Thermodynamic length in open quantum systems
Matteo Scandi and Martí Perarnau-Llobet
Max-Planck-Institut für Quantenoptik, D-85748 Garching, Germany
October 15, 2019
The dissipation generated during a qua-
sistatic thermodynamic process can be
characterised by introducing a metric on
the space of Gibbs states, in such a way
that minimally-dissipating protocols corre-
spond to geodesic trajectories. Here, we
show how to generalize this approach to
open quantum systems by finding the ther-
modynamic metric associated to a given
Lindblad master equation. The obtained
metric can be understood as a perturba-
tion over the background geometry of equi-
librium Gibbs states, which is induced by
the Kubo-Mori-Bogoliubov (KMB) inner
product. We illustrate this construction
on two paradigmatic examples: an Ising
chain and a two-level system interacting
with a bosonic bath with different spectral
densities.
1 Introduction
A central task in finite-time thermodynamics is
to design protocols that maximise the extracted
work while minimising the dissipation during the
process. In the slow driving regime, a powerful
approach consists in equipping the space of ther-
modynamic states with a metric whose geodesics
correspond to minimally dissipative processes.
This geometrical construction was first developed
in the 80s for macroscopic endoreversible thermo-
dynamics in a series of seminal papers [18], and
more recently it was extended to the microscopic
regime [912], leading to several applications in,
e.g., molecular motors [13] and small-scale infor-
mation processing [14, 15].
While this approach is well established for clas-
sical systems, the quantum regime has remained
less explored. The geometry of quantum equilib-
rium Gibbs states has been characterised in [16
20]; however the resulting metric does not take
into account dynamical features of the dissipa-
tion, which are of crucial importance in finite-
time protocols. For arbitrary out-of equilibrium
evolutions, a notion of thermodynamic length has
been put forward in [21, 22]; yet, this approach
requires full knowledge of the global unitary evo-
lution, which makes it difficult to apply in com-
mon situations where the size of the bath allows
only for an effective description of the dynamics.
Finally, Kubo linear-response theory also allows
for describing dissipation near equilibrium [23
25], which in turn allows for defining a notion of
thermodynamic metric [24, 26].
The goal of this article is to provide a general
framework to construct a thermodynamic met-
ric whenever the evolution of the system can be
described by a Lindblad master equation (see
also [27]). The obtained metric can be expressed
as the background equilibrium geometry of [16
19] acted upon by the Drazin inverse of the Lind-
blad operator, which encodes the different equi-
libration timescales of the dissipative dynamics.
This geometrical approach is used to find min-
imally dissipating protocols of a slowly driven
Ising chain in a transverse field, and of a qubit in
contact with a bosonic bath with different spec-
tral densities.
2 Dissipation in quantum systems
Before presenting the construction of the met-
ric, we here define the quantum correspondent
of some important quantities of classical thermo-
dynamics. A quantum system is described by
its density matrix ρ and its Hamiltonian H; if
the state is given by the Gibbs ensemble we will
use the notation ρ ω
β
(H), where ω
β
(H) =
e
βH
/Tr(e
βH
) and β is the inverse tempera-
ture of the surrounding bath. We will sometimes
also use the shorthand notation ω := ω
β
(H). A
functional of key importance is the non equilib-
rium free energy, which is defined by the formula
F (ρ, H) = hHi
ρ
β
1
S(ρ), where we define the
1
arXiv:1810.05583v5 [quant-ph] 14 Oct 2019
average as hAi
ρ
= Tr [], and we denote by S(ρ)
the von Neumann entropy of the state.
Direct calculations show that the non equilib-
rium free energy is connected to the equilibrium
one by the equality
F (ρ, H) = F (ω
β
(H), H) + β
1
S(ρ||ω
β
(H)) (1)
where S(ρ||ω
β
(H)) is the relative entropy. This
quantity is positive definite and it can be under-
stood as a measure of how statistically different
the current state is from the thermal one. More-
over, it corresponds to the extractable work from
ρ and, for this reason, it is sometimes referred to
as availability.
We consider thermodynamic processes in which
the Hamiltonian of a system in contact with a
thermal bath is experimentally varied between
two fixed endpoints, say H
A
and H
B
. Each pro-
tocol is then defined by a curve γ in the space of
controllable parameters, and by its duration T .
The work extracted is defined as
W =
Z
γ
dt Tr
h
ρ
t
˙
H
t
i
. (2)
It is useful for what follows to rewrite the work in
terms of the non equilibrium free energy. Using
the insight coming from (1), we integrate equa-
tion (2) by parts and obtain (see Appendix A):
W = F
n.eq.
β
1
Z
γ
dt [
ρ
t
S(ρ
t
||ω
β
(H
t
))],
(3)
where F
n.eq.
is the difference in the non
equilibrium free energy at the endpoints, and
ρ
t
S(ρ
t
||ω(H
t
)) is given by the limit
lim
ε0
S(ρ
t+ε
||ω
β
(H
t
)) S(ρ
t
||ω
β
(H
t
))
ε
. (4)
The importance of equation (3) is that it makes
manifest that we can isolate a term which only
depends on the endpoints of the protocol, and a
contribution, W
diss
(γ) = (W + F
n.eq.
), which
accounts for the dissipation during the process
and that explicitly depends on the path γ. The
quantity inside the integral can be identified with
the entropy production rate, which is guaranteed
to be positive definite whenever one has Marko-
vian dynamics (see Appendix B). This identifi-
cation is also justified by noting that the heat
supplied to the system can be rewritten as:
βQ = ∆S
Z
γ
dt [
ρ
t
S(ρ
t
||ω
β
(H
t
))], (5)
so that the integral accounts for the correction
to the second law in the presence of dissipation
(Appendix A).
When the driving of H
t
is slow (i.e. the charac-
teristic timescale of the driving is longer than the
relaxation timescales), it is sensible to assume
that the entropy production rate will only depend
on the base point H
t
and on the velocity
˙
H
t
. The
key idea of the classical papers [47] is then to
reduce the task of finding the minimally dissipat-
ing path γ
to the problem of finding the geodesic
of the metric induced by the entropy production
rate on the thermodynamic space. We will now
show how to apply this idea to open quantum
systems.
3 Metric structure in open quantum
systems
We assume that the dissipative dynamics is de-
scribed by a time-dependent Lindblad equation,
˙ρ
t
= L
t
[ρ
t
] (6)
Since a Lindbladian master equation always leads
to Markovian dynamics [28], this assumption
guarantees the positivity of the entropy produc-
tion rate. Furthermore, we assume that the
Lindbladian operator has a unique zero eigen-
state, given by the instantaneous Gibbs state
(L
t
[ω
β
(H
t
)] = 0), and that all the other eigen-
values have strictly negative real part. These two
assumptions are sufficient to ensure that given
any initial conditions the system will thermalise
to the instantenous Gibbs state of H
t
,
lim
τ→∞
e
τL
t
ρ = ω
β
(H
t
). (7)
As a final remark, note that by assuming
L
t
(ω
β
(H
t
)) = 0 we are neglecting non-adiabatic
contributions to L
t
, which is justified whenever
the bath dynamics are fast compared to the driv-
ing rate of the system Hamiltonian [2931].
For slow driving of H
t
, we can assume that
the state is always close to the equilibrium one:
ρ
t
= ω
β
(H
t
) + δω
t
. Plugging this expansion
in equation (6), we obtain an equation for δω
t
as [3234]:
L
t
d
dt
[δω
t
] =
d
dt
ω
β
(H
t
). (8)
2
It is useful to express the derivative on the right
hand side using the derivative for the exponential
of operators [35]:
de
βH
t
dt
= β
Z
1
0
ds e
βsH
˙
H
t
e
β(1s)H
, (9)
obtaining ˙ω
β
(H
t
) = βJ
ω
β
(H
t
)
[
˙
H
t
], where we de-
fined the operator:
J
ρ
[A] :=
Z
1
0
ds ρ
1s
(A Tr[ρA] ) ρ
s
. (10)
In order to solve equation (8) we need to intro-
duce the Drazin inverse of the Lindbladian oper-
ator L
+
t
L
+
t
[A] :=
Z
0
dν e
νL
t
ω
β
(H
t
)Tr [A] A
. (11)
which is the unique operator satisfying the three
conditions (see [36] and Appendix C for de-
tails): (i) commutation with the Lindbladian
(L
t
L
+
t
[A] = L
+
t
L
t
[A] = A ω
β
(H
t
)Tr [A]), (ii)
invariance of the thermal state (L
+
t
[ω
β
(H
t
)] = 0)
and (iii) tracelessness (Tr
h
L
+
t
[A]
i
= 0).
An alternative expression of L
+
t
can be con-
structed in finite dimensions as follows [32]: Con-
sider the space of traceless states, i.e., for any
state ρ
t
we consider the traceless component
given by δω
t
:= ρ
t
ω
β
(H
t
). Thanks to the sta-
tionarity of ω
β
(H
t
), equation (6) can be rewrit-
ten in the instantaneous basis as: δ ˙ω
t
= Λ
t
[δω
t
],
where Λ
t
is the projection of L
t
on the space of
traceless states. Since the only zero eigenstate of
the Lindbladian is traceful, Λ
t
is invertible on this
subspace and we can identify L
+
t
Λ
1
t
, thanks
to the uniqueness of the Drazin inverse. From
this expression it is clear that L
+
t
encodes the
thermalisation timescales of the system.
We can proceed to solve equation (8):
multiplying both sides by L
+
t
and inverting
L
+
t
d
dt
we obtain
ρ
t
=
X
n=0
L
+
t
d
dt
n
ω
β
(H
t
) =
= ω
β
(H
t
) βL
+
t
h
J
ω
β
(H
t
)
[
˙
H
t
]
i
+ . . . (12)
In the slow driving limit each
˙
H
t
is of order
O(1/T ) (recall that T is the total time of the pro-
cess), so that we can truncate the perturbative
expansion above at the second term, with cor-
rections of order O
1/T
2
. More precisely, this
Figure 1: A thermodynamic protocol in the slow driving
regime naturally corresponds to a trajectory in the space
of Gibbs states ω
β
(H). The metric introduced lives in
the tangent space of this manifold parametrized by ω
t
:=
ω
β
(H
t
).
approximation implicitly assumes that the trajec-
tory is sufficiently smooth so that derivatives of
H
t
are always bounded, and that the duration of
the protocol is much larger than the equilibration
timescales encoded in L
+
t
.
