Control Strategies for Maintaining Transport Symmetries Far From Equilibrium

David Andrieux
We present two schemes for controlling the transport dynamics of mesoscopic devices. In both
approaches, we manipulate the system’s output — such as particle currents and energy flows —
while maintaining symmetric transport properties, even under far-from-equilibrium conditions. We
provide exact descriptions of the modified processes under each scheme and investigate their char-
acteristics. Notably, one of the schemes is shown to minimize the dissimilarity between the original
and modified processes, as quantified by the Kullback-Leibler divergence. These findings have the
potential to enhance the design and control of mesoscopic systems.
arXiv:2501.05212v1 [cond-mat.stat-mech] 9 Jan 2025

CONTEXT AND OBJECTIVES

Imagine you are operating a mesoscopic device designed to produce a specific output, such as a flow of particles or
energy currents. The device’s operating mode reflects a balance between achieving the desired transport properties
and managing acceptable levels of fluctuations, dissipation, and robustness. Now, consider the problem of adjusting
this output in a controlled way. How can you ensure that the device continues to function as intended?
Controlling these output currents by modifying the system’s parameters can be quite challenging. In many systems,
there are numerous kinetic parameters, and altering any single parameter can impact the overall dynamics, including
fluctuations, dissipation, and interrelated properties. Conversely, some parameters may exhibit ”sloppy” behavior,
meaning that altering their values over a wide range may not significantly change the output. Navigating this complex
parameter space to identify the parameters’ adjustments for achieving the desired output is a challenging task.
In turn, different transport processes are interconnected: changes to one current can affect others. Close to equi-
librium, the Onsager symmetries describe how the response of a system is related to the response in the opposite
direction. This reciprocal relationship has proved essential for understanding the behavior of physical, chemical and
biological systems out of equilibrium. Extending these symmetries to far-from-equilibrium situations could expand
our understanding of energy and matter transfer in various applications, and improve our ability to design and control
mesoscopic devices.
In this paper, we propose two control schemes designed to achieve a system’s desired output together with symmetric
transport properties. Our strategy is to decompose the space of parameters in terms of thermodynamics quantities
and use geometric structures to navigate the space. This approach leads to the development of two distinct control
schemes, referred to as e and m, which are grounded in concepts from information geometry.
We obtain explicit constructions for the resulting equivalence classes in both methods. The e dynamics coincide with
the concept of ”dynamical equivalence classes” introduced to study large current fluctuations. We also demonstrate
that the e dynamics modify the device’s outputs while remaining as close as possible to the original dynamics, as
measured by the Kullback-Leibler divergence. In turn, the m dynamics define new equivalence classes not considered
previously. We derive expressions for the affinity-current manifold and response coefficients for these classes.
These results enable the control of mesoscopic systems by providing the necessary combinations of parameters to
achieve the desired outputs. These findings also reveal hidden structures and symmetries within the parameter space
with a clear thermodynamic interpretation. Beyond its potential applications, this work integrates and extends recent
findings in information geometry, transport fluctuations, response theory, and the generation of Markov chains.

Structure of the paper:

In Section 1, we introduce the dynamics of Markov chains and their associated transport properties. We explore the
relationship between microscopic and macroscopic quantities, as well as the dimensionality of the parameter space.
We also present the framework for nonlinear response theory.
In Sections 2 and 3, we construct the e and m equivalence classes and analyze their properties. We demonstrate
that both classes lead to symmetric transport properties, albeit through different mechanisms.
In Section 4, we apply these concepts to two distinct systems. The first system is a molecular motor toy model
with a single current, and which can be solved analytically. The second system involves two coupled currents and
a higher-dimensional parameter space. We use numerical simulations to illustrate the nonequilibrium symmetries
present in this more complex system.
Finally, in Section 5 we discuss the implications and new perspectives offered by these results.
 2

I. TRANSPORT IN MESOSCOPIC SYSTEMS

A. Markov chain dynamics

The dynamical properties of many biological, chemical and physical mesoscale systems are often well-described by
a Markov dynamics [1, 2]. We thus consider the space Λ of primitive Markov chains on a finite state space of size N
P to continuous-time is straightforward). Each primitive chain P ∈ Λ then admits a unique
(an extension of our results
stationary state π with i πi = 1.
Steady-state transport is characterized by the stationary currents

Jij = πi Pij − πj Pji . (1)

The system is at thermodynamic equilibrium when all currents vanish, i.e. when

πi Pij = πj Pji (detailed balance) (2)

for all pairs (i, j).

The currents Jij are not independent as they must satisfy Krichoff’s law at each node
X X
Jij = Jij = 0 . (3)
j i

As a result, there exist a set of independent currents Jα such that any current Jij can be expressed as a linear
combination of these independent currents. These independent currents are linked to the fundamental cycles cα of
the graph associated with P [3]:
X
Jij = γij (cα )Jα . (4)
α

Here the quantities γij (c) take the value ±1 if the transition i → j belongs to the cycle c in the positive (+) or
negative (−) direction, and 0 otherwise. An introduction to the graph representation of Markov chains and how to
identify the independent currents is presented in Appendix A.
In tandem with the currents, the thermodynamic affinities driving the system out of equilibrium can be measured
by the breaking of detailed balance along cyclic paths c = (i1 , i2 , · · · , in ) as
 
Pi1 i2 · · · Pin i1
Ac = ln . (5)
Pi1 i2 · · · Pin i1

The affinities vanish at equilibrium where detailed balance is satisfied. They do not involve the stationary probabilities
π and can thus be calculated from the transition probabilities directly. Similar to the currents, only a subset Aα of
these affinities are independent. The affinity of an arbitrary cycle can be expressed as a linear combination of the
fundamental affinities (Appendix A).
The currents and affinities capture the anti-symmetric component of the dynamics. The symmetric component is
described by the activities

Yij = πi Pij + πj Pji . (6)

P
These quantities (also called frenesies in the litterature) are not all independent since ij Yij = 2. Here, the analogue
of the affinities are the ”kinetic forces”

Xij = ln(Pij Pji ) . (7)

Eliminating one of the dependent activities Yr leads to effective forces

′
Xij = Xij − Xr . (8)

While less front-and-center than the currents and affinities, these symmetric components play an important role in
shaping transport properties [4, 5].
 3

B. Dimensionality of the parameter space

In the previous section, we characterized the behavior of a mesoscopic system using macroscopic quantities like
currents and affinities, linking them to the higher-dimensional space of microscopic parameters. In this section we
analyze the dimensionality of the parameter space. Understanding the geometry of the parameter space and its
relationship to thermodynamic quantities is crucial for apprehending - and ultimately controlling - the behavior of a
system.
A transition matrix P can be represented by a graph with N states, S self-transitions, and E edges, where each
edge represents an allowed transition between states [3]. The space Λ of chains with a similar graph structure then
has dimension

dim Λ = 2E + S − N . (9)

Since there are E − N + 1 independent currents or affinities [3], the iso-current manifold ΣJ has dimension

dim ΣJ = E + S − 1 . (10)

In particular, this holds for the equilibrium manifold Σ0 = ΣJ =0 . Similarly, the iso-affinities manifolds ΣA have the
same dimension E + S − 1.

