www.nature.

com/scientificreports

OPEN Spectral Analysis of a Non-

Equilibrium Stochastic Dynamics
on a General Network
Received: 12 April 2018 Inbar Seroussi1 & Nir Sochen1,2
Accepted: 10 September 2018
Published: xx xx xxxx Unravelling underlying complex structures from limited resolution measurements is a known problem
arising in many scientific disciplines. We study a stochastic dynamical model with a multiplicative noise.
It consists of a stochastic differential equation living on a graph, similar to approaches used in
population dynamics or directed polymers in random media. We develop a new tool for approximation
of correlation functions based on spectral analysis that does not require translation invariance. This
enables us to go beyond lattices and analyse general networks. We show, analytically, that this general
model has different phases depending on the topology of the network. One of the main parameters
which describe the network topology is the spectral dimension d. We show that the correlation
functions depend on the spectral dimension and that only for d > 2 a dynamical phase transition occurs.
We show by simulation how the system behaves for different network topologies, by defining and
calculating the Lyapunov exponents on the graph. We present an application of this model in the
context of Magnetic Resonance (MR) measurements of porous structure such as brain tissue. This model
can also be interpreted as a KPZ equation on a graph.

Stochastic dynamics on large scale networks has attracted a lot of attention due to its wide occurrence in many disci-
plines, such as social sciences1–3, physics and biology4–6, communication and control theory7. Also of interest is the
population dynamics on networks, which is usually affected by both the topology of the network and some internal
stochastic noise in the system. In many applications the network topology is not known and one can only have access
to some measurable parameters of the system. Therefore, the analysis of critical phenomena and phase transitions
can provide information about the topology of the network8. Since the topology of the network is not known, the
study of spectral properties of the network is important. For example, the spectra and eigenvectors of the Laplacian
of uncorrelated or locally tree-like complex networks reveal interesting properties of the return-probability distri-
bution, the smallest eigenvalue of the Laplacian, as well as localization of the eigenvectors9,10. In additions, one can
characterize the complexity of the network by analysing the spectral dimension. The analysis of random walks on
graphs and their connections to the spectral dimension in equilibrium system is introduced in refs11–13.
The main question we ask is what kind of information about the topology of the graph can we extract based
on the dynamics of measurable functions, such as the correlation function between the sites of the network in a
non-equilibrium model. We show, by working in the spectral domain of the graph, that the spectral dimension
plays an important role in determining the dynamical properties of the system.
The model we consider consists of a system of interacting sites on a graph  with N vertices and E edges
between them. We are interested in the stochastic dynamics of some characteristic property {mi (t )}i ∈ , t ≥0. The
property mi(t) is linked to a physical measurable quantity in the real world and the graph is the underlying geom-
etry/topology in which the property lives and which usually is a complex network of sites. The model is the fol-
lowing collection of stochastic differential equations in the Stratonovich form on the graph  :
dmi (t )
= J ∑ Wij(mj (t ) − mi (t )) + gi (t )mi (t )
dt j∈

= −J ∑ Lijmj (t ) + gi (t )mi (t ),
j∈ (1)

1
Department of Applied Mathematics, School of Mathematical Sciences, Tel Aviv University, Tel Aviv, 69978, Israel.
2
Sagol School of Neuroscience, Tel Aviv University, Tel Aviv, Israel. Correspondence and requests for materials should
be addressed to I.S. (email: [email protected])

Scientific REportS | (2018) 8:14333 | DOI:10.1038/s41598-018-32650-5 1

www.nature.com/scientificreports/

Figure 1.  Schematic picture of a network of pores, where mi is the magnetization in the ith pore, and the
interaction between the pores is described by the transition matrix W.

with the initial condition mi(0) = m0, where J is a constant. The term gi(t) is a multiplicative white noise such that

⟨gi (t )⟩ = 0, ⟨gi (t )gj (t ′)⟩ = σij2(t − t ′) . (2)

