On the Initial-Boundary Value Problem for the Linearized

Boussinesq Equation

By A. Alexandrou Himonas and Dionyssios Mantzavinos

This work is concerned with the initial-boundary value problem for the
Boussinesq equation. By employing the unified transform method of Fokas,
novel solution formulae for the linearized “good” Boussinesq equation on the
half-line with various initial and boundary conditions are obtained. Moreover,
these solution formulae are numerically illustrated in the case of concrete
data. Finally, Boussinesq’s original physical derivation of the so-called “bad”
Boussinesq equation is provided.

1. Introduction

John Scott Russell was a gifted British experimentalist of the nineteenth

century. Although he began his career as a promising academic in natural
sciences—he had assumed a temporary position as a lecturer in the University of
Edinburgh at the young age of 24 years—he eventually focused on applications
of mathematics and science to engineering. In particular, he became interested
in steam engines and even developed his own steam carriage. He also carried
out a series of experiments to determine the optimal design for canal boats.
In 1834, while conducting one of these experiments, he was following a
barge that was being pulled by two horses along a narrow part of the Union
Canal at Edinburgh. Suddenly, the barge stopped moving, causing the formation
of a large solitary wave, which started propagating fast down the canal. Scott
Russell followed this wave riding on horseback, and was impressed by the fact

Address for correspondence: Dionyssios Mantzavinos, Department of Mathematics, University of Notre

Dame, Notre Dame, IN 46556; e-mail: mantzavinos.1@nd.edu

DOI: 10.1111/sapm.12055 62
STUDIES IN APPLIED MATHEMATICS 134:62–100

C 2014 Wiley Periodicals, Inc., A Wiley Company
 On the Initial-Boundary Value Problem for the Linearized Boussinesq Equation 63

that it preserved its initial speed and structure with almost no dissipation for
about two miles, before it got lost in the windings of the canal.
Truly amazed by this phenomenon, which he called the “wave of translation,”
Scott Russell performed further experiments with the help of a wave tank
that he built in his backyard, and was able to create waves that were
consistent with his observations, that is, they were stable and their speed
seemed to be proportional to their height. In 1844, he reported the wave of
translation during the meeting of the British association for the advancement of
science, challenging the scientific community to come up with a mathematical
explanation of the phenomenon. Most of the prominent mathematicians of the
day, including Airy and Stokes, remained skeptical regarding Russell’s remarks
because these were contradicting some earlier works in water wave theory.
A few decades later, however, the French mathematician and physicist
Joseph Boussinesq was the first to provide a mathematical theory in support to
Russell’s observations. Indeed, he derived in Refs. [1, 2] several approximations
to Euler’s equations that admit traveling wave solutions of the type that Russell
had described. Actually, the plethora of choices for the physical variables that
dominate the phenomenon led Boussinesq to a whole set of perturbations of the
linear wave equation which take into account the effects of weak nonlinearity
and dispersion, see discussion in Bona & Smith [4]. In fact, the celebrated KdV
(Korteweg-de Vries) equation was first derived by Boussinesq and appears in
his 700-pages-long essay [3] that presents his theory on water waves. This
essay was published in 1877, nearly two decades before the 1895 work of
Korteweg and de Vries.
A particular case that was considered by Boussinesq in Ref. 1 refers to the
situation where the depth of the channel is small compared to the length of the
wave. Approximating Euler’s equations in this case led him to the equation

u tt − u x x − u x x x x − (u 2 )x x = 0, x ∈ R, t > 0, (1)

which models the propagation of small amplitude, weakly nonlinear, long water
waves, and more importantly, possesses solutions that agree with Russell’s
observations.
Indeed, seeking a traveling wave solution of the form u(x, t) = f (ξ ), where
ξ = x − ct and c is the constant speed of propagation of the wave, we arrive at
the ODE
c2 f  − f  − f  − ( f 2 ) = 0.
Integrating this ODE under the assumption that f and its derivatives tend to
zero as |ξ | → ∞ eventually gives
√ 
3 2 c 2−1
u(x, t) = (c − 1) sech2 (x − ct) + a , a = const. (2)
2 2
64 A. A. Himonas and D. Mantzavinos

This solution describes a wave that travels at a constant speed c, with amplitude
independent of x and t and positively correlated to c. Hence, (2) is a wave of
translation in accordance with Russell’s description. Such waves are nowadays
known as solitons.
Although with a solid physical interpretation, Equation (1) has a major
defect: it is badly ill-posed. This can be seen immediately by seeking small
amplitude solutions of the form
u(x, t) = εe−ikx+iωt , ε  1, x ∈ R, t > 0, k ∈ R, (3)

which make the nonlinear term of (1) negligible. The dispersion relation for
these solutions is ω2 = k 2 − k 4 . Hence, for |k| > 1 the time frequency of (3) is
2
imaginary and we get exponential growth at a rate of about ek t . This explains
why Equation (1) is known as the “bad” Boussinesq (BB) equation.
There are several ways to correct the ill-posed behavior of the BB equation.
These include replacing the fourth order x-derivative by a mixed derivative
u x xtt , which changes the dispersion relation to ω2 = k 2 /(1 + k 2 ) so that |ω| is
bounded, or simply flipping the sign of the fourth order term to obtain the
equation
u tt − u x x + u x x x x − (u 2 )x x = 0, x ∈ R, t > 0, (4)

with dispersion relation ω2 = k 2 + k 4 . The frequency ω is now real and

Equation (4) is well-posed in the case of small amplitude solutions, thus it is
known as the “good” Boussinesq (GB) equation. What is more, using the
transformation u → −u the GB equation assumes the form
u tt − u x x + u x x x x + (u 2 )x x = 0, x ∈ R, t > 0, (5)
which is the one that usually appears in the literature.
Bona & Sachs [5] showed local well-posedness of the following generalized
GB initial-value problem (ivp):
u tt − u x x + u x x x x + ( f (u))x x = 0, x ∈ R, t > 0, f ∈ C ∞, (6a)
u(x, 0) = u 0 (x) ∈ H s+2 (R), u t (x, 0) = u 1 (x) ∈ H s (R), s > 1/2. (6b)
Note that the GB corresponds to f (u) = u 2 . The case of f (u) = |u| p−1 u,
p > 1, was studied by Tsutsumi & Matahashi in Ref. 14, where
local well-posedness was established for (u 0 , u 1 ) ∈ H 1 (R) × H −1 (R). For
1 < p < 5, Linares [13] showed well-posedness for rougher data, namely
for (u 0 , u 1 ) ∈ L 2 (R) × H −2 (R), via Strichartz estimates. Also, Farah [6] and
Farah & Scialom [7] established well-posedness of the GB ivp on the line
and on the circle, respectively, for data (u 0 , u 1 ) ∈ H s × H s−2 and s > −1/4.
Furthermore, sharper well-posedness results have been proved by Kishimoto &
Tsugawa [12] and Kishimoto [11]. Finally, ill-posedness results for generalized
Boussinesq equations can be found in Geba et al. [10].
 On the Initial-Boundary Value Problem for the Linearized Boussinesq Equation 65

The aforementioned results indicate that the Cauchy problem for the GB
equation has been studied extensively by several researchers. This is not the
case, however, concerning the initial-boundary value problem (ibvp) for this
equation, where significantly less progress has been made. The most recent
result in this direction was obtained by Xue [15], who showed that the GB ibvp
u tt − u x x + u x x x x + (|u| p u)x x = 0, 0 < x < ∞, t > 0, 0 < p < 4, (7a)

u(x, 0) = u 0 (x), u t (x, 0) = ∂x u 1 (x), (7b)

u(0, t) = h 0 (t), u x (0, t) = h 1 (t) (7c)

1
is well-posed for (u 0 , u 1 ) ∈ L 2 (R+ ) × H0−1 (R+ ) and (h 0 , h 1 ) ∈ H04 (R+ ) ×
− 12 ,− 14
H0 (R+ ), where
Hr,s (R)  { f : |ξ |r (1 + |ξ |)s−r f̂ (ξ ) ∈ L 2 (R)}.
The analysis in Ref. [15] is based on the solution via Laplace transform of the
corresponding ibvp for the linearized version of Equation (7a) with forcing,
namely for the equation
u tt − u x x + u x x x x = f, f = f (x, t). (8)
After the relevant linear estimates are obtained, the forcing f is replaced by the
nonlinearity (|u| p u)x x , and well-posedness of the nonlinear ibvp is established
for data as specified above by applying a contraction mapping argument in
suitably chosen Bourgain-type spaces. Note that Equation (8) is nothing but
the linear part of the GB Equation (5) with the nonlinearity replaced by the
forcing f and, therefore, we hereafter refer to it as the forced linearized GB
(FLGB) equation.
Our aim is to initiate a project on the study of ibvps for the GB equation,
following the standard contraction mapping argument that was also employed
in Ref. [15]. For this purpose, we derive in this work alternative solution
formulae for the FLGB equation posed on the half-line with various types
of boundary conditions, including the ones of the ibvp (7). These formulae
involve integrals along contours that lie in the complex spectral plane and
hence, the relevant integrands decay exponentially. This feature of the novel
formulae could be an advantage regarding linear estimates, on top of a clear
advantage when it comes to numerical evaluations.

