NOLTA, IEICE

Paper

On the covariance matrix of the stationary

distribution of a noisy dynamical system

Makito Oku 1 ,2 a) and Kazuyuki Aihara 1

1
Institute of Industrial Science, The University of Tokyo
4-6-1 Komaba, Meguro-ku, Tokyo 153-8505, Japan
2
Institute of Natural Medicine, University of Toyama
2630 Sugitani, Toyama-shi, Toyama 930-0194, Japan

a)
[email protected]

Received July 7, 2017; Revised December 10, 2017; Published April 1, 2018

Abstract: In this paper, we analyze the relation between the stability of a noisy dynamical
system based on linear approximation and the covariance matrix of its stationary distribution.
We reformulate the theory of dynamical network biomarkers in terms of the covariance matrix
and clarify the limiting behavior of the covariance matrix when a dynamical system approaches
a bifurcation point. We also discuss the relation between the Jacobian matrix and principal
component analysis. An application to a simple nonlinear network model is also demonstrated.
Key Words: early warning signal, covariance matrix, Lyapunov equation, vector autoregres-
sive model, multivariate Ornstein-Uhlenbeck process, principal component analysis

1. Introduction
Predicting state transitions is important in various fields such as ecology, economics, biology, medicine,
and engineering [1, 2]. Although the precise timing of the occurrence is usually difficult to predict,
some early warning signals such as the increase in variance [3] or autocorrelation [4] of a state variable
let us know the increase of possibility of the occurrence in near future.
From the viewpoint of dynamical systems theory, a stable fixed/equilibrium point with low stability
has high risks of state transitions in two reasons. Here, we use the term stability in a broad sense
regarding both the linear stability and basin sizes. One reason is that a small change of a parameter or
parameters of the target system may cause a bifurcation, and thus the stable point loses its stability.
Another reason is that external perturbations may push the system’s state out of the basin boundary
of the stable point. Therefore, evaluating the stability of dynamical systems is important to avoid
undesirable system-wide events.
However, the precise mathematical expression of the dynamics is usually unknown for real-world
complex systems, and thus we need to characterize the systems that have low stability from observa-
tions of the state variables. Many approaches have been investigated. The simplest way is based on
the values of one or more state variables at a single time point. An anomaly of the target system is
detected if a value is out of the predefined range. In addition, the sample mean values over multiple
time points are also used to improve the reliability. However, this approach fails if the time series

166
Nonlinear Theory and Its Applications, IEICE, vol. 9, no. 2, pp. 166–184 IEICE
c 2018 DOI: 10.1587/nolta.9.166
data of the observations has a slowly changing component and requires a detrending procedure [5].
In such a case, the mean values are shifted to be nearly constant. Moreover, the mean values does
not change before certain types of bifurcations such as a Hopf bifurcation [6]. The second approach is
based on model fitting [7, 8]. The stability of the target system can be evaluated from the estimated
parameters. However, this approach cannot be applied when the mathematical model to be fitted is
unknown or the measurement intervals are too large. The third approach is based on the recovery
rate [9–11]. This method requires to give a perturbation to the target system, and the system’s sta-
bility is estimated from the relaxation process. Therefore, its usage is limited to the systems that can
be easily perturbed.
Another approach which we focus on in this paper is based on the covariance matrix of the state
variables and its related statistics including variance, standard deviation, and Pearson’s correlation
coefficient. The increase in the variance and the standard deviation of state variables has been
observed in many real-world systems before imminent system-wide transitions [1, 3]. The correlation
coefficient between distinct state variables is also shown to serve as an indicator of the stability of
multidimensional dynamical systems [12–16]. The increase in the largest eigenvalue of the covariance
matrix has also been proposed as a possible early warning signal [17–19]. In addition, the eigenvector
corresponding to the largest eigenvalue is also important to identify the state variables that contribute
largely to abnormal fluctuations [15]. While the precise estimation of the covariance matrix from
observations is difficult for high-dimensional systems [20], the estimation of only the largest eigenvalue
and its corresponding eigenvector generally requires much less observations.
The relation between the covariance matrix and the stability of dynamical systems has long been
investigated [21]. Especially, closed-form expressions of the covariance matrix based on the eigenvalues
of the Jacobian matrix have been obtained both for discrete-time and continuous-time dynamical
systems under general assumptions [7, 8]. However, the limiting behavior of the variance, the standard
deviation, and the correlation coefficient when a dynamical system approaches a bifurcation point
has been only recently investigated in the context of early warning signals [15, 22]. The original
analysis in [15], which is called the dynamical network biomarker theory, was mainly for discrete-
time dynamical systems in which the dominant eigenvalue of the Jacobian matrix is real. Although
the theory is also applicable to continuous-time systems by using time discretization, the converted
Jacobian matrix explicitly depend on the discretized time step. In addition, the original analysis
was based on correlation coefficient rather than covariance, and thus division into cases was needed
because the denominator of the correlation coefficient can be zero. Another inconvenience of the
correlation coefficient is that the analysis for the case of complex-valued dominant eigenvalues was
complicated.
The contributions of this paper are as follows. First, we reformulate the theory in [15] in terms of the
covariance matrix and reveal the limiting behavior of the covariance matrix when a dynamical system
approaches a bifurcation point both for discrete-time and continuous-time systems. Our theory also
fully covers the case where the dominant eigenvalues of the Jacobian matrix are a complex conjugate
pair. Second, we derive simple expressions for the covariance matrix when the Jacobian matrix is
normal or symmetric and noise is i.i.d. (independently and identically distributed). These simple
expressions are useful for investigating the quantitative contributions of non-dominant eigenvalues to
the covariance matrix. Although at least a part of those expressions have been obtained previously [19,
23], we would show that all of them can be derived from the discrete-time and continuous-time
Lyapunov equations. Third, we discuss the relation between the eigenvalues and the eigenvectors of
the Jacobian matrix and principal component analysis. This makes it much easier to identify largely
contributing state variables, which are called dynamical network biomarkers [15] or dynamical network
markers [24, 25].
The rest of the paper is organized as follows. In the subsequent four sections, we explain the findings
of the theoretical analysis. Section 2 is for discrete-time linear dynamical systems, Section 3 is for
continuous-time linear dynamical systems, Section 4 is for discrete-time nonlinear dynamical systems,
and Section 5 is for continuous-time nonlinear dynamical systems. In Section 6, we apply the theory
to an example model and show the results of numerical simulations. At last, we state our conclusions

167
and discuss possible extensions in Section 7.

2. Discrete-time linear dynamical systems

2.1 General case
In this section, a discrete-time linear dynamical system of the following form is considered:

xt+1 = Axt + ξt , for t = 0, 1, 2, . . . , (1)

where xt ∈ Rn is the state of the system at time t, A ∈ Rn×n is a time-invariant matrix, and ξt is
noise at time t. In this paper, vectors are in column format by default. This system is also known as
a vector autoregressive (VAR) model of order 1 [26]. The noise term is assumed to be zero-mean and
independent of the system’s state. Let λi ∈ C (i = 1, 2, . . . , n) denote the eigenvalues of A. Then, we
focus on the case where the spectral radius ρ(A) = maxi (|λi |) is less than 1, that is, all the eigenvalues
have magnitude less than 1. In this case, the system’s state fluctuates around the origin due to noise,
and our goal is to clarify the relation between the stability of the system and the covariance matrix
of the stationary distribution of x.
In the beginning, we explain the definition and some properties of covariance matrices. The co-
variance between real-valued random variables X1 and X2 is defined as Cov(X1 , X2 ) = E[(X1 −
E(X1 ))(X2 − E(X2 ))]. Then, the covariance matrix of a n-dimensional real-valued random vec-
tor X = (X1 , X2 , . . . , Xn )T is defined as the matrix whose (i, j) element is Cov(Xi , Xj ). A ma-
trix M ∈ Cn×n is said to be Hermitian if M = M ∗ , where M ∗ = M̄ T denote the conjugate
transpose of M . A Hermitian matrix M ∈ Cn×n is said to be positive-semidefinite and writ-
ten as M  0 if, for any v ∈ Cn , v ∗ M v ≥ 0 holds. If strict inequality holds, M is said to
be positive-definite and written as M  0. It is known that all covariance matrices are positive-
semidefinite. When a covariance matrix is positive-definite, it can be converted to a correlation
matrix whose (i, j) element is the Pearson’s correlation coefficient between Xi and Xj calculated
as Corr(Xi , Xj ) = Cov(Xi , Xj ) Cov(Xi , Xi )−1/2 Cov(Xj , Xj )−1/2 . However, the opposite conversion
from a correlation matrix to a covariance matrix is not always possible.
Let two real positive-definite matrices C, D ∈ Rn×n denote the covariance matrices of the stationary
distributions of x and ξ, respectively. These matrices and A are related as follows:

