρ i MITE U 1 † M k † U k † ( t ) h X U k ( t ) M k U 1 ] . (27)
Because ρ i M I T E \rho_i^{\mathrm{MITE}} ρ i MITE is MITE, if ∑ k U 1 † M k † U k † ( t ) h X U k ( t ) M k U 1 \sum_k U_1^\dagger M_k^\dagger U_k^\dagger(t)h_XU_k(t)M_kU_1 ∑ k U 1 † M k † U k † ( t ) h X U k ( t ) M k U 1 is approximated by an O ( L S ) \mathcal{O}(L^S) O ( L S ) -local observable, then
∑ k ∑ X Tr [ ρ i M I T E U 1 † M k † U k † ( t ) h X U k ( t ) M k U 1 ] ≃ ∑ k ∑ X Tr [ ρ i e n s U 1 † M k † U k † ( t ) h X U k ( t ) M k U 1 ] . (28) \sum_k\sum_X\operatorname{Tr}\left[\rho_i^{\mathrm{MITE}}U_1^\dagger M_k^\dagger U_k^\dagger(t)h_XU_k(t)M_kU_1\right]
\simeq
\sum_k\sum_X\operatorname{Tr}\left[\rho_i^{\mathrm{ens}}U_1^\dagger M_k^\dagger U_k^\dagger(t)h_XU_k(t)M_kU_1\right]. \tag*{(28)} k ∑ X ∑ Tr [ ρ i MITE U 1 † M k † U k † ( t ) h X U k ( t ) M k U 1 ] ≃ k ∑ X ∑ Tr [ ρ i ens U 1 † M k † U k † ( t ) h X U k ( t ) M k U 1 ] . (28)
This establishes Eq. (7) . It is therefore sufficient to show that
∑ k U 1 † M k † U k † ( t ) h X U k ( t ) M k U 1 (29) \sum_k U_1^\dagger M_k^\dagger U_k^\dagger(t)h_XU_k(t)M_kU_1 \tag*{(29)} k ∑ U 1 † M k † U k † ( t ) h X U k ( t ) M k U 1 (29)
is well approximated by an O ( L s ) \mathcal{O}(L^{s}) O ( L s ) -local observable. Although this operator appears nonlocal, the Lieb–Robinson bound controls the error in a local approximation, as explained below (see also Figure S1 ). In the following, we assume that H 1 ( t ) H_1(t) H 1 ( t ) and H k ( t ) H_k(t) H k ( t ) are time-independent for simplicity. Note that the extension to time-dependent cases is straightforward; as pointed out in note [12 ] of Ref. [54 ] , the Lieb–Robinson bound also applies to time-dependent local Hamiltonians, provided that the local interaction strengths are uniformly bounded.
Figure S1. Schematic illustration of operator spreading. (a) A local operator h X h_X h X , initially supported on X X X , spreads under the feedback unitary U k ( t ) U_k(t) U k ( t ) , but the Lieb–Robinson bound allows it to be approximated by h k , X [ l ] ( t ) h_{k,X}^{[l]}(t) h k , X [ l ] ( t ) , defined using the truncated Hamiltonian H k l H_k^l H k l and supported on X l X^l X l , with an exponentially small error when l − r 0 − v t l-r_0-vt l − r 0 − v t is sufficiently large. (b) Under the locality condition in Eq. (32) , summing over the outcomes of the depth-d M d_M d M local measurement circuit M k M_k M k gives G X [ l ] ( t ) : = ∑ k M k † h k , X [ l ] ( t ) M k G_X^{[l]}(t) := \sum_k M_k^\dagger h_{k,X}^{[l]}(t)M_k G X [ l ] ( t ) := ∑ k M k † h k , X [ l ] ( t ) M k , supported on a region supp ( G X [ l ] ( t ) ) ⊆ X R M \operatorname{supp}(G_X^{[l]}(t)) \subseteq X^{R_M} supp ( G X [ l ] ( t )) ⊆ X R M , where R M = l + ℓ + ( D − 1 ) d M R_M = l+\ell+(D-1)d_M R M = l + ℓ + ( D − 1 ) d M . (c) Finally, the time evolution under the premeasurement unitary U 1 ( t ′ ) U_1(t') U 1 ( t ′ ) gives rise to an additional Lieb–Robinson light cone. Consequently, the outcome-averaged operator ∑ k U 1 † ( t ′ ) M k † U k † ( t ) h X U k ( t ) M k U 1 ( t ′ ) \sum_k U_1^\dagger(t')M_k^\dagger U_k^\dagger(t)h_XU_k(t)M_kU_1(t') ∑ k U 1 † ( t ′ ) M k † U k † ( t ) h X U k ( t ) M k U 1 ( t ′ ) can be approximated by an operator supported on X l ′ X^{l'} X l ′ , with an exponentially small error when both l − r 0 − v t l-r_0-vt l − r 0 − v t and l ′ − R M − v ′ t ′ l'-R_M-v't' l ′ − R M − v ′ t ′ are sufficiently large.
Feedback processes (Figure S1 (a))
In the following, we assume that the operator norm of h X h_X h X is bounded by a finite value independent of the system size V V V . First, we set
U k † ( t ) h X U k ( t ) : = h k , X ( t ) , (30) U_k^\dagger(t)h_XU_k(t) := h_{k,X}(t), \tag*{(30)} U k † ( t ) h X U k ( t ) := h k , X ( t ) , (30)
and evaluate
∣ ∑ k ∑ X Tr [ ρ i M I T E U 1 † M k † h k , X ( t ) M k U 1 ] − ∑ k ∑ X Tr [ ρ i M I T E U 1 † M k † h k , X [ l ] ( t ) M k U 1 ] ∣ . (31) \left|\sum_k\sum_X \operatorname{Tr}\left[\rho_i^{\mathrm{MITE}}U_1^\dagger M_k^\dagger h_{k,X}(t)M_kU_1\right]-\sum_k\sum_X \operatorname{Tr}\left[\rho_i^{\mathrm{MITE}}U_1^\dagger M_k^\dagger h_{k,X}^{[l]}(t)M_kU_1\right]\right|. \tag*{(31)} k ∑ X ∑ Tr [ ρ i MITE U 1 † M k † h k , X ( t ) M k U 1 ] − k ∑ X ∑ Tr [ ρ i MITE U 1 † M k † h k , X [ l ] ( t ) M k U 1 ] . (31)
To explain the definition of h k , X [ l ] ( t ) h_{k,X}^{[l]}(t) h k , X [ l ] ( t ) , we now write the feedback Hamiltonian as
H k = ∑ Z h ~ k ( Z ) , Z . H_k=\sum_Z\widetilde{h}_{k(Z),Z}. H k = Z ∑ h k ( Z ) , Z .
Each term h ~ k ( Z ) , Z \widetilde{h}_{k(Z),Z} h k ( Z ) , Z is a local operator supported on a region Z Z Z with diam ( Z ) ≤ r 0 \operatorname{diam}(Z)\le r_0 diam ( Z ) ≤ r 0 , which is independent of L L L . Also, h ~ k ( Z ) , Z \widetilde{h}_{k(Z),Z} h k ( Z ) , Z depends on k ( Z ) k^{(Z)} k ( Z ) , which denotes a set of the measurement outcomes k n , Y k_{n,Y} k n , Y associated with the region Z Z Z . We impose the following locality condition on k ( Z ) k^{(Z)} k ( Z ) , which indicates that the feedback Hamiltonian depends on nearby measurement outcomes. Specifically, to select k n , Y k_{n,Y} k n , Y in k ( Z ) k^{(Z)} k ( Z ) , we assume that the corresponding measurement region Y Y Y must contain at least one site y 0 y_0 y 0 such that
dist ( Z , y 0 ) ≤ ℓ , ℓ = O ( L s ) (32) \operatorname{dist}(Z,y_0)\le\ell,\qquad\ell=\mathcal{O}(L^s) \tag*{(32)} dist ( Z , y 0 ) ≤ ℓ , ℓ = O ( L s ) (32)
for a fixed s s s (0 ≤ s < 1 0\le s<1 0 ≤ s < 1 ).
We now define h k , X [ l ] ( t ) h_{k,X}^{[l]}(t) h k , X [ l ] ( t ) . Let X l : = { x : dist ( x , X ) < l } X^l:=\{x:\operatorname{dist}(x,X)<l\} X l := { x : dist ( x , X ) < l } with l > 0 l>0 l > 0 . We retain only the Hamiltonian terms whose regions Z Z Z are entirely contained in X l X^l X l :
H k X l : = ∑ Z ⊂ X l h ~ k ( Z ) , Z , H_k^{X^l}:=\sum_{Z\subset X^l}\widetilde{h}_{k(Z),Z}, H k X l := Z ⊂ X l ∑ h k ( Z ) , Z ,
U k X l ( t ) : = T exp [ − i ∫ 0 t H k X l d u ] . U_k^{X^l}(t):=\mathcal{T}\exp\left[-i\int_0^t H_k^{X^l}\,du\right]. U k X l ( t ) := T exp [ − i ∫ 0 t H k X l d u ] .
We use this restricted evolution to define the local approximation
h k , X [ l ] ( t ) : = ( U k X l ( t ) ) † h X U k X l ( t ) . (33) h_{k,X}^{[l]}(t):=\left(U_k^{X^l}(t)\right)^\dagger h_XU_k^{X^l}(t). \tag*{(33)} h k , X [ l ] ( t ) := ( U k X l ( t ) ) † h X U k X l ( t ) . (33)
Since every retained term is supported inside X l X^l X l , on the full Hilbert space we can write
H k X l = K k , X l ⊗ 1 X l ‾ , H_k^{X^l}=K_{k,X^l}\otimes\mathbb{1}_{\overline{X^l}}, H k X l = K k , X l ⊗ 1 X l ,
U k X l ( t ) = e − i t K k , X l ⊗ 1 X l ‾ , U_k^{X^l}(t)=e^{-itK_{k,X^l}}\otimes\mathbb{1}_{\overline{X^l}}, U k X l ( t ) = e − i t K k , X l ⊗ 1 X l ,
where K k , X l K_{k,X^l} K k , X l acts on the Hilbert space of X l X^l X l .
We now estimate the error in Eq. (31) . We define
B k , X : = h k , X ( t ) − h k , X [ l ] ( t ) . (34) B_{k,X}:=h_{k,X}(t)-h_{k,X}^{[l]}(t). \tag*{(34)} B k , X := h k , X ( t ) − h k , X [ l ] ( t ) . (34)
Then, Eq. (31) can be rewritten as follows:
Eq. ( S 7 ) = ∣ ∑ X Tr [ ∑ k M k U 1 ρ i M I T E U 1 † M k † B k , X ] ∣ = ∣ ∑ X Tr [ ∑ k P k M k U 1 ρ i M I T E U 1 † M k † P k B k , X ] ∣ , (35) \begin{aligned}
\operatorname{Eq.}\ (\mathrm{S7}) &= \left|\sum_X \operatorname{Tr}\left[\sum_k M_kU_1\rho_i^{\mathrm{MITE}}U_1^\dagger M_k^\dagger B_{k,X}\right]\right| \\
&= \left|\sum_X \operatorname{Tr}\left[\sum_k P_k\frac{M_kU_1\rho_i^{\mathrm{MITE}}U_1^\dagger M_k^\dagger}{P_k}B_{k,X}\right]\right|, \tag*{(35)}
\end{aligned} Eq. ( S7 ) = X ∑ Tr [ k ∑ M k U 1 ρ i MITE U 1 † M k † B k , X ] = X ∑ Tr [ k ∑ P k P k M k U 1 ρ i MITE U 1 † M k † B k , X ] , (35)
where
P k = Tr [ M k U 1 ρ i M I T E U 1 † M k † ] . (36) P_k=\operatorname{Tr}\left[M_kU_1\rho_i^{\mathrm{MITE}}U_1^\dagger M_k^\dagger\right]. \tag*{(36)} P k = Tr [ M k U 1 ρ i MITE U 1 † M k † ] . (36)
We further define
ρ ~ k : = M k U 1 ρ i M I T E U 1 † M k † P k . (37) \widetilde{\rho}_k:=\frac{M_kU_1\rho_i^{\mathrm{MITE}}U_1^\dagger M_k^\dagger}{P_k}. \tag*{(37)} ρ k := P k M k U 1 ρ i MITE U 1 † M k † . (37)
Then,
Eq. ( S 11 ) = ∣ ∑ X Tr [ ∑ k P k ρ ~ k B k , X ] ∣ = ∣ ∑ X ∑ k P k Tr [ ρ ~ k B k , X ] ∣ ≤ ∣ ∑ X max k ∣ Tr [ ρ ~ k B k , X ] ∣ ∣ ≤ ∑ X max k ∥ B k , X ∥ ∞ = ∑ X max k ∥ h k , X ( t ) − h k , X [ l ] ( t ) ∥ ∞ , (38) \begin{aligned}
\operatorname{Eq.}\ (\mathrm{S11}) &= \left|\sum_X\operatorname{Tr}\left[\sum_k P_k\widetilde{\rho}_kB_{k,X}\right]\right| \\
&= \left|\sum_X\sum_kP_k\operatorname{Tr}\left[\widetilde{\rho}_kB_{k,X}\right]\right| \\
&\leq\left|\sum_X\max_k\left|\operatorname{Tr}\left[\widetilde{\rho}_kB_{k,X}\right]\right|\right| \\
&\leq\sum_X\max_k\lVert B_{k,X}\rVert_\infty\\
&= \sum_X\max_k\lVert h_{k,X}(t)-h_{k,X}^{[l]}(t)\rVert_\infty, \tag*{(38)}
\end{aligned} Eq. ( S11 ) = X ∑ Tr [ k ∑ P k ρ k B k , X ] = X ∑ k ∑ P k Tr [ ρ k B k , X ] ≤ X ∑ k max ∣ Tr [ ρ k B k , X ] ∣ ≤ X ∑ k max ∥ B k , X ∥ ∞ = X ∑ k max ∥ h k , X ( t ) − h k , X [ l ] ( t ) ∥ ∞ , (38)
where ∥ ⋅ ∥ ∞ \lVert\cdot\rVert_\infty ∥ ⋅ ∥ ∞ denotes the operator norm. Since h k , X [ l ] ( u ) h_{k,X}^{[l]}(u) h k , X [ l ] ( u ) is supported on X l X^l X l , it commutes with every Hamiltonian term lying entirely outside X l X^l X l .
For the time-independent Hamiltonians H k X l H_k^{X^l} H k X l considered here, the Heisenberg equation gives
d d u h k , X [ l ] ( u ) = d d u [ U k X l ( u ) † h X U k X l ( u ) ] = i [ H k X l , ( U k X l ( u ) ) † h X U k X l ( u ) ] = ( ∵ U k X l ( t ) : = T exp [ − i ∫ 0 t H k X l d u ] ) = i [ H k X l , h k , X [ l ] ( u ) ] . \begin{aligned}
\frac{d}{du}h_{k,X}^{[l]}(u)
&=\frac{d}{du}\left[U_k^{X^l}(u)^\dagger h_XU_k^{X^l}(u)\right] \\
&=i\left[H_k^{X^l},\left(U_k^{X^l}(u)\right)^\dagger h_XU_k^{X^l}(u)\right] \\
&=\left(\because\ U_k^{X^l}(t):=\mathcal{T}\exp\left[-i\int_0^tH_k^{X^l}\,du\right]\right) \\
&=i\left[H_k^{X^l},h_{k,X}^{[l]}(u)\right].
\end{aligned} d u d h k , X [ l ] ( u ) = d u d [ U k X l ( u ) † h X U k X l ( u ) ] = i [ H k X l , ( U k X l ( u ) ) † h X U k X l ( u ) ] = ( ∵ U k X l ( t ) := T exp [ − i ∫ 0 t H k X l d u ] ) = i [ H k X l , h k , X [ l ] ( u ) ] .
Differentiating the following product therefore yields
d d u [ U k † ( t − u ) h k , X [ l ] ( u ) U k ( t − u ) ] = − i U k † ( t − u ) [ H k − H k X l , h k , X [ l ] ( u ) ] U k ( t − u ) . \begin{aligned}
\frac{d}{du}\left[U_k^\dagger(t-u)h_{k,X}^{[l]}(u)U_k(t-u)\right]
=-iU_k^\dagger(t-u)\left[H_k-H_k^{X^l},h_{k,X}^{[l]}(u)\right]U_k(t-u).
\end{aligned} d u d [ U k † ( t − u ) h k , X [ l ] ( u ) U k ( t − u ) ] = − i U k † ( t − u ) [ H k − H k X l , h k , X [ l ] ( u ) ] U k ( t − u ) .
Integrating from 0 0 0 to t t t , taking the operator norm, and using its invariance under unitary operations, we obtain
∥ h k , X ( t ) − h k , X [ l ] ( t ) ∥ ∞ ≤ ∫ 0 t ∥ [ H k − H k X l , h k , X [ l ] ( u ) ] ∥ ∞ d u (39) \left\lVert h_{k,X}(t)-h_{k,X}^{[l]}(t)\right\rVert_\infty
\leq\int_0^t\left\lVert\left[H_k-H_k^{X^l},h_{k,X}^{[l]}(u)\right]\right\rVert_\infty\,du \tag*{(39)} h k , X ( t ) − h k , X [ l ] ( t ) ∞ ≤ ∫ 0 t [ H k − H k X l , h k , X [ l ] ( u ) ] ∞ d u (39)
since
U k † ( t − u ) h k , X [ l ] ( u ) U k ( t − u ) ∣ u = t = h k , X [ l ] ( t ) , U_k^\dagger(t-u)h_{k,X}^{[l]}(u)U_k(t-u)\big|_{u=t}=h_{k,X}^{[l]}(t), U k † ( t − u ) h k , X [ l ] ( u ) U k ( t − u ) u = t = h k , X [ l ] ( t ) ,
U k † ( t − u ) h k , X [ l ] ( u ) U k ( t − u ) ∣ u = 0 = h k , X ( t ) . U_k^\dagger(t-u)h_{k,X}^{[l]}(u)U_k(t-u)\big|_{u=0}=h_{k,X}(t). U k † ( t − u ) h k , X [ l ] ( u ) U k ( t − u ) u = 0 = h k , X ( t ) .