We choose the Stratonovich form, since its solution is a limiting case of a system with a white noise with a short
memory14. As a start, we investigate the case where the system does not have any spatial and temporal memory,
that is, σij2(t − t ′) = σ 2δijδ(t − t ′), The topology of the network is encoded in the Laplacian matrix Lij of the graph
 , defined as Lij = kiδij − Wij; then k i = ∑ j Wij is the degree of site i. The Laplacian matrix is a zero-row-sum
matrix, i.e., ∑ j Lij = 0, for all i with non-negative eigenvalues which are ordered by increasing values (assuming L
is diagonalizable). The graph is assumed to be undirected, with uniformly bounded weights, and with no self
loops and multiple edges. Later on, for the asymptotic behaviour analysis, we will add a uniformly-like assump-
tion, which will lead to the relation with the spectral dimension of the graph. The model consists of two parts: an
interacting part, where diffusion components exchange with strength depending on their location on the graph,
and a non-interacting part, where each component follows a stochastic noise with variance σ2. The first causes
spreading, while the second pushes towards concentration (a.k.a localization or condensation).
The model has numerous applications. We present the model in the context of porous systems such as brain
tissue which can be measured using Magnetic Resonance Imaging (MRI). The sensitivity of the MR signal to
self-diffusion of water molecules can be utilized to extract information about the network describing the topology
of cells (neurons) in the brain. The concept of self-diffusion of molecules in a network of pores was already intro-
duced in refs15–17. This approach is based on the assumption that diffusion within a pore is much faster than diffu-
sion between pores, which allows one to describe the porous structure as a network of interacting pores. On the
measurement time scale the topology is assumed to be fixed. This description goes beyond the traditional models
which assume free diffusion of molecules and isolated pores18–20. Under this description, mi(t) is the magnetization
of the ith pore at time t. The interaction strength Wij between pores i and j is influenced by the molecules diffusing
between pores. The interaction is time-independent, i.e., possible jamming effects are neglected. In order to model
the processes within each pore, we add a stochastic noise coupled to each pore. The variance σij2 depends on many
parameters, such as temperature, the intrinsic geometry of the pore, fluctuations in the geometry of the pore, and
the inhomogeneous magnetic field applied. Figure 1 presents an illustration of a porous medium as a network of
pores. This model can be viewed as a generalization of the analysis introduced in ref.21. The main challenge in this
context is to solve the inverse problem of finding the network topology based on the limited resolution measure-
ments from a single voxel (three-dimensional pixel in an MRI image). A single voxel is a massive average over large
domains of around millimetres size, whereas the sizes one is actually interested in are three orders of magnitude
smaller of micrometers size, such as radii of axons in the brain. An additional worth noting application of this
model is in the context of interface growth; where the property mi (t ) = e hi(t ) can be related to the interface hight,
hi(t), via the Cole-Hopf transformation. In this representation Eq. (1) is transformed to a discrete KPZ equation on
the graph22. Exact results have recently been obtained for the one-dimensional case23,24.
Our novelty is in analyzing the model on a complex network. The network is not embedded in a lattice topology
and therefore does not posses symmetries such as translation invariance. This is important, since in most applica-
tions, a lattice topology is highly non-physical. We develop a tool to perform perturbation theory calculations on
networks without the usually imposed conservation of momentum constraint. This formalism can be applied to
other problems in physics of complex system, quantum disorder, neural science, and more. Using it, we are able to
show the existence of a phase transition in our model for spectral dimension d > 2. This phase transition is well
known on the lattice topology for the Euclidean dimension25–28. A sketch of the phase diagram of the model is
presented in Fig. 2. It is important to note that the critical values are also geometry/topology dependent as we show
later on. Moreover, the connection between this model and brain tissues measured by MRI is also established here
for the first time. In the following subsections, we review known results for some simple topologies.

Scientific REportS | (2018) 8:14333 | DOI:10.1038/s41598-018-32650-5 2

www.nature.com/scientificreports/

- Coupling

The mass is spread Only a few

Transient
equally on the graph with a big mass

→ →∞ →
=

Only a few with a

Recurrent big mass
→
- Noise

Figure 2.  Sketch of the phase diagram of the model.

Mean field topology.  We review the results obtained in the mean field topology, i.e., a fully connected graph
with constant coupling strength J between each two sites. In this topology, one can calculate exactly the equilib-
rium distribution using the Fokker-Plank equation for a normalized model, i.e., Mi(t ) = mi (t )/∑ i mi (t ), such that
∑ i Mi(t ) = 1 for any t1. Under this transformation, all the nodes in the graph are uncorrelated and we can omit
the subscript i, since the distribution is the same for all the nodes:
exp( − (μ − 1)/M )
Peq(M ) = A ,
Mμ
μ
where μ = 1 + 2J2 is a model-dependent parameter, A = (μ − 1) is a normalization factor, and Γ(μ) is the
σ Γ(μ )
gamma function. The equilibrium distribution exhibits a Pareto power-law tail. In this case, there are two phase
transitions, at μ = 1 and at μ = 3. This can be understood by deriving the equations for the dynamics of the first
and the second moments using the forward Fokker-Plank equation, or alternatively the Feynmann-Kac formula27.
With this formalism, one can also derive an equation for any pth moment.
The steady-state solution for the second moment is, then
μ−1
〈M2〉∞ = .
μ−3 (3)
By comparing these results to Fig. 2, we can identify σc22 = 1/J and σc2H → ∞. In this case, there are only two
phases, and equilibrium is always reached (unless the coupling strength is allowed to be negative). The meaning
of these phases in the context of the analysis of porous structure using MRI is as follows: the values Mi describe
the density of magnetization in each pore due to the molecules at the ith pore at time t, and all the pores are con-
nected and interact with pore i in the same way. Therefore, one would expect that the contribution of each pore to
the magnetization is the same. Here we show that even in the simplified topology of a fully connected graph, the
distribution of the magnetization among pores depends on μ, which represents the ratio of the interactions
strength among pores to the noise variance within each pore. If μ < 1 is allowed, then the interpretation is that the
contribution to the magnetization comes mainly from a few pores in the long-time limit. This can be regarded as
a localization regime which is observed in high magnetic field gradients in the case of isolated pores15,29. For
1 < μ < 3 the distribution of the molecules among the pores is relatively uniform, but the fluctuations in the mag-
netization in each pore are very large, whereas above μ = 3 the fluctuations become finite for all the pores.
An additional solvable model is the separable model Wij = bjki, given that ∑ j bj = 1 and ki is the degree of the
node i. Using a similar derivation as in the mean field topology, the critical phases are similar to those in the mean
field model (see Supplementary Information Sec. 1).

