Thermodynamics as a Consequence of Information Conservation

Manabendra Nath Bera,1, ∗ Arnau Riera,1, 2 Maciej Lewenstein,1, 3 and Andreas Winter3, 4
1
ICFO – Institut de Ciencies Fotoniques, The Barcelona Institute of Science and Technology, ES-08860 Castelldefels, Spain
2
Max-Planck-Institut für Quantenoptik, D-85748 Garching, Germany
3
ICREA, Pg. Lluis Companys 23, ES-08010 Barcelona, Spain
4
Departament de Física: Grup d’Informació Quàntica,
Universitat Autònoma de Barcelona, ES-08193 Bellaterra (Barcelona), Spain
Thermodynamics and information have intricate inter-relations. Often thermodynamics is considered to be
the logical premise to justify that information is physical – through Landauer’s principle – thereby links infor-
mation and thermodynamics. This approach towards information, recently, has been instrumental to understand
thermodynamics of logical and physical processes, both in classical and quantum domains. In this work we
formulate thermodynamics as an exclusive consequence of information conservation. The framework can be
applied to most general situations, beyond the traditional assumptions in thermodynamics, where systems and
thermal-baths could be quantum, of arbitrary sizes and even could posses inter-system correlations. Further, it
arXiv:1707.01750v1 [quant-ph] 6 Jul 2017

does not require a priory predetermined temperature associated to a thermal-bath, which does not carry much
sense for finite-size cases. Importantly, the thermal-baths and systems are not treated here differently, rather both
are considered on equal footing. This leads us to introduce a “temperature”-independent formulation of thermo-
dynamics. We rely on the fact that, for a given amount of information, measured by the von Neumann entropy,
any system can be transformed to a state that possesses minimal energy. This state is known as a completely
passive state that acquires a Boltzmann–Gibb’s canonical form with an intrinsic temperature. We introduce the
notions of bound and free energy and use them to quantify heat and work respectively. We explicitly use the
information conservation as the fundamental principle of nature, and develop universal notions of equilibrium,
heat and work, universal fundamental laws of thermodynamics, and Landauer’s principle that connects ther-
modynamics and information. We demonstrate that the maximum efficiency of a quantum engine with a finite
bath is in general different and smaller than that of an ideal Carnot’s engine. We introduce a resource theoretic
framework for our intrinsic-temperature based thermodynamics, within which we address the problem of work
extraction and inter-state transformations. We also extend the framework to the cases of multiple conserved
quantities.

I. INTRODUCTION formulate quantum thermodynamics within the, so called, re-

Thermodynamics constitutes one of the most basic founda-
tions of modern science. It not only plays an important role
in modern technologies, but offers also basic understanding
of the vast range of natural phenomena. Initially, thermody-
namics was developed in an phenomenological way to address
the question on how, and to what extent, heat could be con-
verted into work. But, in the later course with the develop-
ments of statistical mechanics, quantum mechanics and rela-
tivity, thermodynamics along with its fundamental laws has
attained quite a formal and mathematically rigorous form [1].
It finds applications in ultra-large large systems, where it de-
scribes relativistic phenomena in astrophysics and cosmology,
in microscopic systems, where it describes quantum effects, or
in very complex systems in biology and chemistry.
The inter-relation between information and thermodynam-
ics [2] is very intricate, and has been studied in the context of
Maxwell’s demon [3–6], Szilard’s engine [7], and Landauer’s
principle [8–12]. In the recent years classical and quantum
information theoretic [13–15] approaches help us to under-
stand thermodynamics in the domain of small classical and
quantum systems. That stimulated a whole new approach to
tackle and extend thermodynamics beyond the standard clas-
sical domain. In particular, this new approach has helped to
∗ mnbera@gmail.com
source theoretic framework [16], which allows to reproduce
standard thermodynamics in the asymptotic limit, when one
processes infinitely many copies of the system under consid-
eration. In the finite copy limit, commonly known as one-shot
limit, the resource theory reveals that the laws of thermody-
namics have to be modified to dictate the transformations on
the quantum level [17–25]. In addition, information theory
also plays an important role in understanding thermodynamics
in the presence of inter-system and system-bath correlations
[26–28], equilibration processes [29–32], and foundations of
statistical mechanics [33].
One interesting fact to note is that in standard, both clas-
sical and quantum, thermodynamics the thermal bath is con-
sidered to be a system that is considerably large in size. The
bath’s state assumes the Boltzmann-Gibb’s canonical form at
a given temperature. One assumes here that after exchange of
heat, the bath is not supposed to change its temperature (see
[34] for a review on the notion of temperature). However,
as we consider the system to be small and quantum, a priori,
there is no apparent reason to consider the bath to be large!
It could be small as well, and it could even possess quantum
properties, like coherence or correlations. In this regime it is
clear that exchange of heat will certainly alter bath’s initial
temperature. Therefore, for finite baths, one needs to refor-
mulate a temperature-independent, or even bath-independent
version of quantum thermodynamics. That is the goal of the
present work – formulation of thermodynamics as an exclu-
sive consequence of information conservation.

All fundamental physical theories, such as classical and
quantum mechanics, share the property of conserving infor-
mation. That is, they have dynamics, which are deterministic
and they bijectively map the set of possible configurations be-
tween any two instants of time. Thus, non-determinism can
only appear, when some degrees of freedom are ignored, lead-
ing to apparent information loss. In classical physics this in-
formation loss is due to deterministic chaos and mixing in
nonlinear dynamics. In quantum mechanics this loss is in-
trinsic, and occurs due to measurement processes and non-
local correlations [35]. In information theory, the quantity
that measures the amount of information hidden in a system
is the entropy. In the view of the common feature of informa-
tion conservation for all microscopic underlying dynamics, it
is natural to consider thermodynamics, restricting to the op-
erations that are entropy, or information preserving. In this
paper we show that the laws of thermodynamics are a logi-
cal consequence of the information conservation, or in other
words, any microscopic theory which conserves information
will fulfill the laws of thermodynamics.
II. ENTROPY PRESERVING OPERATIONS, ENTROPIC
EQUIVALENCE CLASS AND INTRINSIC TEMPERATURE
The set of operations that we consider in this framework
is the set of, so called, entropy preserving (EP) operations.
Given a system initially in a state ρ, the set of entropy preserv-
ing operations are all the operations that arbitrarily change the
state, but keep its entropy constant
ρ→σ : S (ρ) = S (σ) , (1)
where S (ρ) B −Tr (ρ log ρ) is the von Neumann entropy. Im-
portantly, an operation that acts on ρ and produces a state with
the same entropy, not necessarily preserves entropy when act-
ing on other states. In fact, such entropy preserving operations
are in general not linear, since they have to be constraint to
some input state. It was shown in Ref. [36] that a quantum
channel Λ(·) that preserves entropy and, at the same time, re-
spects linearity, i. e. Λ(pρ1 +(1− p)ρ2 ) = pΛ(ρ1 )+(1− p)Λ(ρ2 ),
has to be a unitary.
However, the entropy preserving operations can be micro-
scopically described by global unitaries in the limit of many
copies [37]. Given any two states ρ and σ with equal p entropies
S (ρ) = S (σ), there exists an additional system of O( n log n)
ancillary qubits and a global unitary U such that
 
lim kTr anc Uρ⊗n ⊗ ηU † − σ⊗n k1 = 0 , (2)
n→∞
where the partial trace is performed on the ancillary qubits.
Here k · k1 is the one-norm. As shown in Theorem 4 of
Ref. [37], the reverse statement is also true. In other words, if
two states are related as in Eq. (2) then they also have equal
entropies.
As we see later it is important to restrict entropy preserv-
ing operations that are also be energy preserving. The energy
and entropy preserving operations can also be implemented
using a global energy preserving unitary in the many copy
2
limit. More explicitly, in Theorem 1 of Ref. [37], it is shown
that two states ρ and σ with equal entropies and energies,
i.e. (S (ρ) = S (σ) and E(ρ) = E(σ)), is equivalent to the
existence of some
p energy preserving U and an additional sys-
tem A with O( n log n) of ancillary qubits with Hamiltonian
kHA k 6 O(n2/3 ) in some state η for which (2) is fulfilled. Note
that the amount of energy and entropy of the ancillary system
per copy vanishes in the large n limit.
Once the the entropy preserving operations have been mo-
tivated and introduced, let us classify the set of states of a
system in different equivalence classes depending on their en-
tropy. Thereby, we establish a hierarchy of states according to
their information content. More formally,
Definition 1 (Entropic equivalence class). Two states ρ and
σ on any quantum system of dimension d are equivalent and
belong to the same entropic equivalence class if and only if
both have the same Von Neumann entropy,
ρ∼σ iff S (ρ) = S (σ) . (3)
Assuming that the system has some fixed Hamiltonian H,
one can take as a representative element of every class the
state that minimizes the energy within it, i.e.,
γ(ρ) B arg min E(σ), (4)
σ : S (σ)=S (ρ)
where E(σ) B Tr (Hσ) is the energy of the state σ.
The maximum-entropy principle [38, 39] identifies the ther-
mal state as the state that maximizes the entropy for a given
energy. Conversely, one can show that, for a given entropy,
the thermal state also minimizes the energy. We refer to this
complementary property as min-energy principle [26, 40, 41].
Thus, the min-energy principle identifies thermal states as the
representative elements of every class, which is
e−β(ρ)H
γ(ρ) =
. (5)
Tr e−β(ρ)H
The inverse temperature β(ρ) is the parameter that labels the
equivalence class, to which the state ρ belongs. We denote
β(ρ) as the intrinsic inverse temperature associated to ρ.
The state γ(ρ) is, often, termed as the completely passive
(CP) state [26] with the minimum internal energy, for the same
information content. The CP state, with the form γ(HS , βS ),
has the following interesting properties [40, 41]:
(P1) For a given entropy, it minimizes the energy.
(P2) Both energy and entropy monotonically increase (de-
crease) with the decrease (increase) in βS , and vice
versa.
(P3) For non-interacting Hamiltonians, HT = NX=1 I⊗X−1 ⊗
P
⊗N−X
HX ⊗ I , the joint complete passive state is ten-
sor product of individual ones, i.e, γ(HT , βT ) =
⊗NX=1 γ(HX , βT ), for identical βT [40, 41].

III. BOUND AND FREE ENERGIES
Now with the notion of entropic classes and CP states, let us
identify two relevant forms of internal energy: the free and the
bound energies. The bound energy is defined as the amount of
internal energy that cannot be accessed in the form of work. It
is quantified in the following.
Definition 2 (Bound energy). For a state ρ with the system
Hamiltonian H, the bound energy in it is
B(ρ) B min E(σ) = E(γ(ρ)), (6)
σ : S (σ)=S (ρ)
where γ(ρ) is the CP state, with minimum energy, within the
equivalence class, to which ρ belongs.
Indeed, B(ρ) is the amount of energy that cannot be ex-
tracted further, by exploiting any entropy preserving opera-
tions, as guaranteed by the min-energy principle. The above
definition of bound energy also has strong connection with in-
formation content in the state. It can be easily seen that one
could have access to this energy (in the form of work) only, if
it allows an outflow of information from the system and vice
versa.
In contrast to bound energy, free energy is the part of the
internal energy that can be accessed with entropy preserving
operations:
Definition 3 (Free energy). For a system ρ with system Hamil-
tonian HS , the free energy stored in the system is given by
F(ρ) B E(ρ) − B(ρ), (7)
where B(ρ)) is bound energy in ρ.
Note that the free energy as defined in Eq. (7) does not have
a preferred temperature, unlike the standard out of equilib-
rium Helmholtz free energy FT (ρ) B E(ρ) − T S (ρ) where the
temperature T is decided beforehand with choice of a ther-
mal bath. Nevertheless, our definition of free energy can be
written in terms of both the relative entropy and the out of
equilibrium free energy as
F(ρ) = T (ρ)D(ρ||γ(ρ)) = FT (ρ) (ρ) − FT (ρ) (γ(ρ)) , (8)
for the intrinsic temperature T (ρ) B β(ρ)−1 that labels the
equivalence class that contains ρ. Here the
relative entropy is
defined as D(ρ||σ) = Tr ρ log ρ − ρ log σ .
The notions of bound and free energy, as we consider
above, can be extended beyond single systems. If fact, in
multi-particle (multipartite) systems they exhibit several inter-
esting properties. For example, they can capture the presence
and/or absence of inter-party correlations. To highlight these
features, we consider a bipartite system below.
Lemma 4 (Bound and free energy properties). Given a bipar-
tite system with non-interacting Hamiltonian HA ⊗ I + I ⊗ HB
in an arbitrary state ρAB and a product state ρA ⊗ ρB with
marginals ρA/B B Tr B/A (ρAB ). Then, the bound and the free
energy fulfill the following properties:
3
(P4) Bound energy and correlations:
B(ρAB ) ≤ B(ρA ⊗ ρB ) . (9)
(P5) Bound energy of composite systems:
B(ρA ⊗ ρB ) ≤ B(ρA ) + B(ρB ) . (10)
(P6) Free energy and correlations:
F(ρA ⊗ ρB ) ≤ F(ρAB ) . (11)
(P7) Free energy of composite systems:
F(ρA ) + F(ρB ) ≤ F(ρA ⊗ ρB ) . (12)
In Eqs. (9) and (11) the equality is saturated if and only if A
and B are uncorrelated ρAB = ρA ⊗ ρB . Equations (10) and
(12) become an equality if and only if β(ρA ) = β(ρB ).
Proof. To prove (P4), we have used the fact that for a fixed
Hamiltonian the bound energy is monotonically increasing
with entropy, i.e.
B(ρ) < B(σ) iff S (ρ) < S (σ) , (13)
together with the subaddivity property S (ρAB ) ≤ S (ρA ⊗ ρB ).
The proof of (P5) relies on the definition of the bound en-
ergy, where one exploits entropy preserving (EP) operations
to minimize energy. For composite systems, the set of joint
EP operations when applied on product states ρA ⊗ ρB are
strictly
n ep o larger than the
o independent local EP operations, i.e.
n
ΛAB ! Λep A ⊗ Λ ep
B . This immediately implies Eq. (10). To
see that bound in Eq. (10) is saturated iff the states pose iden-
tical internal temperature, i.e., β(ρA ) = β(ρB ). For details, we
refer the reader to Definition 6 and Lemma 7.
The other properties can be easily proven by noting that the
total internal energy is sum of individual ones and E(ρAB ) =
E(ρA ) + E(ρB ), irrespective of inter-system correlations. Then
with the definition of free energy is F(ρX ) = E(ρX ) − B(ρX ),
(P4) and (P5), we could easily arrive at properties (P6) and
(P7). 
The above properties allow us to give an additional opera-
tional meaning to the free energy F(ρ). Let us remind first that
the work that can be extracted from a system in a state ρ when
having at our disposal and infinitely large bath at inverse tem-
perature β and global entropy preserving operations is given
by
W = Fβ (ρ) − Fβ (γ(β)) , (14)
where Fβ (ρ) = E(ρ) − β S (ρ) is the standard free energy
−1
with respect to the inverse temperature β and γ(β) is the ther-
mal state at which the system is left once the work has been
extracted.
Lemma 5 (Free energy vs. β-free energy). The free energy
F(ρ) corresponds to work extracted by using a bath at the
worst possible temperature,
 