We now consider the integral ∫ 0 t ∥ [ H k − H k X l , h k , X [ l ] ( u ) ] ∥ ∞ d u \int_0^t\left\|\left[H_k-H_k^{X^l},h_{k,X}^{[l]}(u)\right]\right\|_\infty\,du ∫ 0 t [ H k − H k X l , h k , X [ l ] ( u ) ] ∞ d u in Eq. (39) . The only terms that can fail to commute with h k , X [ l ] ( u ) h_{k,X}^{[l]}(u) h k , X [ l ] ( u ) are those supported entirely within X l X^l X l or crossing the boundary of X l X^l X l . By the definition of the truncated Hamiltonian, H k − H k X l = ∑ Z ⊈ X l h ~ k ( Z ) , Z H_k-H_k^{X^l}=\sum_{Z\not\subseteq X^l}\tilde{h}_{k(Z),Z} H k − H k X l = ∑ Z ⊆ X l h ~ k ( Z ) , Z , so H k − H k X l H_k-H_k^{X^l} H k − H k X l contains no terms supported entirely within X l X^l X l . Thus, the only nonzero contributions to [ H k − H k X l , h k , X [ l ] ( u ) ] \left[H_k-H_k^{X^l},h_{k,X}^{[l]}(u)\right] [ H k − H k X l , h k , X [ l ] ( u ) ] come from terms crossing the boundary of X l X^l X l . Applying the triangle inequality to their sum gives
∥ h k , X ( t ) − h k , X [ l ] ( t ) ∥ ∞ ≤ ∑ Z ∩ X l ≠ ∅ Z ⊈ X l ∫ 0 t ∥ [ h ~ k ( Z ) , Z , h k , X [ l ] ( u ) ] ∥ ∞ d u . \left\|h_{k,X}(t)-h_{k,X}^{[l]}(t)\right\|_\infty
\leq
\sum_{\substack{Z\cap X^l\neq\varnothing\\ Z\not\subseteq X^l}}
\int_0^t\left\|\left[\tilde{h}_{k(Z),Z},h_{k,X}^{[l]}(u)\right]\right\|_\infty\,du. h k , X ( t ) − h k , X [ l ] ( t ) ∞ ≤ Z ∩ X l = ∅ Z ⊆ X l ∑ ∫ 0 t [ h ~ k ( Z ) , Z , h k , X [ l ] ( u ) ] ∞ d u .
For the evolution generated by the truncated Hamiltonian H k X l H_k^{X^l} H k X l , the Lieb–Robinson bound reads
∥ [ h ~ k ( Z ) , Z , h k , X [ l ] ( u ) ] ∥ ∞ ≤ c L R ∥ h ~ k ( Z ) , Z ∥ ∞ ∥ h X ∥ ∞ e − κ [ dist ( X , Z ) − v u ] . \left\|\left[\tilde{h}_{k(Z),Z},h_{k,X}^{[l]}(u)\right]\right\|_\infty
\leq c_{\mathrm{LR}}\|\tilde{h}_{k(Z),Z}\|_\infty\|h_X\|_\infty e^{-\kappa[\operatorname{dist}(X,Z)-vu]}. [ h ~ k ( Z ) , Z , h k , X [ l ] ( u ) ] ∞ ≤ c LR ∥ h ~ k ( Z ) , Z ∥ ∞ ∥ h X ∥ ∞ e − κ [ dist ( X , Z ) − v u ] .
Here, v ≥ 0 v\geq0 v ≥ 0 is a Lieb–Robinson velocity associated with the feedback Hamiltonian H k X l H_k^{X^l} H k X l , and κ > 0 \kappa>0 κ > 0 is the spatial decay rate of the Lieb–Robinson tail. c L R c_{\mathrm{LR}} c LR , v v v and κ \kappa κ are independent of the system size L L L . Note that this is the commutator form of Theorem 1 of Ref. [62 ] for Hamiltonian dynamics; see also Ref. [63 ] . The factors depending on the uniformly bounded sizes of X X X and Z Z Z are included in c L R c_{\mathrm{LR}} c LR .
For each such region Z Z Z , choose a site z o u t ∈ Z ∖ X l z_{\mathrm{out}}\in Z\setminus X^l z out ∈ Z ∖ X l . By the definition of X l X^l X l , dist ( z o u t , X ) ≥ l \operatorname{dist}(z_{\mathrm{out}},X)\geq l dist ( z out , X ) ≥ l . The condition diam ( Z ) ≤ r 0 \operatorname{diam}(Z)\leq r_0 diam ( Z ) ≤ r 0 means that the distance between any two sites in Z Z Z is at most r 0 r_0 r 0 . Since z o u t z_{\mathrm{out}} z out is a site in Z Z Z , we have dist ( z , z o u t ) ≤ r 0 \operatorname{dist}(z,z_{\mathrm{out}})\leq r_0 dist ( z , z out ) ≤ r 0 for every site z ∈ Z z\in Z z ∈ Z . The triangle inequality therefore gives
dist ( z , X ) ≥ dist ( z o u t , X ) − dist ( z , z o u t ) ≥ l − r 0 . \operatorname{dist}(z,X)\geq\operatorname{dist}(z_{\mathrm{out}},X)-\operatorname{dist}(z,z_{\mathrm{out}})
\geq l-r_0. dist ( z , X ) ≥ dist ( z out , X ) − dist ( z , z out ) ≥ l − r 0 .
Taking the minimum over z ∈ Z z\in Z z ∈ Z , we obtain dist ( X , Z ) ≥ l − r 0 > 0 \operatorname{dist}(X,Z)\geq l-r_0>0 dist ( X , Z ) ≥ l − r 0 > 0 for l > r 0 l>r_0 l > r 0 . Thus,
∥ [ h ~ k ( Z ) , Z , h k , X [ l ] ( u ) ] ∥ ∞ ≤ c L R ∥ h ~ k ( Z ) , Z ∥ ∞ ∥ h X ∥ ∞ e − κ ( l − r 0 ) e κ v u . \left\|\left[\tilde{h}_{k(Z),Z},h_{k,X}^{[l]}(u)\right]\right\|_\infty
\leq c_{\mathrm{LR}}\|\tilde{h}_{k(Z),Z}\|_\infty\|h_X\|_\infty e^{-\kappa(l-r_0)}e^{\kappa vu}. [ h ~ k ( Z ) , Z , h k , X [ l ] ( u ) ] ∞ ≤ c LR ∥ h ~ k ( Z ) , Z ∥ ∞ ∥ h X ∥ ∞ e − κ ( l − r 0 ) e κ v u .
We next bound the sum of the norms of the Hamiltonian terms crossing the boundary of X l X^l X l . For a lattice site x x x , the sum ∑ Z ∋ x \sum_{Z\ni x} ∑ Z ∋ x runs over all local Hamiltonian terms whose regions Z Z Z contain x x x . On the fixed lattice, the bound diam ( Z ) ≤ r 0 \operatorname{diam}(Z)\leq r_0 diam ( Z ) ≤ r 0 , with r 0 r_0 r 0 independent of L L L , bounds the number of such terms independently of L L L . So, we assume that their norms are uniformly bounded, and ∑ Z ∋ x ∥ h ~ k ( Z ) , Z ∥ ∞ \sum_{Z\ni x}\|\tilde{h}_{k(Z),Z}\|_\infty ∑ Z ∋ x ∥ h ~ k ( Z ) , Z ∥ ∞ has an upper bound independent of the system size. We can therefore choose a finite constant J J J , independent of L L L , k k k , such that
∑ Z ∋ x ∥ h ~ k ( Z ) , Z ∥ ∞ ≤ J . \sum_{Z\ni x}\|\tilde{h}_{k(Z),Z}\|_\infty\leq J. Z ∋ x ∑ ∥ h ~ k ( Z ) , Z ∥ ∞ ≤ J .
We now sum this bound over the sites x ∈ X l x\in X^l x ∈ X l . In this double sum, a term with region Z Z Z is counted once for each site in Z ∩ X l Z\cap X^l Z ∩ X l . Every boundary-crossing region contains at least one site in X l X^l X l , so each boundary term is counted at least once. Repeated counting and the inclusion of terms entirely inside X l X^l X l can only increase the sum, since all norms are nonnegative. Hence,
∑ Z ∩ X l ≠ ∅ Z ⊈ X l ∥ h ~ k ( Z ) , Z ∥ ∞ ≤ ∑ x ∈ X l ∑ Z ∋ x ∥ h ~ k ( Z ) , Z ∥ ∞ ≤ ∑ x ∈ X l J = J ∣ X l ∣ . \sum_{\substack{Z\cap X^l\ne\varnothing\\ Z\not\subseteq X^l}}\|\tilde h_{k(Z),Z}\|_\infty
\le\sum_{x\in X^l}\sum_{Z\ni x}\|\tilde h_{k(Z),Z}\|_\infty
\le\sum_{x\in X^l} J = J|X^l|. Z ∩ X l = ∅ Z ⊆ X l ∑ ∥ h ~ k ( Z ) , Z ∥ ∞ ≤ x ∈ X l ∑ Z ∋ x ∑ ∥ h ~ k ( Z ) , Z ∥ ∞ ≤ x ∈ X l ∑ J = J ∣ X l ∣.
where ∣ X l ∣ |X^l| ∣ X l ∣ is the number of lattice sites in X l X^l X l . Substituting these estimates into the preceding integral bound on the truncation error, we obtain
∥ h k , X ( t ) − h k , X [ l ] ( t ) ∥ ∞ ≤ c L R J ∥ h X ∥ ∞ ∣ X l ∣ e − κ ( l − r 0 ) ∫ 0 t e κ v u d u . \|h_{k,X}(t)-h_{k,X}^{[l]}(t)\|_\infty
\le c_{\mathrm{LR}}J\|h_X\|_\infty|X^l|e^{-\kappa(l-r_0)}\int_0^t e^{\kappa vu}\,du. ∥ h k , X ( t ) − h k , X [ l ] ( t ) ∥ ∞ ≤ c LR J ∥ h X ∥ ∞ ∣ X l ∣ e − κ ( l − r 0 ) ∫ 0 t e κ v u d u .
We first estimate the number of sites in X l X^l X l . On a d d d -dimensional hypercubic lattice, all sites at distance less than l l l from a given site x x x lie inside a box centered at x x x with side length 2 l 2l 2 l in each spatial direction. The box therefore contains at most ( 1 + 2 l ) d (1+2l)^d ( 1 + 2 l ) d lattice sites. Since one such box centered at each site of X X X covers X l X^l X l , we obtain
∣ X l ∣ ≤ ∣ X ∣ ( 1 + 2 l ) d ≤ c 0 ( 1 + 2 l ) d . |X^l|\le|X|(1+2l)^d\le c_0(1+2l)^d. ∣ X l ∣ ≤ ∣ X ∣ ( 1 + 2 l ) d ≤ c 0 ( 1 + 2 l ) d .
The constant c 0 c_0 c 0 is a uniform upper bound on ∣ X ∣ |X| ∣ X ∣ and is independent of L L L . Next, for κ , v > 0 \kappa,v>0 κ , v > 0 and t ≥ 0 t\ge0 t ≥ 0 , the time integral is
∫ 0 t e κ v u d u = e κ v t − 1 κ v ≤ e κ v t κ v . \int_0^t e^{\kappa vu}\,du=\frac{e^{\kappa vt}-1}{\kappa v}
\le\frac{e^{\kappa vt}}{\kappa v}. ∫ 0 t e κ v u d u = κ v e κ v t − 1 ≤ κ v e κ v t .
Substituting both estimates into the preceding bound gives, for each measurement record k k k ,
∥ h k , X ( t ) − h k , X [ l ] ( t ) ∥ ∞ ≤ c L R J c 0 κ v ∥ h X ∥ ∞ ( 1 + 2 l ) d e − κ ( l − r 0 ) e κ v t . \|h_{k,X}(t)-h_{k,X}^{[l]}(t)\|_\infty
\le\frac{c_{\mathrm{LR}}Jc_0}{\kappa v}\|h_X\|_\infty(1+2l)^d e^{-\kappa(l-r_0)}e^{\kappa vt}. ∥ h k , X ( t ) − h k , X [ l ] ( t ) ∥ ∞ ≤ κ v c LR J c 0 ∥ h X ∥ ∞ ( 1 + 2 l ) d e − κ ( l − r 0 ) e κ v t .
Combining the exponential factors and setting C : = c L R J c 0 / ( κ v ) C:=c_{\mathrm{LR}}Jc_0/(\kappa v) C := c LR J c 0 / ( κ v ) , we obtain
max k ∥ h k , X ( t ) − h k , X [ l ] ( t ) ∥ ∞ ≤ C ∥ h X ∥ ∞ ( 1 + 2 l ) d e − κ ( l − r 0 − v t ) . (40) \max_k\|h_{k,X}(t)-h_{k,X}^{[l]}(t)\|_\infty
\le C\|h_X\|_\infty(1+2l)^d e^{-\kappa(l-r_0-vt)}. \tag*{(40)} k max ∥ h k , X ( t ) − h k , X [ l ] ( t ) ∥ ∞ ≤ C ∥ h X ∥ ∞ ( 1 + 2 l ) d e − κ ( l − r 0 − v t ) . (40)
Combining Eqs. (38) and (40) , we find that the total feedback approximation error in Eq. (31) obeys
Eq. (S7) ≤ C ( 1 + 2 l ) d e − κ ( l − r 0 − v t ) ∑ X ∥ h X ∥ ∞ . (41) \text{Eq. (S7)}\le C(1+2l)^d e^{-\kappa(l-r_0-vt)}\sum_X\|h_X\|_\infty. \tag*{(41)} Eq. (S7) ≤ C ( 1 + 2 l ) d e − κ ( l − r 0 − v t ) X ∑ ∥ h X ∥ ∞ . (41)
The sum ∑ X ∥ h X ∥ ∞ \sum_X\|h_X\|_\infty ∑ X ∥ h X ∥ ∞ on the right-hand side contains O ( L d ) O(L^d) O ( L d ) terms. By assumption, each norm is bounded by the same constant, independent of L L L . The sum is therefore O ( L d ) O(L^d) O ( L d ) . Since C C C is independent of L L L , l l l , and t t t , we obtain
C ( 1 + 2 l ) d e − κ ( l − r 0 − v t ) ∑ X ∥ h X ∥ ∞ = O ( L d ) ( 1 + 2 l ) d e − κ ( l − r 0 − v t ) . C(1+2l)^d e^{-\kappa(l-r_0-vt)}\sum_X\|h_X\|_\infty
=O(L^d)(1+2l)^d e^{-\kappa(l-r_0-vt)}. C ( 1 + 2 l ) d e − κ ( l − r 0 − v t ) X ∑ ∥ h X ∥ ∞ = O ( L d ) ( 1 + 2 l ) d e − κ ( l − r 0 − v t ) .
Therefore,
Eq. (S7) ≤ O ( L d ) ( 1 + 2 l ) d e − κ ( l − r 0 − v t ) . (42) \text{Eq. (S7)}\le O(L^d)(1+2l)^d e^{-\kappa(l-r_0-vt)}. \tag*{(42)} Eq. (S7) ≤ O ( L d ) ( 1 + 2 l ) d e − κ ( l − r 0 − v t ) . (42)
It follows that, if l l l satisfies
l = O ( L s ) ( 0 < s ) & t ≪ l − r 0 v , (43) l = \mathcal{O}(L^{s}) \quad(0 < s) \ \&\ t \ll\frac{l-r_{0}}{v}, \tag*{(43)} l = O ( L s ) ( 0 < s ) & t ≪ v l − r 0 , (43)
where t t t is the duration of the feedback control, then in the thermodynamic limit, the bound in Eq. (42) is exponentially suppressed. Therefore, the operator h k , X ( t ) h_{k,X}(t) h k , X ( t ) in the first term of Eq. (31) can be approximated by h k , X [ l ] ( t ) h_{k,X}^{[l]}(t) h k , X [ l ] ( t ) :
∑ k , X Tr [ ρ i M I T E U 1 † M k † h k , X ( t ) M k U 1 ] ∼ ∑ k , X Tr [ ρ i M I T E U 1 † M k † h k , X [ l ] ( t ) M k U 1 ] . (44) \sum_{k,X} \operatorname{Tr}\left[\rho_i^{\mathrm{MITE}}U_1^\dagger M_k^\dagger h_{k,X}(t)M_kU_1\right]
\sim\sum_{k,X} \operatorname{Tr}\left[\rho_i^{\mathrm{MITE}}U_1^\dagger M_k^\dagger h_{k,X}^{[l]}(t)M_kU_1\right]. \tag*{(44)} k , X ∑ Tr [ ρ i MITE U 1 † M k † h k , X ( t ) M k U 1 ] ∼ k , X ∑ Tr [ ρ i MITE U 1 † M k † h k , X [ l ] ( t ) M k U 1 ] . (44)
The preceding argument assumes s > 0 s>0 s > 0 for l = O ( L s ) l=\mathcal{O}(L^s) l = O ( L s ) and therefore does not directly apply to the case s = 0 s=0 s = 0 , which corresponds to the ordinary MITE. Nevertheless, Eq. (44) remains valid for s = 0 s=0 s = 0 , as we now show. For the work considered here, the notation W ≃ W e n s W\simeq W_{\mathrm{ens}} W ≃ W ens means that the difference between W W W and W e n s W_{\mathrm{ens}} W ens is subextensive:
W ≃ W e n s ⟺ ∣ W − W e n s ∣ = o ( V ) ⟺ lim L → ∞ ∣ W − W e n s ∣ L d = 0. (45) W \simeq W_{\mathrm{ens}} \quad\Longleftrightarrow\quad|W-W_{\mathrm{ens}}|=o(V) \quad\Longleftrightarrow\quad\lim_{L\to\infty}\frac{|W-W_{\mathrm{ens}}|}{L^d}=0. \tag*{(45)} W ≃ W ens ⟺ ∣ W − W ens ∣ = o ( V ) ⟺ L → ∞ lim L d ∣ W − W ens ∣ = 0. (45)
Equivalently, this condition can be expressed as
∀ ϵ > 0 ∃ L ϵ ∀ L ≥ L ϵ , ∣ W − W e n s ∣ L d < ϵ . (46) \forall\epsilon>0\quad\exists L_\epsilon\quad\forall L\geq L_\epsilon,\qquad\frac{|W-W_{\mathrm{ens}}|}{L^d}<\epsilon. \tag*{(46)} ∀ ϵ > 0 ∃ L ϵ ∀ L ≥ L ϵ , L d ∣ W − W ens ∣ < ϵ . (46)
It is therefore sufficient to show that the error defined in Eq. (31) can be made arbitrarily small per unit volume. Equation (S18) implies that
E q . ( S 7 ) L d ≤ ( 1 + 2 l ) d e − κ [ l − r 0 − v t ] . (47) \frac{\mathrm{Eq.}\ (\mathrm{S7})}{L^d}\leq(1+2l)^d e^{-\kappa[l-r_0-vt]}. \tag*{(47)} L d Eq. ( S7 ) ≤ ( 1 + 2 l ) d e − κ [ l − r 0 − v t ] . (47)
For t = O ( L 0 ) t=\mathcal{O}(L^0) t = O ( L 0 ) and any ϵ > 0 \epsilon>0 ϵ > 0 , one can choose l ϵ l_\epsilon l ϵ independent of L L L such that
( 1 + 2 l ϵ ) d e − κ [ l ϵ − r 0 − v t ] < ϵ . (48) (1+2l_\epsilon)^d e^{-\kappa[l_\epsilon-r_0-vt]}<\epsilon. \tag*{(48)} ( 1 + 2 l ϵ ) d e − κ [ l ϵ − r 0 − v t ] < ϵ . (48)
Consequently,
∀ ϵ > 0 ∃ l ϵ ∃ L ϵ ∀ L ≥ L ϵ , E q . ( S 7 ) L d < ϵ . (49) \forall\epsilon>0\quad\exists l_\epsilon\quad\exists L_\epsilon\quad\forall L\geq L_\epsilon,\qquad\frac{\mathrm{Eq.}\ (\mathrm{S7})}{L^d}<\epsilon. \tag*{(49)} ∀ ϵ > 0 ∃ l ϵ ∃ L ϵ ∀ L ≥ L ϵ , L d Eq. ( S7 ) < ϵ . (49)
Since ϵ > 0 \epsilon>0 ϵ > 0 is arbitrary, the error in Eq. (31) is subextensive. Hence, Eq. (44) remains valid for s = 0 s=0 s = 0 . Combining the two cases, we find that Eq. (44) holds when
l = O ( L s ) , t = O ( L s ) , 0 ≤ s < 1 , (50) l=\mathcal{O}(L^s),\qquad t=\mathcal{O}(L^s),\qquad0\leq s<1, \tag*{(50)} l = O ( L s ) , t = O ( L s ) , 0 ≤ s < 1 , (50)
where, for s > 0 s>0 s > 0 , the auxiliary length l l l is chosen such that l − r 0 − v t ≥ c L s l-r_0-vt\geq cL^s l − r 0 − v t ≥ c L s for some constant c > 0 c>0 c > 0 . For s = 0 s=0 s = 0 , l l l is chosen sufficiently large but independently of L L L , as explained above.