Lattice model and other topologies.  The lattice topology is studied in various contexts in the mathemat-
ics and physics literature. In mathematics, this model is known as the time-dependent Parabolic Anderson Model
(PAM), or in its continuum form as the Stochastic Heat Equation (SHE)30, and is closely related to the KPZ equa-
tion31. In the physics literature, one encounters it in many physical systems, such as directed polymers in random
media, a case which was analyzed also on trees2,6, random interface growth, and turbulence4. The discrete lattice
case was investigated thoroughly in refs26,27, and a generalization with inhomogeneous coupling is given in ref.25.
In the discrete case, the graph  is embedded in d space, and with the corresponding Laplacian, the governing
stochastic equation reads:
dmx (t )
= ∑ J|x−y|(mx (t ) − my (t )) + gx (t )mx (t );
dt |x− y |= 1

Scientific REportS | (2018) 8:14333 | DOI:10.1038/s41598-018-32650-5 3

www.nature.com/scientificreports/

note that in this form the coupling J|x−y| is assumed to be translation invariant. The case where J|x−y| = J was
analyzed mainly in refs26,27 and in the continuum settings, see for example ref.32. The analysis in the continuum
limit is based on perturbation theory calculations and renormalization group arguments. In this work, we extend
this derivation to a general network topology. A generalization to non-constant J|x−y| was done in ref.25. An addi-
tional variation of the lattice model where the weights J|x−y| are multiplied by a time-dependent random process
is introduced in ref.7. It is also possible to consider a model in which the noise is not delta correlated (white
noise) and it has some correlation form, for example, exponentially decaying with some correlation time, e.g., an
Ornstein-Uhlenbeck process. This adds another stochastic equation for the noise in each site. This extension of
the model is analyzed in refs2,25,26,33.

Results and Discussion

In this section, we show how the correlation functions provide information about the topology of the underlying
graph and the phase transitions of the system. As shown in the methods section (Sec. 4), using perturbation the-
ory we can calculate to all orders the two-point correlation function in the spectral domain. Asymptotic proper-
ties of this quantity can be found by analysing the vertex function Γ (s). The vertex function can be expressed in
terms of an infinite series of collision matrices Iij(s) = Pij ∗ Pij(s), where Pij(s) is the Laplace transform of the
transition probability from site i to j. Namely,
L
(s) = lim 2σ 2 l
Γ ∑ 2 ∑
[I(s)l−1]i1il .
L →∞ l =1 N i1il (4)
The convergence properties of this infinite series depends on the topology of the graph. The information about
the graph’s topology is encoded by the collision matrices. In the case of an undirected infinite graph, we assume
that the graph is uniformly-like, i.e. the long-time limit of the return probability does not depend on the node
location. We can connect the collision matrix in the time domain I(t) to the spectral dimension, using a property
of the Hadamard product (see Supplementary Information Sec. 2):

ln ∑k Iik(t ) ln Pii(t ) d
lim = lim =− ,
t →∞ ln t t →∞ ln t 2 (5)
where Pii(t) is the return probability to the site i; its Laplace transform, Pii(s), is the generating function of the
process. Thus the spectral dimension d is defined by the limit of Pii (if it exists)34–37. It can be understood intui-
tively as the dimension a random walker “experiences” in a diffusion process on the graph, it measures to what
extent the graph is recurrent. In the context of dynamics of molecules, it measures the probability that molecules
that started at the ith pore will return to this pore. Note that, Eq. (5) refers to the local spectral dimension (for the
difference between local and average spectral dimension see ref.38). Given Eq. (5), in the limit s → 0, we can
approximate any row of the matrix I(s) as follows:
~
~ ~ ~ 1 ~ s  B d~ −1 + d
∑ Ikl−1kl(s) = ∑ Pkl−1kl ∗ Pkl−1kl = P 
2  2 kl−1kl−1
∼
2−d
~s 2,
kl kl (6)
where B d is a constant in s (non diverging for d < 4). This approximation is used to analyze the behaviour of the
vertex function. The analysis of the vertex function is done separately for the two type of topologies: recurrent and
transient graphs.
For recurrent graphs, since for each i, lim Pii(s) = ∞, the geometric series in Eq. (4) is divergent. Therefore, if
s →0
a spectral dimension exists, then d ≤ 2 and the vertex function diverges, i.e., all the nodes in the graph are
strongly correlated. This corresponds to only one phase of the system in which the magnetization is spread rela-
tively uniformly among the pores.
For transient graphs, since for each i, lim Pii(s) < ∞, and under the constraint that
s →0

σ2 s  σ2
σ 2I(s)F = ∑ Pii  = ∑ s − 2 Jλ < 1,
2 i  2  α α (7)
it is possible to take the geometric sum in Eq. (4). Here, λα are the eigenvalues of the Laplacian matrix. Note
that, in the translation invariant case, an equality in Eq. (7) coincides with the self-consistent equation analyzed
in refs26,27. The analysis here is at the point where this self-consistent equation breaks down, i.e., close to a phase
transition. In the limit of L → ∞ the infinite series in Eq. (4) converges to the following function, when plugging
in the expression (6) recursively:

~ 2σ 2
Γ(s) = ~ .
B d~ −1 + d
1 − σ2 ~s 2
(8)
2−d

In the general case, one can write the vertex function in terms of the eigenvalues of the collision matrix:

 (s ) = 2σ 2 2σ 2 2σ 2 1
Γ N ∑ [(1 − σ 2I(s))−1]ii = N
Tr [(1 − σ 2I(s))−1] = N ∑ 1 − σ 2λ α(I(s )) ,
i α