Disordered ring example: Consider a ring with N states, N edges and no self-transitions (S = 0). There are
2N − N = N independent microscopic parameters and N − N + 1 = 1 independent current and affinity. As a result,
the equilibrium manifold, as well as all the iso-current manifolds, have dimension N − 1. The iso-affinity manifolds
also have dimension N − 1. This implies that one can change a combination of N − 1 parameters while keeping the
affinity or the mean current unchanged!

The dimensionality of the iso-currents and is-affinity manifolds reflects the non-unique, many-to-one mapping
of microscopic parameters to thermodynamic s(macroscopic) quantities. This observation is reminiscent of sloppy
models, where the behavior of the system is invariant despite variations in the microscopic parameters [6, 7]. In terms
of control, this indicates that there is no unique way to navigate the parameter space to generate a given output. It
also reveals that additional conditions are required to obtain a well-defined response theory. This last point will be
detailed in the next sections.
Finally, it is worth noting that the space Λ can be decomposed in terms the currents Jα and affinities Aα as well
as the activities Ye and kinetic forces Xe . Together with an appropriate divergence measure, these elements form an
embedding of the parameter space with a Minkowski structure [5, 8].

C. Nonequilibrium fluctuations and response theory

A),
To study the behavior of a system, a typical quantity of interest is the system’s current-affinity relationship Jα (A
which expresses how the currents vary as a function of the driving forces or affinities. However, since there are many
more microscopic parameters than affinities, any such function implicitly assumes a specific underlying trajectory in
the parameter space. Put differently, the current-affinity manifolds are not uniquely defined unless all the microscopic
parameters are specified [2, 9].
This is an important and subtle point, with theoretical and practical implications. As will we see below, both the e
A) between currents and affinities by prescribing the
and m equivalence classes generate well-defined relationships Jα (A
values of all microcopic parameters. The freedom to choose how to vary the microscopic parameters will also allow
us to generate additional symmetries out of equilibrium.
Assuming for now that Jα (A A) is well defined in the above sense, we can study the behavior of the system as the
affinities are varied. The response of the currents is defined as

∂Jα
A) =
Rαβ (A A) .
(A (11)
∂Aβ

Some investigators instead introduce affinity-dependent responses defined as [10–12]

X
Jα = A)Aβ .
σαβ (A (12)
β
 4

However, as noted by Peterson [13], in the nonlinear region Jα /Aβ does not have the physical relevance that ∂Jα /∂Aβ
has, although σ has an intuitive appeal because of our familiarity with the linear form.
The responses Rαβ (A A) are symmetric at equilibrium, Rαβ (00) = Rβα (00). However, this symmetry doesn’t hold
outside equilibrium in general. To further describe the change in response outside equilibrium, it is often convenient
to expand the currents as functions of the affinities around A = 0:
∞
X 1
A) =
Jα (A Lα,β1 ...βl Aβ1 · · · Aβl , (13)
l!
l=1

where we sum over repeated indices. The response coefficients Lα,β1 ...βl describe the response of the system increasingly
far from equilibrium. Equivalently, the affinities can be expressed as functions of the currents, Aα (J J ). Expanding
around J = 0 leads to the series expansion
∞
X 1
J) =
Aα (J Mα,β1 ...βl Jβ1 · · · Jβl . (14)
l!
l=1

We will use both formulations in our analysis. The expansion (13) will appear naturally when studying the e equiva-
lence classes while expression (14) will appear in the m equivalence classes.
The current fluctuations also provide important information about the systems’ behavior. They are characterized
by the generating function
* n
!+
1 X
G(ηη ) = lim ln exp η · J (k) . (15)
n→∞ n
k=1

The cumulants Kα1 ...αm can then be obtained by successive derivations of Q with respect to the counting parameters
η . In particular, the first-order cumulant Kα = Jα while the higher-order cumulants capture higher-order statistics.
The cumulants Kα1 ...αm also vary with the affinities. Expanding the cumulants around A = 0 leads to
∞
X 1
A) =
Kα1 ...αm (A Lα ...α ,β ...β Aβ · · · Aβl . (16)
l! 1 m 1 l 1
l=1

The coefficients Lα1 ...αm ,β1 ...βl are not independent as the fluctuation theorem imposes linear relationships between
coefficients of the same order m+l. These constraints generate the full (Onsager) symmetries at first order and partial
symmetries at higher orders [14, 15].
In the next sections, we will describe our strategies to change the microscopic parameters of the system to control
its output while ensuring that the system’s nonlinear response remains symmetric away from equilibrium.

II. E-CONTROL SCHEME

A. Affinity-based representation of Markov chains

To effectively control a Markov system or to generate a Markov chain with specific properties requires a parametriza-
tion of the space of transition matrices. Here we use a representation that generates Markov chains with given affinities.
It will be useful to analyze the e equivalence classes.
From an equilibrium chain E, we can generate a Markov chain P with affinities Aα through the mapping [17]

P = s[E ◦ Z(A
A)] . (17)

Here ◦ denotes the Hadamard product (P ◦ Q)ij = Pij Qij and

(
exp (±Aα /2) if i → j is a fundamental chord α in the positive (negative) orientation,
A) ≡
Zij (A (18)
1 otherwise.

The mapping s transforms a positive matrix Q into a stochastic one as

1
s[Q] = diag(x)−1 Q diag(x) , (19)
ρ
 5

where ρ is the largest eigenvalue of Q and x its corresponding right eigenvector.

