Waves and their modulations

Sylvie Benzoni-Gavage

To cite this version:

Sylvie Benzoni-Gavage. Waves and their modulations. 2020. �hal-02881881�

HAL Id: hal-02881881

https://hal.archives-ouvertes.fr/hal-02881881
Preprint submitted on 26 Jun 2020

HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est

archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents
entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non,
lished or not. The documents may come from émanant des établissements d’enseignement et de
teaching and research institutions in France or recherche français ou étrangers, des laboratoires
abroad, or from public or private research centers. publics ou privés.
 Waves and their modulations
Sylvie Benzoni-Gavage
June 26, 2020

Abstract
The words of the title, waves and modulations, are ubiquitous in science and tech-
nology. They are in particular the purpose of an active field of research in applied
mathematics. Modulation ’theory’ is fifty years old and nevertheless conceals many
open questions. An overview of it is given along a journey in the landscape of a few
famous equations in mathematical physics.

1 Oscillations
Our life is bound to many periodic phenomena, starting with the rhythm of our heart beat
to the rotation on itself and the orbiting of the Earth around the Sun.
Mathematically speaking, periodic signals may have all sorts of form but the most basic
ones are represented by (co)sine functions. For instance, one may consider an oscillating signal
represented as a function of time t by y = cos(ω0 t), where ω0 is a parameter determining
how often the signal goes back to its original value. More precisely, the period of this signal
is T0 = 2π/ω0 , and the inverse of T0 is called the frequency of the signal. In terms of music,
the larger the frequency, the more high-pitched the note.
For instance a concert A produced by a tuning fork has a - fundamental - frequency of
440 Hz, meaning that it - almost - corresponds to a sinusoidal signal of period 1/440 ≃ 0.0023
seconds. However, a purely sinusoidal signal of 440 Hz or any other frequency is not at all
pleasant to hear. Musical instruments produce much richer sounds, involving in particular
’modulations’ that make for example bells ring different from guitars.
Before going to the actual meaning of the word ’modulation’ in science, let us mention
that it has been used in technology for decades, modulation being used to encode information
in telecommunications. The word appears in particular, more or less hidden, on standard
radio sets. The F that can be seen in the abbreviation FM on a radio set is indeed for
frequency, the M meaning modulation. We can also see AM on a radio set, which means
amplitude modulation and is slightly easier to formulate than frequency modulation.
Performing amplitude modulation of a reference signal amounts to considering a modified
signal of the form y = a(t) cos(ω0 t), where the amplitude a = a(t) varies over large time scales
compared to the period of the reference signal. For instance we may consider a = cos(ωt)
with ω being much smaller than ω0 . An example is plotted on Figure 1.

1
Figure 1: Plot of y = cos(20πt) (top), y = cos(2πt) (purple), and y = cos(2πt) cos(20πt)
(red).

Frequency modulation of a reference signal y = cos(ω0 t) amounts to modifying its linear

phase ω0 t into a phase function φ = φ(t) in such a way that φ′ (t) oscillates around ω0 over
large time scales compared to T0 = 2π/ω0 . For instance we may consider φ = ω0 t + h cos(ωt)
with ω0 being a multiple of ω so as to keep a periodic signal, namely of period T = 2π/ω,
which is then a multiple of T0 , and with h a parameter tuning the height of frequency
oscillations. An example is plotted on Figure 2.

Figure 2: Plot of y = cos(20πt) (top), y = cos(2πt) (purple), and y = cos(20πt + 3 cos(2πt))

(orange).

2
2 Linear waves
Let us now consider a function of both time t and spatial position x of the form u = cos(k0 x−
ω0 t), assuming for the time being that both k0 and ω0 are positive, tuning parameters.
Depending on whether u is viewed as a function of t or x, it has a time period T0 = 2π/ω0
as before, but also another period in space called the wavelength, X0 = 2π/k0 .
If we plot u as a function of x at each time t we receive a movie of a sinusoidal signal
propagating at velocity c0 = ω0 /k0 , meaning that at some time t we get a signal that is
just the initial one shifted by a distance c0 t. This may represent for instance the pressure
wave carrying - at a velocity of about 340 m per second in the air - a pure musical note of
frequency 1/T0 .
Such a wave is a particular solution of the partial differential equation (PDE)

ut + c0 ux = 0, (1)

called a transport equation, in which the subscripts stand for partial derivatives - as will be
the case everywhere t or x appear in a subscript. It also solves the wave equation

utt − c20 uxx = 0,

which admits in addition solutions that propagate in the opposite direction - at the same
speed c0 .
There are many other PDEs admitting solutions that are oscillatory traveling waves. In
particular, one may consider the Airy equation1

ut + αuxxx = 0, (2)

which admits solutions of the form u = cos(kx − ωt) provided that the wave number k is
linked to ω through the dispersion relation

ω + αk 3 = 0. (3)

Here α is just a parameter introduced for the sake of physical consistency. Whether it is
positive or negative determines the sign of ω, and thus that of the phase speed c = ω/k. This
speed depends on k and for this reason we say that the Airy equation (2) is dispersive, since
waves of distinct wave numbers propagate at different speeds and thus disperse from each
other as time goes on.
Equations (1) and (2) have in common that they are linear, and as a consequence any
two solutions of them may be added to yield a third solution. If we stick to solutions of the
form u = cos(kx − ωt), and take two of them, u1 = cos(k1 x − ω1 t) and u2 = cos(k2 x − ω2 t),
we get a third one by taking their mean value
1
u= (cos(k1 x − ω1 t) + cos(k2 x − ω2 t)),
2
1
Named after George Biddell Airy [1801–1892].

3
By trigonometry this equivalently reads

u = cos(kx − ωt) cos(k0 x − ω0 t)

with k = (k1 − k2 )/2 (assuming for instance that k1 > k2 to ensure a positive wave number
k), ω = (ω1 − ω2 )/2, k0 = (k1 + k2 )/2, ω0 = (ω1 + ω2 )/2. If k is much smaller than k0 , at
each time t the resulting function of x is a modulation of the carrier wave cos(k0 x − ω0 t)
with amplitude cos(kx − ωt). Note that k happens to be much smaller than k0 when the
wave numbers k1 and k2 are chosen close enough to each other.
There is a notable difference between solutions to (1) and solutions to (2). Indeed, in
the former all waves propagate at speed c0 , so that ω/k = ω0 /k0 . Both the carrier wave
and the amplitude and therefore also the modulated wave propagate at speed c0 , whereas
for solutions to the Airy equation (2) the two speeds ω/k = −αk 2 and ω0 /k0 = −αk02 are
distinct as soon as k1 ̸= k2 .
Would we watch the movie of a modulated wave solution to (2) we could see that the
higher frequency oscillations of the carrier wave cos(k0 x − ω0 t) move inside the envelope
determined by the amplitude cos(kx − ωt).
Moreover, let us point out that the speed of propagation of the amplitude

ω/k = (ω1 − ω2 )/(k1 − k2 ) = −α(k13 − k23 )/(k1 − k2 )

goes to −3αk02 when k1 and k2 both approach k0 . This value −3αk02 is obtained by differen-
tiating ω with respect to k at k0 in the dispersion relation (3). It is called the group velocity
of the carrier wave, which is here the triple of its phase speed ω0 /k0 = −αk02 . The fact that
the phase speed differs from the group velocity, which is equivalent to the fact that the phase
speed is not constant as a function of the wave number, is a manifestation of the already
mentioned dispersive feature of (2).
By using the elementary solutions described above and Fourier analysis2 , one can find a
representation formula for all solutions of linear equations such as (1) and (2). Alternatively,
the resolution of (1) actually reduces to merely observing, without any Fourier analysis, that
u(x, t) = u(x − c0 t, 0) for all x and t and any solution u of (1). The resolution of (2) is more
delicate, and involves an intriguing function named as the equation after Airy. Dwelling on
this resolution is not the purpose of this paper though.
Rather, we aim at considering some nonlinear PDEs, which are not amenable to Fourier
analysis, and that still admit periodic waves propagating at constant speed.