Plugging expression (12) into
ρ
t
S(ρ
t
||ω(H
t
))
gives the following formula for the dissipation (see
Appendix D):
W
diss
= β
Z
γ
dt Tr
h
˙
H
t
L
+
t
h
J
ω
β
(H
t
)
[
˙
H
t
]
ii
. (13)
It should be noticed that this expression could
be obtained directly by plugging the expansion
of the state in the definition of the work (2), but
we preferred to keep the discussion independent
of the particular expansion (12), so to empha-
sise the connection between W
diss
and the entropy
production rate.
At this point the introduction of a met-
ric structure on the thermodynamic space is
quite straightforward. Without loss of general-
ity, we decompose the system Hamiltonian as
H =
P
λ
i
t
X
i
, where
λ
i
t
are the time-dependent
externally controllable parameters, and {X
i
} are
the corresponding observables. Then, equa-
tion (13) can be rewritten as:
W
diss
= β
Z
γ
dt
˙
λ
i
t
m
L
ω
β
(H
t
)
(X
i
, X
j
)
˙
λ
j
t
(14)
where we introduced the bilinear form:
m
L
ω
(A, B) =
1
2
Tr
h
A L
+
ω
[J
ω
[B]] + B L
+
ω
[J
ω
[A]]
i
(15)
which is symmetric by construction, positive def-
inite, thanks to the positivity of the entropy pro-
duction rate, and it depends smoothly on the
3
base point ω. These are the defining properties
of a metric. Note that, while we assumed Lind-
blad master equation to ensure the positivity of
(15), the positivity also holds for more general
maps [37], for which this approach can be natu-
rally generalised.
Equation (14) can be interpreted as the energy
functional, or the action, of the curve γ({λ
t
}).
This denomination can be understood thinking
of W
diss
as the action of a system of point parti-
cles moving freely along γ({λ
t
}), with coordinate
dependent mass encoded in m
L
ω
. Then, one can
use Euler-Lagrange equations of motion [38, 39]
to construct curves in the space of parameters
which minimize the dissipation at first order in
O(1/T ). This means that the geodesics induced
by the metric (15) correspond to minimally dissi-
pative protocols in the regime of slow processes.
3.1 Connection with classical results
The connection between the entropy produc-
tion rate in the slowly driven regime and the ge-
ometry of Gibbs states has been studied for clas-
sical systems since the 70s [17]. Historically, the
formalism of thermodynamic length has been ini-
tially developed in the context of discrete per-
fectly thermalising protocols, in which case the
thermodynamic metric can be obtained as the
second derivative of the partition function with
respect to the control parameters [6, 9]. Let us
now show how to recover this metric within our
framework.
Let us assume that the parameters of the sys-
tem are changed in discrete steps from
λ
i
0
to
λ
i
N
. After each step we wait enough time for
the system to perfectly equilibrate. In the qua-
sistatic limit (N 1) we can use the approxima-
tion
H
i+1
= H
i
+ τ
˙
H
i
+ O
τ
2
, (16)
where τ is the time spent at each step, so that
the total time of the protocol is given by T = Nτ.
Using the derivative of the exponential (9) we also
have up to order O
τ
2
that:
ω
β
(H
i+1
) = ω
β
(H
i
) τβJ
ω
β
(H
i
)
[
˙
H
i
]. (17)
Comparing the expression just obtained with
the expansion (12), it is straightforward to see
that we could obtain the same result starting
from the master equation,
˙ρ
t
= τ
1
(ω
β
(H
t
) ρ
t
). (18)
Moreover, the entropy production rate for dis-
crete perfectly thermalising protocols coincides
with the one coming from this fictitious evolu-
tion (see Appendix E). This connection is a con-
sequence of the fact that for discrete protocols
details about the dynamics are disregarded, since
at each step the state completely reaches equilib-
rium. This is translated into a master equation
which has the least possible structure: in fact, the
evolution (18) for a fixed t leads to an exponen-
tial relaxation for all observables with the same
timescale τ.
In this context, we see that the metric (15) re-
duces to the Kubo-Mori-Boguliobov (KMB) inner
product
m
KMB
ω
(A, B)
τ
= cov
ω
(A, B) = Tr [A J
ω
[B]] , (19)
also known as generalised covariance in the con-
text of linear response theory [40, 41]. This quan-
tity is directly related to the partition function by
differentiation:
β
2
m
KMB
i,j
=
2
λ
i
λ
j
log Z
H
, (20)
where we indicate m
KMB
i,j
m
KMB
ω
(X
i
, X
j
),
and Z
H
is the partition function defined as:
Z
H
= Tr[e
βH
].
In the classical case the generalised covari-
ance (19) is substituted by the standard covari-
ance hABi
ω
hAi
ω
hBi
ω
[9]. Indeed, when the
protocol does not generate coherence between dif-
ferent energy levels, that is when [H
t
,
˙
H
t
] = 0,
the former reduces to the latter. Hence, it is only
in the case of non-commuting protocols, meaning
[H
t
,
˙
H
t
] 6= 0, that we expect qualitatively differ-
ent results with respect to classical systems.
Similarly to the classical case, the metric (19)
can be interpreted as a quantum Fisher infor-
mation matrix. The corresponding Riemannian
structure has been studied both from the math-
ematical and the physical point of view [16, 17,
20, 42]. In particular, (20) makes clear the re-
lation between phase transitions, statistical dis-
tinguishability and dissipation: in fact, close to a
phase transition the eigenvalues of the Fisher in-
formation matrix corresponding to the order pa-
rameters diverge. Thanks to the extreme sensi-
bility of the system, it is indeed easier to estimate
4
the values of the external parameters in this re-
gion. In the same way, the abrupt changes in the
system makes it more difficult to stay within the
quasistatic limit, so that for any N the dissipa-
tion will diverge close to the critical point.
In conclusion, the metric (15) extends the
standard thermodynamic metric (20) by includ-
ing non-trivial relaxation dynamics through the
Drazin inverse L
+
t
. This modifies the underly-
ing equilibrium geometry to account for the spe-
cific nature of the dissipative dynamics: since
observables closer to their equilibrium value dis-
sipate less when they are manipulated, the in-
troduction of L
+
t
can be interpreted as a way
to favour parameters which thermalise faster. A
similar extension to non-trivial dynamics of the
geometric framework for classical systems was ob-
tained in [12, 13] and the construction presented
here constitutes the natural generalisation to the
quantum regime. In fact, if we consider the par-
ticular case in which the controllable observables
satisfy the relaxation dynamics:
D
˙
X
i
E
ρ
t
= τ
1
i
hX
i
i
ω
t
hX
i
i
ρ
t
, (21)
then the metric takes the simple form (see Ap-
pendix F):
m
L
ω
(X
i
, X
j
) =
τ
i
+ τ
j
2
m
KMB
i,j
, (22)
in complete analogy with the classical result [12],
which gives the thermodynamic length as the
Hadamard product between a matrix encoding
the equilibration timescales of the system and the
Fisher information matrix.
4 Constructing minimally dissipating
trajectories
We now discuss how to find minimially dissipa-
tive processes from the metric (15), which is then
applied to an Ising chain in a transverse field,
which equilibrates through (18), and for a qubit
in contact with a bosonic bath.
Before starting with the calculations some gen-
eral remarks need to be made. First, it should be
noticed that, in the case of partial control over the
Hamiltonian H =
P
λ
i
t
X
i
, the metric m
i,j
will
be of the same dimension as the number of con-
trollable parameters {λ
i
t
} (see Appendix F for an
explicit expression). The optimal trajectory can
then be obtained by solving the geodesic equa-
tion:
¨
λ
i
t
+ Γ
i
j,k
λ
t
˙
λ
j
t
˙
λ
k
t
= 0, (23)
where Γ denotes the Christoffel symbols, which
can be computed from the formula:
Γ
i
j,k
|
λ
t
=
1
2
m
i,l
(
j
m
l,k
+
k
m
j,l
l
m
j,k
) |
λ
t
,
(24)
where m
i,l
is the inverse of the metric, and
we use the shorthand notation
i
m
j,k
|
λ
t
(m
j,k
/∂λ
i
)|
λ=λ
t
.
In this way, the machinery presented here pro-
vides an automatic method to obtain differential
equations for the optimal trajectory directly in
the space of parameters, which is usually of a di-
mension much smaller than the one of the full
Hilbert space. Nonetheless, it is important to
keep in mind that the computational cost still
scales with the total dimension, where the most
expensive operation is the computation of m
i,j
itself.
In the case of the master equation (18), there
are some important simplifications coming from
the fact that the KMB metric is directly related
to the partition function. In fact, one can obtain
simply by differentiation:
β
3
Γ
i
j,k
|
λ
t
=
1
2
m
i,l
3
λ
j
λ
k
λ
l
log Z
H
!
. (25)
This makes this approach particularly simple
when an analytical expression for Z
H
is known,
as we will illustrate in the following example.
4.1 Ising chain in a transverse field
We now consider the system to be given by an
Ising chain in a transverse field, which equili-
brates through (18). The Hamiltonian of the sys-
tem is given by:
H
I
= J
n
X
i=1
σ
z
i
σ
z
i+1
+ gσ
x
i
(26)
where J > 0 set a measure for the energy scale
and g is a dimensionless coupling parameter,
which can be interpreted as an external mag-
netic field. For simplicity, in the following we
will measure times in units of τ and energies
in units of J. The spectrum of the system can
5
be computed analytically for periodic boundary
conditions through a Jordan-Wigner transforma-
tion, and in the thermodynamic limit the parti-
tion function is given by the integral expression
(see e.g. [43])
lim
n→∞
1
N
log Z =
Z
2π
0
dk log
2 cosh
β
ε
k
2

,
(27)
where ε
k
is the eigenvalue corresponding to the
momentum k, which reads:
ε
k
= 2J
q
1 + g
2
2g cos k. (28)
This system presents a phase transition at zero
temperature and g = 1 from an ordered ferromag-
netic phase to a quantum paramagnetic phase.
