Sharp Well-posedness Results for the BBM Equation

J.L. Bona and N. Tzvetkov

Abstract
The regularized long-wave or BBM equation

ut + ux + uux − uxxt = 0

was derived as a model for the unidirectional propagation of long-crested, surface water
waves. It arises in other contexts as well, and is generally understood as an alternative
to the Korteweg-de Vries equation. Considered here is the initial-value problem wherein
u is specified everywhere at a given time t = 0, say, and inquiry is then made into its
further development for t > 0. It is proven that this initial-value problem is globally well
posed in the L2 -based Sobolev class H s if s ≥ 0. Moreover, the map that associates the
relevant solution to given initial data is shown to be smooth. On the other hand, if s < 0,
it is demonstrated that the correspondence between initial data and putative solutions
cannot be even of class C 2 . Hence, it is concluded that the BBM equation cannot be
solved by iteration of a bounded mapping leading to a fixed point in H s -based spaces for
s < 0. One is thus led to surmise that the initial-value problem for the BBM equation is
not even locally well posed in H s for negative values of s.

1 Introduction
The propagation of unidirectional, one-dimensional, small-amplitude long waves in nonlinear
dispersive media is sometimes well approximated by the Korteweg-de Vries equation or its
regularized counterpart the BBM-equation (see e.g. [8], [16], [26]). Starting in the latter half
of the 1960’s and in the 1970’s, the mathematical theory for such nonlinear, dispersive wave
equations came to the fore as a major topic within nonlinear analysis. Much effort has been
expended on various aspects of the pure initial value problems

ut + ux + uux + uxxx = 0 (1)

or
ut + ux + uux − uxxt = 0 (2)
with

u(x, 0) = u0 (x), (3)

though these are not always the most physically relevant formulations (see the discussions
in [3], [4], [5] [8], [10]). For the Korteweg-de Vries equation (1), recent theory has shown

1
the pure initial-value problem to be globally well posed in the L2 -based Sobolev classes H s
if s > − 43 (see [15]). This result turns out to be sharp in the sense that below this value
of s, the initial-value problem cannot be solved by a Picard iteration, as is shown in the
recent work of Christ, Colliander and Tao [14] (see also [19, 12, 22] and further comments
in the concluding section). The prospect in view here is a similar theory for the initial-value
problem for (2)-(3).
First, it will be established in Section 2 that (2) is globally well posed in H s provided only
that s ≥ 0. Let U be the mapping that associates to a given function u0 = u0 (x) the unique
solution u = u(t, x) of (2) with that initial data. The proof of local well-posedness leads to
the conclusion that U is smooth as a mapping of the relevant Banach spaces. In particular,
U is well defined on L2 and hence on a dense subset of any Sobolev space H s for s < 0. In
Section 3, it is demonstrated that when considered as a putative mapping of H s for s < 0, U
cannot be even C 2 . It is thus projected that s = 0 is the sharp value for which well-posedness
holds for (2). Section 4 contains a short conclusion and additional commentary.
Previously, the initial-value problem for (2) was known to be globally well-posed in H k
for integers k ≥ 1 (see Benjamin el al. [2]). The argument for local well-posedness in L2
follows the analysis of Boussinesq systems by Bona, Chen and Saut [6, 7]. Our result follows
from an elaboration of their argument using a short- and long-wavelength decomposition as
in Bourgain [13]. Analyticity of the solution map is an elementary consequence of the local
existence theory, but appears not to have been noted in the literature. For (1), analyticity of
the solution map was first established by Zhang [23, 24, 25].

