Sei sulla pagina 1di 17

Elementary Thermal Operations and the Universality of Thermodynamic Constraints

Matteo Lostaglio

arXiv:1607.00394v1 [quant-ph] 1 Jul 2016

Department of Physics, Imperial College London, London SW7 2AZ, United Kingdom
To what extent is the resource theory approach to thermodynamics relevant for realistic experimental scenarios? We deconstruct this framework, showing that each transition among incoherent
states allowed in the theory can be obtained by sequentially applying elementary thermal operations.
These only couple the bath to single transitions in the system and satisfy detailed balance. We show
that most of them can be realised within one of the simplest thermodynamic models, a resonant
Jaynes-Cummings interaction with a single mode bosonic bath, and discuss the link to thermalisation
models. No external control of the Hamiltonian or ancillae are needed. We also present an extension
of Birkhoffs theorem for bistochastic matrices to thermodynamics, and discuss its consequences for
the problem of simulating arbitrary thermal processes. Finally, we show that the constraints found
within thermal operations for transitions among incoherent states are to a large extent universal, i.e.
they apply to a broad class consisting of very different thermodynamic frameworks.

Physicists agree that thermodynamics is one of the


pillars of physics, but disagree on its foundations and
scope. The recent surge in interest in the quantum
information/single-shot approach to thermodynamics
[123] is no exception and has attracted some criticism.
The main aim of this paper is to present a universal
set of thermodynamic gates, i.e., a simple set of transformations that allows one to perform any thermal processing achievable by thermal operations among incoherent states. This set is given by elementary processes
involving two-level subsystems and satisfying detailed
balance. Going beyond a previous result [24], we do not
require any external control of the Hamiltonian or ancillae.
This has multiple consequences. First, we show that
the gates in the universal set can be approximated
within the simplest Jaynes-Cummings model in the
rotating-wave approximation (RWA), paving the way
for possible experimental realisations. On the other
hand, we address the connection to fluctuation theorems. In fact, we show that even though every thermodynamic transition can be realised by a process falling
within the assumptions of Crooks theorem [32], there
are also indecomposable processes for which the theorem
cannot be applied. On the theoretical side, this leads to
a generalisation of a famous theorem by Birkhoff on bistochastic maps to thermodynamics, with implications
for the simulation of a generic thermal process and for
the geometry of thermalisation.
Another aim of this work is to clarify the scope of the
resource theory results. In this respect, we find that,
at least for transformations among incoherent states,
the resource theory constraints actually apply to a wide
range of thermodynamic models. Moreover, we highlight connections between the resource theory language
and other approaches, such as collision models and the
master equation formalism, with the aim of developing
a common language. As an application, we discuss constraints that can be deduced from resource theory for
thermalisation models.

Criticism of Thermal Operations

The resource theory approach to thermodynamics is


based on the concept of Thermal Operations (TOs) [1, 3].
These are energy-preserving interactions of a system
with a thermal bath. In particular one assumes that
given a system in an initial state and with Hamiltonian HS , this can be coupled to a bath at temperature TB
with arbitrary Hamiltonian HB by any unitary U satisfying [U, HS + HB ] = 0. So any
Operation
 Thermal
 T can
HB

be written as T ()
=
Tr
U
(

e
/Z
)U
. Here
B
B

ZB = Tr eHB , = (kTB )1 ; HS and HB should
be understood as the non-interacting Hamiltonians of
system and bath, and U as the overall effect of an interaction Hamiltonian which is turned on at some time
ti and off at some later time tf . The resource theory approach to thermodynamics assumes that this is the set of
transformations that can be implemented for free, without any external resource being consumed. In particular,
no classical work source is involved.1
This framework has been criticised in at least two,
somehow opposing, ways, which we now present. TOs
clearly contain a large set of transformations that can
be implemented only through a complete control of the
system-bath interaction U . This in general requires a
fine control of the microscopic degrees of freedom of the
bath, on top of the assumption that any bath Hamiltonian HB can be engineered; hence, achieving the full set
of TOs seems completely unrealistic from an experimental point of view [24].
One may respond that we are interested in fundamental limitations, so that by allowing such a large set
of transformations we are able to embrace a large variety of experimental situations; which is to say, the constraints derived will still provide necessary conditions,

A battery may be needed to turn on and off the couplings, or a field


may be used to extract work. In the resource theory approach these
can be used, but they must be explicitly included in the description
as quantum-mechanical systems.

2
however weak, to practically relevant scenarios. But
conversely TOs appear, in other ways, overly restrictive.
In particular, the constraint of energy conservation may
seem a serious limitation, as typically it will be violated
as soon as interactions are turned on. A deeper analysis of the connections to experimental implementations
is then crucial.
In this paper we aim to make a step towards overcoming the above criticism and clarifying how to connect
the resource theory approach to more experimentallyfriendly frameworks. Given aPfinite-dimensional system with Hamiltonian HS = i ~i |iihi|, we will focus on transitions among incoherent states and i.e.
[, HS ] = [, HS ] = 0. Let p and q be the eigenvalues of
and , representing the occupations of each energy level.
Denote by T the set of thermal stochastic processes, i.e.,
all stochastic matrices T with Ti|j = hi| T (|jihj|) |ii for
some TO T (we say that T is induced by T ). These represent the random energy exchanges between system and
bath and do not carry a description of how superpositions among different energy levels evolve. Since, by assumption, no such superpositions are present in and
, a TO T with T () = exists if and only if there exists
a thermal process T with T p = q (see Appendix A.1).

Elementary Thermal Operations

Among all stochastic processes describing the random heat exchanges between system and bath, a subset stands out as particularly simple and physically motivated. These are all stochastic processes E satisfying
two conditions:
1. Only two energy levels of the system are involved.
2. Detailed balance, Ei|j = e~(i j ) Ej|i , is satisfied.
This set of stochastic matrices will be called elementary
detailed balanced processes (EDPs) and denoted by E . One
can check that E T , i.e., all EDPs can be induced by
means of TOs (Appendix B.2). We then call the set of
TOs generating such simple stochastic processes Elementary Thermal Operations, or ETOs. Any subset of ETOs
that generates all EDPs will be called complete.
The heat exchanges described by E are much simpler
than the generic process in T . In fact, each EDP is defined by the two levels it acts on plus a single parameter.
Of course, mathematical simplicity is not by itself sufficient. Two natural questions arise: can we find a complete set of ETOs that can be realised within a simple
physical model? What transitions between incoherent
states can be achieved with a complete set?

Physical realisations
The Jaynes-Cummings model approximates a complete set of ETOs

One of the core models describing the interaction of


matter and radiation is the Jaynes-Cummings model,
and so we start investigating the previous questions
within it. For simplicity, consider a system whose
relevant transition frequencies are all well separated
from each other. Assume one is allowed to couple every transition of the system to a single-mode
bosonic bath in a thermal state through a resonant
Jaynes-Cummings coupling in RWA approximation,
HJC = g(+ a + a ), for a time t. Here a , a
are creation/annihilation operators on the bath and
excite/de-excite a transition := 1 0 > 0 on the system. This model gives TOs. A standard calculation provides the induced stochastic processes (Appendix C.1):
J1|0 (s) =

X
n=1

sin2 (s n)gn ,

J0|1 (s) =

sin2 (s n)gn1 ,

n=1

where gn = e~n /ZB is the occupation of the n-th energy level of the bath and s = gt/~.2 Also, J0|0 = 1J1|0 ,
J1|1 = 1 J0|1 . One has J1|0 /J0|1 = e~ , i.e. J is an
EDP. So the model generates ETOs defined by a single
control, s. The question then arises, to what extent can
all E E be achieved?
The independent parameter for a generic EDP on
two levels can be chosen to be the de-exciting probability E0|1 . If, for every E0|1 [0, 1], there exists an
s such that J0|1 (s) = E0|1 , then all EDPs can be
achieved. In Appendix C.1 we show that in the JaynesCummings model one can reach all de-exciting probabilities in the interval [0, Jmax (~)], where Jmax (~)
lies in the region presented in Figure 1. In particular
when (the zero temperature limit) Jmax 1,
so the Jaynes-Cummings model gives a complete set of
ETOs. But also the finite temperature scaling is rather
favourable: at room temperature TB = 300K and frequencies of 1013 Hz or above, and at millikelvin temperatures with frequencies of around 108 Hz, one has
Jmax (~) > 0.98.
We conclude that, for a wide interval of reasonable parameters, almost all EDPs can be realised within a simple subset of TOs corresponding to a physically reasonable model, with no need to implement general unitaries
or engineering thermal states with arbitrary Hamiltonian. Of course, one still requires considerable frequency control as we must be able to separately couple
to every transition frequency of the system. This is feasible only if we are interested in the thermodynamics of
small enough systems.

For any s, one can take g small enough and t large enough to avoid
the short times/strong couplings regime where the model breaks
down.