C = E(xt+1 xTt+1 ),
= E((Axt + ξt )(Axt + ξt )T ),
= E(Axt xTt AT ) + E(Axt ξtT ) + E(ξt xTt AT ) + E(ξt ξtT ),
= ACAT + D, (2)

where we use the fact that all of E(xt ), E(ξt ), E(xt ξtT ), and E(ξt xTt ) are 0. This equation is known
as the discrete-time Lyapunov equation [27], and it is known to have a unique solution C with C  0
for any D  0 if and only if ρ(A) < 1. The same equation appears when ξt is replaced with a control
input But in which BB T = D and ut ∈ Rn . In that case, the solution C is called the controllability
Gramian.
Moreover, since C is symmetric, Eq. (2) can be rewritten as a system of n(n + 1)/2 linear equations
as follows:

Cij = Aik Ckl ATlj + Dij ,
k,l
 
= Aik Ajk Ckk + (Aik Ajl + Ail Ajk )Ckl + Dij , for i ≤ j. (3)
k k<l

Therefore, if A and D are given, C can be obtained by solving Eq. (3) numerically or symbolically,
even if A is non-diagonalizable. If we forget that C is symmetric, Eq. (3) can also be rewritten by
using the Kronecker product. A few alternative expressions of C for general cases are also known [27]
and presented in Appendix A. They are useful for some theoretical analyses such as investigating the

168
limiting behavior explained in the next subsection. However, simpler expressions can be obtained for
particular cases, which gives us deeper insight into the behavior of the system.
It should be noted that once C is obtained, covariances of x between different time steps can be
calculated as follows:

C(m) = E(xt+m xTt ) = Am C, for m = 0, 1, 2, . . . . (4)

2.2 Case of diagonalizable matrix

One useful case is where A is diagonalizable, that is, A can be diagonalized by an invertible matrix
P ∈ Cn×n as A = P ΛP −1 , where Λ ∈ Cn×n is a diagonal matrix whose diagonal elements are
the eigenvalues of A. The kth column vector of P , which is denoted as pk ∈ Cn , is an eigenvector
corresponding to the kth eigenvalue. We assume that each column vector is normalized as pk  =
√ ∗
p p = 1. Thus, any pair of column vectors of P corresponding to a pair of complex conjugate
eigenvalues of A is complex conjugate to each other.
Let us rewrite Eq. (2) as follows:

C = ACA∗ + D. (5)

Then, by left-multiplying P −1 and right-multiplying (P ∗ )−1 = (P −1 )∗ to the both side, we get

P −1 C(P ∗ )−1 = ΛP −1 C(P −1 )∗ Λ∗ + P −1 D(P ∗ )−1 ,

C̃ = ΛC̃Λ∗ + D̃, (6)

where C̃ = P −1 C(P ∗ )−1 and D̃ = P −1 D(P ∗ )−1 [7, 8]. The two matrices C̃, D̃ ∈ Cn×n are positive-
definite because for any invertible matrix U ∈ Cn×n and any positive-definite matrix V ∈ Cn×n ,
U ∗ V U  0 holds.
Since Λ is diagonal, Eq. (6) can be solved in an element-wise manner as follows [7, 8]:

C̃ij = λi λ̄j C̃ij + D̃ij ,

D̃ij
C̃ij = . (7)
1 − λi λ̄j

This expression can be rewritten by using the Hadamard product (element-wise multiplication) as
follows:

C̃ = L ◦ D̃, (8)
1
Lij = . (9)
1 − λi λ̄j

Thus, we obtain

C = P C̃P ∗ ,
= P (L ◦ D̃)P ∗ ,
= P (L ◦ (P −1 D(P ∗ )−1 ))P ∗ . (10)

From this expression, we can see how the system behaves when ρ(A) is close to 1. If the dominant
eigenvalue λd that has the largest magnitude is real and its algebraic multiplicity is 1 (that is, A has
exactly one eigenvalue equals to λd ), we get

lim (1 − λ2d )C = D̃dd pd pTd , (11)

λd →±1

where pd ∈ Rn is the dominant eigenvector corresponding to λd . Here, D̃dd is real and positive because
it is a diagonal element of a positive-definite matrix. We can see from Eq. (11) that the covariance
matrix C becomes proportional to the outer product of the dominant eigenvector with itself when λd

169
is close to 1 or −1. In addition, all the off-diagonal elements of the correlation matrix approach either
1 or −1 unless pd contains any element that is strictly 0. The case where pd contains 0-elements has
been investigated in [15].
Similarly, if the dominant eigenvalues are a complex conjugate pair λd , λe = λ̄d and both have
algebraic multiplicity 1, we get

lim (1 − |λd |2 )C = D̃dd pd p∗d + D̃ee pe p∗e ,

|λd |→1

= 2D̃dd Re(pd p∗d ),

 
= 2D̃dd Re(pd )Re(pd )T + Im(pd )Im(pd )T , (12)

where pd , pe = p̄d ∈ Cn are the dominant eigenvectors corresponding to λd and λe , respectively.

Here, a relation D̃dd = D̃ee is used. A proof of this is presented in Appendix B, in which the
linear independence of Re(pd ) and Im(pd ) is also shown. We can see from Eq. (12) that C becomes
proportional to a rank-2 matrix determined only by the dominant eigenvectors when |λd | = |λe | ≈ 1.
In fact, Eqs. (11) and (12) hold even if A is nondiagonalizable. This can be seen from the general
form of C based on the Jordan normal form of A shown in Appendix A.

2.3 Case of normal matrix and i.i.d. noise

The expression for C shown in Eq. (10) can be further simplified when A is normal (that is, AA∗ =
A∗ A) and the noise is i.i.d. In this case, A can be diagonalized by a unitary matrix P ∈ Cn×n .
Therefore, P −1 = P ∗ holds. In addition, the covariance matrix of the noise term becomes D = σ 2 I,
where σ > 0 is the standard deviation of noise and I is an identity matrix of size n.
Under these assumptions, the covariance matrix of x becomes as follows:

C = P (L ◦ (P ∗ (σ 2 I)P ))P ∗ ,
= σ 2 P (L ◦ I)P ∗ ,
= σ 2 P (I − ΛΛ̄)−1 P ∗ , (13)
2 ∗ −1
= σ (I − AA ) ,
2 T −1
= σ (I − AA ) . (14)

Especially, Eq. (13) is useful to investigate the contribution of each eigenvalue and eigenvector. Pre-
cisely, it can be rewritten as follows:

n
pk p∗k
C = σ2 . (15)
k=1
1 − |λk |2

It should be noted that complex eigenvectors always appear in complex conjugate pairs because A is
real. In addition, we have assumed that pk  = 1 for all k. Thus, if pk is not a real vector, there
exists pl such that pl = p̄k and pk p∗k + pl p∗l = 2 Re(pk p∗k ).