F(ρ) = min Fβ (ρ) − Fβ (γ(β)) . (15)
β
The inverse temperature achieves the minimum is the inverse
intrinsic temperature β(ρ).

Proof. The proof is a corollary of Lemma 14 together with
(P7).  linear bound O(n2/3 ).
IV. ENERGY-ENTROPY DIAGRAM
Let us describe here the energy-entropy diagram. Given
a system described by a Hamiltonian H, a state ρ is repre-
sented in the energy-entropy diagram by a point with coor-
dinates xρ B (E(ρ), S (ρ)), as shown in Fig. 1. All physical
states reside in a region that is lower bounded by the hori-
zontal axis (i.e., S = 0) corresponding to the pure states, and
upper bounded by the convex curve (E(β), S (β)) which repre-
sents the thermal states of both positive and negative temper-
atures. Let us denote such a curve as the thermal boundary.
The inverse temperature associated to one point of the ther-
mal boundary is given by the slope of the tangent line in such
a point, since
dS (β)
= β. (16)
dE(β)
S energy.
β=0
β(ρ)
S(ρ)
B(ρ) F (ρ) xρ := (E(ρ), S(ρ))
β → −∞
E
Emin E(ρ) Emax
Figure 1. Energy-entropy diagram. Any quantum state ρ is repre-
sented in the diagram as a point with coordinates xρ B (E(ρ), S (ρ)).
The free energy F(ρ) is the distance in the horizontal direction from
the thermal boundary. The bound energy B(ρ) is the distance in the
horizontal direction between the thermal boundary and the energy
reference.
In Fig. 1, the free energy and the bound energy are plotted
given a state ρ. Its free energy F(ρ) can be seen from the di-
agram as the horizontal distance from the thermal boundary.
This is the part of internal energy which can be extracted with-
out altering system’s entropy. The slope of the tangent line of
the thermal boundary in that point is the intrinsic temperature,
β(ρ), of the state ρ. The bound energy B(ρ) is the distance in
the horizontal direction between the thermal boundary and the
energy reference and it can no way be extracted with entropy
preserving operations.
Note that in general a point in the energy-entropy diagram
represents multiple states, since different quantum states can
have identical entropy and energy. In any case, all these
states with same entropy and energy are thermodynamically
equivalent and, as it was shown in [37], they can be unitar-
ily transformed into each other in the limit of many copies
n → ∞ with an ancilla at our disposal which is of sublinear
4
p
size O( n log n) with a Hamiltonian upper bounded by sub-
V. EQUILIBRIUM AND ZEROTH LAW
The zeroth law of the thermodynamics establishes the ab-
solute scaling of temperature, in terms of thermal equilibrium.
It says that if a system A is in thermal equilibrium with B and
again, B is in thermal equilibrium with C, then A will also
be in thermal equilibrium with C. All the systems that are
in mutual thermal equilibrium can be classified to thermody-
namically equivalent class and each state in the class can be
assigned with a unique parameter called temperature. In other
words, at thermal equilibrium, temperatures of the individual
systems will exactly be equal to each other and also to their
arbitrary collections, as there would be no spontaneous heat or
energy flow in between. Further, when an arbitrary state, out
of thermal equilibrium, is brought in contact to a large ther-
mal bath, then the system tends to acquire an thermally equi-
librium state with the temperature of the bath. In this equili-
bration process, the system could exchange both energy and
entropy with bath and thereby minimizes its Helmholtz free
However, in the new setup, where a system cannot have
access to considerably large thermal bath or in absence of a
thermal bath, the equilibration process is expected to be con-
siderably different than that of the cases of large baths. In the
following, we introduce a formal definition of equilibration,
based on information preservation and intrinsic temperature.
Definition 6 (Equilibrium and zeroth law). Given a collec-
tion of systems A1 , . . . , An with non-interacting Hamiltonians
H1 , . . . Hn in a joint state, ρA1 ...An , we call them to be mutually
at equilibrium if and only if they “jointly” minimize the free
energy as defined in (7), i.e.,
F(ρA1 ...An ) = 0 . (17)
Let us consider the states ρA and ρB , with corresponding
Hamiltonians HA and HB respectively. The equilibrium is
achieved when they jointly attain an iso-informatic state with
minimum energy. Then the corresponding equilibrium state
is a completely passive (CP) state γ(HAB , βAB ) with the joint
Hamiltonian HAB = HA ⊗ I + I ⊗ HB . Further the joint CP state,
following (P3), is
γ(HAB , βAB ) = γ(HA , βAB ) ⊗ γ(HB , βAB ), (18)
where the local systems also assume the completely passive
states but with same βAB . Here we recall that γ(HX , βY ) =
e−βY HX /Tr (e−βY HX ). It also says that if two arbitrary CP states
γ(HA , βA ) and γ(HB , βB ) have βA , βB , then their combined
equilibrium state γ(HAB , βAB ) can, still, jointly reduce bound-
energy without altering the total information content. There-
fore joint CP state acquires a unique βAB , i.e.,
E(γ(HA , βA )) + E(γ(HB , βB )) > E(γ(HAB , βAB )), (19)
which is immediately followed from min-energy principle and
(P5). Moreover if βA ≥ βB , then Eq. (19) dictates a bound on
βAB , as expressed in Lemma 7:

Lemma 7. Any iso-informatic equilibration process between
γ(HA , βA ) and γ(HB , βB ), with βA ≥ βB , leads to a mutually
equilibrium state γ(HAB , βAB ), irrespective of non-interacting
systems’ Hamiltonians, where βAB satisfies
βA ≥ βAB ≥ βB . (20)
Proof. Note for βA = βB , (P3) and (P5) immediately lead to
βA = βAB = βB . What we need to show is that, for initial
βA > βB , the equilibration leads to βA > βAB > βB . It can
be seen from information preservation condition on the equi-
libration process. Say after the equilibration the final temper-
ature is βAB = βA > βB . That implies the state γ((HA , βA ))
remains unchanged, while γ(HB , βB ) changed to a state with
higher entropy as a consequence of (P2). Therefore, it de-
mands S (γ(HA , βA ) ⊗ γ(HB , βB )) < S (γ(HAB , βAB )), which is
forbidden by the information preservation condition. Same is
also true for the case with βAB > βA ≥ βB . Now consider the
other extreme where one has βA > βB = βAB . Then the (P2)
leads to S (γ(HA , βA ) ⊗ γ(HB , βB )) > S (γ(HAB , βAB )), which is
also true for βA > βB > βAB . In both the cases the informa-
tion preservation condition is not respected. Therefore only
possibility would be βA > βAB > βB .  Note that the spontaneous equilibration temperature β̃(ρ)
Now with the clear notion of equilibration and equilibrium
state which has minimum internal energy for a fixed informa-
tion content, we could restate zeroth law in terms of internal
temperature. By definition, the global CP state has minimum
internal energy and one cannot extract free energy by using
global EP operations. Further a global CP state not only as-
sures that the individual states are also CP but also confirms
that they share the same intrinsic temperature, i.e., β and van-
ishing inter-system correlations. Therefore we may argue that
the individual systems are in mutual equilibrium, but deter-
mined by an intrinsic temperature.
We may recover the traditional notion of thermal equilib-
rium, as well as, zeroth law. In traditional thermodynamics,
the thermal bath is considerably large (compared to the sys-
tems under consideration) and assumes CP state with a prede-
fined temperature. Mathematically a thermal bath, with bath
Hamiltonian HB , can be expressed as γB = e−βB HB /Tr (e−βB HB ),
where |γB | → ∞. Now if a finite state ρS (with |ρS |  ∞), with
Hamiltonian HS is brought in contact with thermal bath, then
the global thermal equilibrium state will be a CP state, i.e.,
Λep
γB ⊗ρS −−→ γ0B ⊗γS ,
with a global inverse equilibrium temper-
ature, say βe . With |γB |  |ρS | and |γB | → ∞, one could easily
see that γ0B → γB , βe → βB and γS → e−βB HS /Tr (e−βB HS ).
VI. MAX-ENTROPY PRINCIPLE VS. MIN-ENERGY
PRINCIPLE
Let us mention that our framework based on information
conservation is suited for the work extraction. It assumes that
the system has mechanisms to release energy to some bat-
tery or classical field, but it cannot interchange entropy with
it. This contrasts with other situations in spontaneous equi-
libration in which the system is assumed to evolve keeping
5
constant the conserved quantities. In such scenario, the equi-
librium state is described by the principle of maximum en-
tropy, that is, the state the maximizes the entropy given the
conserved quantities.
In the context of the maximum entropy principle, let us in-
troduce the absolute athermality as:
Definition 8 (Athermality). The athermality of a system in a
state ρ and Hamiltonian H is the amount of entropy (informa-
tion) that the system can still accommodate without increasing
its energy, i. e.
A(ρ) B max S (σ) − S (ρ), (21)
σ : E(σ)=E(ρ)
where the maximization is made over all states σ with energy
E(ρ).
The state σ that maximizes (21) is obviously a thermal
state. In particular, it is the equilibrium state of an equilibra-
tion process guided by the maximum entropy principle. Let us
call the temperature of such thermal state as the spontaneous
equilibration temperature and denote it by β̃(ρ). The ather-
mality corresponds to the amount of entropy produced during
such equilibration.
always differs from the intrinsic temperature β(ρ) unless ρ
is thermal. In other words, the maximum entropy and min-
imum energy principles lead to different equilibrium tempera-
tures. As both entropy and energy are monotonically increas-
ing functions of temperature, the equilibrium temperature de-
termined by the minimum energy principle is always smaller
than the one determined by the maximum entropy principle.
This is represented in the energy entropy diagram of Fig. 2,
where the athermality can be seen as the vertical distance from
a state ρ to the thermal boundary. The point in which the ther-
mal boundary intersects with the vertical line E = E(ρ) repre-
sents the equilibrium thermal state given by the max-entropy
principle. Its temperature β̃(ρ) corresponds to the slope of the
tangent line of the thermal boundary at that point.
S
β̃(ρ)
β(ρ) > β̃(ρ)
A(ρ)
β(ρ)
S(ρ)
F (ρ) xρ
E
E(ρ)
Figure 2. Representation of the free energy F(ρ) and the athermality
A(ρ) of a state ρ in the energy-entropy diagram. The intrinsic temper-
ature β(ρ) and the spontaneous equilibration temperature β̃(ρ) fulfil
β(ρ) > β̃(ρ) due to the concavity of the thermal boundary.
We have chosen the name “athermality” for the quantity
defined in 8 in agreement with Ref. [37]. There, a quantity
called β-athermality and defined by
Aβ (ρ) B βE(ρ) − S (ρ) + log Zβ (22)

with Zβ = Tr (e−βH ) the partition function, is introduced to
characterise the energy-entropy diagram as the set of all points
with positive entropy S (ρ) ≥ 0 and positive β-athermality
Aβ (ρ) ≥ 0 for all β ∈ R.
Lemma 9 (Athermality vs. β-athermality). The athermality
of a state ρ can be written in terms of the β-athermalities as
A(ρ) = min Aβ (ρ) , (23)
β
where the minimum is attained by the spontaneous equilibra-
tion temperature β̃(ρ).
Proof. We prove the lemma by giving a geometric interpreta-
tion to the β-athermality in the energy-entropy diagram. In the
energy entropy diagram represented in Fig. 3, let us consider
the tangent line to the thermal boundary with slope β. Such
straight line is formed the set of all points which fulfil
S − S β = β(E − Eβ ) , (24)
where S β and Eβ are respectively the entropy and the energy
of the the point in the thermal boundary that belongs to the
line. The vertical distance (difference in entropy) of such line
from a point with coordinates (E(ρ), S (ρ)) reads
S ∗ − S (ρ) = β(E(ρ) − Eβ ) − S (ρ)
(25)
= βE(ρ) − S (ρ) − β(Eβ − T S β )
where S ∗ is the entropy of the line at E = E(ρ) and T = β−1 .
By noticing that log Zβ = β(Eβ − T S β ), we get that the β-
athermality of ρ, A(ρ), is the vertical distance of xρ from the
line with slope β tangent to the thermal boundary. Equation
(23) trivially follows from this.  ent. In Fig. 4, we represent in the energy-entropy diagram for
S
β boundary, the equilibrium temperature given by the constant
Aβ (ρ)
A(σ)
xρ
xσ
E
Figure 3. Representation of the β-athermality Aβ (ρ) and the ather-
mality A(σ) in the energy-entropy diagram.
Note that the β-athermality can be written as
 