3
Jmax
1.00

Proof. For E to be a valid stochastic


matrix, L must satP
isfy Li|j 0 for i 6= j and i Li|j = 0 for all j [31].
Detailed balance is equivalent to Li|j = e~(i j ) Lj|i .
A direct calculation then gives Eq. (1).

0.98
0.96

Elementary Thermal Operations are universal for


incoherent processes

0.94
0.92

Sequential application of ETOs is universal


0.90
0.6

0.8

1.0

1.2

1.4

FIG. 1. Realising EDPs in the simplest Jaynes-Cummings model:


to achieve every elementary detailed balanced process (EDP),
one needs to be able to obtain all de-exciting probabilities
Ei|j [0, 1], = j i > 0. In Appendix B.2 we
show that one can realise every transition probability within
[0, Jmax (~)], where Jmax (~) lies somewhere within the
shaded area. Hence, far from the high temperature limit, almost all EDPs can be achieved within the Jaynes-Cummings
model.
Thermalisation models are Markovian ETO

To bridge the gap between resource theories and applications, we investigate the connection to thermalisation models and the master equation formalism, widely
used in quantum thermodynamics [25].
Consider a collision model of thermalisation of a system in the presence of a large thermal bath [26]. A thermal particle from the bath approaches, interacts with a
two-level subsystem and scatters away; the interaction
conserves energy and the particle is lost in the bath [27].
Each collision is then an ETO. Reasoning as in [27],
one finds that the combined effect of many identical
weak interactions on each two-level subsystem, initially
described by a population p(0), leads to an exponential
relaxation to the thermal distribution g (Appendix C.2):
p(t) = et/ p(0) + N (1 et/ )g,

(1)

where > 0 and N is the normalisation of p(0).


Alternatively, from the master equation point of view,
we can consider Davies maps, which are routinely used
as a simple model of a system weakly interacting with a
large thermal bath [28], or in the low-density limit [29].
On two-level systems, one can check that they induce
the evolution of Eq. (1).
Now, what is the connection with the resource theory
approach? This can be clarified as follows. Markovian
ETOs are defined here as the subset of ETOs whose induced processes E E are embeddable stochastic matrices, i.e. there is some generator L and t 0 such that
E = eLt [30]. Then
Proposition 1. Markovian ETOs induce the same evolution (1) of collision models and Davies maps.

The Jaynes-Cummings model approximates a complete set of ETOs. But what thermodynamic transformations can we achieve with a complete set of ETOs?
We show that we can realise every incoherent transition
possible under Thermal Operations:
Theorem 2. Let [, HS ] = 0. Then all transitions
that can be achieved through TOs can be obtained by a finite sequence of ETOs drawn from a complete set and energypreserving unitaries on the system alone.
The proof can be found in Appendix D.1. If p and
q are the eigenvalues of and , the proof consists in
showing that the thermal process T T such that
T p = q can always be chosen to be a sequence of EDPs.
From the previous analysis of the Jaynes-Cummings
model we conclude that, by resonant interactions between two level subspaces and single-mode bosonic
baths and by energy-preserving unitaries on the system
alone, we can realise every incoherent thermal processing that could be achieved by arbitrary TOs.

Crooks theorem and indecomposable thermal processes

Theorem 2 ensures that, if there exists a thermal process T T such that T p = q, then we can achieve the
same final state through a sequence of steps, each satisfying the detailed balance condition. In other words, we
can always choose the thermodynamic process to satisfy Crooks theorem [32]. More subtly, however, this
does not imply that all thermal processes T T can
be decomposed as a sequence of EDPs (see Fig. 2). An
explicit example of such indecomposable process can be
constructed for trivial Hamiltonians, see [33], Chapter 2.

Thermal Birkhoffs theorem and simulation of thermal processes

The existence of indecomposable maps implies that


not all thermal processes can be simulated as a sequence
of EDPs. What is needed then to simulate generic thermal processes? The answer to this problem for bistochastic processes is a classic theorem by Birkhoff: every bistochastic matrix can be decomposed as a convex
mixture of permutations [33]. Is there a thermodynamic
analogue?

FIG. 2. Indecomposable thermal processes: If there exists a thermal


process T p = q, then there exists also T 0 T that achieves the
same thermodynamic transition and is composed of elementary steps satisfying Crooks theorem. However, T may not be
decomposable in this way.

Recall that every permutation is made up of a sequence of transpositions. Thermodynamically, however, transpositions are not allowed processes due to
the unavoidable dissipation of heat in the bath. Given
any two energy levels 0 and 1, with 1 0 = > 0,
we can construct a process E moving all the population
from the upper to the lower level, E0|1 = 1; but detailed
balance limits the probability of the inverse process to
E1|0 = e~ . Stochastic processes involving only two
levels and with the previous transition probabilities, as
well as the identity, are the thermodynamic analogue
of transpositions. As such, they will be called thermotranspositions. By the same logic, sequences of thermotranspositions will be called thermo-permutations. Similarly to permutations, these constitute a finite set. We
then have the following extension of Birkhoffs theorem:
Theorem 3. T coincides with the set of T that can be written
as a convex combination of thermo-permutations {P i }:
X
T =
i P i ,
(2)
i

where i 0 and

FIG. 3. Geometric characterisation of thermal cone: the set Tp


of states accessible from p through thermal processes (orange cone) is the convex combination of those that can be obtained by means of a finite set of thermo-permutations P i (red
dots at the bottom and p itself, P 0 being the trivial thermopermutation).

As an application, we show how this makes the proof


of the core result [5] in the resource theory of thermodynamics straightforward (Appendix D.3). Moreover, we discuss consequences for the optimisation of
coherence preservation under Thermal Operations (Appendix D.4).
Geometric characterisation of the thermal cones

To characterise the set of states accessible through


Thermal Operations from an incoherent with spectrum p, the central non-trivial problem is to obtain Tp ,
the thermal cone of p. This is defined as Tp := {T p :
T T }, i.e. the set of final populations q that can be
achieved from p by means of thermal processes. The
thermal cone has the following geometric characterisation (see Fig. 3):
Corollary 6. Tp is the convex hull of the states that can be
reached from p by means of thermo-permutations.

i = 1.

For the proof, see Appendix D.2. In terms of ETOs,


Corollary 4. Every thermal process T T can be simulated
by finite sequences of ETOs drawn from a complete set and
classical randomness.
As already discussed, a complete set of ETOs can be
approximated within the Jaynes-Cummings model. So
this simple physical model is not only sufficient to (approximately) obtain all final states permitted by thermal
processes; more strongly it also allows, when coupled to
classical randomness, to (approximately) simulate arbitrary thermal processes.
An intermediate lemma in proving the previous theorem shows that the set of thermal processes T coincides
with the set G of stochastic maps that preserve the Gibbs
distribution g:
Lemma 5. G = T .

Comparison to previous work

A comparison with the recent work [24] is useful (also


see the related work [4]). In Ref [24] the following operations are allowed: (1) Partially thermalise any two
levels of the system, as in Eq. (1); (2) Externally control
in an arbitrary manner the energy levels of the system
Hamiltonian, while paying the required work; (3) Bring
in an extra qubit system in a thermal state with arbitrary
Hamiltonian; (4) Apply steps (1) and (2) on system plus
ancilla.
One can readily check from the previous discussion that Partial Level Thermalisations coincide with
the thermal processes induced by Markovian ETOs;
from Eq. (1) this implies a maximum probability of deexciting a transition equal to (1 + e~ )1 . However
within the Jaynes-Cummings model described above

5
this limit is overcome for a large set of parameters (in
Appendix F we show this for ~ (0.1, 6.4)). The advantage of [24] that only requires Markovian ETO
is offset by the need for arbitrary external control of
each energy level in the system Hamiltonian, as well as
the need to generate and control ancillary qubit thermal
states with arbitrary Hamiltonians. Moreover, both here
and in [24] a high degree of control is assumed anyway,
because one must be able to selectively couple any transition in the system to the environment. When such control is present, the greater freedom of ETOs as compared
to Markovian ETOs seems justified; in fact, it is achieved
in the simplest model for such controlled situations a
Jaynes-Cummings interaction in the RWA approximation.

FIG. 4. Thermodynamic universality and thermal processes:


thermo-majorisation, the constraint characterising achievable
transitions with thermal processes T , also characterises all
frameworks F containing elementary thermal processes and
not allowing for a perpetuum mobile (Corollary 8). In fact, the
largest of these sets coincides with T (Theorem 7).

Equivalent thermodynamic frameworks


Are Thermal Operations too restrictive?