3 Nonlinear periodic waves

Generally speaking, any function of the form u = U (x − ct) is called a traveling wave. At any
time t, the graph of u(·, t) is the one of U shifted by ct. The function U is called the profile
of the wave.
2
Fourier analysis is a domain of mathematics with numerous applications nowadays that was funded by
the 19th century scientist Joseph Fourier [1768–1830].

4
 The perhaps most well-known nonlinear PDE that admits periodic traveling wave solu-
tions is the Korteweg–de Vries equation3
ut + uux + αuxxx = 0. (4)
This is a model equation for long water waves, which can be viewed as a ’mixture’ of the
Airy equation (2) and of a nonlinear transport equation
ut + uux = 0. (5)
In the latter, the speed of propagation is given by the unknown u itself, by contrast with the
constant velocity c in (1). The periodic traveling wave solutions to (4) involve Jacobi elliptic
functions, which we shall not use here. Some of them are called cnoidal waves, displaying
sharper crests and flatter troughs than in a sine wave (see Figure 3 for an artist view).

Figure 3: Cnoidal wave profile - plot of elliptic cosine cn(·, 0.9). CC BY-SA 3.0 Kraaiennest.

Another well-known nonlinear PDE with similar features and even more elementary pe-
riodic traveling wave solutions is the nonlinear Schrödinger equation4
zt = iαzxx − iγz|z|2 , (6)
in which the unknown z takes complex values, and α, γ are real parameters. This model is
used in various fields of mathematical physics, ranging from quantum mechanics to nonlinear
optics and water waves on deep water.
We can see that for any positive numbers r and k, the function z = rei(kx−ωt) is a solution
of (6) if and only if
ω = αk 2 + γr2 . (7)
This relation between the frequency ω, the wave number k and the amplitude r is called a
nonlinear dispersion relation. Solutions of the form z = rei(kx−ωt) are called harmonic waves,
their real and imaginary parts being as the (co)sine waves discussed in Section 2. They are
not the only periodic traveling wave solutions to (6). This nonlinear PDE also admits more
complicated periodic traveling wave solutions - namely, cnoidal waves again - that will not
be described here.
Let us give a last example of a nonlinear PDE admitting periodic traveling wave solutions.
This one can be viewed as a ’damped’5 modification of (6). It is called the complex Ginzburg–
Landau equation6 and reads
zt = (1 + iα)zxx + z − (1 + iγ)z|z|2 . (8)
3
Named after Diederik Johannes Korteweg [1848–1941] and Gustav de Vries [1866–1934]. For more infor-
mation see [9].
4
Named after Erwin Schrödinger [1887–1961].
5
This word is not essential here and we shall not try to explain it.
6
Named after Vitaly Ginzburg [1916–2009] and Lev Landau [1908–1968], the complex Ginzburg–Landau
equation is the main example considered in the seminal work [11] on the topic we address here.

5
This equation arises for instance in the modeling of wakes past obstacles in viscous fluids
flows. We can see that z = rei(kx−ωt) is a solution of (8) if and only if

r2 = 1 − k2 , ω = γ + (α − γ)k 2 . (9)

We thus still find harmonic waves, but with only one degree of freedom, the wave number k
determining both ω and r (up to a ± sign). By comparison, there are two degrees of freedom
for harmonic wave solutions to (6), namely, the wave number k and the modulus r, which
are both free and determine ω through (7).
Because of nonlinearity, we cannot add together harmonic waves to build modulated wave
solutions to those PDEs. Yet it is natural to ask whether they admit modulated wave trains,
say of the form z = r(x, t)eiφ(x,t) where the amplitude r, the local wave number φx and the
local frequency φt vary significantly only over large scales. This turns out to be a tough
question in general.
More precisely, this topic involves the following series of natural questions, starting from
the most basic one up to the really tough one.
1. For a given nonlinear PDE, what are the equations governing the large scale varia-
tions of the amplitude and the phase of modulated wave train solutions, at least in an
approximate manner?

2. Do these equations, referred to as modulated equations, actually have solutions on suf-

ficiently large scales, in particular for a long enough time?

3. Once we have a positive answer to the previous question, to what extent does it give
information on actual solutions of the original PDE? In other words, how can we build
actual modulated wave train solutions from this approach?
The case when modulated equations have long-term solutions is the good one as regards
mathematical analysis of wave trains. From an applied mathematics point of view, it can be
more interesting to address the following.
4. What happens when the modulated equations fail to have long-term solutions?
Answering question 1 requires little creativity. As a matter of fact the derivation of
modulated equations, which are expected to provide an asymptotic model for large scale
variations, is just a matter of calculus, as we exemplify in the next two sections.
The way Question 2 can be dealt with depends on the form of modulated equations. In
the simplest cases modulated equations can turn out to be classical - systems of - PDEs, as in
the examples given in Section 4, so that it is not difficult to guess under which conditions(s)
they have long-term solutions.
However, as explained in Section 5, modulated equations in general take the form of
averaged equations involving mean values of periodic wave profiles and nonlinear functions
of these profiles. Since the wave profiles are not known ’explicitly’ in general - and at best
known through rather complicated special functions for a few special PDEs - , determining
whether modulated equations have long-term solutions is already an issue.

6
 Question 3 has been mostly unanswered for fifty years, except for a couple of special
PDEs, while Question 4 has been partly answered. Section 6 is devoted to giving the flavor
of a rule of thumb about Question 4 that became a mathematical result in various frameworks
in the last two decades.
To finish with this overview, let us stress that besides a rather complete theory for the
Korteweg–de Vries equation and a more limited one for the nonlinear Schrödinger equation,
there have been recent breakthroughs regarding Question 3 for some classes of ’dissipative’
PDEs. Section 7 gives a glimpse of the known results and of the remaining issues in general.

4 Some explicit modulated equations

The study of modulated wave trains was initiated in the mid 1960s by Whitham [31], espe-
cially for dispersive PDEs such as (4). More precisely, Gerald Whitham [1927–2014] pointed
out several methods to derive modulated equations that should govern, at least approximately,
the evolution of modulated wave trains, should they exist. One of those methods relies on a
possible underlying Lagrangian for the reference PDE, and is thus by nature not applicable
to dissipative equations such as (8). Another method is what he called the ’two-timing’
method. This one is more systematic and is actually based on a formal, multi-scale expan-
sion. The machinery for deriving modulated equations in this way is a little bit tedious but
not complicated. Let us exemplify it with two simple cases.