2.4 Case of symmetric matrix and i.i.d. noise

A symmetric matrix is a special case of a normal matrix. If A is symmetric (that is, AT = A),
all of the vectors and the matrices given so far become real-valued, and thus the situation is much
simplified. For i.i.d. noise, we can use the following three equivalent expressions:

C = σ 2 P (I − Λ2 )−1 P T , (16)
n
pk pTk
= σ2 , (17)
k=1
1 − λ2k
= σ 2 (I − A2 )−1 . (18)

A similar expression to the right hand side of Eq. (17) has been obtained in [23].

170
 Among the three expressions, the first one is particularly important because it is of the form of the
eigendecomposition of the covariance matrix C in the real space. Therefore, it is related to principal
component analysis (PCA) and singular value decomposition (SVD) [28], which are associated with
the eigendecomposition of sample covariance matrices. Let x1 , x2 , . . . , xN ∈ Rn denote N observations
of a n-dimensional vector. For the case of a dynamical system, they are observations made at different

time steps. Let m ∈ Rn denote the sample mean vector defined as m = (1/N ) N k=1 xk . Then, the
data matrix X ∈ RN ×n is calculated as follows:
T
X = [x1 − m, x2 − m, . . . , xN − m] . (19)

The sample covariance matrix Ĉ ∈ Rn×n corresponding to X is defined as follows:

1
Ĉ = X T X, (20)
N −1
1 
N
Ĉij = ((xk )i − mi )((xk )j − mj ). (21)
N − 1 k=1

On the other hand, PCA is related to the eigendecomposition of X T X written as follows:

X T X = W RW T , (22)

where W ∈ Rn×n is an orthogonal matrix whose columns are called weight vectors or loading vectors,
and R ∈ Rn×n is a diagonal matrix whose diagonal elements are nonnegative and are arranged in
descending order. By using W , the data matrix X is transformed as follows:

T = XW, (23)

where T ∈ RN ×n is called a score matrix in which the kth column vector corresponds to the kth
principal component.
Usually, the eigendecomposition of X T X is not explicitly calculated, but SVD is performed instead
as follows:

X = U SW T , (24)

where U ∈ RN ×N is another orthogonal matrix, and S ∈ RN ×n is a rectangular diagonal matrix

whose diagonal elements s1 ≥ s2 ≥ . . . ≥ 0 are called singular values. Since R = S T S holds, the
eigenvalues of X T X are equal to the squared singular values.
When N is sufficiently large and all the eigenvalues of C are distinct, we can relate the analytical
eigendecompositions of C shown in Eq. (16) and the empirical eigendecomposition of X T X shown in
Eq. (22). By comparing the two equations, we see that the kth eigenvector of A corresponds to the
kth loading vector, and the eigenvalues are related as follows:
σ2 s2i
= , for i = 1, 2, . . . , n. (25)
1 − λ2i N −1

Since s2i increases in proportion to N for large N , the right hand side does not go to 0 as N → ∞.
These relations indicate that we can estimate both the eigenvalues and the eigenvectors of A from
observations if we know that A is symmetric and D = σ 2 I.
In addition, the dominant eigenvectors can be estimated by using PCA and SVD for general A
with ρ(A) ≈ 1 and D  0. In the case of Eq. (11), most variance in data is explained by the first
principal component, and the first loading vector approaches the dominant eigenvector pd . This does
not mean that the stationary distribution of x shrinks in the other directions. Rather, it expands in
the direction of pd so that its relative contribution to C becomes dominant. In the case of Eq. (12),
the first and the second principal components explain most variance, and the 2-dimensional linear
subspace spanned by the first and the second loading vectors approaches the subspace spanned by
Re(pd ) and Im(pd ).

171
3. Continuous-time linear dynamical systems
3.1 General case
In this section, a continuous-time linear dynamical system of the following form is considered:

dxt = Axt dt + B dwt , for t ≥ 0, (26)

where xt ∈ Rn is the state of the system at time t, A, B ∈ Rn×n are time-invariant matrices, and
wt ∈ Rn is a vector of independent Wiener processes. This system is also known as a multivariate
Ornstein-Uhlenbeck (OU) process [29]. Let λi ∈ C (i = 1, 2, . . . , n) denote again the eigenvalues of A.
Then, we assume that Re(λi ) < 0 for all λi . In this case, A is called an asymptotically stable matrix
or a Hurwitz matrix.
This continuous-time dynamics can be converted to a discrete-time dynamics by using Euler-
Maruyama method [30] as follows:
√
xt+Δt = xt + Δt Axt + Δt ξt , for t = 0, Δt, 2Δt, . . . , (27)

where Δt > 0 is a discretized time step, and ξt ∈ Rn is a random vector following N (0, D). Here, we
set D = BB T .
Let C ∈ Rn×n denote again the covariance matrix of x. Then, we obtain the following relation:

∂C E(xt+Δt xTt+Δt ) − E(xt xTt )

= lim ,
∂t Δt→0 Δt
√ √
E((xt + Δt Axt + Δt ξt )(xt + Δt Axt + Δt ξt )T ) − E(xt xTt )
= lim ,
Δt→0 Δt
Δt E(Axt xTt ) + Δt E(xt xTt AT ) + (Δt)2 E(Axt xTt AT ) + Δt E(ξt ξtT )
= lim ,
Δt→0 Δt
= AC + CAT + D = 0, (28)

where we use the fact that all of E(xt ), E(ξt ), E(xt ξtT ), and E(ξt xTt ) are 0. This equation is known as
the continuous-time Lyapunov equation [27], and it is known that C with C  0 uniquely exists for
any D  0 if and only if A is asymptotically stable. The solution C again becomes the controllability
Gramian when the noise term is replaced with a control input. Since C is symmetric, Eq. (28) can
be rewritten as a system of n(n + 1)/2 linear equations as follows:
 
0= Aik Ckj + Cik ATkj + Dij ,
k k
   
= Aik Ckj + Aik Cjk + Cki Ajk + Cik Ajk + Dij , for i ≤ j. (29)
k<j k≥j k<i k≥i

This system of linear equations can be rewritten by using the Kronecker product where all n2 elements
of C are treated as unknown variables.
It should also be noted that covariances of x at different time points can be written by using C as
follows [29]:

C(s) = E(xt+s xTt ) = exp(As)C, for s ≥ 0, (30)

∞ k
where exp(As) is the matrix exponential of As, that is, exp(As) = k=0 (As) /k!. The equation
above can be derived from ∂C(s)/∂s = AC(s).

3.2 Case of diagonalizable matrix

If matrix A is diagonalizable, there exists an invertible matrix P ∈ Cn×n satisfying A = P ΛP −1 ,
where Λ ∈ Cn×n is diagonal. The column vectors of P are the eigenvectors of A, and the diagonal
elements of Λ are the eigenvalues of A. The norm of each column vector of P is normalized to 1.
Let us rewrite Eq. (28) as follows:

172
 AC + CA∗ + D = 0. (31)

Then, by left-multiplying P −1 and right-multiplying (P ∗ )−1 = (P −1 )∗ to the both side [7], we obtain

ΛP −1 C(P ∗ )−1 + P −1 C(P −1 )∗ Λ∗ + P −1 D(P ∗ )−1 = 0,

ΛC̃ + C̃Λ∗ + D̃ = 0, (32)

where C̃ = P −1 C(P ∗ )−1 and D̃ = P −1 D(P ∗ )−1 . The two matrices C̃, D̃ ∈ Cn×n are positive-definite
again.
Since Λ is diagonal, Eq. (32) can be solved in an element-wise manner as follows [7]:

λi C̃ij + λ̄j C̃ij + D̃ij = 0,

D̃ij
C̃ij = − . (33)
λi + λ̄j
This equation can be rewritten by using the Hadamard product as follows:

