Lévy on-off intermittency

Adrian van Kan,∗ Alexandros Alexakis,† and Marc-Etienne Brachet‡

Laboratoire de Physique de l’Ecole normale supérieure, ENS, Université PSL,
CNRS, Sorbonne Université, Université de Paris, F-75005 Paris, France
(Dated: February 18, 2021)
We present a new form of intermittency, Lévy on-off intermittency, which arises from multiplica-
tive α-stable white noise close to an instability threshold. We study this problem in the linear and
nonlinear regimes, both theoretically and numerically, for the case of a pitchfork bifurcation with
fluctuating growth rate. We compute the stationary distribution analytically and numerically from
arXiv:2102.08832v1 [cond-mat.stat-mech] 13 Feb 2021

the associated fractional Fokker-Planck equation in the Stratonovich interpretation. We character-

ize the system in the parameter space (α, β) of the noise, with stability parameter α ∈ (0, 2) and
skewness parameter β ∈ [−1, 1]. Five regimes are identified in this parameter space, in addition to
the well-studied Gaussian case α = 2. Three regimes are located at 1 < α < 2, where the noise has
finite mean but infinite variance. They are differentiated by β and all display a critical transition at
the deterministic instability threshold, with on-off intermittency close to onset. Critical exponents
are computed from the stationary distribution. Each regime is characterised by a specific form of
the density and specific critical exponents, which differ starkly from the Gaussian case. A finite
or infinite number of integer-order moments may converge, depending on parameters. Two more
regimes are found at 0 < α ≤ 1. There, the mean of the noise diverges, and no critical transition
occurs. In one case the origin is always unstable, independently of the distance µ from the deter-
ministic threshold. In the other case, the origin is conversely always stable, independently of µ. We
thus demonstrate that an instability subject to non-equilibrium, power-law-distributed fluctuations
can display substantially different properties than for Gaussian thermal fluctuations, in terms of
statistics and critical behavior.

I. INTRODUTION white noise, hf (t)i = 0, hf (t)f (t0 )i = 2δ(t − t0 ), in terms

of the ensemble average h·i. In this study, we adopt
On-off intermittency is a common phenomenon in the Stratonovich interpretation [22] of eqn. (1), unless
nonequilibrium physical systems, which is characterized stated otherwise. We may take X to be non-negative,
by an aperiodic switching between a large-amplitude since sign changes are incompatible with the exact
“on” state and a small-amplitude “off” state. It was solution of (1) given in [23] and in appendix A. For
originally studied theoretically in the context of low- Gaussian noise, the exact stationary probability density
γ 2
dimensional deterministic chaos and nonlinear maps [1–4] function (PDF) is known to be p(x) = N x−1+µ e− 2 x
and has since then been observed in numerous experi- with normalisation N , for µ > 0 [24]. For µ ≤ 0, the
mental setups ranging from electronic devices [5], spin- distribution approaches δ(x) at late times, with the
wave instabilities [6], liquid crystals [7, 8] and plasmas [9] cumulative distribution function (CDF) (the integral of
to multistable laser fibers [10], sediment transport [11], the PDF up to x) converging to 1 for all x > 0. In that
human balancing motion [12, 13] and blinking quantum case, all moments of the stationary density vanish. For
dots in semiconductor nanocrystals [14, 15]. On-off in- µ > 0 the moments of X scale as hX n i ∝ µcn with the
termittency has also been observed in numerical simula- critical exponents cn = 1 for all n > 0, see [23]. The cn
tions of turbulence in thin layers [16, 17] and magneto- for Gaussian noise are different from their deterministic
hydroydnamic dynamo flows [18–20]. “mean field” values, which are cn = n/2. This is an
From a theoretical perspective, on-off intermittency instance of anomalous scaling, a phenomenon which has
arises in the presence of multiplicative noise close to an received much attention in various areas of physics, in
instability threshold. Therefore it is natural to study it particular in the context of continuous phase transitions
using appropriate stochastic models, such as at equilibrium and critical phenomena [25, 26], as well
as in turbulence [27, 28].
dX
= (f (t) + µ)X − γX 3 , (1)
dt Here, we introduce Lévy on-off intermittency as the
i.e. a supercritical pitchfork bifurcation [21] with a case where f (t) is given by Lévy white noise, whose PDF
fluctuating growth rate, where µ is the deterministic is an α-stable distribution featuring power-law tails as-
growth rate, and f (t) is usually zero-mean, Gaussian, sociated with extreme events in terms of noise amplitude
[29, 30]. The Gaussian distribution (which is a special
case of α-stable distributions) is of fundamental impor-
tance due to its stability: by the central limit theorem
∗ [email protected] [31], it constitutes an attractor in the space of PDFs
† [email protected] with finite variance. Similarly, by the generalized cen-
‡ [email protected] tral limit theorem [32, 33], non-Gaussian α-stable dis-
 2

tributions constitute an attractor in the space of PDFs Here, we show theoretically and numerically that for
whose variance does not exist. Non-Gaussian fluctu- Lévy white noise, the phenomenology of equation (1) can
ations, which may often be modeled as α-stable, are differ starkly from the case of Gaussian white noise. In
found in incompletely thermalized systems or, in gen- some cases, the origin never changes stability – there is
eral, in systems driven away from thermal equilibrium: no critical point. When there is a critical point, the crit-
non-equilibrated heat reservoirs can be considered as a ical behavior and the properties of on-off intermittency
source of non-Gaussian noise [29, 34]. near onset depend non-trivially on the parameters of the
If X(t) solves equation (1) with f (t) being Lévy white Lévy noise. It is shown that in stationary state a finite
noise, then Y = log X(t) is said to perform a Lévy or infinite number of integer-order moments may exist,
flight in a particular anharmonic potential. Lévy flights depending on the parameters of the noise.
were first introduced by Mandelbrot in [35] and have The remainder of this paper is structured as follows.
since found numerous applications, such as anomalous In section II, we present the theoretical background of
diffusion, for instance in different fluid flows, [36–39], this study. In section III, we analyse the linear (γ = 0)
the statistics of 2-D fluid turbulence [40], plasma tur- regime. In section IV, we present analytical results on the
bulence [41], finance [42], climatology [43, 44], animal nonlinear (γ > 0) statistically stationary state and verify
foraging [45, 46], human mobility [47] (although a de- our results against numerical solutions of the stationary
bate about the applicability in the latter two cases is fractional Fokker-Planck and Langevin equations. Fi-
ongoing [48, 49]), COVID-19 spreading [50], human bal- nally in section V, we discuss our results and conclude.
ancing motion [51] and more [52, 53]. Fluctuations obey-
ing heavy-tailed distributions have also been observed for
neuron activity patterns in the human brain [54]. More- II. THEORETICAL BACKGROUND
over, Lévy walks, a class of random processes similar
to Lévy flights with increments following a heavy-tailed Here, we introduce aspects of the theory of stable PDFs
PDF, but with each step taking finite time [55, 56], have and describe how they are related to Lévy flights.
been proposed as a model of blinking quantum dots in
semiconductor nanocrystals [57, 58]. We highlight that
blinking quantum dots and human balancing motion are A. Properties of α-stable probability densities
two examples which exhibit both Lévy statistics and on-
off intermittency. Furthermore, very recently, in an ide- For parameters α ∈ (0, 2], β ∈ [−1, 1], the α-stable
alised model of three-dimensional perturbations in two- PDF for a random variable Y is denoted by ℘α,β (y) and
dimensional flows, described in the companion paper to defined by its characteristic function (i.e. Fourier trans-
this study, [59], it was found that the perturbation am- form),
plitude obeyed equation (1) with an approximately white ( )
noise whose PDF had power-law tails due to the power-
law structure of the velocity fields involved. The findings ϕα,β (k) = exp − |k|α [1 − iβsgn(k)Φ(k)] , (2)
of the companion paper originally motivated the present
study and suggested a rationale for numerically observed with
jump-like growth signals of three-dimensional perturba- (
tions in rapidly rotating turbulence [60]. tan πα


2 α 6= 1
Φ(k) = , (3)
A significant body of theoretical literature is devoted to − π2 log(|k|) α=1
Lévy flights in potentials, driven by additive Lévy noise,
[61–68], as well as to stochastic processes driven by mul- see [33]. A standard method for simulating stable ran-
tiplicative Lévy noise [69–73]. For additive noise, it has dom variables is given in [94]. Note that (2) is not the
been shown that Lévy flights in a quartic or steeper po- most general form possible: there may be a scale param-
tential possess finite mean and variance, for all param- eter in the exponential, which we set equal to one. One
eters of the Lévy noise [63]. Many classical problems refers to α as the stability parameter. For α = 2, where
which are well studied for Gaussian noise have been re- β is irrelevant since Φ = 0, one recovers the Gaussian
visited using Lévy noise, such as the escape from a poten- distribution. In the following, we consider α < 2. The
tial well [74–78], noise-induced transitions and stochastic parameter β, known as the skewness parameter, measures
resonance [79–83], oscillators under the influence of noise the asymmetry of of the distribution, where β = 0 corre-
[62, 84, 85], the Verhulst model [86] and the Lévy ra- sponds to a symmetric PDF, while |β| = 1 is referred to
chet [87]. However, despite this impressive body of work, as maximally skewed. We highlight the symmetry rela-
while the impact of colored noise [23, 88–92] and higher tion
dimensions [93] on on-off intermittency have received at-
tention, the theory of on-off intermittency due to multi- ℘α,β (y) = ℘α,−β (−y), (4)
plicative Lévy noise close to an instability threshold has
not been studied systematically before, to the best of our which follows directly from the definition. Importantly,
knowledge. there are two different possible asymptotic behaviors that
 3