Measurement processes (Figure S1 )
Next, we consider
∑ k ∑ X Tr [ ρ i M I T E U 1 † M k † h k , X [ l ] ( t ) M k U 1 ] . (51) \sum_k\sum_X \operatorname{Tr}\left[\rho_i^{\mathrm{MITE}}U_1^\dagger M_k^\dagger h_{k,X}^{[l]}(t)M_kU_1\right]. \tag*{(51)} k ∑ X ∑ Tr [ ρ i MITE U 1 † M k † h k , X [ l ] ( t ) M k U 1 ] . (51)
We first set
G X [ l ] ( t ) : = ∑ k M k † h k , X [ l ] ( t ) M k . (52) G_X^{[l]}(t):=\sum_k M_k^\dagger h_{k,X}^{[l]}(t)M_k. \tag*{(52)} G X [ l ] ( t ) := k ∑ M k † h k , X [ l ] ( t ) M k . (52)
We now clarify the condition imposed on the measurement operator M k M_k M k . We represent the measurement operator corresponding to outcome k k k by a depth-d M d_{\mathrm{M}} d M local measurement circuit (Figure S2 ):
Figure S2. Schematic illustration of a local measurement circuit of depth d M = 3 d_{\mathrm{M}}=3 d M = 3 in a one-dimensional system. The measurement operator is M k = m k 3 3 m k 2 2 m k 1 1 M_k=m^3_{k_3}m^2_{k_2}m^1_{k_1} M k = m k 3 3 m k 2 2 m k 1 1 , with m k 1 1 = m k 1 , Y 1 1 , Y 1 m k 1 , Y 2 1 , Y 2 m k 1 , Y 3 1 , Y 3 m^1_{k_1}=m^{1,Y_1}_{k_1,Y_1}m^{1,Y_2}_{k_1,Y_2}m^{1,Y_3}_{k_1,Y_3} m k 1 1 = m k 1 , Y 1 1 , Y 1 m k 1 , Y 2 1 , Y 2 m k 1 , Y 3 1 , Y 3 , m k 2 2 = m k 2 , Y 1 ′ 2 , Y 1 ′ m k 2 , Y 2 ′ 2 , Y 2 ′ m^2_{k_2}=m^{2,Y'_1}_{k_2,Y'_1}m^{2,Y'_2}_{k_2,Y'_2} m k 2 2 = m k 2 , Y 1 ′ 2 , Y 1 ′ m k 2 , Y 2 ′ 2 , Y 2 ′ , and m k 3 3 = m k 3 , Y 1 ′ ′ 3 , Y 1 ′ ′ m k 3 , Y 2 ′ ′ 3 , Y 2 ′ ′ m k 3 , Y 3 ′ ′ 3 , Y 3 ′ ′ m^3_{k_3}=m^{3,Y''_1}_{k_3,Y''_1}m^{3,Y''_2}_{k_3,Y''_2}m^{3,Y''_3}_{k_3,Y''_3} m k 3 3 = m k 3 , Y 1 ′′ 3 , Y 1 ′′ m k 3 , Y 2 ′′ 3 , Y 2 ′′ m k 3 , Y 3 ′′ 3 , Y 3 ′′ . The diameters of the corresponding regions are ( 1 , 3 , 1 ) (1,3,1) ( 1 , 3 , 1 ) , ( 2 , 2 ) (2,2) ( 2 , 2 ) , and ( 1 , 1 , 1 ) (1,1,1) ( 1 , 1 , 1 ) for ( Y 1 , Y 2 , Y 3 ) (Y_1,Y_2,Y_3) ( Y 1 , Y 2 , Y 3 ) , ( Y 1 ′ , Y 2 ′ ) (Y'_1,Y'_2) ( Y 1 ′ , Y 2 ′ ) , and ( Y 1 ′ ′ , Y 2 ′ ′ , Y 3 ′ ′ ) (Y''_1,Y''_2,Y''_3) ( Y 1 ′′ , Y 2 ′′ , Y 3 ′′ ) , respectively.
M k = m k d M d M m k d M − 1 d M − 1 ⋯ m k 1 1 , (53) M_k=m^{d_{\mathrm{M}}}_{k_{d_{\mathrm{M}}}}m^{d_{\mathrm{M}}-1}_{k_{d_{\mathrm{M}}-1}}\cdots m^1_{k_1}, \tag*{(53)} M k = m k d M d M m k d M − 1 d M − 1 ⋯ m k 1 1 , (53)
m k n n = ⨂ Y ∈ F n m k n , Y n , Y . (54) m^n_{k_n}=\bigotimes_{Y\in\mathcal{F}_n}m^{n,Y}_{k_n,Y}. \tag*{(54)} m k n n = Y ∈ F n ⨂ m k n , Y n , Y . (54)
Here, m k n , Y n , Y m^{n,Y}_{k_n,Y} m k n , Y n , Y is supported on the region Y Y Y , as explained in the main paper. The local completeness relation is
∑ k n , Y ( m k n , Y n , Y ) † m k n , Y n , Y = 1 Y . \sum_{k_n,Y}\left(m^{n,Y}_{k_n,Y}\right)^\dagger m^{n,Y}_{k_n,Y}=1_Y. k n , Y ∑ ( m k n , Y n , Y ) † m k n , Y n , Y = 1 Y .
Each F n \mathcal{F}_n F n is a collection of mutually disjoint local regions Y Y Y . We define the diameter of the region Y Y Y by
diam ( Y ) : = max { dist ( y 1 , y 2 ) : y 1 , y 2 ∈ Y } , (55) \operatorname{diam}(Y):=\max\{\operatorname{dist}(y_1,y_2):y_1,y_2\in Y\}, \tag*{(55)} diam ( Y ) := max { dist ( y 1 , y 2 ) : y 1 , y 2 ∈ Y } , (55)
where dist ( y 1 , y 2 ) \operatorname{dist}(y_1,y_2) dist ( y 1 , y 2 ) is the Euclidean distance, and
D : = max n max Y ∈ F n diam ( Y ) + 1. (56) \mathcal{D}:=\max_n\max_{Y\in\mathcal{F}_n}\operatorname{diam}(Y)+1. \tag*{(56)} D := n max Y ∈ F n max diam ( Y ) + 1. (56)
Let supp ( G X [ l ] ( t ) ) \operatorname{supp}(G_X^{[l]}(t)) supp ( G X [ l ] ( t )) be a region on which G X [ l ] ( t ) G_X^{[l]}(t) G X [ l ] ( t ) is supported. By definition, the truncated Hamiltonian H k X l H_k^{X^l} H k X l contains only the Hamiltonian terms h ~ k ( Z ) , Z \widetilde{h}_{k(Z),Z} h k ( Z ) , Z with Z ⊆ X l Z\subseteq X^l Z ⊆ X l . The term h ~ k ( Z ) , Z \widetilde{h}_{k(Z),Z} h k ( Z ) , Z may depend on the measurement outcome k n , Y k_{n,Y} k n , Y only if there is a site y 0 ∈ Y y_0\in Y y 0 ∈ Y such that dist ( Z , { y 0 } ) ≤ ℓ \operatorname{dist}(Z,\{y_0\})\leq\ell dist ( Z , { y 0 }) ≤ ℓ . Such a site y 0 y_0 y 0 lies in X l + ℓ X^{l+\ell} X l + ℓ . Thus, h k , X [ l ] ( t ) h_{k,X}^{[l]}(t) h k , X [ l ] ( t ) is independent of k n , Y k_{n,Y} k n , Y whenever Y Y Y is disjoint from X l + ℓ X^{l+\ell} X l + ℓ . When summing over the outcomes, we must account for both the support of h k , X [ l ] ( t ) h_{k,X}^{[l]}(t) h k , X [ l ] ( t ) and its dependence on those outcomes. For measurement regions disjoint from X l + ℓ X^{l+\ell} X l + ℓ , this operator is independent of the corresponding outcomes and commutes with the measurement operators. In the first layer processed, measurement regions disjoint from X l + ℓ X^{l+\ell} X l + ℓ can therefore be removed by local completeness. Consider the measurement regions in this layer that overlap X l + ℓ X^{l+\ell} X l + ℓ . After summing over the outcomes of this layer, the resulting operator is supported within the union of these regions and X l + ℓ X^{l+\ell} X l + ℓ . In the next layer, measurement regions disjoint from the enlarged region can be removed in the same way. Only regions overlapping the enlarged region can extend it further, and we repeat this procedure for the remaining layers.
Within each layer, the measurement regions are mutually disjoint and have diameter at most D \mathcal{D} D . Each layer can therefore increase the distance from X X X by at most D − 1 \mathcal{D}-1 D − 1 . After d M d_{\mathrm{M}} d M layers, every site in supp ( G X [ l ] ( t ) ) \operatorname{supp}(G_X^{[l]}(t)) supp ( G X [ l ] ( t )) is contained in X R M X^{R_{\mathrm{M}}} X R M , where
R M : = l + ℓ + [ D − 1 ] d M (57) R_{\mathrm{M}}:=l+\ell+[\mathcal{D}-1]d_{\mathrm{M}} \tag*{(57)} R M := l + ℓ + [ D − 1 ] d M (57)
is the radius of X R M X^{R_{\mathrm{M}}} X R M . That is,
supp ( G X [ l ] ( t ) ) ⊆ X R M . (58) \operatorname{supp}(G_X^{[l]}(t))\subseteq X^{R_{\mathrm{M}}}. \tag*{(58)} supp ( G X [ l ] ( t )) ⊆ X R M . (58)
In particular, the diameter of supp ( G X [ l ] ( t ) ) \operatorname{supp}(G_X^{[l]}(t)) supp ( G X [ l ] ( t )) is bounded as
diam ( supp ( G X [ l ] ( t ) ) ) ≤ 2 R M = 2 ( l + ℓ ) + 2 [ D − 1 ] d M . (59) \operatorname{diam}\left(\operatorname{supp}(G_X^{[l]}(t))\right) \le2R_M = 2(l+\ell) + 2[\mathcal{D}-1]d_M. \tag*{(59)} diam ( supp ( G X [ l ] ( t )) ) ≤ 2 R M = 2 ( l + ℓ ) + 2 [ D − 1 ] d M . (59)
For consistency with the extended MITE condition discussed below, we assume that M k M_k M k satisfies
[ D − 1 ] d M = O ( L s ) , 0 ≤ s < 1. (60) [\mathcal{D}-1]d_M = O(L^s), \qquad0 \le s < 1. \tag*{(60)} [ D − 1 ] d M = O ( L s ) , 0 ≤ s < 1. (60)
In other words, we assume that the measurement operator is a tensor product of low-depth and local measurements. Under this assumption, R M = O ( L s ) R_M=O(L^s) R M = O ( L s ) , so G X [ l ] ( t ) G_X^{[l]}(t) G X [ l ] ( t ) acts only on an O ( L s ) O(L^s) O ( L s ) -local region.