Equivalence of thermodynamic frameworks

When we perform a Thermal Operation T on a system, its population undergoes a thermal process T T .
More generally, a classical framework can be defined as
a set of stochastic maps F that can be induced on the energy occupations without paying any work. We proved
that E T is already sufficient to achieve all the final
states that would be possible using the full set T . Conversely, would it be possible to allow for a wider thermodynamic framework in which a set of maps F T
can be induced at no cost? A negative answer is given
by the following result:

The discussion above naturally leads to the concept


of equivalent thermodynamic frameworks. Two classical
frameworks, each represented by a set of stochastic matrices F and F 0 , are equivalent if they induce the same
transitions; i.e. for any p and q there exists F F such
that F p = q if and only if there exists F 0 F 0 such that
F 0 p = q.
A priori the specific choice of F may significantly
modify what transformations are possible within the
thermodynamic theory. If this was the case, the constraints found in the resource theory approach would
have a narrow scope. Contrary to this, we find as a
corollary of Theorems 2 and 7 and that a large class of
classical thermodynamic framework are equivalent:

Theorem 7. Let F E be a classical thermodynamic framework that forbids work extraction from a single heat bath.
Then F T .
This shows that the stochastic maps induced by TOs
are hardly a too restrictive set. In fact, they lie precisely
at the boundary of processes that would generate a perpetuum mobile.
This may seem puzzling in the light of the criticism
presented in the introduction. The reason is in part a
misunderstanding about the role of TOs. These are often compared to frameworks in which a classical field
can be applied to the system as it is often the case
in thermodynamic protocols. Compared to such frameworks, TOs seem unduly limited. The crucial point is,
however, that TOs claim to provide a framework for a
careful accounting of thermodynamic resources, not to
set the limits of what can be achieved in the lab. For
example, if a resource state is available, representing
a large enough coherent field, every quantum operation
Q can be approximated on a system using TOs T via
Q() := Tr2 [T ( )]. The caveat is that, because
cannot be prepared freely within TOs, an appropriate
accounting of its deterioration is necessary [11, 34]. This
includes, but is not limited to, an analysis of the energy
flows to and from .

Corollary 8. All classical thermodynamic frameworks


F E that forbid work extraction from a single heat bath
are equivalent. In particular, there exists F F such that
F p = q if and only if p T q.
For a definition of the thermo-majorisation constraint,
T , see Appendix A.2. Corollary 8 says that even if two
physicists use very different, but consistent, thermodynamic frameworks, the resource theory results will apply to them both as far as they agree on allowing EDPs
(see Fig. 4). Theorem 8 is also a result on the universality
of thermo-majorisation within a large class of thermodynamic models, for transitions among classical states.
Somehow inverting the logic of resource theories, it suggests a focus on the analysis of classes of universal constraints, instead of the detailed definition of the framework.
Conclusions

We presented a direct link between resource theory results and experimental implementations, showing that

6
to induce any thermal process one only needs a universal set of thermodynamic gates, which in turn can be
well approximated within a simple Jaynes-Cummings
model. Thermodynamic equivalence, moreover, enlarged the scope and applicability of the results. In Appendix D.1 we discuss how to explicitly find the set of
transformations of Theorem 2 and in Appendix G we
discuss extensions to the fully quantum case.
However, we wish to conclude by showing that the
framework is flexible enough to provide results beyond
the resource theory domain. Consider for example thermalisation models, which describe situations in which
one has little or no control on the system. Drawing on
the idea of analysing classes of constraints, can we find
a stronger requirement than thermo-majorisation, relevant for these situations? In this spirit, we introduce
the following definition: we say that q is a thermalisation of p if and only if p T q and the -ordering of p
coincides with that of q (recall that a sequence p is said
to be -ordered if for all elements pi /gi pi+1 /gi+1 ).
This constraint is immediately seen to be relevant for
thermalisation models. Let p be the energy occupations
of the initial, incoherent, state. Assume thermalisation
can be described through a sequence of collisions as explained before. Then the final occupation q is given by a
sequence of thermal processes on p induced by Markovian ETOs. Hence, p T q. Moreover each of these
processes, being a Partial Level Thermalisation, cannot
change the -ordering [24]. So q is a thermalisation of
p. The converse is also true: if q is a thermalisation
of p, then there exists a sequence of Markovian ETOs
transforming p into q (this is a simple corollary of Theorem 12 in [24]). In other words, the constraint that q is
a thermalisation of p is the strongest statement we can
make, unless further information about the thermalisation model is provided.
A final comment on the applicability of the results of
the present investigation beyond small-scale thermodynamics should be made. Relevant systems for quantum
thermodynamics may be mesoscopic and any control
of the microscopic degrees of freedom of the system unrealistic. However, stochastic processes satisfying detailed balance can be realised at very different scales
(see, e.g., [38]). In fact, thermo-majorisation has already
been applied in very different situations from the thermodynamics of single quantum systems, such as diffusion of a diluted, ideal solution and chemical reactions
of ideal gas mixtures [39]. Some thermodynamic constraints may then be universal not only in the sense of
being applicable within a wide range of models, but also
throughout a wide range of scales. It is after all universality that makes thermodynamics so useful and its
study so appealing.
Acknowledgments. I would like to warmly thank
K. Korzekwa for insightful discussions and for suggesting that a result in the spirit of Eq. (2) may hold. I
also thank D. Jennings and T. Rudolph for their contin-

uous support. I thank the people above and A. Milne,

R. Uzdin, M. Muller,
A. Levy for useful comments on a
previous draft. Finally, I would like to thank the abovementioned people and N. Yunger Halpern, M. Plenio,
S. F. Huelga, M. Christandl, R. Kosloff, P. Salamon for
discussions at various stages of my PhD on the merits
and limitations of Thermal Operations. I am supported
in part by EPSRC and COST Action MP1209.

7
APPENDIX

About notation: We will denote probabilities with Roman bold letters x, y,..., leaving the Greek letters , , ...
to denote quantum states, described by trace one, nonnegative operators.
A.
1.

Preliminary facts

Proof. Unitaries US on the system satisfying

[US , HS ] = 0 are TOs. Let US,1 () = US,1 ()US,1


and

US,2 () = US,2 ()US,2 , where US,1 , US,2 are unitaries


diagonalising the energy blocks of and , respectively.
Then if T T is such that T p = q, let T be a TO
that acts as T in the basis of the eigenstates of . Then

T = US,2
T US,1 is a TO and satisfies T () = . The
converse also follows easily.
So we can focus on the existence of a thermal process
between two arbitrary probability distributions. This
leads to the following considerations.

Thermal Operations, thermal processes T

Let HS be the Hamiltonian of the system. Then


Definition 1 (Thermal Operations (TO) [1, 3]). Thermal
Operations in the presence of an environment at temperature TB are the set of quantum maps T that can be
written as
 
 
eHB
T () = TrB U
U ,
(3)
Tr [eHB ]
where
1. HB is an arbitrary bath Hamiltonian,
2. U is an arbitrary energy-preserving unitary, i.e.
[U, HS + HB ] = 0,
3. TrB represents a partial trace over the bath degrees
of freedom,
4. = (kTB )1 , with k Boltzmanns constant.
T is trace-preserving, Tr [T ()] = Tr [], and maps the
energy eigenbasis into itself. So T induces a stochastic
matrix on the set of probability distributions representing the occupation of different energy levels through the
transition probabilities
Ti|j = hi| T (|jihj|) |ii ,

(4)

if |ii are the eigenstates of HS .3 This set of stochastic


matrices will be denoted by T :
Definition 2 (Thermal processes T ). The set of stochastic maps that can be realised as in Eq. (4), where T is
a Thermal Operation, will be denoted by T and called
thermal processes.
We have the following reduction result for TO among
incoherent states:
Proposition 9. Let and quantum states with [, HS ] =
[, HS ] = 0. Denote by p and q the eigenvalues of and ,
respectively. Then there exists a TO T such that T () =
if and only if there exists a thermal process T T such that
T p = q.

More generally, one has projectors i on the energy eigenspace ~i .

2.

Thermo-majorisation

Let p, q denote n-dimensional probability distributions. Let g = (g1 , ..., gn ) be the thermal Gibbs distribution, i.e.
gi = e~i /Z,

(5)

where {~i } are the eigenstates of the Hamiltonian HS


of the system, = 1/kTP
B is the inverse temperature of
~i
the environment, Z =
is the partition funcie
tion of HS . We now review a relation between probability vectors that generalises the notion of majorisation
[40], which found widespread applications in quantum
information, especially in the context of entanglement
theory [41]. In the same way in which majorisation measures how mixed a probability distribution is (i.e. how
close to the uniform distribution), thermo-majorisation
measures how thermal a distribution is (i.e. how close
to g):
Definition 3 (Thermo-majorisation [5]). p T q if
and if and only if the piecewise
curve obP linear P

k
k
tained by joining the points
g
,
p
i=1 (i)
i=1 (i)
lies
all
over
the
curve
obtained
joining
the
points
P

Pk
k
Here
i=1 g 0 (i) ,
i=1 q 0 (i) , where k = 1, .., n.
(respectively, 0 ) is the permutation ensuring that
p(i) /g(i) (respectively, q(i) /g(i) ) are sorted in decreasing order.
It is worth noticing that this is a particular instance
of the notion of dmajorisation ([33], Chapter 14). Majorisation (denoted by ) is recovered by taking g to be
the uniform distribution. Definition 3 gives a geometrical interpretation to thermo-majorisation by means of
curves called thermo-majorisation curves, a generalisation of the so-called Lorenz curves. The thermal state g
is described by a straight line connecting the origin (0, 0)
to the point (1, 1) (an equivalent definition uses the unnormalised Gibbs-state, so that the final point is (Z, 1)).
Any other curve will lie above this one. There exists p, q
that are incomparable, in the sense that neither p T q
nor p T q. These are the distributions whose thermomajorisation curves cross.