a stable distribution can display. When |β| < 1, there are step dt, we let f (t)dt = dt1/α F (t), where F (t) obeys
two long (“heavy”) power-law tails, at y → ±∞, the alpha-stable PDF ℘α,β (F ), defined by (2), and is
drawn independently for any time t, [39]. Since the
℘α,β (|y| → ∞) ∝ {1 + βsign(y)}|y|−1−α . (5) Langevin equation (1) involves a multiplicative noise
The presence of power-law tails implies that the stable term, one needs to decide on an interpretation thereof.
PDF has a finite mean (equal to zero), but a diverging As has been discussed in the literature, [70, 73], like in
variance for 1 < α < 2, while both mean and variance the Gaussian case, the two standard interpretations are
diverge for α ≤ 1. For β = ±1, the asymptotics given in the Stratonovich [22] interpretation, which preserves the
(5) break down on one side. In this case, there is a short rules of standard calculus, and the non-anticipating Itô
exponential tail on the side where the power law breaks [95] interpretation. According to the choice of interpreta-
down and only a single long power-law tail remains. For tion, the probability density will be governed by a differ-
1 ≤ α < 2, ℘α,β=±1 (y) is supported on R. By contrast, ent form of the (space-)fractional Fokker-Planck equation
for α < 1 and β = ±1, the probability density is one- (FFPE), so called since it involves fractional derivatives
sided, with the exponential tail vanishing at the origin, in the state variable. First consider the Stratonovich in-
such that ℘α,β=1 (y) = 0 at y ≤ 0 and ℘α,β=−1 (y) = 0 at terpretation, such that Y = log(X) obeys the following
y ≥ 0, which is consistent with the symmetry (4). Both equation with additive noise,
for 1 < α < 2, β = −1 as y → +∞, and for α < 1,
dY
β = 1 as y → 0+ , the leading-order asymptotic form = µ − γe2Y + f (t), (7)
of the short tail of the stable PDF can be obtained by dt
Laplace’s method and is given by which says that Y (t) performs a Lévy flight in the poten-
1−α/2 α
tial V (Y ) = −µY + γ2 e2Y . The density associated with
℘α,β (y) ∼ c0 y α−1 exp −c1 y α−1 , (6) Y (t), denoted by py (y, t), then obeys the FFPE
where c0 , c1 are positive, α-dependent constants, cf. the-
∂t py (y, t) = −∂y µ − γe2y py (y, t) +Dyα,β py (y, t), (8)


orem 4.7.1 in [33]. Note that this reduces to a Gaussian
when α = 2. By the symmetry of eq. (4), the same result
[96], where the fractional derivative operator
holds, with y replaced by −y, for 1 < α < 2, β = +1 as
y → −∞ and at α < 1, β = −1 as y → 0− . The differ- α
(1 + β)D+ α
g(y) + (1 − β)D− g(y)
ent behaviors are illustrated for three cases in figure 1. Dyα,β g(y) = − απ

, (9)
Unfortunately, useful explicit expressions for the stable 2 cos 2
PDF only exist in a small number of special cases.
for an arbitrary function g(y), is known as the Riesz-
Feller fractional derivative of order α and skewness β
[97]. It can be expressed in terms of the left and right
0.4 α = 2/3, β = 0 Riemann-Liouville fractional derivatives, which for 1 <
α = 2/3, β = 1 α < 2 are given by [64, 98],
α = 1.5, β = 1
0.3 α 1 d2
Z y
g(z)dz
(D+ g)(y) = (10)
℘α,β (y)

Γ(2 − α) dy 2 −∞ (y − z)α−1
0.2 long tails
and
∞
0.1 d2
Z
α 1 g(z)dz
short tails (D− g)(y) = . (11)
Γ(2 − α) dy 2 y (z − y)α−1
0.0
For 0 < α < 1, the definitions are similar [98],
−10 −5 0 5 10
y α 1 d
Z y
g(z)dz
(D+ g)(y) = + , (12)
Γ(1 − α) dy −∞ (y − z)α
FIG. 1. Illustration of long power-law tails and short expo- and
nential tails in the stable distributions discussed in the text. Z ∞
α 1 d g(z)dz
(D− g)(y) = − . (13)
Γ(1 − α) dy y (z − y)α

B. Lévy flights and the space-fractional For α = 2, one has Dyα,β = ∂y2 . The Riemann-Liouville
Fokker-Planck equation fractional derivatives have a simple Fourier transform,
α
F[D± f ](k) = (±ik)α F[f ](k), see chapter 7 of [98], which
Consider the Langevin equation (1) with f (t) being is often invoked. However, our analysis will be performed
white “Lévy” noise. More precisely, for a given time mostly in physical space. Once the solution to equation
 4

(8) is known, then the probability density px (x, t) asso- A. The Gaussian case
ciated with the original variable X(t) is given by
First, for illustration, consider the Gaussian case α = 2
1 in (16), which gives the log-normal PDF for X
px (x, t) = py (log(x), t). (14)
x
−(log(x) − µt)2
 
1
px (x, t) = √ exp . (18)
x 2πt 2t
If, instead of the Stratonovich interpretation, one adopts
the Itô interpretation, then the FFPE is given by The probability P (x < χ) to find the system at x < χ
after time t is given by the CDF in eq. (17), which here
equals
∂t px (x, t) = −∂x (µx − γx3 )px (x, t) + Dxα,β (xα px (x, t)),
 
(15) 1
 
log(χ) − µt

as derived in [96]. P (x < χ) = 1 + erf √ , (19)
2 2t
We continue in the Stratonovich interpretation. In the where erf(x) is the error function. Considering the limit
absence of nonlinearity, when γ = 0, one can solve in of late times t → ∞ for fixed χ, using that erf(x →
Fourier space for a delta-peaked initial condition, e.g. ±∞) = ±1, one deduces that P (x < χ) → 1 if µ < 0,
X(0) = 1, which leads to the fundamental solution given while P (x < χ) → 0 if µ > 0. This indicates that at
in [97], µ = 0 the origin x = 0 goes from asymptotically stable
to unstable.

log(x)−µt
 Alternatively, one might attempt to determine the sta-
℘α,β t1/α bility of the origin by studying the moments of X as a
px (x, t) = (16) function of time. For α = 2, the FFPE in (8) reduces to
t1/α x
the the ordinary Fokker-Planck equation
where ℘α,β (·) is the α-stable PDF whose Fourier trans-
∂t py (y, t) = −µ∂y py (y, t) + ∂y2 py (y, t). (20)
form is given in eq. (2). The corresponding cumulative
probability distribution (CDF) is Multiplying by exp(ny) = xn and integrating over y, one
arrives, upon integrating by parts, at the relation
log(χ) − µt
 
P (x < χ) = Pα,β , (17) ∂t hX n i = µn + n2 hX n i,


t1/α (21)

Rz which implies that

in terms of the α-stable CDF Pα,β (z) = −∞ ℘α,β (z 0 )dz 0 .
Clearly, equation (16) holds for the Gaussian case of α = hX n (t)i = X0n eλn (µ)t , (22)
2, the familiar log-normal distribution. By analogy with
for an initial condition X(0) = X0 , with the growth rate
the latter, for 0 < α < 2 the PDF in eq. (16) is known
λn (µ) = (nµ+n2 ). Importantly, the value of µ where the
as the log-stable PDF and the associated process as the
growth rate of hX n i vanishes, denoted µc (n), depends on
log-stable process. For 0 < t < ∞, the moments of the
n and is given by µc (n) = −n. We have shown based on
log-stable PDF are only finite for β = −1. This is because
the CDF that the system is stable for µ < 0. However,
it is the only case where the α-stable PDF does not have
equation (22) indicates that for n large enough, hX n i
a heavy tail of the form (5) at +∞. When a heavy tail is
grows exponentially in time even for µ < 0. This is
present (β > −1), then averaging over eny = xn for any
due to rare transient excursions to large y, which give
n > 0 does not give a finite result. For this reason the
a non-negligible contribution since eny is large. Thus,
associated stochastic process with β = −1 is also known
the moments are not the correct indicator for stability in
as the finite-moment log-stable process. It is well known,
the system (1) with γ = 0 and one needs to be careful
in particular in finance, see [99] (there, only 1 < α < 2 is
when concluding stability based on them. However, as
considered).
discussed in [100], the limit of µc (n) as n → 0 does indi-
cate the correct threshold, namely µ = 0. This is because
that limit is related to the growth of hlog(X(t))i, which
weighs large-X contributions less strongly.
III. LINEAR THEORY

Here we study the late-time limit of solution (16) corre- B. The general α-stable case: moments
sponding to eq. (1) with γ = 0, starting from a localized
initial condition at x > 0, to determine the stability of In the general α-stable case the solution is the log-
the origin x = 0. This will be helpful later for interpret- stable distribution given in (16). When β > −1, mo-
ing the nonlinear (γ > 0) results. ments hX n i diverge for any n > 0, as described above.
 5

Thus no stability criterion can be derived based on the

moments.
For the special case β = −1, the moments hX n i exist
and can be calculated. While the moments have been
given in the literature before in the Itô interpretation,
see [99], we give a novel (to our knowledge) derivation in
the Stratonovich interpretation. For β = −1 the FFPE
reads
Dyα,β=−1 f (y) = − sec(απ/2)D−
α
f (y), (23)
α
with D− given by (10). Following the steps made in the
Gaussian case, we multiply (8) by eny and integrate over
y. Fractional integration by parts obeys
Z ∞ Z ∞
α α
f (y)(D+ g)(y)dy = (D− f )(y)g(y)dy, (24)
−∞ −∞

for sufficiently well-behaved functions f and g such that

the fractional derivatives and integrals exist, [98]. Here,
this requires β = −1. Furthermore, note that
α ny
D+ (e ) = nα eny , (25)
which for 0 < α < 1 and 1 < α < 2 follows directly from
α
the definition of D+ in eqs. (10), (12) upon changing
integration variables to u = y − z. One obtains
∂t hX n i = [nµ − sec(πα/2)nα ] hX n i, (26)
such that
S
h(X(t))n i = X0n eλn t , (27)
with
λSn = [nµ − sec(απ/2)nα ] . (28)
FIG. 2. CDF of log-stable law (16) for β = 0, α = 1.5 (top
Hence, the value of µ where the growth rate vanishes panel) and α = 0.5 (bottom panel) with µ = 1 and time t
depends on n, increasing in the order red, orange, green, blue, cyan, grey.
Clearly, the CDF shifts to the right as time increases in the
µc (n) = sec(απ/2)nα−1 . (29) top panel indicating that probability is leaking to +∞, but
takes the constant value 0.5 in the bottom panel, indicating
For α = 2, this reduces to the Gaussian result. For that the probability leaking to both +∞ and −∞.
completeness, we note that the growth rate in the Itô
interpretation given in [99] is similar (see their eq. (8)),