Moreover, at finite but low enough temperatures,
there is a rich variety of different physical regimes
(summarised in Fig. 2), characterised by different
scaling for the correlation length and the equili-
bration timescale of the system (see e.g. [43] for
details).
We study here the case in which one has control
only over g, so that the metric and the Christoffel
symbols become a scalar, which we name m and
Γ, respectively. In particular, we see that it is
sufficient to differentiate (27) with respect to g
twice, to obtain:
m(g) =
Z
2π
0
dk
¨ε
k
2
tanh
β
ε
k
2
+
˙ε
k
2
sech
β
ε
k
2

2
!
,
(29)
Similarly, we can compute Γ using (25). Both
m and Γ are shown in Fig. 2 for different tem-
peratures. Finally, we also compute numerically
optimal thermodynamic processes through the
geodesic equations (23). The results are also
shown in Fig. 2.
Since the metric is connected with the free
energy of the system, it is interesting to notice
how the Riemannian structure is affected by the
presence of a phase transition at zero tempera-
ture. Comparing the metric and, in particular,
the Christoffel symbol with the phase diagram of
the system, we can see that the change in be-
haviour of these geometric quantities retrace a
change in the underlying physical properties. Ad-
ditionally, note that m increases close to g = 1 as
Figure 2: The top two figures show the metric and the
Christoffel symbol for the Ising chain at different temper-
atures as a function of the coupling g. On the bottom,
on the left we present the behaviour of minimally dis-
sipating trajectories defined by the boundary conditions
g(0) = 0 to g(1) = 5; on the right, the phase diagram
of the system.
the temperature is decreased, illustrating how the
dissipation increases in the presence of a phase
transition. This behaviour reflects in the shape of
the geodesics, as
˙
H
t
decreases close to the phase
transition in order to compensate for the larger
dissipation.
4.2 Qubit in contact with a bosonic bath
We now move to treat an example in which L
+
t
is
non-trivial, a two level system in contact with a
bosonic bath with spectral density J(ω) = γ
0
ω
α
.
The Lindbladian we use can be obtained through
a microscopic derivation [44]; more details and
the particular form of the dynamics are given in
appendix G. The parameter α characterises the
ohmicity of the environment, and we have that
for α = 1, α > 1, α < 1, the bath is ohmic,
superohmic and subohmic, respectively. We as-
sume full control on the Hamiltonian of the two-
level system, which is parametrised by spherical
coordinates
H = r cos ϕ sin θ ˆσ
x
+ r sin ϕ sin θ ˆσ
y
+ r cos θ ˆσ
z
,
(30)
where (r, θ, ϕ) are the control parameters.
Thanks to the convenient choice of coordinates,
the metric takes the particularly simple form (see
Appendix G)
m
L
=
1
r
α
diag
n
λ
d
, λ
q
r
2
, λ
q
r
2
sin
2
θ
o
, (31)
6
Figure 3: From left to right, from top to bottom: de-
pendence of the classical and quantum component of
the metric λ
d,q
on the energy spacing r; geodesics in
the radial direction (r(0) = 0.1, r(1) = 5) for con-
stant θ and φ; geodesics in Cartesian coordinates for
non commuting Hamiltonians ({x(0), z(0)} = {0, 1},
{x(1), z(1)} = {1, 1}) and y 0; comparison of the
first order contribution to the dissipation during a linear
protocol with respect to the optimal one as a function of
the energy spacing. The energies are measured in units
of β.
where we can identify the expression of
the Euclidean metric in spherical coordinates
(diag{1, r
2
, r
2
sin
2
(θ)}), and the two eigenvalues
are given by:
λ
d
=
tanh(r)
cosh
2
(r)
, λ
q
=
2 tanh
2
(r)
r
. (32)
The existence of different eigenvalues for the ra-
dial direction and the solid angle reflects the
physical fact that the system dissipates differently
if only the energy spacing is moved, or if coher-
ence between the two levels is created. Moreover,
it is straightforward to verify that the ratio λ
q
d
diverges for r . This illustrates a general be-
haviour: the existence of exponentially more dis-
sipative parameters in thermodynamic systems.
In this case, there is a simple physical explana-
tion: changing the energy spacing only affects
an exponentially small fraction of the population,
whereas the whole system has to be manipulated
in order to create coherence. This example how-
ever illustrates the fact that inspecting the metric
eigenvalues provides us with a powerful descrip-
tion of the physics of dissipation, which is crucial
in more complex systems where a simple intuitive
understanding is out of reach.
The ohmicity of the bath contributes to the
metric with a factor r
α
, which is the overall de-
pendence of the equilibration timescales on the
energy spacing r. Owing to this dependence, op-
timal protocols will tend to spend more time in
the region with low r. This effect is illustrated in
the right top panel of Fig. 3, where we construct
optimal trajectories in the radial direction. As ex-
pected, this effect is more pronounced for higher
αs.
Our results also allow for constructing optimal
trajectories between non-commuting endpoints.
Then we necessarily have that [
˙
H
t
, H
t
] 6= 0,
so that quantum coherence between energy lev-
els is created along the protocol. Optimal non-
commuting trajectories are shown in the bottom
left Figure of 3. Interestingly, we see that the
ohmicity of the bath qualitatively changes the be-
haviour of the optimal trajectories.
In order to illustrate the improvement obtained
by using geodesic trajectories, we compare them
to a naive choice where the parameters are lin-
early modified in time. We consider boundary
conditions H
A
= 0 and H
B
= E
f
σ
z
(energy is
measured in units of β). Restricting to trivial
dynamics (18), which allows for a simple analytic
solution, we find at first order in O(1/T ):
W
lin
dis
= E
f
tanh E
f
, (33)
W
KMB
dis
=
1
4
π 2 tan
1
(csch(E
f
))
2
. (34)
In the limit of E
f
1 the first term diverges
linearly in E
f
, while W
KMB
dis
π
2
/4. This dif-
ference makes clear how for big energy gaps or,
equivalently, at low temperature, using a geodesic
trajectory gives an increasingly bigger advantage,
making the use of optimal trajectories particu-
larly relevant in the quantum regime. These con-
siderations are also relevant for experiments, as
current demonstrations of the Landauer principle
rely on a linear increase of E(t) [4548].
5 Comparison to other approaches
Whereas the presented results are valid in the
slow driving limit, exact solutions for minimis-
ing dissipation also exist in the literature [49
52]. In particular, optimal finite-time protocols
for two-level systems have been treated in a va-
riety of settings [5054]. Besides specific solvable
systems, a general approach for minimising dissi-
pation based on optimal control theory has been
7
put forward in [51, 52, 54], which requires solving
a system of D non-linear differential equations,
where D is the Hilbert space dimension (see [51]
and Appendix H).
Conversely, the geometric approach provides
an approximation of the optimal solution (which
becomes exact in the slow driving limit) by a
set of d non-linear differential equations, where
d is the number of controllable parameters in the
Hamiltonian (see Eq. (23)). This makes this ap-
proach particularly useful in more complex sys-
tems, where d D, as we have illustrated in the
Ising chain example where d = 1 and D .
This gives the geometric approach a wide range
of applicability, which has also been illustrated in
classical systems of different nature and complex-
ity [47, 1315, 5558]. It is however important
to keep in mind that in general building the met-
ric (15) requires diagonalising the Hamiltonian of
the system of interest, as discussed in Appendix
F. A class of dynamics where the computation of
the metric notably simplifies is when the expected
value of each externally controlled observable (i.e.
hX
i
i
ρ
t
)) relaxes to its thermal expectation value
with a well defined time scale τ
i
as in (21); as
in this case the corresponding metric (22) can be
directly computed from the partition function.
Besides providing an efficient way of finding
minimally dissipative paths, there are other ad-
vantages of working in the quasistatic regime:
1. By construction, the protocols developed are
independent of the total time T . That is, one
does not need to optimise the path for every
duration of the protocol.
2. When dealing with small systems, it is not
only important to minimise the dissipated
work, but also to minimise fluctuations. In
the quasistatic limit, minimising dissipation
guarantees the minimisation of fluctuations
for classical systems [9], and for commuting
protocols in the quantum regime [59].
3. The introduction of a Riemannian structure
in the space of parameters provides an auto-
matic way to infer the dissipative properties
of the system, just by inspecting the eigen-
values of the metric.
Finally, in Appendix H we also qualitatively
compare the geometric approach with exact ap-
proaches for a two-level system finding good
agreement between them even when the total
time T of the protocol is similar to the time scale
of relaxation, see also [26].
6 Conclusions and outlook
We developed a general framework to define a
thermodynamic metric for systems that evolve
under a Lindblad master equation, in such a
way that geodesics correspond to minimally-
dissipative paths in the quasistatic regime. The
connection with the non-equilibrium free energy
makes it extremely versatile and an extension
to more general settings seems to be foreseeable
with only minor modifications. For example, it
should be possible to extend our approach to gen-
eralised Gibbs ensembles [6063] simply by re-
defining the non equilibrium free energy (1) so
to take into account the extra conserved quan-
tities. We can even consider a dynamics which
equilibrates to a non equilibrium steady states
by simply substituting the thermal state in the
definition (4) of the entropy production rate with
the appropriate steady state π(H) [33]. For non-
equilibrium steady states, an interesting future
direction is to consider relations between this ge-
ometric approach and quantum thermodynamic
uncertainty relations [42, 64]. Moreover, despite
the fact that we assume a Lindblad master equa-
tion, and therefore weak coupling, thanks to the
reaction-coordinate mapping, an extension to the
strong coupling regime is possible [6572]. In this
context, since the spectral density of the bath is
modified under the mapping, we expect that in
the strong coupling regime there will be qualita-
tively changes in the behaviour of optimal tra-
jectories, similarly to what we showed in Fig. 3.
Similarly, the same framework can also be applied
to other types of thermalisation such as in colli-
sional models [73]. Another interesting research
direction is to investigate the connection between
optimal thermodynamic processes and shortcuts
to thermalisation [74]. Overall, it seems that the
introduction of a metric structure on the space of
thermodynamic states is not only natural, but it
is worth of more investigation in the future.