4.1 Basic example

Let us start by giving its flavor for (8) - we refer to [11] for more details. We recall that
harmonic wave solutions to (8) are of the form
√ 2
z = 1 − k 2 ei(kx−(γ+(α−γ)k )t) ,

according to the nonlinear dispersion relation (9).

For simplicity - as in [11, § 3.3] - we concentrate on modulations of harmonic waves with
k = 0. Let us start by noticing that there are actually harmonic wave solutions to (8) with
arbitrarily small wave numbers k = εK, namely
√ 2 2
zε = 1 − ε2 K 2 ei(−γt+εKx−(α−γ)ε K t) (10)

for any fixed K and an arbitrarily small positive parameter ε . The first reason why we
introduce K here (instead of just considering the case K = 1) is because we think of ε as
being nondimensional (that is, with no unit of measurement contrary to k and K, which are
homogeneous the inverse of a length).
By construction, (10) defines a family of solutions to (8) parametrized by ε that are ’slow’
modulations over ’large’ spatial scales of the limiting one z0 = e−iγt . Indeed, for any positive
ε the function zε in (10) is equal to z0 multiplied by a harmonic wave of wavelength 2π/(Kε)
and time period 2π/((α − γ)K 2 ε2 ), both going to infinity when ε goes to zero.

7
 A deeper reason why we have introduced K is that we want to look for more general
modulations of z0 in terms of the rescaled variables χ = εx and τ = ε2 t, in which K will be
varying as a local wave number with respect to χ and τ . More precisely, we seek solutions
to (8) of the form
2
zε = (1 + ε2 Rε (εx, ε2 t)) ei(−γt+Φε (εx,ε t)) (11)
We already know special solutions of this form, namely those given in (10), which correspond
to Φε (χ, τ ) = Kχ − (α − γ)K 2 τ for a fixed K, and
√
Rε = ( 1 − ε2 K 2 − 1)/ε2 = −K 2 /2 + O(ε2 )
by Taylor expansion. For those special solutions the phase function Φε is linear in (χ, τ ) and
actually independent of ε.
Let us now seek solutions to (8) of the form (11) by just assuming that Rε and Kε = Φε,χ
admit asymptotic expansions
Rε = R + O(ε), Kε = K + O(ε)
that depend smoothly on the variables (χ, τ ). Then, by plugging (11) in (8) we find through
a calculus exercise the following equations for the lower order terms R, K
R = −K 2 /2 − αKχ /2 ,
Kτ + 2(α − γ)KKχ = (1 + αγ)Kχχ , (12)
in which the subscripts in τ and χ stand for partial derivatives. The latter is the sought
modulated equation associated with (8). The properties of this equation depend crucially on
the sign of the right-hand side coefficient 1 + αγ. The ’good’ case is when 1 + αγ is positive.
Then (12) is a parabolic PDE called Burgers equation7 , a ’mixture’ of the heat equation
Kτ = (1 + αγ)Kχχ
and the nonlinear transport equation
Kτ + 2(α − γ)KKχ = 0.
The fact that the modulated equation (12) is parabolic is reminiscent of the parabolic nature
of the Ginzburg-Landau equation (8). The term parabolic for the latter merely refers to the
fact that the set of pairs (ω, k) satisfying the second relation in (9) is a parabola. The reason
for classifying the heat equation as parabolic is similar, and by extension we also say the
Burgers equation is parabolic.
Parabolic PDEs are known to have solutions for all positive times. So the answer to
Question 2 is positive for the modulated equation (12) when 1 + αγ is positive. We will
come back in Section 7 to Question 3 regarding the link between solutions to (12) and actual
solutions to (8).
We have seen so far an example of a PDE (8) associated with a single modulated equation
(12). This is actually the simplest situation.
7
Named after Jan Burgers [1895–1981].

8
4.2 Another example
In general, there are several modulated equations associated with a single PDE. As an ex-
ample, let us derive modulated equations associated with (6).
Keeping in mind that in view of its nonlinear dispersion relation (7) Equation (6) admits
harmonic wave solutions of the form
2 +αk 2 )t)
z = r ei(kx−(γr ,
we may again investigate modulations of z0 = e−iγt , corresponding to k = 0 and r = 1. We
thus seek solutions to (6) of the form
zε = (1 + εRε (εx, εt)) ei(−γt+Φε (εx,εt)) (13)
for a small positive parameter ε. The chosen scaling is again consistent with the form of the
explicit solutions we already know
2 +αε2 K 2 )t)
(1 + εR) ei(εKx−(γ(1+εR) ,
which are of the form (13) for any R and K, with Rε = R and
Φε (χ, τ ) = Kχ − γτ (2R + εR2 ) − αεK 2 τ = Kχ − 2γτ R + O(ε),
where we have denoted χ = εx as before, and here τ = εt. Then, just assuming that Rε and
Kε = Φε,χ admit asymptotic expansions
Rε = R + O(ε), Kε = K + O(ε)
that depend smoothly on the variables (χ, τ ), by plugging (13) in (6) we find (through a
calculus exercise very much similar to the derivation of (12), with actually fewer terms to
deal with) the following equations for the lower order terms R, K
Rτ + αKχ = 0, Kτ + 2γRχ = 0. (14)
The fact that we receive a system of PDEs instead of a single PDE as in (12) comes from the
fact that harmonic wave solutions to (6) have one more degree of freedom than those of (8).
We can see that smooth enough solutions of System (14) are such that K satisfies a
single, second order PDE. Indeed, we can eliminate R by differentiating both equations, and
we receive
Kτ τ − 2αγKχχ = 0,
which we recognize as a wave equation provided that αγ be positive (this one is called
hyperbolic merely because the set of solutions of its dispersion relation is a hyperbola). The
case αγ > 0 is thus the ’good’ one, for which the modulated equations have solutions for all
times.
In [13], a slightly more complicated modulated system is considered, which governs the
leading order part of more general solutions of the form
zε = rε (εx, εt) eiφε (εx,εt)/ε . (15)