Notation. We denote by b· or F the Fourier transform and by F −1 the inverse transform. In

case the variable(s) on either side of this transformation need emphasis, the notation Fx7→ξ
or F(x,t)7→(ξ,τ ) will be employed. In general, the norm in a Banach space X is denoted k · kX .
Thus the symbol k · kLp denotes the norm in the Lebesgue space Lp while k · kH s is the norm
in the L2 – based Sobolev space H s = H s (R). The notation A ≈ B means that there exists
a constant c ≥ 1 such that 1c |A| ≤ |B| ≤ c|A|. For any positive A and B, the notation
A . B (resp. A & B) means that there exists a positive constant c such that A ≤ cB (resp.
1
A ≥ cB). The symbol hξi connotes (1+ξ 2 ) 2 . The characteristic function of an interval I ⊂ R
is written 1lI . The convolution over R of two functions f and g is written f ? g. The only
inner product intervening in our analysis is that of L2 . If f, g ∈ L2 , then hf, gi is the L2 -
inner product of f and g.

2 Well-posedness for data in H s , s ≥ 0

As already mentioned, the initial-value problem (2)-(3) is known to be globally well posed
in H k for any integer k = 1, 2, · · ·. A straightforward application of nonlinear interpolation
theory (see [9] and the references therein) extends this result to H s for any s ≥ 1. Thus our
interest devolves to the range 0 ≤ s ≤ 1.

2
 2.1 Bilinear Estimates
The goal of this subsection is to prove two helpful inequalities (see also [7]).

Lemma 1 Let u, v ∈ H s (R), s ≥ 0. Then

kϕ(Dx )(uv)kH s . kukH s kvkH s , (4)

ξ \
where ϕ(ξ) = 1+ξ 2
and ϕ(Dx ) is the Fourier multiplier operator defined by ϕ(Dx )u(ξ) =
ϕ(ξ)b
u(ξ).

Proof. Expressing ϕ(Dx )uv in terms of Fourier transformed variables and using duality and
a polarization argument, one may write (4) in the equivalent form
Z ∞ Z ∞
ξhξis

 

2 s s
û(ξ1 )v̂(ξ − ξ1 )ŵ(ξ)dξdξ1  . kukL2 kvkL2 kwkL2 . (5)
−∞ −∞ (1 + ξ )hξ1 i hξ − ξ1 i


For s ≥ 0, hξis . hξ1 is hξ − ξ1 is . In consequence, the combination hξis /hξ1 is hξ − ξ1 is is

ξ
bounded and may be ignored. Set w1 (ξ) = 1+ξ 2 ŵ(ξ). With this notation, the left-hand side

of (5) is simply hû ? v̂, w̄1 i. Define u1 to be the reverse of û, so u1 (x) = û(−x) for all x.
Clearly, u and u1 have the same L2 -norm. Then hû ? v̂, w̄1 i = hu1 ? w1 , v̂i ¯ and hence one
may bound (5) via the Cauchy-Schwarz inequality by ku1 ? w1 kL2 kvkL2 . A use of Young’s
ku1 ? w1 kL2 ≤ kukL2 kw1 kL1 . The proof is completed by estimating the
inequality then gives
ξ
1
L -norm of w1 by 1+ξ 2
2
kwkL2 . 2
Lξ

Remark. The estimate (5) is not valid for s < 0. This becomes clear upon letting û
be the characteristic function of the interval [N − 1, N + 1], v̂ the characteristic function of
the interval [−N − 1, −N + 1] and ŵ(ξ) the characteristic function of the interval [−1, 1] in
(5). Then the left-hand side of (5) behaves as N −2s , while the right-hand side is a constant
independent of N . Hence, if s < 0, no matter how large is the implied constant, (5) fails for
N  1.
1
Lemma 2 Let f ∈ H r (R) and g ∈ H s (R) for some r and s with 0 ≤ s ≤ r and 2 < r. Then,
ϕ(Dx )f g ∈ H s+1 (R) and there is a constant C = C(r, s) such that

kϕ(Dx )(f g)kH s+1 (R) ≤ Ckf kH r (R) kgkH s (R) , (6)

where ϕ is as in Lemma 1.
1
Proof. Since r > 2 and r ≥ s ≥ 0, the elements of H r (R) are multipliers in H s (R), which is
to say,

kf gkH s (R) ≤ Ckf kH r (R) kgkH s (R)

where C = C(r, s). Since ϕ(Dx ) smooths by exactly one derivative in the L2 -based Sobolev
classes H s (R), the result follows. 2

3
 2.2 Local Well-posedness
Write (2)-(3) in the form

iut = ϕ(Dx )u + 21 ϕ(Dx )u2 ,


(7)
u(0, x) = u0 (x).