C̃ = L ◦ D̃, (34)
1
Lij = − . (35)
λi + λ̄j
Thus, we obtain

C = P C̃P ∗ ,
= P (L ◦ D̃)P ∗ ,
= P (L ◦ (P −1 D(P ∗ )−1 ))P ∗ . (36)

Although this equation seems to be the same as Eq. (10), the definition of L is different.
From this expression, we can see how the system behaves when maxi Re(λi ) is close to 0. If the
dominant eigenvalue λd that has the largest real part is real and its algebraic multiplicity is 1, we get

lim (−2λd )C = D̃dd pd pTd , (37)

λd →0

where pd ∈ Rn is the dominant eigenvector corresponding to λd . Here, D̃dd is real and positive again
because it is a diagonal element of a positive-definite matrix. We can see from Eq. (37) that the
covariance matrix C becomes proportional to the outer product of the dominant eigenvector with
itself when Re(λd ) is close to 0.
Similarly, if the dominant eigenvalues are a complex conjugate pair λd , λe = λ̄d and both have
algebraic multiplicity 1, we get

lim (−2Re(λd ))C = D̃dd pd p∗d + D̃ee pe p∗e ,

Re(λd )→0

= 2D̃dd Re(pd p∗d ),

 
= 2D̃dd Re(pd )Re(pd )T + Im(pd )Im(pd )T , (38)

where pd , pe = p̄d ∈ Cn are the dominant eigenvectors corresponding to λd and λe , respectively. Here,
the relation D̃dd = D̃ee is used again. We can see from Eq. (38) that C becomes proportional to a
rank-2 matrix determined only by the dominant eigenvectors when Re(λd ) = Re(λe ) ≈ 0.
In addition, Eqs. (37) and (38) also hold even when A is not diagonalizable. This can be shown in
a similar manner to the discrete-time case.

3.3 Case of normal matrix and i.i.d. noise

The expression for C shown in Eq. (36) can be further simplified when A is normal and D = σ 2 I.
Here, P becomes unitary so that P −1 = P ∗ holds. In this case, the covariance matrix of x becomes
as follows:

173
 C = P (L ◦ (P ∗ (σ 2 I)P ))P ∗ ,
= σ 2 P (L ◦ I)P ∗ ,
= −σ 2 P (Λ + Λ̄)−1 P ∗ , (39)
2 ∗ −1
= −σ (A + A ) ,
2 T −1
= −σ (A + A ) . (40)

The third line can be rewritten as

n
pk p∗k
C = −σ 2 , (41)
λ + λ̄k
k=1 k

where pk ∈ Cn is the pth column vector of matrix P .

3.4 Case of symmetric matrix and i.i.d. noise

If A is symmetric and D = σ 2 I, C becomes as follows:
σ2
C=− P Λ−1 P T , (42)
2
σ 2  pk pT
n
=− k
, (43)
2 k=1
λk
σ 2 −1
=− A , (44)
2
where both P and Λ are real. The third equation (44) is a special case of the expression for the
covariance matrix over a time window derived in [19].
Relations to PCA and SVD described in Eqs. (19)–(24) are as follows. When the number of
observations N is sufficiently large and the eigenvalues of C are all distinct, the kth eigenvector of A
corresponds to the kth loading vector, and the eigenvalues are related as follows:
σ2 s2i
− = , for i = 1, 2, . . . , n, (45)
2λi N −1
where si is the ith singular value.

4. Discrete-time nonlinear dynamical systems

In this section, a discrete-time autonomous nonlinear dynamical system with additive noise is consid-
ered. It is described as follows:

xt+1 = f (xt ) + ξt , for t = 0, 1, 2, . . . , (46)

where xt ∈ Rn represents a state of the system at time t, f : Rn → Rn is a nonlinear map, and ξt

is noise at time t. We assume that if the noise term is removed, the system has an asymptotically
stable (that is, Lyapunov stable and attracting [31]) fixed point a ∈ Rn that satisfies f (a) = a, and
that f is continuously differentiable at a. This means that f is differentiable at a and the derivative
is continuous at a. We also assume that the fixed point a is hyperbolic, that is, the Jacobian matrix
of f at a has no eigenvalue on the unit circle in the complex plane. It should be noted that if
the asymptotically stable fixed point is non-hyperbolic (for example, the origin of a one-dimensional
map xt+1 = tanh xt ), our knowledge described in Section 2 cannot be directly applied because the
discrete-time Lyapunov equation (2) does not have a solution. The noise term is again assumed to be
zero-mean and independent of the system’s state. An additional assumption is that the noise intensity
is sufficiently small so that we can perform the linear approximation of f around a as follows:

f (x) ≈ f (a) + J(a)(x − a), (47)

where J(a) ∈ Rn×n is the Jacobian matrix of f at a defined as follows:

174
 
∂fi 
(J(a))ij = . (48)
∂xj x=a

Since a is asymptotically stable and hyperbolic, ρ(J(a)) < 1 holds. By using the approximation,
Eq. (46) becomes as follows:

xt+1 = a + J(a)(xt − a) + ξt ,
x̃t+1 = J(a)x̃t + ξt , (49)

where a new variable x̃t = xt − a is introduced. Thus, all the findings in Section 2 can be applied.
There are three representative cases related to codimension-1 bifurcations. The first case is just
before a saddle-node bifurcation (also known as a fold bifurcation or a tangential bifurcation), in
which a real eigenvalue is close to 1. The second case is just before a period-doubling bifurcation,
in which a real eigenvalue is close to −1. The last case is right before a Neimark-Sacker bifurcation,
where a pair of eigenvalues that are complex conjugate of each other are near the unit circle. The
first and the second cases correspond to Eq. (11), and the third case corresponds to Eq. (12). Based
on these equations, we can characterize the systems that have high risks of system-wide changes and
estimate the dominant eigenvectors from observations.

5. Continuous-time nonlinear dynamical systems

In this section, a continuous-time autonomous nonlinear dynamical system with additive noise is
considered. It is represented as follows:

dxt = f (xt )dt + B dwt , (50)

The definition of xt and f is the same to that in Section 4. The definition of wt and B is the same
to that in Section 3. We assume that the noiseless system ẋt = f (xt ) has an asymptotically stable
equilibrium point a ∈ Rn that satisfies f (a) = 0, and that f is continuously differentiable at a. We
also assume that a is hyperbolic, that is, the Jacobian matrix of f at a has no eigenvalue with the
zero real part, and a has no center manifold. An example of non-hyperbolic equilibrium points is
the origin of a one-dimensional differential equation ẋt = −x3t . The assumption of hyperbolicity is
needed in order for the continuous-time Lyapunov equation (28) to have a solution. Then, the linear
approximation of f around a as shown in Eq. (47) gives

dxt = J(a)(xt − a)dt + B dwt ,

dx̃t = J(a)x̃t dt + B dwt , (51)

where a variable transformation x̃t = xt − a is used. Here, all the eigenvalues of J(a) have negative
real parts because a is asymptotically stable and hyperbolic. Thus, all the findings in Section 3 can
be applied.
The relation between the Jacobian matrix J(a) and the covariance matrix C gives us a way to
characterize the situation where at least one eigenvalue of J(a) has a real part close to 0. There are
two representative cases. The first case is just before either a saddle-node bifurcation, a transcritical
bifurcation, or a pitchfork bifurcation. In this case, one real eigenvalue is located near 0. This case
corresponds to Eq. (37). The second case is right before a Hopf bifurcation, in which two eigenvalues
that are complex conjugate of each other have real parts close to 0. This case corresponds to Eq. (38).