9
One of its two equations pertains to a very general class of modulated equations, sometimes
called the conservation of waves equation - also called Eikonal equation in geometric optics.
To explain its derivation, let us assume that rε = r + O(ε) and that the phase function φε is
such that
φε,χ = k + O(ε) , φε,τ = −ω + O(ε).
We thus find that for zε of the form in (15) to solve (6), r, k, and ω must satisfy (7) as
functions of (χ, τ ). Moreover, writing the equality of cross derivatives φ0,χτ = φ0,τ χ we
readily obtain the modulated equation
kτ + (αk 2 + γr2 )χ = 0. (16)
The other modulated equation is obtained by plugging (15) in (6) and equating to zero the
terms of order ε in (6). It turns out to read
rτ + αrkχ + 2αkrχ = 0. (17)
The system (16)-(17) is now nonlinear. However one may observe that its linearized version
about (k, r) = (0, 1) - obtained by keeping only first order approximations of quadratic
quantities - coincides with (14). It is interesting to note that the ’good’ case for (16)-(17) is
still αγ > 0, as for (14). This condition ensures indeed that (16)-(17) is strictly hyperbolic at
any point (k, r), not only at (0, 1), as we explain below. As a matter of fact, the hyperbolicity
of (16)-(17) is to be checked on its quasilinear form
( ) ( )( ) ( )
kτ 2αk 2γr kχ 0
+ = , (18)
rτ αr 2αk rχ 0
which is by definition equivalent to the set of equations (16)-(17) as long as we stick to
smooth solutions. Remarkably enough, the transport operator ∂τ + 2αk∂χ appears in both
rows of (18), and the speed 2αk happens to be the group velocity, obtained by differentiating
ω with respect to k in the nonlinear dispersion relation (7). However the nondiagonal terms
in the characteristic matrix ( )
2αk 2γr
M :=
αr 2αk
are also important to determine the properties of (18). The latter is said to be strictly
hyperbolic if the characteristic matrix M has distinct real eigenvalues8 , which is the case
precisely when αγ is positive. Otherwise, and more precisely if αγ were negative the initial
value problem for (18) would be ill-posed.
Without trying to dwell on the ill-posedness issues here, let us mention that the word ’hy-
perbolic’ is used as a generalization of what happens for the wave equation, even though the
implications of hyperbolicity are less strong for nonlinear equations than for linear ones. In
general, nonlinear hyperbolic systems involve indeed finite-time blow-up of solutions. Local-
in-time solutions can nevertheless give valuable information, since we are talking about the
rescaled time τ = εt. See Section 6 for more details.
8
The eigenvalues X of M are characterized by the existence of directions in which M acts as the multi-
plication by X. They are found as the roots of the polynomial X 2 − 4αkX + 4α2 k 2 − 2αγr2 .

10
5 More modulated equations
Modulated equations are actually not restricted to the study of wave trains of the form (15).
They may serve as a tool for investigating wave trains depending on a phase φε (εx, εt) in
a more general manner than just through cosine/sine functions as in (15). The drawback
of general modulated equations is that they are not as explicit as the ones described in the
previous section.
Let us exemplify this on the seminal case of the Korteweg–de Vries equation (4). We seek
solutions of the form
uε = Uε (εx, εt; φε (εx, εt)/ε) (19)
with ε a small positive parameter and Uε being periodic - but not a priori harmonic - in its
last argument, denoted by θ in what follows. We still use the notation χ = εx and τ = εt
for the rescaled variables. Without loss of generality we may assume the period in θ to be
equal to one. Denoting kε = φε,χ and ωε = −φε,τ , by the chain rule we have ∂t = ∂τ − ωε ∂θ ,
∂x = ∂χ + kε ∂θ . So, assuming that

kε = k(χ, τ ) + O(ε) , ωε = ω(χ, τ ) + O(ε),

Uε = U (χ, τ ; θ) + εV (χ, τ ; θ) + O(ε2 ),

we find that for Uε to solve (4) we must have

−ωUθ + kU Uθ + αk 3 Uθθθ = 0,

where the subscripts in θ stand once more for partial derivatives. This relation is obtained
by equating to zero the leading order term in powers of ε when the ansatz (19) is plugged
into (4). By straightforward integration it implies the existence of λ = λ(χ, τ ) such that

− ωU + 12 kU 2 + αk 3 Uθθ = λ k. (20)

(The factor k is introduced here above in the constant of integration for convenience, it will
soon be factored out.) This second order differential equation admits Uθ as an integrating
factor, which means that by multiplying (20) by Uθ we can integrate it at once. This implies
that for any solution to (20) there must exist µ = µ(χ, τ ) such that
1
2
αk 2 Uθ2 = µ + λU + 12 cU 2 − 16 U 3 . (21)

with c = ω/k. We readily see

{ from the integrated
} form (21) of (20) that its solutions describe
curves in the phase plane (U, U̇ = kUθ ) of equation

1
2
αU̇ 2 + 16 U 3 − 12 cU 2 − λU = µ. (22)

We find thus find families of curves parametrized by µ at fixed (c, λ), see Figure 4.
In particular, when the parameters (c, λ) are such that the third order polynomial
1 3
6
U − 12 cU 2 − λU

11
Figure 4: Top: Plot of polynomial 16 U 3 − 21 cU 2 − λU (in blue) and some its value levels
(µ = −2/3, −0.5, −0.2, 0, 0.5). Bottom: Corresponding level curves of 12 U̇ 2 + 16 U 3 − 12 cU 2 −λU .
For c = 1 and λ = 0.

has a strict local minimum, say U0 = U0 (c, λ), we find closed curves around (U0 , 0) that
solve (22) for values of µ greater than and sufficiently close to the minimal value of that
polynomial.
These closed curves thus yield a family of periodic profiles U parametrized by (c, λ, µ). We
also see on Figure 4 a so-called homoclinic loop based at the point (0, 0), which corresponds
to a solitary wave profile. Solitary waves are traveling waves u = U (x − ct) of a special
kind, with U going exponentially fast to a same constant at both ±∞. Remarkably enough
the Korteweg–de Vries equation was precisely derived in the 19th century to explain the
occurrence of solitary waves on the surface of water9 . Solitary wave profiles may be viewed
as a limiting case of periodic profiles when their wavelength goes to infinity.
As to periodic profiles, they are the building blocks of modulated wave trains. Since they
have three degrees of freedoms here, namely the three parameters (c, λ, µ), we seek three
modulated equations associated with the Korteweg–de Vries equation (4).
One of the modulated equations comes somehow for free. This is the conservation of
waves, as in the previous section
kτ + ωχ = 0,
9
These were first reported on and called ’great waves of translation’ by the naval engineer John Scott
Russell [1808–1882].

12
or equivalently
kτ + (ck)χ = 0. (23)
Let us now explain where the other two modulated equations come from. They are based on
averaging in the variable θ.
By taking the average over one period in θ, we infer from (20) that

− c⟨U ⟩ + ⟨ 12 U 2 ⟩ = λ, (24)

where the brackets ⟨·⟩ stand for the average. This suggests that we try and find modulated
equations involving ⟨U ⟩ and ⟨U 2 /2⟩.
Using the chain rule as before and equating to zero the terms of order ε when we plug
(19) into (4) we obtain

Uτ + U Uχ + (αk 2 Uθθ )χ − ωVθ + k(U V + 2αkUθχ + αk 2 Vθθ )θ = 0.