Acknowledgements. We thank G. Crooks, S.
Deffner, M. Ueda, and V. Cavina for insightful
8
comments. M.P.-L. acknowledges support from
the Alexander von Humboldt Foundation. This
project has received funding from the European
Union’s Horizon 2020 research and innovation
programme under the Marie Skłodowska-Curie
grant agreement No 713729, and from Span-
ish MINECO (QIBEQI FIS2016-80773-P, Severo
Ochoa SEV-2015-0522), Fundacio Cellex, Gener-
alitat de Catalunya (SGR 1381 and CERCA Pro-
gramme). This research was also supported in
part by the National Science Foundation under
Grant No. NSF PHY-1748958.
References
[1] F. Weinhold. Metric geometry of equilibrium
thermodynamics. The Journal of Chemical
Physics, 63:2479, 1975. URL https://doi.
org/10.1063/1.431689.
[2] F. Weinhold. Metric geometry of equilib-
rium thermodynamics. III. Elementary for-
mal structure of a vector-algebraic represen-
tation of equilibrium thermodynamics. The
Journal of Chemical Physics, 63(6):2488–
2495, 1975. DOI: 10.1063/1.431636. URL
https://doi.org/10.1063/1.431636.
[3] F. Schlögl. Thermodynamic metric
and stochastic measures. Zeitschrift
für Physik B Condensed Matter, 59
(4):449–454, Dec 1985. ISSN 1431-
584X. DOI: 10.1007/BF01328857. URL
https://doi.org/10.1007/BF01328857.
[4] P. Salamon and R. S. Berry. Thermody-
namic length and dissipated availability.
Phys. Rev. Lett., 51:1127–1130, Sep 1983.
DOI: 10.1103/PhysRevLett.51.1127. URL
https://link.aps.org/doi/10.1103/
PhysRevLett.51.1127.
[5] P. Salamon, B. Andresen, P. D. Gait, and
R. S. Berry. The significance of Weinhold’s
length. The Journal of Chemical Physics,
73(2):1001–1002, 1980. URL https://doi.
org/10.1063/1.440217.
[6] J. Nulton, P. Salamon, B. Andresen, and
Q. Anmin. Quasistatic processes as step
equilibrations. The Journal of Chemical
Physics, 83:334, 1985. URL https://doi.
org/10.1063/1.449774.
[7] B. Andresen, R. S. Berry, R. Gilmore,
E. Ihrig, and P. Salamon. Thermody-
namic geometry and the metrics of Wein-
hold and Gilmore. Phys. Rev. A, 37:
845–848, Feb 1988. DOI: 10.1103/Phys-
RevA.37.845. URL https://link.aps.
org/doi/10.1103/PhysRevA.37.845.
[8] L. Diosi, K. Kulacsy, B. Lukacs, and
A. Racz. Thermodynamic length, time,
speed, and optimum path to minimize en-
tropy production. The Journal of chemical
physics, 105(24):11220–11225, 1996. URL
https://aip.scitation.org/doi/abs/
10.1063/1.472897.
[9] G. E. Crooks. Measuring ther-
modynamic length. Phys. Rev.
Lett., 99:100602, Sep 2007. DOI:
10.1103/PhysRevLett.99.100602. URL
https://link.aps.org/doi/10.1103/
PhysRevLett.99.100602.
[10] P. R. Zulkowski, D. A. Sivak, G. E. Crooks,
and M. R. DeWeese. Geometry of ther-
modynamic control. Phys. Rev. E, 86:
041148, Oct 2012. DOI: 10.1103/Phys-
RevE.86.041148. URL https://link.aps.
org/doi/10.1103/PhysRevE.86.041148.
[11] E. H. Feng and G. E. Crooks. Far-
from-equilibrium measurements of thermo-
dynamic length. Phys. Rev. E, 79:
012104, Jan 2009. DOI: 10.1103/Phys-
RevE.79.012104. URL https://link.aps.
org/doi/10.1103/PhysRevE.79.012104.
[12] D. A. Sivak and G. E. Crooks. Thermo-
dynamic metrics and optimal paths. Phys.
Rev. Lett., 108:190602, May 2012. DOI:
10.1103/PhysRevLett.108.190602. URL
https://link.aps.org/doi/10.1103/
PhysRevLett.108.190602.
[13] D. A. Sivak and G. E. Crooks. Thermo-
dynamic geometry of minimum-dissipation
driven barrier crossing. Phys. Rev. E, 94:
052106, Nov 2016. DOI: 10.1103/Phys-
RevE.94.052106. URL https://link.aps.
org/doi/10.1103/PhysRevE.94.052106.
[14] P. R. Zulkowski, D. A. Sivak, and M. R.
DeWeese. Optimal control of transitions
between nonequilibrium steady states. PloS
one, 8(12):e82754, 2013. URL https:
//journals.plos.org/plosone/article?
id=10.1371/journal.pone.0082754.
[15] P. R. Zulkowski and M. R. DeWeese.
Optimal control of overdamped sys-
tems. Phys. Rev. E, 92:032117, Sep 2015.
DOI: 10.1103/PhysRevE.92.032117. URL
9
https://link.aps.org/doi/10.1103/
PhysRevE.92.032117.
[16] D. Petz and G. Toth. The Bogoliubov in-
ner product in quantum statistics. Let-
ters in Mathematical Physics, 27(3):205–
216, Mar 1993. ISSN 1573-0530. DOI:
10.1007/BF00739578. URL https://doi.
org/10.1007/BF00739578.
[17] P. W. Michor, D. Petz, and A. Andai.
On the curvature of a certain Rieman-
nian space of matrices. Infinite Dimen-
sional Analysis, Quantum Probability and
Related Topics, 03(02):199–212, 2000. DOI:
10.1142/S0219025700000145. URL https:
//doi.org/10.1142/S0219025700000145.
[18] D. Petz. Covariance and Fisher informa-
tion in quantum mechanics. Journal of
Physics A: Mathematical and General, 35(4):
929, 2002. URL http://stacks.iop.org/
0305-4470/35/i=4/a=305.
[19] D. Petz and C. Ghinea. Introduction
to quantum Fisher information, pages
261–281. World Scientific, 2011. DOI:
10.1142/9789814338745_0015. URL
https://www.worldscientific.com/doi/
abs/10.1142/9789814338745_0015.
[20] R. Balian. The entropy-based quan-
tum metric. Entropy, 16(7):3878–
3888, 2014. ISSN 1099-4300. DOI:
10.3390/e16073878. URL http:
//www.mdpi.com/1099-4300/16/7/3878.
[21] S. Deffner and E. Lutz. Generalized
clausius inequality for nonequilib-
rium quantum processes. Phys. Rev.
Lett., 105:170402, Oct 2010. DOI:
10.1103/PhysRevLett.105.170402. URL
https://link.aps.org/doi/10.1103/
PhysRevLett.105.170402.
[22] S. Deffner and E. Lutz. Thermo-
dynamic length for far-from-equilibrium
quantum systems. Phys. Rev. E, 87:
022143, Feb 2013. DOI: 10.1103/Phys-
RevE.87.022143. URL https://link.aps.
org/doi/10.1103/PhysRevE.87.022143.
[23] M. Campisi, S. Denisov, and P. Hänggi.
Geometric magnetism in open quantum
systems. Phys. Rev. A, 86:032114, Sep 2012.
DOI: 10.1103/PhysRevA.86.032114. URL
https://link.aps.org/doi/10.1103/
PhysRevA.86.032114.
[24] T. V. Acconcia, M. V. S. Bonança,
and S. Deffner. Shortcuts to adia-
baticity from linear response theory.
Physical Review E, 92(4):042148, 2015.
URL https://journals.aps.org/pre/
abstract/10.1103/PhysRevE.92.042148.
[25] M. F. Ludovico, F. Battista, F. von Op-
pen, and L. Arrachea. Adiabatic re-
sponse and quantum thermoelectrics for ac-
driven quantum systems. Phys. Rev. B,
93:075136, Feb 2016. DOI: 10.1103/Phys-
RevB.93.075136. URL https://link.aps.
org/doi/10.1103/PhysRevB.93.075136.
[26] M. V. S. Bonança and S. Deffner. Minimal
dissipation in processes far from equilib-
rium. Phys. Rev. E, 98:042103, Oct 2018.
DOI: 10.1103/PhysRevE.98.042103. URL
https://link.aps.org/doi/10.1103/
PhysRevE.98.042103.
[27] Patrick R. Zulkowski and Michael R. De-
Weese. Optimal protocols for slowly driven
quantum systems. Phys. Rev. E, 92:
032113, Sep 2015. DOI: 10.1103/Phys-
RevE.92.032113. URL https://link.aps.
org/doi/10.1103/PhysRevE.92.032113.
[28] V. Gorini, A. Kossakowski, and E. C. G.
Sudarshan. Completely positive dynamical
semigroups of n-level systems. Journal
of Mathematical Physics, 17(5):821–825,
1976. DOI: 10.1063/1.522979. URL
https://aip.scitation.org/doi/abs/
10.1063/1.522979.
[29] Tameem Albash, Sergio Boixo, Daniel A
Lidar, and Paolo Zanardi. Quantum
adiabatic markovian master equations.
New Journal of Physics, 14(12):123016,
December 2012. DOI: 10.1088/1367-
2630/14/12/123016. URL https://doi.
org/10.1088/1367-2630/14/12/123016.
[30] Makoto Yamaguchi, Tatsuro Yuge, and
Tetsuo Ogawa. Markovian quantum master
equation beyond adiabatic regime. Phys.
Rev. E, 95:012136, Jan 2017. DOI:
10.1103/PhysRevE.95.012136. URL
https://link.aps.org/doi/10.1103/
PhysRevE.95.012136.
[31] Roie Dann, Amikam Levy, and Ronnie
Kosloff. Time-dependent markovian quan-
tum master equation. Phys. Rev. A, 98:
052129, Nov 2018. DOI: 10.1103/Phys-
RevA.98.052129. URL https://link.aps.
org/doi/10.1103/PhysRevA.98.052129.