1.1. The unified transform method

Our approach to the FLGB ibvp on the half-line uses the so-called unified
transform method, which was introduced in 1997 by Fokas [8, 9]. Although
motivated by ibvps for integrable nonlinear equations, Fokas’ new method
also applies to linear ibvps; in particular, it can be employed for the solution
66 A. A. Himonas and D. Mantzavinos

of linear evolution equations of any spatial order, formulated either on the

half-line or on the finite interval.
For those problems which can be solved via the traditional techniques (e.g.,
classical transform method, Green’s functions, etc.), the unified transform
method yields novel solution representations which involve integrals along
contours in the complex spectral plane. If desired, these representations can be
reduced to the classical ones via Cauchy’s theorem.
It should be noted, however, that for evolution equations of spatial order
higher than three the classical transform method in the spatial variable is
not applicable, while the Laplace transform in the temporal variable usually
requires solving a complicated indicial equation. On the other hand, as it will
be made evident in Section 2, the unified transform method applies to higher
order problems in the same way as to lower order ones. Moreover, it can
accommodate boundary conditions of various types, like Dirichlet, Neumann,
Robin, and more.
For the FLGB equation on the half-line, the unified transform method
involves the following steps.

1. Expression of the equation in divergence forms. Rewrite the FLGB

Equation (8) in the divergence forms
 −ikx+iωt 
e (u t − iωu) t
  
+ e−ikx+iωt u x x x + iku x x − (1 + k 2 )u x − ik(1 + k 2 )u x
= e−ikx+iωt f, k ∈ C, (9)

and
 −ikx−iωt 
e (u t + iωu) t
  
+ e−ikx−iωt u x x x + iku x x − (1 + k 2 )u x − ik(1 + k 2 )u x
= e−ikx−iωt f, k ∈ C, (10)
.  1
where ω = k 2 + k 4 2 .
2. Integration of the divergence forms. Define the half-line Fourier transform
of u in the spatial variable by
 ∞
.
û(k, t) = e−ikx u(x, t)d x, k ∈ C− = {k ∈ C : Im(k) ≤ 0} . (11)
0

Note that, in contrast to the full line, the half-line Fourier transform (11)
is valid in all of the lower half of the complex spectral plane and not
just for k ∈ R. Then, integrate the divergence forms of Step 1 over the
given physical domain to obtain the so-called global relation, which is
 On the Initial-Boundary Value Problem for the Linearized Boussinesq Equation 67

an equation coupling the Fourier transform û(k, t) of the solution with

appropriate transforms of the initial data and of the boundary values.
3. Escape to the complex k-plane. Solve the global relation for û(k, t)
and use the inverse Fourier transform to express u(x, t) in terms of
transforms of the initial data and of the boundary values. Subsequently,
deform the contour of integration with respect to k from the real line
to appropriate contours in the complex k-plane to obtain an integral
representation of u(x, t). This representation is not yet effective because
it involves transforms of unknown boundary values.
4. Elimination of the unknowns. Use the global relation, as well as three
more equations which are obtained from the global relation under
appropriate symmetry transformations, to eliminate the transforms of
the unknown boundary values.

Note that Step 4 turns the integral representation of Step 3 turns into an
effective solution formula, in the sense that it only involves transforms of the
given data. In this connection, we emphasize that the main novelty of the unified
transform method in the case of linear equations is the exploitation, through
the deformations of Step 3, of the complex k-plane. Without performing
these deformations, the symmetries used in Step 4 for the elimination of the
unknowns from the integral representation could not have been employed.

1.2. Organization of the paper

Novel explicit solution formulae for the FLGB equation on the half-line are
produced in Section 2. Section 3 contains certain numerical evaluations using
MATHEMATICA. A direct verification of the solution formulae obtained in Section
2 is provided in Appendix D. Finally, the derivation of the “bad” Boussinesq
equation according to Boussinesq himself is presented in Appendix B.

2. Explicit solution formulae via the unified transform method

Consider the linearized “good” Boussinesq equation with forcing (FLGB)

formulated on the half-line:
u tt − u x x + u x x x x = f, 0 < x < ∞, t > 0, u = u(x, t), f = f (x, t). (12a)
For a well-posed ibvp, one should prescribe two initial conditions, which we
denote by
u(x, 0) = u 0 (x), u t (x, 0) = u 1 (x), 0 < x < ∞, (12b)
together with any two of the boundary values
∂xj u(0, t), t > 0, j = 0, 1, 2, 3, (12c)
68 A. A. Himonas and D. Mantzavinos

k k

i
ϕ2

−i ϕ1

Figure 1. The branch cut for ω.

as boundary conditions at x = 0. At the moment we do not specify which

two of the above boundary values are given, as the unified transform method
applies in the same way for any boundary conditions. Later, however, we
examine two particular cases that are common in the literature.
We now employ Steps 1–4 of the unified transform method, as described in
the Introduction.

1. Expression of the equation in divergence forms. We couple Equation

(12a) with the homogeneous version of its formal adjoint equation,
namely
ũ tt − ũ x x + ũ x x x x = 0. (13)
Multiplying (12a) by ũ t , (13) by u t , and subtracting the resulting
equations yields
(ũu t − u ũ t )t + (ũu x x x − ũ x u x x + ũ x x u x − ũ x x x u − ũu x + ũ x u)x
= ũ f. (14)
Equation (13) admits the one-parameter families of solutions ũ + and ũ −
given by
ũ ± (x, t) = e−ikx±iωt , k ∈ C, (15)
where
1
ω = ω(k)  (k 2 + k 4 ) 2 . (16)
Introducing the branch cut shown in Figure 1, so that
 1
1 + k 2 2 = |1 + k 2 | ei(ϕ1 +ϕ2 −π)/2 , (17)
On the Initial-Boundary Value Problem for the Linearized Boussinesq Equation 69

ϕ1 , ϕ2 ∈ [0, 2π ),
we may write ω in the form

ω = k |1 + k 2 | ei(ϕ1 +ϕ2 −π)/2 , (18)
ϕ1 , ϕ2 ∈ [0, 2π ).
Note that, according to the above choice of branch cut, for large |k| we
have
1 1  
ω = k 2 + − 2 + O k −4 . (19)
2 8k
Inserting the particular solutions (15) into Equation (14), we find the
following divergence forms for FLGB:
 −ikx+iωt 

e (u t − iωu) t + e−ikx+iωt u x x x + iku x x

− (1 + k 2 )u x − ik(1 + k 2 )u = e−ikx+iωt f, k ∈ C, (20)
x
and
 

e−ikx−iωt (u t + iωu) t + e−ikx−iωt u x x x + iku x x

−(1 + k 2 )u x − ik(1 + k 2 )u
x
−ikx−iωt
=e f, k ∈ C. (21)
2. Integration of the divergence forms. Define the spatial Fourier transform
pair on the half-line by
 ∞
û(k, t) = e−ikx u(x, t)d x, k ∈ C−  {k ∈ C : Im(k) ≤ 0}, (22a)
0

 ∞
1
u(x, t) = eikx û(k, t)dk, 0 < x < ∞. (22b)
2π −∞

Note that, because 0 < x < ∞, the spectral variable k can reside in C−
and not just in R. Taking the Fourier transform of the divergence forms
(20) and (21) yields
 iωt  
e [û t (k, t) − iωû(k, t)] t = eiωt u x x x (0, t) + iku x x (0, t)
− (1 + k 2 )u x (0, t) − ik(1 + k 2 )u(0, t)

+ fˆ(k, t)
and
 −iωt  
e (û t (k, t) + iωû(k, t)] t = e−iωt u x x x (0, t) + iku x x (0, t)
70 A. A. Himonas and D. Mantzavinos

− (1 + k 2 )u x (0, t) − ik(1 + k 2 )u(0, t)


+ fˆ(k, t) .

Integrating the above equations with respect to t, we respectively

obtain


−iωt
  t
û t (k, t) − iωû(k, t) = e û 1 (k) − iωû 0 (k) + eiω(τ −t) fˆ(k, τ )dτ
0

+ e−iωt g3+ (ω, t) + ikg2+ (ω, t) − (1 + k 2 )g1+ (ω, t)

− ik(1 + k 2 )g0+ (ω, t) , k ∈ C− , (23a)

and

  t
û t (k, t) + iωû(k, t) = e iωt
û 1 (k) + iωû 0 (k) + e−iω(τ −t) fˆ(k, τ )dτ
0

+ eiωt g3− (ω, t) + ikg2− (ω, t) − (1 + k 2 )g1− (ω, t)


− ik(1 + k 2 )g0− (ω, t) , k ∈ C− , (23b)

where the functions g ±j (ω, t) are defined by

 t
g ±j (ω, t)  e±iωτ ∂xj u(0, τ )dτ, j = 0, 1, 2, 3, (24)
0

and depend on the spectral variable k only through ω (see Equation (16)).
Subtracting (23a) from (23b) gives the global relation



2iω û(k, t) = eiωt û 1 (k) + iωû 0 (k) − e−iωt û 1 (k) − iωû 0 (k) (25)
 t  t
−iω(τ −t)
+ e fˆ(k, τ )dτ − eiω(τ −t) fˆ(k, τ )dτ
0 0

− e−iωt g3+ (ω, t) + ikg2+ (ω, t) − (1 + k 2 )g1+ (ω, t)