Conversely, an equilibrium chain can be generated from P through the mapping
h i
(1/2)
P (e) = s (P ◦ P ∗ ) , (20)

where P ∗ is the time-reversed chain

Pij∗ = Pji (πj /πi ) (21)

and Q(1/2) is the element-wise exponentiation. The chain P (e) is the equilibrium chain that has the closest projection
distance to P in the following information-theoric sense [19]:

P (e) = arg min D(Q|P ) , (22)

Q∈Σ0

where the minimization is done over the space Σ0 of compatible equilibrium chains [25]. This hints at the central role
that P (e) will play in the characterization of the e equivalence classes. The chain P (e) is also related to the entropy
production of the system [26].
The representation (17) allows a direct parametrization in terms of the affinities. By varying the equilibrium chain
E we can generate the entire space of Markov chains [20]. However, while the affinities are directly available in this
representation, no simple expression exists for the corresponding currents.

B. Definition of the e-equivalence classes

We are now in position to define our e control scheme. Starting from the reference dynamics P , the e-control
scheme is defined as

Q(e) (λ
λ) = s[P ◦ Z(λ
λ)] . (23)

Here the parameters λ directly control the system’s affinities: If the chain P has affinity A , the chain Q(e) (λλ) has
affinities A + λ . The set of dynamics Q(e) forms the e equivalence classes. These classes were initially introduced to
study the large current fluctuations [17, 24, 27].

The class Q(e) forms a manifold with dimension equals to the number of independent currents or affinities. The
chains Q(e) (λ
λ) share similar characteristics for the symmetric part of their dynamics. Indeed, noting that Zij Zji = 1,
we see that two dynamics P and Q in the same class satisfy

Pij Pji ∝ Qij Qji . (24)

The conditions (24) imply that the kinetic forces are shifted by a constant along an e equivalence classes, ln Pij Pji =
ln Qij Qji + C for all (i, j). The effective forces (8) are thus invariant along an e equivalence class,

XP′ = XQ
′
. (25)

When looking at the entire e class of P , it is convenient to parametrize its elements by the affinities instead of the
parameters λ . We then have that
h i
Q(e) (AA) = s P (e) ◦ Z(A
A) . (26)

The class Q(e) is then obtained by varying the affinities A from ±∞. Since the mapping s is non-linear, the e manifold
depends non-linearly on the microscopic parameters (Figure 1). It expands the concept of e-geodesic introduced in
information geometry [16, 28].

C. Transport and fluctuation symmetries of the e-class

As discussed previously, the nonlinear response is meaningful only when both the macroscopic affinities and the
complete set of microscopic parameters are specified along the path they trace in parameter space. Within the e
equivalence class, there exists a unique mapping between affinities (and currents) and the corresponding dynamics;
 6

all other microscopic parameters are then determined by the appartenance to Q(e) . Transport is thus well-defined
along an e equivalence class.
We can thus examine the transport and fluctuation properties of the system under the e-control scheme (23). We
start by looking at the mean currents. By construction, the e-class will span the entire range of affinities with a
unique equilibrium dynamics P (e) . Also, since the class Q(e) always contains a dynamics and its time-reversed one,
the current-affinity relationship will be anti-symmetric with respect to the affinities,

A) = −Jα (A
Jα (−A A) . (27)

This relationship already imposes a symmetry that does not necessarily hold when varying a system’s parameters
along arbitrary paths [9].
Next, we consider the current reponses (11). As demonstrated in [35], they satisfy the relation

∂Jα ∂Jβ
A) =
(A A) ,
(A (28)
∂Aβ ∂Aα

demonstrating symmetry in indices (α, β) for any A , i.e. both near and far from equilibrium [35]. As a result, the
response coefficients (13) are also symmetric in indices (α, β1 , ..., βl ). Moreover, they can be expressed in terms of
equilibrium cumulants:

Lα,β1 ...βl = 0 if l is even (29)

 l
1
Lα,β1 ...βl = Kαβ1 ...βl (00) if l is odd. (30)
2

In particular, this leads to the Onsager symmetry Lα,β = Lβ,α and the corresponding Green-Kubo formula
Lα,β = (1/2)Kαβ (00). The symmetries (29)-(30), however, extend beyond those of the fluctuation theorem [14, 15, 35].
Similar expressions and symmetries can be derived for the higher-order cumulants and their nonequilibrium response.
A) [35].
Furthermore, all cumulants can be obtained by deriving the current-affinity manifold Jα (A

Where does the symmetry (28) come from? Its origin can be traced back to the expression (23) and realizing
that P ◦ Z(η) forms the generator of the current generating function (15). Combining this observation with the
representation (26), the current generating function of P can be expressed in terms of the equilibrium generating
function of the equilibrium chain P (e) :

GP (ηη ) = GP (e) (ηη + A /2) − GP (e) (A

A/2) . (31)

In other words, the nonequilibrium current fluctuations of P are entirely determined by the equilibrium fluctuations
of P (e) [24]. Since the equilibrium fluctuations are symmetric, these symmetries are shared by all elements of the
corresponding e class.

D. Minimizing distortion in the control of mesocopic transport

When controlling a system, a possible strategy is to modify the device’s outputs while remaining as close as
possible to the original dynamics. Remarkably, in addition to generating symmetric transport responses, the e class
also minimizes the distortion of the modified processes with respect to the original process.
This can be formalized as follows. To quantify the differences between the original and adjusted dynamics, we employ
the Kullback-Leibler divergence, a statistical measure that captures the dissimilarity between the two processes:
X Pij
D(P |Q) = πi Pij ln ≥ 0. (32)
ij
Qij

The KL divergence is always non-negative and equals zero when P and Q are identical.
In standard situations where the KL divergence is used, if P represents the true underlying process and Q represents
a model or approximation of that process, the KL divergence indicates how well the model Q captures the behavior
of the true process P . Another interpretation is that D(P |Q) reflects the probability rate of observing a typical
trajectory generated by the process Q in the original process P .
Using the KL divergence as our measure of dissimilarity, we then have the following result:
 7

FIG. 1. Schematic representation of the e and m equivalence classes in parameter space.. Both e and m manifolds
have dimension E−N +1 (here shown as 1-dimensional). They contain the time-reversed dynamics P ∗ , and cross the equilibrium
manifold Σ0 at P (e) and P (m) , respectively. The equilibrium dynamics P (e) and P (m) are in general distinct from each other.
The e-equivalence class typically forms a non-linear manifold while the m class is linear in parameter space.