Aβ (ρ) = β Fβ (ρ) − Fβ (γ(β)) , (26)
that is, it is β times the work that can be potentially ex-
tracted from a state ρ when having a bath at temperature β
(see Eq. (14)). Hence, thanks to lemmas 5 and 9 both the
free energy and the athermality can be expressed in terms of a
minimization over standard β-free energies
 
F(ρ) = min Fβ (ρ) − Fβ (γ(β))
β
   (27)
A(ρ) = min β Fβ (ρ) − Fβ (γ(β))
β
6
S
βB
SB
βS
S̄ βE
SA βA
Ē := 12 (EA + EB )
S̄ := 12 (SA + SB )
E
EA Ē EB
Figure 4. Energy-entropy diagram of a system with Hamiltonian H.
Two systems with the same Hamiltonian H and initially at different
temperatures βA and βB equilibrate to different temperature depend-
ing on the approach taken: minimum energy principle vs. maximum
entropy principle. The equilibrium temperature when entropy is con-
served β−1
S is always smaller than the equilibrium temperature when
energy is conserved β−1
E , i. e. βS > βE .
where the minimum is respectively attained by the intrinsic
temperature β(ρ) and the spontaneous equilibration tempera-
ture β̃(ρ).
Note that, while for positive temperatures both the ather-
mality and the free energy are a measure of out of equilib-
rium, for negative temperatures, states with small athermality
are highly active and have huge free energies.
Let us finally show that in situations as the equilibration of
a hot body in contact with a cold one, the intrinsic temperature
and the spontaneous equilibration one are not so much differ-
the particular case of the equilibration of a system composed
of two identical subsystems initially at different temperatures
βA and βB . As expected and due to the concavity of the thermal
energy constrain (max. entropy principle) β−1 E is larger that
the one given by the constant entropy constrain (min. energy
principle) β−1
S , i. e. βE < βS . The difference in bound energies
of these two thermal states corresponds precisely to the work
extracted in the constant entropy scenario.
VII. WORK, HEAT AND THE FIRST LAW
In thermodynamics, the first law deals with the conserva-
tion of energy. In addition, it dictates the distribution of en-
ergy over work and heat, that are the two forms of thermody-
namically relevant energy transfer.
Let us define a thermodynamic process involving a system
A and a bath B as a transformation ρAB → ρ0AB that conserves
the global entropy S (ρAB ) = S (ρ0AB ). In standard thermody-
namics, the bath is by construction assumed to be initially
in a thermal state and completely uncorrelated from the sys-
tem. In such scenario, the heat dissipated in a process is usu-
ally defined as the change in the internal energy of the bath
∆Q = E(ρ0B ) − E(ρB ) where ρ(0) B = Tr A ρAB is the reduced
(0)
state of the bath. This definition, however, may have some
limitations and an alternative definitions have been discussed
 7

S Thus, definition 10 becomes ∆Q = T ∆S B in the limit of small

∆Q entropy changes in which T (S B ) ' T (S 0B ), while it differs from
the common definition and ∆Q , ∆E B .
0
TB ∆SB In the case that the initial state of the environment is ther-
SB
x0ρB mal ρB = e−HB /T /Tr (e−HB /T ) with Hamiltonian HB at temper-
ature T , then the heat definition is lower and upper bounded
SB
xρ B by
∆EB
E T ∆S B ≤ ∆Q ≤ ∆E B , (31)
EB 0
EB
where the three quantities have been represented in the
energy-entropy diagram in Fig. 5. If, on top of that, the pro-
Figure 5. Representation in the energy-entropy diagram of the dif-
ferent notions of heat: ∆Q as change of the bound energy of the bath, cess slightly perturbs the environment ρ0B = ρB +δρB , then, the
∆E B as the change in internal energy, and T B ∆S B . In this example, three definitions of heat coincide, in the sense that
the bath is initially thermal but this does not have to be necessarily
the case. T ∆S B + O(δρ2B ) = ∆Q = ∆E B − O(δρ2B ) . (32)

recently [28]. In particular, in [28], an information theoretic
approach suggests heat to be defined as ∆Q = T B ∆S B with T B
being the temperature and ∆S B = S (ρ0B ) − S (ρB ) is the change
in von Neumann entropy of the bath.
In the present approach, we go beyond the restriction that
the environment has to have a definite predefined temperature,
and be in the state of the Boltzmann-Gibbs form. For an ar-
bitrary environment, which could even be athermal, the heat
transfer is defined in terms of bound energies, as in the fol-
lowing.
Definition 10 (Heat). Given a system A and its environment B,
Λep
the heat dissipated by the system A in the process ρAB −−→ ρ0AB
is defined as the change in the bound energy of the environ-
ment B, i.e.
∆Q B B(ρ0B ) − B(ρB )
where B(ρ(0)
B ) is the initial (final) bound energy of the bath B.
Note that heat is a process dependent quantity in the sense
that there might be different processes with the same initial
and final marginal state for A, but different marginals for B.
Because the global process is entropy preserving, processes
that lead to the same marginal for A, but different marginals
with different entropies for B, necessarily imply enabling a
different amount of correlations between A and B measured
by the mutual information I(A : B) B S A + S B − S AB . That is,
∆S A + ∆S B = ∆I(A : B).
Lemma 11 (Connections among heat definitions). Given a
system A and its environment B, the heat absorbed by A in
Λep
a globally entropy preserving process ρAB −−→ ρ0AB can be
written as
Z S 0B
∆Q = T (s)ds, (29)
SB
where S (0)B = S (ρ B ) is the initial (final) entropy of B, and
(0)
T (S ) is the intrinsic temperature of the thermal state of B with
entropy S . Because of the continuity of T (S ), there exists a
s∗ ∈ [S B , S 0B ] such that
∆Q = T (s∗ )∆S B .
That is, in the limit of large thermal baths, the three definitions
become equivalent.
Proof. Equation (29) is a consequence of the definitions of
heat (28) and bound energy (6), together with Eq. (5). Equa-
tion (30) is proven by using the mean value theorem for the
function T (s). The lower bound in (31) is due to the concav-
ity of the thermal boundary together with the reminder theo-
rem of a first order Taylor expansion of the thermal bound-
ary (see Fig. 5). The upper bound in (31) is a consequence
of the positivity of the free-energy. Finally, Eq. (32) follows
from the fact that the thermal state ρB is a minimum of the
standard free-energy FT (ρ) = E(ρ) − T S (ρ), which implies
∆FT,B = O(δρ2B ), and ∆E B = T ∆S B + O(δρ2B ). 
Once the heat has been introduced and related to the stan-
dard definitions of heat, let us define work.
(28)
Definition 12 (Work). For an arbitrary entropy preserving
transformation involving a system A and its environment B,
ρAB → ρ0AB , with fixed non-interacting Hamiltonians HA and
HB , the worked performed on the system A is defined as,
∆WA B W − ∆F B (33)
where W = ∆E A + ∆E B is the work cost of implementing the
global transformation and ∆F B = F(ρ0B ) − F(ρB ).
Now equipped with the notions of heat and work, the first
law takes the form of a mathematical identity.
Lemma 13 (First law). For an arbitrary entropy preserving
transformation involving a system A and its environment B,
ρAB → ρ0AB , with fixed non-interacting Hamiltonians HA and
HB , the change in energy for A is distributed as
∆E A = ∆WA − ∆Q . (34)
Proof. The proof follows from the definitions of work and
heat. 
Clearly, in a process, −∆WA is the amount of “pure” energy
that can be transferred and stored in a battery. While, the ∆QA
is the change in energy due to flow of information from/into
(30) the system.

VIII. SECOND LAW
The second law of thermodynamics is formulated in many
different forms: an upper bound on the extracted work, the im-
possibility of converting heat into work completely etc. In this
section we show how all these formulations are a consequence
of the principle of entropy conservation.
A. Work extraction
Major concern in thermodynamics is to convert any form of
energy into work, which can be used for any application with
certainty.
Lemma 14 (Work extraction). For an arbitrary composite
system ρ, the extractable work by any entropy preserving pro-
cess ρ → ρ0 , W = E(ρ) − E(ρ0 ) is upper-bounded by the free
energy
W ≤ F(ρ) (35)
where the equality is saturated if and only if ρ0 = γ(ρ).
If the system has the particular structure ρ = ρA ⊗ γB (T B )
with γB (T B ) being thermal at temperature T B , then
W ≤ FT B (ρA ) − FT B (γA (T B )) (36)
where FT (·) is the standard out of equilibrium free energy and
the equality is saturated in the limit of an infinitely large bath
(infinite heat capacity).
Proof. By using E(ρ0 ) ≥ B(ρ0 ), one gets
W = E(ρ) − E(ρ0 ) ≤ E(ρ) − B(ρ0 ) = F(ρ), (37)
where note that B(ρ0 ) = B(ρ).
When the system has a composite structure ρ = ρA ⊗γB (T B ),
then
F(ρA ⊗ γB (T B )) = E(ρA ) + E(γB (T B ))
(38)
− (B(γA (T AB )) + B(γB (T AB )))
where T AB is the intrinsic temperature of the joint state ρA ⊗
γB (T B ). By considering that thermal states have zero free en-
ergy, E(γ) = B(γ), and reshafling the terms above, we recover
the first law of thermodynamics
F(ρA ⊗ γB (T B )) = −∆E A − ∆QA (39)
where we have identified ∆E A = E(γA (T AB )) − E(ρA ) and
∆QA = B(γB (T AB )) − B(γB (T B )). From Eq. (31) in Lemma
11, we have that ∆Q ≥ T B ∆S B . By using now that the whole
process is entropy preserving and that initially and finally
the subsystems A and B are in a product state, we have that
∆S A = −∆S B , and thus,
F(ρA ⊗ γB (T B )) ≤ −∆E A − T B ∆S B = − (∆E A − T B ∆S A ) (40)
which proves (36). Finally, in the limit of an infinitely large
bath, the intrinsic temperature T AB tends to the bath temper-
ature T B and ∆Q = T B ∆S B , which saturates the bound in
Eq. (40).
8
To see that the intrinsic temperature T AB tends to T B in the
limit of large baths, let us increase the bath size by adding
several copies of it. The entropy change of such bath reads
∆S B = S (γ⊗n ⊗n
B (T AB )) − S (γ B (T B ))
(41)
= n (S (γB (T AB )) − S (γB (T B ))) .
By considering the bound ∆S B = −∆S A ≤ |A|, we get
|A|
S (γB (T AB )) − S (γB (T B ) ≤ , (42)
n
which, together with the the continuity of S (γ(T )), proves
limn→∞ T AB = T B . 
One of most important questions in thermodynamics is if
heat can be converted into work. Any form of energy can be
transformed into heat and then finally be converted in “me-
chanical” work. Therefore it was very much fundamental to
understand, not only qualitatively but also quantitatively, to
what extend or efficiency the heat can be transformed into
work. These quests led to various other formulations of sec-
ond law in standard thermodynamics, like Clausius statement,
Kelvin-Planck statement and Carnot statement, to mention a
few.
Similar questions can be posed in the frame-work consid-
ered here. That is if bound energy can be converted into free
energy and if so, to what extend it can be done. In the fol-
lowing, analogous forms of second laws that consider these
questions qualitatively as well as quantitatively.
B. Clausius statement
Second law, in thermodynamics, not only dictates the direc-
tion of state transformations, but also put fundamental bound
on extractable work from such transformations. Here we first
concentrate on the bounds on extractable work and introduce
analogous versions of second law in our setup.
Lemma 15 (Clausius statement). No iso-informatic equili-
bration process is possible whose sole result is the transfer
of bound energy (i.e., heat) from an equilibrium state with βC
to another equilibrium state with βH , where βC > βH .
Proof. In order to prove, we need to show that the iso-
informatic equilibration process cannot lead to an increase of
∆β = βH − βC . Further any iso-informatic process that leads
to such increase must requires a supply of energy. First con-
sider the case where βC = βH (∆β = 0), which means the
corresponding states γC and γH are in equilibrium between
each other and jointly acquire the min-energy state. The min-
energy principle says that any iso-informatic transformation
to make βC , βH is bound to increase the sum of their indi-
vidual min-energy. Therefore it requires a inflow of energy,
which is nothing but introduction of work.
Note, for a completely passive state γ with given Hamilto-
nian H and β, the change in bound energy, which is nothing

but heat in our definition, due to an infinitesimal change in
entropy dS (γ), is given by
1 working body is left in the same state and uncorrelated with
dE(γ) = dS (γ). (43)
β the bath(s). In other words, under a global entropy preserving
Therefore larger the β, smaller be the change in internal en-
ergy for a fixed infinitesimal change in entropy in the state.
Now consider the case where βC > βH . Any flow of heat from
γC to γH will lead to a reduction of entropy in the former. That
will also lead to an equal increase of of the same in the latter.
A very small amount of bound energy will result in a infinites-
imal entropy flow, say dS from γC to γH . However the change
in internal energy (dE(γC/H ) = βdSC/H
) would not be equal and
that is −dE(γC ) < dE(γH ), for βC > βH . As a consequence,
in this iso-informatic process the overall energy is bound to
increase, which is not possible without an influx of energy or
work. Therefore, without external work the process will never
take place spontaneously.