Premeasurement processes (Figure S1 (c))
With the definition in Eq. (52) ,
Eq. (S27) = ∑ X Tr [ ρ i M I T E U 1 † G X [ l ] ( t ) U 1 ] . (61) \text{Eq. (S27)} = \sum_X \operatorname{Tr}\left[\rho_i^{\mathrm{MITE}} U_1^\dagger G_X^{[l]}(t)U_1\right]. \tag*{(61)} Eq. (S27) = X ∑ Tr [ ρ i MITE U 1 † G X [ l ] ( t ) U 1 ] . (61)
The unitary operator U 1 U_1 U 1 , unlike the feedback unitary operator, is independent of the measurement outcome and can be written as
U 1 : = U 1 ( t ′ ) = T exp ( − i ∫ 0 t ′ H 1 ( u ) d u ) , (62) U_1 := U_1(t') = \mathcal{T}\exp\left(-i\int_0^{t'} H_1(u)\,\mathrm{d}u\right), \tag*{(62)} U 1 := U 1 ( t ′ ) = T exp ( − i ∫ 0 t ′ H 1 ( u ) d u ) , (62)
where t ′ t' t ′ denotes the duration for which the unitary operator U 1 U_1 U 1 is applied. As we mentioned before, we assume that H 1 ( u ) = H 1 H_1(u)=H_1 H 1 ( u ) = H 1 is time-independent, although the generalization to time-dependent cases is straightforward. We choose l ′ l' l ′ such that
R M ≤ l ′ . (63) R_M \le l'. \tag*{(63)} R M ≤ l ′ . (63)
We now define an operator supported on X l ′ X^{l'} X l ′ by
G X [ l ′ , l ] ( t , t ′ ) : = 1 Tr X l ′ 1 X l ′ ⊗ Tr X l ′ ‾ [ U 1 † G X [ l ] ( t ) U 1 ] . (64) G_X^{[l',l]}(t,t') := \frac{1}{\operatorname{Tr}_{X^{l'}}\mathbf{1}_{X^{l'}}}\otimes\operatorname{Tr}_{\overline{X^{l'}}}\left[U_1^\dagger G_X^{[l]}(t)U_1\right]. \tag*{(64)} G X [ l ′ , l ] ( t , t ′ ) := Tr X l ′ 1 X l ′ 1 ⊗ Tr X l ′ [ U 1 † G X [ l ] ( t ) U 1 ] . (64)
In this case,
∣ ∑ X Tr [ ρ i M I T E U 1 † G X [ l ] ( t ) U 1 ] − ∑ X Tr [ ρ i M I T E G X [ l ′ , l ] ( t , t ′ ) ] ∣ = ∣ ∑ X Tr [ ρ i M I T E ( U 1 † G X [ l ] ( t ) U 1 − G X [ l ′ , l ] ( t , t ′ ) ) ] ∣ ≤ ∑ X ∥ U 1 † G X [ l ] ( t ) U 1 − G X [ l ′ , l ] ( t , t ′ ) ∥ ∞ . (65) \begin{aligned}
\left|\sum_X \operatorname{Tr}\left[\rho_i^{\mathrm{MITE}}U_1^\dagger G_X^{[l]}(t)U_1\right]-\sum_X\operatorname{Tr}\left[\rho_i^{\mathrm{MITE}}G_X^{[l',l]}(t,t')\right]\right| \\
&=\left|\sum_X\operatorname{Tr}\left[\rho_i^{\mathrm{MITE}}\left(U_1^\dagger G_X^{[l]}(t)U_1-G_X^{[l',l]}(t,t')\right)\right]\right| \\
&\le\sum_X\left\|U_1^\dagger G_X^{[l]}(t)U_1-G_X^{[l',l]}(t,t')\right\|_\infty. \tag*{(65)}
\end{aligned} X ∑ Tr [ ρ i MITE U 1 † G X [ l ] ( t ) U 1 ] − X ∑ Tr [ ρ i MITE G X [ l ′ , l ] ( t , t ′ ) ] = X ∑ Tr [ ρ i MITE ( U 1 † G X [ l ] ( t ) U 1 − G X [ l ′ , l ] ( t , t ′ ) ) ] ≤ X ∑ U 1 † G X [ l ] ( t ) U 1 − G X [ l ′ , l ] ( t , t ′ ) ∞ . (65)
Applying the Lieb–Robinson bound [53 , 54 , 55 ] to the additional spreading generated by U 1 U_1 U 1 , and using the fact that G X [ l ] ( t ) G_X^{[l]}(t) G X [ l ] ( t ) is already contained in X R M X^{R_M} X R M , we obtain
Eq. (S41) ≤ ∑ X c 2 ′ e − κ ′ ( l ′ − R M − v ′ t ′ ) . (66) \text{Eq. (S41)} \le\sum_X c'_2 e^{-\kappa'(l'-R_M-v't')}. \tag*{(66)} Eq. (S41) ≤ X ∑ c 2 ′ e − κ ′ ( l ′ − R M − v ′ t ′ ) . (66)
where
c 2 ′ = c 2 ∥ U 1 † G X [ l ] ( t ) U 1 ∥ ∞ ∣ X l ′ ∣ . (67) c'_2=c_2\left\|U_1^\dagger G_X^{[l]}(t)U_1\right\|_\infty|X^{l'}|. \tag*{(67)} c 2 ′ = c 2 U 1 † G X [ l ] ( t ) U 1 ∞ ∣ X l ′ ∣. (67)
Here, c 2 c_2 c 2 is a constant independent of the system size, v ′ v' v ′ is the Lieb-Robinson velocity associated with H 1 H_1 H 1 , κ ′ > 0 \kappa' > 0 κ ′ > 0 is the spatial decay rate of the Lieb-Robinson tail, and ∣ X l ′ ∣ |X^{l'}| ∣ X l ′ ∣ denotes the number of sites in the region X l ′ X^{l'} X l ′ . In spatial dimension d d d , we can estimate ∣ X l ′ ∣ ∼ ( 1 + 2 l ′ ) d |X^{l'}| \sim(1+2l')^d ∣ X l ′ ∣ ∼ ( 1 + 2 l ′ ) d . Moreover,
∥ U 1 † G X [ l ] ( t ) U 1 ∥ ∞ = ∥ G X [ l ] ( t ) ∥ ∞ = ∥ ∑ k M k † h k , X [ l ] ( t ) M k ∥ ∞ . (68) \left\|U_1^\dagger G_X^{[l]}(t)U_1\right\|_\infty=\left\|G_X^{[l]}(t)\right\|_\infty
=\left\|\sum_k M_k^\dagger h_{k,X}^{[l]}(t)M_k\right\|_\infty. \tag*{(68)} U 1 † G X [ l ] ( t ) U 1 ∞ = G X [ l ] ( t ) ∞ = k ∑ M k † h k , X [ l ] ( t ) M k ∞ . (68)
Since M k † h k , X [ l ] ( t ) M k M_k^\dagger h_{k,X}^{[l]}(t)M_k M k † h k , X [ l ] ( t ) M k is Hermitian,
Eq. (68)
= sup ∥ ∣ ψ ⟩ ∥ = 1 ∣ ⟨ ψ ∣ ∑ k M k † h k , X [ l ] ( t ) M k ∣ ψ ⟩ ∣ , (69) =\sup_{\|\lvert\psi\rangle\|=1}\left|\left\langle\psi\left|\sum_k M_k^\dagger h_{k,X}^{[l]}(t)M_k\right|\psi\right\rangle\right| , \tag*{(69)} = ∥ ∣ ψ ⟩ ∥ = 1 sup ⟨ ψ k ∑ M k † h k , X [ l ] ( t ) M k ψ ⟩ , (69)
where ∣ ψ ⟩ \lvert\psi\rangle ∣ ψ ⟩ is a unit vector. This expression can be further transformed as
Eq. (69)
≤ sup ∥ ∣ ψ ⟩ ∥ = 1 ∑ k ∥ M k ∣ ψ ⟩ ∥ 2 ∣ ⟨ ψ ∣ M k † ∥ M k ∣ ψ ⟩ ∥ h k , X [ l ] ( t ) M k ∣ ψ ⟩ ∥ M k ∣ ψ ⟩ ∥ ∣ ≤ sup ∥ ∣ ψ ⟩ ∥ = 1 ∑ k ∥ M k ∣ ψ ⟩ ∥ 2 max k ∣ ⟨ ψ ∣ M k † ∥ M k ∣ ψ ⟩ ∥ h k , X [ l ] ( t ) M k ∣ ψ ⟩ ∥ M k ∣ ψ ⟩ ∥ ∣ = sup ∥ ∣ ψ ⟩ ∥ = 1 max k ∣ ⟨ ψ ∣ M k † ∥ M k ∣ ψ ⟩ ∥ h k , X [ l ] ( t ) M k ∣ ψ ⟩ ∥ M k ∣ ψ ⟩ ∥ ∣ ≤ max k ∥ h k , X [ l ] ( t ) ∥ ∞ . (70) \begin{aligned}
&\leq\sup_{\|\lvert\psi\rangle\|=1}\sum_k\|M_k\lvert\psi\rangle\|^2
\left|\frac{\langle\psi|M_k^\dagger}{\|M_k\lvert\psi\rangle\|}
h_{k,X}^{[l]}(t)
\frac{M_k|\psi\rangle}{\|M_k|\psi\rangle\|}\right|\\
&\leq\sup_{\|\lvert\psi\rangle\|=1}\sum_k\|M_k\lvert\psi\rangle\|^2
\max_k\left|\frac{\langle\psi|M_k^\dagger}{\|M_k\lvert\psi\rangle\|}
h_{k,X}^{[l]}(t)
\frac{M_k|\psi\rangle}{\|M_k|\psi\rangle\|}\right|\\
&= \sup_{\|\lvert\psi\rangle\|=1}\max_k
\left|\frac{\langle\psi|M_k^\dagger}{\|M_k\lvert\psi\rangle\|}
h_{k,X}^{[l]}(t)
\frac{M_k|\psi\rangle}{\|M_k|\psi\rangle\|}\right|\\
&\leq\max_k\left\|h_{k,X}^{[l]}(t)\right\|_\infty. \tag*{(70)}
\end{aligned} ≤ ∥ ∣ ψ ⟩ ∥ = 1 sup k ∑ ∥ M k ∣ ψ ⟩ ∥ 2 ∥ M k ∣ ψ ⟩ ∥ ⟨ ψ ∣ M k † h k , X [ l ] ( t ) ∥ M k ∣ ψ ⟩ ∥ M k ∣ ψ ⟩ ≤ ∥ ∣ ψ ⟩ ∥ = 1 sup k ∑ ∥ M k ∣ ψ ⟩ ∥ 2 k max ∥ M k ∣ ψ ⟩ ∥ ⟨ ψ ∣ M k † h k , X [ l ] ( t ) ∥ M k ∣ ψ ⟩ ∥ M k ∣ ψ ⟩ = ∥ ∣ ψ ⟩ ∥ = 1 sup k max ∥ M k ∣ ψ ⟩ ∥ ⟨ ψ ∣ M k † h k , X [ l ] ( t ) ∥ M k ∣ ψ ⟩ ∥ M k ∣ ψ ⟩ ≤ k max h k , X [ l ] ( t ) ∞ . (70)
Thus,
∥ U 1 † G X [ l ] ( t ) U 1 ∥ ∞ ≤ max k ∥ h k , X [ l ] ( t ) ∥ ∞ . (71) \left\|U_1^\dagger G_X^{[l]}(t)U_1\right\|_\infty
\leq\max_k\left\|h_{k,X}^{[l]}(t)\right\|_\infty. \tag*{(71)} U 1 † G X [ l ] ( t ) U 1 ∞ ≤ k max h k , X [ l ] ( t ) ∞ . (71)
By the triangle inequality,
Eq. (71)
≤ max k [ ∥ h k , X ( t ) ∥ ∞ + ∥ h k , X ( t ) − h k , X [ l ] ( t ) ∥ ∞ ] = max k [ ∥ U k † ( t ) h X U k ( t ) ∥ ∞ + ∥ h k , X ( t ) − h k , X [ l ] ( t ) ∥ ∞ ] = max k [ ∥ h X ∥ ∞ + ∥ h k , X ( t ) − h k , X [ l ] ( t ) ∥ ∞ ] . (72) \begin{aligned}
&\leq\max_k\left[\left\|h_{k,X}(t)\right\|_\infty+
\left\|h_{k,X}(t)-h_{k,X}^{[l]}(t)\right\|_\infty\right]\\
&=\max_k\left[\left\|U_k^\dagger(t)h_XU_k(t)\right\|_\infty+
\left\|h_{k,X}(t)-h_{k,X}^{[l]}(t)\right\|_\infty\right]\\
&=\max_k\left[\left\|h_X\right\|_\infty+
\left\|h_{k,X}(t)-h_{k,X}^{[l]}(t)\right\|_\infty\right]. \tag*{(72)}
\end{aligned} ≤ k max [ ∥ h k , X ( t ) ∥ ∞ + h k , X ( t ) − h k , X [ l ] ( t ) ∞ ] = k max [ U k † ( t ) h X U k ( t ) ∞ + h k , X ( t ) − h k , X [ l ] ( t ) ∞ ] = k max [ ∥ h X ∥ ∞ + h k , X ( t ) − h k , X [ l ] ( t ) ∞ ] . (72)
As in the derivation of Eq. (40) , the Lieb-Robinson bound gives
Eq. (72) ≤ \leq ≤
∥ h X ∥ ∞ + C ∥ h X ∥ ∞ ( 1 + 2 l ) d e − κ ′ ( l − r 0 − v t ) . (73) \left\|h_X\right\|_\infty+C\left\|h_X\right\|_\infty(1+2l)^d e^{-\kappa'(l-r_0-vt)} . \tag*{(73)} ∥ h X ∥ ∞ + C ∥ h X ∥ ∞ ( 1 + 2 l ) d e − κ ′ ( l − r 0 − v t ) . (73)
Thus, when l l l satisfies Eq. (50) , where t t t is the duration of the feedback control, the second term vanishes in the thermodynamic limit. Consequently,
∥ U 1 † G X [ l ] ( t ) U 1 ∥ ∞ ≲ ∥ h X ∥ ∞ . (74) \left\|U_1^\dagger G_X^{[l]}(t)U_1\right\|_\infty\lesssim\left\|h_X\right\|_\infty. \tag*{(74)} U 1 † G X [ l ] ( t ) U 1 ∞ ≲ ∥ h X ∥ ∞ . (74)
Therefore, from Eqs. (S42), (S43), and (S50),
Eq. (S41) ≤ ∑ X c 2 ∥ U 1 † G X [ l ] ( t ) U 1 ∥ ∞ ∣ X l ′ ∣ e − κ ′ ( l ′ − R M − v ′ t ) \text{Eq. (S41)}\leq\sum_X c_2\left\|U_1^\dagger G_X^{[l]}(t)U_1\right\|_\infty|X^{l'}|e^{-\kappa'(l'-R_M-v't)} Eq. (S41) ≤ X ∑ c 2 U 1 † G X [ l ] ( t ) U 1 ∞ ∣ X l ′ ∣ e − κ ′ ( l ′ − R M − v ′ t )
≲ ∑ X c 2 ∥ h X ∥ ∞ ( 1 + 2 l ′ ) d e − κ ′ ( l ′ − R M − v ′ t ) . (75) \lesssim\sum_X c_2\left\|h_X\right\|_\infty(1+2l')^d e^{-\kappa'(l'-R_M-v't)} . \tag*{(75)} ≲ X ∑ c 2 ∥ h X ∥ ∞ ( 1 + 2 l ′ ) d e − κ ′ ( l ′ − R M − v ′ t ) . (75)
Now, as well as Eq. (50) , we assume for l ′ l' l ′ that
l ′ = O ( L s ) , R M ≤ l ′ , t ′ = O ( L s ) , 0 ≤ s < 1 , (76) l' = O(L^s), \qquad R_M \le l', \qquad t' = O(L^s), \qquad0 \le s < 1, \tag*{(76)} l ′ = O ( L s ) , R M ≤ l ′ , t ′ = O ( L s ) , 0 ≤ s < 1 , (76)
where, for s > 0 s>0 s > 0 , the auxiliary length l ′ l' l ′ is chosen such that l ′ − R M − v ′ t ′ ≥ c ′ L s l' - R_M - v't' \ge c'L^s l ′ − R M − v ′ t ′ ≥ c ′ L s for some constant c ′ > 0 c' > 0 c ′ > 0 . For s = 0 s=0 s = 0 , l ′ l' l ′ is chosen sufficiently large but independently of L L L , in a similar manner to Eq. (50) . Then, in the thermodynamic limit the operator U 1 † G X [ l ] ( t ) U 1 U_1^\dagger G_X^{[l]}(t)U_1 U 1 † G X [ l ] ( t ) U 1 in the first term of Eq. (65) can be approximated by G X [ l ′ , l ] ( t , t ′ ) G_X^{[l',l]}(t,t') G X [ l ′ , l ] ( t , t ′ ) :
∑ X Tr [ ρ i M I T E U 1 † G X [ l ] ( t ) U 1 ] ≃ ∑ X Tr [ ρ i M I T E G X [ l ′ , l ] ( t , t ′ ) ] . (77) \sum_X \operatorname{Tr}[\rho_i^{\mathrm{MITE}}U_1^\dagger G_X^{[l]}(t)U_1] \simeq\sum_X \operatorname{Tr}[\rho_i^{\mathrm{MITE}}G_X^{[l',l]}(t,t')]. \tag*{(77)} X ∑ Tr [ ρ i MITE U 1 † G X [ l ] ( t ) U 1 ] ≃ X ∑ Tr [ ρ i MITE G X [ l ′ , l ] ( t , t ′ )] . (77)
We next consider
∑ X Tr [ ρ i M I T E G X [ l ′ , l ] ( t , t ′ ) ] . (78) \sum_X \operatorname{Tr}[\rho_i^{\mathrm{MITE}}G_X^{[l',l]}(t,t')]. \tag*{(78)} X ∑ Tr [ ρ i MITE G X [ l ′ , l ] ( t , t ′ )] . (78)
Because ρ i M I T E \rho_i^{\mathrm{MITE}} ρ i MITE is an MITE, and from Eq. (76) , the support radius l ′ l' l ′ of G X [ l ′ , l ] ( t , t ′ ) G_X^{[l',l]}(t,t') G X [ l ′ , l ] ( t , t ′ ) satisfies
l ′ = O ( L s ) ( 0 ≤ s < 1 ) , (79) l' = O(L^s) \quad(0 \le s < 1), \tag*{(79)} l ′ = O ( L s ) ( 0 ≤ s < 1 ) , (79)
so, in the thermodynamic limit,
∑ X Tr [ ρ i M I T E G X [ l ′ , l ] ( t , t ′ ) ] ≃ ∑ X Tr [ ρ i e n s G X [ l ′ , l ] ( t , t ′ ) ] . (80) \sum_X \operatorname{Tr}[\rho_i^{\mathrm{MITE}}G_X^{[l',l]}(t,t')] \simeq\sum_X \operatorname{Tr}[\rho_i^{\mathrm{ens}}G_X^{[l',l]}(t,t')]. \tag*{(80)} X ∑ Tr [ ρ i MITE G X [ l ′ , l ] ( t , t ′ )] ≃ X ∑ Tr [ ρ i ens G X [ l ′ , l ] ( t , t ′ )] . (80)
We now summarize the conditions on Z Z Z , l l l , R M R_M R M , and l ′ l' l ′ that have appeared so far. From Eqs. (32) , (50) , (57) , (60) , (63) , and (76) ,
dist ( Z , { y 0 } ) ≤ ℓ , ℓ = O ( L s ) , (81) \operatorname{dist}(Z,\{y_0\}) \le\ell, \qquad\ell= O(L^s), \tag*{(81)} dist ( Z , { y 0 }) ≤ ℓ , ℓ = O ( L s ) , (81)
l = O ( L s ) , t = O ( L s ) , (82) l = O(L^s), \qquad t = O(L^s), \tag*{(82)} l = O ( L s ) , t = O ( L s ) , (82)
R M = l + ℓ + [ D − 1 ] d M = O ( L s ) , (83) R_M = l + \ell+ [\mathcal{D}-1]d_M = O(L^s), \tag*{(83)} R M = l + ℓ + [ D − 1 ] d M = O ( L s ) , (83)
l ′ = O ( L s ) , R M ≤ l ′ , t ′ = O ( L s ) (84) l' = O(L^s), \qquad R_M \le l', \qquad t' = O(L^s) \tag*{(84)} l ′ = O ( L s ) , R M ≤ l ′ , t ′ = O ( L s ) (84)
for 0 ≤ s < 1 0 \le s < 1 0 ≤ s < 1 . These conditions mean that the feedback light cone, the measurement causal cone, and the additional light cone generated by U 1 U_1 U 1 all remain within sublinear regions, when we consider the operator spreading of h X h_X h X . Note that l l l and l ′ l' l ′ are auxiliary scales introduced for the proof, while ℓ \ell ℓ , t t t , t ′ t' t ′ , D \mathcal{D} D , and d M d_M d M are the parameters given in the setup.