Theorem. The e control scheme (23) solves the minimization problem

h X i
Q(e) (λ
λ) = arg min D(Q|P ) + λα Jα [Q] , (33)
Q
α

where the minimization is performed over the space of compatible chains [25].

Demonstration: To obtain the solution (23), an explicit calculation can be performed using techniques from rate
distortion theory, in particular the Blahut-Arimoto algorithm [29–31]. However, a simpler and more direct demon-
stration can be found in Ref. [17] based on a result from Ref. [32]. □

In other words, the equivalence class (23) describes how to change the system’s parameters to achieve a given output
while minimizing the distortion D(Q|P ) with respect to the original system. This ensures that the adjustments made
to the system are as targeted as possible, and minimize unrelated deviations or disruptions.

III. M-CONTROL SCHEME

A. Current-based representation of Markov chains

This representation generates Markov chains with given currents and activities, which will be used to define the m
equivalence classes. In this representation, we consider the flux matrix
Fij = πi Pij , (34)
which describes the steady state probability fluxes between states. F satisfies
X X
Fij = πi and Fij = πj (35)
j i
P
so that ij Fij = 1. Every such matrix F corresponds to a unique chain P and vice versa.
The currents (1) and activities (6) are obtained from the symmetric and anti-symmetric parts of F , respectively:
Yij = Fij + Fij (36)
Jij = Fij − Fij . (37)
The affinities of a cycle c = (i1 , i2 , · · · , in ) can be calculated from the flux matrix as
 
Fi1 i2 · · · Fin i1
Ac = ln (38)
Fi1 i2 · · · Fin i1
 8

since the stationary probabilities in the numerator and denominator cancel each other for a cyclic path.
The flux matrix can be decomposed as
(F + F T ) (F − F T ) Y +J
F = + = . (39)
2 2 2
In particular, the matrix F is symmetric at equilibrium where J = 0. More generally, the time-reversed dynamics
F ∗ is obtained by taking the transpose F T = (Y − J)/2. A flux matrix F can also be decomposed as a convex
combination of cycle matrices [21–23].
From a flux matrix F we can define the equilibrium system
F + FT Y
F̄ = = . (40)
2 2
The corresponding transition matrix
P + P∗
P (m) = , (41)
2
where we used that π is the stationary distribution of both P and P ∗ . The chain P (m) is the equilibrium chain that
has the closest projection distance to P in the following information-theoric sense [19]:
P (m) = arg min D(P |Q) , (42)
Q∈Σ0

where the minimization is done over the space Σ0 of compatible equilibrium chains. This anticipates the important
role that P (m) will play in the characterization of the m equivalence classes. The chain P (m) is also related to the
entropy production of the system [26].

B. Definition of the m-control scheme

We are now in position to define our m control scheme. In this scheme, the output of the system is controled by
varying the currents Jα . Starting from the flux matrix FP associated with P , the m-control scheme is defined as
1X
F (m) (∆J
J ) = FP + ∆Jα ℓα (Cα − CαT ) . (43)
2 α

Here ∆Jα = Jα′ − Jα is the change in output current and Cα is the cycle matrix associated with the cycle α (Appendix
(m)
A). The expression (43) is a flux matrix only when 0 ≤ Fij ≤ 1. The m class is then obtained by varying the change
in currents ∆JJ in their allowed range.
When looking at the entire m class of P , it is often convenient to parametrize its elements by the absolute currents
instead of their variations ∆J J . We then have that
1X Y 1X
F (m) (J
J ) = F̄ + Jα ℓα (Cα − CαT ) = + Jα ℓα (Cα − CαT ) . (44)
2 α 2 2 α

The class F (m) forms a manifold with dimension equals to the number of independent currents. From this expression,
we see that the m manifold contains the equilibrium chain P (m) and the time-reversed chain P ∗ .
The chains F (m) share similar characteristics forP the symmetric part of their dynamics. Indeed, using that the
matrices (Cα − CαT ) are anti-symmetric and satisfy j (Cα − CαT )ij = 0, we see that the stationary distribution π is
invariant as the J s are varied, and that F + F T is independent of J . That is, two dynamics P and Q in the same m
class have the same activities,
YP = YQ , (45)
P
and the same stationary distribution, i πi Qij = πj .
Since the stationary distribution π is invariant along the equivalence class m, the class (44) leads to the following
transition probabilities:
(m) 1 X
Qij (J J ) = Pij(m) + Jα ℓα (Cα − CαT ) . (46)
2πi α

The parameters (transition probabilities) vary linearly with the currents. The m control scheme is thus achieved by
a linear variation of the parameters (Figure 1). The m manifold expands the concept of m-geodesic introduced in
information geometry [16, 28].
 9

C. Transport symmetries of the m-class

Similar to the e class, the m class has well-defined transport properties as all the microscopic parameters are uniquely
determined within the class. We can thus examine the transport properties of the system under the m-control scheme
(43).
From expression (38), the affinity-current relationship for an m class can be written as
P !
Y Ye + ϵα (e) ρ Jρ ϵρ (e)
J ) = ln
Aα (J P . (47)
e
Ye − ϵα (e) ρ Jρ ϵρ (e)

Here, ϵα (e) = ±1 if cα contains the edge e in the positive (+) or negative orientation (−), and 0 otherwise (Appendix
A). In this expression, the activities Ye are fixed and the currents can vary in their allowed range. These affinity-current
relationships are anti-symmetric, i.e.