Scientific REportS | (2018) 8:14333 | DOI:10.1038/s41598-018-32650-5 4

www.nature.com/scientificreports/

where the λ α(I(s)) are the α eigenvalues of the collision matrix I(s).
Renormalization group calculations, similar to the analysis in ref.32, show that when d < 2 there is only one
phase, in which all the nodes in the graphs are strongly correlated, and that when d > 2 there exists a critical value
σc2, which separates between any two phases, one with a finite correlation between two nodes in the network, and
the other with infinite correlation. The perturbative calculations, although carried out to all orders, are not capa-
ble of capturing the high noise critical point. At the point d = 2, the RG parameter diverges, which may also
indicates that there is a unique point, which we are unable to probe. It may be that a more careful analysis of the
asymptotic behaviour of Eq. (2) can lead to a better description of the phases of the system. If we compare the
stated results to known results on systems with translation invariance25 and to the mean field scenario, the pertur-
bation theory analysis reveals that the system is intermittent for all σ2 if d ≤ 2, whereas for d > 2 the system is
intermittent only for σ 2 ≤ σc2.

The graph Lyapunov exponents.  The moment Lyapunov exponents are an important tool. A large amount
of information about the topology of the system can be extracted by analyzing theses exponents. They are relevant
parameters in the context of porous media, since for instance they are closely related to what we can measure using
MRI, and can provide insight about the hidden porous structure and complexity. For example, the first Lyapunov
exponent can be viewed as a measure of the diffusivity of the medium. In this sub-section, we define and calculate
numerically the moment/annealed Lyapunov exponents for simple topologies. The moment Lyapunov exponents
can be calculated analytically in the mean field topology27. In the lattice topology, one can derive lower and upper
bounds26,27. The definition of the Lyapunov exponents can be generalized to the case of a general graph topology,
where, since the system is not translation invariant and the correlations are node dependent, one needs to average
over all the nodes. We define two graph Lyapunov exponents (if the indicated limits exist):

ln mp(x , t )
γp(L, σ ) = lim ,
t →∞ t (9)

ln(mp(x , t ))
γp(L, σ ) = lim ,
t →∞ t (10)
where mp(x , t ) = ⟨m(x1, t )m(x 2, t )m(x p, t )⟩ is the pth moment, such that x = (x1, x 2, … , x p) = (x , x , … , x ),
where x ∈  . The overline stands for the average over all the nodes in the graph, e.g., mp(x , t ) = lim 1 ∑ x mp(x , t )
N
N →∞
and the brackets stands for the average over all possible paths of the random walk up to time t. The order of the
limits is important. We call the quantity in Eq. (9) the graph sample Lyapunov exponent and the one in Eq. (10) the
graph moment Lyapunov exponent. Using Jensen’s inequality, it is clear that γp(L, σ ) ≤ γp(L, σ ) for any p and any
graph. In order to examine the behaviour of these exponents, we calculate them numerically for different topolo-
gies and noise values. The calculation can be done by solving the equations governing the moments dynamics,
which can be derived by using the Stratonovich Fokker-Plank equation for the density of all the nodes p(m, t):

∂p †
p = J∑ 
(
∂ k i mi − ∑ j Wijmj p )
 + σ
2
∂  ∂(mi p) 
∂t
= L FP
∂mi
∑ mi
2 i ∂mi 
.
∂mi 
i (11)
Alternatively, one can calculate the moments dynamics based on the backward Fokker-Plank using the Feynmann-Kac
formula (see ref.27). Since the behaviour of higher moments and the intermittency property are highly dependent
on the second moment26,27, we focus on the second-moment dynamics for insight into interesting dynamical
properties of the model:

d⟨mr ml ⟩  

= 2δrlσ 2⟨mr ml ⟩ − J∑ Llj⟨mr mj ⟩ + ∑ Lrj⟨ml mj⟩ = H2⟨mr ml ⟩,
dt  j j  (12)
where Hp can be identified, as a deterministic Schrödinger operator26, in this example, for the second moment
p = 2. Eqs (11) and (12) are valid for any σ as opposed to the perturbative analysis in Secs 2 and 4, which is valid
only for small σ. Since we do not assume translation invariance in our model, Eq. (12) cannot be reduced to a
one-dimensional problem as in the lattice case26. Therefore, we use a numerical solution to show the behaviour of
the solutions for different graphs. For the numerical calculation, we present Eq. (12) in a vector form, containing
all the possible pairs of correlations. The Lyapunov exponents, γ2 and γ2, are then calculated using the above defi-
nitions Eqs (9) and (10) assuming uniform initial conditions. Note that in the lattice case the moment Lyapunov
exponents are usually defined in two ways, the first as the long-time limit of a solution to Eq. (12), and the second
as the spectral radius of the Schrödinger operator H226,27. These two definitions are equivalent for the
d-dimensional lattice and the mean field topologies26, but it is not clear whether this is the case for general net-
works. We believe, following similar arguments as in ref.26, that this is the case for all transitive graphs and for
graphs of constant degree. We show numerically that it is indeed true for a d-regular graph. In general, for a
non-trivial topology the spectral radius is a lower bound on the Lyapunov exponent γ2(L,σ). An interesting ques-
tion worth investigating is that of the gap between the two definitions Eqs (9) and (10) with respect to the relation
between space (the number of nodes) and time. We compare numerically the two definitions and the largest
eigenvalue of the equation λmax (see Supplementary Information Sec. 3 for details). We also use as an additional