− ik(1 + k 2 )g0+ (ω, t)

+eiωt g3− (ω, t) + ikg2− (ω, t) − (1 + k 2 )g1− (ω, t)


−ik(1 + k 2
)g0− (ω, t) , k ∈ C− ,
On the Initial-Boundary Value Problem for the Linearized Boussinesq Equation 71

which expresses the Fourier transform û(k, t) of the solution in terms of

the Fourier transforms û 0 (k), û 1 (k), fˆ(k, t) of the initial data and the
forcing, as well as of the functions g ±j (ω, t).
3. Escape to the complex k-plane. For k ∈ R, it is possible to invert the
global relation by means of the inverse Fourier transform (22b) to
obtain
 ∞ ikx 
1 e  
u(x, t) = eiωt û 1 (k) + iωû 0 (k)
2π −∞ 2iω
 
− e−iωt û 1 (k) − iωû 0 (k) dk
 ∞ ikx   t
1 e
+ e iωt
e−iωτ fˆ(k, τ )dτ
2π −∞ 2iω 0
 t 
−iωt
− e e iωτ ˆ
f (k, τ )dτ dk
0

1 eikx  iωt  −
∞
+ e g3 (ω, t) + ikg2− (ω, t)−(1 + k 2 )g1− (ω, t)
2π−∞ 2iω

− ik(1 + k 2 )g0− (ω, t)

−e−iωt g3+ (ω, t) + ikg2+ (ω, t) − (1 + k 2 )g1+ (ω, t)

−ik(1 + k 2 )g0+ (ω, t) dk. (26)

The first two terms in the above expression depend only on the given
initial conditions and forcing, while the third term involves all four
boundary values ∂ j u(0, t), j = 0, 1, 2, 3, through the functions g ±j (ω, t).
For a well-posed ibvp, however, only two of these boundary values can
be specified as boundary conditions. Thus, Equation (26) contains two
unknowns that have to be eliminated.

To achieve this, we begin by deforming the contour of integration with

respect to k from the real line to appropriate contours in the complex
k-plane. These contours can be determined by studying the exponentials
involved in (26). We first note that, because 0 < x < ∞, the exponential
eikx is analytic and bounded for all k ∈ C+ . Moreover, (19) implies that

 
Re (iω)
−Im k 2 = −|k|2 sin(2 arg k), |k|  1. (27)

Hence, because 0 < t − τ < t the exponentials eiω(t−τ ) and e−iω(t−τ )

are bounded for arg k ∈ [0, π/2] and arg k ∈ [π/2, π ], respectively, as
72 A. A. Himonas and D. Mantzavinos

D2 D1
i

−i

Figure 2. The regions D1 , D2 .

|k| → ∞ in C+ . Consequently, the contours of integration of the terms

involving g +j and g −j in Equation (26) can be deformed to the positively
oriented boundaries of the regions D2 and D1 respectively, shown in
Figure 2.

REMARK 1. Because 0 < x < ∞ is unbounded and 0 < t − τ < t is

bounded, the exponential eikx is bounded only for k ∈ C+ while the exponentials
e±iω(t−τ ) are also bounded for all finite k ∈ C. Therefore, the deformed contours
of integration ∂ D1 and ∂ D2 may be any contours in the upper half of the
complex k-plane which, as |k| → ∞, coincide with the boundaries of the first
and the second quadrant, respectively.

We thus obtain the following integral representation for u(x, t):


1 eikx  iωt 
∞   
u(x, t) = e û 1 (k) + iωû 0 (k) − e−iωt û 1 (k) − iωû 0 (k) dk
2π −∞ 2iω
 ∞ ikx   t  t 
1 e −iωτ ˆ −iωt
+ e iωt
e f (k, τ )dτ − e e f (k, τ )dτ dk
iωτ ˆ
2π −∞ 2iω 0 0

1 eikx−iωt
+
− g3 (ω, t) + ikg2+ (ω, t) − (1 + k 2 )g1+ (ω, t)
2π ∂ D1 2iω

2 +
− ik(1 + k )g0 (ω, t) dk
On the Initial-Boundary Value Problem for the Linearized Boussinesq Equation 73

L2
L1

i
L2 L1

−i

Figure 3. The contours L 1 and L 2 .

Figure 4. The solution (48) depicted for (x, t) ∈ [0, 5] × [0, 6].
74 A. A. Himonas and D. Mantzavinos

Figure 5. The solution (42) for the initial data (49) and the boundary data (50), depicted for
(x, t) ∈ [0, 2] × [0, 3] with a = 3, b = 2.

Figure 6. The solution (42) for the initial data (49) and the boundary data (50), depicted for
(x, t) ∈ [0, 2] × [0, 3] with a = 2.5, b = 4.
 On the Initial-Boundary Value Problem for the Linearized Boussinesq Equation 75

Figure 7. The solution (47) for zero initial data and the boundary data (51), depicted for
(x, t) ∈ [0, 2] × [0, 3] and b = 5.


1 eikx+iωt
−
+ g3 (ω, t) + ikg2− (ω, t) − (1 + k 2 )g1− (ω, t)
2π ∂ D2 2iω

− ik(1 + k 2 )g0− (ω, t) dk. (28)

4. Elimination of the unknowns. Let

 1
ν = i 1 + k2 2 . (29)
Through the definition (18), we identify the following symmetry transformations
of ω:
k → −k ⇒ ω → ω, (30a)

k → ν ⇒ ω → −ω, (30b)

k → −ν ⇒ ω → −ω. (30c)
Applying these transformations to the global relation (25) gives rise to the
following equations:
76 A. A. Himonas and D. Mantzavinos

Figure 8. The solution (47) for zero initial data and the boundary data (51), depicted for
(x, t) ∈ [0, 2] × [0, 3] and b = 10.

z
z = h(x, t)

w
H
u
x

Figure 9. Long water waves in a horizontal channel of infinite length.

   
2iω û(−k, t) = eiωt û 1 (−k) + iωû 0 (−k) − e−iωt û 1 (−k) − iωû 0 (−k)
 t  t
−iω(τ −t) ˆ
+ e f (−k, τ )dτ − eiω(τ −t) fˆ(−k, τ )dτ
0 0

−e −iωt
g3+ (ω, t) − ikg2+ (ω, t) − (1 + k 2 )g1+ (ω, t)
 On the Initial-Boundary Value Problem for the Linearized Boussinesq Equation 77


+ ik(1 + k 2 )g0+ (ω, t)

+eiωt g3− (ω, t) − ikg2− (ω, t) − (1 + k 2 )g1− (ω, t)


+ik(1 + k 2 )g0− (ω, t) , k ∈ C+ , (31a)



− 2iω û(ν, t) = e−iωt û 1 (ν) − iωû 0 (ν) − eiωt û 1 (ν) + iωû 0 (ν)
 t  t
+ eiω(τ −t) fˆ(ν, τ )dτ − e−iω(τ −t) fˆ(ν, τ )dτ
0 0

− eiωt g3− (ω, t) + iνg2− (ω, t) − (1 + ν 2 )g1− (ω, t)


− iν(1 + ν 2 )g0− (ω, t)

+e−iωt g3+ (ω, t) + iνg2+ (ω, t) (1 + ν 2 )g1+ (ω, t)


−iν(1 + ν 2 )g0+ (ω, t) , k ∈ D2 , (31b)

and



− 2iω û(−ν, t) = e−iωt û 1 (−ν) − iωû 0 (−ν) − eiωt û 1 (−ν) + iωû 0 (−ν)
 t  t
iω(τ −t) ˆ
+ e f (−ν, τ )dτ − e−iω(τ −t) fˆ(−ν, τ )dτ
0 0

− eiωt g3− (ω, t) − iνg2− (ω, t) − (1 + ν 2 )g1− (ω, t)


+ iν(1 + ν 2 )g0− (ω, t)

+e−iωt g3+ (ω, t) − iνg2+ (ω, t) − (1 + ν 2 )g1+ (ω, t)


+iν(1 + ν 2 )g0+ (ω, t) , k ∈ D1 . (31c)

Note the adjustment of the domain of validity of the global relation in each
one of the above cases, according to the transformation applied.
The three relations (31a)–(31c) can be employed for the elimination of
the unknowns from the integral representation (28). Next, we illustrate this
procedure by considering two ibvps that appear often in the literature.
 78 A. A. Himonas and D. Mantzavinos

2.1. The canonical problem

Suppose that the prescribed boundary conditions are

u(0, t) = h 0 (t), u x (0, t) = h 1 (t), t > 0, (32)

for some given functions h 0 and h 1 . Then, the functions g0± , g1± can be
computed through Equation (24), while the functions g2± , g3± are unknown.
However, for k ∈ D2 we can use Equations (31a) and (31b) to characterize
g2− and g3− in terms of given functions. Indeed, these two equations can be
combined into the 2 × 2 linear system
⎛ iωt ⎞ ⎛ iωt ⎞
⎛ ⎞ e g− ⎛     ⎞ e −
−1 ik ⎜2iω 3 ⎟ − 1 + k 2 ik 1 + k 2 ⎜2iω g1 ⎟
⎝ ⎠⎜
⎜
⎟ ⎝
⎟= ⎠⎜ ⎟
⎝ ⎠     ⎜⎝
⎟
⎠
−1 −iν e iωt
− 1 + ν 2
−iν 1 + ν 2 e iωt
g2− g0−
2iω 2iω
+ N (k, t) + U (k, t), (33)

where the known vector N (k, t) is given by

N (k, t)  N0 (k, t) + N f (k, t) + N g (k, t) (34)

with
⎛ ⎞
eiωt e−iωt
⎜ 2iω 1[û (−k) + iω û 0 (−k)] − [û 1 (−k) − iω û 0 (−k)] ⎟
⎜ 2iω ⎟
N0 (k, t) = ⎜ ⎟, (35a)
⎝ eiωt e−iωt ⎠
[û 1 (ν) + iωû 0 (ν)] − [û 1 (ν) − iωû 0 (ν)]
2iω 2iω