8
The interest in this condition lies in the fact that it
characterises the constraints present in various thermodynamic models. In particular we have defined the set
of thermal processes T . Another interesting set of transformations is that of Gibbs-preserving quantum maps:
Definition 4 (Gibbs-preserving stochastic processes G ).
For a system with thermal distribution g, the set of
stochastic matrices G such that Gg = g is denoted by G .
One can check T G . Moreover, these two sets
induce the same transitions and are characterised by
thermo-majorisation:
Lemma 10. Let p and q be two probability distributions.
Then the following are equivalent:
1. There exists G G such that G(p) = q.

3. p T q
Proof. For 1 3 see [42]. For 2 3 see [1, 5]. Later we
will give an alternative, direct proof of this equivalence,
showing that actually T = G (see Lemma 24)
Essentially this generalises the classic result that p  q
if and only if there exists a bistochastic matrix M such
that M p = q (bistochastic means stochastic and preserving the uniform distribution) [40].
For technical purposes it will be useful to use an
equivalent definition of thermo-majorisation:
Lemma 11. p T q if and only if


X

n
qj
pj


gj a
gj a
gj
gj
j=1
j=1

Definition 5. Let g be a thermal state with rational entries and let d be defined as above. Then the embedding
map is an application d : Rn RD such that
d (p) = ni=1 pi i ,

where i is a di -dimensional uniform distribution and


(p, p0 ) (q, q 0 ) := (p, p0 , q, q 0 ).
Note that, by definition, d (g) = , where is a Ddimensional uniform distribution. We are now in the
position of proving a useful technical lemma:

Proof. Using Lemma 11 for a trivial Hamiltonian in a Ddimensional space (so that plays the role of g in RD ),
one has d (p)  d (q) if and only if
D
X

|D(q)j a|

(6)

The original proof can be found, in a more general


context, in [42, 43]. Note that, if g = (1/n, ..., 1/n) is the
uniform distribution, then the condition of Lemma 11
reduces to a characterisation of majorisation, p  q ([40],
Chapter 2). Lemma 11 can also be derived from the
above mentioned result on majorisation using the techniques introduced in the next section.

D
X

|D(p)j a| a 0.

Taking into account the structure of the embedding


map, both sides of (9) can be split into two distinct sums:
one over i, running from 1 to n and, for every i, a sum
Pi
over j from ji1 to ji , where ji = k=0 dk (here d0 := 0
by convention). Then,
|D(p)j a| =

j=1

n
X
i=1



ji
n
X
X
pi

D a
di

i=1 j=j
i1

X



n

pi

pi



Dgi a ,
di D a =
di
gi
i=1

where in the last line we used Eq. (7). From Eq. (9) and
the previous equation one can see that d (x)  d (y) if
and only if
n
X

3.

Embedding

Let us now introduce the concept of embedding map,


first introduced in [7]. The definition is given for thermal states whose energies are rational numbers, i.e. for
which there exists a vector of integers d = (d1 , ..., dn )
with


n
X
d1
dn
g=
, ...,
, D=
di .
(7)
D
D
i=1
This can be assumed without loss of generality because
real numbers can be approximated to any precision by

(9)

j=1

j=1

D
X

a 0.

(8)

Lemma 12. p T q if and only if d (p)  d (q).

2. There exists T T such that T (p) = q.

n
X

rationals and, at any rate, the case in which energies


are irrationals can formally be dealt with [7, 16]. From
now on we will then assume that g can be written as in
Eq. (7):


X


n
qi


pi



gi a
gi a
gi
gi
j=1
j=1

a 0.

Using again Lemma 11, this time for the Hamiltonian of


the system, i.e. for the n-dimensional probability distribution g, we obtain the result we wanted.
T-transformations

In the context of bistochastic maps, i.e. stochastic matrices that preserve the uniform distribution, a special
role is played by the so-called T -transformations or T transforms ([33], Chapter 2):

9
Definition 6 (T -transformation). A T -transform is
stochastic map T with
T = I + (1 )Q,

[0, 1],

(10)

where Q is a transposition of any two elements of the


probability distribution and I is the identity matrix.
One then has the following Lemma (B.1 of [33]), due

to Muirhead, Hardy, Littlewood, Polya:


Lemma 13. p  q if and only if p can be achieved from q by
only applying a sequence of T-transformations. The number
of T -transformation required is bounded by the dimension of
the probability vectors.
Proof. A proof can be found in [33] (Lemma B.1) as well
as [6] (Lemma 19) and [40].
B.

Elementary Thermal Operations (ETOs)

In this section we introduce the concept of Elementary Thermal Operations. We discuss their appealing
physical properties and show that most of them can be
realised if we can resonantly couple any transition of
the system to a single-mode bosonic bath by a standard
Jaynes-Cummings coupling. We also study the Markovian limit of ETO, showing that they generate the same
dynamics of Davies maps.
1.

Definition

In the context of thermodynamics, a central role is


played by the detailed balance condition [32, 44]:
Definition 7 (Detailed Balance). A stochastic map M
is called detailed balanced with respect to the thermal
state g if
Mi|j
gi
= e~(i j ) =
Mj|i
gj

(11)

Every detailed balanced map is in G , i.e. M g = g:


X
X
(M g)j =
Mj|i gi =
Mi|j gj = gj
(12)
i

However the opposite implication is not true (see [45],


Appendix A). Also, T contains transformations that do
not satisfy detailed balance ([5], Supplementary Note 6).
A particularly simple set of detailed balanced stochastic
maps is defined as follows:
Definition 8 (Elementary Detailed balanced process
(EDP)). An EDP is a stochastic map E satisfying two
properties:
1. The transformation only couples two energy levels, i.e. there are two energies ~i , ~j of HS , with
i 6= j , such that Ek|m = 0 if k, m 6 {i, j}.

2. E is detailed balanced.
The set of all EDP will be denoted by E . Elementary
Thermal Operations are defined as Thermal Operations
that induce such elementary stochastic processes:
Definition 9 (Elementary Thermal Operation (ETO)).
An ETO is a Thermal Operation T whose induced
stochastic map T is an EDP.
More explicitly, every E E has the especially simple
structure

..
..
..
..
..
.
.
.
.
.

. . . Ei|i . . . Ei|j . . .

E=
. . . . . . . . . . . . . . .
. . . Ej|i . . . Ej|j . . .

..
.. . .
. ..
..
.
.
.
.
where Ei|j = Ej|i e~(i j ) and dots are all zeros for
off-diagonal elements and 1 for diagonal elements. Due
to stochasticity, if HS is known then each EDP is fully
determined by the labels i, j and a single transition
probability. This can be chosen to be Ei|j [0, 1], if
j > i .
2.

Proof E T

Lemma 14. E T , i.e., for any EDP there exists a TO


inducing it. Conversely, not all thermal processes are elementary.
Proof. To prove this, let |0 i, |1 i be two eigenstates of
HS with energies ~0 and ~1 , with 1 > 0 . Let E
be an EDP involving them, fully characterised by the
transition probability E0|1 [0, 1]. We now show that
E T , by explicitly constructing a TO inducing it.
Consider P
a single-mode thermal bath with Hamilto
nian HB = n=0 n~ |nihn|, with = 1 0 . Let the
bath be in the state B , with B = eHB /ZB , where
ZB = (1 e~ )1 . Take now the energy-preserving
unitary (appearing in [4])
U = |0 ih0 ||0ih0|+

1
X
X

vkk0 |k ihk0 ||n ki n k0 ,

n=1 k,k0 =0

where V (whose elements are vkk0 ) is a 2 by 2 unitary.


Hence, by Definition 1, the transformation


7 TrB U ( B )U := T (),
is a TO T . One can readily compute the induced transition probabilities of T , as defined in Eq. (4). These give
T1|0 = e~ |v10 |2 ,

T0|1 = |v01 |2 ,

the others being determined by the condition of stochasticity (that follow from the unitarity of V ). We can take
V of the form


cos(x) i sin(x)
V =
.
i sin(x) cos(x)

10
Then for any E0|1 [0, 1] we can choose x such that
sin2 (x) = E0|1 , so that T0|1 = E0|1 . Hence we realised
every EDP through a TO, which implies E T . The
inclusion is strict because, as mentioned just after Definition 7, there exists TOs whose induced stochastic process does not satisfy the detailed balance condition.