C. Kelvin-Planck statement
Lemma 16 (Kelvin-Planck statement). No iso-informatic
equilibration process is possible whose sole result is the ab-
sorption of bound energy (heat) from an equilibrium state and
its complete conversion into work.
Proof. To prove Kelvin-Planck statement, we consider fol-
lowing example of two CP states γA and γB with βA and βB
respectively, which undergo an iso-informatic equilibration
transformation as
ΛEP
γA ⊗ γB ⊗ |0ih0|W −−−−→ γAB ⊗ |WihW|W ,
ABW
(44)
γAB is the final joint equilibrium state with βAB . Since the ini-
tial states are CP states, a non-zero W could only result from
heat transfer. Let say βA < βB . Then βA < βAB < βB . In
this case the heat is flown out from γA , say ∆QA , with an as-
sociated decrease in its entropy. Again the information con-
servation of whole process guarantees an entropy increase in
γB . Therefore, there has to be an associated increase in bound
energy (heat) content in γB . As a result, a part of ∆QA could
be converted to work and that is
W = ∆QA − ∆QB . (45)
Note any transfer of heat, i.e., −∆QA > 0, also bounds
∆QB > 0. As a consequence, heat cannot be converted into
work completely.  η=
D. Carnot statement
A heat engine extracts work from having a situation in
which two, possibly “finite”, systems A and B have different
intrinsic temperatures T A and T B . The work extraction is usu-
ally implemented in practice by means of a working body S
9
which cyclically interacts with A and B. Here we do not mind
on how the working medium interacts with the baths A and
B, but just assume that at the beginning of every cycle the
process, the working body only absorbs heat from a bath and
releases to other one, and at the end of the cycle it comes back
to its initial state. Therefore it is irrelevant concerning the dis-
cussion of energy-entropy flows between the working body
and the bath(s). From this perspective, a heat engine com-
posed to non-interacting systems (A and B), locally at equi-
librium with two different intrinsic temperatures (T A and T B ,
respectively) and an arbitrary working body.
Now, after one cycle, driven by a global entropy preserving
operation, the states change to ρAB → ρ0AB . Without loss of
generality, let us also consider that the initial and final states
are uncorrelated. Then efficiency of work extraction, defined
as η = W/Q, where Q is the heat absorbed and W is the work
extracted, is bounded as in the following:
Lemma 17 (Carnot statement). For an engine following iso-
informatic process, working with two locally equilibrium
states γA and γB with intrinsic temperatures T A > T B , the
maximum efficiency of work extraction is given by
∆BB
η=1− , (46)
∆BA
where ∆BA and ∆BB are the change in bound energies of the
systems A and B respectively.
Proof. Here the systems could be small in size. Then the max-
imum extractable work due to the transformation ρAB → ρ0AB
is given by
W = F(ρAB ) − F(ρ0AB ) > 0 . (47)
Lets us consider that the initial and final states are uncorre-
lated and remains, locally, in complete passive states. This,
in fact, leads to maximum work extraction. Further bound
energy could only flow from A to B as restricted by the Clau-
sius statement and −∆BA > ∆BB . Then we could recast the
extractable work, with T A > T B , as
W = −∆BA − ∆BB . (48)
Therefore the working body absorbs −∆BA of bound energy,
which is analogous to heat absorption. Therefore, the maxi-
mum efficiency of conversion of bound energy to work is
W ∆BB
=1− . (49)
| − ∆BA | ∆BA

Now let us consider the limits ∆BA ≪ BA and ∆BB ≪ BB .
This can be achieved if we alter the systems infinitesimally or
if the systems are considerably large such that letting it inter-
act with the working body affects the system minimally. In
this case, the maximum conversion efficiency will converge to

the one, analogous to Carnot, but with internal temperature of
the systems, as
TB
ηC = 1 − . (50)
TA
Therefore we recover Carnot statement of second law in
the large bath limit. But, for finite baths, in every steps
(through iso-information operation) the temperatures of the
bath change and the ratio TT AB increases. For an n-step process,
T B(n−1) T B(n)
the ratio T A(n−1)
6 T A(n)
6 1, where T X(i) is the temperature before
the ith step starts, and asymptotically converges to 1. There-
fore, it is easy to follow that for finite baths, the work extrac-
tion efficiency decreases in every step and remains smaller that
the one of Carnot.
IX. THIRD LAW
The third law of thermodynamics establishes the impossi-
bility of attaining the absolute zero temperature, or according
to Nernst, is stated as: “it is impossible to reduce the entropy
of a system to its absolute-zero value in a finite number of
operations”. The third law of thermodynamics has been very
recently proven in Ref. [42]. The unattainability of absolute
zero entropy is a consequence of the unitarity character of the
transformations considered. For instance, consider the trans-
formation
ρS ⊗ ρB → |0ih0| ⊗ ρ0B (51)
where ρB is a thermal state of the bath B and S is the system to
be cooled down (erased), initially in a state ρS with rank(ρS ) >
1. Here the dimension of the Hilbert space of the bath, dB , is
considered to be arbitrarily large but finite. As ρB is a full-rank
state, the left hand side and the right hand side of Eq. (51) have
different ranks, and they cannot be transformed via a unitary
transformation. Thus, irrespective of work supply, one cannot
attain the absolute zero entropy state.
The zero entropy state can only be achieved for infinitely
large baths and a sufficient work supply. Assuming a locality
structure for the bath’s Hamiltonian, this would take an in-
finitely long time. However, in the finite dimensional case, a
quantitative bound on the achievable temperature given a fi-
nite amount of resources (e. g. work, time) is derived in [42].
Note that our set of entropy preserving operations is more
powerful than unitaries. In this case, transformation (51) is
possible given that the
S (ρS ) 6 log dB − S (ρB ) (52)
the required work to do so will be W = F(|0ih0| ⊗ ρ0B ) − F(ρS ⊗
ρB ). As it has been discussed in Sec. II, entropy preserving
operations can be implemented as global unitaries acting on
infinitely many copies. Thus, the attainability of the absolute
zero entropy state by means of entropy preserving operations
is in agreement with the cases of infinitely large baths and
unitary operations [42].
10
In conclusion, the third law of thermodynamics is a conse-
quence of the microscopy reversibility (unitarity) of the trans-
formation and is not respected by entropy preserving opera-
tions.
X. A TEMPERATURE INDEPENDENT RESOURCE
THEORY OF THERMODYNAMICS
In this section we connect the framework developed above
with the standard approach of thermodynamics as a resource
theory. The main ingredients of a resource theory are the state
space, which is usually compatible with a composition oper-
ation – for instance, quantum states together with the tensor
product– and a set of of allowed state transformations. In
a first attempt to have a resource theory of thermodynamics
from our framework, we look like to be bounded to transfor-
mations within an equivalence class, that is, between states
with equal entropies and same Hilbert spaces. In that case,
the set of allowed operations seem reasonable to be defined
by the non-increasing energy transformations and the mono-
tone to be given by both energy and entropy. To compare
and explore inter-convertibility between states with different
entropies or dimensions, it is necessary to extend the set of
operations, and in particular to add a composition rule. In or-
der to be able to connect states with different entropies and/or
number of particles it is natural to consider different number
of copies.
Let us restrict first to the case in which the set of operations
is entropy preserving regardless the energy. We wonder what
is the rate of transforming ρ into σ and consider without loss
of generality that S (ρ) ≤ S (σ). Then, there is a large enough
n such that
S (ρ⊗n ) = S (σ⊗m ). (53)
With such a trick of considering a different number of copies,
we have brought ρ and σ which had different entropies into
the same manifold of equal entropy. There is a subtlety here
which is that ρ⊗n and σ⊗m live in general in spaces of different
dimension. This can be easily circumvented by adding some
product state, i. e.
 
Λ ρ⊗n = σm ⊗ |0ih0|n−m (54)
where ρ⊗n and σ⊗m ⊗ |0ih0|n−m lie now in spaces of the same
dimension. Equation (54) can also be understood as a pro-
cess of randomness compression in which the information in
n copies of ρ is compressed to m < n copies of σ and n − m
systems have been erased. Thus, in the case of having only an
entropy preserving constraint, the transformation rate can be
determined from (53)
m S (ρ)
rB = . (55)
n S (σ)
In thermodynamics, however, energy must also be taken
into account. When one does that, it could well be that the
process (54) is not energetically favourable since E(ρ⊗n ) <
 11

S XI. THERMODYNAMICS WITH MULTIPLE

ρ⊗n → σ ⊗m ⊗ φ⊗(n−m) CONSERVED QUANTITIES

xσ The formalism for thermodynamics developed above can be

xρ easily extended to situations with multiple conserved quanti-
xφ ties.
a m a
r= n = b
b
E
A. The charges-entropy diagram

Figure 6. Representation in the energy-entropy diagram of how

the energy and entropy conservation constraints in the transformation The energy-entropy diagram then is extended to include the
ρ⊗n → σ⊗m ⊗ φ⊗n−m imply xρ , xσ and xφ to be aligned and xρ to lie in additional conserved quantities. Let us introduce the charges-
between xσ and xφ . entropy diagram as:
Definition 18 (Charges-entropy diagram). Let us consider a
E(σ ). In such a case, we would need more copies of ρ,
⊗m system with q conserved quantities (not necessarily commut-
in order for the transformation to be energetically favourable. ing) Qk for k = 0, . . . , q − 1 and Q0 = H the Hamiltonian.
This would force us to add entropic ancillas, since otherwise, The charges-entropy diagram is the subset of points in Rq+1
the initial and final state of such a process would not have the produced by all the states ρ ∈ B(H) with coordinates
same entropy xρ = (Q0 (ρ), Q1 (ρ), . . . , Qq−1 , S (ρ)) . (61)
Λ(ρ ) = σ
⊗n ⊗m
⊗φ ⊗n−m
(56) The upper boundary of the charges-entropy diagram is de-
where the number of copies n and m have to fulfil the energy scribed by the Generalized Gibbs Ensemble (GGE) instead of
and entropy conservation constraints the canonical ensemble, i. e.