Scientific REportS | (2018) 8:14333 | DOI:10.1038/s41598-018-32650-5 5

www.nature.com/scientificreports/

Figure 3.  A numerical solution of the graph sample (green star) and moment (purple circle) Lyapunov
exponents (Eqs (9) and (10)) and the largest eigenvalue of Eq. (12) (horizontal blue line) for four types of
graphs: (a) Mean field (b) 1D lattice (c) regular graph of degree 4 (d) 3D lattice. The black dashed line represents
the analytical lower bound presented in Eq. (13).

reference the following theoretical lower bound on the second graph moment and sample Lyapunov exponent,
where equality is achieved in the mean field:

γ2(L, σ ) ≥ 2σ 2 − 2J〈k〉 ; (13)

here 〈k〉 denotes the average degree of the graph. This bound is proved in refs26,27 for the d-dimensional lattice
case where 〈k〉 = 2d. The proof for a general graph is similar (see SI Sec. 3). This lower bound shows that for
graphs where the average degree distribution (e.g., graphs with power law degree distribution) is diverging, the
second graph moment and sample Lyapunov exponent also diverge.
The numerical solution for the mean field Laplacian is presented in Fig. 3a for different values of the ratio J/σ2.
Each point was obtained by solving Eq. (12) for different values of the noise variance. The critical point is obtained
as expected (Eq. (3)) at J/σ2 = 1. In this case, γ2 → λmax which is also equal to the predicted analytical value Eq.
(13) 27. Figure 3b,d present the Lyapunov exponents for a 1D and 3D lattice, respectively, with periodic boundary
conditions. The simulation shows that, as, predicted by theory26,27, in 1D there is no phase transition. On the other
hand, in 3D, we see two phase transitions. The two definitions of the Lyapunov exponent Eqs (9) and (10) coincide
in these cases. Figure 3c presents the values of the Lyapunov exponent on a d-regular graph of degree d = 4. In this
case, γˆ2 → γ2 → λ max .
The theoretical analysis for simple topologies, such as the mean field and the lattice model1,25–27, suggests the
existence of an additional phase transition in the high noise limit in the transient case. This phase transition could
not be detected using the perturbative analysis. The numerical solutions in Fig. 3 show some divergence, for high
noise values, but it is not clear yet whether this is due to numerical instability or the high noise phase transition.

Conclusions
We have presented a stochastic model describing diffusion on a graph  with an additional multiplicative stochas-
tic noise. The dependence of the large-scale behaviour of this model on the topology/geometry of the graph is
analyzed. The problem we address is whether one can determine the topology of the network based on measure-
ments of some observables. This problem arises in different contexts and disciplines. We show that one of the
main structure parameters determining the asymptotic behaviour of the system is the spectral dimension. It is a
measure of the complexity of the network which indicates to what extent the network is recurrent. We established
the relation between the spectral dimension and the two-point function. This was obtained by calculating the

Scientific REportS | (2018) 8:14333 | DOI:10.1038/s41598-018-32650-5 6

www.nature.com/scientificreports/

correlation functions perturbatively, in the strength of the noise. This calculation is done in the spectral domain
without translation invariance assumption and it shows a direct connection to the return probability distribution.
In the transient case, of spectral dimension d > 2, there exists a phase transition from a phase when the average
correlation between any two nodes is finite and exponentially decaying with time, to a phase in which the average
correlation between each two nodes has a power law tail in time. We also present the moment and sample
Lyapunov exponents for a general graph. We calculate them numerically in simple topologies. We investigate the
connection between the two definitions considered above of the moment Lyapunov exponents on the graph. The
connection between these two definitions is an interesting issue and deserves a further analysis. This is also true
concerning the connection between the size of the graph and time.
Our analysis and results can be used to study other complex systems, especially, our spectral generalization of
the Janssen-De Dominicis technique; using this technique one can calculate moment functions which are related
to average observables in networks. In the context of MRI this model is rather a new proposition for a simplified
model which takes into accounts the effect of interaction among pores. Our analysis of the correlation functions
may lead the way to MRI measurements of the spectral dimension and the Lyapunov exponents. It provides a
new way to analyze and understand MR data for porous systems and can lead to new experiments and observable
parameters that may reveal exciting structural properties of our brain.