6. Numerical simulations
6.1 Example model
In this section, we demonstrate how to apply the theory to nonlinear dynamical systems in practice
by using an example model. Let us consider a network system of n nodes in which each node has its
intrinsic dynamics and also interacts with its neighbors. The intrinsic dynamics of node i is based on
the following function:

175
 Fi (xi ) = xi − x3i − μi , (52)

where xi ∈ R is the state of node i, and μi ∈ R is a parameter. This function is the negative
derivative of a double-well potential Ui (xi ) = −x2i /2 + x√
4
i /4 + μi xi . The isolated dynamical system
ẋi = Fi (xi ) exhibits saddle-node bifurcations at μi = ±2 3/9. We choose this model for simplicity,
but more realistic ecological models with a few parameters are often used in the studies of early
warning signals [10, 12, 25, 32–34].
Let G = (V, E) denote the network structure of a random geometric graph [35] in which V is the
vertex set and E ⊆ V × V is the edge set. The node positions are randomly located in [0, 1]2 , and two
nodes are connected by an edge if the Euclidean distance between them is less than or equal to the
threshold d. Let LG ∈ Zn×n denote the Laplacian matrix of G. LG is a symmetric matrix whose ith
diagonal element is the degree of node i (that is, the number of edges incident to node i), and (i, j)
element for i = j is −1 if node i and node j are adjacent and 0 otherwise.
Then, the dynamics of the network system is described by the following stochastic differential
equation:

dx(t) = [F (x(t)) + αM x(t)] dt + B dw(t), (53)

where x = (x1 , . . . , xn )T is the state of the system, F (x) = (F1 (x1 ), . . . , Fn (xn ))T corresponds to
the intrinsic dynamics, α ≥ 0 is the coupling strength, M = −LG is the sign-reversed Laplacian
matrix of G, and w ∈ Rn is a vector of independent Wiener processes. It should be noted that the
 
ith element of M x(t) can be rewritten as j Mij xj (t) = j,(i,j)∈E (xj (t) − xi (t)), which is known as
diffusive coupling. The covariance matrix of noise D = BB T is defined as Dij = σ 2 exp(−γri −rj 2 ),
where σ, γ > 0 are parameters, and ri ∈ [0, 1]2 represents the position of node i. This means that
close nodes receive much correlated noise than distant nodes. B is calculated from D by using the
Cholesky decomposition. It should be noted that D  0 holds because it is a Gram matrix of the
Gaussian kernel [36].
We induce system-wide state transitions in this network model according to the following procedure.
Let xon and xoff denote the higher and the lower stable equilibrium points of the isolated system
ẋi = Fi (xi ), respectively. Initially, all nodes are set to xon . Then, in each iteration, we randomly
select a node that has not been selected yet, deactivate it by setting its value to xoff , and update a
with the stable equilibrium point that the system without noise reaches eventually. Let noff denote
the number of deactivated nodes. This does not include the nodes that are not selected but fall down
due to interaction.

6.2 Calculation of the covariance matrix

The theoretical values of the covariance matrix C is obtained in three steps. First, we need to
numerically calculate the position of a stable equilibrium point a that satisfies F (a) + αM a = 0.
Next, the Jacobian matrix at a is calculated as
∂F1 (a1 ) ∂Fn (an )
J(a) = diag( ,..., ) + αM. (54)
∂x1 ∂xn
Here, ∂Fi (ai )/∂xi = 1 − 3a2i . Finally, C is calculated according to Eq. (36).
On the other hand, the sample covariance matrix Ĉ is obtained as follows. First, we need to know
again the position of a stable equilibrium point a. Next, a trajectory of the stochastic differential
equation (53) is calculated numerically. In this example, we use the Euler-Maruyama method [30] to
discretize time as follows:
√
x(τ + 1) = x(τ ) + Δt [F (x(τ )) + αM x(τ )] + Δt B (τ ), for τ = 0, 1, 2, . . . , τmax , (55)

where we set τ = t/Δt, and (τ ) is a random vector following N (0, I). The initial state x(0) is set to
a. Since the multivariate OU process is correlated in the time domain, we only take N points each
of which is separated by τmax /N steps. Next, a data matrix X is constructed according to Eq. (19).
Then, the sample covariance matrix is calculated according to Eq. (20).

176
6.3 Simulation results
In the following simulations, we set μi = 0.25 for all i. Other parameters are n = 100, d = 0.25,
α = 0.02, σ = 0.001, γ = 10, Δt = 0.01, τmax = 105 , and N = 500.
Figure 1 shows an example of system-wide state transitions in the network model. In this case,
most nodes fall down at noff = 13. Since the precise timing varies according to the node positions
and the order of the node selection, we focus on this particular case henceforth. Notice that the mean
value of a clearly decreases until noff = 12, and thus the phenomenon is possibly used as an early
warning signal, too.
Figure 2 shows the covariance matrix at the beginning (noff = 0) and right before the system-wide
state transition (noff = 12). The theoretical predictions and the results of the numerical simulations

Fig. 1. Relation between the number of deactivated nodes (noff ) and the
average of the state variables evaluated at a.

Fig. 2. Covariance matrices of the example model. (A) C at noff = 0, (B) Ĉ

at noff = 0, (C) C at noff = 12, and (D) Ĉ at noff = 12. The nodes are sorted
based on the Euclidean distance between their positions by using hierarchical
clustering with the average linkage method.

177
match well in both cases. At noff = 12, a part of nodes show particularly high variances and also
high covariances between them. Notice that the maximum values of the color scales in Figs. 2(C) and
2(D) are several times larger than those in Figs. 2(A) and 2(B).
On the other hand, it would be beneficial if a statistic can characterize the change of the covariance
matrix. As an example, Fig. 3 shows the change in the largest eigenvalue of the covariance matrix.
Its value tends to increase until noff = 12 both in theory and in the numerical simulations, although
the simulation result is noisy. This supports the idea that the increase in the largest eigenvalue of the
sample covariance matrix, which can be calculated only from measurements of the system, may serve
as an early warning signal [17–19].
Another important aspect provided by the covariance matrix is which nodes contribute largely to
the fluctuation just before the state transition. Figure 4 shows the state of the network at noff = 12.
It can be seen that many active nodes (close to xon ) around the five inactive nodes (close to xoff ) at
the bottom right region show large contribution. In addition, high degree nodes tend to have large
contribution in this model. This kind of information may help to take countermeasures for preventing
system-wide changes.

Fig. 3. Relation between the number of deactivated nodes (noff ) and the
largest eigenvalue of the covariance matrix.

Fig. 4. Network view of the example model at noff = 12. Blue nodes are
active (close to xon ) and red nodes are inactive (close to xoff ). The diameter of
each node is proportional to the absolute value of the corresponding element
of the dominant eigenvector of C.