Let S(t) = exp(−itϕ(Dx )) be the unitary group defining the associated free evolution; that
is, S(t)u0 solves the linear initial-value problem

iut = ϕ(Dx )u,
(8)
u(0, x) = u0 (x).

Then, (7) may be rewritten as the integral equation

Z t
i
u(t, x) = S(t)u0 (x) − S(t − t0 )ϕ(Dx )u2 (t0 , x)dt0 = A(u, u0 )(x, t). (9)
2 0

This latter integral equation may be solved locally in time by performing a Picard iteration
(e.g. by a fixed-point argument) in the space XTs of continuous functions defined on [−T, T ]
with values in H s , equipped its usual norm

kukXTs = sup ku(t, ·)kH s .

t∈[−T,T ]

More precisely, argue as follows. The H s norm is clearly preserved by the free evolution since
its symbol has absolute value equal to 1. Thus, for any t ≥ 0 and s ∈ R,

kS(t)u0 kH s = ku0 kH s ,

and consequently, for any T > 0,

kS(t)u0 kXTs = ku0 kH s .

The second term on the right-hand side of (9) may be bounded by using again the fact that
S(t) is a unitary operator in H s (R) for each value of its argument and Lemma 1, viz.
Z t
1 S(t − t0 )ϕ(Dx )u2 (t0 , x)dt0 ≤ 1 Cs T kuk2X s

(10)
2 0
s
X 2 T
T

where Cs is a constant depending only on s. Similarly there obtains

Z t 
1 S(t − t0 )ϕ(Dx ) u2 (t0 , x) − v 2 (t0 , x) dt0 ≤ 1 Cs T ku − vkX s ku + vkX s .

(11)
2 0
s
X 2 T T
T

These inequalities imply the following local well-posedness result.

4
Theorem 3 Fix s ≥ 0. For any u0 ∈ H s (R), there exists a T = T (u0 ) > 0 and a unique
solution u ∈ XTs of the initial-value problem (2)-(3). The maximal existence time T = Ts for
the solution has the property that
1
Ts ≥ (12)
4Cs ku0 kH s (R)
where the positive constant Cs depends only on s.
For R > 0, let BR denote the ball of radius R centered at the origin in H s (R) and let
T = T (R) > 0 denote a uniform existence time for (2)- (3) with u0 ∈ BR . Then the
correspondence u0 7→ u that associates to u0 the solution u of (2)-(3) with initial value u0 is
a real analytic mapping of BR to XTs .

Proof. Existence, uniqueness and the lower bound on the existence time follow from the
contraction mapping principle applied to the closed ball BM centered at the origin in XTs ,
where we may choose
1 1
M = 2ku0 kH s (R) and T = = , (13)
2Cs M 4Cs ku0 kH s (R)

for example. It follows readily from (10) and (11) that the mapping A in (9) is a contraction
mapping of BM and thus A has a unique fixed point which is easily seen to comprise a
solution of (2)-(3) on the time interval [0, T ]. Clearly Ts ≥ 1/4Cs ku0 kH s (R) . (By solution,
we mean in the first instance a solution in the sense of tempered distributions, say, but it
is straightforward to ascertain that in fact, all the terms in the differential equation lie in
XTs−2 and that the equation is satisfied identically at least in this space.) Note that in this
case, the equivalence of the integral equation and the initial value problem in the space XTs
is clear, at least for s ≥ 0.
Attention is now turned to the smoothness issue for the solution map U. This result is
local in the sense that if it can be established for T sufficiently small, it is generally true
because of uniqueness of solutions to the initial-value problem and the semigroup property.
Let Λ : H s × XTs −→ XTs be defined as