Methods
Perturbation theory analysis.  The model introduced in Eqs (1) and (2) for a general network is hard to
solve exactly. Here, we demonstrate how methods adopted from field theory can provide a unifying framework
to produce perturbative approximations for measurable physical quantities. Any stochastic or even deterministic
system can be described in terms of a path integral, to which asymptotic methods can be applied systematically.
Often of interest are observable quantities such as correlation functions (moments) of m or the probability density
function of the process p(m, t). Path integral methods provide a convenient tool for computing quantities such as
moments and transition probabilities perturbatively using Feynmann diagrams. They also make renormalization
group methods (see ref.32 Chap. 5) available when perturbation theory breaks down. The strategy of path integral
methods is to derive a generating function or functional for the moments and response functions. For stochastic
differential equation this can be done using the Janssen-De Dominicis formalism32,39. The action derived from
Eq. (1) by using this formalism is as follows (see Supplementary Information Sec. 4 for more details). Here, for
convenience we write the equation in the equivalent Itô’s form
~ , m] =  [m
[m ~ , m] +  [m ~ , m]
0 int
 σ 2 
~  ∂mi + J 1 2 2 2
= ∫
dt ∑ m
i
i
 ∂t
∑
j
Lijmj + mi  −
2  2 ∫ dt ∑i σ m~i mi ,
where the first term is the free theory and the second term is the “interacting” non-quadratic term. Here the initial
conditions have been omitted since our focus is on stationary states and a loss of memory of the initial conditions
is assumed. The Janssen-De Dominicis formalism allows us to calculate the correlation functions to all orders
using perturbation theory by expanding around the non-interaction term. The analysis here is similar to the lat-
tice case32, a difference being that we use the eigenbasis of the Laplacian of the graph. The eigenvalues and eigen-
vectors of the Laplacian operator are defined by the equation Lφα = λαφα and ∑ i φα† iφαi ′ = δαα′, λα ≥ 0. The zero
eigenvalue corresponds to the uniform eigenfunction: φ0i = 1 , ∀i. The difficulty in this expansion is that, in
N
contrast to the Fourier eigenfunctions, momentum is not assumed to be preserved, and neither is translation
i j
invariance. Therefore, ∑ α φαφα ≠ δij . Keeping this in mind, we can now calculate the correlation function in the
Laplacian basis. We define the transforms of the fields with respect to the eigenspace by m ∼ =∑ m ∼ i
i α αφα, and by
mi = ∑ α mαφα† i the inverse transformation: m∼ =∑m ∼ φ † i and m = ∑ m φ i . The free propagator of the process
α i i α α i i α
is diagonal in this basis:
∼ (t ′)〉 = δ exp( − Jλ (t − t ′)) Θ(t − t ′) = G (t − t ′)δ .
〈mα(t )mα′ 0 αα ′ α 0α αα ′

Let us also introduce the transition probability matrix:

Pki(t ) = ∑ G0α(t )φαkφα†i = ∑ φαkφα†i exp( − Jλ αt ) Θ(t )

α α (14)
∼ (t ′)〉 = 〈m (t )m
It can be shown, by using similar arguments as for the lattice topology, that 〈mα(t )m ∼ (t ′)〉 , i.e.,
α′ α α′ 0
in this theory the propagator is not renormalized. Note also that due to the symmetric form of the interaction, any
correlation function with odd combination of m and m ∼ vanishes. Therefore an interesting quantity to investigate
is the two-point correlation function. We analyze this quantity in the diagonal basis of the Laplacian. The expan-
sion of the two-point correlation function is then written as
~ (t )m
⟨mη (tη )mχ(tχ)m ~ (t )⟩
ρ ρ μ μ
= ⟨mη (tη )mχ(tχ)m ~ (t )m
~ (t )⟩
ρ ρ μ μ 0
∞
1 ~ , m])l m (t )m (t )m
+ ∑ ⟨ ( − int[m η η
~ ~
χ χ ρ(tρ)mμ (tμ )⟩0 .
l =1 l !

The actions in the basis of the Laplacian reads

Scientific REportS | (2018) 8:14333 | DOI:10.1038/s41598-018-32650-5 7

www.nature.com/scientificreports/

Figure 4.  A diagrammatic representation of the second-order in perturbation theory Eq. (15).

σ2
~ , m] =
0[m ∫ dτ ∑α m~α(τ )(m˙ α(τ ) + Jλ αmα(τ ) + 2
mα(τ )),

and

σ2 ~ ( τ ) φ † iφ † iφ i φ i .
~ , m] = −
int[m
2
∑ ∫ ~ (τ )m
dτmα1(τ )mα2(τ )m α3 α4 ∑ α1 α2 α3 α4
α1α 2 α 3α4 i

In order to calculate perturbatively the two-point function, we use the concept of the Feynmann diagrams. For
convenience, we look at the cumulative function. We sum over all the fully connected diagrams using methods
similar to those customarily applied in the lattice case (see Chap. 9 in ref.32). The difference here is that conser-
vation of momentum is not assumed. We calculate the two-point cumulative correlation function in the time
domain in the eigenbasis of the Laplacian:
~ (t )m
⟨mη (tη)mχ (tχ)m ~
ρ ρ μ (tμ)⟩c

 σ2 

=2 ∫ dτG0ρ(τ − tρ)G0η(tη − τ) exp  − 2 Δtηχρμ Γηχρμ(τ)G0μ(τ − tμ)G0χ(tχ − τ).
where we define Δtηχρμ = tη + tχ − tρ − tμ. This function Γηχρμ(τ) is called the vertex function in dynamic field
theory terms32. The vertex function is calculated below using perturbation theory. We show the derivation for
the first three orders; the general case follows by induction. The first-order expansion (l = 1) yields, using Wick’s
theorem,

Γηχρμ(τ , τ )[1] = 2σ 2 ⋅ ∑ φμ†iφρ†iφχiφηi.

i

The second-order expansion (l = 2) yields

Γηχρμ(τ1, τ2)[2] = 2 ⋅ 2! ⋅ σ 4 ⋅ exp( − σ 2(τ2 − τ1))

× ∑ φμ† iφρ† iφχjφηjP(τ2 − τ1)ij P(τ2 − τ1)ij .
i ,j (15)
Figure 4 shows a diagrammatic representation of the second-order perturbative calculation of Eq. (15) in the
eigenbasis domain. The third order expansion (l = 3) can derive in the same way:

Γηχρμ(τ1, τ3)[3] = 2 ⋅ 3! ⋅ σ 6 exp ( − σ 2(τ3 − τ1))

†i †i j j
× ∫ dτ2 i∑
,j ,k
φμ φρ φχ φη Pik(τ2 − τ1)Pkj(τ3 − τ2)Pkj(τ3 − τ2)Pik(τ2 − τ1) .