178
7. Discussion and Conclusions
In this paper, we have investigated how the covariance matrix of the stationary distribution of the state
of a dynamical system is associated with the eigenvalues and the eigenvectors of the Jacobian matrix
of the system. Our major contributions are as follows. First, we have reformulated the theory in [15]
in terms of the covariance matrix and revealed the asymptotic behavior of the covariance matrix when
a dynamical system approaches a bifurcation point both for discrete-time and continuous-time cases,
and both for real-valued and complex-valued dominant eigenvalues. Second, we have derived several
simple expressions for the covariance matrix from the discrete-time and continuous-time Lyapunov
equations when the Jacobian matrix is normal or symmetric and noise is i.i.d. Third, we have pointed
out that the eigenvalues and the eigenvectors of the Jacobian matrix correspond to the singular values
of SVD and the loading vectors of PCA, respectively, when the Jacobian matrix is symmetric and noise
is i.i.d. In addition, we have also applied the theory to a nonlinear network model and demonstrated
how the covariance matrix changes before a system-wide state transition.
We make some remarks regarding the extension of this work. First, we only considered dynamical
noise that perturbs the true state of the system but did not consider observation noise that only
distorts the measurement. However, the effect of observation noise in its simplest form is easy to
incorporate into the theory. Let yt = xt + ηt denote the observation at time t where ηt is the
observation noise independent of xt . Then, the covariance matrix of the stationary distribution of yt
is simply the summation of the covariance matrices of xt and ηt .
Second, the current theory is based on linear approximation and cannot be used when noise is large
or the asymptotically stable fixed/equilibrium point is non-hyperbolic. This is because the linear
approximation as shown in Eq. (47) is based on a multivariate Taylor expansion of a nonlinear map
f at a stable fixed/equilibrium point a, and the higher order terms are neglected. Let us assume here
that a is hyperbolic for simplicity. Then, this approximation is accurate as long as the state variable
x stays close to a so that the higher order terms are negligible. However, the accuracy becomes worse
as x moves away from a because the influence of the higher order terms becomes large. Therefore,
the theory based on the linear approximation cannot be applied to the cases where the position of x
is disturbed by large noise [37].
In that case, the stationary distribution is no longer a normal distribution, and we need to take
into account the effect of nonlinearity. Analytic expressions of the stationary distribution have been
obtained for particular cases by solving the Fokker-Planck equation [38, 39]. In addition, the con-
trollability Gramian which corresponds to the covariance matrix of an uncontrolled system has been
extended to incorporate nonlinear dynamics [40]. The relative size of the basin of attraction [41]
and its extension to finite retention time [42] are also important aspects to characterize the stability
of dynamical systems. A phenomenon called flickering can also occur when noise is large [34, 43].
Integrating these knowledge to construct a robust indicator for predicting system-wide changes in
complex systems is a challenging future problem.
Other directions of future works regarding the theoretical underpinning include state-dependent
noise, time delay, nonstationarity, transient dynamics, and other types of attractors such as periodic
orbits. On the other hand, development of indicators both for general and specific purposes is also
important.
Last but not least, practical applications of the theoretical framework refined in this study is a
matter of high priority. Although the original theory in [15] has been already applied to several
real data such as type 1 diabetes [44], type 2 diabetes [45], DNA methylation [46], foreign exchange
market [24], and cell differentiation [47, 48], it is our important future work to utilize our new findings
to broaden the applicability of the theory.

Acknowledgments
We would like to thank Dr. Koji Iwayama for pointing out the relation between dynamical network
markers and principal component analysis. We would also like to thank Dr. Yuji Okamoto for sug-
gesting counterparts in the field of control theory. This research is partially supported by JSPS

179
KAKENHI Grant No. JP15H05707.

Appendix
A. Alternative forms of C
Here, we present a few alternative expressions of C as the solution of the discrete-time Lyapunov
equation (2) for ρ(A) < 1 and D  0. The first one is useful to see that Eqs. (11) and (12) hold in
general. Let Λ ∈ Cn×n be a Jordan normal form of A that satisfies A = P ΛP −1 , where P ∈ Cn×n is
an invertible matrix. Each column vector of P is a generalized eigenvector of A, and we assume that
their norms are normalized to 1. It should be noted that complex generalized eigenvectors always
appear in complex conjugate pairs. Λ is a block diagonal matrix written as follows:
⎡ ⎤
Λ1
⎢ ⎥
⎢ Λ2 ⎥
Λ=⎢ ⎢ ⎥, (A-1)
.. ⎥
⎣ . ⎦
Λm
⎡ ⎤
λk 1
⎢ ⎥
⎢ λk . .
. ⎥
⎢ ⎥
Λk = ⎢ ⎥ , for k = 1, 2, . . . , m, (A-2)
⎢ .. ⎥
⎣ . 1 ⎦
λk

where m is the number of the blocks and λk is the eigenvalue related to the kth block. Let n(k)
denote the size of Λk . From similar calculations to Eq. (6), we get

C̃ = ΛC̃Λ∗ + D̃, (A-3)

where C̃ ∈ Cn×n and D̃ ∈ Cn×n satisfy C = P C̃P ∗ and D = P D̃P ∗ . Let us partition C̃ and D̃ into
blocks as follows:
⎡ ⎤
C11 C12 ··· C1m
⎢ ⎥
⎢ C21 C22 ··· C2m ⎥
C̃ = ⎢
⎢ .. .. .. .. ⎥,
⎥ (A-4)
⎣ . . . . ⎦
Cm1 Cm2 ··· Cmm
⎡ ⎤
D11 D12 ··· D1m
⎢ ⎥
⎢ D21 D22 ··· D2m ⎥
D̃ = ⎢
⎢ .. .. .. .. ⎥,
⎥ (A-5)
⎣ . . . . ⎦
Dm1 Dm2 ··· Dmm

where submatrices Ckl and Dkl have n(k) rows and n(l) columns. By using them, Eq. (A-3) can be
rewritten in a blockwise manner as follows [27]:

Ckl = Λk Ckl Λ∗l + Dkl , for 1 ≤ k, l ≤ m. (A-6)

Thus, we can solve it for each block separately. Element-wise equations are given as follows [27]:

(Ckl )ij = λk λ̄l (Ckl )ij + λk (Ckl )i,j+1 + λ̄l (Ckl )i+1,j + (Ckl )i+1,j+1 + (Dkl )ij ,
for 1 ≤ i ≤ n(k), 1 ≤ j ≤ n(l), (A-7)

where we define (Ckl )n(k)+1,j = (Ckl )i,n(l)+1 = (Ckl )n(k)+1,n(l)+1 = 0. It can be seen that (Ckl )ij is
determined if λk λ̄l = 1 and all of (Ckl )i,j+1 , (Ckl )i+1,j , and (Ckl )i+1,j+1 have been calculated. Since
|λk | < 1 for all k and (Ckl )ij values can be determined iteratively from the right and the bottom
edges, all elements of Ckl can be calculated. Moreover, all elements of Ckl are rational functions of
λk and λ̄l , and their denominators are powers of (1 − λk λ̄l ). Since every Ckl is determined similarly,
C̃ is constructed by combining them, and the solution of Eq. (2) is obtained as C = P C̃P ∗ .

180
 Next, we explain another expression for C, which is also valuable for some analysis. From Eq. (2),
we get the following expression:

C = ACAT + D, (A-8)
T T
= A(ACA + D)A + D, (A-9)
m T m T m−1 T m−1
= A C(A ) + D + ADA + . . . + A D(A ) , for m = 1, 2, . . . . (A-10)

By taking the limit of m → ∞, the first term vanishes because all the eigenvalues of A have magnitude
less than 1 (for a rigorous proof of this, see [49]). Thus, C becomes a series written as follows:
∞

C= Am D(AT )m . (A-11)
m=0

We already know that this converges because otherwise C cannot be determined. From this expression,
it can be seen that C is real and positive-definite because (i) D is real and positive-definite, (ii) A is
real, (iii) for any matrix U ∈ Cn×k and any positive-definite matrix V ∈ Cn×n , U ∗ V U  0 holds, and
(iv) a summation of a positive-definite matrix and a positive-semidefinite matrix is positive-definite.
Similarly, alternative expressions of C as the solution of the continuous-time Lyapunov equation (28)
are also known. A blockwise expression based on the Jordan normal form of A can be similarly
calculated. From it, we can see that Eqs. (37) and (38) hold in general. The expression shown in
Eq. (A-11), which is based on a summation over infinite number of matrices, becomes an integral over
an infinite range as follows [27]:
∞
C= exp(At )D exp(AT t )dt . (A-12)
0

B. Proof of D̃dd = D̃ee

Here, we provide a proof for D̃dd = D̃ee .
Let S1 = {p1 , . . . , pm , p̄1 , . . . , p̄m , w1 , . . . , wn−2m } denote a set of n-dimensional vectors, and let i
denote the imaginary unit only in this section. The first 2m vectors are complex-valued written as
pk = uk + i vk ∈ Cn (uk , vk ∈ Rn ). The remaining vectors are real-valued, that is, wk ∈ Rn . Let
S2 = {u1 , . . . , um , v1 , . . . , vm , w1 , . . . , wn−2m } denote another set of vectors, in which all element are
real vectors.
Lemma. If the vectors in S1 are linearly independent, then the vectors in S2 are linearly independent.