E(ρ⊗n ) = E(σ⊗m ⊗ φ⊗n−m ) 1 P

γ(~β) B e− k βk Qk , (62)
(57) Z~β
S (ρ⊗n ) = S (σ⊗m ⊗ φ⊗n−m ) .
where Z~β B Tr (e− k βk Qk ) is the generalized partition func-
P
The above conditions can be easily written as a geometric
equation of the points xψ = (E(ψ), S (ψ)) with ψ ∈ {ρ, σ, φ} tion. These states are precisely the states of maximum en-
in the energy-entropy diagram tropy among all with prescribed expectation values for hQk i,
k = 0, ..., q − 1. We call this boundary the equilibrium bound-
xρ = r xσ + (1 − r) xφ , (58) ary.
The equilibrium boundary is mathematically described by
where r B m/n is the conversion rate, and we have only used
~ for which there exists a ~β ∈ Rq such that
all the points Q(ρ)
the extensivity of both entropy and energy in the number of
copies, e.g. E(ρ⊗n ) = nE(ρ). q−1
X
Equation (58) implies that the three points xρ,σ,φ need to be βk Qk (ρ) − S (ρ) = − log Zβ . (63)
aligned. In addition, the fact that 0 ≤ r ≤ 1 implies that xρ lies k=0
in the segment xσ and xφ (see Fig. 6). The conversion rate r
has then a geometric interpretation. It is the relative Euclidean Note that, given some ~β, the points Q(ρ)
~ that fulfil
distance between xφ and xρ over the total distance kxσ − xφ k,
q−1
or in other words, is the fraction of the path that one has run X
when going form xφ to xσ and reaches xρ (see Fig. 6). βk Qk (ρ) − S (ρ) = C (64)
In order for the conversion rate r from ρ and σ to be max- k=0

imum, the state φ needs to lie in the boundary of the energy- belong to a hyperplane of dimension q. The hyperplane with
entropy diagram, that is, it needs to be either a thermal or a C = − log Zβ is the one that is tangent to the equilibrium
pure state. This allows us to quantitatively determine the con- boundary. The normal vector to this family of hyperplanes
version rate r by first finding the temperature of the thermal (64) is precisely (~β, −1). This agrees with the fact that the
state φ which satisfies tangent plane has a slope βk in the k-th direction, i. e.
S (σ) − S (φ) E(σ) − E(φ) ∂S
= (59) = βk ∀ k ∈ {0, . . . , q − 1} , (65)
S (σ) − S (ρ) E(σ) − E(ρ) ∂Qk
and then where here the Qk need to be understood as coordinates in
m S (ρ) − S (φ) the charges-entropy diagram (not matrices). This is shown in
r= = . (60) Fig. 8. Equation (65) can also be written in a vectorial form
n S (σ) − S (φ)
as
See the geometrical interpretation in Fig. 6.
Note how (60) becomes (55) in the case of φ being pure. ~ ,
~β = ∇S (66)
 12

where the partial trace is performed on the ancillary

qubits A, k·k1 is the trace-norm and Q0k 6 O(n2/3 )
with k · k the operator-norm. Here the global charge is
defined by
 n−1 
X 
Q̄k B  Ii ⊗ Qk ⊗ In−i−1  ⊗ IA + In ⊗ Q0k . (68)
i=0

The proof is shown in the appendix A and is a straightfor-

ward generalization of Theorem 6 in [37]. The above theorem
can be extended to the case of non-commuting charges relying
on the fact that the global charges asymptotically commute.

B. Athermality and free entropy

Figure 7. Charges-entropy diagram of a system with 2 conserved
quantities E and L. Proceeding as above, let us introduce some quantities that
will be relevant in the following.

S Definition 20 (~β-athermality). The ~β-athermality of a state ρ

is defined as
∂S
βk = ∂Qk
q−1
X
Aβ~ (ρ) A~β (ρ) B βk Qk (ρ) − S (ρ) + log Z~β . (69)
k=0
(βk , −1) A(σ)
xρ
xσ The ~β-athermality of a point with coordinates (Q(ρ),
~ S (ρ))
can be interpreted geometrically in the charges-entropy dia-
Qk
gram as the vertical distance from the hyperplane tangent to
the equilibrium boundary with normal vector (~β, −1). This is
Figure 8. Transverse section of the charges-entropy diagram in the
represented in Fig. 8. Note that the ~β-athermality is also intro-
plane Qk − S . The normal vector to the tangent plane has coordinates duced in [43] as as the free entropy.
(~β, −1), which in the section appears as (βk , −1). The ~β-athermality, The ~β-athermality can also be written as
A~β (ρ), and the absolute athermality, A(σ), are respectevely repre-
sented for two states ρ and σ. A~β (ρ) = D(ρ||γ(~β)) , (70)
with D(ρ||σ) B Tr (ρ log ρ − ρ log σ) being the relative en-
tropy. Hence, because of the positivity of the relative entropy
and ~β corresponds to the direction of maximal variation of the D(ρ||σ) > 0, we have that
entropy.
In order to microscopically justify the charges-entropy dia- q−1
X
gram, we need to show that states with the same expectation S (ρ) 6 βk Qk (ρ) + log Zβ . (71)
values for the conserved quantities can be unitarily connected k=0
in the limit of many copies. This is proven in the following Note now that the right hand side corresponds to the entropy
theorem. coordinate of the hyperplane with normal vector ~β tangent to
Theorem 19. Given any two states ρ and σ and a set of q mu- the equilibrium boundary. Thus, any tangent hyperplane with
tually commuting conserved quantities {Qk }q−1 normal vector ~β leaves all the points of the charges entropy di-
k=0 , the following agram below it. This proves that the charges entropy diagram
two statements are equivalent.
is a convex set.
• The states ρ and σ have equal entropies S (ρ) = S (σ) In a similar way we define the absolute athermality or sim-
and expectation values of the conserved quantities ply athermality in the following.
Qk (ρ) = Qk (σ) for any k. Definition 21 (Absolute athermality). The absolute athermal-
p ity of a state ρ is defined as
• There exists one additional system A of O( n log n)
qubits, which has all q charges Q0k and a global uni- A(ρ) B min A~β (ρ) . (72)
~β∈Rq
tary U commuting with the global conserved quantities,
[U, Q̄k ] = 0, such that The athermality of a state ρ can geometrically be under-
  stood as the vertical distance from the thermal boundary (see
lim kTr A Uρ⊗n ⊗ ηA U † − σ⊗n k1 = 0 , (67) Fig. 8).
n→∞

C. Charge extraction
A first relevant scenario of thermodynamics with multiple
conserved quantities is the extraction of a charge of the system
while keeping constant the other charges. For such set of op-
erations the system is restricted to move along a straight line
in the direction of the extracted charge within the iso-entropic
hyperplane (see Fig. 9).
The bound charge for such scenario is given by
Bk (ρ) B min Qk (σ) = Qk (γk (ρ)) ,
σ : S (σ)=S (ρ);
Qi (σ)=Qi (ρ)∀i,k
where γk (ρ) is the GGE state that attains the minimum and
corresponds to the point of the equilibrium boundary which is
intersected by the straight-line with direction k which passes
through ρ. This is represented in Fig. 9 (left). The ~β corre-
sponding to γk can be geometrically determined by the normal
vector of the tangent plane to the equilibrium surface in that
point.
The free charge Fk (ρ) is then defined by
Fk (ρ) = Qk (ρ) − Bk (ρ)
and corresponds to the maximum amount of charge that can
be extracted given the restrictions of constant entropy and
charges. This is represented in Fig. 9 (left) for a case of 2
conserved quantities.
We study the single charge extraction by means of unitary
conserving operations in the many copy limit in Appendix B.
In particular, we prove the following theorem.
Theorem 22. The asymptotically completely passive states
for work extraction under asymptotically conserved charges
Q0 , . . . , Qq−1 are precisely the generalised grand canonical
Gibbs states
 q−1 
1  X 
γE,Q~ = exp − β j Q j  ,
Z j=0
or limit points in the closure of the exponential family, which
correspond to diverging inverse temperature β0 and/or chem-
ical potentials β j .
Furthermore, the tensor product of several states is asymp-
totically passive if and only if they all are individually asymp-
totically passive and they share the same intensive parameters
~β.
D. Extraction of a Generalized Potential
An alternative but also relevant scenario is when the set of
operations allow for the variation of the charges and the aim
is the extraction of a generalized potential
q−1
X the single charge situation. Both first and second laws can be
Vµ̂ (ρ) B µk Qk (ρ)
k=0
13
E E
ρ
F (ρ )
F (ρ )
γ
ρ
γ
μ
L L
(73)
Figure 9. Constant entropy section of a charges-entropy diagram
with 2 charges E and L. (Left) Scenario of extraction of a single
charge while keeping constant the other charges. The system is con-
strained to move in one dimensional line and free-charge F(ρ) of a
state ρ is the distance from the boundary in such direction. (Right)
Scenario in which the charges are allowed to change and the aim is
the extraction of a potential Vµ̂ (ρ) = q−1
P
k=0 µk Qk (ρ). The parallel pur-
ple lines represent equipotential surfaces of Vµ̂ . The GGE state γ is
the state which minimizes V and belongs to the hyperplane which is
tan tangent to the equilibrium boundary.
(74)
where µ̂ B (µ0 , . . . , µq−1 ) specifies the weight µk of every
charge Qk in the generalized potential Vµ̂ . For convenience,
µ̂ is assumed to be normalized kµ̂k = 1 in the Euclidean norm
and its components to be positive µk > 0. An alternative nor-
malization could be to take the coefficient of the Hamiltonian
µ0 = 1 as it is usually made in the Grand-canonical ensemble.
In this setting, the bound potential is defined as
Bµ̂ (ρ) B min Vµ̂ (σ) = Vµ̂ (γµ̂ (ρ)) , (76)
σ : S (σ)=S (ρ)
where γµ̂ (ρ) is the state that attains the minimum which is
again a GGE. Note that the ~β values of the state γµ̂ (ρ) need to
be proportional to the unitary vector µ̂
~β = βµ̂ (77)
with β a scalar that is determined by the equal entropies condi-
tion S (ρ) = S (γµ̂ (ρ)). This can be seen geometrically in Fig. 9
(right) or analytically by making the simple observation that,
given a new Hamiltonian H̃ = Vµ̂ , the min energy principle
singles out e−βH̃ as the state that attains the minimum.
Analogously, the free potential is given by
Fµ̂ (ρ) = Vµ̂ (ρ) − Bµ̂ (ρ) , (78)
and corresponds to the maximum amount of generalized po-
tential Vµ̂ that can be extracted under entropy preserving oper-
ations. This situation has been diagrammatically represented
in Fig. 9.
The scenario with a generalized potential Vµ̂ is analogue to
(75) stated as in Secs. VII and VIII but replacing the free energy
F(ρ) by Fµ̂ (ρ).

E. Second law
In order to give express the second law irrespective of a
particular choice of the generalized potential, let us formulate
it as an inequality that
Any process (not necessarily entropy preserving) that
brings an initial generalized Gibbs state γB (~β) out of equilib-
rium fulfils
X sected by the straight line that goes from σ to ρ (see Fig. 6 for
βk ∆QkB > ∆S B , (79)
k
where the inequality is only saturated in the limit of large
baths (small variations of entropy) and when final state is also
at equilibrium (GGE). Geometrically (79) is a trivial conse-
quence that the charges entropy diagram is upper bounded by
any of its tangent planes.
In the particular case of an entropy preserving process on
an initial bipartite state ρA ⊗ γB (~β), the change in mutual in-
formation between the subsystems A and B
∆I = ∆S A + ∆S B > 0 ,
which together with Eq. (79) implies
X acquires the same temperature, while it achieves thermal equi-
βk ∆QkB > −∆S A .
k sumption and considers bath to be a finite and small system,
Note that the above equation (81) is equivalent to the second
law formulated in [43]. We see here that such a law is also
valid for baths of arbitrary small sizes. Equation (81) is only
saturated in the limit of large system sizes and in absence of
correlations in the final state ρ0AB = ρ0A ⊗ γ0B .
F. Interconvertibility rates
The convertibility rate from a quantum state ρ’s to a state σ
Λ(ρ⊗n ) = σ⊗m ⊗ φ⊗n−m
given the constraints on the conservation of the entropy and
the charges Qk
Qk (ρ⊗n ) = Qk (σ⊗m ⊗ φ⊗n−m ) ∀ k = 0, . . . , q − 1
S (ρ⊗n ) = S (σ⊗m ⊗ φ⊗n−m ) .
The above conditions can be again written as a geometric
equation of the points xψ = (Q0 (ψ), . . . , Qq−1 (ψ), S (ψ)) with
ψ ∈ {ρ, σ, φ} in the energy-entropy diagram
xρ = r xσ + (1 − r) xφ ,
where we have merely used the extensivity of both entropy
and the conserved quantities in the number of copies, e.g.
S (ρ⊗n ) = nS (ρ).
Thus, by means of the same argument used in the previous
section, the convertibility rate from ρ to σ reads
m S (ρ) − S (φ)
r= = .
n S (σ) − S (φ)
14
where the state φ is determined from the following set of q
equations
S (σ) − S (φ) Qk (σ) − Qk (φ)
= k = 0, . . . , q − 1 . (86)
S (σ) − S (ρ) Qk (σ) − Qk (ρ)
The state φ can be found geometrically in the charges-entropy
diagram as the point in the equilibrium boundary that is inter-
the single charge example).
XII. DISCUSSION
In the recent years, thermodynamics, and in particular work
extraction from non-equilibrium states, has been studied in
the quantum domain, giving rise to radically new insights into
quantum thermal processes. However, in standard endeavour
of thermodynamics, be it classical or quantum, thermal baths
are considered to be considerably large in size. That is why
(80) if a system is attached with a bath, and allowed to exchange
energy and entropy, the bath stays intact and its temperature
remains unchanged. For the same reason the system finally
(81) librium with the bath. However, if one goes beyond this as-
then the traditional thermodynamics breaks down. This situa-
tion is very much relevant for thermodynamics in the quantum
regime, where both the system and the bath may be small. The
first problem that appears in such a situation is the notion of
temperature itself, since the finite bath may go out of thermal
equilibrium due to the exchange of energy with the system.
Therefore, it is absolutely necessary to develop a tempera-
ture independent universal thermodynamics, in which the bath
could be small or large, and would not get any special status.
In this work, we have formulated temperature independent
thermodynamics as an exclusive consequence of information
(82) conservation. We have relied on the fact that, for a given
amount of information, measured by the von Neumann en-
tropy, any system can be transformed to a state that possesses
minimal energy. This state is known as a completely passive
state that acquires a Boltzmann–Gibb’s canonical form with
an intrinsic temperature. We introduce the notions of bound
(83)
and free energy and use them to quantify heat and work re-
spectively.
Our findings show that:
• Zeroth law emerges as the consequence of information
conservation.
(84)
• First and second laws emerge as the consequence of en-
ergy conservation, together with information conserva-
tion.
• Third law emerges as the consequence of "strict" infor-
mation conservation (microscopic reversibility or uni-
tarity). Therefore there is no third law if one considers
(85) "average" information conservation.