If these conditions are satisfied, U 1 † M k † U k ( t ) † h X U k ( t ) M k U 1 U_1^\dagger M_k^\dagger U_k(t)^\dagger h_X U_k(t)M_kU_1 U 1 † M k † U k ( t ) † h X U k ( t ) M k U 1 can be safely approximated as an O ( L s ) O(L^s) O ( L s ) -local observable, and
Tr [ H ρ f ] = ∑ k ∑ X Tr [ ρ i M I T E U 1 † M k † U k ( t ) † h X U k ( t ) M k U 1 ] (85) \operatorname{Tr}[H\rho_f] = \sum_k \sum_X \operatorname{Tr}[\rho_i^{\mathrm{MITE}}U_1^\dagger M_k^\dagger U_k(t)^\dagger h_X U_k(t)M_kU_1] \tag*{(85)} Tr [ H ρ f ] = k ∑ X ∑ Tr [ ρ i MITE U 1 † M k † U k ( t ) † h X U k ( t ) M k U 1 ] (85)
≃ ∑ X Tr [ ρ i e n s G X [ l ′ , l ] ( t , t ′ ) ] (86) \simeq\sum_X \operatorname{Tr}[\rho_i^{\mathrm{ens}}G_X^{[l',l]}(t,t')] \tag*{(86)} ≃ X ∑ Tr [ ρ i ens G X [ l ′ , l ] ( t , t ′ )] (86)
can be used as an approximation. If ρ i M I T E \rho_i^{\mathrm{MITE}} ρ i MITE in Eq. (85) is replaced by ρ i e n s \rho_i^{\mathrm{ens}} ρ i ens , the same argument leading to Eq. (86) applies. Hence,
∑ k ∑ X Tr [ ρ i e n s U 1 † M k † U k ( t ) † h X U k ( t ) M k U 1 ] ≃ ∑ X Tr [ ρ i e n s G X [ l ′ , l ] ( t , t ′ ) ] . (87) \begin{aligned}
\sum_k \sum_X \operatorname{Tr}[\rho_i^{\mathrm{ens}}U_1^\dagger M_k^\dagger U_k(t)^\dagger h_X U_k(t)M_kU_1] \\
&\simeq\sum_X \operatorname{Tr}[\rho_i^{\mathrm{ens}}G_X^{[l',l]}(t,t')]. \tag*{(87)}
\end{aligned} k ∑ X ∑ Tr [ ρ i ens U 1 † M k † U k ( t ) † h X U k ( t ) M k U 1 ] ≃ X ∑ Tr [ ρ i ens G X [ l ′ , l ] ( t , t ′ )] . (87)
From Eqs. (86) and (87) ,
Tr [ H ρ f ] ≃ ∑ k ∑ X Tr [ ρ i e n s U 1 † M k † U k ( t ) † h X U k ( t ) M k U 1 ] (88) \operatorname{Tr}[H\rho_f] \simeq\sum_k \sum_X \operatorname{Tr}[\rho_i^{\mathrm{ens}}U_1^\dagger M_k^\dagger U_k(t)^\dagger h_X U_k(t)M_kU_1] \tag*{(88)} Tr [ H ρ f ] ≃ k ∑ X ∑ Tr [ ρ i ens U 1 † M k † U k ( t ) † h X U k ( t ) M k U 1 ] (88)
holds. Therefore, Eq. (25) becomes
W M I T E = Tr [ H ρ i M I T E ] − Tr [ H ρ f ] ≃ Tr [ H ρ i e n s ] − ∑ k Tr [ ρ i e n s U 1 † M k † U k ( t ) † H U k ( t ) M k U 1 ] . (89) W^{\mathrm{MITE}} = \operatorname{Tr}[H\rho_i^{\mathrm{MITE}}] - \operatorname{Tr}[H\rho_f]
\simeq\operatorname{Tr}[H\rho_i^{\mathrm{ens}}] - \sum_k \operatorname{Tr}[\rho_i^{\mathrm{ens}} U_1^\dagger M_k^\dagger U_k(t)^\dagger H U_k(t) M_k U_1]. \tag*{(89)} W MITE = Tr [ H ρ i MITE ] − Tr [ H ρ f ] ≃ Tr [ H ρ i ens ] − k ∑ Tr [ ρ i ens U 1 † M k † U k ( t ) † H U k ( t ) M k U 1 ] . (89)
This coincides with the work extractable from the canonical state, and therefore proves Eq. (7) . We summarize the preceding discussion as a theorem.
Details for the V-qubit system
We provide the details of work extraction from the V V V -qubit system by measurement and feedback control, as discussed in the main text. As stated in the main text, we consider a cycle and take the Hamiltonian as
H = g ∑ j = 1 V σ j 3 . (91) H = g\sum_{j=1}^{V}\sigma_j^3. \tag*{(91)} H = g j = 1 ∑ V σ j 3 . (91)
Here, σ j 3 \sigma_j^3 σ j 3 denotes the operator in which σ 3 \sigma^3 σ 3 acts only on the j j j th site:
σ j 3 = σ 0 ⊗ σ 0 ⊗ ⋯ ⊗ σ 3 ⊗ σ 0 ⊗ ⋯ ⊗ σ 0 . (92) \sigma_j^3 = \sigma^0 \otimes\sigma^0 \otimes\cdots\otimes\sigma^3 \otimes\sigma^0 \otimes\cdots\otimes\sigma^0. \tag*{(92)} σ j 3 = σ 0 ⊗ σ 0 ⊗ ⋯ ⊗ σ 3 ⊗ σ 0 ⊗ ⋯ ⊗ σ 0 . (92)
Note that there is no interaction between sites. We take the initial state to be the canonical state:
ρ i e n s = e − β H Z = ( ( e − g β 2 cosh ( g β ) 0 0 e g β 2 cosh ( g β ) ) ) ⊗ V . (93) \rho_i^{\mathrm{ens}} = \frac{e^{-\beta H}}{Z} = \left(
\begin{pmatrix}
\frac{e^{-g\beta}}{2\cosh(g\beta)} & 0 \\
0 & \frac{e^{g\beta}}{2\cosh(g\beta)}
\end{pmatrix}
\right)^{\otimes V}. \tag*{(93)} ρ i ens = Z e − β H = ( ( 2 c o s h ( g β ) e − g β 0 0 2 c o s h ( g β ) e g β ) ) ⊗ V . (93)
We apply the operators U 1 U_1 U 1 , M k M_k M k , and U k U_k U k to the initial state ρ i e n s \rho_i^{\mathrm{ens}} ρ i ens . The final state is then
ρ f e n s = ∑ k U k M k U 1 ρ i e n s U 1 † M k † U k † . (94) \rho_f^{\mathrm{ens}} = \sum_k U_k M_k U_1 \rho_i^{\mathrm{ens}} U_1^\dagger M_k^\dagger U_k^\dagger. \tag*{(94)} ρ f ens = k ∑ U k M k U 1 ρ i ens U 1 † M k † U k † . (94)
For simplicity, we take the premeasurement unitary operator to be U 1 = σ 0 ⊗ σ 0 ⊗ ⋯ ⊗ σ 0 U_1 = \sigma^0 \otimes\sigma^0 \otimes\cdots\otimes\sigma^0 U 1 = σ 0 ⊗ σ 0 ⊗ ⋯ ⊗ σ 0 . We describe the measurement operator by the following Kraus operators M ± , ± , … , ± M_{\pm,\pm,\ldots,\pm} M ± , ± , … , ± and assume that it is written as a tensor product of the measurement operators m k l ( η ) m_{k_l}(\eta) m k l ( η ) on the individual qubits:
M ± , ± , … , ± = ⨂ l = 1 V m k l ( η ) . (95) M_{\pm,\pm,\ldots,\pm} = \bigotimes_{l=1}^{V} m_{k_l}(\eta). \tag*{(95)} M ± , ± , … , ± = l = 1 ⨂ V m k l ( η ) . (95)
Here, k l k_l k l corresponds to the measurement outcome of σ 3 \sigma^3 σ 3 at site l l l and can take the value + + + or − - − . Also,
m k l ( η ) = 1 2 ( 1 + η 2 ) ( σ 0 ± η σ 3 ) ( 0 ≤ η ≤ 1 ) . (96) m_{k_l}(\eta)=\frac{1}{\sqrt{2(1+\eta^2)}}\left(\sigma^0\pm\eta\sigma^3\right)\qquad(0\leq\eta\leq1). \tag*{(96)} m k l ( η ) = 2 ( 1 + η 2 ) 1 ( σ 0 ± η σ 3 ) ( 0 ≤ η ≤ 1 ) . (96)
This quantum measurement contains an error when η < 1 \eta<1 η < 1 , and + + + or − - − denotes the measurement outcome of σ 3 \sigma^3 σ 3 in the presence of this error [57 ] . We take the corresponding feedback operator to be
U ± , ± , … , ± = ⨂ l = 1 V u k l , (97) U_{\pm,\pm,\ldots,\pm}=\bigotimes_{l=1}^{V}u_{k_l}, \tag*{(97)} U ± , ± , … , ± = l = 1 ⨂ V u k l , (97)
where u k l u_{k_l} u k l is a unitary operator acting on site l l l . The explicit forms of u + u_+ u + and u − u_- u − will be given below. The definition of k l k_l k l is as given above.
Next, consider the one-qubit density operator
ρ i 1 : = e − β g σ 3 2 cosh ( g β ) = 1 2 cosh ( g β ) ( e − g β 0 0 e g β ) . (98) \rho_i^1:=\frac{e^{-\beta g\sigma^3}}{2\cosh(g\beta)}=\frac{1}{2\cosh(g\beta)}
\begin{pmatrix}
e^{-g\beta} & 0\\
0 & e^{g\beta}
\end{pmatrix}. \tag*{(98)} ρ i 1 := 2 cosh ( g β ) e − β g σ 3 = 2 cosh ( g β ) 1 ( e − g β 0 0 e g β ) . (98)
We use this state in the following. We first consider the case in which the measurement operator m − m_- m − is applied. Then,
m − ρ i 1 m − † = 1 4 ( 1 + η 2 ) cosh ( g β ) [ ( 1 − η ) 2 e − g β ∣ ↑ ⟩ ⟨ ↑ ∣ + ( 1 + η ) 2 e g β ∣ ↓ ⟩ ⟨ ↓ ∣ ] . (99) m_-\rho_i^1m_-^\dagger=\frac{1}{4(1+\eta^2)\cosh(g\beta)}
\left[(1-\eta)^2e^{-g\beta}|\uparrow\rangle\langle\uparrow|+(1+\eta)^2e^{g\beta}|\downarrow\rangle\langle\downarrow|\right]. \tag*{(99)} m − ρ i 1 m − † = 4 ( 1 + η 2 ) cosh ( g β ) 1 [ ( 1 − η ) 2 e − g β ∣ ↑ ⟩ ⟨ ↑ ∣ + ( 1 + η ) 2 e g β ∣ ↓ ⟩ ⟨ ↓ ∣ ] . (99)
Next, we determine the feedback unitary operator u − u_- u − that minimizes
Tr [ u − m − ρ i 1 m − † u − † g σ 3 ] (100) \operatorname{Tr}\left[u_-m_-\rho_i^1m_-^\dagger u_-^\dagger g\sigma^3\right] \tag*{(100)} Tr [ u − m − ρ i 1 m − † u − † g σ 3 ] (100)
with respect to u − u_- u − . For this purpose, we should make the probability of obtaining ∣ ↓ ⟩ ⟨ ↓ ∣ |\downarrow\rangle\langle\downarrow| ∣ ↓ ⟩ ⟨ ↓ ∣ larger than that of obtaining ∣ ↑ ⟩ ⟨ ↑ ∣ |\uparrow\rangle\langle\uparrow| ∣ ↑ ⟩ ⟨ ↑ ∣ . Therefore, the feedback operator u − u_- u − to be constructed may be chosen as
u − = c σ 0 (101) u_-=c\sigma^0 \tag*{(101)} u − = c σ 0 (101)
with ∣ c ∣ 2 = 1 |c|^2=1 ∣ c ∣ 2 = 1 . Under the feedback operation implemented in this way,
u − m − ρ i 1 m − † u − † = 1 4 ( 1 + η 2 ) cosh ( g β ) [ ( 1 − η ) 2 e − g β ∣ ↑ ⟩ ⟨ ↑ ∣ + ( 1 + η ) 2 e g β ∣ ↓ ⟩ ⟨ ↓ ∣ ] (102) u_-m_-\rho_i^1m_-^\dagger u_-^\dagger=\frac{1}{4(1+\eta^2)\cosh(g\beta)}
\left[(1-\eta)^2e^{-g\beta}|\uparrow\rangle\langle\uparrow|+(1+\eta)^2e^{g\beta}|\downarrow\rangle\langle\downarrow|\right] \tag*{(102)} u − m − ρ i 1 m − † u − † = 4 ( 1 + η 2 ) cosh ( g β ) 1 [ ( 1 − η ) 2 e − g β ∣ ↑ ⟩ ⟨ ↑ ∣ + ( 1 + η ) 2 e g β ∣ ↓ ⟩ ⟨ ↓ ∣ ] (102)
is obtained.
Next, consider the case in which the measurement operator m + m_+ m + is applied. Then,
m + ρ i 1 m + † = 1 4 ( 1 + η 2 ) cosh ( g β ) [ ( 1 + η ) 2 e − g β ∣ ↑ ⟩ ⟨ ↑ ∣ + ( 1 − η ) 2 e g β ∣ ↓ ⟩ ⟨ ↓ ∣ ] . (103) m_+\rho_i^1m_+^\dagger=\frac{1}{4(1+\eta^2)\cosh(g\beta)}
\left[(1+\eta)^2e^{-g\beta}|\uparrow\rangle\langle\uparrow|+(1-\eta)^2e^{g\beta}|\downarrow\rangle\langle\downarrow|\right]. \tag*{(103)} m + ρ i 1 m + † = 4 ( 1 + η 2 ) cosh ( g β ) 1 [ ( 1 + η ) 2 e − g β ∣ ↑ ⟩ ⟨ ↑ ∣ + ( 1 − η ) 2 e g β ∣ ↓ ⟩ ⟨ ↓ ∣ ] . (103)
Therefore, the optimal feedback operator to be constructed depends on the following cases.
(i) Case ( 1 + η ) 2 e − g β ≥ ( 1 − η ) 2 e g β (1+\eta)^2e^{-g\beta}\geq(1-\eta)^2e^{g\beta} ( 1 + η ) 2 e − g β ≥ ( 1 − η ) 2 e g β (high temperature or high measurement accuracy)
In this case, the feedback unitary operator is
u + = c ( ∣ ↓ ⟩ ⟨ ↑ ∣ + ∣ ↑ ⟩ ⟨ ↓ ∣ ) = c σ 1 , (104) u_+=c\left(|\downarrow\rangle\langle\uparrow|+|\uparrow\rangle\langle\downarrow|\right)=c\sigma^1, \tag*{(104)} u + = c ( ∣ ↓ ⟩ ⟨ ↑ ∣ + ∣ ↑ ⟩ ⟨ ↓ ∣ ) = c σ 1 , (104)
where ∣ c ∣ 2 = 1 |c|^2=1 ∣ c ∣ 2 = 1 . This feedback unitary operator can be realized by applying a magnetic field, as explained below. If we set
H + = h ′ σ 1 , (105) H_+=h'\sigma^1, \tag*{(105)} H + = h ′ σ 1 , (105)
then
u + = exp ( − i ∫ 0 t H + d t ′ ) = cos ( h ′ t ) σ 0 − i sin ( h ′ t ) σ 1 . (106) u_+=\exp\left(-i\int_0^t H_+\,dt'\right)=\cos(h't)\sigma^0-i\sin(h't)\sigma^1. \tag*{(106)} u + = exp ( − i ∫ 0 t H + d t ′ ) = cos ( h ′ t ) σ 0 − i sin ( h ′ t ) σ 1 . (106)
Thus, if the external field h ′ h' h ′ is adjusted according to the feedback-control time t t t so as to satisfy
cos ( h ′ t ) = 0 , sin ( h ′ t ) = ∓ 1 , (107) \cos(h't)=0,\qquad\sin(h't)=\mp1, \tag*{(107)} cos ( h ′ t ) = 0 , sin ( h ′ t ) = ∓ 1 , (107)
then the desired unitary operator is obtained, with c = ± i c=\pm i c = ± i . Under the feedback operation implemented in this way,
u + m + ρ i 1 m + † u + † = 1 4 ( 1 + η 2 ) cosh ( g β ) [ ( 1 + η ) 2 e − g β ∣ ↓ ⟩ ⟨ ↓ ∣ + ( 1 − η ) 2 e g β ∣ ↑ ⟩ ⟨ ↑ ∣ ] . (108) u_{+}m_{+}\rho_i^{1}m_{+}^{\dagger}u_{+}^{\dagger}=\frac{1}{4(1+\eta^{2})\cosh(g\beta)}\left[(1+\eta)^{2}e^{-g\beta}|\downarrow\rangle\langle\downarrow|+(1-\eta)^{2}e^{g\beta}|\uparrow\rangle\langle\uparrow|\right]. \tag*{(108)} u + m + ρ i 1 m + † u + † = 4 ( 1 + η 2 ) cosh ( g β ) 1 [ ( 1 + η ) 2 e − g β ∣ ↓ ⟩ ⟨ ↓ ∣ + ( 1 − η ) 2 e g β ∣ ↑ ⟩ ⟨ ↑ ∣ ] . (108)
Therefore,
ρ f 1 : = u + m + u 1 ρ i 1 u 1 † m + † u + † + u − m − u − 1 ρ i 1 u − 1 † m − † u − † = 1 2 ( 1 + η 2 ) [ ( 1 − η ) 2 ∣ ↑ ⟩ ⟨ ↑ ∣ + ( 1 + η ) 2 ∣ ↓ ⟩ ⟨ ↓ ∣ ] . (109) \begin{aligned}
\rho_f^{1}:={}&u_{+}m_{+}u_{1}\rho_i^{1}u_{1}^{\dagger}m_{+}^{\dagger}u_{+}^{\dagger}+u_{-}m_{-}u_{-1}\rho_i^{1}u_{-1}^{\dagger}m_{-}^{\dagger}u_{-}^{\dagger}\\
={}&\frac{1}{2(1+\eta^{2})}\left[(1-\eta)^{2}|\uparrow\rangle\langle\uparrow|+(1+\eta)^{2}|\downarrow\rangle\langle\downarrow|\right]. \tag*{(109)}
\end{aligned} ρ f 1 := = u + m + u 1 ρ i 1 u 1 † m + † u + † + u − m − u − 1 ρ i 1 u − 1 † m − † u − † 2 ( 1 + η 2 ) 1 [ ( 1 − η ) 2 ∣ ↑ ⟩ ⟨ ↑ ∣ + ( 1 + η ) 2 ∣ ↓ ⟩ ⟨ ↓ ∣ ] . (109)
and hence
Tr [ H f 1 ρ f 1 ] = − 2 η g 1 + η 2 , (110) \operatorname{Tr}[H_f^{1}\rho_f^{1}]=-\frac{2\eta g}{1+\eta^{2}}, \tag*{(110)} Tr [ H f 1 ρ f 1 ] = − 1 + η 2 2 η g , (110)
where
H f 1 : = g σ 3 . (111) H_f^{1}:=g\sigma^{3}. \tag*{(111)} H f 1 := g σ 3 . (111)
Since we consider noninteracting V V V qubits, we have
Tr [ ρ f H f ] = − 2 η g 1 + η 2 V . (112) \operatorname{Tr}[\rho_f H_f]=-\frac{2\eta g}{1+\eta^{2}}V. \tag*{(112)} Tr [ ρ f H f ] = − 1 + η 2 2 η g V . (112)
Moreover,
Tr [ ρ i e n s H i ] = − g V tanh ( g β ) . (113) \operatorname{Tr}[\rho_i^{\mathrm{ens}}H_i]=-gV\tanh(g\beta). \tag*{(113)} Tr [ ρ i ens H i ] = − g V tanh ( g β ) . (113)
Therefore, the extractable work is
W o p t , l o c = Tr [ H i ρ i ] − Tr [ H f ρ f ] = − g V tanh ( g β ) + 2 η g 1 + η 2 V . (114) W_{\mathrm{opt,loc}}=\operatorname{Tr}[H_i\rho_i]-\operatorname{Tr}[H_f\rho_f]=-gV\tanh(g\beta)+\frac{2\eta g}{1+\eta^{2}}V. \tag*{(114)} W opt , loc = Tr [ H i ρ i ] − Tr [ H f ρ f ] = − g V tanh ( g β ) + 1 + η 2 2 η g V . (114)
Next, we compute the corresponding information quantities, namely, the mutual information, the Holevo information, and the relative entropy. We also compute their sum, which is the right-hand side of Eq. (8) , and check that it indeed agrees with the extractable work.