λIn = λSn + n sec(απ/2). (30) C. The general α-stable case: the CDF

We have verified equation (27) for both α > 1 and

α < 1 by computing the moments of the exact solution Consider the log-stable CDF given in eq. (17). Figure
(16) numerically (not shown). However, the moments 2 shows the time evolution of the CDF for β = 0 and
which we just computed for β = −1 are ill-suited for α = 1.5 (top panel), α = 0.5 (bottom panel), both for
studying the linear stability problem. This is because, µ = 1. One observes that for α > 1, probability shifts to
as in the Gaussian case, the moments are dominated by the right due to the drift, indicating leakage to positive
rare large-amplitude events. However, taking the limit infinity. Conversely, for α < 1, the CDF approaches
n → 0 in µc (n) following [100], where large amplitudes a constant value, strictly larger than zero and strictly
are weighted less strongly, one predicts the threshold to smaller than one, indicating that probability is leaking
be at µ = 0 for α > 1 and at µ = ∞ for α < 1. In the to both positive and negative infinity.
following section, we consider the CDF of the log-stable
process to deduce the asymptotic stability of the origin More precisely, for 1 < α < 2 and β < 1, µ > 0 one
and show in particular that the n → 0 predictions are may use eqns. (5) and (17) to show that at late times
correct. (t → ∞), for any given χ, the probability for x < χ, is
 6

given by β
−1 (−1, 1) 1
α
(1 − β)t (1, 2] sign(µ)∞ sign(µ)∞ sign(µ)∞
P (x < χ) ∝ ∝ (1 − β)t1−α . (31) 1 +∞ & − ∞ +∞ & − ∞ +∞ & − ∞
|µt − log(χ)|α
(0, 1) −∞ +∞ & − ∞ +∞
Thus P (x < χ) decreases as time progresses, in agree-
TABLE I. Summary of the late-time behavior of the (linear)
ment with our conclusion based the top panel of figure
log-stable process (16). For a given combination of α and β,
2. A similar argument for µ < 0 and the same range of it is indicated where the weight of the probability will leak to
α shows that in this case P (x > χ) decreases in time. in terms of the variable Y = log(X).
For β = 1 and the same range of α, taking the same
limit, for µ > 0, t → ∞ and χ fixed, one finds using (6)
and (17) that In summary, translating the results back to the original
1−α
α variable x, we have shown that in the log-stable process,
P (x < χ) ∝ t 2α e−c1 µ α−1 t
, (32) for 1 < α < 2, for any β ∈ [−1, 1], all the probability
leaks to x → +∞ for µ > 0, while for µ < 0 all the
which also decays, this time exponentially fast, as t probability accumulates at the origin x = 0. On the
increases. Similarly for µ < 0 and the same range of other hand, for 0 < α < 1, the probability leaks both
α, one can show that P (x > χ) decreases in time. In x → 0 and to x → ∞ independently of µ, except for
short, we find that for any α in the range 1 < α < 2, the one-sided noise at β = ±1. There, all the probability
probability leaks to log(x) → sign(µ)∞ as t → ∞ for leaks to x = 0 for β = −1 and to x → ∞ for β = 1. At
the linear (γ = 0) problem. α = 1, the probability leaks to both x = 0 and x = ∞,
independently of β and µ. Table I summarises the late-
If 0 < α < 1, then for any µ and fixed χ as t → ∞, the time behavior of the linear solution.
argument of the CDF in (17), (log(χ) − µt)t−1/α → 0, Finally, we point out that in the only case where the
such that moments exist, at β = −1, they do not straightforwardly
indicate asymptotic stability. For 1 < α < 2, and µ < 0,
P (x < χ) → Pα,β (0), (33) the origin is stable. Yet, moments of sufficiently high
order will grow. For 0 < α ≤ 1, the origin is stable
where the right-hand side is the α-stable CDF evaluated independently of µ, but there also, high-order moments
at zero, which is a µ-independent constant. For β = 0, grow. However, taking the moment order n → 0 predicts
the constant is 0.5 by symmetry, as illustrated in figure the correct thresholds µ = 0 for α > 1 and µ = ∞ (no
2, but in general, it will depend on β in a continuous instability at any finite µ) for α < 1.
way. In particular, for β = 1, Pα,β=1 (0) = 0, since the
stable PDF is only supported at positive values in this
case. On the other hand, for β = −1, the constant is
IV. NONLINEAR THEORY
Pα,β=−1 (0) = 1, since the PDF is only supported at neg-
ative values. In short, we find that for any α in the range
0 < α < 1 the probability leaks to both log(x) → −∞ In this section, we study the effect of the nonlinear
and log(x) → +∞, with the exceptions of β = ±1, where term in equation (1) with γ > 0 in the development of
probability leaks to log(x) → β∞. the instability. The nonlinearity will prevent the leakage
In the marginal case α = 1, the fact that (log(χ) − of probability to x → ∞ that was observed for many
µt)/t → −µ for any fixed χ implies cases in the linear regime, thus leading to a stationary
distribution that we try to estimate here.
P (x < χ) → Pα=1,β (−µ), (34) For illustration, typical solutions of the nonlinear
Langevin equation (1) are shown in figure 3. The re-
where the right-hand side is the α-stable CDF evaluated alizations are generated efficiently by integrating eq. (1)
at −µ, which is a positive constant for any finite µ and using its exact solution given in appendix A. Three dif-
any β ∈ [−1, 1]. Hence, at α = 1, probability leaks to ferent cases are shown: in panel a) α = 1.5, β = 0, in
both log(x) → −∞ and log(x) → ∞ for all µ. Only the panel b) α = 0.5, β = 1, in panel c) α = 0.5, β = −1.
fraction of the weight escaping in each direction depends For each case, two typical time series are shown, one for
on µ. positive µ and one for negative µ at fixed γ = 1. In panel
We note that all of the results obtained above from the a), at negative µ, X decays to zero. At positive µ, there
exact linear (γ = 0) solution can be understood in terms is on-off intermittency: X fluctuates over many orders
of a competition between the drift µt and the widening of magnitude, but does not decay. There is a qualitative
of the PDF, which goes as t1/α . For α > 1, the drift change of behavior between µ > 0 and µ < 0. Typical
is dominant over the widening and probability leaks to trajectories at β 6= 0, 1 < α < 2 resemble those in panel
log(x) → sign(µ)∞. On the other hand, for 0 < α ≤ 1, a). In panel b), the origin is unstable for both positive
the drift no longer dominates and probability spreads out and and negative µ. In panel c) the origin is stable for
to both log(x) → ±∞, except for one-sided noise. both positive and negative µ.
 7

A. Exact relation for the second moment

One important property of equation (1) is that if hf (t)i

exists (i.e. for 1 < α ≤ 2), then for X > 0 it implies that

d
hlog(X)i = µ − γhX 2 i + hf (t)i. (35)
dt
Assuming hf (t)i = 0, then for γ > 0, µ < 0 the right-
hand side is negative, resulting in

hlog(X)i ≤ µt, (36)

which tends to −∞ as t → ∞. However, if µ > 0, then

a stationary state is reached for which dhlog(X)i/dt = 0
and the second moment satisfies

hX 2 i = µ/γ. (37)

By contrast with the linear regime, for which it was

shown above that moments are not a reliable indicator of
stability, moments in the nonlinear regime are pertinent
to the stability of the origin. This is due to the fact
that the nonlinearity in equation (1) impedes excursions
to large amplitudes, which are the reason why high
moments may grow exponentially in the linear case
(γ = 0), even when the origin is stable. Equation (37)
thus already indicates that for 1 < α ≤ 2 the system is
unstable when µ > 0 in agreement with the predictions
of section III C. Note, however that stability or instabil-
ity cannot be concluded from (37) for 0 < α ≤ 1, since
eq. (37) is not valid there.

B. Asymptotics of the PDF at large x

In this section, we study the fractional Fokker-Planck

equation under the assumption of stationarity to derive
the asymptotics of the stationary density for x → ∞.
Here, we need to consider β > −1 and β = −1 separately.