J ) = −Aα (J
Aα (−J J) . (48)

This symmetry doesn’t necessarily hold when tracing an arbitrary path in parameter space [9], indicating that the m
classes also display special transport properties.
Next, a direct calculation shows that the response takes the form

∂Aα X Ye ϵα (e)ϵβ (e)

J) = 2
(J P , (49)
∂Jβ e
2
Ye − ( ρ ϵρ (e)Jρ )2

demonstrating symmetry in indices (α, β):

∂Aα ∂Aβ
J) =
(J J) .
(J (50)
∂Jβ ∂Jα

This symmetry holds for any J , i.e. both near and far from equilibrium.
In terms of the reponse coefficients (14), the symmetry (50) imposes that all even-order coefficients are symmetric in
the indices (α, β1 , . . . , βl ) while all odd-order coeficients vanish. Using the result (49), we can obtain exact expressions
for the response coefficients:

X  1 l
Mα,β1 ...βl = 2 ϵα (e)ϵβ1 (e) . . . ϵβl (e) if l is odd, (51)
e
Ye
Mα,β1 ...βl = 0 if l is even. (52)

The response coefficients are symmetric in the indices (α, β1 , . . . , βl ) and expressed in terms of the equilibrium activ-
ities.

Where does the symmetry (50) come from? In contrast to the e equivalence class, there is no obvious connection
to the underlying equilibrium fluctuations. The symmetry of the nonlinear response thus seem to emerge from the
linear structure (43) and the structure of formula (49).
Also, in contrast with the e control scheme, the m class does not appear to follow a minimization principle similar to
(33), despite the structure suggested by (42). An interpretation of the m control scheme in terms of information-theoric
quantities remains to be explored.

IV. APPLICATIONS

We now illustrate the construction and transport symmetries of the e and m equivalence classes on two systems.
We start by considering a molecular motor model with one independent current and a 2-dimensional parameter space.
This allows us to provide analytical expressions for both the e- and m-classes and visualize their shapes in parameter
space, as well as deriving the current-affinity curves.
Next, we consider a model of ions transport through a membrane with two coupled currents and a 8-dimensional
parameter space. While some analytical results are available, we focus on showing numerically that the symmetries
(28) and (48) hold far away from equilibrium.
 10

A. Molecular motor model

We consider a Markov chain representing a molecular motor with 2ℓ states corresponding to different conformations
of the protein complex. These states form a cycle of periodicity 2ℓ corresponding to a revolution by 360° for a rotary
motor or a reinitialization step for a linear motor. The motor alternates between two types of states according to the
transition matrix [33]

0 p1 1 − p1
 
1 − p2 p2 
1 − p1 p1
 
 
P = .. ..  .

 . 0 . 

 1 − p1 0 p1 
p2 1 − p2 0 2ℓ×2ℓ

The matrix P is doubly stochastic, so that its stationary state is given by the uniform distribution π = (1, 1, · · · , 1)/2ℓ
for all parameters (p1 , p2 ). The independent current J and affinity A take the form

1
J= (p1 + p2 − 1) (53)
2ℓ
and
p1 p2
A = ℓ ln . (54)
(1 − p1 )(1 − p2 )

The system is at equilibrium when J = A = 0 (i.e., when p2 = 1 − p1 ).

The activities read Y1 = (p1 − p2 + 1)/2ℓ and Y2 = (p2 − p1 + 1)/2ℓ depending if the transition is odd or even,
respectively. However, only one edge is independent due to the symmetry of the model and the constraint Y1 +Y2 = 1/ℓ.
Taking

1
Y = Y1 = (p1 − p2 + 1) (55)
2ℓ
as the independent activity leads to an effective kinetic force

p1 (1 − p2 )
X = ℓ (X1 − X2 ) = ℓ ln . (56)
p2 (1 − p1 )

1. Transport along e-equivalence classes

The e equivalence classes are obtained by varying the affinity A at fixed value of X. Using expressions (54) and
(56), the model parameters can be expressed as

e(A+X)/2ℓ e(A−X)/2ℓ
p1 = , p2 = . (57)
1 + e(A+X)/2ℓ 1 + e(A−X)/2ℓ

Therefore, each e equivalence class traces a nonlinear path in parameter space (Figure 2a). Along these equivalence
classes, the current-affinity curves take the form

1 (eA/ℓ − 1)
JX (A) = . (58)
2ℓ (1 + e(A+X)/2ℓ )(1 + e(A−X)/2ℓ )

Within these equivalence classes, the current-affinity curves are well-defined (Figure 2b). The currents are anti-
symmetric, reach the maximal and minimal possible values ±1/2ℓ when the affinity reaches ±∞, and are invariant
under the transformation X → −X. The current-affinity curves (58) also capture fluctuation at all orders, i.e. deriving
JX (A) with respect to A generates the various cumulants for any value of the affinity (not shown) [9].
 11

FIG. 2. e-equivalence classes and corresponding current-affinity curves for the molecular motor. (a) Equivalence
classes for different values of X = [−20, −10, −5, 0, 5, 10, 20]. Each class traces a non-linear path in parameter space except
(e)
for the class X = 0. Each class intersects the equilibrium manifold (dashed line) at a unique point PX . (b) Current-affinity
curves (58) for different equivalence classes X. Each curve is anti-symmetric and invariant under the transformation X → −X.
Each curve spans the entire range of possible current values ±1/2ℓ.

2. Transport along m-equivalence classes

The m-equivalence classes can be obtained from the construction (44). Since the system has a uniform steady state,
the flux matrix F = P/2ℓ. It can be decomposed as

0 p1 1 − p1 0 Y1 Y2 0 1 −1
     
1 − p2 p2  Y1 Y2  −1 1 
1 − p1 p1 Y2 Y1 −1 1
     
1   1  J 
F =  . .  =  . .  +  . . 
2ℓ 
 . . 0 . .  2
  . . 0 . .  2
  . . 0 . . 

 1 − p1 0 p1   Y2 0 Y1   −1 0 1 
p2 1 − p2 0 Y2 Y1 0 1 −1 0

The m equivalence classes are obtained by fixing the activities Yi (remember that Y1 +Y2 = 1/ℓ) and varying the current
J. Since the transition probabilities satisfy 0 ≤ pi ≤ 1, the current varies between − min(Y1 , Y2 ) ≤ J ≤ min(Y1 , Y2 ).
In terms of the (p1 , p2 ) variables, we have

(p1 , p2 ) = ℓ (Y1 + J, Y2 + J) . (59)

That is, the m equivalence classes form straight lines in the parameter space (Figure 3a) [34]. While not a general
property of m classes, in this system they satisfy a minimization principle akin to (33) but for the reversed divergence
D(P |Q) instead of D(Q|P ).
The affinity-current curves are well-defined in the equivalence classes and take the form
  
Y1 + J Y2 + J
AY (J) = ℓ ln (60)
Y1 − J Y2 − J

where the Yi are fixed. The response curves are monotonic, anti-symmetric in J, and invariant under the transforma-
tion Y → 1/ℓ − Y . However, in contrast with the e-classes, the currents do not span the entire possible range ±1/2ℓ
(Figure 3b).