=I
1. E
2. Or there are indexes i,j with j > i and
i|j = 1,
E

We also introduce
Definition 10 (Complete set of ETOs). A set of ETOs is
called complete if all E E can be induced by them.
The previous result has shown that such sets exist.

3.

Definition 11 (Thermo-transposition). A thermo such that either


transposition is a stochastic map E

Further properties of EDPs, thermo-transpositions

Detailed balance is enjoyed by the commonly used


operations modelling reversible systems coupled with a
heat bath (see [28, 29, 44, 46, 47] and references therein).
Detailed balance is sometimes known under the name
of microscopic reversibility and it is used by Crooks
to derive his famous theorem [44], and more recently in
single-shot generalisations [45]. Hence this property not
only makes the set of transformations more tightly connected to practical realisations, but it is also of great theoretical interest because it ensures that fluctuation theorems are satisfied by processes built composing them.
For completeness, we show the following:

j|i = e~ ,
E

As expected by analogy to the theory of bistochastic


processes, EDP are the thermodynamic analogue of Ttransformations. The following property of EDP makes
clear that they are the thermodynamic analogue of T transformations:
Lemma 16. Any E E can be decomposed as a convex com
bination of the identity matrix and a thermo-transposition E:

E = (1 )I + E,

q0 (n) = q0 (0)(1 Z)n +

N
(1 (1 Z)n )
Z

Then q0 () := limn q0 (n) = N/Z, independently of


the initial distribution. Because the map is stochastic,
q() = N g
where g = (1/Z, 11/Z) is a thermal distribution within
the two-level subsystem.
Transpositions are stochastic maps acting only on two
levels and having transition probabilities E0|1 = 1,
E1|0 = 1. Such processes are not thermal processes.
The closest we can get to such process while satisfying
the condition that the thermal state is a fixed point is
E0|1 = 1, E1|0 = e~ , where is the transition under
consideration. We then define

[0, 1].

(13)

Proof. By direct inspection, choosing = E0|1 .


This should be compared with Eq. (10). We now analyse if all EDPs can be induced by Thermal Operations
corresponding to some simple physical toy model (the
transformations used to prove Lemma 14 are not good
for this purpose, as they involve rather non-standard
couplings; see Appendix E.3 of [11]).

Lemma 15. Let E be an EDP on two given energy levels.


Then, limn E n maps the subsystem it acts on to a distribution proportional to the thermal state.
Proof. Let the relevant two-level subsystem, with energies ~0 , ~1 , be described by the (possibly unnormalised) distribution q(0) := (q0 (0), q1 (0)). Without loss of generality, take 0 = 0 and 1 = .
Let q(n) := E n q(0).
One can easily compute
q0 (1) = q0 (0)(1 Z) + N , where Z = 1 + e~ , N =
q0 (0) + q1 (0) and := E0|1 . By recursion,

k|m = km , (k, m) 6= (i, j).


E

C.
1.

Physical realisations

The Jaynes-Cummings model

One of the simplest toy models of open system dynamics is the Jaynes-Cummings model [48, 49], describing the interaction of a two-level system with a single
mode of the electromagnetic field, for example in an optical cavity. If the field is taken to be in a thermal state,
can we achieve all EDPs within this model? In other
words, do we obtain a complete set of ETOs?
Lemma 17. Let HS be a non-degenerate Hamiltonian with
no equal spacings. Then, the Jaynes-Cummings interactions
in RWA approximation with a single-mode bosonic bath are
ETOs. In fact, in the low-temperature limit they are a complete set of ETOs.
Proof. Let |k i, k = 0, 1 be two eigenstates in HS , with
energy ~k , 1 > 0 and 1 0 = . Consider the
Jaynes-Cummings Hamiltonian HJC resonantly interacting these two levels with a single mode bosonic bath
of frequency in RWA approximation, i.e.
HJC = g(+ a + a ),
where [a, a ] = IB and g is a coupling constant. Notice
that we work in interaction picture. Here HB = ~ a a is

11

the bath Hamiltonian, + = |1 ih0 | and = +


. One
can compute ([11], Appendix E.2)

it
~ HJC

X
1
X

(n)

uk,k0 (s) |k ihk0 | |n kihn k 0 |

n=1 k,k0 =0

+ |0 ih0 | |0ih0| ,
(n)

where s = gt/~ and, for each n, uk,k0 (s) are matrix elements of the unitary
U (n) (s) =

n) i sin(s
cos(s
n)
i sin(s n) cos(s n).

Take system and bath to be in the initial state


eHB /ZB , where ZB = (1 e~ )1 . One then
has that the dynamics of the system is described as
i
h it
it
7 TrB e ~ HJC ( eHB /ZB )e ~ HJC ,
so it is a Thermal Operation. This induces a stochastic
process J whose transition probabilities are:
J1|0 (s) =

J0|1 (s) =

e~n
sin2 (s n)
,
ZB
n=1

e~(n1)
,
sin2 (s n)
ZB
n=1

J0|0 = 1 J1|0 ,

J1|1 = 1 J0|1 . (14)

So J1|0 /J0|1 = e~ , i.e. the transformation is detailed


balanced. This shows that every transformation within
the resonant Jaynes-Cummings model in RWA approximation gives an ETO.
To investigate the low temperature limit notice that
we only need to show that in such limit J0|1 (s) can
achieve all values in [0, 1]. For every s we have
J0|1 (s) (1 e~ ) sin2 (s)
At s = 0, in fact, equality holds. By continuity of J0|1 (s)
and the previous inequality, we see that all values in
[0, 1] can be achieved, varying s, in the limit .
Hence, at zero temperature all EDPs can be induced
within the Jaynes-Cummings model. Hence, varying s,
we get a complete set of ETOs.
One may be tempted to hope that all EDPs can be
achieved within the Jaynes-Cummings model at all temperatures. Analytically one immediately notices that
J0|1 (s) < 1 for all times s. However this does not exclude that there is a sequence of si such that J0|1 (si ) gets
arbitrarily close to 1, i.e. that we can approximate EDPs
arbitrarily well within the Jaynes-Cummings model. As
the dynamics in s is very chaotic, this cannot be excluded numerically. What we need is a bound on J0|1
independent of s. In the following lemma we prove such
bound:

Lemma 18. For the Jaynes-Cummings model, set = ~.


Then for every s


1 
8e e2 + e3 + 8 , for [0, log(4)/3]
16

J0|1 (s) e4 e3 + 1
for log(4)/3.

J0|1 (s)

The usefulness of these bounds is that they prove that


for any fixed temperature (0, 1) one will have a constant C (0, 1) such that J0|1 < 1 C for all s. So, the
Jaynes-Cummings model cannot achieve all EDPs nor
arbitrary approximations of them:
Corollary 19. For any fixed (0, 1), the JaynesCummings cannot induce arbitrary close approximations of
every EDP.
Proof of Lemma 18. Consider the bound on J0|1 (s) constructed following these steps:
1. Split the series in Eq. (14) into a finite sum F up
to m, plus a residue R. In the residue, bound each

factor sin2 (s n) with 1. Then one has R = em .


2. In F , bound every oscillating term with irrational

frequency as sin2 (s n) 1.
This provides a bound of J0|1 (s) with a periodic function of s, which can then be simply analysed. For m = 4,
the function has one global maximum at s = /2 when
log(4)/3. On the other hand, when < log(4)/3

it has two global maxima at s = arccos(e3 /4)/2.


The values achieved at the maxima give the bounds of
Lemma 18.
On the other hand, a rough lower bound (good when
temperatures are not too high) of the achievable J0|1 can
be obtained by truncating the sum in Eq. (14) at some finite mmax , and setting every other term to zero. We take
mmax = 12. Then numerics suggest to choose s = 98.92.
This, together with the upper bounds of Lemma 18, provides a region where the curve of maximum achievable
probability J0|1 must lie as a function of (Fig. 1 in the
all values from 0 to this limit can
main text). For any ,
be achieved, because J0|1 (s) is a continuous function in
s and it takes the value 0 at s = 0. Hence Figure 1 shows
that, even though not every EDPs can be realised within
the Jaynes-Cummings model, a vast majority of them
can be achieved, especially at lower temperatures. In
other words, we realised within a physically reasonable
model an approximation to a complete set of ETOs.
2.

Thermalisation models: proof of Eq. (1)

Following the first part of the proof of Lemma 15, we


obtain that the effect of n interactions on the occupation
of the lower energy state is
q0 (n) = q0 (0)(1 Z)n +

N
(1 (1 Z)n ),
Z

(15)

12
where = E0|1 , Z = 1 + e~ , N = q0 (0) + q1 (0). We
can define n = t/tint , with tint the duration of a single
interaction. Then we will take the limit tint 0, 0,
while keeping Z/tint 0 finite (Z is a constant
that we absorbed in the definition of ). Then one can
check that (1 Z)n et/ and hence
q(t) = et/ q(0) + N (1 et/ )g.