1. The case β > −1

Let β > −1, and consider the FFPE in the

Stratonovich interpretation, i.e. equation (8). For y →
∞, we neglect Dα−  Dα+ and µ  γ exp(2y) to find the
following equation for the stationary distribution associ-
FIG. 3. Semi-logarithmic plots of time series X(t) (see eq.
(1)). Panel a): α = 1.5, β = 0, γ = 1. Orange (top): µ = 0.2 ated with the process Y = log(X), denoted by py,st (y),
– X(t) varies over 20 orders of magnitude, displaying on-off
intermittency. Blue (bottom): µ = −0.2 – X decays to zero. d
Ry py,st (z)
A critical transition occurs between the two, at µ = 0. Panel 2y
(1 + β) dy −∞ (y−z)α−1
dz
b): α = 0.5, β = 1, γ = 1. Orange (top): µ = 1. Blue γe py,st (y) ≈ . (38)
2 cos(πα/2)Γ(2 − α)
(bottom): µ = −1. The origin is unstable for all µ. Panel c):
same as panel b) but with β = −1. Here, the origin is stable Asymptotically, the integral is dominated by z  y.
for all µ. Hence (y − z)α−1 ≈ α−1
R yy . The remaining
R ∞ integral can be
approximated as −∞ py,st (z)dz ≈ −∞ py,st (z)dz = 1.
 8

The resulting equation implies the following asymptotic 2. The case β = −1

behavior for the stationary density at large y,
The asymptotics in (41) and (43) break down for β =
C(1 + β) −α −1, which is the nonlinear version of the finite-moment
py,st (y) ∼ y exp(−2y). (39)
γ log-stable process. For Gaussian noise, α = 2, the sta-
γ 2y
tionary PDF in y is known to be py,st (y) = N eµy e− 2 e ,
The prefactor is given by ny
which decays faster than e at large y for any n > 0.
For 1 < α < 2 and β = −1, the stable PDF has a
C = sin(πα/2)Γ(α)/π, (40) short tail +∞, decaying faster than a Gaussian (since
α/(α − 1) > 2 in eqn. (6)). This implies that the station-
which has been simplified using Euler’s reflection formula ary PDF under such Lévy noise will decay faster than
Γ(α)Γ(1 − α) = π/ sin(πα). In terms of the stationary in the Gaussian case. Hence, the moments of x for any
distribution px,st (x) associated with the original process order n > 0 exist there also. In terms of py,st (y), one
X, this gives expects a double-exponential function as in the Gaussian
case. However, unfortuantely, we cannot derive these
C(1 + β) large-x asymptotics explicitly as we did for β > −1, since
px,st (x) ∼ log(x)−α x−3 , (41)
γ the Riemann-Liouville derivative of such functions is not
known in simple terms. Rather, we will rely on numeri-
for x → ∞. While the above derivation is valid for 1 < cal solutions to confirm that the PDF of y indeed decays
α < 2, one may repeat the same steps for 0 < α < 1 faster than exponentially at y → +∞. At 0 < α < 1,
with the corresponding fractional derivative from eqns. β = −1, since the origin is stable for all µ in the lin-
(12), and finds the same result. For α > 1, there is ear regime, it will continue to be stable in the nonlinear
both a finite mean and a finite variance. For α < 1, regime (the nonlinearity in equation (1) is strictly nega-
the variance in infinite, but the mean is finite. We note tive). Thus the stationary PDF is δ(x) in this case, i.e.
that the derivation given here is inspired by a similar the CDF converges to 1 for all x > 0 in the long-time
argument from [64]. limit.
Further, if one chooses the Itô interpretation, then one
may derive the large-x asymptotics in a similar way. One
begins by considering the stationary solutions of the as- C. Asymptotics of the PDF at x → 0
sociated Itô FFPE for px (x, t), i.e. equation (15). Then
α α
one takes the limit x → ∞, assuming D+  D− , and We now investigate the asymptotic behavior of the sta-
3
using µx  γx to find tionary density for x → 0. Here we need to distinguish
between the cases β = 1 and β < 1.
d
Rx z α px,st (z)dz
(1 + β) dx −∞ (x−z)α−1
γx3 px,st (x) ≈ . (42)
2 cos(πα/2)Γ(2 − α) 1. The case β = 1
R x zα px,st (z)dz 1
R∞ α
Now, −∞ (x−z) α−1 ≈ xα−1 −∞
z px,st (z)dz by a sim- Consider β = 1 and µ > 0. The FFPE (8) in steady
ilar reasoning as for the Stratonovich case. The remain- state, taking y → −∞ and neglecting γe2y  µ, reads
ing integral cannot be performed explicitly, but it is an
x-independent constant. Hence, one finds the asymptotic 1
0 = −µ∂y py,st (y) − (Dα f )(y). (44)
proportionality cos(απ/2) +

pItô −3−α Making an exponential ansatz py,st (y) ∝ eAy and using
x,st (x) ∝ (1 + β)x , (43)
the fractional derivative of the exponential given in equa-
tion (25) leads to
for large x. This result is remarkable, since the power
law matches exactly the one found for additive noise in a 1
quartic potential, [64]. In particular, the third moment 0 = −µAeAy − µAα eAy , (45)
cos(απ/2)
is finite in the Itô interpretation for 1 < α < 2 (where it
diverges in the Stratonovich case), and the variance is fi- implying
nite for 0 < α < 1 (where it diverges in the Stratonovich 1/(α−1)
case). The observation that the asymptotic form of A ≡ Aα (µ) = (−µ cos(απ/2)) (46)
the tails of the stationary PDF are altered by a state-
In terms of the original variable x, this corresponds to
dependent Lévy noise amplitude in the Stratonovich in-
terpretation, but not in the Itô interpretation, has been px,st (x) ∝ x−1+Aα (µ) , (47)
made in previous studies with different functional forms
of multiplicative noise [69, 70]. For the remainder of this which for α > 1 is integrable when µ > 0. The term
paper, we will adopt the Stratonovich interpretation. integrable is used here to mean that the integral of a given
 9

function over its domain converges. When α > 1 and and µ > 0, on the other hand, (49) gives a consistent,
µ ≤ 0, on the other hand, the solution is non-integrable, integrable stationary PDF.
which is associated with the absence of a steady-state
solution in the space x > 0. In that case, the stationary
density is given by δ(x), with the CDF converging to 1 β px,st (x → 0) px,st (x → ∞)
for all x in the long-time limit. For α < 1 and −∞ < −1 C(µx)−1 log−α (1/x) exponential decay
µ < 0, the same result (47) holds with A > 0, such (−1, 1) C(µx)−1 log−α (1/x) Cγ −1 x−3 log−α (x)
that the solution is integrable. For α < 1, µ > 0, the 1 ∝ x−1+Aα (µ) Cγ −1 x−3 log−α (x)
solution again ceases to be integrable. From the Langevin
equation (7) one deduces that this is due to the fact that TABLE II. Summary of the different asymptotic behaviors
of px,st (x) obtained in the previous sections for α < 2. The
in this case Ẏ > 0 for Y < 21 ln (µ/γ), since the noise is domain of validity of the formulas is discussed in the text.
strictly positive, and thus the probability to be at Y ≤ The constants C and Aα (µ) are given in eqns. (40), (46).
1
2 ln(µ/γ) vanishes at late times. The stationary PDF is
thus only supported at values of y above the deterministic
saturation point 12 ln(µ/γ) and vanishes for all smaller Table II summarises the different asymptotic behav-
y. Hence, the exponential ansatz is inappropriate and iors obtained above. The closest resemblance with the
breaks down. This indicates that for α < 1, µ > 0, we α = 2 case is seen in the large-x exponential decay at
cannot neglect the nonlinear term in equation (1), since β = −1 (which is Gaussian for α = 2), and the small-
it is the only one that decreases Y . x µ-dependent power-law at β = 1 (where A2 (µ) = µ).
The above discussion confirms the intuition based on The asymptotics in the remaining cases are qualitatively
the linear solution (including an arrest of the leakage of different from the Gaussian case.
probability to +∞ by the nonlinear term in eq. (1)). For Summarising, a transition occurs at µ = 0 for 1 < α <
β = 1, α > 1 a critical transition occurs at µ = 0, from 2, from a stable origin at µ < 0 to an unstable origin at
all weight of the stationary PDF being at x = 0 (origin µ > 0. This is as predicted in the linear theory. When
stable) to non-zero weight at x > 0 (origin unstable). On α < 1, β < 1, the origin is stable for all µ, and for α < 1,
the other hand, for 0 < α < 1, β = 1, the origin is always β = 1, the origin is always unstable. These results are
unstable. also consistent with the linear theory, taking into account
saturation by the nonlinearity.

2. The case β < 1

D. PDFs and Moments
Let β < 1 and µ > 0. We follow once again the argu-
ments of [64], starting from eq. (8). Consider y → −∞, Here we attempt to deduce the moments based on the
such that µ  γe2y , and neglect D+ α
 D−α
to find asymptotic behavior of the PDFs discussed in the previ-
R∞ py,st (z)dz
ous sections complemented by numerical solutions of the
d
(1 − β) dy y (z−y)α−1 stationary FFPE, using a heuristic approach. The nu-
− µpy,st (y) = . (48) merical solutions are computed using the finite-difference
2 cos(πα/2)Γ(2 − α)
formulation described in appendix B.
Using normalisation of the PDF, as for the large-x limit
at β > −1, we find that the stationary PDF is asymp-
totically given by 1. The case 1 < α < 2, β = 1
(1 − β)C
py,st (y) ∼ (−y)−α , (49) We begin by showing the results from the numerical
µ
solution of the FFPE. Figure 4 indicates an agreement
for y → −∞, where C is given by (40). In terms of the with the theoretical results of the previous section, for
original variable, this corresponds to x  1 (i.e. y → −∞) and x  1 (i.e. y → ∞). In order
to calculate the scaling with µ of the different moments
C(1 − β) −1 we can model the PDF as
px,st (x) ∼ x (log(1/x))−α (50)
µ (
1 x−1+Aα (µ) : x < x∗
for x → 0. While the above derivation is for 1 < α < 2, px,st (x) ≈ , (51)
the case 0 < α < 1 leads to the same result. Clearly, this N Bx−3 log−α (x) : x ≥ x∗
solution breaks down at negative µ, since the predicted
PDF ceases to be positive. In this case, the stationary where x∗ and N are model parameters, Aα (µ) is as given
2+A (µ)
distribution is δ(x). For α < 1, the fact that the solution in equation (46) and B = x∗ α logα (x∗ ) for conti-
(49) is not integrable at y = −∞ implies that there is nuity. To determine the two unknowns N and x∗ , we
no stationary state at x > 0. Instead, the stationary impose normalisation of the PDF and the second mo-
density in that case is δ(x), for all µ. For 1 < α < 2 ment identity (37). At small µ, the dominant part of the
 10

implies,
p
−4 x∗ ≈ 2µ/(Aα (µ)γ), (54)
10
such that
10−8
Py,st(y)

α
hXi ≈ Aα (µ)x∗ ∝ µ 2(α−1) , (55)
−12
10
where Aα (µ) was inserted from equation (46). Figure
5 shows that this agrees with the numerical solution of
10−16 the stationary FFPE for the examplary case α = 1.5,
β = 1. We note that (55) can simply be extended to
hX n i ∝ µcn with 0 < n ≤ 2, with cn = 2−n(2−α)
2(α−1) which
−75 −50 −25 0 1
y varies continuously with n. In particular cn ≈ α−1 for
α
small n, c1 = 2(α−1) and, by construction, c2 = 1.