First, we compute ∑ k p k S ( ρ k ′ ) \sum_k p_k S(\rho'_k) ∑ k p k S ( ρ k ′ ) . Since we consider noninteracting tensor-product structure, we find
P k = Tr [ M k M k † ρ i e n s ] : = ∏ l = 1 V p k l . (115) P_k=\operatorname{Tr}[M_kM_k^{\dagger}\rho_i^{\mathrm{ens}}]:=\prod_{l=1}^{V}p_{kl}. \tag*{(115)} P k = Tr [ M k M k † ρ i ens ] := l = 1 ∏ V p k l . (115)
with
p + = Tr [ m + † m + ρ i 1 ] = ( 1 + η ) 2 e − g β + ( 1 − η ) 2 e g β 4 ( 1 + η 2 ) cosh ( g β ) , (116) p_{+}=\operatorname{Tr}[m_{+}^{\dagger}m_{+}\rho_i^{1}]=\frac{(1+\eta)^{2}e^{-g\beta}+(1-\eta)^{2}e^{g\beta}}{4(1+\eta^{2})\cosh(g\beta)}, \tag*{(116)} p + = Tr [ m + † m + ρ i 1 ] = 4 ( 1 + η 2 ) cosh ( g β ) ( 1 + η ) 2 e − g β + ( 1 − η ) 2 e g β , (116)
p − = 1 − p + = ( 1 − η ) 2 e − g β + ( 1 + η ) 2 e g β 4 ( 1 + η 2 ) cosh ( g β ) . (117) p_{-}=1-p_{+}=\frac{(1-\eta)^{2}e^{-g\beta}+(1+\eta)^{2}e^{g\beta}}{4(1+\eta^{2})\cosh(g\beta)}. \tag*{(117)} p − = 1 − p + = 4 ( 1 + η 2 ) cosh ( g β ) ( 1 − η ) 2 e − g β + ( 1 + η ) 2 e g β . (117)
Also,
S ( ρ k ′ ) = S ( U k M k ρ i e n s M k † U k † P k ) = ∑ l = 1 V S ( u k l m k l ρ i 1 m k l † u k l † p k l ) = ∑ l = 1 V S ( ρ k l ′ ) . (118) S(\rho'_k)=S\left(\frac{U_kM_k\rho_i^{\mathrm{ens}}M_k^{\dagger}U_k^{\dagger}}{P_k}\right)=\sum_{l=1}^{V}S\left(\frac{u_{kl}m_{kl}\rho_i^{1}m_{kl}^{\dagger}u_{kl}^{\dagger}}{p_{kl}}\right)=\sum_{l=1}^{V}S(\rho'_{kl}). \tag*{(118)} S ( ρ k ′ ) = S ( P k U k M k ρ i ens M k † U k † ) = l = 1 ∑ V S ( p k l u k l m k l ρ i 1 m k l † u k l † ) = l = 1 ∑ V S ( ρ k l ′ ) . (118)
Thus,
∑ k P k S ( ρ k ′ ) = V ( ∑ k 1 = ± p k 1 S ( ρ k 1 1 ) ) . (119) \sum_k P_k S(\rho'_k)=V\left(\sum_{k_1=\pm}p_{k_1}S(\rho^1_{k_1})\right). \tag*{(119)} k ∑ P k S ( ρ k ′ ) = V ( k 1 = ± ∑ p k 1 S ( ρ k 1 1 ) ) . (119)
From Eq. (108) ,
ρ + 1 = u + m + u 1 ρ i 1 u 1 † m + † u + † p + = 1 4 p + ( 1 + η 2 ) cosh ( g β ) [ ( 1 + η ) 2 e − g β ∣ ↓ ⟩ ⟨ ↓ ∣ + ( 1 − η ) 2 e g β ∣ ↑ ⟩ ⟨ ↑ ∣ ] . (120) \rho^1_+=\frac{u_+m_+u_1\rho^1_i u_1^\dagger m_+^\dagger u_+^\dagger}{p_+}
=\frac{1}{4p_+(1+\eta^2)\cosh(g\beta)}\left[(1+\eta)^2e^{-g\beta}|\downarrow\rangle\langle\downarrow|+(1-\eta)^2e^{g\beta}|\uparrow\rangle\langle\uparrow|\right]. \tag*{(120)} ρ + 1 = p + u + m + u 1 ρ i 1 u 1 † m + † u + † = 4 p + ( 1 + η 2 ) cosh ( g β ) 1 [ ( 1 + η ) 2 e − g β ∣ ↓ ⟩ ⟨ ↓ ∣ + ( 1 − η ) 2 e g β ∣ ↑ ⟩ ⟨ ↑ ∣ ] . (120)
Similarly, from Eq. (102) ,
ρ − 1 = u − m − u 1 ρ i 1 u 1 † m − † u − † p − = 1 4 p − ( 1 + η 2 ) cosh ( g β ) [ ( 1 − η ) 2 e − g β ∣ ↑ ⟩ ⟨ ↑ ∣ + ( 1 + η ) 2 e g β ∣ ↓ ⟩ ⟨ ↓ ∣ ] . (121) \rho^1_-=\frac{u_-m_-u_1\rho^1_i u_1^\dagger m_-^\dagger u_-^\dagger}{p_-}
=\frac{1}{4p_-(1+\eta^2)\cosh(g\beta)}\left[(1-\eta)^2e^{-g\beta}|\uparrow\rangle\langle\uparrow|+(1+\eta)^2e^{g\beta}|\downarrow\rangle\langle\downarrow|\right]. \tag*{(121)} ρ − 1 = p − u − m − u 1 ρ i 1 u 1 † m − † u − † = 4 p − ( 1 + η 2 ) cosh ( g β ) 1 [ ( 1 − η ) 2 e − g β ∣ ↑ ⟩ ⟨ ↑ ∣ + ( 1 + η ) 2 e g β ∣ ↓ ⟩ ⟨ ↓ ∣ ] . (121)
Hence, after some calculation, we obtain
∑ k P k S ( ρ k ′ ) = V ( ∑ k 1 = ± p k 1 S ( ρ k 1 1 ) ) = V ( − ( 1 + η ) 2 e − g β 4 ( 1 + η 2 ) cosh ( g β ) log [ ( 1 + η ) 2 e − g β ( 1 + η ) 2 e − g β + ( 1 − η ) 2 e g β ] − ( 1 − η ) 2 e − g β 4 ( 1 + η 2 ) cosh ( g β ) log [ ( 1 − η ) 2 e − g β ( 1 − η ) 2 e − g β + ( 1 + η ) 2 e g β ] − ( 1 − η ) 2 e g β 4 ( 1 + η 2 ) cosh ( g β ) log [ ( 1 − η ) 2 e g β ( 1 + η ) 2 e − g β + ( 1 − η ) 2 e g β ] − ( 1 + η ) 2 e g β 4 ( 1 + η 2 ) cosh ( g β ) log [ ( 1 + η ) 2 e g β ( 1 − η ) 2 e − g β + ( 1 + η ) 2 e g β ] ) . (122) \begin{aligned}
\sum_k P_kS(\rho'_k)&=V\left(\sum_{k_1=\pm}p_{k_1}S(\rho^1_{k_1})\right)\\
&=V\left(-\frac{(1+\eta)^2e^{-g\beta}}{4(1+\eta^2)\cosh(g\beta)}\log\left[\frac{(1+\eta)^2e^{-g\beta}}{(1+\eta)^2e^{-g\beta}+(1-\eta)^2e^{g\beta}}\right]\right.\\
&\quad-\frac{(1-\eta)^2e^{-g\beta}}{4(1+\eta^2)\cosh(g\beta)}\log\left[\frac{(1-\eta)^2e^{-g\beta}}{(1-\eta)^2e^{-g\beta}+(1+\eta)^2e^{g\beta}}\right]\\
&\quad-\frac{(1-\eta)^2e^{g\beta}}{4(1+\eta^2)\cosh(g\beta)}\log\left[\frac{(1-\eta)^2e^{g\beta}}{(1+\eta)^2e^{-g\beta}+(1-\eta)^2e^{g\beta}}\right]\\
&\quad\left.-\frac{(1+\eta)^2e^{g\beta}}{4(1+\eta^2)\cosh(g\beta)}\log\left[\frac{(1+\eta)^2e^{g\beta}}{(1-\eta)^2e^{-g\beta}+(1+\eta)^2e^{g\beta}}\right]\right). \tag*{(122)}
\end{aligned} k ∑ P k S ( ρ k ′ ) = V ( k 1 = ± ∑ p k 1 S ( ρ k 1 1 ) ) = V ( − 4 ( 1 + η 2 ) cosh ( g β ) ( 1 + η ) 2 e − g β log [ ( 1 + η ) 2 e − g β + ( 1 − η ) 2 e g β ( 1 + η ) 2 e − g β ] − 4 ( 1 + η 2 ) cosh ( g β ) ( 1 − η ) 2 e − g β log [ ( 1 − η ) 2 e − g β + ( 1 + η ) 2 e g β ( 1 − η ) 2 e − g β ] − 4 ( 1 + η 2 ) cosh ( g β ) ( 1 − η ) 2 e g β log [ ( 1 + η ) 2 e − g β + ( 1 − η ) 2 e g β ( 1 − η ) 2 e g β ] − 4 ( 1 + η 2 ) cosh ( g β ) ( 1 + η ) 2 e g β log [ ( 1 − η ) 2 e − g β + ( 1 + η ) 2 e g β ( 1 + η ) 2 e g β ] ) . (122)
Next, we compute S ( ρ i e n s ) S(\rho_i^{\mathrm{ens}}) S ( ρ i ens ) :
S ( ρ i e n s ) = V S ( ρ i 1 ) = V ( − e − g β 2 cosh ( g β ) log [ e − g β 2 cosh ( g β ) ] − e g β 2 cosh ( g β ) log [ e g β 2 cosh ( g β ) ] ) . (123) S(\rho_i^{\mathrm{ens}})=VS(\rho_i^1)=V\left(-\frac{e^{-g\beta}}{2\cosh(g\beta)}\log\left[\frac{e^{-g\beta}}{2\cosh(g\beta)}\right]-\frac{e^{g\beta}}{2\cosh(g\beta)}\log\left[\frac{e^{g\beta}}{2\cosh(g\beta)}\right]\right). \tag*{(123)} S ( ρ i ens ) = V S ( ρ i 1 ) = V ( − 2 cosh ( g β ) e − g β log [ 2 cosh ( g β ) e − g β ] − 2 cosh ( g β ) e g β log [ 2 cosh ( g β ) e g β ] ) . (123)
We can also compute S ( ρ f ) S(\rho_f) S ( ρ f ) , represented as
S ( ρ f ) = V S [ ∑ k 1 = ± u k 1 m k 1 ρ i 1 m k 1 † u k 1 † ] . (124) S(\rho_f)=VS\left[\sum_{k_1=\pm}u_{k_1}m_{k_1}\rho_i^1m_{k_1}^\dagger u_{k_1}^\dagger\right]. \tag*{(124)} S ( ρ f ) = V S [ k 1 = ± ∑ u k 1 m k 1 ρ i 1 m k 1 † u k 1 † ] . (124)
Using Eqs. (108) and (102) , we have
S ( ρ f ) = V S [ 1 4 ( 1 + η 2 ) cosh ( g β ) ( 2 ( 1 + η ) 2 cosh ( g β ) ∣ ↓ ⟩ ⟨ ↓ ∣ + 2 ( 1 − η ) 2 cosh ( g β ) ∣ ↑ ⟩ ⟨ ↑ ∣ ) ] = V [ − ( 1 + η ) 2 2 ( 1 + η 2 ) log ( ( 1 + η ) 2 2 ( 1 + η 2 ) ) − ( 1 − η ) 2 2 ( 1 + η 2 ) log ( ( 1 − η ) 2 2 ( 1 + η 2 ) ) ] . (125) \begin{aligned}
S(\rho_f)&=VS\left[\frac{1}{4(1+\eta^2)\cosh(g\beta)}\left(2(1+\eta)^2\cosh(g\beta)|\downarrow\rangle\langle\downarrow|+2(1-\eta)^2\cosh(g\beta)|\uparrow\rangle\langle\uparrow|\right)\right]\\
&=V\left[-\frac{(1+\eta)^2}{2(1+\eta^2)}\log\left(\frac{(1+\eta)^2}{2(1+\eta^2)}\right)-\frac{(1-\eta)^2}{2(1+\eta^2)}\log\left(\frac{(1-\eta)^2}{2(1+\eta^2)}\right)\right]. \tag*{(125)}
\end{aligned} S ( ρ f ) = V S [ 4 ( 1 + η 2 ) cosh ( g β ) 1 ( 2 ( 1 + η ) 2 cosh ( g β ) ∣ ↓ ⟩ ⟨ ↓ ∣ + 2 ( 1 − η ) 2 cosh ( g β ) ∣ ↑ ⟩ ⟨ ↑ ∣ ) ] = V [ − 2 ( 1 + η 2 ) ( 1 + η ) 2 log ( 2 ( 1 + η 2 ) ( 1 + η ) 2 ) − 2 ( 1 + η 2 ) ( 1 − η ) 2 log ( 2 ( 1 + η 2 ) ( 1 − η ) 2 ) ] . (125)
From the above results, the mutual information I Q C I_{\mathrm{QC}} I QC is
I Q C = S ( ρ i e n s ) − ∑ k P k S ( ρ k ′ ) = V ( − e − g β 2 cosh ( g β ) log [ e − g β 2 cosh ( g β ) ] − e g β 2 cosh ( g β ) log [ e g β 2 cosh ( g β ) ] ) − V ( − ( 1 + η ) 2 e − g β 4 ( 1 + η 2 ) cosh ( g β ) log [ ( 1 + η ) 2 e − g β ( 1 + η ) 2 e − g β + ( 1 − η ) 2 e g β ] − ( 1 − η ) 2 e − g β 4 ( 1 + η 2 ) cosh ( g β ) log [ ( 1 − η ) 2 e − g β ( 1 − η ) 2 e − g β + ( 1 + η ) 2 e g β ] − ( 1 − η ) 2 e g β 4 ( 1 + η 2 ) cosh ( g β ) log [ ( 1 − η ) 2 e g β ( 1 + η ) 2 e − g β + ( 1 − η ) 2 e g β ] − ( 1 + η ) 2 e g β 4 ( 1 + η 2 ) cosh ( g β ) log [ ( 1 + η ) 2 e g β ( 1 − η ) 2 e − g β + ( 1 + η ) 2 e g β ] ) . (126) \begin{aligned}
I_{\mathrm{QC}}&=S(\rho_i^{\mathrm{ens}})-\sum_kP_kS(\rho'_k)\\
&=V\left(-\frac{e^{-g\beta}}{2\cosh(g\beta)}\log\left[\frac{e^{-g\beta}}{2\cosh(g\beta)}\right]-\frac{e^{g\beta}}{2\cosh(g\beta)}\log\left[\frac{e^{g\beta}}{2\cosh(g\beta)}\right]\right)\\
&\quad-V\left(-\frac{(1+\eta)^2e^{-g\beta}}{4(1+\eta^2)\cosh(g\beta)}\log\left[\frac{(1+\eta)^2e^{-g\beta}}{(1+\eta)^2e^{-g\beta}+(1-\eta)^2e^{g\beta}}\right]\right.\\
&\qquad-\frac{(1-\eta)^2e^{-g\beta}}{4(1+\eta^2)\cosh(g\beta)}\log\left[\frac{(1-\eta)^2e^{-g\beta}}{(1-\eta)^2e^{-g\beta}+(1+\eta)^2e^{g\beta}}\right]\\
&\qquad-\frac{(1-\eta)^2e^{g\beta}}{4(1+\eta^2)\cosh(g\beta)}\log\left[\frac{(1-\eta)^2e^{g\beta}}{(1+\eta)^2e^{-g\beta}+(1-\eta)^2e^{g\beta}}\right]\\
&\qquad\left.-\frac{(1+\eta)^2e^{g\beta}}{4(1+\eta^2)\cosh(g\beta)}\log\left[\frac{(1+\eta)^2e^{g\beta}}{(1-\eta)^2e^{-g\beta}+(1+\eta)^2e^{g\beta}}\right]\right). \tag*{(126)}
\end{aligned} I QC = S ( ρ i ens ) − k ∑ P k S ( ρ k ′ ) = V ( − 2 cosh ( g β ) e − g β log [ 2 cosh ( g β ) e − g β ] − 2 cosh ( g β ) e g β log [ 2 cosh ( g β ) e g β ] ) − V ( − 4 ( 1 + η 2 ) cosh ( g β ) ( 1 + η ) 2 e − g β log [ ( 1 + η ) 2 e − g β + ( 1 − η ) 2 e g β ( 1 + η ) 2 e − g β ] − 4 ( 1 + η 2 ) cosh ( g β ) ( 1 − η ) 2 e − g β log [ ( 1 − η ) 2 e − g β + ( 1 + η ) 2 e g β ( 1 − η ) 2 e − g β ] − 4 ( 1 + η 2 ) cosh ( g β ) ( 1 − η ) 2 e g β log [ ( 1 + η ) 2 e − g β + ( 1 − η ) 2 e g β ( 1 − η ) 2 e g β ] − 4 ( 1 + η 2 ) cosh ( g β ) ( 1 + η ) 2 e g β log [ ( 1 − η ) 2 e − g β + ( 1 + η ) 2 e g β ( 1 + η ) 2 e g β ] ) . (126)
For example, for η = 1 \eta= 1 η = 1 (an error-free measurement), the mutual information I Q C I_{\mathrm{QC}} I QC becomes equal to the von Neumann entropy S ( ρ i ) S(\rho_i) S ( ρ i ) of the initial state. For η = 0 \eta= 0 η = 0 (a maximally erroneous measurement), the mutual information I Q C I_{\mathrm{QC}} I QC becomes vanishing.