10
[32] V. Cavina, A. Mari, and V. Giovan-
netti. Slow dynamics and thermodynamics
of open quantum systems. Phys. Rev.
Lett., 119:050601, Aug 2017. DOI:
10.1103/PhysRevLett.119.050601. URL
https://link.aps.org/doi/10.1103/
PhysRevLett.119.050601.
[33] D. Mandal and C. Jarzynski. Analysis
of slow transitions between nonequilib-
rium steady states. Journal of Statis-
tical Mechanics: Theory and Experi-
ment, 2016(6):063204, jun 2016. DOI:
10.1088/1742-5468/2016/06/063204. URL
https://doi.org/10.1088/1742-5468/
2016/06/063204.
[34] Gavin E Crooks. On the Drazin inverse of
the rate matrix. 2018.
[35] F. Hiai and D. Petz. Introduction to Ma-
trix Analysis and Applications. Springer In-
ternational Publishing, Cham, 2014. ISBN
978-3-319-04150-6. DOI: 10.1007/978-3-319-
04150-6_3. URL https://doi.org/10.
1007/978-3-319-04150-6_3.
[36] T. L. Boullion and P. L. Odell. Generalised
inverse matrices. Wiley-Interscience, New
York, 1971.
[37] A. Müller-Hermes and D. Reeb. Mono-
tonicity of the quantum relative entropy
under positive maps. Annales Henri
Poincaré, 18(5):1777–1788, jan 2017. DOI:
10.1007/s00023-017-0550-9. URL https://
doi.org/10.1007/s00023-017-0550-9.
[38] B. O’Neill. Semi-Riemannian geometry with
applications to relativity. Pure and Applied
Mathematics. Elsevier Science, 1983. ISBN
9780080570570.
[39] V. I. Arnold. Lagrangian mechanics on
manifolds, pages 75–97. Springer New
York, New York, NY, 1989. ISBN 978-1-
4757-2063-1. DOI: 10.1007/978-1-4757-2063-
1_4. URL https://doi.org/10.1007/
978-1-4757-2063-1_4.
[40] F. J. Dyson, E. H. Lieb, and B. Simon.
Phase transitions in quantum spin systems
with isotropic and nonisotropic interactions.
Journal of Statistical Physics, 18(4):335–
383, Apr 1978. ISSN 1572-9613. DOI:
10.1007/BF01106729. URL https://doi.
org/10.1007/BF01106729.
[41] G. Roepstorff. Correlation inequali-
ties in quantum statistical mechanics and
their application in the Kondo prob-
lem. Comm. Math. Phys., 46(3):253–262,
1976. URL https://projecteuclid.org:
443/euclid.cmp/1103899639.
[42] G. Guarnieri, G. T. Landi, S. R. Clark,
and J. Goold. Thermodynamics of precision
in quantum non equilibrium steady states.
arXiv preprint arXiv:1901.10428, 2019.
[43] S. Sachdev. Quantum phase transitions.
Handbook of Magnetism and Advanced Mag-
netic Materials, 2007.
[44] H. P. Breuer and F. Petruccione. The theory
of open quantum systems. Oxford University
Press, Great Clarendon Street, 2002.
[45] A. Bérut, A. Arakelyan, A. Petrosyan,
S. Ciliberto, R. Dillenschneider, and E. Lutz.
Experimental verification of Landauer’s
principle linking information and thermo-
dynamics. Nature, 483(7388):187, 2012.
URL https://www.nature.com/articles/
nature10872.
[46] Y. Jun, M. Gavrilov, and J. Bechhoe-
fer. High-precision test of Landauer’s
principle in a feedback trap. Phys.
Rev. Lett., 113:190601, Nov 2014. DOI:
10.1103/PhysRevLett.113.190601. URL
https://link.aps.org/doi/10.1103/
PhysRevLett.113.190601.
[47] J. V. Koski, V. F. Maisi, J. P. Pekola,
and D. V. Averin. Experimental realiza-
tion of a Szilard engine with a single elec-
tron. Proceedings of the National Academy
of Sciences, 111(38):13786–13789, 2014.
URL http://www.pnas.org/content/111/
38/13786.
[48] R. Gaudenzi, E. Burzurí, S. Maegawa,
H. S. J. Zant, and F. Luis. Quantum
Landauer erasure with a molecular nano-
magnet. Nature Physics, 14(6):565, 2018.
URL https://www.nature.com/articles/
s41567-018-0070-7.
[49] T. Schmiedl and U. Seifert. Optimal finite-
time processes in stochastic thermodynam-
ics. Phys. Rev. Lett., 98:108301, Mar 2007.
DOI: 10.1103/PhysRevLett.98.108301. URL
https://link.aps.org/doi/10.1103/
PhysRevLett.98.108301.
[50] M. Esposito, R. Kawai, K. Lindenberg, and
C. Van den Broeck. Finite-time thermo-
dynamics for a single-level quantum dot.
EPL (Europhysics Letters), 89(2):20003,
11
2010. URL http://iopscience.iop.org/
article/10.1209/0295-5075/89/20003.
[51] V. Cavina, A. Mari, A. Carlini, and V. Gio-
vannetti. Optimal thermodynamic control
in open quantum systems. Phys. Rev. A,
98:012139, Jul 2018. DOI: 10.1103/Phys-
RevA.98.012139. URL https://link.aps.
org/doi/10.1103/PhysRevA.98.012139.
[52] Paul Menczel, Tuomas Pyhäranta, Chris-
tian Flindt, and Kay Brandner. Two-stroke
optimization scheme for mesoscopic refrig-
erators. Phys. Rev. B, 99:224306, Jun 2019.
DOI: 10.1103/PhysRevB.99.224306. URL
https://link.aps.org/doi/10.1103/
PhysRevB.99.224306.
[53] S. Deffner. Optimal control of a qubit in
an optical cavity. Journal of Physics
B: Atomic, Molecular and Optical
Physics, 47(14):145502, 2014. URL
http://iopscience.iop.org/article/
10.1088/0953-4075/47/14/145502/meta.
[54] Vasco Cavina, Andrea Mari, Alberto Car-
lini, and Vittorio Giovannetti. Vari-
ational approach to the optimal con-
trol of coherently driven, open quantum
system dynamics. Phys. Rev. A, 98:
052125, Nov 2018. DOI: 10.1103/Phys-
RevA.98.052125. URL https://link.aps.
org/doi/10.1103/PhysRevA.98.052125.
[55] M. V. S. Bonança and S. Deffner. Opti-
mal driving of isothermal processes close
to equilibrium. The Journal of chemi-
cal physics, 140(24):244119, 2014. URL
https://aip.scitation.org/doi/abs/
10.1063/1.4885277?journalCode=jcp.
[56] G. M. Rotskoff and G. E. Crooks. Op-
timal control in nonequilibrium systems:
dynamic Riemannian geometry of the Ising
model. Phys. Rev. E, 92:060102, Dec 2015.
DOI: 10.1103/PhysRevE.92.060102. URL
https://link.aps.org/doi/10.1103/
PhysRevE.92.060102.
[57] T. R. Gingrich, G. M. Rotskoff, G. E.
Crooks, and P. L. Geissler. Near-
optimal protocols in complex nonequi-
librium transformations. Proceedings
of the National Academy of Sciences,
113(37):10263–10268, aug 2016. DOI:
10.1073/pnas.1606273113. URL https:
//doi.org/10.1073/pnas.1606273113.
[58] G. M. Rotskoff, G. E. Crooks, and
E. Vanden-Eijnden. Geometric approach
to optimal nonequilibrium control: min-
imizing dissipation in nanomagnetic spin
systems. Phys. Rev. E, 95:012148, Jan 2017.
DOI: 10.1103/PhysRevE.95.012148. URL
https://link.aps.org/doi/10.1103/
PhysRevE.95.012148.
[59] Harry JD Miller, Matteo Scandi, Janet An-
ders, and Martí Perarnau-Llobet. Work fluc-
tuations in slow processes: quantum signa-
tures and optimal control. arXiv preprint
arXiv:1905.07328, 2019.
[60] Y. Guryanova, S. Popescu, A. J. Short,
R. Silva, and P. Skrzypczyk. Thermodynam-
ics of quantum systems with multiple con-
served quantities. Nature communications,
7:ncomms12049, 2016. URL https://www.
nature.com/articles/ncomms12049/.
[61] M. Lostaglio, D. Jennings, and T. Rudolph.
Thermodynamic resource theories,
non-commutativity and maximum en-
tropy principles. New Journal of
Physics, 19(4):043008, 2017. URL
http://iopscience.iop.org/article/
10.1088/1367-2630/aa617f/meta.
[62] N. Y. Halpern, P. Faist, J. Oppenheim, and
A. Winter. Microcanonical and resource-
theoretic derivations of the thermal state
of a quantum system with noncommut-
ing charges. Nature communications, 7:
12051, 2016. URL https://www.nature.
com/articles/ncomms12051.
[63] M. Perarnau-Llobet, A. Riera, R. Gal-
lego, H. Wilming, and J. Eisert. Work
and entropy production in generalised
Gibbs ensembles. New Journal of
Physics, 18(12):123035, 2016. URL
http://iopscience.iop.org/article/
10.1088/1367-2630/aa4fa6/meta.
[64] André M. Timpanaro, Giacomo Guarnieri,
John Goold, and Gabriel T. Landi. Ther-
modynamic uncertainty relations from
exchange fluctuation theorems. Phys.
Rev. Lett., 123:090604, Aug 2019. DOI:
10.1103/PhysRevLett.123.090604. URL
https://link.aps.org/doi/10.1103/
PhysRevLett.123.090604.
[65] M. F. Gelin and M. Thoss. Thermody-
namics of a subensemble of a canonical
ensemble. Phys. Rev. E, 79:051121, May
2009. DOI: 10.1103/PhysRevE.79.051121.
12
URL https://link.aps.org/doi/10.
1103/PhysRevE.79.051121.
[66] M. Campisi, P. Talkner, and P. Hänggi.
Fluctuation theorem for arbitrary
open quantum systems. Phys. Rev.