⎛   ⎞
eiωt t −iωτ e−iωt t
⎜ 2iω e fˆ(−k, τ )dτ − e iωτ
f (−k, τ )dτ ⎟
ˆ
⎜ 0 2iω 0 ⎟
N f (k, t) = ⎜
⎜ iωt  t
⎟ , (35b)
⎟
⎝e  t ⎠
e−iωt
e−iωτ fˆ(ν, τ )dτ − e f (ν, τ )dτ
iωτ ˆ
2iω 0 2iω 0

⎛ ⎞
e−iωt   + +

⎜ 2iω 1 + k g1 − ikg0 ⎟
2

⎜ ⎟
N g (k, t) = ⎜ ⎟, (35c)
⎝ e−iωt     ⎠
1 + ν 2 g1+ + iνg0+
2iω
 On the Initial-Boundary Value Problem for the Linearized Boussinesq Equation 79

and the unknown vector U (k, t) is equal to

⎛ −iωt ⎞
e  + +

⎜ 2iω −g3 + ikg2 − û(−k, t) ⎟
⎜ ⎟
U (k, t)  ⎜ ⎟. (36)
⎝ e−iωt   ⎠
−g3+ − iνg2+ − û(ν, t)
2iω
g2− and 2iω g3− and inserting the resulting
iωt iωt
e e
Solving system (33) for 2iω
expressions in (28) yields an expression for the solution in which the only
unknowns come from the term U (k, t) on the right-hand side of (33). However,
these unknowns yield zero contributions when integrated along the contour ∂ D2 .
For example, consider the integral
   
eikx e−iωt  + +

ν g (ω, t) + ikg2 (ω, t) − û(−k, t)
∂ D2 k + ν 2iω 3
 −iωt 
e  + +

−k g (ω, t) − iνg2 (ω, t) − û(ν, t) dk, (37)
2iω 3

which arises after substituting for g3− in Equation (28) via Equation (33). First,
note that k + ν = 0 for k ∈ D2 . Moreover, by the definitions (22a) and (24) we
see that the integrand of (37) involves the exponentials eik(x+ξ ) , eikx−iω(t−τ ) and
eikx e−iνξ . Because x + ξ ≥ 0, the first exponential is analytic and bounded for
all k ∈ C+ . Regarding the second exponential, because t − τ > 0 and finite,
it is bounded for all finite k ∈ C+ , while as |k| → ∞ we need Re(iω) ≥ 0.
Thus, the second exponential is analytic and bounded in D2 . Finally, it follows
from (18) that the third exponential is also analytic and bounded for k ∈ D2 .
Hence, because ω = 0 in D2 we conclude by Cauchy’s theorem and Jordan’s
lemma that the integral (37) vanishes. Similarly, the integral that arises from
substituting for g2− in Equation (28) via Equation (33) also vanishes.
Consequently, we can write system (33) in the form
⎛ iωt ⎞ ⎛ iωt ⎞
⎛ ⎞ e g− ⎛     ⎞ e g−
−1 ik ⎜2iω 3 ⎟ − 1 + k ik 1 + k
2 2
⎜ 2iω 1 ⎟
⎝ ⎠⎜
⎜
⎟ ⎝
⎟
⎠⎜
⎜
⎟
⎟ + N (k, t), (38)
⎝ ⎠     ⎝
−1 −iν eiωt − − 1 + ν 2 −iν 1 + ν 2 eiωt −⎠
g g
2iω 2 2iω 0
where the use of “
” denotes that the functions which yield zero contributions
in the integral representation (28) have been omitted from the right-hand side
of (33). Actually, some of the terms involved in the known vector N (k, t) also
yield zero contributions.
Solving system (38), we obtain
 
g2−
− i ν − k g1− − iωg0−
80 A. A. Himonas and D. Mantzavinos


i    
+ − û 1 (−k) + iωû 0 (−k) + û 1 (ν) + iωû 0 (ν)
k+ν
 t  t 
−iωτ ˆ −iωτ ˆ
− e f (−k, τ )dτ + e f (ν, τ )dτ , k ∈ D2 , (39a)
0 0

and

g3−
(1 + iω)g1− − ω (ν − k) g0−

1    
− ν û 1 (−k) + iωû 0 (−k) + k û 1 (ν) + iωû 0 (ν)
k+ν
 t  t 
−iωτ ˆ −iωτ ˆ
+ν e f (−k, τ )dτ + k e f (ν, τ )dτ , k ∈ D2 .
0 0
(39b)

The unknowns g2+ and g3+ which are involved in the integral along ∂ D1 of
Equation (28) can be eliminated similarly to the functions g2− and g3− , combining
this time Equations (31a) and (31c) for k ∈ D1 . Analogous computations yield

g2+
i(ν + k)g1+
i    
+iωg0+ + û 1 (−ν) − iωû 0 (−ν) − û 1 (−k) − iωû 0 (−k)
k−ν
 t  t
+ e f (−ν, τ )dτ − eiωτ fˆ(−k, τ )dτ ,
iωτ ˆ
0 0
k ∈ D1 , (40a)

and
1    
g3+
(1 − iω)g1+ − ω(ν + k)g0+ + ν û 1 (−k) − iωû 0 (−k) − k û 1 (−ν)
k−ν
  t
 t
−iωû 0 (−ν) +ν eiωτ fˆ(−k, τ )dτ −k eiωτ fˆ(−ν, τ )dτ ,
0 0
k ∈ D1 . (40b)

Overall, employing expressions (39a), (39b), (40a), and (40b) in the integral
representation (28), we obtain
 ∞ ikx 
1 e    
u(x, t) = eiωt û 1 (k) + iωû 0 (k) − e−iωt û 1 (k) − iωû 0 (k) dk
2π −∞ 2iω
 
1 eikx+iωt k − ν  
+ û 1 (−k) + iωû 0 (−k)
2π ∂ D2 2iω k+ν
 On the Initial-Boundary Value Problem for the Linearized Boussinesq Equation 81


2k  
− û 1 (ν) + iωû 0 (ν) dk
k+ν
 
1 eikx−iωt k + ν 
− û 1 (−k) − iωû 0 (−k)
2π ∂ D1 2iω k−ν

2k  
− û 1 (−ν) − iωû 0 (−ν) dk
k−ν
 ∞   t  t 
1 eikx −iωτ ˆ −iωt
+ e iωt
e f (k, τ )dτ − e e f (k, τ )dτ dk
iωτ ˆ
2π −∞ 2iω 0 0
  
1 eikx+iωt k−ν t
+ e−iωτ fˆ(−k, τ )dτ
2π ∂ D2 2iω k+ν 0
 t 
2k −iωτ ˆ
− e f (ν, τ )dτ dk
k+ν 0
  
1 eikx−iωt k + ν t iωτ ˆ
− e f (−k, τ )dτ
2π ∂ D1 2iω k−ν 0
 t 
2k
− e f (−ν, τ )dτ dk
iωτ ˆ
k−ν 0

1 eikx+iωt

+ (iω − k 2 )g1− (ω, t) + ω(k − ν)g0− (ω, t) dk
2π ∂ D2 iω
 ikx−iωt

1 e
+ (iω + k 2 )g1+ (ω, t) + ω(k + ν)g0+ (ω, t) dk. (41)
2π ∂ D1 iω
Because all the terms on the right-hand side depend on the given initial and
boundary data and the known forcing, expression (41) is an explicit solution
formula for the canonical problem of the FLGB on the half-line.
The solution formula (41) can be further simplified after observing that the
1
sign of the function (1 + k 2 ) 2 changes from negative, on the part of ∂ D2 that
lies on the left of the branch cut, to positive, on the part of ∂ D1 that lies on the
right of the branch cut.
Indeed, recalling the definition (17) we have that ϕ1 = π , ϕ2 = 2π as k
approaches the cut along ∂ D2 so that
 1
1 + k 2 2 = |1 + k 2 | ei(π+2π−π)/2 = − |1 + k 2 |,
while ϕ1 = π , ϕ2 = 0 as k approaches the cut along ∂ D1 so that
 1
1 + k 2 2 = |1 + k 2 | ei(π+0−π)/2 = |1 + k 2 |.
Consequently, both ν and ω change sign across the branch cut. Hence, the
integrals in (41) over the parts of the contours ∂ D2 and ∂ D1 that lie on the left
82 A. A. Himonas and D. Mantzavinos

and on the right of the branch cut, respectively, cancel and we arrive at the
following proposition.
PROPOSITION 1 (Solution to the canonical problem). The solution of the
FLGB Equation (8) posed on the half-line with the initial conditions (12b)
and the boundary conditions (32) is given by formula (41). Furthermore, this
formula may be written in the simplified form
 ∞ ikx 
1 e    
u(x, t) = eiωt û 1 (k) + iωû 0 (k) − e−iωt û 1 (k) − iωû 0 (k) dk
2π −∞ 2iω
 