Following this derivation, we can provide the lth-order result in perturbation theory:

Γηχρμ(τ1, τL)[l] = 2 ⋅ l ! ⋅ σ 2l exp ( − σ 2(τL − τ1))

 l− 1 
 tr + 1 

× ∑ φμ† i1φρ† i1φχilφηil ∏  ∑ dtr I (tr +1 − tr )i ri r+1 
∫ 
 i1, il i tr 
r =2 
 r 

2
= 2 ⋅ l ! ⋅ exp ( − σ (τL − τ1))
× ∑ φμ† i1φρ† i1φχilφηil σ 2l dτl(I ∗ I ∗ …I )(τl − τ1) ,
∫
i1 , il  i1il (16)
where we denoted Iij(t) = Pij(t)Pij(t). Note that, in the integration over time we put t2 = τ1 and tl = τl. Figure 5
shows a diagrammatic representation of the geometric series presented in Eq. (17). Since there is no conservation

Scientific REportS | (2018) 8:14333 | DOI:10.1038/s41598-018-32650-5 8

www.nature.com/scientificreports/

Figure 5.  A diagrammatic representation of the sum of geometric series of concatenated two-point loop
diagrams (Dyson’s equation); this sum produces the self-consistent equation for the full vertex.

of momentum, each vertex in the sum has to include a product of four eigenvectors. The function Γηχρμ(τ) can be
calculated exactly to all orders in perturbation theory, yielding
L
1
Γηχρμ(τ ) = lim ∑ l! Γηχρμ(τ1, τL)[l] .
L →∞ l =0 (17)
The matrix I(t) is defined as the entry-wise square of the transition matrix P(t):

Iik(t ) = Pik(t )2 ,
where I(0) = 1. We call the matrix I(t) the collision matrix. Taking ρ = μ = η = χ = 0, and tρ = tη = 0, tμ = tχ = t,
such that φη† i = φχ† j = φρi = φμj = 1 and applying the Laplace transform Γ (s) = ∫ ∞ Γ(t )e−stdt , we obtain for
N 0
the vertex function
L
(s) = lim 2σ 2 l l− 1 L
2σ 2 l
Γ
L →∞
∑∑ N2
∏ ∑ I(s)irir+1 = Llim
→∞
∑ N 2 ∑
[I(s)l−1]i1il .
l = 1 i1il r =2 ir l =1 i1il (18)
In the results section, we present an analysis of the convergence of the vertex function for some graph topologies
based on the asymptotic properties of the collision matrix I(s). This allows us to derive the phase diagram of the
model.

References
1. Bouchaud, J. P. & Mezard, M. Wealth condensation in a simple model of economy. Physica A: Stat. Mech. Appl. 282, 536–545 (2000).
2. Gueudré, T., Dobrinevski, A. & Bouchaud, J. P. Explore or exploit? A generic model and an exactly solvable case. Phys. Rev. Lett. 112,
050602 (2014).
3. Solomon, S. & Richmond, P. Power laws of wealth, market order volumes and market returns. Physica A Stat. Mech. Appl. 299,
188–197 (2001).
4. Halpin-Healy, T. & Zhang, Y. C. Kinetic roughening phenomena, stochastic growth, directed polymers and all that. aspects of
multidisciplinary statistical mechanics. Phys. Rep. 254, 215–414 (1995).
5. König, W. The Parabolic Anderson Model: Random Walk in Random Potential (Birkhäuser, 2016).
6. Derrida, B. & Spohn, H. Polymers on disordered trees, spin glasses, and traveling waves. J. Stat. Phys. 51, 817–840 (1988).
7. Allegra, N., Bamieh, B., Mitra, P. P. & Sire, C. Phase transitions in distributed control systems with multiplicative noise. arXiv
preprint arXiv:1610.00653 (2016).
8. Dorogovtsev, S. N., Goltsev, A. V. & Mendes, J. F. Critical phenomena in complex networks. Reviews of Modern Physics 80, 1275
(2008).
9. Samukhin, A., Dorogovtsev, S. & Mendes, J. Laplacian spectra of, and random walks on, complex networks: Are scale-free
architectures really important? Physical Review E 77, 036115 (2008).
10. Mitrović, M. & Tadić, B. Spectral and dynamical properties in classes of sparse networks with mesoscopic inhomogeneities. Physical
Review E 80, 026123 (2009).
11. Cassi, D. Phase transitions and random walks on graphs: A generalization of the mermin-wagner theorem to disordered lattices,
fractals, and other discrete structures. Physical Review Letters 68, 3631 (1992).
12. Burioni, R. & Cassi, D. Universal properties of spectral dimension. Physical Review Letters 76, 1091 (1996).
13. Merkl, F. & Wagner, H. Recurrent random walks and the absence of continuous symmetry breaking on graphs. Journal of Statistical
Physics 75, 153–165 (1994).
14. Van Kampen, N. G. Itô versus Stratonovich. J. Stat. Phys. 24, 175–187 (1981).
15. Callaghan, P. T. Translational Dynamics and Magnetic Resonance: Principles of Pulsed Gradient Spin Echo NMR (Oxford University
Press, 2011).
16. Callaghan, P. T. et al. Diffusion in porous systems and the influence of pore morphology in pulsed gradient spin-echo nuclear
magnetic resonance studies. J. Chem. Phys. 97, 651–662 (1992).
17. Laun, F. B., Müller, L. & Kuder, T. A. NMR-based diffusion lattice imaging. Physical Review E 93, 032401 (2016).
18. Callaghan, P. T. A simple matrix formalism for spin echo analysis of restricted diffusion under generalized gradient waveforms. J.
Magn. Reson. 129, 74–84 (1997).
19. Grebenkov, D. S. NMR survey of reflected brownian motion. Rev. Mod. Phys. 79, 1077–1137 (2007).
20. Bar, L. & Sochen, N. A spectral framework for nmr signal with restricted diffusion. Concept. Magn. Reson. A (2015).
21. Seroussi, I., Grebenkov, D. S., Pasternak, O. & Sochen, N. Microscopic interpretation and generalization of the Bloch-Torrey
equation for diffusion magnetic resonance. J. Magn. Reson. 277, 95–103 (2017).
22. Barabási, A.-L. & Stanley, H. E. Fractal Concepts in Surface Growth (Cambridge university press, 1995).
23. Hairer, M. Solving the KPZ equation. Annals of Mathematics 559–664 (2013).
24. Corwin, I. Kardar-Parisi-Zhang universality. Notices of the AMS 63, 230–239 (2016).
25. Greven, A. & den Hollander, F. Phase transitions for the long-time behavior of interacting diffusions. Ann. Probab. 35, 1250–1306
(2007).
26. Carmona, R. & Molchanov, S. A. Parabolic Anderson problem and intermittency, vol. 518 (Memoirs of the American Mathematical
Society, 1994).
27. Molchanov, S. A. Ideas in the theory of random media. Acta Applicandae Mathematicae 22, 139–282 (1991).
28. Shiga, T. Ergodic theorems and exponential decay of sample paths for certain interacting diffusion systems. Osaka J. Math. 29,
789–807 (1992).
29. Grebenkov, D. S. Diffusion MRI/NMR at high gradients: Challenges and perspectives. Micro. Meso. Mat. (2017).