B. Model of ions transport across a membrane

Let’s now turn to a model of ion transport as an archetype of systems with two independent currents. Following
Hill [2], consider a cell surrounded by a membrane that separates the cell’s interior (In) from its environment (Out).
 12

FIG. 3. m equivalence classes and corresponding current-affinity curves for the molecular motor. (a) m
equivalence classes for different values of Y = Y1 = [0.03, 0.08, 0.12, 0.17, 0.21, 0.26, 0.3]. Each class forms a straight line in
(m)
parameter space, and intersects the equilibrium manifold (dashed line) at a unique point PY . (b) Current-affinity curves (60)
for different equivalence classes Y . Each curve is anti-symmetric and invariant under the transformation Y → 1/ℓ − Y . The
currents reach different limiting values depending on the value of Y .

A complex E, which can exist in two conformations E and E*, has binding sites for ions L and M. These sites are
accessible to inside molecules in configuration E only, and to outside molecules in configuration E* only. L can be
bound only when M is already bound on its site (Figure 4a).

FIG. 4. Ion transport model with two independent currents. (a) Mechanism for the transport of ions M and L across
the membrane. In the normal mode of operation, molecule M has a larger concentration inside than outside, [Mi ] > [Mo ], while
the opposite holds for molecule L, [Lo ] > [Li ]. The complex E acts as a free energy transducer and utilizes the M concentration
gradient to drive molecules of L from inside to outside against its concentration gradient. For example, in the case of the
Na/K-ATPase complex M and L would correspond to K+ and Na+ , and transport would be coupled to ATP consumption.
(b) Kinetic diagram. (c) Cycle decomposition. The cycles a and b are chosen as the two independent cycles. The positive
orientation is chosen counterclockwise (adapted from Hill [2]).

The system has N = 6 states connected by the kinetic diagram depicted in Figure 4b. The parameter space Λ is
8-dimensional while the set of equilibrium dynamics forms an 6-dimensional manifold. There are two independent
affinities or currents, here measured along cycles a and b (Figure 4c).
 13

1. Transport along the e-equivalence classes

We illustrate the transport dynamics in a given equivalence class e. Each equivalence class corresponds to a 2-
dimensional manifold in the 8-dimensional parameter space. As expected, these currents are coupled and nonlinear, i.e.
a change in Aa will drive the current Jb and vice versa (Figure 5). Whithin an e equivalence class, the anti-symmetry
A) = −Jα (A
Jα (−A A) of the current-affinity manifold is verified.
In addition, and despite their nonlinear behavior, the currents responses display the ’hidden symmetry’ (28) both
near and far from equilibrium [36]:
∂Ja ∂Jb
A) =
(A A)
(A (61)
∂Ab ∂Aa
for all values of the affinities A in a given equivalence class (Figure 6). The symmetry (62) is also confirmed for
all other e equivalence classes (not shown). Higher-order fluctuations can be obtained by further derivation of the
currents and their responses are also symmetric [35].

FIG. 5. Coupled currents of the ion transport model along an e equivalence class. Currents Ja and Jb and their
isolines (solid curves) as a function of the affinities Aa and Ab in a given equivalence class. The equivalence class is defined
by the equilibrium dynamics P (e) with p12 = 0.62, p16 = 0.38, p21 = 0.57, p23 = 0.13, p25 = 0.30, p32 = 0.41, p34 = 0.59, p43 =
0.46, p45 = 0.54, p52 = 0.42, p54 = 0.32, p56 = 0.26, p61 = 0.66, and p16 = 0.34.

FIG. 6. The currents response is symmetric both in the linear and nonlinear regimes. The response coefficients
J /∂A
∂J A display a complex structure in the nonlinear regime. Nonetheless, the cross-response coefficients ∂Ja /∂Ab = ∂Jb /∂Aa
are identical both near and far from equilibrium. Parameters take the same values as in Figure (5).
 14

2. Transport along the m-equivalence classes

Each m equivalence class also corresponds to a 2-dimensional manifold in the 8-dimensional space. The m classes
also display a nonlinear behavior and, beyond the anti-symmetry Aα (−J J ) = −Aα (J
J ) characteristic of equivalence
classes, no other symmetry is apparent (Figure 7).
Despite the nonlinearities, the affinity-current responses display the ’hidden symmetry’ (50) both near and far from
equilibrium:
∂Aa ∂Ab
J) =
(J J)
(J (62)
∂Jb ∂Ja
for all values of the currents J in a given equivalence class (Figure 8). The symmetry (62) is also confirmed for all
other m equivalence classes. The parameters are varied lineraly to generate this equivalence class (not shown).

FIG. 7. Affinities of the ion transport model along an m equivalence class. Affinities Aa and Ab and their isolines
(solid curves) as a function of the currents Ja and Jb in a given equivalence class. The equivalence class is defined by the
equilibrium dynamics P (m) with the same parameters as in Figure 5.

FIG. 8. The affinity response is symmetric both in the linear and nonlinear regimes. The response coefficients
A/∂J
∂A J display a complex structure in the nonlinear regime. Nonetheless, the cross-response coefficients ∂Aa /∂Jb = ∂Ab /∂Ja
are identical both near and far from equilibrium. Parameters take the same values as in Figure (7).
 15