1 eikx+iωt k − ν  
+ û 1 (−k) + iωû 0 (−k)
2π L 2 2iω k+ν

2k  
− û 1 (ν) + iωû 0 (ν) dk
k+ν
 
1 eikx−iωt k + ν  
− û 1 (−k) − iωû 0 (−k)
2π L 1 2iω k−ν

2k  
− û 1 (−ν) − iωû 0 (−ν) dk
k−ν
 ∞ ikx   t
1 e
+ e iωt
e−iωτ fˆ(k, τ )dτ
2π −∞ 2iω 0
 t 
−iωt
−e e f (k, τ )dτ dk
iωτ ˆ
0
  
1 e ikx+iωt
k−ν t
+ e−iωτ fˆ(−k, τ )dτ
2π L2 2iω k+ν 0
 t 
2k −iωτ ˆ
− e f (ν, τ )dτ dk
k+ν 0
  
1 eikx−iωt k + ν t iωτ ˆ
− e f (−k, τ )dτ
2π L1 2iω k−ν 0
 t 
2k
− eiωτ fˆ(−ν, τ )dτ dk
k−ν 0

1 eikx+iωt

+ (iω − k 2 )g1− (ω, t) + ω(k − ν)g0− (ω, t) dk
2π L2 iω

1 eikx−iωt

+ (iω + k 2 )g1+ (ω, t) + ω(k + ν)g0+ (ω, t) dk, (42)
2π L1 iω
where the contours L 1 , L 2 are shown in Figure 3, the quantity ν is defined by
Equation (29), the functions û 0 , û 1 , fˆ are the Fourier transforms as defined
 On the Initial-Boundary Value Problem for the Linearized Boussinesq Equation 83

by (22a) of the initial conditions u 0 , u 1 and of the forcing f , and the functions
g1± , g0± are defined by Equations (24).

2.2. Given u and u x x at the boundary

As an alternative to the canonical ibvp of Section 2.1, we supplement the
FLGB equation with the boundary conditions
u(0, t) = h 0 (t), u x x (0, t) = h 2 (t), t > 0, (43)
where h 0 and h 2 are given functions. The unified transform method has the
advantage that it requires only small algebraic modifications to yield the
solution of the FLGB ibvp in this case.
The unknowns are now the functions g1+ , g3+ and g1− , g3− . As in Section 2.1,
by combining Equations (31a) and (31c) for k ∈ D1 we can express the former
pair in terms of known functions, while the latter pair can be eliminated via
Equations (31a) and (31b) for k ∈ D2 . We thus find
i  + 
g1+
− g2 − iωg0+
k+ν

1    
+ û 1 (−k) − iωû 0 (−k) − û 1 (−ν) − iωû 0 (−ν)
1 + 2k 2
 t  t 
+ e f (−k, τ )dτ − e
iωτ ˆ
f (−ν, τ )dτ ,
iωτ ˆ
0 0
k ∈ D1 , (44a)
i  
g3+
(iω − 1)g2+ − ω2 g0+
k+ν

1   
+ (1 + ν 2 ) û 1 (−k) − iωû 0 (−k) − (1 + k 2 ) û 1 (−ν)
1 + 2k 2
 t

−iωû 0 (−ν) + (1 + ν 2 ) eiωτ fˆ(−k, τ )dτ
0
 t 
−(1 + k ) 2
e iωτ
fˆ(−ν, τ )dτ , k ∈ D1 , (44b)
0

and i  − 
g1−
− g2 + iωg0−
k−ν

1    
+ û 1 (−k) + iωû 0 (−k) − û 1 (ν) + iωû 0 (ν)
1 + 2k 2
 t  t 
+ e iωτ ˆ
f (ν, τ )dτ − e f (−k, τ )dτ , k ∈ D2 , (45a)
iωτ ˆ
0 0
i  
g3−
(1 + iω) g2− − ω2 g0−
k−ν
84 A. A. Himonas and D. Mantzavinos


1    
+ (1 + ν 2
) û 1 (−k) + iω û 0 (−k) − (1 + k 2
) û 1 (ν) + iω û 0 (ν)
1 + 2k 2
 t
+ (1 + ν ) 2
e−iωτ fˆ(−k, τ )dτ
0
 t 
−iωτ
−(1 + k ) 2
e ˆf (ν, τ )dτ , k ∈ D2 . (45b)
0

Inserting these expressions in the integral representation (28) gives the analogue
of the solution formula (41), which again can be simplified to the analogue of
formula (42). In the particular case of zero initial conditions and zero forcing,
the expressions for g1± , g3± derived above become
i  + 
g1+
− g2 − iωg0+ , k ∈ D1 , (46a)
ν+k

i  
g3+
− (1 − iω) g2+ + ω2 g0+ , k ∈ D1 , (46b)
ν+k

i  − 
g1−
g2 + iωg0− , k ∈ D2 , (46c)
ν−k

i  
g3−
(1 + iω) g2− + ω2 g0− , k ∈ D2 , (46d)
ν−k
and the solution is given by

1 eikx+iωt
− 
u(x, t) = 1
g2 (ω, t) − (1 + k 2 )g0− (ω, t) dk
2π L 2 (1 + k 2 ) 2

1 eikx−iωt
+ 
− 1
g2 (ω, t) − (1 + k 2 )g0+ (ω, t) dk. (47)
2π L 1 (1 + k 2 ) 2

3. Numerical Evaluations

In the following examples, we provide graphical representations of the solution

(42) in the case of zero forcing and certain typical initial and boundary
conditions whose transforms can be computed explicitly. Numerical integration
is only required for the integrals with respect to the spectral variable k and is
done via a simple code in MATHEMATICA.
Example 1. For the initial conditions u 0 (x, 0) = u 1 (x, 0) = 0 and the
boundary conditions u(0, t) = te− 2 , u x (0, t) = 0, the solution formula (42)
t
 On the Initial-Boundary Value Problem for the Linearized Boussinesq Equation 85

becomes

1
u(x, t) = eikx+iωt (k − ν)g0− (ω, t)dk
2iπ L2

1
+ eikx−iωt (k + ν)g0+ (ω, t)dk (48)
2iπ L1

where the spectral functions g0± can be expressed in closed form as

   
± 1 1 ( ±iω− 21 )t
g0 (ω, t) =  2 ±iω − t −1 e +1 .
±iω − 12 2

The graph of this solution is provided in Figure 4.

Example 2. Suppose that the initial conditions are

u 0 (x) = xe−a x , u 1 (x) = x 2 e−a x ,

2 2
a ∈ R, (49)

and the boundary conditions are

u(0, t) = te− 2 ,
t
u x (0, t) = sin bt, b ∈ R. (50)

Then, we have
1 2
û 0 (k) = , û 1 (k) =
(ik + a 2 )2 (ik + a 2 )3
and also
   
1 1
g0± (ω, t)= 2 ±iω − t −1 e (±iω− 12 )t
+1 ,
±iω − 12 2
 
± b ±iωt iω ±iωt
g1 (ω, t) = 2 1−e cos bt ± e sin bt .
b − ω2 b
The solution formula (42) in the case of conditions (49) and (50) is depicted in
Figures 5 and 6 below for different values of a and b.
Example 3. Suppose that u 0 (x) = u 1 (x) = 0 and

u(0, t) = te− 2 ,
t
u x x (0, t) = sin bt, b ∈ R. (51)

Then, we have
   
1 1
g0± (ω, t) = 2 ±iω − t −1 e ( ±iω− 21 )t
+1 ,
±iω − 12 2
86 A. A. Himonas and D. Mantzavinos

and
 
b iω ±iωt
g2± (ω, t) = 2 1−e ±iωt
cos bt ± e sin bt .
b − ω2 b
The solution (47) for b = 5 and b = 10 is depicted Figures 7 and 8.

Acknowledgments

This work was partially supported by a grant from the Simons Foundation
(#246116 to Alex Himonas) and an AMS-Simons Travel Grant to Dionyssios
Mantzavinos. The authors would like to thank the referees of the paper for
constructive comments that led to its improvement.