All these results have been illustrated by means of the energy-
entropy diagram. Furthermore, we give a geometric interpre-
tation in the diagram to the relevant thermodynamic quantities
as well as the inter-convertibility rate between quantum states
under entropy and energy preserving operations.
We have demonstrated that the maximum efficiency of a
quantum engine with a finite bath is in general different and
smaller than that of an ideal Carnot’s engine. We have intro-
duced a resource theoretic framework for our intrinsic ther-
modynamics, within which we address the problem of work
extraction and inter-state transformations.
The information conservation based framework for ther-
modynamics, as well as the resource theory and the energy-
entropy diagram, is also extended to multiple conserved quan-
tities. In this case, the energy-entropy diagram becomes the
charges-entropy diagram and allows us to understand thermo-
dynamics in a geometrical way. In particular, we have stud-
ied the extraction of a single charge while keeping the other
charges conserved as well as the extraction of a generalized
potential. In the first scenario, we have shown that passiv-
ity under work extraction with conserved quantities yields a
consistent theory of grand canonical Gibbs states, even in the
generalised setting of non-commuting “conserved charges”,
as the unique states from which no work can be extracted. In
fact, and more generally, we have shown that the maximum
work extractable from any state by operations that asymptoti-
cally conserve the given charges is the difference between the
free energy of the state and that of the iso-entropic generalised
grand canonical Gibbs state.
Even in the commuting case, the theory we have presented
above is more complex than the Pusz/Woronowicz-Lenard
theory of passivity of energy alone. In particular, we cannot
[1] J. Gemmer, M. Michel, and G. Mahler, Lecture Notes in
Physics: Quantum thermodynamics, Vol. 748 (Berlin, Heidel-
berg: Springer, 2009).
[2] J. M. R. Parrondo, J. M. Horowitz, and T. Sagawa, Nature
Physics 11, 131 (2015).
[3] J. C. Maxwell, Theory of heat (Longmans, Green, and Co.:
London, New York, Bombay, 1908).
[4] H. S. Leff and A. F. Rex, Maxwell’s Demon: Entropy, Informa-
tion, Computing (London: Taylor and Francis, 1990).
[5] H. S. Leff and A. F. Rex, Maxwell’s Demon 2: Entropy, Classi-
cal and Quantum Information, Computing (London: Taylor and
Francis, 2002).
[6] K. Maruyama, F. Nori, and V. Vedral, Reviews in Modern
Physics 81, 1 (2009).
[7] L. Szilard, Zeitschrift für Physik 53, 840 (1929).
[8] R. Landauer, IBM Journal of Research and Development 5, 183
(1961).
[9] C. H. Bennett, International Journal of Theoretical Physics 21,
905 (1982).
[10] M. B. Plenio and V. Vitelli, Contemporary Physics 42, 25
(2001).
[11] L. del Rio, J. Aberg, R. Renner, O. C. O. Dahlsten, and V. Ve-
dral, Nature 474, 61 (2011).
15
proceed from passivity to complete passivity by first consider-
ing single systems (passivity) and then increasing numbers of
replicas of the system and copies of the state (complete pas-
sivity). Rather, we have to go directly to an asymptotic setting
because a strict conservation of the Q j could (and generically
will) trivialise the set of operations to extract work, hence we
discussed asymptotic charge conservation and asymptotic pas-
sivity.
Concerning the extraction of a generalized potential (a lin-
ear combination of charges), we have shown that it is anal-
ogous to the work extraction (the single charge case). Fi-
nally, we have determined the inter-convertibility rates be-
tween states with different entropy and charges.
ACKNOWLEDGEMENTS
We thank P. Faist, K. Gawedzki, N. Y. Halpern, R. B. Har-
vey, J. Kimble, S. Maniscalco, Ll. Masanes, J. Oppenheim, V.
Pellegrini, M. Polini, C. Sparaciari, A. Vulpiani and R. Zam-
brini for useful discussions and comments in both theoreti-
cal and experimental aspects of our work. We acknowledge
financial support from the European Commission (FETPRO
QUIC (H2020-FETPROACT-2014 No. 641122) and STREP
EQuaM), the European Research Council (AdG OSYRIS and
AdG IRQUAT), the Spanish MINECO (grants no. FIS2008-
01236, FISICATEAMO FIS2016-79508-P, FIS2013-40627-P,
FIS2016-86681-P, and Severo Ochoa Excellence Grant SEV-
2015-0522) with the support of FEDER funds, the Gen-
eralitat de Catalunya (grants no. SGR 874, 875 and 966),
CERCA Program/Generalitat de Catalunya and Fundació Pri-
vada Cellex. AR also thanks support from the CELLEX-
ICFO-MPQ fellowship.
[12] D. Reeb and M. M. Wolf, New Journal of Physics 16, 103011
(2014).
[13] C. E. Shannon, Bell System Technical Journal 27, 379 (1948).
[14] M. L. Nielsen and I. L. Chuang, Quantum Computation
and Quantum Information (Cambridge: Cambridge University
Press, 2000).
[15] T. M. Cover and J. A. Thomas, Elements of Information Theory,
2nd ed. (John Wiley and Sons, Inc., 2005).
[16] F. G. S. L. Brandão, M. Horodecki, J. Oppenheim, J. M. Renes,
and R. W. Spekkens, Physical Review Letters 111, 250404
(2013).
[17] O. C. O. Dahlsten, R. Renner, E. Rieper, and V. Vedral, New
Journal of Physics 13, 053015 (2011).
[18] J. Aberg, Nature Communications 4, 1925 (2013).
[19] M. Horodecki and J. Oppenheim, Nature Communications 4,
2059 (2013).
[20] P. Skrzypczyk, A. J. Short, and S. Popescu, Nature Communi-
cations 5, 4185 (2014).
[21] F. G. S. L. Brandao, M. Horodecki, N. Ng, J. Oppenheim, and
S. Wehner, Proceedings of the National Academy of Sciences
112, 3275 (2015).
[22] P. Ćwikliński, M. Studziński, M. Horodecki, and J. Oppen-
heim, Physical Review Letters 115, 210403 (2015).
 16

[23] M. Lostaglio, K. Korzekwa, D. Jennings, and T. Rudolph,

Physical Review X 5, 021001 (2015).
[24] D. Egloff, O. C. O. Dahlsten, R. Renner, and V. Vedral, New
Journal of Physics 17, 073001 (2015).
[25] M. Lostaglio, D. Jennings, and T. Rudolph, Nature Communi-
cations 6, 6383 (2015).
[26] R. Alicki and M. Fannes, Phys. Rev. E 87, 042123 (2013).
[27] M. Perarnau-Llobet, K. V. Hovhannisyan, M. Huber,
P. Skrzypczyk, N. Brunner, and A. Acín, Phys. Rev. X 5,
041011 (2015).
[28] M. N. Bera, A. Riera, M. Lewenstein, and A. Winter,
arXiv:1612.04779 (2016).
[29] A. J. Short, New Journal of Physics 13, 053009 (2011).
[30] J. Goold, M. Huber, A. Riera, L. del Rio, and P. Skrzypczyk,
Journal of Physics A: Mathematical and Theoretical 49, 143001
(2016).
[31] L. del Rio, A. Hutter, R. Renner, and S. Wehner, Physical Re-
view E 94, 022104 (2016).
[32] C. Gogolin and J. Eisert, Reports on Progress in Physics 79,
056001 (2016).
[33] S. Popescu, A. J. Short, and A. Winter, Nature Physics 2, 745
(2006).
[34] A. Puglisi, A. Sarracino, and A. Vulpiani, to be published
(2017).
[35] M. N. Bera, A. Acín, M. Kuś, M. Mitchell, and M. Lewenstein,
arXiv:1612.04779 (2016).
[36] F. Hulpke, U. V. Poulsen, A. Sanpera, A. Sen(De), U. Sen, and
M. Lewenstein, Foundations of Physics 36, 477 (2006).
[37] C. Sparaciari, J. Oppenheim, and T. Fritz, arXiv:1607.01302
(2016).
[38] E. T. Jaynes, Phys. Rev. 106, 620 (1957).
[39] E. T. Jaynes, Phys. Rev. 108, 171 (1957).
[40] W. Pusz and S. L. Woronowicz, Communications in Mathemat-
ical Physics 58, 273 (1978).
[41] A. Lenard, Journal of Statistical Physics 19, 575 (1978).
[42] L. Masanes and J. Oppenheim, Nature Communications 8,
14538 (2017).
[43] Y. Guryanova, S. Popescu, A. J. Short, R. Silva, and
P. Skrzypczyk, Nature Communications 7, 12049 (2016).
[44] A. Lenard, Journal of Statistical Physics 19, 575 (1978).
[45] W. Pusz and S. L. Woronowicz, Communications in Mathemat-
ical Physics 58, 273 (1978).
[46] S. D. Bartlett, T. Rudolph, and R. W. Spekkens, Rev. Mod.
Phys. 79, 555 (2007).
[47] A. Kitaev, D. Mayers, and J. Preskill, Phys. Rev. A 69, 052326
(2004).
[48] N. Yunger Halpern and J. M. Renes, Phys. Rev. E 93, 022126
(2016).
[49] N. Y. Halpern, arXiv[quant-ph]:1409:7845 (2014).
[50] N. Yunger Halpern, P. Faist, J. Oppenheim, and A. Winter,
Nature Communications 7, 12051 (2016).
[51] A. Winter, IEEE Transactions on Information Theory 45, 2481
(1999).
 17

SUPPLEMENTAL MATERIAL

Appendix A: Asymptotic equivalence of states with equal expectation values of the charges

In this section we show that any two states with equal expectation values of the charges {Qk }q−1
k=0 can be related in the limit of
many copies by a global unitary that commutes with the charges. In order for the unitary to commute with the global charge it
needs to also act on an ancillary system with a sublinear number of qubits. This result is a straightforward extension of Theorem
6 in [37] to the scenario with multiple mutually commuting conserved charges.
In the case of non-commuting charges, the proof of the asymptotic equivalence also holds, but is much more involved as
it relies on the asymptotic commutativity of non-commuting observables. It will be updated in a subsequent version of the
manuscript.
The main idea in Theorem 6, and in particular in Lemma 5, in Ref. [37] is that an energy non-conserving unitary can be turned
into energy conserving when acting on a larger system where a sufficiently large ancillary system has been added. The energy
conservation in the larger system is achieved thanks to the ancillary system that plays the role of a battery providing/storing the
lack/excess of energy in the system.
The essential
√ point to have asymptotic equivalence is that the ancillary system can be taken of a sublinear size in the number of
copies e.g. n, that is, the energy added/subtracted per copy vanishes in the limit of many copies n → ∞. The ancillary system
can be taken so “small” thanks to the fact that the global
√ energy fluctuations and, equivalently, the energy range of the relevant
energy subspace (also called typical
√ subspace) scale as n. Hence, the amount of energy that needs to be provided/stored by the
ancilary qubits also scales as n, which is sublinear in n.
More explicitly, in the following theorem, we present the above result with the scalings for the ancillary system and operator
norm of the charges taken analogous to [37]:
Theorem 19. Given any two states ρ and σ and a set of q mutually commuting conserved quantities {Qk }q−1
k=0 , the following two
statements are equivalent.
(a): The states ρ and σ have equal entropies S (ρ) = S (σ) and expectation values of the conserved quantities Qk (ρ) = Qk (σ) for
any k.
p
(b): There exists one additional system A of O( n log n) qubits, which has all q charges Q0k and a global unitary U commuting
with the global conserved quantities, [U, Q̄k ] = 0, such that
 
lim kTr A Uρ⊗n ⊗ ηA U † − σ⊗n k1 = 0 , (A1)
n→∞

where the partial trace is performed on the ancillary qubits A, k·k1 is the trace-norm and Q0k 6 O(n2/3 ) with k · k the
operator-norm. Here the global charge is defined by
 n−1 
X 
i
Q̄k B  I ⊗ Qk ⊗ I
 n−i−1  ⊗ IA + In ⊗ Q0 . (A2)
 k
i=0

Proof. The proof (b) implies (a) is straightforward and a consequence of kHA k being sublinear in n. We prove the reverse
implication ((a) implies (b)) by induction. The case of a single charge energy has been proven in [37]. Let us consider that there
is already a unitary U that conserves the first m − 1 charges {Qk }m−1
k=0 . Lemma 5 in [37] implies that by adding and ancillary
p
system Am of size O( n log n), there exists a new unitary Ũ that acts on the whole system plus ancillary systems and conserves
the charge Qm . At the end, the total ancillary system A is composed of the different ancillary subsystems A = ⊗q−1 k=0 Ak , where
each Ak is guaranteeing the conservation of the charge Q̄k . 