Lett., 102:210401, May 2009. DOI:
10.1103/PhysRevLett.102.210401. URL
https://link.aps.org/doi/10.1103/
PhysRevLett.102.210401.
[67] S. Hilt, B. Thomas, and E. Lutz. Hamil-
tonian of mean force for damped quantum
systems. Phys. Rev. E, 84:031110, Sep 2011.
DOI: 10.1103/PhysRevE.84.031110. URL
https://link.aps.org/doi/10.1103/
PhysRevE.84.031110.
[68] R Gallego, A Riera, and J Eisert. Ther-
mal machines beyond the weak coupling
regime. New Journal of Physics, 16(12):
125009, 2014. URL http://stacks.iop.
org/1367-2630/16/i=12/a=125009.
[69] D. Gelbwaser-Klimovsky and A. Aspuru-
Guzik. Strongly coupled quantum heat
machines. The journal of physical chem-
istry letters, 6(17):3477–3482, 2015. URL
https://pubs.acs.org/doi/abs/10.
1021/acs.jpclett.5b01404.
[70] P. Strasberg, G. Schaller, N. Lambert,
and T. Brandes. Nonequilibrium ther-
modynamics in the strong coupling and
non-Markovian regime based on a reac-
tion coordinate mapping. New Journal
of Physics, 18(7):073007, 2016. URL
http://iopscience.iop.org/article/
10.1088/1367-2630/18/7/073007/meta.
[71] D. Newman, F. Mintert, and A. Nazir.
Performance of a quantum heat engine at
strong reservoir coupling. Phys. Rev. E,
95:032139, Mar 2017. DOI: 10.1103/Phys-
RevE.95.032139. URL https://link.aps.
org/doi/10.1103/PhysRevE.95.032139.
[72] M. Perarnau-Llobet, H. Wilming, A. Riera,
R. Gallego, and J. Eisert. Strong coupling
corrections in quantum thermodynamics.
Phys. Rev. Lett., 120:120602, Mar 2018.
DOI: 10.1103/PhysRevLett.120.120602.
URL https://link.aps.org/doi/10.
1103/PhysRevLett.120.120602.
[73] Elisa Bäumer, Martí Perarnau-Llobet,
Philipp Kammerlander, Henrik Wilm-
ing, and Renato Renner. Imperfect
Thermalizations Allow for Optimal Ther-
modynamic Processes. Quantum, 3:
153, June 2019. ISSN 2521-327X. DOI:
10.22331/q-2019-06-24-153. URL https:
//doi.org/10.22331/q-2019-06-24-153.
[74] Roie Dann, Ander Tobalina, and Ron-
nie Kosloff. Shortcut to equilibration
of an open quantum system. Phys.
Rev. Lett., 122:250402, Jun 2019. DOI:
10.1103/PhysRevLett.122.250402. URL
https://link.aps.org/doi/10.1103/
PhysRevLett.122.250402.
13
A Non equilibrium free energy and work extraction
In this section we give the details on how to derive equation (1) and (3) of the main text. First, it is
straightforward to show, using the defining formula F (ρ, H) = hHi
ρ
β
1
S(ρ), that the free energy
of a thermal state ω
β
(H) is given by
F (ω, H) = hHi
ρ
β
1
(β hHi
ρ
+ log Z
H
) = β
1
log Z
H
(35)
in complete analogy with classical thermodynamics. Then, equation (1) is obtained by direct compu-
tation, rewriting the relative entropy as:
β
1
S(ρ||ω
β
(H)) = β
1
Tr [ρ(log ρ log ω
β
(H))] =
=
β
1
S(ρ) + hHi
ρ
+ β
1
log Z
H
=
= F (ρ, H) F (ω
β
(H), H), (36)
which gives the result after rearranging the terms. For what regards equation (3), it is useful to first
give the expression of the total derivative of the non equilibrium free energy along a trajectory (ρ
t
, H
t
).
This can be decomposed in the two terms:
d
dt
F (ρ
t
, H
t
) = (Tr [ ˙ρ
t
H
t
] + β
1
d
dt
Tr [ρ
t
log ρ
t
]) + Tr[ρ
t
˙
H
t
]. (37)
The last trace can be rewritten as:
Tr[ρ
t
˙
H
t
] = lim
ε0
F (ρ
t
, H
t+ε
) F (ρ
t
, H
t
)
ε
=:
H
t
F (ρ
t
, H
t
). (38)
For what regards the term inside the parenthesis, it should be first noticed that we can rewrite:
Tr [ ˙ρ
t
H
t
] = β
1
(Tr
h
˙ρ
t
log e
βH
t
i
log Z
H
Tr [ ˙ρ
t
]) =
= β
1
Tr [ ˙ρ
t
log ω
β
(H
t
)] , (39)
where in the first line we use the fact that the derivative of a state is traceless. Then, we obtain for
the whole expression:
d
dt
Tr [ρ
t
log ρ
t
] Tr [ ˙ρ
t
log ω
β
(H
t
)] = lim
ε0
S(ρ
t+ε
||ω
β
(H
t
)) S(ρ
t
||ω
β
(H
t
))
ε
=:
ρ
t
S(ρ
t
||ω
β
(H
t
)), (40)
where we recognise the quantity defined in (4). In this way, equation (37) can be expressed in the
compact form:
d
dt
F (ρ
t
, H
t
) =
H
t
F (ρ
t
, H
t
) + β
1
ρ
t
S(ρ
t
||ω
β
(H
t
)) (41)
This equality can be used to obtain equation (3). In fact, starting from the definition of the work
in (2), a simple integration by parts gives:
W =
Z
γ
dt Tr
h
ρ
t
˙
H
t
i
=
Z
γ
dt
H
t
F (ρ
t
, H
t
) =
= F
n.eq.
β
1
Z
γ
dt [
ρ
t
S(ρ
t
||ω
β
(H
t
))] (42)
This concludes the derivation of (3).
Finally, one can obtain equation (5) simply by using the first law of thermodynamics, plugging in
the definition of F
n.eq.
:
Q = ∆U + W = ∆U F
n.eq.
β
1
Z
γ
dt [
ρ
t
S(ρ
t
||ω
β
(H
t
))]
=
U
U + β
1
S β
1
Z
γ
dt [
ρ
t
S(ρ
t
||ω
β
(H
t
))]. (43)
Multiplying both sides by β, we obtain (5).
14
B The entropy production rate
We now study the main features of the entropy production rate. Equation (42) suggests that the
definition of the instantaneous entropy production rate is [44]:
˙σ(ρ
t
) :=
ρ
t
S(ρ
t
||ω
β
(H
t
)) = lim
ε0
S(ρ
t+ε
||ω
β
(H
t
)) S(ρ
t
||ω
β
(H
t
))
ε
. (44)
Before investigating the positivity of this quantity, it is useful to show a weaker equation for ρ
t
which
are induced by CPTP maps in the case in which there is no driving. In particular, assuming the system
to be initially uncorrelated with the environment (ρ
UN IV
(0) = ρ
S
0
ρ
E
0
), and that the Gibbs state is
stationary under the evolution, one has that:
S(ρ
t
||ω
β
(H)) = S(Tr
E
h
U
t
(ρ
S
0
ρ
E
0
)U
t
i
||Tr
E
h
U
t
(ω
β
(H) ρ
E
0
)U
t
i
)
S(U
t
(ρ
S
0
ρ
E
0
)U
t
||U
t
(ω
β
(H) ρ
E
0
)U
t
) = S(ρ
S
0
ρ
E
0
||ω
β
(H) ρ
E
0
) =
= S(ρ
S
0
||ω
β
(H)) +
S(ρ
E
0
||ρ
E
0
), (45)
where we used the fact that the relative entropy decreases when one traces out part of the system,
together with its invariance under unitary evolution. Denoting with V
t
the dynamical map which bring
the reduced density matrix of the system from the original state ρ
S
0
to ρ
t
, we can rewrite the equation
as:
S(ρ
S
0
||ω
β
(H)) S(V
t
ρ
S
0
||ω
β
(H)) 0, (46)
which ensures that on any finite time the integral of
ρ
t
S(ρ
t
||ω
β
(H)) is positive. Unfortunately,
this result does not give any information about the monotonicity of the change of entropy: it is
sufficient to consider an exactly recurrent system, with periodicity T, for which the the quantity
S(ρ
S
0
||ω
β
(H)) S(ρ
t
||ω
β
(H)) will first increase, to only go back to zero at time T .
Nonetheless, when V
t
is a dynamical semigroup, meaning that for any t, s 0 the identity V
t+s
= V
t
V
s
holds, one can further prove that:
S(V
t
0
ρ||ω
β
(H)) = S(V
t
0
t
V
t
ρ||ω
β
(H)) S(V
t
ρ||ω
β
(H)) for t
0
t. (47)
This result implies that the entropy production rate in (44) is positive. As it was stated above, this
result is true in general only if the dynamics can be described by a dynamical semigroup, that is, if
the evolution is Markovian.
The generalization to the case in which the system is slowly driven as considered here follows directly
from the fact that we are dealing with adiabatic master equations for which L(ω
β
(H
t
)) = 0 t, and
hence we can restrict ourselves for each time t to an infinitesimal region around H
t
and apply the same
reasoning.
C Drazin inverse of the Lindblad operator
Here we derive the integral expression for the Drazin inverse L
+
t
in the space of trace-class operators.