Appendix A
Verification of the Solution

In the applied mathematics literature, a PDE is solved under the assumption

of existence. Indeed, because the solution is obtained by applying an appropriate
transform, this procedure makes sense only if one assumes that the solution
exists and has certain decay and smoothness properties. To eliminate this
assumption, one should prove a posteriori that the expression obtained by
this approach satisfies the PDE and the given initial and boundary conditions
(one then has to address independently the question of uniqueness). However,
this is not done in the applied literature. Actually, such a verification is not
straightforward because any representation obtained via the usual transforms
does not converge uniformly at the boundary. A major advantage of the unified
transform method is that it produces a solution which is uniformly convergent
at the boundary. Thus, through merely algebraic manipulations one can, at least
formally, verify that the solution satisfies the PDE as well as the given data.
Next, we perform this verification for the solution formula (41) of the
canonical problem 2.1. First, we note that for any fixed T > t, Cauchy’s
theorem and Jordan’s lemma imply
   
eikx+iωt k − ν T −iωτ ˆ
e f (−k, τ )dτ dk = 0 (52)
∂ D2 2iω k+ν t
and also,
   
eikx+iωt   T −iωτ
iω − k 2
e u x (0, τ )dτ dk = 0. (53)
∂ D2 iω t

The same is true for all the terms of (41) that depend on t and are integrated
along the contours ∂ D1 and ∂ D2 , that is, we can replace the upper limit of the
 On the Initial-Boundary Value Problem for the Linearized Boussinesq Equation 87

τ -integration by T > t. Hence, we obtain the following alternative version of

the solution formula (41):
 ∞ ikx  
1 e   −iωt
 
u(x, t) = e û 1 (k) + iωû 0 (k) − e
iωt
û 1 (k) − iωû 0 (k) dk
2π −∞ 2iω
 
1 eikx+iωt k − ν  
+ û 1 (−k) + iωû 0 (−k)
2π ∂ D2 2iω k+ν

2k  
− û 1 (ν) + iωû 0 (ν) dk
k+ν
 
1 eikx−iωt k + ν  
− û 1 (−k) − iωû 0 (−k)
2π ∂ D1 2iω k−ν

2k  
− û 1 (−ν) − iωû 0 (−ν) dk
k−ν
 ∞ ikx   t  t 
1 e −iωτ ˆ −iωt
+ e iωt
e f (k, τ )dτ − e e f (k, τ )dτ dk
iωτ ˆ
2π −∞ 2iω 0 0
  
1 eikx+iωt k − ν T −iωτ ˆ
+ e f (−k, τ )dτ
2π ∂ D2 2iω k+ν 0
 T 
2k
− e−iωτ fˆ(ν, τ )dτ dk
k+ν 0
  
1 eikx−iωt k + ν T iωτ ˆ
− e f (−k, τ )dτ
2π ∂ D1 2iω k−ν 0
 T 
2k
− e iωτ ˆ
f (−ν, τ )dτ dk
k−ν 0

1 eikx+iωt

+ (iω − k 2 )g1− (ω, T ) + ω(k − ν)g0− (ω, T ) dk
2π ∂ D2 iω

1 eikx−iωt

+ (iω + k 2 )g1+ (ω, T ) + ω(k + ν)g0+ (ω, T ) dk, (54)
2π ∂ D1 iω
where g ±j (ω, T ) are defined analogously to g ±j (ω, t) of Equation (24) by
 T
±
g j (ω, T ) = e±iωτ ∂xj u(0, τ )dτ, j = 0, 1. (55)
0

The FLGB equation. Observe that, with the exception of the forcing
terms that are integrated along R, the only dependence in x and t in the
solution formula (54) is due to the exponential terms. These exponentials,
however, are particular solutions of the FLGB, and hence, the corresponding
88 A. A. Himonas and D. Mantzavinos

integrals automatically satisfy it. Moreover, a straightforward application of

the Fundamental Theorem of Calculus yields
 ∞ ikx   t  t 
1 e
eiωt ∂t e−iωτ fˆ(k, τ )dτ − e−iωt eiωτ fˆ(k, τ )dτ dk = 0.
2π −∞ 2iω 0 0
(56)
Thus, the FLGB equation is satisfied by (54) and, equivalently, by (41).
The initial conditions. We verify the initial conditions through the solution
formula (41), because evaluating this formula at t = 0 results in major
simplifications that yield
 ∞
1 eikx    
u(x, 0) = û 1 (k) + iωû 0 (k) − û 1 (k) − iωû 0 (k) dk
2π −∞ 2iω
 
1 eikx k − ν  
+ û 1 (−k) + iωû 0 (−k)
2π ∂ D2 2iω k + ν

2k  
− û 1 (ν) + iωû 0 (ν) dk
k+ν
 
1 eikx k + ν  
− û 1 (−k) − iωû 0 (−k)
2π ∂ D1 2iω k − ν

2k  
− û 1 (−ν) − iωû 0 (−ν) dk. (57)
k−ν

Note that, by the definition of D1 and D2 , the exponentials eikx , e−iνx , and eiνx
are analytic and bounded for k in C+ , D2 , and D1 , respectively. It then follows
from Cauchy’s theorem and Jordan’s lemma that the second and the third term
of Equation (57) vanish, that is,
 ∞
1 eikx    
u(x, 0) = û 1 (k) + iωû 0 (k) − û 1 (k) − iωû 0 (k) dk
2π −∞ 2iω
= u 0 (x) (58)

by the Fourier transform inversion. Thus, the initial condition u(x, 0) = u 0 (x)
has been verified. The initial condition u t (x, 0) = u 1 (x) can be checked
similarly.
The boundary conditions. Evaluating expression (41) at x = 0 gives
 ∞
1 1  iωt    
u(0, t) = e û 1 (k) + iωû 0 (k) − e−iωt û 1 (k) − iωû 0 (k) dk
2π −∞ 2iω
 
1 eiωt k − ν  
+ û 1 (−k) + iωû 0 (−k)
2π ∂ D2 2iω k + ν
 On the Initial-Boundary Value Problem for the Linearized Boussinesq Equation 89


2k  
− û 1 (ν) + iωû 0 (ν) dk
k+ν
 
1 e−iωt k + ν 
− û 1 (−k) − iωû 0 (−k)
2π ∂ D1 2iω k−ν

2k  
− û 1 (−ν) − iωû 0 (−ν) dk
k−ν
 ∞   t  t 
1 1 −iωτ ˆ −iωt
+ e iωt
e f (k, τ )dτ − e e f (k, τ )dτ dk
iωτ ˆ
2π −∞ 2iω 0 0
  
1 eiωt k − ν t −iωτ ˆ
+ e f (−k, τ )dτ
2π ∂ D2 2iω k + ν 0
 t 
2k
− e−iωτ fˆ(ν, τ )dτ dk
k+ν 0
  
1 e−iωt k + ν t iωτ ˆ
− e f (−k, τ )dτ
2π ∂ D1 2iω k−ν 0
 t 
2k
− e iωτ ˆ
f (−ν, τ )dτ dk
k−ν 0

1 eiωt

+ (iω − k 2 )g1− (ω, t) + ω(k − ν)g0− (ω, t) dk
2π ∂ D2 iω

1 e−iωt

+ (iω + k 2 )g1+ (ω, t) + ω(k + ν)g0+ (ω, t) dk. (59)
2π ∂ D1 iω

Let k → −k in the first and fourth term, and then deform the contours of
integration from the real line to either ∂ D2 or ∂ D1 to get

1 eiωt  
u(0, t) = û 1 (−k) + iωû 0 (−k) dk
2π ∂ D2 2iω

1 e−iωt  
− û 1 (−k) − iωû 0 (−k) dk
2π ∂ D1 2iω
 
1 eiωt k − ν  
+ û 1 (−k) + iωû 0 (−k)
2π ∂ D2 2iω k + ν

2k  
− û 1 (ν) + iωû 0 (ν) dk
k+ν
 
1 e−iωt k + ν  
− û 1 (−k) − iωû 0 (−k)
2π ∂ D1 2iω k − ν
90 A. A. Himonas and D. Mantzavinos


2k  
− û 1 (−ν) − iωû 0 (−ν) dk
k−ν
  t
1 eiωt
+ e−iωτ fˆ(−k, τ )dτ dk
2π ∂ D2 2iω 0
 
1 e−iωt t iωτ ˆ
− e f (−k, τ )dτ dk
2π ∂ D1 2iω 0
  
1 eiωt k − ν t −iωτ ˆ
+ e f (−k, τ )dτ
2π ∂ D2 2iω k + ν 0
 t 
2k −iωτ ˆ
− e f (ν, τ )dτ dk
k+ν 0
  
1 e−iωt k + ν t iωτ ˆ
− e f (−k, τ )dτ
2π ∂ D1 2iω k − ν 0
 t 
2k
− e iωτ ˆ
f (−ν, τ )dτ dk
k−ν 0

1 eiωt

+ (iω − k 2 )g1− (ω, t) + ω(k − ν)g0− (ω, t) dk
2π ∂ D2 iω

1 e−iωt

+ (iω + k 2 )g1+ (ω, t) + ω(k + ν)g0+ (ω, t) dk. (60)
2π ∂ D1 iω
Consider now the second part of the third term above, namely the integral

e−iωt 2k  
û 1 (−ν) − iωû 0 (−ν) dk. (61)
∂ D1 2iω k − ν
1
The change of variables k = −i(1 + l 2 ) 2 implies that ω(k) = −ω(l) and
ν(k) = ν(l). Moreover, ∂ D1 is mapped to ∂ D2 and kdk = −ldl, hence

eiωt 2l  
(61) = − û 1 (−l) + iωû 0 (−l) dl. (62)
∂ D2 2iω l + ν

Adding (62) to the first part of the second term of (60), we obtain the integral

1 eiωt  
− û 1 (−k) + iωû 0 (−k) dk, (63)
2π ∂ D2 2iω
which cancels with the first integral of (60). Similar computations establish the
cancelation of the remaining terms that involve initial conditions or forcing.
Thus, Equation (60) becomes