Taking the average over one period we thus get the much simpler equation

⟨U ⟩τ + ⟨ 12 U 2 ⟩χ = 0. (25)

It remains to find an equation for ⟨U 2 /2⟩. This can be done by observing that any smooth
enough solution to (4) satisfies the additional conservation law10
( 1 2) (1 3 )
2
u t
+ 3
u + αuu xx − 1
2
αu 2
x x = 0. (26)

By the same kind of computation as for the derivation of (25), this yields the averaged
equation
⟨ 21 U 2 ⟩τ + ⟨ 13 U 3 + αk 2 U Uθθ − 12 αk 2 Uθ2 ⟩χ = 0.
By using (20) and (21), we can rewrite the latter in a simpler way as

⟨ 12 U 2 ⟩τ + ⟨ 12 cU 2 − µ⟩χ = 0. (27)

The system of modulated equations for the Korteweg–de Vries equation (4) is thus made
of the three equations (23), (25), and (27). It is not clear at first glance whether it is in closed
form, since it apparently involves more than three unknowns, namely c, µ, k, ⟨U ⟩, and ⟨ 12 U 2 ⟩.
However, thanks to (24) we may substitute λ for ⟨ 21 U 2 ⟩ as an unknown. In addition, as long
as periodic profiles U are properly parametrized by (c, λ, µ), their mean value ⟨U ⟩ and wave
number k are well defined in terms of (c, λ, µ). This implies that the system (23) (25) (27)
is in closed form in the ’variables’ (c, λ, µ).
Even though in closed form, the system (23) (25) (27) is not explicit enough. It is in
particular not obvious to check whether it is hyperbolic, a property that is necessary to at
least have local-in-time solutions for reasonably smooth initial data. It was an astonishing
achievement by Whitham not only to show that it is hyperbolic but also that this system
10
In fact, the Korteweg–de Vries has an infinity of conservation laws, as was discovered in the mid 1960s,
see for instance [12, Chap. 1] for a narrative of this series of discoveries.

13
admits a whole set of special quantities called Riemann invariants. By definition, for a first
order system of PDEs Riemann invariants are quantities that satisfy a nonlinear transport
equation in which the speed is an eigenvalue of the characteristic matrix - arising in the
system’s quasilinear form. Having a whole set of Riemann invariants is equivalent to being
able to write the system in diagonal form, up to a change of coordinates in the space of
unknowns.
Whitham figured out more precisely that some Riemann invariants for (23) (25) (27) just
read as the sums
v1 + v2 , v1 + v3 , v2 + v3
of the three roots v1 , v2 , and v3 of the polynomial
µ + λU + 12 cU 2 − 16 U 3
appearing in the right-hand side of (21) (these roots are just the abscissas of the intersection
points between the blue graph and the horizontal lines, corresponding to several values of µ,
on the upper picture on Figure 4).
It was understood in the 1980s that the existence of Riemann invariants for the modulated
equations associated with the Korteweg–de Vries equation is linked to deep algebraic proper-
ties, in connection with its infinite number of conservation laws. See for instance the survey
paper by Lax, Levermore and Venakides [24] for some explanation and further references.
As to the nonlinear Schrödinger equation (6), it actually admits periodic wave profiles
depending on four parameters. Therefore, the most general modulated equations for (6)
include four equations. It turns out that these modulated equations also admit a complete
set of Riemann invariants, see e.g. [23]. This is a very special feature linked to integrability
properties of (6) analogous to those of (4).
As said in [24], “Less is known about nonintegrable cases. While they sometimes share
properties exhibited by the integrable cases, new phenomena arise that have yet to be com-
pletely understood.” Almost a quarter of century later, this is still the case that we need to
understand the nonintegrable cases.
For general dispersive equations, modulated equations do not have any reason to admit
Riemann invariants. It is not even known whether they are hyperbolic. A large range of
numerical experiments has shown that this depends on the nonlinearities involved, and that
it seems to hardly ever be the case that for not completely integrable PDEs modulated
equations are hyperbolic in the whole range of parameters [25].

6 Modulational instability
At this point of the exposition, some questions about the properties of modulated equations
arise naturally.
Why is hyperbolicity of modulated equations so important, and what does it imply when
it fails?
In fact, the failure of hyperbolicity of modulated equations has been thought of from the
very beginning of the development of modulation theory as being somehow equivalent to

14
the instability of periodic traveling wave solutions. The story is in particular recounted by
Zakharov and Ostrovsky in [32].
The importance of the so-called modulation(al) instability and the underlying phenomena
is obvious when we search the Web, with more than 120 000 results on Google, 16 000 on
Google Scholar and 600 on arXiv (among which 435 posts since 2010) as of January 3rd,
2020.
The exact meaning of the term modulation(al) instability may vary according to authors
though. To make things clear, we use in what follows the term modulational instability as a
synonym, for a given wave, for the modulated equations to fail to be hyperbolic at the point
corresponding to this wave in the set of parameters. For instance, if we go back to § 4 and
periodic traveling wave solutions to the nonlinear Schrödinger equation (6) of the form
2 +αk 2 )t)
z = r ei(kx−(γr ,

such a wave is said to display modulational instability if the modulated system (16)-(17) is
not hyperbolic at point (k, r). In this case, this equivalently means that the characteristic
matrix ( )
2αk 2γr
M=
αr 2αk
has nonreal eigenvalues. For non zero r, this happens if and only if αγ is negative, irrespective
of the actual values of r and k. Thus in this special case, the modulational instability criterion
does not depend on the specific point considered in the set of wave parameters. In general
it does, even though convincing examples are out of the scope of the present paper - the
only other system of modulated presented here being the everywhere hyperbolic system
(23) (25) (27).
The most famous example of modulational instability, which both served to explain a
physical phenomenon and paved the way for further theory, is known as Benjamin–Feir
instability. It was shown indeed by Benjamin and Feir [1, 2] that for the periodic water waves
known as Stokes waves, over water of depth h modulational instability occurs when kh >
1.363. Roughly speaking, this means that waves of sufficiently small wavelength compared
to depth are modulationally unstable. This explained why Stokes waves are difficult to
reproduce in wave tanks such as the one pictured on Figure 5.
In practice, modulational instability is expected to imply the eventual breakup of the
waveform into a train of pulses. More precisely, modulational instability is linked to what is
called sideband instability. As said in the survey paper by Zakharov and Ostrovsky [32], “In
its simplistic version, the effect of modulation instability is the result of interaction between
a strong carrier harmonic wave at a frequency ω, and small sidebands ω ± Ω.”
It is only rather recently that rigorous results confirmed this point of view, in various
frameworks - starting with dissipative ones in [26, 27, 30], and going on with dispersive
PDEs in [22, 7, 21].
Before saying a bit more on these results, let us exemplify sideband instability by consid-
ering harmonic wave solutions to (6)

zr = r ei(kx−ωt) .

15
Figure 5: Model testing with periodic waves at the Wave/Tow Tank of the Jere A. Chase
Ocean Engineering Laboratory, University of New Hampshire.

Recall that the amplitude r, the wavenumber k and the time frequency ω must satisfy the
nonlinear dispersion relation in (7) for such a wave to solve (6). (The harmonic wave zr
thus also depends on k but for the sake of readability we refrain from introducing a double
subscript.)
Sideband instability is actually a specific case of linear instability. Linear stability or
instability have to do with perturbations zr + w of the reference solution zr that are approxi-
mate solutions up to O(w2 ) terms. As regards sideband instability, it is both more convenient
and more insightful to seek the perturbation with zr factorized, which amounts to looking
for an approximate solution of the form (1 + W )zr for small |W |.
Plugging z = (1 + W )zr in (6), using (7) and dropping quadratic terms in W or its
conjugate W ∗ , we receive the following linear equation

Wt + 2αkWx = iαWxx − iγr2 (W + W ∗ ). (28)

A first interesting remark is that the group velocity 2αk associated with zr shows up again,
namely in the transport operator ∂t + 2αk∂x in the left-hand side. We say that zr is linearly
unstable if there are solutions to (28) that are unbounded when t grows.
Let us in particular seek solutions to (28) of the form

W (x, t) = E(t) eiK(x−2αkt)

for real numbers K. The complex amplitude E = U + iV must satisfy

Et = −iαK 2 E − iγr2 (E + E ∗ ), (29)

which is equivalent to the system of real ordinary differential equations (ODE)

{
Ut = αK 2 V
Vt = −(αK 2 + 2γr2 )U,

The behavior of its solutions critically depends on the sign of

D := αK 2 (αK 2 + 2γr2 ).