(16)

Here 0 is a free parameter. So the state decays exponentially to a Gibbs distribution (whenever > 0).

3.

Thermalisation models: comparison to Davies maps

An alternative, but equivalent, point of view is given


by Davies maps. These maps emerge as the Markovian
description of a system weakly coupled to an infinite
thermal bath [28]. Let us begin with a definition [47]:
Definition 12. A Davies map M is a map such that
1. M = exp(Lt).
Here L = H + L0 , with
H() = [, H], H hermitian operator and L0 a general Lindbladian.
2. [H, L0 ] = 0 (covariance).




3. Tr L0 (BeH )A = Tr BL0 (eH A) for all A, B
(quantum detailed balance).
The action of the most general qubit Davies map on a
two-level system was studied in [47], Section 3.
Notice that by taking A = |iihi|, B = |jihj| in the third
property we can show that the induced stochastic map
(defined through Eq. (4)) satisfies detailed balance in the
sense of Definition 7. The covariance condition ensures
that every Davies map is time-translation symmetric, in
the sense discussed in Appendix G.
As expected, acting with Davies maps on a state
whose population is initially described by q(0), one gets
back again Eq. (1). Every Davies map induces an ETO.
In particular, as shown in the main text, it gives Markovian ETOs.

D.

Elementary Thermal Operations are universal for


incoherent processes
1.

Proof. From Lemma 10 we need to equivalently prove


p T q if and only if q can be obtained from p through
a sequence of EDPs. From Lemma 12, p T q if
and only if d (p)  d (q) (recall that p Rn and
d : Rn RD ). Now, due to Lemma 13, we know that
d (q) can be obtained from d (p) through a sequence of
T -transformations T i :

Proof of Theorem 2

We can now prove one of the main results, showing


the sufficiency of ETOs. TOs commute with dephasing
in energy, so we get [, HS ] = 0. Hence, using Proposition 9, Theorem 2 can be rewritten as follows:
Theorem 20. A thermal process T T such that T (p) = q
exists if and only if q can be obtained from p through a finite
sequence of EDPs.

d (q) = T 1 . . . T D d (p),

(17)

perhaps some of them trivial. Here we used the fact that


the number of T -transformations required is finite and
bounded by the dimension of the probability distribution.
i
Next define E i = 1
d T d . We will prove the following:
1. q = E 1 . . . E D (p)
2. Each E i is an EDP.
The first point is immediately checked by substituting
T i = d E i 1
d in Eq. (17). Regarding the second
point note that, from Def. 6,
i
E i = I + (1 )E

[0, 1],

(18)

i = 1 Qi d
where I is the identity matrix in Rn , E
d
i
D
and Q is a transposition in R . Because the convex
combination of an EDP and the identity is readily seen
i is
to be just another EDP, we only need to show that E
an EDP.
i on a probability distribution
Consider the action of E
i

p. Defining q = E (p), we get


qk =



1
1
1
pk + pj ,
dk
dj

qj =



1
1
pk + 1
pj ,
dk
dj

where we used the definition of embedding map, Def. 5.


One can immediately see that the choice of i fixes the
two indexes k, j (but many choices of i in general correspond to the same k,j). The previous expression immediately shows that the transformation involves only two
energy levels. Moreover, using di = gi D (see Eq. (7))
and the previous equation, we obtain for the transition
i,
probabilities of E
i = gk /gj = e~(k j ) ,
i /E
E
k|j
j|k

i.

(19)

i is detailed balanced. By construction, the


Hence, E
transition is achieved using a finite number of EDPs.
In other words, EDPs are the analogue in thermodynamics of T-transforms, and Theorem 2 is the generalisation to thermodynamics of Lemma 13 by Muirhead,

Hardy, Littlewood, Polya.

13
Explicit construction of EDPs sequence in Theorem 2

Because the proof of Lemma 13 is constructive, i.e. it


explicitly gives a prescription to build each T i , the above
embedding gives a way to explicitly construct the maps
i = 1 T i d . Suppose one wants to construct a
E
d
sequence of transformations (18) achieving the transition from p to q. By assumption a thermal process transforming p into q exists, so p T q from Lemma 10. But
then from Lemma 12 we know that d (p)  d (q), so
we can run the construction of [33] (Lemma B.1), also
reported in [6] (Lemma 19). This construction gives
the sequence of T -transformations mapping d (p) into
d (q). Sort both p and q in such a way that pi /gi and
qi /gi are non-increasing.4 This is called -ordering in
[5] and will be denoted by p , q . One can then take
T i = I + (1 )Qi , where
i

1. Q is the transposition between the index jex


and jdf . Here jex (using the notation of [6]) is
the largest index j such that d (p )j > d (q )j
and jdf is the smallest index j such that
d (p )j < d (q )j .
The existence of both
these indexes can be proven as shown in [33],
Lemma B.1.
2. Defining
= min{d (p )jex d (q )jex , d (q )jdf d (p )jdf },

one transfers from jex to jdf , i.e.


d (x )jex 7 d (x )jex ,

d (x )jdf 7 d (x )jdf +,

by taking = 1 /(d (p )jex d (p )jdf ).


As proved in [33] (Lemma B.1), iterating the reasoning above one finds a finite sequence of T transforms
mapping d (p) into d (q). As should be clear from the
proof of the previous theorem, the corresponding se i = 1 T i d , transquence of maps (18), where E
d
forms p into q.
An interesting line of research is to go beyond this
construction and find shorter sequences of transformations; one obvious reason is that there are many T i that
give an EDP among the same two levels (k, l). However, composing EDPs among the same levels gives yet
another EDP, so all of these can be grouped into a single step. This can cut massively the number of required
elementary steps. Another idea may be to require that
the target state is obtained only within a given level of
precision (which is physically reasonable).

In case for some i, j pi /gi = pj /gj , i will precede j if pi pj .

2.

Proof of Theorem 3, Lemma 5, Corollary 4

A second result showing the theoretical importance of


EDPs, and in particular thermo-permutations (Def. 11),
is the following. It is a famous theorem of Birkhoff
that bistochastic maps are the convex hull of permutations. This generalises to thermal processes (the set T
of Def. 2) as follows (Theorem 3 in the main text):
Theorem 21. The set T coincides with the convex hull of
thermo-permutations, i.e. with the set of T that can be written
as
X
X
T =
i P i , i 0,
i = 1,
(20)
i

where each P i is a thermo-permutation.


The result above is an immediate consequence of two
Lemmas, the second of which has some interest by itself as we will discuss. Recall the definition of Gibbspreserving processes G given in Def. 4.
Lemma 22. The set G coincides with the maps that can be
decomposed as in Eq. (20).
Proof. Let G G . Then the map d G 1
d is bistochastic. In fact, it is a composition of stochastic maps
and it is immediate to prove that the uniform distribution is preserved using the properties of d and the fact
that G is Gibbs-preserving. So, using Birkhoff theorem
this map can be decomposed as a convex combination of
k
permutations M k . Hence, defining Dk = 1
d M d
we obtain
X
X
G=
xk Dk , xk 0,
xk = 1.
(21)
k

Each Dk is a convex combination of thermopermutations. In fact, M k = Qk1 . . . Qknk , where Qkj are
transpositions. So Dk = E1k . . . Enkk , where Ejk = 1
d
Qkj d . As we have seen in the second part of the proof
of Theorem 2, Ejk are EDPs. So
G=

xk E1k . . . Enkk .

(22)

k,
Moreover, by Lemma 16, Ejk = (1 kj )I + kj E
j
k is a thermo-transposition. Hence, after subwhere E
j
stitution we have an expression of each Dk as a convex combination of thermo-permutations. After
P reari
ranging we obtain a final expression G =
i i P ,
for some probability distribution {i } and thermopermutations P i .
P
Conversely, let G = i i P i for thermo-permutations
i
i
P . Each P is a composition of thermo-transpositions.
Each thermo-transposition is an element of G , as it is
their composition. G is clearly a convex set, so G G .