1 t
Z
Λ(u0 , v(t)) = v(t) − S(t)u0 − S(t − t0 )ϕ(Dx )v 2 (t0 )dt0 ,
2 0
where the spatial variable x has been suppressed throughout. Due to Lemma 1, Λ is a smooth
map from H s × XTs to XTs , provided s ≥ 0. Let Λ(u0 , u(t)) = 0, which is to say, suppose u(t)
is a solution of (2) with initial data u0 . Then the Fréchet derivative of Λ with respect to the
second variable is calculated to be the linear map
Z t
Λ0u (u0 , u(t))[h] = h − S(t − t0 )ϕ(Dx )u(t0 )h(t0 )dt0 .
0

Since, as in the proof of Theorem 3,

Z t
S(t − t0 )ϕ(Dx )h(t0 )u(t0 )dt0

. T kukXTs khkXTs ,
0 XTs

5
it is deduced that for T small enough Λ0u (u0 , u(t)) is invertible since it is of the form I + K
and
kKkB(XTs ,XTs ) < 1,
where B(XTs , XTs ) is the Banach space of bounded linear operators on XTs . Thus the second
assertion of Theorem 3 follows from the Implicit Function Theorem. 2

Remark The argument for smoothness of the flow map U is quite general and can be found
in many particular contexts (see, e.g. Bekiranov [1], Kenig, Ponce and Vega [18, 19], Bona,
Sun and Zhang [10] and Zhang [23, 24, 25] to name but a few). Indeed, whenever one solves a
partial differential equation in a suitable functional framework by a Picard iteration scheme
applied to an integral equation formulation, 1 there automatically obtains strong regularity
information on the corresponding flow map. This is in contrast with the methods and results
for quasi-linear partial differential equations which provide local existence, uniqueness and
continuity of the flow map, but the flow map may lack smoothness; in fact, it may not even
be uniformly continuous on bounded sets.

2.3 Global Well-Posedness

In this subsection, the local theory is extended using a low-frequency–high-frequency
decomposition. The outcome is a satisfactory global well-posedness theorem.

Theorem 4 In Theorem 3, the value of T may be taken arbitrarily large and hence the
Cauchy problem (2)–(3) is globally well-posed in H s for any s ≥ 0.

Proof of Theorem 4. Fix T > 0. The aim is to show that corresponding to any initial data
u0 ∈ H s , there is a unique solution u of (2) that lies in XTs , and that u depends continuously
upon u0 . Because of Theorem 3, this result is clear for data that is small enough in H s .
Moreover, as continuous dependence, uniqueness and the analytic dependence on the data
of the flow map are all properties that are local in time, the issue is to prove existence of a
solution corresponding to initial data of arbitrary size. Fix u0 ∈ H s and let N  1 be such
that
Z
u0 (ξ)|2 dξ . T −2 .
hξi2s |c (14)
|ξ|≥N

Such values of N exist since hξi2s |c

u0 (ξ)|2 is an L1 -function. With N fixed as above, define
Z
v0 (x) := eixξ u
c0 (ξ)dξ .
|ξ|≥N

Using Theorem 3, a solution v ∈ XTs of (2) is adduced having as initial data v0 . Split the
initial datum u0 into two pieces, namely

u0 = v0 + w0 ,
1
Exactly as we do in the fundamental Cauchy-Lipschitz theorem for ODE’s

6
and consider the initial-value problem

wt − wxxt + wx + wwx + (vw)x = 0,
(15)
w(0, x) = w0 (x).