16
More precisely, if D is positive, all solutions of (29) are bounded, but if D is negative the
system above admits exponentially growing solutions
( ) √
( 2
)
U t −D √αK
=e .
V −D

This implies that for negative D Equation (28) has solutions of the form
√
−D iK(x−2αkt)
W (x, t) = E0 et e , (30)

hence the linear instability of the wave zr .

When αγ is positive, D is positive as soon as K is non zero. The interesting case regarding
instability is thus αγ < 0. For in this case D < 0 for non zero wave numbers K such that

α2 K 2 < −2αγr2 .
√
In other words, if αγ is negative all non zero K such that |K| ∈ (0, r −2γ/α) trigger
instabilities of zr . This is why we speak of sideband instability, the perturbed approximate
solution
√ (1+W √ )zr involving the perturbed wave number k +K with non zero K in the ’band’
(−r −2γ/α, r −2γ/α).
Of course, the fact that W as given by (30) is exponentially growing with t implies that
(1 + W )zr cannot be a valid approximate solution to the nonlinear equation (6) for all times.
To say it more simply, linear instability does not necessarily imply nonlinear instability for
solutions of PDEs - unlike what happens for ODEs - but we are not going to dwell on this
topic here.
What we are going to explain is the link with modulational instability. This has √ to do
with small values of |K|, for√which we note that the growth rate in the exponential −D
is approximately equal to r −2αγ|K|. Thus for small |K| the total factor √ of t in (30) is
approximately equal to the complex number −iKC± where C± = 2αk ± ir −2αγ and the
± sign is that of K.
Interestingly enough, the values C± happen to be precisely the eigenvalues of the charac-
teristic matrix ( )
2αk 2γr
M= ,
αr 2αk
which are complex conjugate when αγ is negative.
This is the sought link between modulational instability and sideband instability. As a
matter of fact, the calculation above shows that the complex conjugate eigenvalues C± of M
are associated with approximate solutions of the original equation (6) of the form

zK (x, t) = r (1 + E0 ei(Kx−Ωt) )ei(kx−ωt)

√
with Ω = i −D so that

Ω ∼ C+ K when K ↘ 0 , Ω ∼ C− K when K ↗ 0 .

17
 In the physics literature the two notions of modulational instability and sideband insta-
bility are most often used as synonyms. Mathematically speaking there is actually a single
implication, which is not obvious at all to prove. This one says that modulational insta-
bility implies sideband instability in the same way as in the example here above. Namely,
whenever the characteristic matrix of the modulated equations has complex eigenvalues C± ,
with ± denoting the sign of their imaginary parts, the original equation linearized about the
underlying wave has solutions involving a growth factor of the form
ei(Kx−Ωt) with Ω ∼ C+ K when K ↘ 0 , Ω ∼ C− K when K ↗ 0 .
The actual proof of this result depends on the framework, and has been achieved rather
recently [7, 21, 22, 26, 27, 30].
Let us be a little more precise about the statement. For the sake of clarity, we stick to the
example of Equation (6). Given a reference wave solution zr , the linearized equation about
zr is obtained by looking for perturbed solutions zr + Z and by dropping quadratic terms in
Z or its conjugate Z ∗ . This yields the equation
Zt = iαZxx − 2iγ|zr |2 Z − iγzr2 Z ∗ . (31)

What has been found here above is that for a harmonic wave zr = r ei(kx−ωt) , in the case
αγ < 0, Equation (31) has solutions of the form
Z(x, t) = ei(Kx−Ωt) ei(kx−ωt) (32)

where Ω depends on K and is of positive imaginary part when |K| ↘ 0. This Z is just the
product W zr with W solution to (28). The advantage of (28) over (31) is merely that the
former has constant coefficients, so that we could more easily find W than directly Z.
Let us now reinterpret the existence of these solutions (32) to (31) in a more abstract
manner, in order to appreciate the analogous results proved in the aforementioned references.
We first invoke a usual trick when dealing with traveling waves. We make a change of
frame so that the wave of interest becomes steady. We set y = x − ct - recalling that ω = ck
- and consider
e t) := Z(x, t).
Z(y,
Then
Zt = Zet − cZey , Zx = Zey ,
so that Ze solves the new equation
Zet = iαZeyy + cZey − 2iγ|zer |2 Ze − iγ zer 2 Ze∗ (33)
with zer = r eiky . Furthermore, (32) reads in the new frame
e t) = eiK(y+(c−C)t) eiky ,
Z(y,
with C := Ω/K, so that (33) implies

iK(c − C)Ze = MZ,

e

18
where M is the differential operator in the y variable appearing in the right-hand side of
(33). To be more specific, M is the operator associating to any smooth enough complex
valued function Y of y the new function

MY := iαYyy + cYy − 2iγ|zer |2 Y − iγ zer 2 Y ∗ .

The equality deduced from (33) is true in particular at t = 0 so that we have

iK(c − C)Y = MY

for Y (y) := ei(k+K)y . At this stage it is tempting to say that λ := iK(c − C) is an eigenvalue
of M associated with the direction Y . However, the notion of eigenvalues is more subtle for
differential operators than for matrices. It depends on the chosen functional setting.
Recalling that both k and K are real numbers, we may say indeed that λ is an eigenvalue
of M in the space of C ∞ functions that are bounded as well as all their derivatives, since Y
pertains to this class of functions. Moreover, recalling that c is real and KC is of positive
imaginary part for small |K|, we infer that λ is of positive real part for such K. The operator
M having an eigenvalue of positive real part implies what is called spectral instability 11 for
the wave zr .
However, we would still like to say a little bit more, which will lead us back to the notion
of sideband instability. For K = 0 the function Y reduces to y 7→ eiky , which is periodic of
period ℓ := 2π/k. This period ℓ happens to be - and this is no chance - the wavelength of
the original wave zr , and it is also the period of zer . In other words, the differential operator
M has periodic coefficients of period ℓ.
For general K we have Y (ℓ) = eiℓK = eiℓK Y (0). Since the operator M has ℓ-periodic
coefficients, this implies together with the fact that λY = MY that λ belongs to the spectrum
of M on the space of square integrable functions. It would lead us too far to give a precise
meaning to this statement, not even speaking of proving it - the interested reader may refer
for instance to [28, Chap. I] and [29, Theorem XIII.89].
Let us just mention that in this framework K can be called a Floquet exponent. Since λ
is of positive real part for small |K|, we can summarize the meaning of the existence of Y by
saying that the operator M has unstable spectrum associated with small Floquet exponents.
This can be considered as a mathematical definition for sideband instability.
Now we are ready to go back to the rigorous link between modulational instability and
sideband instability. It has been proved indeed in the various frameworks quoted above
[7, 21, 22, 26, 27, 30] that, as for our chosen simple example, modulational instability due to
complex eigenvalues C± of the modulated equations implies sideband instability due to some
unstable spectrum
λ = iK(c − C± ) + O(K 2 )
associated with small Floquet exponent K.
11
The word ’spectral’ comes from the notion of spectrum, which contains eigenvalues and possibly other
complex numbers for differential operators, see for instance for definitions [28, Chap. I].