FIG. 4. Semi-log plot of numerically obtained stationary PDF

for α = 1.5, β = 1.0, varying µ = 0.1, 0.33, 0.55, 0.78, 1.0, at
γ = 1 fixed. The dashed line on the right is the theoreti-
cal prediction (41) for the the cut-off by non-linearity. The
dashed lines on the left show shows the prediction py,st (y) ∝
exp(A(µ)y) from (47).

FIG. 6. Log-log plot of the numerically obtained stationary

PDF py,st (y) versus |y| for α = 1.5, β = −1. Five different
values of µ are shown (0.01, 0.025, 0.085, 0.3, 1.0) for γ = 1.
The tail at y → −∞ fits the prediction (49) (dashed lines
on the right). Before that limiting scaling is observed at
y → −∞, an intermediate, flatter power-law range occurs
at y < 0, whose exponent is independent of µ, but whose am-
FIG. 5. First and second moment of X for α = 1.5, β = 1 plitude decreases and whose range increases as µ decreases.
versus µ at γ = 1. The first moment scales as predicted in At large positive y, there is a faster-than-exponential decay
(55), shown by the curved dashd line, and the second moment as predicted (compare with figure 8 at y > 0).
is linear, in agreement with (37).

weight is at negative y, i.e. at small x, as visible in figure 2. The case 1 < α < 2, β = −1
4. This implies
A (µ)
x∗ α Figure 6 shows the numerically obtained PDF. It
N∼ , (52) matches the theoretically predicted asymptotics in the
Aα (µ)
y → ±∞ limits. In addition, an intermediate, shallower
as µ → 0+ . In addition, range is observed at intermediate negative y before the
  predicted asymptotic behavior at y → −∞ is realized.
1 1 1−α 2+A (µ)
2+Aα (µ) + α−1 log (x ∗ ) x∗ α Figure 6 suggests that this intermediate range is a power
hX 2 i = (53) law. A close inspection shows that it is only approx-
N imately a power law since it has a finite curvature in
and moments of order higher than two diverge. By equa- the log-log diagram. Notwithstanding this caveat, we
tion (37) we have hX 2 i = µ/γ. For 0 < µ  1, this propose a simplistic model approximately describing the
 11

y are exponentially suppressed:

Z 0
0.4 hX n i = heny i ≈ D (−y)−λ eny dy
−∞
∝µ (61)
hX ni

0.2 Note that the critical exponent in the final result is in-
dependent of the value of λ. The result of eq. (61) is
confirmed in figure 7, where the integer moments up to
order four, determined from the numerical solution of the
stationary FFPE, are all shown to scale linearly with µ.
0.0
0.0 0.1 0.2
µ

FIG. 7. Moments for α = 1.5, β = −1 versus µ at γ = 1

from the numerically computed steady-state PDF. Symbols
represent n = 1 (diamond), n = 2 (hexagon), n = 3 (circles),
n = 4 (triangles). The dashed lines show linear scaling.

numerical result

0
 :y≥0
py,st (y) ≈ D(−y)−λ : 0 > y ≥ y∗ , (56)
2C/µ (−y)−α

: y∗ > y

where C = Γ(π) sin(πα/2)/π. The portion of the PDF

at y > 0 makes a negligible contribution to its nor-
malisation and any moments of X, due to the faster- FIG. 8. Log-log plot of the numerically obtained stationary
PDF py,st (y) versus |y| for α = 1.5, β = 0. Five different
than-exponential decay at y > 0. The value of λ can
values of µ are shown (0.01, 0.025, 0.085, 0.3, 1.0) for γ = 1.
only be determined numerically, with relatively large er- The tail at y → −∞ fits the prediction of (49) shown in
rorbars. Thus fixing λ numerically (e.g. λ ≈ 0.6 for dashed straight lines on the right. At intermediate y < 0, a
α = 1.5, β = −1 in figure 6), there are two unknowns D flatter power-law range is seen, with exponent independent
and y∗ which we determine by imposing normalisation of ν of µ (ν ≈ 0.25 here), but with amplitude decreasing and
the PDF and the second moment identity (37). Formally, width increasing as µ decreases. The tail at y  1 matches
the prediction (41), shown by the curved dashed line.
D 2C
1= (−y∗ )1−λ + (−y∗ )1−α , (57)
1−λ µ(α − 1)
Z 0
2C y∗
Z
µ
=D (−y)−λ e2y dy + (−y)−α e2y dy. (58)
γ y∗ µ −∞
3. The case 1 < α < 2, |β| < 1
Figure 6 suggests that y∗ → −∞ as µ → 0+ . In eq. (58),
this implies that the second integral, from −∞ to y∗ , is Figure 8 shows that the PDF, which matches the pre-
exponentially suppressed for small µ. For large |y∗ |, the dicted asymptotics at y → ±∞, strongly resembles the
lower limit of the first integral may be replaced by −∞. case of β = −1 in that in addition to the asymptotic
This leads to power-law range, an intermediate, approximately power-
21−λ law range is seen at negative y. Again, close inspection
D≈ µ, (59) shows that the intermediate range shows small deviations
γΓ(1 − λ)
 1/(α−1) from a power law. However, the most marked difference
2C from the case β = −1 is that the decay for |β| < 1 is
y∗ ≈ − µ−1/(α−1) . (60)
(α − 1) only exponential in y at positive y, not faster than ex-
ponential as for β = −1. In particular, the asymptotics
Note that, as expected, y∗ → −∞ as µ → 0+ . The at positive y imply a slow, power-law convergence of the
two results (59), (60) imply that the moments of X of second moment since pst (y)e2y ∝ y −α at y  1. Bearing
arbitrary order n > 0 scale linearly, since large negative this in mind, we nonetheless employ the same approxi-
 12

Note that, by contrast with the case β = −1, the critical

exponent in (67) depends on the exponent ν of the inter-
mediate power-law range at negative y, which we have
not determined theoretically as a function of α, β, but
only measured numerically. The critical scaling of the
first moment predicted in eq. (67) is shown to be consis-
tent with the numerically obtained moments in figure 9
for the case α = 1.5, β = 0. The prediction (67) for the
critical exponent at n = 1 was also verified for different
values of α at β = 0 (not shown).

4. The case 0 < α < 1, β = 1

In this case the origin is unstable for all µ, and a non-

trivial stationary state exists due to the nonlinearity in
equation (1). The point-vortex model presented in the
FIG. 9. First and second moment of X for α = 1.5, β = companion paper [59], where α = 2/3, β = 1, falls into
0 versus µ at γ = 1. The scaling of the first moment is
this parameter range. The asymptotic theoretical results
compatible with the prediction of (67), shown by the curved
dashed line, with ν ≈ 0.25 from figure 8. The second moment
suggest that for µ < 0, the PDF can be modeled as
is linear in µ, satisfying the identity (37). (
BeAα (µ)y : y < y∗
py,st (y) = 2C −2y −α , (68)
γ e y : y ≥ y∗
mate form for the PDF as for β = −1,
 where C = sin(απ/2)Γ(α)/π, Aα (µ) given by (46), and
0
 : y≥0 the two unknowns B and y∗ are in principle determined
py,st (y) ≈ E(−y) −ν
: 0 > y ≥ y∗∗ , (62) by continuity at y∗ and normalisation. We note that
(1 − β)C/µ|y|−α : y < y the second moment does not exist because y −α is not

∗∗
integrable at infinity for α < 1. Moments of order higher
where once more C = Γ(α) sin(απ/2)/π. We may again than two also diverge. However, hX n i does exist for all
determine ν numerically, albeit with significant uncer- 0 < n < 2.
tainty. In figure 8, where α = 1.5, β = 0, we observe For illustration, we consider the special case α = 1/2,
ν ≈ 0.25. As for β = −1, the portion of the PDF at γ = 1 and take the limit µ → 0− , where Aα (µ) → ∞.
y > 0 makes a subdominant contribution to the normal- Clearly then py,st (y) → 0 at y < 0. Further, since
isation and the moments of order n < 2. To determine Z ∞
E, ν, we need two conditions. By contrast with the case
2C e−2y y −1/2 dy = 1, (69)
β = −1, we impose continuity at y∗∗ instead of (37), 0
normalisation of the PDF. Formally,
taking y∗ ≈ 0 as µ → 0− gives a consistently normalized
(1 − β) C model of the PDF. For this special case, α = 1/2, γ = 1,
E(−y∗∗ )−ν = (−y∗∗ )−α , (63)
µ the n-th moment of the PDF for 0 < n < 2 may be
E (1 − β)C computed to be
1= (−y∗∗ )1−ν + (−y∗∗ )1−α .
1−ν µ(α − 1) Z ∞ r
(64) n (n−2)y −α 2
hX i = 2C e y dy = , (70)
0 2−n
Solving these two equations gives
  −1/(α−1) for µ small and negative. Note that the result is inde-
1 1 pendent of µ. For n = 1, equation (70) was found to
y∗∗ = + (1 − β)Cµ (65)
1−ν α−1 be satisfied to within a few percent relative error by av-
2α−ν−1 eraging over sample trajectories (not shown) using the
ν−α (1 − β)C {(1 − β)C} α−1 1−ν
E =y∗∗ = h 1 µ α−1 . (66) method in appendix A.
µ 1 1
i α−1
1−ν + α−1

This implies that the first moment exhibits the anoma- V. CONCLUSIONS
lous scaling
Z 0 We have studied the stochastic process obeying the
1−ν Langevin equation (1) with Lévy white noise. The theory
hXi = hey i ≈ E ey (−y)−ν dy ∝ µ α−1 . (67)
−∞ of on-off intermittency was generalized, from the known
 13

1−ν
hX n i ∝ µcn , with c1 ≈ α−1 and c2 = 1.