Next, we compute the Holevo information.
χ = S ( ρ f ) − ∑ k P k S ( ρ k ′ ) = V [ − ( 1 + η ) 2 2 ( 1 + η 2 ) log ( ( 1 + η ) 2 2 ( 1 + η 2 ) ) − ( 1 − η ) 2 2 ( 1 + η 2 ) log ( ( 1 − η ) 2 2 ( 1 + η 2 ) ) ] − V ( ( 1 + η ) 2 e − g β 4 ( 1 + η 2 ) cosh ( g β ) log [ ( 1 + η ) 2 e − g β ( 1 + η ) 2 e − g β + ( 1 − η ) 2 e g β ] + ( 1 − η ) 2 e − g β 4 ( 1 + η 2 ) cosh ( g β ) log [ ( 1 − η ) 2 e − g β ( 1 − η ) 2 e − g β + ( 1 + η ) 2 e g β ] + ( 1 − η ) 2 e g β 4 ( 1 + η 2 ) cosh ( g β ) log [ ( 1 − η ) 2 e g β ( 1 + η ) 2 e − g β + ( 1 − η ) 2 e g β ] + ( 1 + η ) 2 e g β 4 ( 1 + η 2 ) cosh ( g β ) log [ ( 1 + η ) 2 e g β ( 1 − η ) 2 e − g β + ( 1 + η ) 2 e g β ] ) . (127) \begin{aligned}
\chi&= S(\rho_f)-\sum_k P_k S(\rho'_k)\\
&=V\left[-\frac{(1+\eta)^2}{2(1+\eta^2)}\log\left(\frac{(1+\eta)^2}{2(1+\eta^2)}\right)-\frac{(1-\eta)^2}{2(1+\eta^2)}\log\left(\frac{(1-\eta)^2}{2(1+\eta^2)}\right)\right]\\
&\quad-V\left(\frac{(1+\eta)^2e^{-g\beta}}{4(1+\eta^2)\cosh(g\beta)}\log\left[\frac{(1+\eta)^2e^{-g\beta}}{(1+\eta)^2e^{-g\beta}+(1-\eta)^2e^{g\beta}}\right]\right.\\
&\qquad\quad+\frac{(1-\eta)^2e^{-g\beta}}{4(1+\eta^2)\cosh(g\beta)}\log\left[\frac{(1-\eta)^2e^{-g\beta}}{(1-\eta)^2e^{-g\beta}+(1+\eta)^2e^{g\beta}}\right]\\
&\qquad\quad+\frac{(1-\eta)^2e^{g\beta}}{4(1+\eta^2)\cosh(g\beta)}\log\left[\frac{(1-\eta)^2e^{g\beta}}{(1+\eta)^2e^{-g\beta}+(1-\eta)^2e^{g\beta}}\right]\\
&\qquad\quad\left.+\frac{(1+\eta)^2e^{g\beta}}{4(1+\eta^2)\cosh(g\beta)}\log\left[\frac{(1+\eta)^2e^{g\beta}}{(1-\eta)^2e^{-g\beta}+(1+\eta)^2e^{g\beta}}\right]\right). \tag*{(127)}
\end{aligned} χ = S ( ρ f ) − k ∑ P k S ( ρ k ′ ) = V [ − 2 ( 1 + η 2 ) ( 1 + η ) 2 log ( 2 ( 1 + η 2 ) ( 1 + η ) 2 ) − 2 ( 1 + η 2 ) ( 1 − η ) 2 log ( 2 ( 1 + η 2 ) ( 1 − η ) 2 ) ] − V ( 4 ( 1 + η 2 ) cosh ( g β ) ( 1 + η ) 2 e − g β log [ ( 1 + η ) 2 e − g β + ( 1 − η ) 2 e g β ( 1 + η ) 2 e − g β ] + 4 ( 1 + η 2 ) cosh ( g β ) ( 1 − η ) 2 e − g β log [ ( 1 − η ) 2 e − g β + ( 1 + η ) 2 e g β ( 1 − η ) 2 e − g β ] + 4 ( 1 + η 2 ) cosh ( g β ) ( 1 − η ) 2 e g β log [ ( 1 + η ) 2 e − g β + ( 1 − η ) 2 e g β ( 1 − η ) 2 e g β ] + 4 ( 1 + η 2 ) cosh ( g β ) ( 1 + η ) 2 e g β log [ ( 1 − η ) 2 e − g β + ( 1 + η ) 2 e g β ( 1 + η ) 2 e g β ] ) . (127)
For example, for the infinite-temperature state (β = 0 \beta= 0 β = 0 ), the Holevo information becomes zero.
Next, we compute the relative entropy. From Eq. (112) ,
D ( ρ f e n s ∥ ρ i e n s ) = − S ( ρ f ) + β Tr [ H f ρ f ] + log Z = − V [ − ( 1 + η ) 2 2 ( 1 + η 2 ) log ( ( 1 + η ) 2 2 ( 1 + η 2 ) ) − ( 1 − η ) 2 2 ( 1 + η 2 ) log ( ( 1 − η ) 2 2 ( 1 + η 2 ) ) ] + β − 2 η g 1 + η 2 V + V log ( 2 cosh ( g β ) ) . (128) \begin{aligned}
D(\rho_f^{\mathrm{ens}}\Vert\rho_i^{\mathrm{ens}})&=-S(\rho_f)+\beta\operatorname{Tr}[H_f\rho_f]+\log Z\\
&=-V\left[-\frac{(1+\eta)^2}{2(1+\eta^2)}\log\left(\frac{(1+\eta)^2}{2(1+\eta^2)}\right)-\frac{(1-\eta)^2}{2(1+\eta^2)}\log\left(\frac{(1-\eta)^2}{2(1+\eta^2)}\right)\right]+\beta\frac{-2\eta g}{1+\eta^2}V+V\log(2\cosh(g\beta)). \tag*{(128)}
\end{aligned} D ( ρ f ens ∥ ρ i ens ) = − S ( ρ f ) + β Tr [ H f ρ f ] + log Z = − V [ − 2 ( 1 + η 2 ) ( 1 + η ) 2 log ( 2 ( 1 + η 2 ) ( 1 + η ) 2 ) − 2 ( 1 + η 2 ) ( 1 − η ) 2 log ( 2 ( 1 + η 2 ) ( 1 − η ) 2 ) ] + β 1 + η 2 − 2 η g V + V log ( 2 cosh ( g β )) . (128)
For example, for an infinite-temperature state with an error-free measurement (β = 0 \beta= 0 β = 0 , η = 1 \eta= 1 η = 1 ), the relative entropy becomes
D ( ρ f e n s ∥ ρ i e n s ) = V log 2. (129) D(\rho_f^{\mathrm{ens}}\Vert\rho_i^{\mathrm{ens}})=V\log2. \tag*{(129)} D ( ρ f ens ∥ ρ i ens ) = V log 2. (129)
Finally, we compute the sum of the information quantities, i.e., the right-hand side of Eq. (8) , and check that it indeed agrees with the extractable work.
I Q C − χ − D ( ρ f e n s ∥ ρ i e n s ) = V ( − e − g β 2 cosh ( g β ) log [ e − g β 2 cosh ( g β ) ] − e g β 2 cosh ( g β ) log [ e g β 2 cosh ( g β ) ] ) − β − 2 η g 1 + η 2 V − V log ( 2 cosh ( g β ) ) = V ( − g β tanh ( g β ) + log ( 2 cosh ( g β ) ) ) − β − 2 η g 1 + η 2 V − V log ( 2 cosh ( g β ) ) = β [ − g V tanh ( g β ) + 2 η g 1 + η 2 V ] . (130) \begin{aligned}
I_{\mathrm{QC}}-\chi-D(\rho_f^{\mathrm{ens}}\Vert\rho_i^{\mathrm{ens}})
&=V\left(-\frac{e^{-g\beta}}{2\cosh(g\beta)}\log\left[\frac{e^{-g\beta}}{2\cosh(g\beta)}\right]-\frac{e^{g\beta}}{2\cosh(g\beta)}\log\left[\frac{e^{g\beta}}{2\cosh(g\beta)}\right]\right)-\beta\frac{-2\eta g}{1+\eta^2}V-V\log(2\cosh(g\beta))\\
&=V(-g\beta\tanh(g\beta)+\log(2\cosh(g\beta)))-\beta\frac{-2\eta g}{1+\eta^2}V-V\log(2\cosh(g\beta))\\
&=\beta\left[-gV\tanh(g\beta)+\frac{2\eta g}{1+\eta^2}V\right]. \tag*{(130)}
\end{aligned} I QC − χ − D ( ρ f ens ∥ ρ i ens ) = V ( − 2 cosh ( g β ) e − g β log [ 2 cosh ( g β ) e − g β ] − 2 cosh ( g β ) e g β log [ 2 cosh ( g β ) e g β ] ) − β 1 + η 2 − 2 η g V − V log ( 2 cosh ( g β )) = V ( − g β tanh ( g β ) + log ( 2 cosh ( g β ))) − β 1 + η 2 − 2 η g V − V log ( 2 cosh ( g β )) = β [ − g V tanh ( g β ) + 1 + η 2 2 η g V ] . (130)
which agrees with the extractable work in Eq. (114) .
(ii) Case ( 1 + η ) 2 e − g β ≤ ( 1 − η ) 2 e g β (1+\eta)^2e^{-g\beta}\leq(1-\eta)^2e^{g\beta} ( 1 + η ) 2 e − g β ≤ ( 1 − η ) 2 e g β (low temperature or low measurement accuracy)
In this case, the feedback unitary operator u + u_+ u + is
u + = c σ 0 , (131) u_+=c\sigma^0, \tag*{(131)} u + = c σ 0 , (131)
where ∣ c ∣ 2 = 1 \lvert c\rvert^2=1 ∣ c ∣ 2 = 1 . Under this feedback operation,
u + m + ρ i 1 m + † u + † = 1 4 ( 1 + η 2 ) cosh ( g β ) [ ( 1 + η ) 2 e − g β ∣ ↑ ⟩ ⟨ ↑ ∣ + ( 1 − η ) 2 e g β ∣ ↓ ⟩ ⟨ ↓ ∣ ] . (132) u_+m_+\rho_i^1m_+^\dagger u_+^\dagger=\frac{1}{4(1+\eta^2)\cosh(g\beta)}\left[(1+\eta)^2e^{-g\beta}\lvert\uparrow\rangle\langle\uparrow\rvert+(1-\eta)^2e^{g\beta}\lvert\downarrow\rangle\langle\downarrow\rvert\right]. \tag*{(132)} u + m + ρ i 1 m + † u + † = 4 ( 1 + η 2 ) cosh ( g β ) 1 [ ( 1 + η ) 2 e − g β ∣ ↑ ⟩ ⟨ ↑ ∣ + ( 1 − η ) 2 e g β ∣ ↓ ⟩ ⟨ ↓ ∣ ] . (132)
Therefore,
ρ f 1 : = u + m + u 1 ρ i 1 u 1 † m + † u + † + u − m − u 1 ρ i 1 u 1 † m − † u − † = 1 2 cosh ( g β ) [ e − g β ∣ ↑ ⟩ ⟨ ↑ ∣ + e g β ∣ ↓ ⟩ ⟨ ↓ ∣ ] (133) \rho_f^1 := u_+m_+u_1\rho_i^1u_1^\dagger m_+^\dagger u_+^\dagger+ u_-m_-u_1\rho_i^1u_1^\dagger m_-^\dagger u_-^\dagger= \frac{1}{2\cosh(g\beta)}\left[e^{-g\beta}|\uparrow\rangle\langle\uparrow|+e^{g\beta}|\downarrow\rangle\langle\downarrow|\right] \tag*{(133)} ρ f 1 := u + m + u 1 ρ i 1 u 1 † m + † u + † + u − m − u 1 ρ i 1 u 1 † m − † u − † = 2 cosh ( g β ) 1 [ e − g β ∣ ↑ ⟩ ⟨ ↑ ∣ + e g β ∣ ↓ ⟩ ⟨ ↓ ∣ ] (133)
and hence
Tr [ H f 1 ρ f 1 ] = − g tanh ( g β ) , (134) \operatorname{Tr}[H_f^1\rho_f^1]=-g\tanh(g\beta), \tag*{(134)} Tr [ H f 1 ρ f 1 ] = − g tanh ( g β ) , (134)
where
H f 1 : = g σ 3 . (135) H_f^1 := g\sigma^3. \tag*{(135)} H f 1 := g σ 3 . (135)
By performing the same calculation as in Case (i), we obtain the extractable work as
W o p t , l o c = 0. (136) W_{\mathrm{opt,loc}}=0. \tag*{(136)} W opt , loc = 0. (136)
Thus, no work can be extracted when the system temperature or the measurement accuracy is below a certain threshold.
Next, we compute the corresponding information quantities, namely, the mutual information, the Holevo information, and the nonequilibrium contribution. We also compute their sum, which is the right-hand side of Eq. (8) , and check that it indeed agrees with the extractable work.
First, we compute ∑ k P k S ( ρ k ′ ) \sum_k P_kS(\rho'_k) ∑ k P k S ( ρ k ′ ) , where
P k : = ∏ l = 1 V p k l (137) P_k := \prod_{l=1}^{V}p_{kl} \tag*{(137)} P k := l = 1 ∏ V p k l (137)
with
p + = Tr [ m + † m + ρ i 1 ] = ( 1 + η ) 2 e − g β + ( 1 − η ) 2 e g β 4 ( 1 + η 2 ) cosh ( g β ) , (138) p_+ = \operatorname{Tr}[m_+^\dagger m_+\rho_i^1] = \frac{(1+\eta)^2e^{-g\beta}+(1-\eta)^2e^{g\beta}}{4(1+\eta^2)\cosh(g\beta)}, \tag*{(138)} p + = Tr [ m + † m + ρ i 1 ] = 4 ( 1 + η 2 ) cosh ( g β ) ( 1 + η ) 2 e − g β + ( 1 − η ) 2 e g β , (138)
p − = 1 − p + = ( 1 − η ) 2 e − g β + ( 1 + η ) 2 e g β 4 ( 1 + η 2 ) cosh ( g β ) . (139) p_- = 1-p_+ = \frac{(1-\eta)^2e^{-g\beta}+(1+\eta)^2e^{g\beta}}{4(1+\eta^2)\cosh(g\beta)}. \tag*{(139)} p − = 1 − p + = 4 ( 1 + η 2 ) cosh ( g β ) ( 1 − η ) 2 e − g β + ( 1 + η ) 2 e g β . (139)
Also,
S ( ρ k ′ ) = ∑ l = 1 V S ( ρ k l 1 ) (140) S(\rho'_k)=\sum_{l=1}^{V}S(\rho_{kl}^1) \tag*{(140)} S ( ρ k ′ ) = l = 1 ∑ V S ( ρ k l 1 ) (140)
and
∑ k P k S ( ρ k ′ ) = V ( ∑ k 1 = ± p k 1 S ( ρ k 1 1 ) ) . (141) \sum_k P_kS(\rho'_k)=V\left(\sum_{k_1=\pm}p_{k_1}S(\rho_{k_1}^1)\right). \tag*{(141)} k ∑ P k S ( ρ k ′ ) = V ( k 1 = ± ∑ p k 1 S ( ρ k 1 1 ) ) . (141)
From Eq. (132) ,
ρ + 1 = u + m + u 1 ρ i 1 u 1 † m + † u + † p + = 1 4 p + ( 1 + η 2 ) cosh ( g β ) [ ( 1 + η ) 2 e − g β ∣ ↑ ⟩ ⟨ ↑ ∣ + ( 1 − η ) 2 e g β ∣ ↓ ⟩ ⟨ ↓ ∣ ] , (142) \rho_+^1=\frac{u_+m_+u_1\rho_i^1u_1^\dagger m_+^\dagger u_+^\dagger}{p_+}=\frac{1}{4p_+(1+\eta^2)\cosh(g\beta)}\left[(1+\eta)^2e^{-g\beta}|\uparrow\rangle\langle\uparrow|+(1-\eta)^2e^{g\beta}|\downarrow\rangle\langle\downarrow|\right], \tag*{(142)} ρ + 1 = p + u + m + u 1 ρ i 1 u 1 † m + † u + † = 4 p + ( 1 + η 2 ) cosh ( g β ) 1 [ ( 1 + η ) 2 e − g β ∣ ↑ ⟩ ⟨ ↑ ∣ + ( 1 − η ) 2 e g β ∣ ↓ ⟩ ⟨ ↓ ∣ ] , (142)
and from Eq. (102) ,
ρ − 1 = u − m − u 1 ρ i 1 u 1 † m − † u − † p − = 1 4 p − ( 1 + η 2 ) cosh ( g β ) [ ( 1 − η ) 2 e − g β ∣ ↑ ⟩ ⟨ ↑ ∣ + ( 1 + η ) 2 e g β ∣ ↓ ⟩ ⟨ ↓ ∣ ] . (143) \rho_-^1=\frac{u_-m_-u_1\rho_i^1u_1^\dagger m_-^\dagger u_-^\dagger}{p_-}=\frac{1}{4p_-(1+\eta^2)\cosh(g\beta)}\left[(1-\eta)^2e^{-g\beta}|\uparrow\rangle\langle\uparrow|+(1+\eta)^2e^{g\beta}|\downarrow\rangle\langle\downarrow|\right]. \tag*{(143)} ρ − 1 = p − u − m − u 1 ρ i 1 u 1 † m − † u − † = 4 p − ( 1 + η 2 ) cosh ( g β ) 1 [ ( 1 − η ) 2 e − g β ∣ ↑ ⟩ ⟨ ↑ ∣ + ( 1 + η ) 2 e g β ∣ ↓ ⟩ ⟨ ↓ ∣ ] . (143)
Hence,
∑ k P k S ( ρ k ′ ) = V ( ∑ k 1 = ± p k 1 S ( ρ k 1 1 ) ) = V ( − ( 1 + η ) 2 e − g β 4 ( 1 + η 2 ) cosh ( g β ) log [ ( 1 + η ) 2 e − g β ( 1 + η ) 2 e − g β + ( 1 − η ) 2 e g β ] − ( 1 − η ) 2 e − g β 4 ( 1 + η 2 ) cosh ( g β ) log [ ( 1 − η ) 2 e − g β ( 1 − η ) 2 e − g β + ( 1 + η ) 2 e g β ] \begin{aligned}
\sum_k P_k S(\rho_k')=V\left(\sum_{k_1=\pm}p_{k_1}S(\rho_{k_1}^{1})\right) \\
&=V\left(-\frac{(1+\eta)^2e^{-g\beta}}{4(1+\eta^2)\cosh(g\beta)}\log\left[\frac{(1+\eta)^2e^{-g\beta}}{(1+\eta)^2e^{-g\beta}+(1-\eta)^2e^{g\beta}}\right]-\frac{(1-\eta)^2e^{-g\beta}}{4(1+\eta^2)\cosh(g\beta)}\log\left[\frac{(1-\eta)^2e^{-g\beta}}{(1-\eta)^2e^{-g\beta}+(1+\eta)^2e^{g\beta}}\right]\right.