19
7 Modulation theory, fifty years on and beyond
Since its inception modulation theory has been investigated in quite different ways by math-
ematicians and mathematical physicists, for various types of PDEs.
The earliest and most complete theory is for the Korteweg–de Vries equation (4). Thanks
to deep geometric properties of this equation that cannot be described here (see [16]) both
Question 3 and Question 4 were answered through a series of work in the 1980s-1990s (see
[24], references therein and [10] as well), in connection with the related topic of zero dispersion
limit.
The link between modulation theory and the zero dispersion limit comes from a rescaling
argument, a function u being solution to (4) if and only if the rescaled function u e(χ, τ ) =
2
u(χ/ε, τ /ε) solves the Korteweg–de Vries equation with parameter ε α.
Formally, Equation (4) goes to the nonlinear transport equation (5) when α goes to zero.
This is a singular limit though, the nature of the two PDEs (4) and (5) being very different.
One of the important features of (5) is that it admits shock wave solutions. A shock wave
is a discontinuous traveling wave u = U (x − ct) whose profile U is a step function. Shock
waves are to be sought as weak solutions to (5), defined as satisfying the integral equation
∫
2u dx − u2 dt = 0 (34)
R

for all rectangle R of the space-time plane, Equation (34) being formally obtained from (5)
by Green’s theorem. By considering a rectangle R = [0, T /c] × [0, T ], we find that for the
profile {
u− , x < 0,
U (x) =
u+ , x > 0,
to yield a weak solution to (5), (u− , u+ , c) must satisfy the Rankine–Hugoniot condition

2c = u− + u+ .

A natural question is whether shock wave solutions to (5) can be achieved as limits of solutions
to (4) when α goes to zero. The answer is no, and this can be explained through the concept
of dispersive shocks.
Dispersive shocks can be loosely defined as unsteady patterns

• that are close to the underlying shock wave for large |x|,

• that are oscillatory in a region that expands proportionally to time t.

Dispersive shocks are thus more complicated wave trains than traveling waves. A concrete
example of such waves is given by tidal bores. These are known to happen on dozens of rivers
in the world and can be fun for surfers, but can also dangerous when they are too strong
- as for instance on the Qiantang River in China. For the Korteweg–de Vries equation it has
been shown that such wave trains do exist and that they are well approximated by the wave
trains based on modulated equations constructed by Gurevich and Pitaevskii [17].

20
 Figure 6: From [17], plots at two different times of a dispersive shock.

The oscillations in these wave trains have shorter and shorter wavelengths when α goes
to zero, in such a way that they eventually fill a whole region within an envelope that can
be computed from the modulated equations. It can be shown that the weak limit 12 of a
dispersive shock is not a mere step function, and that is not a solution of (5). Fore more
details see for instance [15, Section 3.1] and references therein.
If modulation theory is well understood for the Korteweg–de Vries equation, which has
received a lot of attention in the last decades, this is not so for other types of PDEs. Even
for the nonlinear Schrödinger equation, which has also been receiving a lot of attention,
modulation theory has not completely been clarified yet. It is of course also linked to the
zero dispersion limit (α → 0 in (6)), better called semiclassical limit in that framework. An
important contribution regarding the semiclassical limit in the defocusing case (αγ > 0) was
brought in the 1990s by Jin, Levermore and Mc Laughlin [19], but it does not really make
modulation theory a rigorous way of characterizing wave train solutions to (6). One rigorous
result was obtained regarding modulation of harmonic waves - corresponding to the second
example introduced in Section 4 - by Düll and Schneider in 2009 [13]. Irrespective of the sign
of αγ they proved ”that slow modulations in time and space of periodic wave trains of the NLS
[nonlinear Schrödinger] equation can be approximated via solutions of Whithams equations
associated with the wave train”. Their approach is based on a change of variables and the
use of the Cauchy-Kowalewskaya theorem 13 to circumvent the ill-posedness of modulated
equations in the case αγ < 0.
Remarkably enough, both the Korteweg–de Vries equation and the nonlinear Schrödinger
equation themselves can be viewed as modulation equations for the water wave equations 14 .
A rigorous justification was achieved in particular by Craig, Sulem and Sulem [8], who proved
that modulations of harmonic waves as governed by the nonlinear Schrödinger equation do
yield approximate solutions of the water wave equations.
A thorough study of modulated equations has been undertaken in a series of work [5, 6, 7]
for a large class of nonlinear dispersive PDEs that includes the Korteweg–de Vries equation, a
fluid formulation of the nonlinear Schrödinger equation, and more generally Euler–Korteweg
equations, which are dispersive modifications of the usual equations for compressible fluids
that are used in various fields of mathematical physics. It is in particular motivated by the
study of dispersive shock waves in non integrable systems, as initiated by El in [14]. It has
12
We say that u is the weak limit of a sequence
∫ of functions
∫ (un ) if for any infinitely smooth function φ
that vanishes outside a finite interval we have un φ → uφ.
13
This theorem shows the existence of analytic solutions for general classes of PDEs with analytic data.
14
The water wave equations consist of a free interface version of the Euler equations for incompressible
fluids.

21
already turned out that the Gurevich-Pitaevskii problem is in general much more singular
than expected.
Another field of research in connection with modulation theory is the stability of periodic
traveling waves with respect to ’localized’ perturbations. If we think of a water wave, a
localized perturbation could be induced for instance by a pebble thrown into the water. As
regards dispersive models, the stability of periodic traveling waves with respect to ’localized’
perturbations is a widely open problem15 , even though we can find numerical evidence -
by computing approximate solutions through carefully chosen algorithms - that modulation
equations do play a key role in describing the perturbed solutions.
Some breakthroughs have been achieved rather recently on this topic for dissipative mod-
els. This started with the work by Doelman, Sandstede, Scheel, and Schneider [11], who
addressed the case of reaction-diffusion equations 16 , and dealt in particular with the first ex-
ample described in Section 4. Then Johnson, Noble, Rodrigues, and Zumbrun [20] managed
to develop a similar theory for systems of viscous conservation laws, which are generalized
versions of the Burgers equation arising for instance in the modeling of viscous compressible
fluids. Their analysis is in particular based on higher order modulated equations, like (12)
for the Ginzburg-Landau equation (8). The counterpart of these higher order modulated
equations for dispersive PDEs like (6) would be third order systems of PDEs instead of just
first order systems like (18).
To conclude this overview, let us stress that all the aforementioned work is basically
dealing with waves in one space dimension. Of course most waves in practical applications
are not unidimensional, as we can see for instance on the picture of a tidal bore on Figure 7.