1 eiωt

u(0, t) = (iω − k 2 )g1− (ω, t) + ω(k − ν)g0− (ω, t) dk (64a)
2π ∂ D2 iω
 On the Initial-Boundary Value Problem for the Linearized Boussinesq Equation 91


1 e−iωt

− (iω + k 2 )g1+ (ω, t) + ω(k + ν)g0+ (ω, t) dk. (64b)
2π ∂ D1 iω

1
What is more, the change of variables k = i(1 + l 2 ) 2 implies that
ω(k) = −ω(l) and ν(k) = −ν(l) and maps the contour ∂ D2 to the contour
∂ D1 , that is,


e−iωt   ldl
(64a) = (−iω + 1 + l 2 )g1+ (ω, t) − ω(l + ν)g0+ (ω, t) 1
∂ D1 iω i(1 + l 2 ) 2

e−iωt  
= − (l 2 + iω)g1+ (ω, t) + (il 3 − lω)g0+ (ω, t) dl. (65)
∂ D1 iω

Inserting (65) in (64), we find


1 e−iωt  
u(0, t) = − (l 2 + iω)g1+ (ω, t) + (il 3 − lω)g0+ (ω, t) dl
2π ∂ D1 iω

1 e−iωt  
+ (iω + k 2 )g1+ (ω, t) + ω(k + ν)g0+ (ω, t) dk
2π ∂ D1 iω

1 e−iωt
= ik(1 + 2k 2 )g0+ (ω, t)dk. (66)
2π ∂ D1 iω

1 1
Finally, letting l = k(1 + k 2 ) 2  ω, which implies that (1 + k 2 ) 2 dl =
(1 + 2k 2 )dk, we have

 
e−iωt
ik(1 + 2k 2 )g0+ (ω, t)dk = e−ilt g0+ (l, t)dl, (67)
∂ D1 iω R

which, recalling the definition (24) of g0+ , the boundary condition (32) and the
inverse Fourier transform formula, implies

1
u(0, t) = e−ilt g0+ (l, t)dl
2π R
  t
1 −ilt
= e eilτ h 0 (τ )dτ dl = h 0 (t). (68)
2π R 0

Thus, the Dirichlet boundary condition u(0, t) = h 0 (t) has been checked. The
verification of the boundary condition u x (0, t) = h 1 (t) is similar.
92 A. A. Himonas and D. Mantzavinos

Appendix B
The Derivation of Boussinesq

In this section, we present Boussinesq’s derivation of the so-called “bad”

Boussinesq Equation (1). The complete derivation in its original form is
contained in Ref. [2], pages 64–75.
As mentioned in the Introduction, Equation (1) was the outcome of
Boussinesq’s effort to explain the soliton observed by Scott Russell in 1834.
Thus, we consider waves traveling down a horizontal channel whose length is
much greater than its width, so that it can be regarded as infinite (see Figure 9).
The conservation of mass for a fluid of density ρ that moves under a velocity
field

u = (u, v, w)

is expressed by the equation

∂ρ
+ ∇ · (ρu) = 0. (69)
∂t
In the case of water, the flow may be assumed to be incompressible, in the
sense that every parcel of water retains the same density as it moves with the
flow. Mathematically, this can be expressed by the equation
Dρ
= 0, (70)
Dt
where the material derivative operator is defined by
D ∂
= + u · ∇. (71)
Dt ∂t
Hence, for an incompressible flow the conservation of mass Equation (69) is
equivalent to

∇ · u = 0. (72)

In addition, the conservation of momentum equation reads

Du
ρ = −∇ p + F, (73)
Dt
where p is the pressure and F is the total body force. Letting the acceleration
due to F be denoted by a, we have
F
a , (74)
ρ
 On the Initial-Boundary Value Problem for the Linearized Boussinesq Equation 93

and the conservation of momentum equation can be rearranged as

1 Du
∇p = a− . (75)
ρ Dt
Furthermore, assuming that the flow down the channel is irrotational, that is
∇ × u = 0, (76)
we deduce that the velocity field is a gradient field, that is, it can be written as
u = ∇φ (77)
for some potential φ = φ(x, y, z). Immediately, due to the conservation of
mass (72) we see that φ is harmonic, that is,
∇ 2 φ = 0. (78)
Also, because the only present body force is gravity, we have F = (0, 0, −g).
Thus, multiplying Equation (75) by the vector dx = (d x, dy, dz) yields
1 Du
∇ p · dx = (0, 0, −g) · dx − · dx. (79)
ρ Dt
By Equation (77), we have
Du ∂
= (∇φ) + (∇φ · ∇)∇φ (80)
Dt ∂t
and integrating by parts we find
 
Du ∂φ
· dx = + (∂i φ)∂i ∂ j φ d x j
Dt ∂t

∂φ 1
= + ∂ j (∂i φ)2 d x j
∂t 2

∂φ 1
= + ∇(∇φ)2 · dx
∂t 2
∂φ 1
= + (∇φ)2 . (81)
∂t 2
Hence, integrating Equation (79) by the Fundamental Theorem of Calculus
yields
p ∂φ 1
= −gz − − (∇φ)2 + f (t) (82)
ρ ∂t 2
for some arbitrary function f (t).
The impermeability of the bottom of the channel imposes the “no-penetration”
condition
∂φ
w(x, y, 0, t) = (x, y, 0, t) = 0. (83)
∂z
94 A. A. Himonas and D. Mantzavinos

Integrating with respect to z gives

 z z
∂ 2φ
φ(x, y, z, t) = φ0 (x, y, t) + (x, y, ζ, t)dζ dz  , (84)
0 0 ∂ζ 2
where
φ0 (x, y, t)  φ(x, y, 0, t). (85)
Combining this expression with Equation (78) gives the representation
 z  z
φ(x, y, z, t) = φ0 (x, y, t) − φ(x, y, ζ, t)dζ dz  . (86)
0 0

where
∂2 ∂2
= + (87)
∂ x 2 ∂ y2
is the Laplacian in the (x, y) coordinates.
Because the soliton’s velocity components u and v do not vary much with z,
we may assume at a first approximation that

u(x, y, z, t)
u(x, y, 0, t)  u 0 (x, y, t), (88a)

v(x, y, z, t)
v(x, y, 0, t)  v0 (x, y, t), (88b)

where u 0 and v0 are the relevant velocity components at the bottom of the
channel. Equivalently,

∂φ ∂φ0
(x, y, z, t)
(x, y, t), (89a)
∂x ∂x

∂φ ∂φ0
(x, y, z, t)
(x, y, t). (89b)
∂y ∂y
Then, Equation (86) yields
 z z
φ(x, y, t)
φ0 (x, y, t) − φ0 (x, y, t)dζ dz 
0 0

z2
= φ0 − φ0 (x, y, t) . (90)
2!
Inserting this approximation back in Equation (86) gives an improved
approximation. By repeating this iterative approximations scheme, we obtain
 On the Initial-Boundary Value Problem for the Linearized Boussinesq Equation 95

φ in the form of the power series

z2 z4 z6
φ(x, y, z, t) = φ0 (x, y, t) − φ0 + 2 φ 0 − 3 φ 0 + ... . (91)
2! 4! 6!
Let p0 be the atmospheric pressure, which we take to be constant.
Furthermore, let H denote the constant depth of calm water and denote the
depth of the agitated water by H + h, for some quantity h = h(x, y, t) which
is assumed to be small. Note that this assumption is valid provided that the
wave amplitude is small compared to the depth of the channel. Also, denote by
P the nonhydrostatic part of the pressure p. Then,
p = p0 + ρg (H + h − z) + P. (92)
and Equation (91) becomes
P ∂φ 1
= −gh − − (∇φ)2 + F(t), (93)
ρ ∂t 2
where F is some arbitrary function of time. However, far downstream the
potential φ is constant and equal to zero, that is,
lim φ(x, y, z, t) ≡ 0. (94)
x→∞

Thus, F ≡ 0 and (93) becomes

 
∂φ 1
P = −ρ gh + + (∇φ)2 . (95)
∂t 2
Also, at the free surface P = 0, while the vertical component of the
superficial fluid parcels equals the rate of change of h with respect to time,
that is, w = dh/dt, which implies
∂h ∂h d x ∂h dy
w= + + . (96)
∂t ∂ x dt ∂ y dt
Thus, we have the free surface conditions
∂h ∂h ∂h
P = 0 and w= +u +v at z = 0. (97)
∂t ∂x ∂y
What is more, the fact that the channel has infinite length suggests to assume
that h and φ do not change with y, that is,
h = h(x, t), φ = φ(x, z, t), φ0 = φ0 (x, t), (98)
from which we deduce that
v≡0 and u 0 = u 0 (x, t). (99)
96 A. A. Himonas and D. Mantzavinos

Also, as noted earlier far downstream we have u 0 = 0 and φ0 = 0, hence

 ∞  ∞
∂φ0  
φ0 (x, t) = − 
(x , t)d x = − u 0 (x  , t)d x  (100)
x ∂ x x
and Equation (91) becomes
 ∞
∂u 0 z 2 ∂ 3 u 0 z 4
φ(x, z, t) = − u 0 (x  , t)d x  − + + ... . (101)
x ∂ x 2! ∂ x 3 4!
Overall, combining Equations (95) and (101) with the free surface conditions
(97) yields the following system of equations for the agitation h and the
x-component u 0 of the velocity at the bottom of the channel:
  ∞ 
∂  (H + h)2 ∂u 0
gh + − u0 d x − + ...
∂t x 2! ∂x
 2  2 
1 (H + h)2 ∂ 2 u 0 ∂u 0
+ u0 − + . . . + −(H + h) + ...
2 2! ∂x2 ∂x
=0 (102a)