Appendix B: Extraction of a single charge

1. Setting and preliminaries. We have a Hamiltonian Q0 = H and other conserved “charges” Q1 , . . . , Qq−1 on a Hilbert
space H of finite dimension d = |H|. This assumption is used in the proof, but it is hoped that for the final result it can be lifted
to yield a similar statement for separable Hilbert spaces. We shall without loss of generality assume that the Q j are linearly
independent as matrices.
We shall be concerned with the generalised grand canonical Gibbs state (GGS) γE,v , which is the state maximising the von
Neumann entropy under the constraints hQ j i = v j ( j = 1, . . . , q − 1) and hHi ≤ E; we denote as v = (v1 , . . . , vq−1 ) the tuple of the
expectation values of the conserved quantities that are not the energy. As long as E ≤ d1 Tr H, also Tr γE = E, i.e. all constraints
are attained with equality [38, 39].
 18

As is well-known, and easily derived using Lagrange multipliers,

 q−1

1  X 
γE,v = exp −βH −
 µ j Q j  , (B1)
Z j=1

 
where Z = Tr exp −βH − q−1 j=1 µ j Q j is the partition function, and the real numbers µ j are uniquely specified by the require-
P
ments Tr γv Q j = v j . (The parameter β = µ0 would be the inverse temperature in the commuting case, and the other µ j chemical
potentials.) We shall use the above form (B1) in a formal sense even if γE,v actually is in the closure of the above exponential
family.
In contrast to the usual Gibbs state, which is interpreted as the thermal equilibrium state for a system that is in thermal contact,
i.e. can exchange energy, with a bath, the grand canonical state is the equilibrium if system and bath can exchange energy, but
not the other charges (think of particle numbers of different species), which have to be conserved on the system. At least, this is
an accepted description for a Hamiltonian commuting with the conserved charges.
We take a similar point of view when considering work extraction: Here, we have a battery system and no bath, but the
philosophy remains the same that system and battery can exchange energy (work), but the other charges have to be conserved
both on the system and the battery.
Remark The crucial property of the GGS that we shall use is that it is the unique entropy maximiser under constraints on the
Q j , namely hQ j i = v j ( j = 1, . . . , q − 1) and hHi ≤ v0 =: E, and at the same time the energy minimiser under all states with the
same or larger entropy as γE,v :

γE,v = argmax S (ρ) s.t. Tr ρQ j = v j , Tr ρH ≤ E,

= argmin Tr ρH s.t. Tr ρQ j = v j , S (ρ) ≥ S (γE,v ).

2. Passivity, complete passivity and asymptotic passivity. Our setting will be the extraction of work onto a battery system.
In principle we can also allow other systems which have to be returned in their original state, but we can ignore those for
the moment; we shall only consider thermal machines consisting of the system under consideration and a battery. All global
operations will have to conserve the quantities H and Q j , but while we assume that the system and the battery can exchange
energy quanta freely, we shall demand that charges cannot.
2a. Warm-up: Information theoretic view on passivity à la Alicki/Fannes. We start by reviewing the information theoretic
1 −βH
proof by Alicki and Fannes [26] of the complete passivity
P of the usual Gibbs states Z e , i.e. without any conserved charges.
In the presence of only a simple Hamiltonian H = i Ei |iihi|, a state ρ is called passive for H if for any unitary on the system,

Tr ρH ≤ Tr UρU † H.

From this one can easily derive that ρ has to commute with H, ρ = i pi |iihi| and the eigenvalues pi are non-increasing in the
P
energy. The motivation for this definition is that the difference Tr ρH − Tr UρU † H can be extracted as work onto a battery in
a suitable initial state [44, 45]. The state is called completely passive for H if for all N, ρ⊗N is passive for the macroscopic
extension
N−1
1 X ⊗`
H= 11 ⊗ H ⊗ 11⊗N−1−` .
N `=0

From this one sees directly that the Gibbs states Z1 e−βH , for β ≥ 0 are completely passive.
To deal with the converse, assume that ρ is not of the Gibbs form (including the limit β = ∞ corresponding to the ground
state). If ρ does not commute with H, it cannot be passive, so we may assume that [ρ, H] = 0. I.e., ρ, H and Gibbs states are
diagonalised in the same basis:
X X
H= Ei |iihi|, ρ = pi |iihi|.
i i

Let E = Tr ρH be the average energy of ρ; that ρ is not a Gibbs state is equivalent to saying that it is not the energy minimiser
among all states with entropy S (ρ):

ρ , γE0 := argmin Tr σH s.t. S (σ) ≥ S (ρ),

 19

with E0 = Tr γE0 H < E. By continuity, we have for sufficiently small η > 0 that

E0 + η = Tr γE0 +η H ≤ E − η and S (γE0 +η ) ≥ S (ρ) + η. (B2)

Also the Gibbs state is diagonal in the same basis as the previous matrices:
X
γE0 +η = qi |iihi|.
i

For integer N, we write iN = i1 . . . iN for a generic string of basis indices, and

X X
ρ⊗N = piN |iN ihiN |, γ⊗N
E0 +η = qiN |iN ihiN |,
iN iN

with piN = pi0 · · · piN−1 and qiN = qi0 · · · qiN−1 . We use Shannon’s theorem on data compression: Define, for δ > 0, the frequency-
typical sequences
(   )
1
T pN := iN :  j(iN ) − p j  ≤ δ ,
N

where j(iN ) is the number of times the index j appears among i0 , i1 , . . . , iN−1 ; TqN is defined likewise. The (elementary) properties
of the typical sequences we shall need are that for δ > 0 small enough and N large enough,

p⊗N (T pN ) ≥ 1 − ,
q⊗N (TqN ) ≥ 1 − ,
log |T pN | ≤ N S (ρ) + η/2 ,

log |TqN | ≥ N S (γE0 +η ) − η/2 ,

where p⊗N (and q⊗N ) are the N-fold product probability distributions of p (and q, resp.) on the sequences of length N, and –
importantly –,

1 X
∀iN ∈ T pN Tr |iN ihiN |H = Ei` ≥ E − η/3,
N `
1 X
∀iN ∈ TqN Tr |iN ihiN |H = Ei` ≤ E0 + η + η/3,
N `

so that in particular, T pN ∩ TqN = ∅. Since |TqN | ≥ |T pN |, we can choose a subset T0 ⊂ TqN of cardinality |T pN |, and a bijection
f : T pN → T0 ⊂ TqN . We claim that the unitary
X X
| f (iN )ihiN | + |iN ih f (iN )| + |iN ihiN |,


U :=
iN ∈T pN iN <T pN ∪T0

which swaps the linear span of T pN with that of T0 , decreases the energy of ρ⊗N . Indeed, by the above properties we have,
X
Tr ρ⊗N H ≥ piN Tr |iN ihiN |H − kHk
iN ∈T pN

≥ E − 2kHk − η/3,
X
Tr Uρ ⊗N
U H≤†
piN Tr | f (iN )ih f (iN )|H + kHk
iN ∈T pN

≤ E0 + 2kHk + 4η/3.

In total,

Tr ρ⊗N H − Tr Uρ⊗N U † H ≥ E − E0 − 4kHk − 2η,

 20

which can be made arbitrarily close to E − E0 for large N, because η > 0 and  > 0 are arbitrarily small. 
2b. General case of non-commuting observables. We now take these simple information-theoretic lessons from the standard
case, to show that the GGS are all passive with respect to work extraction, while at the same time respecting the conservation
of the charges in a suitable sense, and that they are the only passive states. The argument will appear a bit more abstract, and is
more complex, than the hands-on construction we have seen above. At the same time, its generality will bring out the essential
structures more prominently.
Encouraged by the results on reference frames [46, 47], we shall restrict the discussion to so-called implicit batteries. This
means that (a) we start off the battery in a state product with the system, (b) we consider only the effective map on the system, a
conjugation by a unitary or unital cptp map, and (c) we measure the extracted energy by the difference in expected energy before
and after the action of the map. Regarding point (c), in general this difference is an upper bound, simply by energy conservation.
That the whole amount can also be extracted, at least for the case of a unitary effected on the system, is possible in the presence
of a suitable reference frame.
We now have to discuss the restrictions imposed by conservation of the charges Q j . Clearly, the ideal would be that, for,
say, a unitary U on the system, UQ j U † = Q j for all charges j = 1, . . . , q − 1, i.e. [U, Q j ] = 0. Note that this imposes strong
restrictions on the unitaries, since it implies that U commutes with the algebra generated by the Q j . For example, in the simplest
case of a Hamiltonian and a single additional conserved quantity Q1 = Q commuting with H, assume that Q1 has no degenerate
eigenvalues: Then, U has to fix each eigenvector of H and can only attach phases to each of them. In particular, U conserves
also the energy, and so cannot extract work! If we go to N replicas of the system, U should commute with the extension of the
conserved quantity,
N−1
1 X ⊗`
Q= 11 ⊗ Q ⊗ 11⊗N−1−` ,
N `=0

and work is extracted with respect to the extended Hamiltonian H. This looks like good news, since Q has degenerate eigen-
values. However, if the eigenvalues of Q are rationally independent, the degeneracy of Q is restricted to permutations of the
N systems, thus again U acting on H ⊗N cannot change the energy of the system (it can act nontrivially only on the energy
eigenspaces), and so still cannot extract work. There are at least two objections to the reasonableness of asking for exact conser-
vation (which we shall discuss shortly), beyond the pragmatic requirement that we would like to get a fruitful theory for all sets
of charges, in particular one that makes sense also for non-commuting charges.
For the subsequent treatment we shall adopt the following notion.
Definition 23. We call a sequence of unitaries U on H ⊗N asymptotically Q-conserving if

lim Q − UQU † k = 0.
N→∞

More generally, we call a sequence of cptp maps Φ (with adjoint map Φ∗ on the observables, which are completely positive and
unital, cpup) asymptotically Q-preserving, if

lim Q − Φ∗ (Q)k = 0.
N→∞

As it turns out, this notion will give rise to a complete and satisfying, and above all: consistent, theory of passivity.
As for physical justification, why would we allow transformations that only approximately conserve the “conserved” quanti-
ties? One answer to this comes from looking at the energy, and how we overcome energy conservation by a reference frame in
the battery, which together allow us to implement any unitary on the system, while extracting any energy difference as work. As
a matter of fact, we could imagine that there are other batteries for the extraction of charge Q, with its own Q-reference system,
that would allow overcoming the charge conservation as well [48, 49]. From [43] we know that in that case, one can trade the
extraction of work (H) against that of the charges (Q). This highlights the resource character of having a reference frame, and
it suggests that a conservation law is simply the expression of lacking a reference frame [46, 47]. But then one can discuss
partial, even little, access to this resource. Having a very poor reference system will not allow extraction of the charge quantity,
and generally only actions that nearly conserve the charge observable Q. The second, related, argument which should make us
suspicious of exact charge conservation is due to stability [44]. A physically meaningful requirement would be that the thermo-
dynamics of a system with Hamiltonian H and charges Q j is similar to that of H 0 ≈ H and Q0j ≈ Q j . In our above example of
Q with rationally independent eigenvalues, there are Q0 arbitrarily close by whose eigenvalues are integer multiples of the same
quantity, and for each such approximation Q0 suddenly there are many nontrivial, energy-non-conserving unitaries on N copies
of the system. However, depending on the precise rational approximation of the eigenvalues of Q, the combinatorics of what is
allowed for each N differs subtly, and stability requires that these details should have a small effect on the thermodynamics of
the system.
 21

Definition 23 addresses these two concerns, giving a notion that is both robust under small changes of the charge observables,
and consistent with the idea that an asymptotically useless reference frame is present, or that charges are exchanged with the
battery to a small degree.
2c. Two technical propositions containing basically the whole story. First, we show that a naive notion of thermal operations
associated to work extraction cannot get any work from a GGS, at least asymptotically. Namely, with an implicit battery, the
operation results in a well-defined unital cptp map on the system which is shown to never decrease the energy of the system.
Definition 24. For any state ρ, with E = Tr ρH and v j = Tr ρQ j , let

E0 := min Tr σH s.t. S (σ) ≥ S (ρ), Tr σQ j = v j ,

and γ = γE0 ,v the corresponding GGS. Note that E0 ≤ E.

Then, for any sequence of of entropy-nondecreasing maps Φ on the system such that Tr Φ(ρ⊗N )Q j → v j , it follows that

lim inf Tr Φ(ρ⊗N )H ≥ E0 .