Figure 7: Tidal bore in Alaska.

Research is going on. In particular, a semester will be devoted to ’Dispersive hydrody-

namics: mathematics, simulation and experiments, with applications in nonlinear waves’ at
the Isaac Newton Institute (Cambridge, UK) - originally scheduled for 2020, this thematic
semester has unfortunately been postponed to 2022 due to the health crisis.
15
By contrast, numerous results have been obtained regarding the stability of periodic traveling waves with
respect to perturbations of the same period, see [3] and references therein.
16
A reaction-diffusion equation is a ’mixture’ of the ordinary differential equation ut = f (u) (the reaction
part) and of the heat equation ut = αuxx (with α > 0 being linked to heat conductivity), and thus reads
ut = αuxx + f (u). Such equations are widely used in mathematical biology.

22
Acknowledgment. L. Miguel Rodrigues is warmly thanked for sharing his insight, for his
careful reading of successive versions of the manuscript, and for all the great discussions we
have been having on this topic.

References
[1] T.B. Benjamin. Instability of periodic wavetrains in nonlinear dispersive systems. Proc.
Roy. Soc. A. 299 (1967) 5975.
[2] T. B. Benjamin, J. E. Feir. The disintegration of wave trains on deep water. Part 1.
Theory. J. Fluid Mech. 27 (1967) 417430.
[3] S. Benzoni-Gavage, C. Mietka, and L. M. Rodrigues. Co-periodic stability of periodic
waves in some hamiltonian PDEs. Nonlinearity, 29(11):3241, 2016.
[4] S. Benzoni-Gavage, C. Mietka, and L.M. Rodrigues. Stability of periodic waves in hamil-
tonian PDEs of either long wavelength or small amplitude. Indiana Univ Math J., To
appear.
[5] S. Benzoni-Gavage, C. Mietka, and L.M. Rodrigues. Modulated equations of Hamiltonian
PDEs and dispersive shocks. arXiv 1911.10067, 2019.
[6] S. Benzoni-Gavage, P. Noble, and L.M. Rodrigues. Stability of periodic waves in Hamil-
tonian PDEs. In GDR Analyse des EDP. cedram, 2013.
[7] S. Benzoni-Gavage, P. Noble, and L. M. Rodrigues. Slow modulations of periodic waves in
Hamiltonian PDEs, with application to capillary fluids. J. Nonlinear Sci., 24(4):711–768,
2014.
[8] W. Craig, C. Sulem, P.-L. Sulem. Nonlinear modulation of gravity waves: a rigorous
approach. Nonlinearity, 5(2):497–522, 1992.
[9] E.M. de Jager. On the origin of the Korteweg–de Vries equation. Forum der Berliner
Mathematischen Gesellschaft, Band 19, Dezember 2011, pp. 171-195.
[10] P. Deift, S. Venakides, and X. Zhou. New results in small dispersion KdV by an extension
of the steepest descent method for Riemann-Hilbert problems. Internat. Math. Res.
Notices, (6):286–299, 1997.
[11] A. Doelman, B. Sandstede, A. Scheel, and G. Schneider. The dynamics of modulated
wave trains. Mem. Amer. Math. Soc., 199(934):viii+105, 2009.
[12] P.G. Drazin and R.S. Johnson. Solitons: an introduction. Cambridge University Press,
1996.
[13] W.-P. Düll, and G. Schneider. Validity of Whithams Equations for the Modulation of
Periodic Traveling Waves in the NLS Equation. J Nonlinear Sci (2009) 19: 453466.

23
[14] G. A. El. Resolution of a shock in hyperbolic systems modified by weak dispersion.
Chaos, 15(3):037103, 21, 2005.

[15] G. A. El. and M. A. Hoefer. Dispersive shock waves and modulation theory. Physica D,
333:11–65, 2016.

[16] H. Flaschka, M. G. Forest, and D. W. McLaughlin. Multiphase averaging and the

inverse spectral solution of the Korteweg-de Vries equation. Comm. Pure Appl. Math.,
33(6):739–784, 1980.

[17] A. V. Gurevich and L. P. Pitaevskii. Nonstationary structure of a collisionless shock

wave. Zhurnal Eksperimentalnoi i Teoreticheskoi Fiziki, 65:590–604, August 1973.

[18] M. A. Hoefer. Shock waves in dispersive Eulerian fluids. J. Nonlinear Sci., 24(3):525–
577, 2014.

[19] S. Jin, C. D. Levermore, and D. W. McLaughlin. The semiclassical limit of the defocusing
NLS hierarchy. Comm. Pure Appl. Math., 52(5):613–654, 1999.

[20] M. A. Johnson, P. Noble, L .M. Rodrigues, and K. Zumbrun. Behavior of periodic

solutions of viscous conservation laws under localized and nonlocalized perturbations.
Inventiones mathematicae, pages 1–99, 2013.

[21] M. A. Johnson and W. R. Perkins. Modulational Instability of Viscous Fluid Conduit

Periodic Waves. arXiv 1904.06348, 2019.

[22] M. A. Johnson, K. Zumbrun, and J. C. Bronski. On the modulation equations and

stability of periodic generalized Korteweg-de Vries waves via Bloch decompositions. Phys.
D, 239(23-24):2057–2065, 2010.

[23] A. M. Kamchatnov. Nonlinear periodic waves and their modulations. World Scientific
Publishing Co., Inc., River Edge, NJ, 2000. An introductory course.

[24] P. D. Lax, C. D. Levermore, and S. Venakides. The generation and propagation of

oscillations in dispersive initial value problems and their limiting behavior. In Important
developments in soliton theory, 205241, Springer Ser. Nonlinear Dynam., Springer, Berlin,
1993.

[25] C. Mietka. Periodic waves in some Hamiltonian PDEs : stability, modulations and
dispersive shocks. Theses, Université de Lyon, February 2017.

[26] M. Oh and K. Zumbrun. Stability of periodic solutions of conservation laws with vis-
cosity: analysis of the Evans function. Arch. Ration. Mech. Anal., 166(2):99–166, 2003.

[27] M. Oh and K. Zumbrun. Stability of periodic solutions of conservation laws with viscos-
ity: pointwise bounds on the Green function. Arch. Ration. Mech. Anal., 166(2):167–196,
2003.

24
[28] M. Reed and B. Simon. Methods of modern mathematical physics. I Functional analysis.
Academic Press, 1972.

[29] M. Reed and B. Simon. Methods of modern mathematical physics. IV Analysis of

operators. Academic Press, 1978.

[30] D. Serre. Spectral stability of periodic solutions of viscous conservation laws: large
wavelength analysis. Comm. Partial Differential Equations, 30(1-3):259–282, 2005.

[31] G. B. Whitham. Linear and nonlinear waves. John Wiley & Sons Inc., New York, 1999.
Reprint of the 1974 original, A Wiley-Interscience Publication.

[32] V. E. Zakharov, L. A. Ostrovsky. Modulation instability: The beginning. Physica D

238 (2009) 540548.

25