(iii) “Critical 3” with 1 < α < 2, β = −1. At

small x and µ > 0, the PDF is px,st (x) ∝
µ−1 x−1 (log(1/x))−α as for case (ii). For large x,
the PDF px,st (x) decays faster than any power of
x. At x < 1, but not too small, there is an inter-
mediate range similar to that in (ii) where approx-
imately px,st (x) ∝ x−1 (log(x))−λ , with a different
exponent λ which was determined numerically. It
remains an open problem to compute λ theoreti-
cally as a function of α, β. However, the critical
exponents are independent of λ: for small µ > 0,
FIG. 10. The parameter space of (1) with white Lévy noise, one finds hX n i ∝ µcn , with cn = 1 for all n > 0.
α ∈ (0, 2], β ∈ [−1, 1]. A critical transition occurs at µ = 0
for 1 < α < 2, and for the Gaussian case α = 2. For 0 <
α ≤ 1, the origin is either always stable or always unstable, (iv) “Unstable” with 0 < α < 1, β = 1. Since the noise
independently of µ. is strictly positive and has infinite mean, the origin
x = 0 is always unstable, independently of µ. At
small x and for all µ < 0, the PDF is px,st (x) ∝
case of Gaussian noise, to Lévy noise by studying the x−1+Aα (µ) . For µ > 0, the PDF vanishes at x <
FFPE (8) analytically and numerically. First the linear
p
µ/γ. For µ < 0 small, in the special case α = 1/2,
(γ = 0) solution was analysed, which showed leakage of γ = 1, the n-th moment is shown to be hX n i =
the probability to x = 0, x = ∞ or both, depending on p
2/(2 − n) for 0 < n < 2. For n ≥ 2, all moments
the noise parameters α and β. Then we computed the hX n i diverge.
nonlinear (γ > 0) stationary solutions of the stationary
FFPE, for which the leakage of probability to large x is
arrested by the nonlinearity in equation (1). We showed (v) “Stable” with 0 < α < 1, β < 1 or α = 1 for
that for 1 < α ≤ 2 the origin is stable at µ < 0 and any β. The origin is always stable in this case, the
unstable at µ > 0. For 0 < α < 1, the origin is always stationary PDF is δ(x) for all µ as long as γ > 0.
stable, or always unstable, for any µ, due to the diver-
gent mean of the noise. In addition to the Gaussian case
α = 2, where the stationary PDF for µ > 0 is given by In summary, we have shown that instabilities under
γ 2
px,st (x) = N x−1+µ e− 2 x and all critical exponents are the influence of multiplicative heavy-tailed noise, mod-
equal to 1, we identify a total of five qualitatively distinct eled as Lévy white noise, can display anomalous critical
regimes in the parameter space α ∈ (0, 2], β ∈ [−1, 1], il- exponents differing from those for Gaussian noise, where
lustrated in figure 10, cn = 1 for all n. Anomalous critical exponents different
from the Gaussian results have been found previously,
(i) “Critical 1” with 1 < α < 2, β = 1. For µ > for instance for instabilities subject to colored noise [92].
0 and small x, the PDF is px,st (x) ∝ x−1+Aα (µ) , Here, we add the scenario of Lévy white noise, which
1
with Aα (µ) ∝ µ α−1 . This matches the Gaussian leads to several new possibilities of anomalous scaling, as
small-x result for α = 2. At x  1, the PDF is discussed above.
px,st (x) ∝ x−3 (log(x))−α , i.e. hX n i < ∞ for n ≤ 2 Our work serves as a first step in the study of instabil-
but hX n i = ∞ for n > 2. As µ → 0+ , one has ities in the presence of multiplicative Lévy noise. There
α
hX n i ∝ µcn , with c1 = 2(α−1) and c2 = 1. are many directions that can be further pursued. First
of all the values of the power-law exponents λ, ν in eqs.
(ii) “Critical 2” with 1 < α < 2, |β| < 1. The (62), (56) remain unknown leading to only a non-rigorous
PDF is px,st (x) ∝ µ−1 x−1 (log(1/x))−α at small estimate of the scaling exponents of the different mo-
x > 0. This is in stark contrast with the Gaus- ments with µ. Furthermore, the behavior of the system
sian result; the logarithmic term here is crucial for under truncated Lévy noise [42, 63, 64, 101, 102], com-
integrability at x = 0. At x  1, the PDF is bined Lévy-Gaussian noise [103], a finite-velocity Lévy
px,st (x) = Cx−3 (log(x))−α as in case (i), s.t. only walk [104], different nonlinearities [105], higher dimen-
moments of order n ≤ 2 are finite. At x < 1, but sions [93, 106, 107] and its time statistics [4–10, 23, 108]
not too small, there is an intermediate range where would also be interesting to understand. Finally, since
approximately px,st (x) ∝ x−1 (log(x))−ν , where the Lévy statistics are found in many physical systems, we
exponent ν was determined numerically. It remains permit ourselves speculate that the anomalous critical
an open problem to compute it theoretically as exponents predicted here for instabilities in the presence
a function of α, β. For small µ > 0, we found of power-law noise may be observable experimentally.
 14

VI. ACKNOWLEDGEMENTS where the fractional derivative is given by

α α
[(1 + β)D+ pst + (1 − β)D− pst ]
The authors acknowledge helpful discussions with Dyα,β pst (y) = − (B2)
François Pétrélis and Stephan Fauve, and thank Olivier 2 cos(πα/2)
Bénichou for bringing the example of blinking quantum [Dα + D− α
+ β(D+ α
− D−α
)]pst
=− + . (B3)
dots to their attention. This work was granted access 2 cos(πα/2)
to the HPC resources of MesoPSL financed by the Re-
gion Ile de France and the project Equip@Meso (refer- We consider 1 < α < 2, for which Riemann-Liouville
ence ANR-10-EQPX-29-01) of the programme Investisse- derivatives are given by (10), and (11). Integrating once
ments d’Avenir supervised by the Agence Nationale in y gives
pour la Recherche and the HPC resources of GENCI-
0 = − (µ − γe2y )pst (y)
TGCC & GENCI-CINES (Project No. A0080511423, Z ∞
A0090506421). This work has also been supported by d pst (z)[1 + βsign(y − z)]
+ Kα dz, (B4)
the Agence nationale de la recherche (ANR DYSTURB dy −∞ |y − z|α−1
project No. ANR-17-CE30-0004). AvK acknowledges
support by Studienstiftung des deutschen Volkes. where Kα = −(2 cos(πα/2)Γ(2 − α))−1 . To simplify
discretization further, it is advantageous to rewrite the
term stemming from the fractional derivative in the the
Appendix A: Solution of the Langevin equation Grünwald-Letnikov form, cf. [109], thereby transferring
the y-derivative into the integral. This gives
Equation (1) is of the form of a Bernoulli differential Z ∞ 0
pst (z)[1 + βsign(y − z)]
equation. Hence, it admits an exact solution, which can 0 = −(µ−γe2y )pst (y)+Kα dz.
be derived by dividing (1) by X 3 and letting Z(t) = −∞ |y − z|α−1
1/X 2 (t), such that (B5)
For discretization we consider a large domain
dZ(t) [ymin , ymax ], meshed by intervals [yn−1 , yn ], whose
+ 2r(t)Z(t) = 2γ. (A1)
dt N + 1 endpoints are yn , where pst (yn ) = pn , with
This gives n = 0, . . . , N . We prescribe an arbitrary initial condition
sign(x0 ) p−1 > 0. Then, using a backward difference scheme
X(t) =   12 , for f 0 (z) and regularising |y − z|α−1 → |y − z|α−1 + 
e−2µt−2L(t)
Rt
+ 2γ 0 e2µ(t0 −t)+2(L(t0 )−L(t)) dt0 (0 <   1), we find a matrix equation
x20
(A2) N
which is non-negative if x0 ≥ 0. We denote dL/dt = f (t)
X
bn = Ln,m pm , (B6)
with f (t) white Levy noise, i.e. L(t) is a free Levy flight. m=0
This solution is also given in [23], where it is stressed that
it holds for any type of noise with L(t) being the integral where n = 1, . . . , N ,
of the noise. By contrast with other nonlinear equations
p−1 ((1 + βsign(yn − y0 )))
involving multiplicative Lévy noise, such as [86], where bn = −Kα (B7)
the analytical solution of the nonlinear Langevin equa- |yn − y0 |α−1 + 
tion gives access to the exact time-dependent PDF, this and
is impossible here since the result depends the integral
of L(t), in addition to L(t). However, the expression is (1 + βsign(y0 − ym )))
Ln,m = − µ + γe2y0 + Kα (B8)
useful for numerical evaluation to generate realisations |yn − ym |α−1 + 
of the random process. It is advantageous over a direct (1 + βsign(yn − ym+1 ))
iterative numerical integration since it does not require +Kα θ(N − 1 − m) , (B9)
|yn − ym+1 |α−1 + 
smaller time steps at large nonlinearity. Nonetheless, for
large values of L(t0 ) − L(t), the exponential in the in- where θ(x) designates the Heaviside function. Finally,
tegrand may produce an overflow error. This can be the steady density pm , m = 1, . . . , N is obtained by in-
avoided by choosing integration step dt and the total in- verting the matrix Ln,m and normalising the result. For
tegration time t not too large. β < 1 we chose a non-uniform grid, composed of a loga-
rithmically spaced grid at y < −O(100), combined with
a uniform grid in the region O(100) > y > −O(100).
Appendix B: Finite-difference numerical solution The total grid size was N = 24000. For β = 1, a uniform
grid was used with N = 16000 (the PDF does not extend
We recall the stationary space-fractional Fokker- to y < 0 as far). Choosing  on the order of the small-
Planck equation in the Stratonovich interpretation in est grid resolution to the appropriate power α − 1 gives
terms of Y = log(X), which reads results consistent with exact theoretical predictions, as
0 = −∂y [(µ − γe2y )pst (y)] + Dyα,β pst (y), (B1) described in the text.
 15