If there is a solution of (15) in XTs , then v + w will be a solution of (2) in XTs and the
result will be established. Observe that w0 ∈ H 1 (R). (Indeed, w0 lies in H r (R) for any r.)
The analogue of the integral equation (9) in this context can be solved locally in time in
XS1 for small values of S, by the same sort of contraction mapping argument used to prove
Theorem 3. The function v is fixed in this discussion of courses, and use has been made of
Lemma 2 with r = 1. If one had in hand an a priori bound on the H 1 -norm of w showing
it was bounded on the interval [−T, T ], then it would follow that the contraction argument
could be iterated and a solution on [−T, T ] thereby obtained.
An a priori bound for w is now provided which implies that the local existence result
continues to hold on any time interval over which kv(t, ·)kL2 remains finite, and in particular
on [−T, T ]. The formal steps to this inequality are as follows. Multiply the first equation in
(15) by w and integrate over the entire real line R. After integrations by parts, there appears
the identity
Z ∞  Z ∞
1d 2 2
(w(t, x) + wx (t, x))dx − v(t, x)w(t, x)wx (t, x)dx = 0. (16)
2 dt −∞ −∞

Due to the Sobolev and Hölder inequalities, the estimate

Z ∞ 
 . kw(t, ·)k2 1 kv(t, ·)kL2
 

 v(t, x)w(t, x)wx (t, x)dx  H (17)
−∞

is valid. Then (16), (17) and Gronwall’s inequality yield

Z t 
0 0
kw(t, ·)kH 1 . kw0 kH 1 exp kv(t , ·)kL2 dt . (18)
0

The key a priori bound (18) now enables one to deduce that w(t) exists at least on the time
interval [−T, T ]. The justification of these formal steps is made by regularizing; for smooth
solutions the calculations are secure. One then finishes by using the continuous dependence
result to infer that (18) remains valid in the limit wherein the smoothing disappears.
This completes the proof of Theorem 4. 2

3 Local “ill-posedness” below L2

Here, an ill-posedness result is established which indicates the sharpness of Theorems 3 and
4. This result suggests that one can not expect to obtain via iteration arguments even local
solutions of the BBM equation for data in H s if s < 0.

Theorem 5 For any s < 0, T > 0 the flow map Φ established in Theorems 3 and 4 is not
of class C 2 from H s to XTs .

7
 The proof is made in several steps outlined below. The analysis turns upon the explicit
arithmetic relation
ξ1 ξ − ξ1 ξ ξξ1 (ξ − ξ1 )(ξ 2 − ξξ1 + ξ12 + 3)
+ − = := θ(ξ, ξ1 ) (19)
1 + ξ12 1 + (ξ − ξ1 )2 1 + ξ 2 (1 + ξ12 )(1 + (ξ − ξ1 )2 )(1 + ξ 2 )
for the symbol ϕ.
3.1 Reduction of Theorem 5 to disproving a bilinear estimate
Consider the Cauchy problem

iut = ϕ(Dx )u + 12 ϕ(Dx )(u2 ),


(20)
u(0, x) = ηu0 (x) ,

where η > 0 is a parameter. As in (9), write (20) as an integral equation, viz.

u2 (t0 , x) 0
Z t
u(t, x) = ηS(t)u0 (x) − i S(t − t0 )ϕ(Dx )( )dt ,
0 2
where, as before, S(t) = exp(−itϕ(Dx )) is the unitary group defining the solution of the
linear BBM-equation. The solution of (20) is a function of three variables, u = u(η, t, x).
Clearly u(0, t, x) = 0. Furthermore, the formal first two derivatives of u(η, t, x) with respect
to η at η = 0 are
∂u
(0, t, x) = S(t)u0 (x) := u1 (t, x)
∂η
and
∂2u t
u21 (t0 , x) 0
Z
(0, t, x) = −2i S(t − t0 )ϕ(Dx )( )dt := u2 (t, x),
∂η 2 0 2
respectively. If it is presumed that the map U exists and is of class C 2 from H s to XTs , then
the estimate

ku2 kXTs . ku0 k2H s (21)

necessarily holds. The strategy used to prove Theorem 5 is to find a u0 such that (21) fails
if s < 0, no matter how small might be T > 0.
3.2 The choice of u0 and a representation for u2
Let u0 be a defined via its Fourier transform as follows;
1
c0 (ξ) = γ − 2 N −s (1lI1 (ξ) + 1lI2 (ξ))
u

where I1 = [−N − γ, −N + γ] and I2 = [N − γ, N + γ]. Here, the positive constant N is large

and the positive constant γ is small. The relation between N and γ will be fixed presently.
It is clear that ku0 kH s ≈ 1. Moreover, notice that u0 is real-valued since u
c0 is an even,
real-valued function.
Consider the first iteration u1 as a function of N , γ and s. Because