Proof. Suppose that the vectors in S2 are not linearly independent. Then, there exists a set of
real-valued coefficients ak , bk , and ck that are not all zero, and


m 
n−2m
(ak uk + bk vk ) + ck wk = 0, (B-1)
k=1 k=1

holds. However, this leads to the following equality:

m 
  n−2m

ak − i bk ak + i bk
( )(uk + i vk ) + ( )(uk − i vk ) + ck wk = 0. (B-2)
k=1
2 2 k=1

This contradicts the supposition that the vectors in S1 are linearly independent.

Next, let us consider a linear combination of the vectors in S1 :


m 
n−2m
q= ((ak + i bk )(uk + i vk ) + (ck + i dk )(uk − i vk )) + (ek + i fk )wk , (B-3)
k=1 k=1

where ak , bk , ck , dk , ek , and fk are real coefficients. Assume that the vectors in S1 are linearly
independent. When q is real, we obtain

181
 
m 
n−2m
Im(q) = ((bk + dk )uk + (ak − ck )vk ) + fk wk = 0. (B-4)
k=1 k=1

Since uk , vk , and wk are linearly independent from the lemma, bk = −dk , ak = ck , and fk = 0
hold for all k. This means that, if a linear combination gives a real value, every pair of coefficients
corresponding to a complex conjugate vector is complex conjugate to each other.
Let us introduce a new variable ζ ∈ Cn that satisfies P ζ = ξ and E(ζζ ∗ ) = P −1 D(P −1 )∗ = D̃. Each
column vector of P corresponds to an eigenvector of A, and we have assumed that each column vector’s
norm is normalized to 1. Since A is real, all complex eigenvectors appear in complex conjugate pairs.
Moreover, since P is invertible, its column vectors are linearly independent. Thus, the discussion
above can be applied by regarding the column vectors of P as S1 . Let ζd and ζe denote the two
elements corresponding to the dominant complex eigenvalues of A. Since ξ is real, they are complex
conjugate to each other. Thus, we obtain

D̃dd = E(ζd ζ̄d ) = E(ζ̄d ζd ) = E(ζe ζ̄e ) = D̃ee . (B-5)

References
[1] M. Scheffer, J. Bascompte, W.A. Brock, V. Brovkin, S.R. Carpenter, V. Dakos, H. Held, E.H.
van Nes, M. Rietkerk, and G. Sugihara, “Early-warning signals for critical transitions,” Nature,
vol. 461, pp. 53–59, September 2009. doi: 10.1038/nature08227.
[2] M. Scheffer, S.R. Carpenter, T.M. Lenton, J. Bascompte, W. Brock, V. Dakos, J. van de Koppel,
I.A. van de Leemput, S.A. Levin, E.H. van Nes, M. Pascual, and J. Vandermeer, “Anticipating
critical transitions,” Science, vol. 338, pp. 344–348, October 2012. doi: 10.1126/science.1225244.
[3] S.R. Carpenter and W.A. Brock, “Rising variance: A leading indicator of ecological transition,”
Ecol. Lett., vol. 9, pp. 311–318, March 2006. doi: 10.1111/j.1461-0248.2005.00877.x.
[4] H. Held and T. Kleinen, “Detection of climate system bifurcations by degenerate fingerprinting,”
Geophys. Res. Lett., vol. 31, L23207, December 2004. doi: 10.1029/2004GL020972.
[5] V. Dakos, M. Scheffer, E.H. van Nes, V. Brovkin, V. Petoukhov, and H. Held, “Slowing down
as an early warning signal for abrupt climate change,” Proc. Natl. Acad. Sci. U.S.A., vol. 105,
pp. 14308–14312, September 2008. doi: 10.1073/pnas.0802430105.
[6] E.A. Gopalakrishnan, Y. Sharma, T. John, P.S. Dutta, and R.I. Sujith, “Early warning signals
for critical transitions in a thermoacoustic system,” Sci. Rep., vol. 6, 35310, October 2016. doi:
10.1038/srep35310.
[7] A. Neumaier and T. Schneider, “Multivariate autoregressive and Ornstein-Uhlenbeck processes:
Estimates for order, parameters, spectral information, and confidence regions,” ACM Trans.
Math. Software, 1998.
[8] A. Neumaier and T. Schneider, “Estimation of parameters and eigenmodes of multivariate
autoregressive models,” ACM Trans. Math. Software, vol. 27, pp. 27–57, March 2001. doi:
10.1145/382043.382304.
[9] S.L. Pimm, “The complexity and stability of ecosystems,” Nature, vol. 307, pp. 321–326, January
1984. doi: 10.1038/307321a0.
[10] E.H. van Nes and M. Scheffer, “Slow recovery from perturbations as a generic indicator of a
nearby catastrophic shift,” Am. Nat., vol. 169, pp. 738–747, June 2007. doi: 10.1086/516845.
[11] A.J. Veraart, E.J. Faassen, V. Dakos, E.H. van Nes, M. Lürling, and M. Scheffer, “Recovery
rates reflect distance to a tipping point in a living system,” Nature, vol. 481, pp. 357–359,
January 2012. doi: 10.1038/nature10723.
[12] V. Dakos, E.H. van Nes, R. Donangelo, H. Fort, and M. Scheffer, “Spatial correlation as
leading indicator of catastrophic shifts,” Theor. Ecol., vol. 3, pp. 163–174, August 2010. doi:
10.1007/s12080-009-0060-6.
[13] J.M. Drake and B.D. Griffen, “Early warning signals of extinction in deteriorating environ-
ments,” Nature, vol. 467, pp. 456–459, September 2010. doi: 10.1038/nature09389.