14
As an intermediate result, we proved Corollary 4:
Corollary 23. Every thermal process can be simulated by finite sequences of ETOs from a complete set and classical randomness.
Proof. Every thermal process is Gibbs-preserving, so it
can be decomposed as in Eq. (22). This gives the statement.
It is known that G allows the same transitions as T
[35] (Lemma 10). However T and G may well be different sets of maps. For example, the set of unital and that
of noisy channels [36] are distinct sets, even if the same
set of final states can be reached from an arbitrary initial
state by either set of maps. However we prove here that
G and T coincide as sets of stochastic maps:
Lemma 24. The set of Gibbs-preserving stochastic maps coincides with the set of thermal processes, i.e. T = G .
Proof. Every thermal process is Gibbs-preserving by direct inspection. Conversely, let G G . From Lemma
22, it can be decomposed as in Eq. (20). Each Pi is a
thermo-permutation. Hence, it can be achieved by composing thermo-transpositions. Thermo-transpositions
are EDPs, and EDPs are thermal processes (Lemma 14).
As the set of thermal processes is closed under composition, thermo-permutations are thermal processes. So
G is a convex combination of thermal processes. However T is a convex set, because TO are a convex set (see
[15], Appendix C). So we conclude that G is a thermal
process. Hence, the two sets coincide.
Note that this result does not hold true in the quantum case [35]. This is for the simple reason that TOs
cannot generate quantum coherence in the energy basis, whereas Gibbs-preserving quantum maps can (see
Appendix G). One can now prove Theorem 21 simply
applying the previous two Lemmas one after the other.

3.

A simple proof of the central result on Thermal Operations

Lemma 24 (Lemma 5 in the main text) greatly simplifies the proof of the central result of the theory of Thermal Operations, Lemma 10 (the result of [5]). This is important because it clarifies the main reasoning, reduces
the number of assumptions and makes it easier to build
over the known result. Previous proofs of this central
lemma are significantly harder and seem to require extra assumptions about the bath (see (i)-(iv) of Supplementary Note 1 of [5]). Here, the proof follows directly
as an immediate corollary of Lemma 24:
Theorem 25 ([5]). Let , be block-diagonal states in the energy eigenbasis, with eigenvalues p and q, respectively. Then
a Thermal Operation T such that T = exists if and only if
p T q.

Alternative proof. First, from Proposition 9, this is the


same as proving that p T q is equivalent to the existence of T T such that T p = q. From Lemma 24,
this can be reformulated as the existence of a map G
such that G(p) = q and G(g) = g. Using the embedding map this is equivalent to the existence of a bistochastic map M such that M (d (p)) = d (q) (to see
this, define M = d G 1
d ). But the answer to this
problem is a classic result in the theory of majorisation
[40]: it is equivalent to d (p)  d (q). However, from
Lemma 12, this is equivalent to p T q.

4.

Lemma 5 and coherence preservation under Thermal


Operations

Lemma 5 is a useful tool to investigate optimal


coherence-preserving processes. Given an arbitrary
quantum state with energy population given by p and
some target population q, it is natural to wonder what is
the Thermal Operation T that maps p into q while preserving the maximum amount of the initial coherence of
. This corresponds to the question: given some heat exchange with the bath, what is the minimal degradation
of its coherent properties that the system must undergo?
This problem can be split into two steps:
1. Find all T T such that there exists T satisfying
Eq. (4) and T p = q.
2. For each T , find a Thermal Operation T that induces T and optimally preserves coherence, according to some given measure
One of the problems that one immediately faces is the
lack of a simple characterisation of T . Lemma 24 provides it: the first step becomes finding all G such that
Gp = q and Gg = g. For the second step, one may use
the inequalities presented in [15].

E.

Comparison to previous work

It is useful to make a connection with a work related


to the present one [24]. There, a set of transformations
called Partial Level Thermalisations are allowed, but
seen to be insufficient to perform every transition. In
agreement with that we have the following connection
with EDPs:
Lemma 26. Partial Level Thermalisations involving two energy levels ~i and ~j , i.e. the transformation:
(xi , xj ) 7 (1 )(xi , xj ) + N g

(23)

with all other xk unchanged are a strict subset of EDPs. Here



1
e~
N = xi + xj , = j i > 0, g = 1+e~
, 1+e
.

15
Proof. One can easily verify that the transformation in
Eq. (23) is characterised by the following transition
probabilities:
pi|j =


,
1 + e~

pj|i =


,
1 + e~

(24)

and pi|i = 1 pj|i , pj|j = 1 pi|j . This and the fact that
pj|i /pi|j = e~ shows that every Partial Level Thermalisation is an EDP. The converse is not true because
p0|1 1/(1 + e~ ) < 1, for all [0, ), whereas
EDP require one to be able to achieve all p0|1 [0, 1].
As mentioned in the main text, Partial Level Thermalisations correspond to embeddable EDP. In fact, for a
wide range of parameters, there are Jaynes-Cummings
interactions that cannot be realised by means of Partial
Level Thermalisations:
Lemma 27. There exist resonant Jaynes-Cummings interactions with a single-mode bosonic bath that cannot be realised
by means of Partial Level Thermalisations.
Proof. Let j > i . From the proof of Lemma 26,
the maximum probability of de-exciting the level is
pmax (~) = (1 + e~ )1 . However, suppose we resonantly couple the two energy levels to a single-mode
bosonic bath via a Jaynes-Cummings coupling in RWA
approximation. Following the calculations in the proof
of Lemma 17, we find a de-exciting probability equal to
Ji|j (s). In Appendix B.2 we obtained a lower bound on
the maximum achievable de-exciting probability within
the Jaynes-Cummings model, Jmax (~). This can
be compared with the maximum achievable transition
probability through Partial Level Thermalisations This
allows us to check that at least for ~ (0.1, 6.4)
Jmax (~) > pmax (~).

for some a N\{0}, b N\{0} and w = (0, 1). The


Hamiltonian of the final state is taken to be Hw =
w |wihw|.
Eq. (26) can be understood as a transformation of heat
into work, in the form of b weight systems in a pure excited state of energy w.
1.

Proof of Theorem 8

First, let us rewrite the theorem in a more precise way:


Theorem 28. Let F a set of stochastic matrices such that
F E and such that it is impossible to obtain arbitrary
approximations of the transformation in Eq. (26). Then,
F T.
Proof. From Lemma 24, we only need to prove the above
statement for G instead of T .
By contradiction, assume G F . Let M F ,
M 6 G . Then there exists x 6= g such that M (g) = x.
Take many copies of g (which can be thought of as a
larger thermal bath composed of many thermal particles). Applying the map M to each system independently we obtain by assumption
M a (g a ) = xa
Consider now the transition
xa wb

(25)

Equivalent thermodynamic frameworks

A classical framework F is defined as a set of stochastic processes that can be induced on diagonal states
without paying any work. Whatever our choice is, it
should lead to a consistent thermodynamic framework
in which work cannot be extracted from a single heat
bath at a given temperature. In other words, we clearly
do not want our theory to allow for the construction of a
perpetuum mobile. This is made precise by the following:
Definition 13 (Perpetuum mobile). A perpetuum mobile (of second kind) is a machine extracting work from
a single heat bath at given temperature. Specifically, if
g is the thermal distribution, it allows to approximate
arbitrarily well the transition
g a wb

(26)

(28)

Taking a , there exists a thermal process T that


achieves with arbitrary precision a number b of
pure excited states with any finite energy w at a rate R
given by Theorem 1 of [3]:
R=

F.

(27)

S(x||g)
> 0,
w

(29)

P
where S(x||g) = i pi log(pi /gi ). R > 0 because for the
relative entropy S(x||g) = 0 if and only if x = g. Equivalently, for every  > 0 there exists a thermal process

T T such that T xa wb . However, from Theorem 2, the thermodynamic transition (28) can be realised
as a sequence of EDPs E 1 ,...,E N . Hence, using only operations in F , we achieved the perpetuum mobile transition of Eq. (26) with b = aS(x||g)/w. This is a contradiction, so it must be F G , that gives F T
2.

Proof of Corollary 8

Proof. This can be proven by showing that all thermodynamic frameworks with the mentioned properties are
equivalent to the T . In fact
1. Equivalence of frameworks is an equivalence relation, so this proves the equivalence of any two
frameworks.

16
2. Because thermal processes are characterised by
thermo-majorisation (Lemma 10 or Theorem 25),
this proves the second part of the claim as well.

scribes scenarios in which a quantum system interacts with a dissipative environment and no classical external field is applied.

Let F be a thermodynamic framework. Let us fix p


and q. If there exists a T T such that T (p) = q, then
there exists a sequence of EDPs mapping p into q (Theorem 20). Because E F , we conclude that the transformation p q can be achieved using only maps in
F.
Conversely, assume that one can map p into q using
only transformations in F . Because F does not allow
for a perpetuum mobile, it is contained in the set T of
thermal processes (Theorem 28). Hence, one can transform p into q by means of thermal processes. We conclude that F is equivalent to T .

M2 Those in which coherence can be freely exchanged


with work, an example being Gibbs-preserving
quantum operations [35]. These models implicitly
assume the possibility of using an external classical field [12] that allows to freely transform coherence into work and convert it back, but not to do
work on the system.
In the models M1, the crucial property is timetranslation symmetry [52]. If HS describes the Hamiltonian part of the evolution, the quantum map E is required to satisfy for all and t
E Ut () = Ut E(),

G.