Fx7→ξ (u1 )(t, ξ) = exp(−itϕ(ξ))c

u0 (ξ),

8
it is immediate that up to a factor of 2π,
Z
− 21 −s
u1 (t, x) = γ N exp(ixξ − itϕ(ξ))dξ.
ξ∈I1 ∪I2

To compute u2 , the following technical lemma is helpful.

Lemma 5 Let F (t, x) be given and define v(t, x) by

Z t
v(t, x) = S(t − t0 )F (t0 , x)dt0
0

where S(t) = exp(−itϕ(Dx )) as before. Then, formally, v may be expressed in the form
∞ ∞
eit(τ +ϕ(ξ)) − 1 b
Z Z
1
v(t, x) = exp(ixξ − itϕ(ξ)) F (τ, ξ)dτ dξ.
4iπ 2 −∞ −∞ τ + ϕ(ξ)

Proof of Lemma 5. We follow closely the proof of Lemma 4 in [20]. By the group law for
S(t), v(t, x) may be written as
Z t
v(t, x) = S(t) S(−t0 )F (t0 , x)dt0 . (22)
0

Set H(t0 , x) = S(−t0 )F (t0 , x). A calculation reveals that

t ∞
eitτ − 1
Z Z
0 0 1 0
S(−t )F (t , x)dt = Ft0 7→τ (H)(τ, x)dτ. (23)
0 2π −∞ iτ

Indeed, both sides of (23) vanish at t = 0, and the derivative of the right-hand side with
respect to t is H(t, x) by the formula of the inverse Fourier transform. Define G by

G(t, t0 , x) = S(t − t0 )F (t0 , x).

Using (22), (23) and the fact that S(t) is, for each t, a Fourier multiplier operator, one arrives
at the representation
Z ∞ itτ
1 e −1
v(t, x) = Ft0 7→τ (G)(t, τ, x)dτ
2π −∞ iτ

for v. The inverse Fourier transform formula with respect to x then yields
Z ∞ Z ∞ itτ
1 e −1
v(t, x) = 2
F(t0 ,x)7→(τ,ξ) (G)(t, τ, ξ)eixξ dτ dξ. (24)
4iπ −∞ −∞ τ

Now, it is clear that

Fx7→ξ (G)(t, t0 , ξ) = exp(−i(t − t0 )ϕ(ξ))Fx7→ξ (F )(t0 , ξ),

9
and therefore
Z ∞
−itϕ(ξ) 0
F(t0 ,x)7→(τ,ξ) (G)(t, τ, ξ) = e e−it (τ −ϕ(ξ)) Fx7→ξ (F )(t0 , ξ)dt0
−∞
−itϕ(ξ)
= e Fb(τ − ϕ(ξ), ξ). (25)

Substituting (25) into (24) gives

Z ∞Z ∞
1 eitτ − 1 b
v(t, x) = exp(ixξ − itϕ(ξ)) F (τ − ϕ(ξ), ξ)dτ dξ.
4iπ 2 −∞ −∞ τ

It remains to perform the change of variable

τ − ϕ(ξ) 7→ τ1

to complete the proof of Lemma 5. 2

Using Lemma 5, the following representation for u2 (t, x) emerges; up to a constant,