[1] H. Fujisaka and T. Yamada, A new intermittency in cou- [20] R. Raynaud and E. Dormy, Intermittency in spherical
pled dynamical systems, Progress of theoretical physics couette dynamos, Physical Review E 87, 033011 (2013).
74, 918 (1985). [21] S. H. Strogatz, Nonlinear dynamics and chaos with stu-
[2] N. Platt, E. Spiegel, and C. Tresser, On-off intermit- dent solutions manual: With applications to physics, bi-
tency: A mechanism for bursting, Physical Review Let- ology, chemistry, and engineering (CRC press, 2018).
ters 70, 279 (1993). [22] R. Stratonovich, A new representation for stochastic in-
[3] E. Ott and J. C. Sommerer, Blowout bifurcations: the tegrals and equations, SIAM Journal on Control 4, 362
occurrence of riddled basins and on-off intermittency, (1966).
Physics Letters A 188, 39 (1994). [23] S. Aumaı̂tre, K. Mallick, and F. Pétrélis, Noise-induced
[4] J. Heagy, N. Platt, and S. Hammel, Characterization of bifurcations, multiscaling and on–off intermittency,
on-off intermittency, Physical Review E 49, 1140 (1994). Journal of Statistical Mechanics: Theory and Experi-
[5] P. W. Hammer, N. Platt, S. M. Hammel, J. F. Heagy, ment 2007, P07016 (2007).
and B. D. Lee, Experimental observation of on-off in- [24] A. Schenzle and H. Brand, Multiplicative stochastic pro-
termittency, Physical review letters 73, 1095 (1994). cesses in statistical physics, Physical Review A 20, 1628
[6] F. Rödelsperger, A. Čenys, and H. Benner, On-off in- (1979).
termittency in spin-wave instabilities, Physical review [25] L. P. Kadanoff, W. Götze, D. Hamblen, R. Hecht,
letters 75, 2594 (1995). E. Lewis, V. V. Palciauskas, M. Rayl, J. Swift, D. Asp-
[7] T. John, R. Stannarius, and U. Behn, On-off inter- nes, and J. Kane, Static phenomena near critical points:
mittency in stochastically driven electrohydrodynamic theory and experiment, Reviews of Modern Physics 39,
convection in nematics, Physical review letters 83, 749 395 (1967).
(1999). [26] N. Goldenfeld, Lectures on phase transitions and the
[8] A. Vella, A. Setaro, B. Piccirillo, and E. Santamato, On- renormalization group (CRC Press, 2018).
off intermittency in chaotic rotation induced in liquid [27] G. Eyink and N. Goldenfeld, Analogies between scal-
crystals by competition between spin and orbital angu- ing in turbulence, field theory, and critical phenomena,
lar momentum of light, Physical Review E 67, 051704 Physical Review E 50, 4679 (1994).
(2003). [28] N. Goldenfeld and H.-Y. Shih, Turbulence as a prob-
[9] D. Feng, C. Yu, J. Xie, and W. Ding, On-off intermit- lem in non-equilibrium statistical mechanics, Journal of
tencies in gas discharge plasma, Physical Review E 58, Statistical Physics 167, 575 (2017).
3678 (1998). [29] M. F. Shlesinger, G. M. Zaslavsky, and U. Frisch, Lévy
[10] G. Huerta-Cuellar, A. N. Pisarchik, and Y. O. Bar- flights and related topics in physics (Springer, 1995).
menkov, Experimental characterization of hopping dy- [30] A. V. Chechkin, R. Metzler, J. Klafter, V. Y. Gonchar,
namics in a multistable fiber laser, Physical Review E et al., Introduction to the theory of lévy flights, Anoma-
78, 035202 (2008). lous Transport , 129 (2008).
[11] S. Benavides, E. Deal, J. Perron, J. Venditti, Q. Zhang, [31] W. Feller, An introduction to probability theory and its
and K. Kamrin, Multiplicative noise and intermittency applications, vol 2 (John Wiley & Sons, 2008).
in bedload sediment transport (2020). [32] B. Gnedenko, A. Kolmogorov, B. Gnedenko, and
[12] J. L. Cabrera and J. G. Milton, On-off intermittency in A. Kolmogorov, Limit distributions for sums of inde-
a human balancing task, Physical Review Letters 89, pendent, Am. J. Math 105 (1954).
158702 (2002). [33] V. V. Uchaikin and V. M. Zolotarev, Chance and stabil-
[13] J. L. Cabrera and J. G. Milton, Stick balancing: On-off ity: stable distributions and their applications (Walter
intermittency and survival times, Nonlinear Studies 11, de Gruyter, 2011).
305 (2004). [34] B. Dybiec and E. Gudowska-Nowak, Resonant activa-
[14] G. Margolin, V. Protasenko, M. Kuno, and E. Barkai, tion in the presence of nonequilibrated baths, Physical
Power law blinking quantum dots: Stochastic and phys- Review E 69, 016105 (2004).
ical models, arXiv preprint cond-mat/0506512 (2005). [35] B. B. Mandelbrot, The fractal geometry of nature, Vol.
[15] P. Frantsuzov, M. Kuno, B. Janko, and R. A. Mar- 173 (WH freeman New York, 1983).
cus, Universal emission intermittency in quantum dots, [36] M. Shlesinger, B. West, and J. Klafter, Lévy dynamics
nanorods and nanowires, Nature Physics 4, 519 (2008). of enhanced diffusion: Application to turbulence, Phys-
[16] S. J. Benavides and A. Alexakis, Critical transitions in ical Review Letters 58, 1100 (1987).
thin layer turbulence, Journal of Fluid Mechanics 822, [37] T. Solomon, E. R. Weeks, and H. L. Swinney, Obser-
364 (2017). vation of anomalous diffusion and lévy flights in a two-
[17] A. van Kan and A. Alexakis, Condensates in thin-layer dimensional rotating flow, Physical Review Letters 71,
turbulence, Journal of Fluid Mechanics 864, 490 (2019). 3975 (1993).
[18] D. Sweet, E. Ott, J. M. Finn, T. M. Antonsen Jr, and [38] R. Metzler and J. Klafter, The random walk’s guide to
D. P. Lathrop, Blowout bifurcations and the onset of anomalous diffusion: a fractional dynamics approach,
magnetic activity in turbulent dynamos, Physical Re- Physics reports 339, 1 (2000).
view E 63, 066211 (2001). [39] A. A. Dubkov, B. Spagnolo, and V. V. Uchaikin, Lévy
[19] A. Alexakis and Y. Ponty, Effect of the lorentz force flight superdiffusion: an introduction, International
on on-off dynamo intermittency, Physical Review E 77, Journal of Bifurcation and Chaos 18, 2649 (2008).
056308 (2008). [40] B. Dubrulle and J.-P. Laval, Truncated lévy laws
and 2d turbulence, The European Physical Journal B-
 16