Fully quantum considerations

(30)

M1 Those in which coherence cannot be freely created


or exchanged with work, an example being Thermal Operations. In these models thermodynamics
can be seen as a particular instance of coherence
theory ([12], Theorem 1, [15, 51]). This better de-

where Ut () = eiHt eiHt . For processes described by


a master equation, this is the same as requiring that the
dissipative and unitary part of the generator of the dynamics commute. In fact, this is a property of standard
thermalisation models, such as Davies maps [28, 47].
Time-translation symmetry makes coherence a resource,
as it is unavoidably dissipated in time, and decouples
its evolution from that of population [53].
On the other hand, models M2 are characterised by
the lack of time-translation symmetry, even though the
condition that the Gibbs state is a fixed point still ensures that external work cannot be used to push a state
farther away from equilibrium. Davies maps are not
recovered in the Markovian limit, so these models appear to describe more controlled situations where one
can drive the system with a classical field (but without
pushing the system more out of equilibrium). It would
be useful to understand if such situation can be realised
in a simple physical model.
Hence, the lack of thermodynamic equivalence
(Corollary 8) in the quantum regime then reflects the different roles coherence can play in the model, but this is
ultimately a question of the physical scenario we wish to
describe. A promising line of research is to find whether
quantum thermodynamic models are equivalent once
we specify one of the two alternatives M1 or M2.
These considerations could also offer a new approach
to the problem of analysing coherence constraints in
thermodynamics. For example, one could compare decomposable and indecomposable thermodynamic processes in terms of their ability to preserve coherence. We
leave this to future investigations.

[1] D. Janzing, P. Wocjan, R. Zeier, R. Geiss, and T. Beth, Int.


J. Theor. Phys. 39, 2717 (2000).
[2] O. Dahlsten, R. Renner, E. Rieper, and V. Vedral, New
Journal of Physics 13, 053015 (2011).

[3] F. G. S. L. Brandao, M. Horodecki, J. Oppenheim, J. M.


Renes, and R. W. Spekkens, Phys. Rev. Lett. 111, 250404
(2013).

[4] J. Aberg,
Nat. Commun. 4, 1925 (2013).

An obvious question is how to generalise the results


of the present investigation to the fully quantum case.
Natural questions are in particular: to what extent all
ETOs can be realised within a natural physical model?
Assuming that all ETOs can be realised, is it possible
to induce any transition allowed by means of Thermal
Operations?
It is interesting also to discuss thermodynamic equivalence in the fully quantum case. Theorem 7 does not
holds, because Gibbs-preserving quantum channels are
a strict superset of Thermal Operations and do not allow for a perpetuum mobile. Moreover, also Corollary 8
requires further discussion. In fact, the framework of
Gibbs-preserving quantum channels is inequivalent to
that of Thermal Operations.
The reason for these differences has to do with the
physics more than to some technical difficulty. In fact,
whereas work is a universal resource in thermodynamics classically, it is not such quantum-mechanically [12].
A second resource in thermodynamics is quantum coherence among energy eigenspaces, and the inequivalence of various thermodynamic models arises from the
different role coherence plays in them.
Broadly speaking, there are two general classes of
models:

17
[5] M. Horodecki and J. Oppenheim, Nat. Commun. 4, 2059
(2013), 10.1038/ncomms3059.

[6] G. Gour, M. P. Muller,


V. Narasimhachar, R. W. Spekkens,
and N. Y. Halpern, Physics Reports 583, 1 (2015).
[7] F. G. S. L. Brandao, M. Horodecki, N. H. Y. Ng, J. Oppenheim, and S. Wehner, Proc. Natl. Acad. Sci. U.S.A. 112,
3275 (2015).
[8] L. Masanes and J. Oppenheim, arXiv 1412.3828 (2014).
[9] V. Narasimhachar and G. Gour, Nature Communications
6, 7689 (2015).
[10] J. M. Renes, Eur. Phys. J. Plus 129, 153 (2014),
arXiv:1402.3496 [math-ph].

[11] J. Aberg,
Phys. Rev. Lett. 113, 150402 (2014).
[12] M. Lostaglio, D. Jennings, and T. Rudolph, Nat. Commun. 6, 6383 (2015).
[13] N. H. Y. Ng, L. Mancinska, C. Cirstoiu, J. Eisert, and
S. Wehner, New Journal of Physics 17, 085004 (2015).

[14] P. Cwikli
nski,
M. Studzinski,
M. Horodecki, and J. Oppenheim, Phys. Rev. Lett. 115, 210403 (2015).
[15] M. Lostaglio, K. Korzekwa, D. Jennings, and T. Rudolph,
Phys. Rev. X 5, 021001 (2015).

[16] M. Lostaglio, M. P. Muller,


and M. Pastena, Phys. Rev.
Lett. 115, 150402 (2015).
[17] J. Gemmer and J. Anders, New Journal of Physics 17,
085006 (2015).
[18] M. P. Woods, N. Ng, and S. Wehner, arXiv 1506.02322
(2015).
M. Alhambra, J. Oppenheim, and C. Perry, arXiv
[19] A.
1504.00020 (2015).
[20] D. Egloff, O. C. O. Dahlsten, R. Renner, and V. Vedral,
New Journal of Physics 17, 073001 (2015).
[21] V. Narasimhachar and G. Gour, arXiv 1604.03185 (2016).
[22] H. Wilming, R. Gallego, and J. Eisert, Phys. Rev. E 93,
042126 (2016).
[23] J. Goold, M. Huber, A. Riera, L. del Rio,
and
P. Skrzypczyk, arXiv 1505.07835 (2015).

[24] C. Perry, P. Cwikli


nski,
J. Anders, M. Horodecki, and
J. Oppenheim, arXiv 1511.06553 (2015).
[25] R. Kosloff, Entropy 15, 2100 (2013).
[26] M. Ziman, P. Stelmachovic, and V. Buzek, Open systems
& information dynamics 12, 81 (2005).
[27] V. Scarani, M. Ziman, P. Stelmachovic, N. Gisin, and
V. Buzek, Phys. Rev. Lett. 88, 097905 (2002).
[28] E. B. Davies, Communications in Mathematical Physics
39, 91 (1974).

[29] R. Dumcke,
Communications in Mathematical Physics 97,
331 (1985).
[30] E. Davies, Electron. J. Probab. 15, 1474 (2010).
[31] E. B. Davies, Linear operators and their spectra, Vol. 106
(Cambridge University Press, 2007).
[32] G. E. Crooks, Journal of Statistical Physics 90, 1481 (1998).
[33] A. W. Marshall, I. Olkin, and B. C. Arnold, Inequalities:
Theory of Majorization and Its Applications (Springer, 2010).
[34] K. Korzekwa, M. Lostaglio, J. Oppenheim, and D. Jennings, New Journal of Physics 18, 023045 (2016).
[35] P. Faist, J. Oppenheim, and R. Renner, New Journal of
Physics 17, 043003 (2015).
[36] M. Horodecki, P. Horodecki, and J. Oppenheim, Phys.
Rev. A 67, 062104 (2003).
[37] L. del Rio, L. Kraemer, and R. Renner, arXiv 1511.08818
(2015).
[38] J. L. England, The Journal of chemical physics 139, 121923
(2013).
[39] E. Ruch and A. Mead, Theoretica chimica acta 41, 95
(1976).
[40] R. Bhatia, Matrix analysis, Vol. 169 (Springer Science &
Business Media, 2013).
[41] M. A. Nielsen and I. L. Chuang, Quantum computation and
quantum information (Cambridge university press, 2010).
[42] E. Ruch, R. Schranner, and T. H. Seligman, J. Math. Analysis and Applications 76, 222 (1980).
[43] E. Ruch, R. Schranner, and T. H. Seligman, J. Chem. Phys.
69 (1978).
[44] G. E. Crooks, Phys. Rev. E 60, 2721 (1999).
[45] N. Y. Halpern, A. J. Garner, O. C. Dahlsten, and V. Vedral,
New Journal of Physics 17, 095003 (2015).
[46] H.-P. Breuer and F. Petruccione, The theory of open quantum
systems (Oxford University Press, 2002).

[47] W. Roga, M. Fannes, and K. Zyczkowski,


Rep. Math.
Phys. 66, 311 (2010).
[48] E. T. Jaynes and F. W. Cummings, Proceedings of the IEEE
51, 89 (1963).
[49] B. W. Shore and P. L. Knight, Journal of Modern Optics 40,
1195 (1993).
[50] R. Rado, Journal of the London Mathematical Society s127, 1 (1952).
[51] I. Marvian and R. W. Spekkens, arXiv 1602.08049 (2016).
[52] I. Marvian and R. W. Spekkens, Nat. Commun. 5, 3821
(2014).
[53] I. Marvian and R. W. Spekkens, Phys. Rev. A 90, 062110
(2014).

Potrebbero piacerti anche