V. SUMMARY AND OUTLOOK

In this study, we present a novel framework for controlling mesoscopic transport. By systematically varying combina-
tions of the system’s parameters, we can effectively regulate the output while ensuring symmetric transport properties
out of equilibrium. Achieving symmetric transport properties away from equilibrium simplifies the behavior of the
system and opens new avenues for apprehending, controlling and designing mesoscopic systems.
We employ two geometric structures as a natural way to navigate the parameter space [16], leading to the develop-
ment of two distinct control schemes. Each scheme generates a corresponding dynamical equivalence class, denoted
as e and m, which describes how parameter adjustments can achieve a desired output. Both schemes lead to sym-
metric transport dynamics, i.e. the system’s response is identical for all reciprocal transport processes, even far from
equilibrium.
These transport symmetries arise through different mechanisms in the e and m classes. While both influence the anti-
symmetric components of the dynamics, the symmetric components are constrained differently in each case. Indeed, an
e-class fixes the kinetic forces, while an m-class maintains constant activities [41]. Future research could explore other
schemes for generating symmetric transport properties, and investigate potential connections with other frameworks,
as explored in [4, 5, 11, 12, 37]. In addition, the role of these equivalence classes in non-stationary scenarios, such as
systems relaxing towards a steady state or driven by time-dependent forces, remain to be investigated.
Despite their conceptual similarities, one scheme might be better suited than the other for a specific system or
application, depending on one’s focus and ability to control the parameters of a system. One the one hand, the
e control scheme presents useful properties, including enhanced symmetries for higher-order fluctuations and the
capacity to achieve the full range of possible currents. It also offers the advantage of modifing the device’s outputs
while remaining as close as possible to the original dynamics, thereby minimizing unrelated deviations or disruptions.
However, the e necessitates nonlinear parameter changes for system control.
The m control scheme, on the other hand, allows for linear parameter adjustments to tune the output, with affinity-
current relationships that can be calculated analytically. However, the range of achievable currents is limited in each
equivalence class and does not encompass the full spectrum of possible currents.
These equivalence classes serve as natural extensions of information geometric concepts in non-equilibrium contexts.
In the linear regime, the parameter space exhibits a well-defined structure based on the Fisher information metric,
which is associated with equilibrium fluctuations [38, 39]. However, the structure of the parameter space outside
the linear regime remains largely unexplored, and our classes provide new frameworks with clear thermodynamic
interpretations [26, 40]. The impact of alternative dissimilarity measures [5, 16, 42] on transport properties also
deserves further examination.
Finally, the principle of adapting a system under a minimization constraint can be extended to meet various
objectives. For instance, a framework for simplifying generative models of stochastic processes has been developed
by adapting the entropy rate under a dissimilarity constraint [43]. Potential applications include enhancing model
quality for sampling tasks and achieving clean models from corrupted data (non-parametric denoising) [44]. We an-
ticipate that these concepts could unlock new applications in the modeling and characterization of mesoscopic systems.

Acknowledgments. I thank the Solvay Institutes for the workshop organized in honor of Prof. Pierre Gaspard,
where some of this research was initiated. I’m also grateful to David Lacoste for encouraging me to write down these
results.

APPENDIX A: FUNDAMENTAL CURRENTS AND AFFINITIES

Schnakenberg’s theory decomposes the thermodynamical properties of stochastic dynamics into their independent
contributions [3].
A graph G is associated with a stochastic dynamics as follows: each state i of the system corresponds to a vertex
or node while the edges represent the different transitions i → j allowed between the states. An orientation is given
to each edge of the graph G. The directed edges are thus defined by i → j. A graph G usually contains cyclic paths.
However, not all such paths are independent. They can be expressed by a linear combination of a smaller subset of
cycles, called the fundamental set, which plays the role of a basis in the space of cycles. To identify all the independent
cycles of a graph we introduce a maximal tree T (G), which is a subgraph of G that satisfies the following properties:
• T contains all the vertices of G;
• T is connected;
 16

• T contains no circuit, i.e., no cyclic sequence of edges.

In general a given graph G has several maximal trees.
The edges l of G that do not belong to T are called the chords of T . For a graph with N vertex and E edges, there
exists M = E − N + 1 chords. If we add to T one of its chords l, the resulting subgraph T + l contains exactly one
circuit, cl , which is obtained from T + l by removing all the edges that are not part of the circuit. Each chord l thus
defines a unique cycle cl called a fundamental cycle. In this paper, we use the convention that the orientation is such
that the cycles are oriented as the chords l.
We can formulate many important thermodynamic concepts in terms of cycles. For instance, the affinity of an
arbitrary cycle c can be expressed as a linear combination of the affinities of a fundamental set:
X
A(c) = ϵl (c)A(cl ) . (63)
l

where the sum extends over all the chords, and ϵl (c) = ±1 if c contains the edge l in the positive (+) or negative
orientation (−), and 0 otherwise. Accordingly, the maximal tree T can be chosen arbitrarily because each cycle cl
can be redefined by linear combinations of the fundamental cycles. The fundamental cycles thus constitute a basis
identifying the independent contributions to the stochastic process.
If multiple cycles cl generate the same macroscopic P currents or affinities, we then group them into a common to
obtain the resulting macroscopic quantities, e.g. Jα = l∈α Jl .
We can also express these relationships in terms of matrices. If the cycle α is the sequence of distinct integers
[a1 , . . . , am ], then the corresponding cycle matrix Cα is the N × N matrix C[ a1 , . . . , am ] given by Ca1 a2 = Ca2 a3 =
. . . = Cam a1 = 1/m and 0 otherwise. The length of cycle α is ℓα = m. The entries of any such cycle matrix sum to
one.
Using these cycle matrices, we have that
X X
Jij = Jα ℓα (Cα − CαT )ij = γij (cα )Jα (64)
α α

where we introduce the matrices γ(cα ) = ℓα (Cα − CαT ). The matrices γ(cα ) are anti-symmetric and Kirchoff’s law (3)
imposes
X X
γij (cα ) = ℓα (Cα − CαT )ij = 0, (65)
j j

and similarly when summing over the indices i.

[1] G. Nicolis and I. Prigogine, Self-Organization in Nonequilibrium Systems (Wiley, 1977).