\end{aligned} k ∑ P k S ( ρ k ′ ) = V ( k 1 = ± ∑ p k 1 S ( ρ k 1 1 ) ) = V ( − 4 ( 1 + η 2 ) cosh ( g β ) ( 1 + η ) 2 e − g β log [ ( 1 + η ) 2 e − g β + ( 1 − η ) 2 e g β ( 1 + η ) 2 e − g β ] − 4 ( 1 + η 2 ) cosh ( g β ) ( 1 − η ) 2 e − g β log [ ( 1 − η ) 2 e − g β + ( 1 + η ) 2 e g β ( 1 − η ) 2 e − g β ]
− ( 1 − η ) 2 e g β 4 ( 1 + η 2 ) cosh ( g β ) log [ ( 1 − η ) 2 e g β ( 1 + η ) 2 e − g β + ( 1 − η ) 2 e g β ] − ( 1 + η ) 2 e g β 4 ( 1 + η 2 ) cosh ( g β ) log [ ( 1 + η ) 2 e g β ( 1 − η ) 2 e − g β + ( 1 + η ) 2 e g β ] ) . (144) \left.-\frac{(1-\eta)^2e^{g\beta}}{4(1+\eta^2)\cosh(g\beta)}\log\left[\frac{(1-\eta)^2e^{g\beta}}{(1+\eta)^2e^{-g\beta}+(1-\eta)^2e^{g\beta}}\right]-\frac{(1+\eta)^2e^{g\beta}}{4(1+\eta^2)\cosh(g\beta)}\log\left[\frac{(1+\eta)^2e^{g\beta}}{(1-\eta)^2e^{-g\beta}+(1+\eta)^2e^{g\beta}}\right]\right). \tag*{(144)} − 4 ( 1 + η 2 ) cosh ( g β ) ( 1 − η ) 2 e g β log [ ( 1 + η ) 2 e − g β + ( 1 − η ) 2 e g β ( 1 − η ) 2 e g β ] − 4 ( 1 + η 2 ) cosh ( g β ) ( 1 + η ) 2 e g β log [ ( 1 − η ) 2 e − g β + ( 1 + η ) 2 e g β ( 1 + η ) 2 e g β ] ) . (144)
Next, we compute S ( ρ i e n s ) S(\rho_i^{\mathrm{ens}}) S ( ρ i ens ) . Since the von Neumann entropy of the initial state is the same as in Case (i), we have
S ( ρ i e n s ) = V ( − e − g β 2 cosh ( g β ) log [ e − g β 2 cosh ( g β ) ] − e g β 2 cosh ( g β ) log [ e g β 2 cosh ( g β ) ] ) . (145) S(\rho_i^{\mathrm{ens}})=V\left(-\frac{e^{-g\beta}}{2\cosh(g\beta)}\log\left[\frac{e^{-g\beta}}{2\cosh(g\beta)}\right]-\frac{e^{g\beta}}{2\cosh(g\beta)}\log\left[\frac{e^{g\beta}}{2\cosh(g\beta)}\right]\right). \tag*{(145)} S ( ρ i ens ) = V ( − 2 cosh ( g β ) e − g β log [ 2 cosh ( g β ) e − g β ] − 2 cosh ( g β ) e g β log [ 2 cosh ( g β ) e g β ] ) . (145)
Next, using Eqs. (S108) and (S78), we obtain
S ( ρ f ) = V S [ 1 4 ( 1 + η 2 ) cosh ( g β ) ( 2 ( 1 + η 2 ) e − g β ∣ ↑ ⟩ ⟨ ↑ ∣ + 2 ( 1 + η 2 ) e g β ∣ ↓ ⟩ ⟨ ↓ ∣ ) ] (146) S(\rho_f)=VS\left[\frac{1}{4(1+\eta^2)\cosh(g\beta)}\left(2(1+\eta^2)e^{-g\beta}|\uparrow\rangle\langle\uparrow|+2(1+\eta^2)e^{g\beta}|\downarrow\rangle\langle\downarrow|\right)\right] \tag*{(146)} S ( ρ f ) = V S [ 4 ( 1 + η 2 ) cosh ( g β ) 1 ( 2 ( 1 + η 2 ) e − g β ∣ ↑ ⟩ ⟨ ↑ ∣ + 2 ( 1 + η 2 ) e g β ∣ ↓ ⟩ ⟨ ↓ ∣ ) ] (146)
= V [ − e − g β 2 cosh ( g β ) log ( e − g β 2 cosh ( g β ) ) − e g β 2 cosh ( g β ) log ( e g β 2 cosh ( g β ) ) ] , (147) =V\left[-\frac{e^{-g\beta}}{2\cosh(g\beta)}\log\left(\frac{e^{-g\beta}}{2\cosh(g\beta)}\right)-\frac{e^{g\beta}}{2\cosh(g\beta)}\log\left(\frac{e^{g\beta}}{2\cosh(g\beta)}\right)\right], \tag*{(147)} = V [ − 2 cosh ( g β ) e − g β log ( 2 cosh ( g β ) e − g β ) − 2 cosh ( g β ) e g β log ( 2 cosh ( g β ) e g β ) ] , (147)
which is equal to the von Neumann entropy S ( ρ i e n s ) S(\rho_i^{\mathrm{ens}}) S ( ρ i ens ) of the initial state:
S ( ρ f ) = S ( ρ i e n s ) . (148) S(\rho_f)=S(\rho_i^{\mathrm{ens}}). \tag*{(148)} S ( ρ f ) = S ( ρ i ens ) . (148)
From the above results, the mutual information I Q C I_{\mathrm{QC}} I QC is
I Q C = S ( ρ i e n s ) − ∑ k P k S ( ρ k ′ ) = V ( − e − g β 2 cosh ( g β ) log [ e − g β 2 cosh ( g β ) ] − e g β 2 cosh ( g β ) log [ e g β 2 cosh ( g β ) ] ) \begin{aligned}
I_{\mathrm{QC}}=S(\rho_i^{\mathrm{ens}})-\sum_kP_kS(\rho_k') \\
&=V\left(-\frac{e^{-g\beta}}{2\cosh(g\beta)}\log\left[\frac{e^{-g\beta}}{2\cosh(g\beta)}\right]-\frac{e^{g\beta}}{2\cosh(g\beta)}\log\left[\frac{e^{g\beta}}{2\cosh(g\beta)}\right]\right)
\end{aligned} I QC = S ( ρ i ens ) − k ∑ P k S ( ρ k ′ ) = V ( − 2 cosh ( g β ) e − g β log [ 2 cosh ( g β ) e − g β ] − 2 cosh ( g β ) e g β log [ 2 cosh ( g β ) e g β ] )
− V ( − ( 1 + η ) 2 e − g β 4 ( 1 + η 2 ) cosh ( g β ) log [ ( 1 + η ) 2 e − g β ( 1 + η ) 2 e − g β + ( 1 − η ) 2 e g β ] -V(-\frac{(1+\eta)^2e^{-g\beta}}{4(1+\eta^2)\cosh(g\beta)}\log[\frac{(1+\eta)^2e^{-g\beta}}{(1+\eta)^2e^{-g\beta}+(1-\eta)^2e^{g\beta}}] − V ( − 4 ( 1 + η 2 ) cosh ( g β ) ( 1 + η ) 2 e − g β log [ ( 1 + η ) 2 e − g β + ( 1 − η ) 2 e g β ( 1 + η ) 2 e − g β ]
− ( 1 − η ) 2 e − g β 4 ( 1 + η 2 ) cosh ( g β ) log [ ( 1 − η ) 2 e − g β ( 1 − η ) 2 e − g β + ( 1 + η ) 2 e g β ] -\frac{(1-\eta)^2e^{-g\beta}}{4(1+\eta^2)\cosh(g\beta)}\log\left[\frac{(1-\eta)^2e^{-g\beta}}{(1-\eta)^2e^{-g\beta}+(1+\eta)^2e^{g\beta}}\right] − 4 ( 1 + η 2 ) cosh ( g β ) ( 1 − η ) 2 e − g β log [ ( 1 − η ) 2 e − g β + ( 1 + η ) 2 e g β ( 1 − η ) 2 e − g β ]
− ( 1 − η ) 2 e g β 4 ( 1 + η 2 ) cosh ( g β ) log [ ( 1 − η ) 2 e g β ( 1 + η ) 2 e − g β + ( 1 − η ) 2 e g β ] -\frac{(1-\eta)^2e^{g\beta}}{4(1+\eta^2)\cosh(g\beta)}\log\left[\frac{(1-\eta)^2e^{g\beta}}{(1+\eta)^2e^{-g\beta}+(1-\eta)^2e^{g\beta}}\right] − 4 ( 1 + η 2 ) cosh ( g β ) ( 1 − η ) 2 e g β log [ ( 1 + η ) 2 e − g β + ( 1 − η ) 2 e g β ( 1 − η ) 2 e g β ]
− ( 1 + η ) 2 e g β 4 ( 1 + η 2 ) cosh ( g β ) log [ ( 1 + η ) 2 e g β ( 1 − η ) 2 e − g β + ( 1 + η ) 2 e g β ] ) , (149) -\frac{(1+\eta)^2e^{g\beta}}{4(1+\eta^2)\cosh(g\beta)}\log[\frac{(1+\eta)^2e^{g\beta}}{(1-\eta)^2e^{-g\beta}+(1+\eta)^2e^{g\beta}}]), \tag*{(149)} − 4 ( 1 + η 2 ) cosh ( g β ) ( 1 + η ) 2 e g β log [ ( 1 − η ) 2 e − g β + ( 1 + η ) 2 e g β ( 1 + η ) 2 e g β ]) , (149)
which is the same as in Case (i).
Next, we compute the Holevo information.
χ = S ( ρ f ) − ∑ k P k S ( ρ k ′ ) = V [ − e − g β 2 cosh ( g β ) log ( e − g β 2 cosh ( g β ) ) − e g β 2 cosh ( g β ) log ( e g β 2 cosh ( g β ) ) ] \begin{aligned}
\chi=S(\rho_f)-\sum_kP_kS(\rho_k') \\
&=V\left[-\frac{e^{-g\beta}}{2\cosh(g\beta)}\log\left(\frac{e^{-g\beta}}{2\cosh(g\beta)}\right)-\frac{e^{g\beta}}{2\cosh(g\beta)}\log\left(\frac{e^{g\beta}}{2\cosh(g\beta)}\right)\right]
\end{aligned} χ = S ( ρ f ) − k ∑ P k S ( ρ k ′ ) = V [ − 2 cosh ( g β ) e − g β log ( 2 cosh ( g β ) e − g β ) − 2 cosh ( g β ) e g β log ( 2 cosh ( g β ) e g β ) ]
− V ( − ( 1 + η ) 2 e − g β 4 ( 1 + η 2 ) cosh ( g β ) log [ ( 1 + η ) 2 e − g β ( 1 + η ) 2 e − g β + ( 1 − η ) 2 e g β ] -V(-\frac{(1+\eta)^2e^{-g\beta}}{4(1+\eta^2)\cosh(g\beta)}\log[\frac{(1+\eta)^2e^{-g\beta}}{(1+\eta)^2e^{-g\beta}+(1-\eta)^2e^{g\beta}}] − V ( − 4 ( 1 + η 2 ) cosh ( g β ) ( 1 + η ) 2 e − g β log [ ( 1 + η ) 2 e − g β + ( 1 − η ) 2 e g β ( 1 + η ) 2 e − g β ]
− ( 1 − η ) 2 e − g β 4 ( 1 + η 2 ) cosh ( g β ) log [ ( 1 − η ) 2 e − g β ( 1 − η ) 2 e − g β + ( 1 + η ) 2 e g β ] -\frac{(1-\eta)^2e^{-g\beta}}{4(1+\eta^2)\cosh(g\beta)}\log\left[\frac{(1-\eta)^2e^{-g\beta}}{(1-\eta)^2e^{-g\beta}+(1+\eta)^2e^{g\beta}}\right] − 4 ( 1 + η 2 ) cosh ( g β ) ( 1 − η ) 2 e − g β log [ ( 1 − η ) 2 e − g β + ( 1 + η ) 2 e g β ( 1 − η ) 2 e − g β ]
− ( 1 − η ) 2 e g β 4 ( 1 + η 2 ) cosh ( g β ) log [ ( 1 − η ) 2 e g β ( 1 + η ) 2 e − g β + ( 1 − η ) 2 e g β ] -\frac{(1-\eta)^2e^{g\beta}}{4(1+\eta^2)\cosh(g\beta)}\log\left[\frac{(1-\eta)^2e^{g\beta}}{(1+\eta)^2e^{-g\beta}+(1-\eta)^2e^{g\beta}}\right] − 4 ( 1 + η 2 ) cosh ( g β ) ( 1 − η ) 2 e g β log [ ( 1 + η ) 2 e − g β + ( 1 − η ) 2 e g β ( 1 − η ) 2 e g β ]
− ( 1 + η ) 2 e g β 4 ( 1 + η 2 ) cosh ( g β ) log [ ( 1 + η ) 2 e g β ( 1 − η ) 2 e − g β + ( 1 + η ) 2 e g β ] ) . (150) -\frac{(1+\eta)^2e^{g\beta}}{4(1+\eta^2)\cosh(g\beta)}\log[\frac{(1+\eta)^2e^{g\beta}}{(1-\eta)^2e^{-g\beta}+(1+\eta)^2e^{g\beta}}]). \tag*{(150)} − 4 ( 1 + η 2 ) cosh ( g β ) ( 1 + η ) 2 e g β log [ ( 1 − η ) 2 e − g β + ( 1 + η ) 2 e g β ( 1 + η ) 2 e g β ]) . (150)
Next, we compute the relative entropy. Since ρ f e n s = ρ i e n s \rho_f^{\mathrm{ens}}=\rho_i^{\mathrm{ens}} ρ f ens = ρ i ens in this case, D ( ρ f e n s ∥ ρ i e n s ) = 0 D(\rho_f^{\mathrm{ens}}\Vert\rho_i^{\mathrm{ens}})=0 D ( ρ f ens ∥ ρ i ens ) = 0 :
D ( ρ f e n s ∥ ρ i e n s ) = − S ( ρ f ) + β Tr [ H f ρ f ] + log Z = − V [ e − g β 2 cosh ( g β ) log ( e − g β 2 cosh ( g β ) ) − e g β 2 cosh ( g β ) log ( e g β 2 cosh ( g β ) ) ] − β g V tanh ( g β ) + V log ( 2 cosh ( g β ) ) = 0. (151) \begin{aligned}
D(\rho_f^{\mathrm{ens}}\Vert\rho_i^{\mathrm{ens}})=-S(\rho_f)+\beta\operatorname{Tr}[H_f\rho_f]+\log Z \\
&=-V\left[\frac{e^{-g\beta}}{2\cosh(g\beta)}\log\left(\frac{e^{-g\beta}}{2\cosh(g\beta)}\right)-\frac{e^{g\beta}}{2\cosh(g\beta)}\log\left(\frac{e^{g\beta}}{2\cosh(g\beta)}\right)\right]-\beta gV\tanh(g\beta)+V\log(2\cosh(g\beta)) \\
&=0. \tag*{(151)}
\end{aligned} D ( ρ f ens ∥ ρ i ens ) = − S ( ρ f ) + β Tr [ H f ρ f ] + log Z = − V [ 2 cosh ( g β ) e − g β log ( 2 cosh ( g β ) e − g β ) − 2 cosh ( g β ) e g β log ( 2 cosh ( g β ) e g β ) ] − β g V tanh ( g β ) + V log ( 2 cosh ( g β )) = 0. (151)
Finally, we compute the sum of the information quantities, i.e., the right-hand side of Eq. (8) , and check that it indeed agrees with the extractable work.
I Q C − χ − D ( ρ f e n s ∥ ρ i e n s ) = V ( e − g β 2 cosh ( g β ) log [ e − g β 2 cosh ( g β ) ] − e g β 2 cosh ( g β ) log [ e g β 2 cosh ( g β ) ] ) + β g V tanh ( g β ) − V log ( 2 cosh ( g β ) ) = 0 , (152) \begin{aligned}
I_{\mathrm{QC}}-\chi-D(\rho_f^{\mathrm{ens}}\Vert\rho_i^{\mathrm{ens}}) \\
&=V\left(\frac{e^{-g\beta}}{2\cosh(g\beta)}\log\left[\frac{e^{-g\beta}}{2\cosh(g\beta)}\right]-\frac{e^{g\beta}}{2\cosh(g\beta)}\log\left[\frac{e^{g\beta}}{2\cosh(g\beta)}\right]\right)+\beta gV\tanh(g\beta)-V\log(2\cosh(g\beta)) \\
&=0, \tag*{(152)}
\end{aligned} I QC − χ − D ( ρ f ens ∥ ρ i ens ) = V ( 2 cosh ( g β ) e − g β log [ 2 cosh ( g β ) e − g β ] − 2 cosh ( g β ) e g β log [ 2 cosh ( g β ) e g β ] ) + β g V tanh ( g β ) − V log ( 2 cosh ( g β )) = 0 , (152)
which agrees with the extractable work in Eq. (136) .