Scientific REportS | (2018) 8:14333 | DOI:10.1038/s41598-018-32650-5 9

www.nature.com/scientificreports/

30. Mukherjee, C., Shamov, A. & Zeitouni, O. Weak and strong disorder for the stochastic heat equation and continuous directed
polymers in d ≥ 3. Electron. Commun. Prob. 21 (2016).
31. Kardar, M., Parisi, G. & Zhang, Y. C. Dynamic scaling of growing interfaces. Phys. Rev. Lett. 56, 889 (1986).
32. Täuber, U. C. Critical Dynamics: a Field Theory Approach to Equilibrium and Non-Equilibrium Scaling Behavior (Cambridge
University Press, 2014).
33. Janssen, H. K., Täuber, U. C. & Frey, E. Exact results for the Kardar-Parisi-Zhang equation with spatially correlated noise. The
European Physical Journal B - Condensed Matter and Complex Systems 9, 491–511 (1999).
34. Alexander, S. & Orbach, R. Density of states on fractals: ≪fractons≫. Journal de Physique Lettres 43, 625–631 (1982).
35. Hattori, K., Hattori, T. & Watanabe, H. Gaussian field theories on general networks and the spectral dimensions. Prog. Theor. Phys.
Suppl. 92, 108–143 (1987).
36. Burioni, R. & Cassi, D. Random walks on graphs: ideas, techniques and results. J. Physics A: Math. Gen. 38, R45 (2005).
37. Rammal, R. & Toulouse, G. Random walks on fractal structures and percolation clusters. Journal de Physique Lettres 44, 13–22
(1983).
38. Cassi, D. Local vs average behavior on inhomogeneous structures: Recurrence on the average and a further extension of mermin-
wagner theorem on graphs. Physical Review Letters 76, 2941 (1996).
39. De Dominicis, C. & Peliti, L. Field-theory renormalization and critical dynamics above Tc: Helium, antiferromagnets, and liquid-gas
systems. Phys. Rev. B 18, 353 (1978).

Acknowledgements
We thank Ofer Pasternak, Asaf Nachmias and Ofer Zeitouni for helpful discussions and comments.

Author Contributions
All authors conceived and performed the research project. All authors wrote the text of the manuscript and
reviewed the manuscript.

Additional Information
Supplementary information accompanies this paper at https://doi.org/10.1038/s41598-018-32650-5.
Competing Interests: The authors declare no competing interests.
Publisher's note: Springer Nature remains neutral with regard to jurisdictional claims in published maps and
institutional affiliations.
Open Access This article is licensed under a Creative Commons Attribution 4.0 International
License, which permits use, sharing, adaptation, distribution and reproduction in any medium or
format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Cre-
ative Commons license, and indicate if changes were made. The images or other third party material in this
article are included in the article’s Creative Commons license, unless indicated otherwise in a credit line to the
material. If material is not included in the article’s Creative Commons license and your intended use is not per-
mitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the
copyright holder. To view a copy of this license, visit http://creativecommons.org/licenses/by/4.0/.

© The Author(s) 2018

Scientific REportS | (2018) 8:14333 | DOI:10.1038/s41598-018-32650-5 10