∂u 0 (H + h)3 ∂ 3 u 0
− (H + h) + + ...
∂x 3! ∂x3
 
∂h ∂h (H + h)2 ∂ 2 u 0
= + u0 − + ... , (102b)
∂t ∂x 2! ∂x2
where we have used the series representations
(H + h)2 ∂ 2 u 0
u = u0 − + ...
2! ∂x2
∂u 0 (H + h)3 ∂ 3 u 0
w = −(H + h) + + .... (103)
∂x 3! ∂x3
The first approximation. Recall that h is small relatively to H . This implies
that u 0 , which is the velocity in the x-direction at the bottom of the channel,
is also small and the relevant derivatives will be even smaller. As a first
approximation, we neglect the terms of (102a) that are small compared to u 0 and
also the terms of (102b) that are small compared to ∂u 0 /∂ x to get the system
 ∞
∂u 0 
gh − dx = 0 (104a)
x ∂t
∂u 0 1 ∂h
=− . (104b)
∂x H ∂t
Differentiating the first equation twice with respect to x, the second once
with respect to t and adding the resulting expressions gives d’Alembert’s wave
 On the Initial-Boundary Value Problem for the Linearized Boussinesq Equation 97

equation
∂ 2h ∂ 2h
= g H . (105)
∂t 2 ∂x2
The solution of this equation is

h(x, t) = f 1 (x − t g H ) + f 2 (x + t g H ). (106)
Hence, Equation (104b) becomes
∂u 0 g
 
= f 1 (x − t g H ) − f 2 (x + t g H ) (107)
∂x H
and integrating, we find
g

u 0 (x, t) = f 1 (x − t g H ) − f 2 (x + t g H ) + F(t), (108)
H
where F is an arbitrary function of time. √ According to Equation (104a),
however, F  (t) = 0 implying that F(t) = g/H c. Thus,
g
! c" ! c "
u 0 (x, t) = f 1 (x − t g H ) + − f 2 (x + t g H ) − . (109)
H 2 2
Moreover, by writing the expression (106) for h in the form
! c" ! c"
h(x, t) = f 1 (x − t g H ) + + f 2 (x + t g H ) − , (110)
2 2
and letting f 1 → f 1 + c/2 and f 2 → f 2 − c/2, we find

h(x, t) = f 1 (x − t g H ) + f 2 (x + t g H ) (111)
and
g

u0 = f 1 (x − t g H ) − f 2 (x + t g H ) . (112)
H
Further, because we are studying waves that propagate in the positive direction,
we impose the conditions u 0 (x, 0) = h(x, 0) = 0 for all x > 0. Then, Equations
(111) and (112) yield the relations
f 1 (x) = − f 2 (x), f 1 (x) = f 2 (x) ∀x > 0 (113)
and hence,
f 1 (x) = f 2 (x) = 0 ∀x > 0. (114)
Therefore, Equations (111) and (112) become

h(x, t) = f 1 (x − t g H ) (115)
98 A. A. Himonas and D. Mantzavinos

and
g g
u0 = f 1 (x − t g H ) = h, (116)
H H
implying that, in the first approximation,
√ the free surface performs a translation
in the x-direction at a speed g H .
The second approximation. Returning to Equations 102, at the level of the
second approximation we have the system
 ∞
∂u 0  H 2 ∂ 2 u 0 u2
gh − dx − + 0 =0 (117a)
x ∂t 2! ∂t∂ x 2

∂u 0 H 3 ∂ 3u0 ∂h ∂h
− (H + h) + = + u0 . (117b)
∂x 3! ∂ x 3 ∂t ∂x
The new terms involved in this system are smaller than the ones that were
present in system (104), thus they can be handled by employing the expressions
(115) and (116) for h and u 0 that were obtained in the first approximation.
Then, we have the system
 ∞
∂u 0  g H 2 ∂ 2 h g 2
gh − dx + + h =0 (118a)
x ∂t 2! ∂ x 2 2H

∂u 0 g ∂h H3 g g ∂ 3h ∂h g ∂h
−H − h + = + h (118b)
∂x H ∂x 3! H H ∂x 3 ∂t H ∂x
which can be rearranged to
 ∞  2 
∂u 0 g h2 2∂ h
d x = gh + +H (119a)
x ∂t 2 H ∂x2
 
∂h ∂u 0 ∂ h2 H 2 ∂ 2h
= −H − gH − . (119b)
∂t ∂x ∂x H 6 ∂x2
Next, we differentiate the second equation with respect to t; note that √ the
bracket on the right-hand side can be treated as a function of (x − g H t)
according to the first approximation and hence, its derivative √ with respect to t
is equal to its derivative with respect to x multiplied by − g H . Also, we
differentiate the first equation twice with respect to x. Overall, we find
 2 
∂ 2u0 ∂ 2h g ∂ 2 h2 2∂ h
− =g 2+ +H (120a)
∂t∂ x ∂x 2 ∂x2 H ∂x2
 
1 ∂ 2h ∂ 2u0 ∂ 2 h2 H 2 ∂ 2h
=− +g 2 − . (120b)
H ∂t 2 ∂t∂ x ∂x H 6 ∂x2
 On the Initial-Boundary Value Problem for the Linearized Boussinesq Equation 99

Adding these two equations gives the “bad” Boussinesq equation in the form
 
∂ 2h ∂ 2 h 3g ∂ 2 h 2 g H 3 ∂ 4h
= g H + + . (121)
∂t 2 ∂x2 2 ∂x2 3 ∂x4
Rescaling. Letting
√
2 3 3g
h = H q, s= x, τ= t (122)
3 H H
so that
√
∂ 3 ∂ ∂ 3g ∂
= , = , (123)
∂x H ∂s ∂t H ∂τ
Equation (121) assumes the form
∂ 2q ∂ 2q ∂ 2 (q 2 ) ∂ 4 q
= + + 4, (124)
∂τ 2 ∂s 2 ∂s 2 ∂s
which is Equation (1).

References

1. J.V. BOUSSINESQ, Théorie de l’ intumescence liquide, appelée onde solitaire ou de

translation, se propageant dans un canal rectangulaire. Comptes rendus de l’ Académie
des Sciences 72:755–759 (1871).
2. J.V. BOUSSINESQ, Théorie des ondes et des remous qui se propagent le long d’un canal
rectangulaire horizontal, en communiquant au liquide contenu dans ce canal des vitesses
sensiblement pareilles de la surface au fond. J. Math. Pures Appl. 17:55–108 (1872).
3. J.V. BOUSSINESQ, Essai sur la théorie des eaux courantes. Mémoires présentés par divers
savants à l’Académie des Sciences 23(1):1–680 (1877).
4. J.L. BONA and R. SMITH, A model for the two-way propagation of water waves in a
channel. Math. Proc. Camb. Phil. Soc. 79:167–182 (1976).
5. J.L. BONA and R.L. SACHS, Global existence of smooth solutions and stability of solitary
waves for a generalized Boussinesq equation. Comm. Math. Phys. 118:15–29 (1988).
6. L.G. FARAH, Local solutions in Sobolev spaces with negative indices for the “good”
Boussinesq equation. Comm. PDE 34:52–73 (2009).
7. L.G. FARAH and M. SCIALOM, On the periodic “good” Boussinesq equation. Proc. AMS
138:953–964 (2010).
8. A.S. FOKAS, A unified transform method for solving linear and certain nonlinear PDEs.
Proc. R. Soc. A 453:1411–1443 (1997).
9. A.S. FOKAS, A unified approach to boundary value problems, SIAM (2008).
10. D. GEBA, A. HIMONAS, and D. KARAPETYAN, Ill-posedness results for generalized
Boussinesq equations. Nonlinear Anal. 95:404–413 (2014).
11. N. KISHIMOTO, Sharp local well-posedness for the “good” Boussinesq equation. J. Diff.
Eq. 254:2393–2433 (2013).
12. N. KISHIMOTO and K. TSUGAWA, Local well-posedness for quadratic nonlinear Schrödinger
equations and the good Boussinesq equation. Diff. Integral Eq. 23:463–493 (2010).
100 A. A. Himonas and D. Mantzavinos

13. F. LINARES, Global existence of small solutions for a generalized Boussinesq equation.
J. Diff. Eq. 106:257–293 (1993).
14. M. TSUTSUMI and T. MATAHASHI, On the Cauchy problem for the Boussinesq type
equation. Math. Japon. 36:371–379 (1991).
15. R. XUE, The initial-boundary value problem for the “good” Boussinesq equation on the
half-line. Nonlin. Anal. 69:647–682 (2008).

UNIVERSITY OF NOTRE DAME

(Received March 16, 2014)
Copyright of Studies in Applied Mathematics is the property of Wiley-Blackwell and its
content may not be copied or emailed to multiple sites or posted to a listserv without the
copyright holder's express written permission. However, users may print, download, or email
articles for individual use.