182
[14] D.J. Fenn, M.A. Porter, S. Williams, M. McDonald, N.F. Johnson, and N.S. Jones, “Temporal
evolution of financial-market correlations,” Phys. Rev. E, vol. 84, 026109, August 2011. doi:
10.1103/PhysRevE.84.026109.
[15] L. Chen, R. Liu, Z.-P. Liu, M. Li, and K. Aihara, “Detecting early-warning signals for sudden
deterioration of complex diseases by dynamical network biomarkers,” Sci. Rep., vol. 2, 342,
March 2012. doi: 10.1038/srep00342.
[16] I.A. van de Leemput, M. Wichers, A.O.J. Cramer, D. Borsboom, F. Tuerlinckx, P. Kuppens,
E.H. van Nes, W. Viechtbauer, E.J. Giltay, S.H. Aggen, C. Derom, N. Jacobs, K.S. Kendler,
H.L.J. van der Maas, M.C. Neale, F. Peeters, E. Thiery, P. Zachar, and M. Scheffer, “Critical
slowing down as early warning for the onset and termination of depression,” Proc. Natl. Acad.
Sci. U.S.A., vol. 111, pp. 87–92, January 2014. doi: 10.1073/pnas.1312114110.
[17] W. Brock and S. Carpenter, “Variance as a leading indicator of regime shift in ecosystem
services,” Ecol. Soc., vol. 11, no. 2, 9, 2006. doi: 10.5751/ES-01777-110209.
[18] V. Dakos and J. Bascompte, “Critical slowing down as early warning for the onset of collapse in
mutualistic communities,” Proc. Natl. Acad. Sci. U.S.A., vol. 111, pp. 17546–17551, December
2014. doi: 10.1073/pnas.1406326111.
[19] H. Moon and T.-C. Lu, “Network catastrophe: Self-organized patterns reveal both the in-
stability and the structure of complex networks,” Sci. Rep., vol. 5, 9450, March 2015. doi:
10.1038/srep09450.
[20] O. Ledoit and M. Wolf, “A well-conditioned estimator for large-dimensional covariance ma-
trices,” J. Multivariate Anal., vol. 88, no. 2, pp. 365–411, 2004. doi: 10.1016/S0047-
259X(03)00096-4.
[21] A.R. Ives, B. Dennis, K.L. Cottingham, and S.R. Carpenter, “Estimating community stability
and ecological interactions from time-series data,” Ecol. Monogr., vol. 73, pp. 301–330, May
2003. doi: 10.1890/0012-9615(2003)073[0301:ECSAEI]2.0.CO;2.
[22] R. Biggs, S.R. Carpenter, and W.A. Brock, “Turning back from the brink: Detecting an im-
pending regime shift in time to avert it,” Proc. Natl. Acad. Sci. U.S.A., vol. 106, pp. 826–831,
January 2009. doi: 10.1073/pnas.0811729106.
[23] T. Nakagawa, “Early warning signals by dynamical network markers,” Bachelor’s thesis, The
University of Tokyo, February 2016. (in Japanese).
[24] S. Oya, K. Aihara, and Y. Hirata, “Forecasting abrupt changes in foreign exchange markets:
Method using dynamical network marker,” New J. Phys., vol. 16, 115015, November 2014. doi:
10.1088/1367-2630/16/11/115015.
[25] T. Nakagawa, M. Oku, and K. Aihara, “Early warning signals by dynamical network mark-
ers,” SEISAN KENKYU, vol. 68, pp. 271–274, May 2016. (in Japanese). doi: 10.11188/
seisankenkyu.68.271.
[26] W.W.S. Wei, Time Series Analysis: Univariate and Multivariate Methods, Pearson Addison
Wesley, 2nd ed., 2006.
[27] Z. Gajic and M.T.J. Qureshi, Lyapunov Matrix Equation in System Stability and Control, Dover
Publications, 2008. (Republication of the Academic Press, San Diego, 1995 edition).
[28] B. Moore, “Principal component analysis in linear systems: Controllability, observability,
and model reduction,” IEEE Trans. Autom. Control, vol. 26, no. 1, pp. 17–32, 1981. doi:
10.1109/TAC.1981.1102568.
[29] C.W. Gardiner, Handbook of Stochastic Methods for Physics, Chemistry and the Natural Sci-
ences, Springer-Verlag Berlin Heidelberg, 2nd ed., 1985.
[30] T. Sauer, “Numerical solution of stochastic differential equations in finance,” in Handbook of
Computational Finance (J.-C. Duan, W.K. Härdle, and J.E. Gentle, eds.), pp. 529–550, Springer
Berlin Heidelberg, 2012. doi: 10.1007/978-3-642-17254-0 19.
[31] S.H. Strogatz, Nonlinear Dynamics and Chaos: With Applications to Physics, Biology, Chem-
istry, and Engineering, Perseus Books Publishing, 1994.
[32] V. Guttal and C. Jayaprakash, “Changing skewness: An early warning signal of regime shifts in
ecosystems,” Ecol. Lett., vol. 11, pp. 450–460, May 2008. doi: 10.1111/j.1461-0248.2008.01160.x.

183
[33] V. Guttal and C. Jayaprakash, “Spatial variance and spatial skewness: Leading indicators of
regime shifts in spatial ecological systems,” Theor. Ecol., vol. 2, pp. 3–12, March 2009. doi:
10.1007/s12080-008-0033-1.
[34] V. Dakos, S.R. Carpenter, W.A. Brock, A.M. Ellison, V. Guttal, A.R. Ives, S. Kéfi, V. Livina,
D.A. Seekell, E.H. van Nes, and M. Scheffer, “Methods for detecting early warnings of critical
transitions in time series illustrated using simulated ecological data,” PLOS ONE, vol. 7, e41010,
July 2012. doi: 10.1371/journal.pone.0041010.
[35] J. Dall and M. Christensen, “Random geometric graphs,” Phys. Rev. E, vol. 66, 016121, July
2002. doi: 10.1103/PhysRevE.66.016121.
[36] C.M. Bishop, Pattern Recognition and Machine Learning, Springer-Verlag New York, 2006.
[37] R. Liu, P. Chen, K. Aihara, and L. Chen, “Identifying early-warning signals of critical transitions
with strong noise by dynamical network markers,” Sci. Rep., vol. 5, 17501, December 2015. doi:
10.1038/srep17501.
[38] R. Graham and H. Haken, “Generalized thermodynamic potential for Markoff systems in de-
tailed balance and far from thermal equilibrium,” Z. Physik, vol. 243, no. 3, pp. 289–302, 1971.
doi: 10.1007/BF01394858.
[39] M. Brics, J. Kaupuzs, and R. Mahnke, “How to solve Fokker-Planck equation treating
mixed eigenvalue spectrum?,” Condensed Matter Physics, vol. 16, no. 1, 13002, 2013. doi:
10.5488/CMP.16.13002.
[40] K. Kashima, “Noise response data reveal novel controllability Gramian for nonlinear network
dynamics,” Sci. Rep., vol. 6, 27300, June 2016. doi: 10.1038/srep27300.
[41] P.J. Menck, J. Heitzig, N. Marwan, and J. Kurths, “How basin stability complements the linear-
stability paradigm,” Nature Phys., vol. 9, pp. 89–92, January 2013. doi: 10.1038/nphys2516.
[42] F. Hellmann, P. Schultz, C. Grabow, J. Heitzig, and J. Kurths, “Survivability of deterministic
dynamical systems,” Sci. Rep., vol. 6, 29654, July 2016. doi: 10.1038/srep29654.
[43] R. Wang, J.A. Dearing, P.G. Langdon, E. Zhang, X. Yang, V. Dakos, and M. Scheffer, “Flicker-
ing gives early warning signals of a critical transition to a eutrophic lake state,” Nature, vol. 492,
pp. 419–422, December 2012. doi: 10.1038/nature11655.
[44] X. Liu, R. Liu, X.-M. Zhao, and L. Chen, “Detecting early-warning signals of type 1 diabetes
and its leading biomolecular networks by dynamical network biomarkers,” BMC Med. Genomics,
vol. 6, Suppl. 2, S8, May 2013. doi: 10.1186/1755-8794-6-S2-S8.
[45] M. Li, T. Zeng, R. Liu, and L. Chen, “Detecting tissue-specific early warning signals for complex
diseases based on dynamical network biomarkers: Study of type 2 diabetes by cross-tissue
analysis,” Brief Bioinform., vol. 15, pp. 229–243, March 2014. doi: 10.1093/bib/bbt027.
[46] A.E. Teschendorff, X. Liu, H. Caren, S.M. Pollard, S. Beck, M. Widschwendter, and L. Chen,
“The dynamics of DNA methylation covariation patterns in carcinogenesis,” PLOS Comput.
Biol., vol. 10, e1003709, July 2014. doi: 10.1371/journal.pcbi.1003709.
[47] P. Chen, R. Liu, L. Chen, and K. Aihara, “Identifying critical differentiation state of MCF-7
cells for breast cancer by dynamical network biomarkers,” Front. Genet., vol. 6, 252, July 2015.
doi: 10.3389/fgene.2015.00252.
[48] A. Richard, L. Boullu, U. Herbach, A. Bonnafoux, V. Morin, E. Vallin, A. Guillemin, N. Papili
Gao, R. Gunawan, J. Cosette, O. Arnaud, J.-J. Kupiec, T. Espinasse, S. Gonin-Giraud, and
O. Gandrillon, “Single-cell-based analysis highlights a surge in cell-to-cell molecular variability
preceding irreversible commitment in a differentiation process,” PLOS Biol., vol. 14, e1002585,
December 2016. doi: 10.1371/journal.pbio.1002585.
[49] R. Oldenburger, “Infinite powers of matrices and characteristic roots,” Duke Math. J., vol. 6,
no. 2, pp. 357–361, 1940. doi: 10.1215/S0012-7094-40-00627-5.

184