[2] T. L. Hill, Free Energy Transduction and Biochemical Cycle Kinetics (Dover, 2005).
[3] J. Schnakenberg, Network theory of microscopic and macroscopic behavior of master equation systems, Rev. Mod. Phys
48, 571 (1976).
[4] C. Maes and M. H. van Wieren, Time-symmetric fluctuations in nonequilibrium systems, Phys. Rev. Lett. 96, 240601
(2006).
[5] D. Andrieux, A Minkowski space embedding to understand Markov models dynamics, arXiv:2404.11238 (2024).
[6] R. Gutenkun, J. Waterfall, F. Casey, et al., Universally sloppy parameter sensitivities in systems biology models, PLoS
Comput Biol 3: e189 (2007).
[7] B. B. Machta, R. Chachra, M. K. Transtrum and J. P. Sethna, Parameter Space Compression Underlies Emergent Theories
and Predictive Models, Science 342, 604 (2013).
[8] H. K. Teoh, K. N. Quinn, J. Kent-Dobias, et al., Visualizing probabilistic models in Minkowski space with intensive
symmetrized Kullback-Leibler embedding, Phys. Rev. Res. 2, 033221 (2020).
[9] D. Andrieux, Making sense of nonequilibrium current fluctuations: A molecular motor example, arXiv:2306.01445 (2023).
[10] K. Tani, A formula of non-linear responses, Prog. Theor. Phys 32, 167 (1964).
[11] H. Vroylandt, D. Lacoste and G. Verley, Degree of coupling and efficiency of energy converters far-from-equilibrium, J.
Stat. Mech. 023205 (2018).
[12] H. Vroylandt, D. Lacoste and G. Verley, An ordered set of power-efficiency trade-offs, J. Stat. Mech. 054002 (2019).
[13] R. Peterson, Formal theory of nonlinear response, Rev. Mod. Phys. 39, 69 (1967).
[14] D. Andrieux and P. Gaspard, Fluctuation theorem and Onsager reciprocity relations, J. Chem. Phys. 48, 571 (2004).
[15] M. Barbier and P. Gaspard, Microreversibility, nonequilibrium current fluctuations, and response theory, J. Phys. A 51,
355001 (2018).
 17

[16] S. Amari, Information Geometry and Its Applications (Springer, 2016).

[17] D. Andrieux, Equivalence classes for large deviations, arXiv:1208.5699 (2012).
[18] H. Nagaoka, The exponential family of Markov chains and its information geometry, The proceedings of the Symposium
on Information Theory and Its Applications 28, 601 (2005).
[19] G. Wolfer and S. Watanabe, Information Geometry of Reversible Markov Chains, Information Geometry 4, 393 (2021).
[20] Using a similar approach, we can generate a Markov chain with given affinities and kinetic forces. Introducing the symmetric
matrix Wij (XX ) = exp(Xij /2), the chain s[W (XX ) ◦ Z(A
A)] has affinities A and kinetic forces X + C. Since the kinetic forces
are defined up to a constant, this completes the construction.
[21] J. E. Cohen, A geometric representation of stochastic matrix: Theorem and conjecture, The Annals of Probability 9, 899
(1981).
[22] S. Alpern, Rotational representations of stochastic matrices, The Annals of Probability 11, 789 (1983).
[23] S. Kalpazidou, Rotational representations of transition matrix functions, The Annals of Probability 22, 703 (1994).
[24] D. Andrieux, Nonequilibrium large deviations are determined by equilibrium dynamics, arXiv:1212.1807 (2012).
[25] The minization problem is performed over the space of matrices compatible with P . A stochastic matrix Q is said to be
compatible with P if it satisfies Qij ≥ 0 if Pij ≥ 0 and Qij = 0 if Pij = 0.
[26] D. Andrieux, Irreversibility as divergence from equilibrium, arXiv:2404.01978 (2024).
[27] G. Verley, Dynamical equivalence classes for Markov jump processes, J. Stat. Mech. 023211 (2022).
[28] The e geodesic between P and P (e) is obtained by varying all parameters λα uniformly, λα = λ for all α. Similarly, the m
geodesic between P and P (m) is obtained by varying all currents Jα uniformly, Jα = J for all α.
[29] R. Blahut, Computation of channel capacity and rate-distortion functions, IEEE Transactions on Information Theory 18,
460 (1972).
[30] S. Arimoto, An algorithm for computing the capacity of arbitrary discrete memoryless channels, IEEE Transactions on
Information Theory 18, 14 (1972).
[31] T. M. Cover, Elements of information theory (Wiley-Interscience, 2006).
[32] S. Shieh, Eigenvalue estimates using the Kolmogorov-Sinai entropy, Entropy 13, 2036 (2011)
[33] D. Andrieux and P. Gaspard, Fluctuation theorems and the nonequilibrium thermodynamics of molecular motors, Phys.
Rev. E 74, 011906 (2006).
[34] The same result is obtained by using the property (45), which shows that Q belongs to the m class of P if q1 − q2 = p1 − p2
as the stationary state is uniform for all parameters.
[35] D. Andrieux, Fully symmetric nonequilibrium response of stochastic systems, arXiv:2205.10784 (2022).
[36] D. Andrieux, Revealing hidden structures and symmetries in nonequilibrium transport, arXiv:2401.14496 (2024).
[37] T. Aslyamov and M. Esposito, Nonequilibrium Response for Markov Jump Processes: Exact Results and Tight Bounds,
Phys. Rev. Lett. 132, 037101 (2024).
[38] D. Sivak and G. E. Crooks, Thermodynamic metrics and optimal paths, Phys. Rev. Lett. 108, 190602 (2012).
[39] P. R. Zulkowski, D. A. Sivak, G. E. Crooks, and M. R. DeWeese, The geometry of thermodynamic control, Phys. Rev. E.
86, 0141148 (2012).
[40] K. N. Quinn, M. C. Abbott, M. K. Transtrum, B. B. Machta and J. P. Sethna, Information geometry for multiparameter
models: new perspectives on the origin of simplicity, Rep. Prog. Phys. 86, 035901 (2023).
[41] This can also be illustrated by calculating the symmetrized Kullback-Leibler divergence DS (P, Q) = D(P |Q) + D(Q|P )
between two models within a given equivalence class. It reduces to 2DS (P, Q) = (A AP − AQ ) · (J
J P − J Q ) + (X
X P − X Q) ·
Y P − Y Q ) = (A
(Y AP − A Q ) · (J
J P − J Q ) when both P and Q belong to either an e- or m-equivalence class. That is, the
symmetric variables X and Y do not contribute to DS (P, Q), which is not the case for two arbitrary dynamics P and Q
[5].
[42] G. E. Crooks, On measures of entropy and information, Tech. Note 009 (2010).
[43] G. E. Henter and W. B. Kleijn, Minimum entropy rate simplification of stochastic processes, IEEE Trans. on Pattern
Analysis and Machine Intelligence 38, 2487 (2016).
[44] As a side note, the simplified models of Henter and Kleijn require increasing the system’s affinities. They do not mention
or explore this consequence of their scheme; I believe this might have interesting implications on model denoising and
simplification.