We first recall the three conditions needed to define the inverse:
L
t
L
+
t
[A] = L
+
t
L
t
[A] = A ω
β
(H
t
)Tr [A] ; (48)
L
+
t
[ω
β
(H
t
)] = 0; (49)
Tr
h
L
+
t
[A]
i
= 0. (50)
We introduce the following trial solution
Λ
+
t
[A] :=
Z
0
dν e
νL
t
ω
β
(H
t
)Tr [A] A
. (51)
15
We first check (49), which gives
Λ
+
t
[ω
β
(H
t
)] =
Z
0
dν e
νL
t
ω
β
(H
t
)Tr [ω
β
(H
t
)] ω
β
(H
t
)
= 0, (52)
which follows from the normalisation of ω
β
(H
t
). For what regards (50) we find
Tr
h
Λ
+
t
[A]
i
=
Z
0
dν Tr
h
e
νL
t
ω
β
(H
t
)Tr [A] A
i
=
=
Z
0
dν
Tr [ω
β
(H
t
)] Tr [A] Tr [A]
= 0, (53)
where we used the fact that the propagator e
νL
t
is trace-preserving. Finally, in order to show that (48)
holds, we first find the following:
L
t
Λ
+
t
[A] =
Z
0
dν
d
dν
e
νL
t
ω
β
(H
t
)Tr [A] A
=
Z
ν=
ν=0
d
e
νL
t

ω
β
(H
t
)Tr [A] A
=
= A ω
β
(H
t
)Tr [A] + lim
ν→∞
e
νL
t
ω
β
(H
t
)Tr [A] A
= A lim
ν→∞
e
νL
t
A =
= A ω
β
(H
t
)Tr [A] , (54)
where we used the fact that t lim
ν→∞
e
νL
t
[B] = ω
β
(H
t
), for any normalised operator B. Moreover,
it also follows from the definition that
Λ
+
t
L
t
[A] =
Z
0
dν e
νL
t
ω
β
(H
t
)Tr [L
t
[A]] L
t
[A]
=
Z
0
dν e
νL
t
L
t
[A] =
=
Z
ν=
ν=0
d
e
νL
t
A = A lim
ν→∞
e
νL
t
A =
= A ω
β
(H
t
)Tr [A] . (55)
We thus conclude that Λ
+
t
= L
+
t
. In fact, assume that there exists another
˜
Λ
+
t
6= Λ
+
t
satisfying
conditions (37-39). Then, A we would have:
˜
Λ
+
t
[A]
(49)
=
˜
Λ
+
t
[A ω
β
(H
t
)Tr [A]]
(48)
=
˜
Λ
+
t
h
L
t
Λ
+
t
[A]
i
=
=
˜
Λ
+
t
L
t
h
Λ
+
t
[A]
i
(48)
= Λ
+
t
[A] ω
β
(H
t
)Tr
h
Λ
+
t
[A]
i
(50)
= Λ
+
t
[A] , (56)
which is in contradiction with the assumption that
˜
Λ
+
t
6= Λ
+
t
. This completes the derivation of the
integral expression (11).
D Metric in open quantum systems
We show here how to obtain equation (13) of the main text from the expression of the entropy pro-
duction rate in (42) and (44). Before starting, it is useful to introduce the derivative of the logarithm
of a density matrix [35]:
J
1
ρ
[δρ] := lim
ε0
log(ρ + εδρ) log(ρ)
ε
=
Z
0
dx (x + ρ)
1
δρ (x + ρ)
1
. (57)
It is worth pointing out that this operator is the inverse of J
ρ
defined in the main text, when one
restricts ρ to density matrices.
We can now pass to first express the entropy production rate (44) for a Lindbladian evolution, and
then plug the quasistatic expansion of the state (12). First notice that we can split at first order ˙σ as:
˙σ(ρ
t
) = lim
ε0
Tr [ρ
t+ε
log ρ
t+ε
] Tr [ρ
t
log ρ
t
]
ε
+
Tr [ρ
t+ε
log ω
β
(H
t
)] Tr [ρ
t
log ω
β
(H
t
)]
ε
=
= Tr [L
t
[ρ
t
](log ρ
t
log ω
β
(H
t
))] lim
ε0
Tr [ρ
t
log ρ
t+ε
] Tr [ρ
t
log ρ
t
]
ε
. (58)
16
where we used the Lindblad equation ˙ρ
t
= L
t
[ρ
t
]. We will now show that the last term is actually zero.
In fact, since for ε 1, ρ
t+ε
' ρ
t
+ εL
t
[ρ
t
] + . . . , we can expand the logarithm as:
lim
ε0
Tr [ρ
t
log ρ
t+ε
] Tr [ρ
t
log ρ
t
]
ε
' lim
ε0
Tr [ρ
t
log ρ
t
] + εTr
h
ρ
t
J
1
ρ
t
[L
t
[ρ
t
]]
i
Tr [ρ
t
log ρ
t
]
ε
=
= Tr
h
ρ
t
J
1
ρ
t
[L
t
[ρ
t
]]
i
= Tr [L
t
[ρ
t
]] = 0, (59)
where in the second line we used the fact that J
1
is self-adjoint with respect to the trace inner
product, meaning that: Tr
AJ
1
[B]
= Tr
J
1
[A]B
. This property can be inferred directly from the
definition (57) and the cyclicity of the trace. Moreover, it is just a matter of computation to see that
J
1
ρ
[ρ] = . Finally, in the last equality we used the fact that L
t
[ρ] is traceless.
We can now pass to give the quasi-static expansion of the entropy production rate. Recall that in
this limit the state is given by:
ρ
t
= ω
β
(H
t
) βL
+
t
h
J
ω
β
(H
t
)
[
˙
H
t
]
i
+ . . . (60)
at first order in O (1/T ). Plugging this expression in (58), we can obtain the result presented in the
main text:
˙σ(ρ
t
) = Tr [L
t
[ρ
t
](log ρ
t
log ω
β
(H
t
))] =
= β Tr
h
L
t
[ρ
t
] J
1
ω
β
(H
t
)
[L
+
t
h
J
ω
β
(H
t
)
[
˙
H
t
]
i
]
i
=
= β
2
Tr
h
J
ω
β
(H
t
)
[
˙
H
t
] J
1
ω
β
(H
t
)
[L
+
t
h
J
ω
β
(H
t
)
[
˙
H
t
]
i
]
i
=
= β
2
Tr
h
˙
H
t
L
+
t
h
J
ω
β
(H
t
)
[
˙
H
t
]
ii
(61)
where in the second line we used the expansion of the logarithm, in the third line we plugged in the
expansion of the state and used the property of the Drazin inverse, and in the last line the fact that
J
1
is self adjoint. This conclude the derivation of (11).
E Work for discrete processes
The natural extension of the definition of extracted work (2) for processes in which the Hamiltonian
is changed through a series of N quenches is given by:
W =
N1
X
i=0
Tr [ρ
i
(H
i
H
i+1
)] , (62)
where H
i
and ρ
i
are respectively the Hamiltonian of the system and its density matrix at the i-th step.
Adding and subtracting at each step S(ρ
i
) and using the definition of nonequilibrium free energy, we
can obtain a splitting analogous to (3):
W =
N1
X
i=0
(Tr [ρ
i
H
i
] β
1
S(ρ
i
)) (Tr [ρ
i
H
i+1
] β
1
S(ρ
i
)) =
=
N1
X
i=0
(F (ρ
i
, H
i
) F (ρ
i
, H
i+1
)) =
= F (ρ
0
, H
0
) F (ρ
N
, H
N
)
N1
X
i=0
(F (ρ
i
, H
i+1
) F (ρ
i+1
, H
i+1
)) =
= F
n.eq.
β
1
N1
X
i=0
(S(ρ
i
||ω
β
(H
i+1
)) S(ρ
i+1
||ω
β
(H
i+1
))), (63)
17
where in the last line we used the relation (1), which connects the nonequilibrium free energy with
the equilibrium one. In this way, we can see that the entropy production rate for a discrete process is
completely analogous to the one defined for a continuous evolution (4).
If we now assume that at each step the system is allowed enough time to thermalise (ρ
i
ω
β
(H
i
))
the dissipation simplifies to:
W
diss
= β
1
N1
X
i=0
S(ω
β
(H
i
)||ω
β
(H
i+1
)). (64)
Before proceeding further, it is useful to point out that the expansion of the relative entropy is given
by [35]:
S(ρ||ρ + εσ) = ε
2
Tr
h
σJ
1
ρ
[σ]
i
+ O
ε
3
. (65)
Then, if we assume to be in the quasistatic regime, so that we can use the expansion (17), W
diss
can
be rewritten up to order O
τ
2
as:
W
diss
= β
1
N1
X
i=0
S(ω
β
(H
i
)||ω
β
(H
i
) τβJ
ω
β
(H
i
)
[
˙
H
i+1
]) =
= βτ
2
N1
X
i=0
Tr
h
J
ω
β
(H
i
)
[
˙
H
i+1
]J
1
ω
β
(H
i
)
[J
ω
β
(H
i
)
[
˙
H
i+1
]]
i
=
= βτ
Z
γ
dt Tr
h
˙
H
t
J
ω
β
(H
t
)
[
˙
H
t
]
i
+ O
τ
2
, (66)
where in the last line we used the definition of Riemann sum to pass from the discrete sum to the
integral expression. As it can be seen, a formally equivalent result could be obtained from the fictitious
master equation (18).
F Expression of the metric in coordinates
In this section we explain how to get an explicit expression of the metric for general Lindbladian
evolutions. As a preliminary remark, it should be notice that any operator A can be treated as a
vector in a linear space via the identification:
A =
X
l,m
A
lm
|lihm| |Ai =
X
lm
A
lm
|lmi, (67)
where |ii is some orthogonal basis. In particular, since any Hermitian operator can be expressed as a
linear combination of:
l
= |li hl |, Σ
x
lm
=
1
2
(|l i hm| + |m i hl |), Σ
y
lm
=
i
2
(|l i hm| |m i hl |), (68)
we can interpret these operators as an orthonormal basis of the corresponding real vector space. More-
over, we can lift the Hilbert-Schmidt inner product to define a scalar product
hA|Bi = Tr[A
B]. (69)
This construction can be used to rewrite any linear superoperator as a d
2
· d
2
, matrix, where d is the
dimension of the original Hilbert space.
Using this identification, we can give an explicit expression for the Drazin inverse of a mixing
Lindbladian. As it was said in the main text, Lindbladians of this type have a unique zero eigenvector,
which we will denote by |ω
β
(H) i, and all the other eigenvalues λ
α
(corresponding to the eigenvectors
18
|αi) have negative real part. Moreover, since L
ω
is trace preserving, all the corresponding operators
α must be traceless. We can give an explicit expression of L
+
ω
as:
L
+
ω
=
X
α
(L
+
ω
)
α
|αi hα| =
X
α
Z
0
dν hα |e
νL
|αi
|αi hα| =
=
X
α
Z
0
dν e
νλ
α
|αi hα| =
X
α
1
λ
α
|αi hα| (70)
This explicit construction is a further proof of the fact that the Drazin inverse simply corresponds to
the usual inverse restricted to the traceless subspace.