N→∞

In particular, in the case E0 = E, or equivalently ρ = γE,v , it follows that GGS’s are asymptotically passive under work
extraction by asymptotically Q j -conserving maps: For any sequence of entropy-nondecreasing maps Φ on the system such that
Tr Φ(γ⊗N
E,v )Q j → v j , it follows that

lim inf Tr Φ(γ⊗N

E,v )H ≥ E.
N→∞

 cptp map Φ on the N-copy system as described, and ρ = Φ(ρ ), and assume that N is large enough so
0 ⊗N
 Consider any
Proof.
that Tr ρ Q j − v j ≤ η, for a given η > 0. On the other hand, since Φ can only increase the entropy, S (ρ0 ) ≥ N S (ρ).
 0 
Thus, we get
 
Tr Φ(ρ⊗N )H ≥ min Tr ωH s.t. Tr ωQ j − v j  ≤ η, S (ω) ≥ N S (γ)
 
= min Tr σH s.t. Tr σQ − v  ≤ η, S (σ) ≥ S (ρ).
j j

The identity in the second line comes from the fact that the minimum in the first line is attained on ω = σ⊗N .
Now, from the maximum entropy characterisation of GGS’s, we know that the minimum in the second line is attained on a
GGS γE 0 ,v0 , where |v0j − v j | ≤ η, and for which min Tr γE 0 ,v0 H = E 0 .
Since for all N, S (γE 0 ,v0 ) ≥ S (ρ), we get also lim inf N→∞ E 0 ≥ E0 , since u0 → u, and so

lim inf Tr Φ(ρ⊗N )H ≥ lim inf E 0 ≥ E0 .

N→∞ N→∞

In the case of ρ = γE,u , E = E0 , showing that the GGS’s are asymptotically completely passive. 

Remark For the above proof to work, we do not have to require that the map Φ is globally entropy non-decreasing, nor that it
is asymptotically Q j -conserving. It is only required that for the particular state ρ on which we are working,

1
lim inf S (Φ(ρ⊗N ) ≥ S (ρ) and lim Tr Φ(ρ⊗N )Q j = v j
N→∞ N N→∞

hold true. We say that Φ is asymptotically entropy-non-decreasing and asymptotically Q j -conserving relative to ρ. 
Second, if a state is not a GGS, then there exists an operation, in fact a unitary, which extracts work from it (hence it cannot be
passive), in fact asymptotically as much as it can be, consistent with the charges and the entropy associated with the first state.
This unitary can be implemented using a reference frame for the energy, and using asymptotically vanishing frameness for the
other charges. We will need the concept of approximate microcanonical subspace [50]:
Definition 25. An approximate microcanonical (a.m.c.) subspace, or more precisely a (, η, η0 , δ, δ0 )-approximate microcanon-
ical subspace, M of H ⊗N , with projector P, for observables Q j and values v j = hQ j i is one that consists, in a certain precise
sense, of exactly the states with “very sharp” values of all the Q j . Mathematically, the following has to hold:

1. Every state ω with support contained in M satisfies Tr ωΠηj ≥ 1 − δ for all j.

2. Conversely, every state ω on H ⊗N such that Tr Πηj ω ≥ 1 − δ0 for all j, satisfies Tr ωP ≥ 1 − .

0
 22

Here, Πηj := v j − ηΣ(Q j ) ≤ Q j ≤ v j + ηΣ(Q j ) is the spectral projector of Q j of values close to v j , and Σ(Q) = λmax (Q) − λmin (Q)

is the spectral diameter of the Hermitian Q, i.e. the diameter of the smallest disc covering the spectrum of Q.
In [50, Thm. 3] it was shown that for every  > (c + 1)δ0 > 0, δ > 0 and η > η0 > 0, and for all sufficiently large N, there exists
a nontrivial (, η, η0 , δ, δ0 )-a.m.c. subspace.
Definition 26. If the state ρ is not a GGS, which is to say Tr ρH = E and Tr ρQ j = v j , but ρ , γE,v , then work can be extracted
from ρ⊗N while asymptotically conserving the charges, for N  1.
To be precise, if E0 is such that S (ρ) = S (γE0 ,v ), then there are asymptotically Q j -conserving unitaries U on H ⊗N , i.e.
kQ j − UQ j U † k → 0, as N → ∞, that extract work

lim Tr ρ⊗N H − Tr Uρ⊗N U † H = ∆E = E − E0 > 0.

N→∞

Proof. According to [50, Thm. 3], there exists an a.m.c. subspace for Q1 , . . . , Qq−1 with values v = (v1 , . . . , vq−1 ); note that we
exclude the energy in its definition. More precisely, for sufficiently large N and subject to the constraints  > (c + 1)δ0 > 0, δ > 0
and η > η0 > 0, we can find an (, η, η0 , δ, δ0 )-a.m.c. subspace M of H ⊗N ; denote its projector by P.
By looking at the proof of Theorem 4 in [50], we see that we can then also add the energy and obtain a.m.c. subspaces ME
M, which now concern the observables Q j (with values v j ) and also H (with value E). Assume furthermore that η > 0 is small
enough such that S (ρ) + η ≤ S (γE0 +η,v ) − η. By construction, for η < 13 ∆E the subspaces for E and E0 + η are orthogonal,
ME ⊥ ME0 +η ; denote their projectors by PE and PE0 +η , respectively.
Now, since Tr ρ⊗N H = E and Tr ρ⊗N Q j = v j for j = 1, . . . , q − 1, and indeed Tr ρH = E and Tr ρQ j = v j , we have by
Hoeffding’s inequality that for sufficiently large N Tr ρ⊗N ΠηE ≥ 1 − δ0 and Tr ρ⊗N Πηj ≥ 1 − δ0 , hence Tr ρ⊗N PE ≥ 1 − .
0 0

This means that we can find a typical subspace T ⊂ ME of ρ⊗N , its projector denoted Pρ , such that Tr ρ⊗N Pρ ≥ 1 − 2 and
η0 η0
log dim T ≤ N S (ρ) + η . Likewise, Tr γ⊗N E0 +η,v ΠE0 +η ≥ 1 − δ and Tr γE0 +η,v Π j ≥ 1 − δ , so Tr γE0 +η,v PE0 +η ≥ 1 − . I.e., ME0 +η
0 ⊗N 0 ⊗N

E0 +η,v , and so by Schumacher’s theorem, log dim ME0 ≥ N S (γE0 +η,v ) − η ≥ log dim T . This
is a high-probability subspace for γ⊗N

means that we can find a subspace T0 of ME0 +η of dimension equal to that of T , with projector P0
Now, we are ready to define the unitary U: On M, U acts by swapping T and T0 , and as the identity on M⊥ , i.e.

U = PUP + P⊥ , UPρ U † = P0 , UP0 U † = Pρ .

We claim that U approximately conserves the Q j (which is intuitive as the unitary acts nontrivially only inside the space M of
sharply defined values of the Q j ), and that it extracts work from ρ⊗N . Starting with the latter, we have Tr ρ⊗N Pρ ≥ 1 − 2 and
√
hence by the gentle measurement lemma [51], kρ⊗N − Pρ ρ⊗N Pρ k1 ≤ 8. Thus, using the triangle inequality,
√
Tr Uρ⊗N U † H ≤ Tr UPρ ρ⊗N Pρ U † H + 8kHk
√
= Tr ωH + 8kHk
√
= Tr ω(ΠηE0 +η + 11 − ΠηE0 +η )H + 8kHk
 √ 
≤ E0 + 2η + δ + 8 kHk,

where ω = UPρ ρ⊗N Pρ U † is the (sub-normalised) state after restricting to Mρ and after the unitary U, which is supported on
M0 ⊂ ME0 +η . Then, in the third and fourth line we use the properties of the a.m.c. subspace to bound the energy expectation.
Since η,  and δ are arbitrarily small, we can get arbitrarily close to E0 for sufficiently large N.
Regarding asymptotic Q j -conservation, consider the following:

Q j − PQ j P − P⊥ Q j P⊥ = PQ j P⊥ + P⊥ Q j P

≤ 2 P⊥ Q j P

= 2 P⊥ (Q j − v j )P

≤ 2 (Q j − v j )P
 
= 2 max (Q j − v j )|ϕi2
|ϕi∈M

= 2 max (Q j − v j )(Πηj + 11 − Πηj )|ϕi2

 
|ϕi∈M
" #
η η
   
≤ 2 max (Q j − v j )Π j |ϕi 2 + (Q j − v j )(11 − Π j )|ϕi 2 .
   
|ϕi∈M
 23

The first norm in the last line is bounded by k(Q j − v j )Πηj k ≤ ηΣ(Q j ). By the property of the a.m.c. subspace, Tr ϕΠηj ≥ 1 − δ,
√ √
i.e. (11 − Πηj )|ϕi has length ≤ δ, so the second norm in the last line is bounded by 2 δΣ(Q j ). All taken together, we get
√
Q j − PQ j P − P⊥ Q j P⊥ ≤ (2η + 4 δ)Σ(Q j ).

As U commutes with P, we get the same bound for UQ j U † :

√
UQ j U † − PUQ j U † P − P⊥ UQ j U † P⊥ ≤ (2η + 4 δ)Σ(Q j ).

By these last two bounds and triangle inequality, we get

√
Q j − UQ j U † ≤ (4η + 8 δ)Σ(Q j ) + P(Q j − UQ j U † )P , (B3)

noting that U acts as the identity on M⊥ . The latter norm can be bounded, using the a.m.c. subspace property as before, by

P(Q j − UQ j U † )P ≤ P(Q j − v j )P + PU(Q j − v j )U † Pk

= 2 (Q j − v j )P
√
≤ (2η + 4 δ)Σ(Q j ).

Thus, from eq. (B3) and this last bound we find

√
Q j − UQ j U † ≤ (6η + 12 δ)Σ(Q j ),

which can be made arbitrarily close to 0 for sufficiently large N. 

Remark The energy difference ∆E = E − E0 equals the difference in free energy between ρ and the isentropic GGS:

D(ρkγE0 ,v ) = −S (ρ) − Tr ρ log γE0 ,v

= −S (γE0 ,v ) − Tr ρ log γE0 ,v
  
  X 
= Tr (γE0 ,v − ρ)−βH −
  µ j Q j 
j

= β Tr ρH − Tr γE0 ,v H

= β(E − E0 ),

since the expectations hQ j i = v j are identical for ρ and γE0 ,v , while the energy expectations are E and E0 , respectively.
This is the maximum we can extract by asymptotically Q j -conserving operations, by Proposition 24. 

3. Main result. We have essentially done the work to show the main result of this write-up.
Theorem 22. The asymptotically completely passive states for work extraction under asymptotically conserved charges
Q1 , . . . , Qq−1 are precisely the generalised grand canonical Gibbs states
 q−1

1  X 
γE,v = exp −βH − µ j Q j  ,
Z j=0

or limit points in the closure of the exponential family, which correspond to diverging inverse temperature β and/or chemical
potentials µ j .
Furthermore, the tensor product of several states is asymptotically passive if and only if they all are individually asymptotically
passive and they share the same intensive parameters (β, µ).

Proof. By Proposition 24, from asymptotically many copies of a GGS γE,v , we cannot extract work with entropy-nondecreasing,
asymptotically Q j -conserving maps; as a matter of fact, we cannot even lower the average energy of the system.
For a general state ρ on the other hand, which has average energy E = Tr ρH, we can asymptotically extract ∆E = E − E0 per
copy of the state, with unitaries that are asymptotically Q j -conserving, and with

E0 = min Tr σH s.t. S (σ) ≥ S (ρ).

 24

The work is positive, ∆E > 0, if and only if ρ is not a GGS.

j (k = 1, . . . , K), on the
As for the composition of systems, given by Hilbert spaces Hk with Hamiltonians H (k) and charges Q(k)
NK
combined system H = k=1 Hk we have the Hamiltonian and charges given by

K
X K
X
H= 11⊗k−1 ⊗ H (k) ⊗ 11⊗K−k , Qj = 11⊗k−1 ⊗ Q j ⊗ 11⊗K−k .
(k)

k=1 k=1

But we know already that the passive states on the composite system are precisely those of the form
 q−1
 K
1  X  O
γE,v = exp −βH − µ j Q j  = γ(k)
E,v ,
Z j=1 k=1

with
 q−1

1  X 
γ(k)
E,v =  (k)
exp −βH − µ j Q j  ,
(k) 
Zk j=1

because the operators on system k commute with all operators on system k0 , k. 

By the remark after Proposition 26, we know more generally exactly how much work can be extracted from many copies of a
general state:
Theorem 27. For any ρ with energy E = Tr ρH and charges v j = Tr ρQ j , the maximum extractable work by asymptotically
Q j -conserving unitaries is

1
∆E = E − E0 = D(ρkγE0 ,v ),
β
where β is the inverse temperature of γE0 ,v , and

E0 = min Tr σH s.t. Tr σQ j = v j , S (σ) ≥ S (ρ).

This amount cannot be increased by going to more general operations, as long as they are asymptotically entropy-non-
decreasing and Q j -conserving relative to ρ.