Condensed Matter and Complex Systems 4, 143 (1998). [59] A. van Kan, A. Alexakis, and M. E. Brachet, Inter-
[41] D. del Castillo-Negrete, B. Carreras, and V. Lynch, mittency of three-dimensional perturbations in a point-
Nondiffusive transport in plasma turbulence: a frac- vortex model, submitted to Physical Review E (2020).
tional diffusion approach, Physical review letters 94, [60] K. Seshasayanan and B. Gallet, Onset of three-
065003 (2005). dimensionality in rapidly rotating turbulent flows, Jour-
[42] C. Schinckus, How physicists made stable lévy processes nal of Fluid Mechanics 901, R5 (2020).
physically plausible, Brazilian Journal of Physics 43, [61] S. Jespersen, R. Metzler, and H. C. Fogedby, Lévy
281 (2013). flights in external force fields: Langevin and fractional
[43] P. D. Ditlevsen, Anomalous jumping in a double-well fokker-planck equations and their solutions, Physical
potential, Physical Review E 60, 172 (1999). Review E 59, 2736 (1999).
[44] P. D. Ditlevsen, Observation of α-stable noise induced [62] A. Chechkin, V. Gonchar, J. Klafter, R. Metzler, and
millennial climate changes from an ice-core record, Geo- L. Tanatarov, Stationary states of non-linear oscillators
physical Research Letters 26, 1441 (1999). driven by lévy noise, Chemical Physics 284, 233 (2002).
[45] G. M. Viswanathan, V. Afanasyev, S. Buldyrev, E. Mur- [63] A. V. Chechkin, J. Klafter, V. Y. Gonchar, R. Metzler,
phy, P. Prince, and H. E. Stanley, Lévy flight search pat- and L. V. Tanatarov, Bifurcation, bimodality, and finite
terns of wandering albatrosses, Nature 381, 413 (1996). variance in confined lévy flights, Physical review E 67,
[46] D. W. Sims, E. J. Southall, N. E. Humphries, G. C. 010102 (2003).
Hays, C. J. Bradshaw, J. W. Pitchford, A. James, M. Z. [64] A. V. Chechkin, V. Y. Gonchar, J. Klafter, R. Metzler,
Ahmed, A. S. Brierley, M. A. Hindell, et al., Scaling laws and L. V. Tanatarov, Lévy flights in a steep potential
of marine predator search behaviour, Nature 451, 1098 well, Journal of Statistical Physics 115, 1505 (2004).
(2008). [65] B. Dybiec, E. Gudowska-Nowak, and I. Sokolov, Sta-
[47] I. Rhee, M. Shin, S. Hong, K. Lee, S. J. Kim, and tionary states in langevin dynamics under asymmetric
S. Chong, On the levy-walk nature of human mobility, lévy noises, Physical Review E 76, 041122 (2007).
IEEE/ACM transactions on networking 19, 630 (2011). [66] S. Denisov, W. Horsthemke, and P. Hänggi, Steady-
[48] M. C. Gonzalez, C. A. Hidalgo, and A.-L. Barabasi, Un- state lévy flights in a confined domain, Physical Review
derstanding individual human mobility patterns, nature E 77, 061112 (2008).
453, 779 (2008). [67] B. Dybiec, I. M. Sokolov, and A. V. Chechkin, Station-
[49] A. M. Edwards, R. A. Phillips, N. W. Watkins, M. P. ary states in single-well potentials under symmetric lévy
Freeman, E. J. Murphy, V. Afanasyev, S. V. Buldyrev, noises, Journal of Statistical Mechanics: Theory and
M. G. da Luz, E. P. Raposo, H. E. Stanley, et al., Re- Experiment 2010, P07008 (2010).
visiting lévy flight search patterns of wandering alba- [68] A. Padash, A. V. Chechkin, B. Dybiec, I. Pavlyukevich,
trosses, bumblebees and deer, Nature 449, 1044 (2007). B. Shokri, and R. Metzler, First-passage properties of
[50] B. Gross, Z. Zheng, S. Liu, X. Chen, A. Sela, J. Li, D. Li, asymmetric lévy flights, Journal of Physics A: Mathe-
and S. Havlin, Spatio-temporal propagation of covid- matical and Theoretical 52, 454004 (2019).
19 pandemics, EPL (Europhysics Letters) 131, 58003 [69] T. Srokowski, Fractional fokker-planck equation for lévy
(2020). flights in nonhomogeneous environments, Physical Re-
[51] J. L. Cabrera and J. G. Milton, Human stick balancing: view E 79, 040104 (2009).
tuning lévy flights to improve balance control, Chaos: [70] T. Srokowski, Multiplicative lévy processes: Itô ver-
An Interdisciplinary Journal of Nonlinear Science 14, sus stratonovich interpretation, Physical Review E 80,
691 (2004). 051113 (2009).
[52] R. Metzler and J. Klafter, The restaurant at the end [71] A. La Cognata, D. Valenti, A. Dubkov, and B. Spagnolo,
of the random walk: recent developments in the de- Dynamics of two competing species in the presence of
scription of anomalous transport by fractional dynam- lévy noise sources, Physical Review E 82, 011121 (2010).
ics, Journal of Physics A: Mathematical and General [72] T. Srokowski, Nonlinear stochastic equations with mul-
37, R161 (2004). tiplicative lévy noise, Physical Review E 81, 051110
[53] D. Applebaum, Lévy processes-from probability to fi- (2010).
nance and quantum groups, Notices of the AMS 51, [73] T. Srokowski, Multiplicative lévy noise in bistable sys-
1336 (2004). tems, The European Physical Journal B 85, 65 (2012).
[54] J. A. Roberts, T. W. Boonstra, and M. Breakspear, The [74] A. V. Chechkin, V. Y. Gonchar, J. Klafter, and R. Met-
heavy tail of the human brain, Current opinion in neu- zler, Barrier crossing of a lévy flight, EPL (Europhysics
robiology 31, 164 (2015). Letters) 72, 348 (2005).
[55] M. F. Shlesinger and J. Klafter, Lévy walks versus lévy [75] B. Dybiec, E. Gudowska-Nowak, and P. Hänggi, Escape
flights, in On growth and form (Springer, 1986) pp. 279– driven by α-stable white noises, Physical Review E 75,
283. 021109 (2007).
[56] V. Zaburdaev, S. Denisov, and J. Klafter, Lévy walks, [76] A. V. Chechkin, O. Y. Sliusarenko, R. Metzler, and
Reviews of Modern Physics 87, 483 (2015). J. Klafter, Barrier crossing driven by lévy noise: Uni-
[57] Y. Jung, E. Barkai, and R. J. Silbey, Lineshape the- versality and the role of noise intensity, Physical Review
ory and photon counting statistics for blinking quantum E 75, 041101 (2007).
dots: a lévy walk process, Chemical Physics 284, 181 [77] T. Koren, M. A. Lomholt, A. V. Chechkin, J. Klafter,
(2002). and R. Metzler, Leapover lengths and first passage time
[58] G. Margolin and E. Barkai, Nonergodicity of blinking statistics for lévy flights, Physical review letters 99,
nanocrystals and other lévy-walk processes, Physical re- 160602 (2007).
view letters 94, 080601 (2005). [78] K. Capala, A. Padash, A. V. Chechkin, B. Shokri,
R. Metzler, and B. Dybiec, Lévy noise-driven escape
 17

from arctangent potential wells, Chaos: An Interdisci- american statistical association 71, 340 (1976).
plinary Journal of Nonlinear Science 30, 123103 (2020). [95] K. Itô, Stochastic integral, Proceedings of the Imperial
[79] L. Zeng, R. Bao, and B. Xu, Effects of lévy noise in ape- Academy 20, 519 (1944).
riodic stochastic resonance, Journal of physics A: Math- [96] S. I. Denisov, W. Horsthemke, and P. Hänggi, General-
ematical and Theoretical 40, 7175 (2007). ized fokker-planck equation: Derivation and exact solu-
[80] B. Dybiec, Lévy noises: Double stochastic resonance in tions, The European Physical Journal B 68, 567 (2009).
a single-well potential, Physical Review E 80, 041111 [97] F. Mainardi, Y. Luchko, G. Pagnini, and D. T. R.
(2009). Gorenflo, The fundamental solution of the space-time
[81] B. Dybiec and E. Gudowska-Nowak, Lévy stable noise- fractional diffusion equation, Fract. Calc. Appl. Anal ,
induced transitions: stochastic resonance, resonant ac- 153.
tivation and dynamic hysteresis, Journal of Statisti- [98] S. G. Samko, A. A. Kilbas, O. I. Marichev, et al.,
cal Mechanics: Theory and Experiment 2009, P05004 Fractional integrals and derivatives, Vol. 1 (Gordon and
(2009). Breach Science Publishers, Yverdon Yverdon-les-Bains,
[82] Y. Xu, J. Li, J. Feng, H. Zhang, W. Xu, and J. Duan, Switzerland, 1993).
Lévy noise-induced stochastic resonance in a bistable [99] P. Carr and L. Wu, The finite moment log stable pro-
system, The European Physical Journal B 86, 198 cess and option pricing, The journal of finance 58, 753
(2013). (2003).
[83] R. Yamapi, R. M. Yonkeu, G. Filatrella, and J. Kurths, [100] K. Seshasayanan and F. Pétrélis, Growth rate distribu-
Lévy noise induced transitions and enhanced stability in tion and intermittency in kinematic turbulent dynamos:
a birhythmic van der pol system, The European Physi- Which moment predicts the dynamo onset?, EPL (Eu-
cal Journal B 92, 152 (2019). rophysics Letters) 122, 64004 (2018).
[84] I. M. Sokolov, W. Ebeling, and B. Dybiec, Harmonic [101] R. N. Mantegna and H. E. Stanley, Stochastic process
oscillator under lévy noise: Unexpected properties in with ultraslow convergence to a gaussian: the truncated
the phase space, Physical Review E 83, 041118 (2011). lévy flight, Physical Review Letters 73, 2946 (1994).
[85] T. Tanaka, Low-dimensional dynamics of phase oscil- [102] I. Koponen, Analytic approach to the problem of con-
lators driven by cauchy noise, Physical Review E 102, vergence of truncated lévy flights towards the gaussian
042220 (2020). stochastic process, Physical Review E 52, 1197 (1995).
[86] A. Dubkov and B. Spagnolo, Verhulst model with lévy [103] W. Zan, Y. Xu, J. Kurths, A. V. Chechkin, and R. Met-
white noise excitation, The European Physical Journal zler, Stochastic dynamics driven by combined lévy–
B 65, 361 (2008). gaussian noise: fractional fokker–planck–kolmogorov
[87] B. Dybiec, E. Gudowska-Nowak, and I. Sokolov, Trans- equation and solution, Journal of Physics A: Mathemat-
port in a lévy ratchet: Group velocity and distribution ical and Theoretical 53, 385001 (2020).
spread, Physical Review E 78, 011117 (2008). [104] P. Xu, T. Zhou, R. Metzler, and W. Deng, Lévy walk
[88] M. Ding and W. Yang, Distribution of the first return dynamics in an external harmonic potential, Physical
time in fractional brownian motion and its application Review E 101, 062127 (2020).
to the study of on-off intermittency, Physical Review E [105] F. Pétrélis and S. Aumaı̂tre, Modification of instability
52, 207 (1995). processes by multiplicative noises, The European Physi-
[89] S. Aumaı̂tre, F. Pétrélis, and K. Mallick, Low-frequency cal Journal B-Condensed Matter and Complex Systems
noise controls on-off intermittency of bifurcating sys- 51, 357 (2006).
tems, Physical review letters 95, 064101 (2005). [106] R. Graham and A. Schenzle, Stabilization by multiplica-
[90] S. Aumaı̂tre, K. Mallick, and F. Pétrélis, Effects of the tive noise, Physical Review A 26, 1676 (1982).
low frequencies of noise on on–off intermittency, Journal [107] K. Mallick and P. Marcq, Stability analysis of a noise-
of statistical physics 123, 909 (2006). induced hopf bifurcation, The European Physical Jour-
[91] A. Alexakis and F. Pétrélis, Critical exponents in zero nal B-Condensed Matter and Complex Systems 36, 119
dimensions, Journal of Statistical Physics 149, 738 (2003).
(2012). [108] E. Bertin, On-off intermittency over an extended range
[92] F. Pétrélis and A. Alexakis, Anomalous exponents at of control parameter, Physical Review E 85, 042104
the onset of an instability, Physical Review Letters 108, (2012).
014501 (2012). [109] F. Liu, V. Anh, and I. Turner, Numerical solution of
[93] A. Alexakis and F. Pétrélis, Planar bifurcation subject the space fractional fokker–planck equation, Journal
to multiplicative noise: Role of symmetry, Physical Re- of Computational and Applied Mathematics 166, 209
view E 80, 041134 (2009). (2004).
[94] J. M. Chambers, C. L. Mallows, and B. Stuck, A method
for simulating stable random variables, Journal of the