We can now pass to the explicit computation of the metric. For simplicity, we choose as a basis on
the original Hilbert space the eigenbasis |ii of H. The only non-zero matrix element of J
ω
β
(H)
in the
basis given by the identification (67) can be explicitly evaluated as:
(J
ω
β
(H)
)
ij
lm
=
1
Z
H
Z
1
0
ds e
(1s)β
i
(δ
i
l
δ
j
m
e
β
l
Z
H
δ
i
j
δ
l
m
)e
j
=
=
1
Z
H
e
β
i
e
β
j
β
j
β
i
!
δ
i
l
δ
j
m
e
β(
i
+
l
)
Z
2
H
!
δ
i
j
δ
l
m
. (71)
Moreover, since we can rewrite the eigenbasis of L
ω
as:
|αi = u
α,ij
|ij i, (72)
where u is a unitary matrix, L
+
ω
is expressed in this basis simply by:
L
+
ω
=
X
α
1
λ
α
|αi hα| =
X
α,i,j
u
ij,α
u
α,lm
λ
α
|ij i hlm|. (73)
Now, assuming control only on a set of observables
n
X
(i)
o
, we can then finally compute the metric
entries as:
(m
L
ω
)
i,j
= m
L
ω
(X
(i)
, X
(j)
) =
=
1
2
X
lm,kn
X
(i)
lm
X
(j)
kn
((L
+
ω
)
ml
xy
(J
ω
β
(H)
)
xy
kn
+ (J
ω
β
(H)
)
ml
xy
(L
+
ω
)
xy
kn
). (74)
However cumbersome at a first look, this expression corresponds to a simple matrix product of the
operator specified in coordinates in (71) and (73). In this way, one gets a matrix expression of m
L
ω
of
the same dimension as the one of controllable parameters.
We can now derive equation (22) in the main text. Considering an evolution equation as in (21),
the Lindbladian can be expressed as:
L
ω
=
X
i
τ
1
i
|X
i
hX
i
i
ω
i hX
i
hX
i
i
ω
|. (75)
Then, we can use equation (70) to find the Drazin inverse of L
ω
, and plugging this expression in (74)
simply gives:
m
L
ω
(X
i
, X
j
) =
1
2
X
lm,kn
X
(i)
lm
X
(j)
kn
(τ
i
(J
ω
β
(H)
)
ml
kn
+ τ
j
(J
ω
β
(H)
)
ml
kn
) =
=
τ
i
+ τ
j
2
m
KMB
ω
(X
i
, X
j
), (76)
which proves the claim.
19
G Two level system coupled to a bosonic bath
In this section the details about the Lindblad master equation of a qubit in contact with a bosonic
bath. We preliminary choose the basis of the Hilbert space in such a way that the Hamiltonian can be
rewritten as:
H =
1
2
rˆσ
z
. (77)
The dynamics induced in the interaction picture by a bosonic bath with spectral density J(r) r
α
is
described by the master equation [44]:
˙ρ
t
= γ
r
(P
r
+ 1)
ˆσ
ρ
t
ˆσ
+
1
2
{ˆσ
+
ˆσ
, ρ
t
}
+ γ
r
P
r
ˆσ
+
ρ
t
ˆσ
1
2
{ˆσ
ˆσ
+
, ρ
t
}
, (78)
where γ
r
and P
r
are given by:
γ
r
= ˜γ
0
r
α
P
r
=
1
e
2βr
1
. (79)
For simplicity, the proper equilibration timescale ˜γ
1
0
is assumed to be one. Additionally, for book-
keeping reasons it is useful to define the quantity Γ
r
= γ
r
(2P
r
+ 1). As it was argued in the main text,
we can divide the density matrix of the system in a traceful and traceless component ρ = ω
β
(H) + δω,
so that the Lindblad equation can be rewritten as ˙ρ
t
= Λ
t
[δω
t
]. In this simple case, we can use the
Pauli matrices to parametrize the traceless space:
z x + iy
x iy z
!
(x, y, z). (80)
and plugging this expansion into (78) we obtain:
˙x(t)
˙y(t)
˙z(t)
= Λ
r
(ρ(t) ω
β
(H)) =
Γ
r
2
0 0
0
Γ
r
2
0
0 0 Γ
r
x(t)
y(t)
z(t)
, (81)
In this way we obtained a definition for Λ
r
, and consequently for its Drazin inverse. Assuming that
the dissipative dynamics only depends on the energy spacing of the Hamiltonian, and not on the unit
vector ˆn = (ˆx, ˆy, ˆz), the Drazin inverse for a generic Hamiltonian is simply obtained by a change of
basis. More explicitly we have:
H(x, y, z) =
1
2
r U ˆσ
z
U
Λ
1
(x,y,z)
= UΛ
1
r
U
. (82)
We now pass to quickly show how to obtain the metric in this case starting from the parametrisation
in spherical coordinates:
H = r cos ϕ sin θ ˆσ
x
+ r sin ϕ sin θ ˆσ
y
+ r cos θ ˆσ
z
. (83)
Using the parametrisation (80) the metric could be directly computed by plugging the Pauli matrices
in equation (14). In spherical coordinates, though, one needs to first find the basis of the tangent space
induced by the parametrisation, which can be computed by simple differentiation of the coordinate
chart:
r
= cos ϕ sin θ ˆσ
x
+ sin ϕ sin θ ˆσ
y
+ cos θ ˆσ
z
θ
= r cos ϕ cos θ ˆσ
x
+ r sin ϕ cos θ ˆσ
y
r sin θ ˆσ
z
ϕ
= r sin ϕ sin θ ˆσ
x
+ r cos ϕ sin θ ˆσ
y
.
(84)
Then, the matrix form of the metric can be obtained directly from (74), using as
n
X
(i)
o
the observ-
ables (84) or, equivalently:
m
i, j
= m
L
ω
(
i
,
j
) =
1
2
Tr
h
L
+
ω
[J
ω
[
j
]]
i
+ L
+
ω
[J
ω
[
i
]]
j
i
. (85)
20
Figure 4: On the left hand side of the picture optimal protocols between (0) = 2β and (1) = 2β are presented
for different times. These results are obtained with the technique presented in [50]. As it can be seen, for short times
the trajectories present a quench at the beginning and at the end of the protocol. In the large time limit, instead,
the optimal protocols become equivalent to the one obtained with the thermodynamic length. This behaviour is
illustrated on the left hand side, where we plot the work required for trajectories obtained by exact minimisation or
by solving the geodesic equation in function of time.
H Comparison with optimal control optimisation
The geometric approach provides an approximate solution which becomes exact in the limit of slow
driving. In this appendix we compare it with the exact solutions reported in [50? ]. First, we compare
in general terms the present method to the approach presented in [? ] to minimise dissipation in open
quantum systems, discussing how complex the two optimisation procedures are. Then, we focus on a
specific protocol involving a driven qubit and compare both approaches quantitatively.
The main idea of [? ] is to use the Pontryagin’s minimum principle (PMP) to minimise the dissipation
given a general Lindbladian master equation. In particular, using the decomposition of the system
Hamiltonian H =
P
λ
i
t
X
i
, we denote by d the number of externally controllable parameters and by
D the dimension of the Hilbert space. Then, following [? ], one needs to solve a differential system of
2D equations (see Eq. (16) in [? ]), constraint by d algebraic equations. This should be contrasted
to the geometric approach reported here, which only requires the solution of d differential equations
(see Eqs. (15), (23) and (24) in the main text). In general, in the optimal control theory there is
no systematic way to reduce the number of differential equations from 2D to d. This is possible in
particular systems, e.g., when the problem can be linearised; however, in thermodynamics the control
parameters usually appear both in H and e
βH
, making such a linearisation not possible. Hence, while
the complexity of the solution will of course depend on the specific system under considerations, it
seems reasonable to say that the geometric approach will be considerably simpler in mesoscopic/many-
body systems, in which usually d D. An extreme case is shown for the Ising chain in the main text,
where d = 1 and D . Another advantage of the geometric approach is that it allows to easily
identify the less dissipating transformations of H
t
by simple inspection of the eigenvalues of the metric.
This was illustrated in the qubit case, where the effect of quantum coherence or of different timescales
in the Lindbladian could be read out from the metric.
It should however be kept in mind that the geometric approach provides an approximate solution,
which becomes only exact in the slowly driven limit. To quantitatively analyse how good the approx-
imation is, we now compare it to the exact solution for a quantum dot presented in [50] (the same
results can be obtained using [? ]) with the one obtained solving the geodesic equation with the same
Lindbladian as above and a varying Hamiltonian of the form: H =
(t)
2
( + σ
z
). For simplicity of
notation, we will measure the energy in units of β and the times in τ. In [50] it is shown that optimal
protocols will present a quench in the Hamiltonian at the beginning and at the end of the protocol
when T is small. The size of the jumps is proportional to the duration of the process, and goes to zero
for T . This behaviour is illustrated on the left hand side of figure 4. Comparing the asymptotic
solution for T with the one provided by the geodesic equation we notice that the two actually
coincide. As expected, this implies that geodesics become optimal in the limit of big T , that is, in the
21
regime in which the quenches at the beginning and at the end of the protocol becomes negligible. This
trend is confirmed by the plot on the right hand side of Fig. 4, where we plot the total work necessary
to excite the quantum dot from (0) = 2β to (1) = 2β in function of the duration of the protocol.
In fact, the two minimisation protocols give results which are reasonably similar already for times of
order T τ , diverging by 10% when T = τ .
Summarising, both the approach reported here and in [? ] allow for minimising dissipation given
a Lindbladian master equation, and they appear to be complementary: [? ] provides exact solutions,
whereas the geometric approach provides approximate solutions (which become exact in the slow
driving limit) with a considerably simpler minimisation. The latter approach appears particularly
suited for mesoscopic/many-body systems, where the number of control parameters in the Hamiltonian
is small but the Hilbert space dimension is exponentially large.
22