Z ∞Z ∞
eit(τ +ϕ(ξ)) − 1
u2 (t, x) = exp(ixξ − itϕ(ξ)) ϕ(ξ) (c
u1 ? u
c1 ) (τ, ξ)dτ dξ,
−∞ −∞ τ + ϕ(ξ)

in which justifying the formal steps leading to the representation presents no difficulty because
of the particular form of what was called F in the Lemma. Because of the explicit formula
u
c1 (τ, ξ) = δ(τ + ϕ(ξ))c u0 (ξ) (δ connoting the Dirac delta function) the convolution u c1 ? u
c1 is
seen to have the form
Z ∞
(c
u1 ? u
c1 ) (τ, ξ) = δ(τ + ϕ(ξ1 ) + ϕ(ξ − ξ1 ))c u0 (ξ − ξ1 )dξ1 .
u0 (ξ1 )c
−∞

Using this relationship and the definition of θ in the arithmetic formula (19) pertaining to
the symbol ϕ, there obtains

e−itθ(ξ,ξ1 ) − 1
Z
−1 −2s
u2 (t, x) = γ N ϕ(ξ) exp(ixξ − itϕ(ξ)) dξdξ1 ,
ξ1 ∈I1 ∪I2 ,ξ−ξ1 ∈I1 ∪I2 θ(ξ, ξ1 )

up to a constant. In Fourier transformed variables, and still ignoring constants of no conse-

quence, this amounts to

e−itθ(ξ,ξ1 ) − 1
Z
−1 −2s
Fx7→ξ (u2 )(t, ξ) = γ N exp(−itϕ(ξ))ϕ(ξ) dξ1
ξ1 ∈I1 ∪I2 ,ξ−ξ1 ∈I1 ∪I2 θ(ξ, ξ1 )
(Z Z )
= γ −1 N −2s exp(−itϕ(ξ))ϕ(ξ) ... + ...
A1 (ξ) A2 (ξ)
:= g1 (t, ξ) + g2 (t, ξ)

where n o
A1 (ξ) = ξ1 : ξ1 ∈ I1 , ξ − ξ1 ∈ I1 or ξ1 ∈ I2 , ξ − ξ1 ∈ I2

10
and n o
A2 (ξ) = ξ1 : ξ1 ∈ I1 , ξ − ξ1 ∈ I2 or ξ1 ∈ I2 , ξ − ξ1 ∈ I1 .

Let fj = Fξ7−1
→x (gj ), j = 1, 2. Then, we see that

u2 = f1 + f2

and, due to the support properties of the gj (t, ξ), j = 1, 2, it transpires that

ku2 (t, ·)kH s (R) ≈ kf1 (t, ·)kH s + kf2 (t, ·)kH s . (26)

3.3 Proof of Theorem 5

As we will see momentarily, with a proper choice of γ and N , the main contribution to
the H s (R) norm of u2 (t, ·) arises from f2 . If ξ1 ∈ I1 and ξ − ξ1 ∈ I2 or ξ1 ∈ I2 and ξ − ξ1 ∈ I1 ,
then |ξ1 | ≈ |ξ − ξ1 | ≈ N and |ξ| ≤ 2γ. An examination of the formula for θ in (19) together
with the preceding orders of magnitude reveals that |θ(ξ, ξ1 )| . γ. Let 0 < ε  1 and let
γ = N −ε . With this choice, if ξ1 ∈ I1 and ξ − ξ1 ∈ I2 or ξ1 ∈ I2 and ξ − ξ1 ∈ I1 , one has
 
 e−itθ(ξ,ξ1 ) − 1 
+ it . γt2 ,
 
 θ(ξ, ξ1 )



provided N  1. It follows that

nZ o1/2
−1 −2s
kf2 (t, .)k Hs & |t| γ N |ϕ(ξ)|2 hξi2s γ 2 dξ
|ξ|≈γ
nZ o1/2
& |t| γ −1 N −2s γ |ξ|2 dξ
|ξ|≈γ
3
−1 −2s
& |t| γ N γγ 2
3
= |t| N −2s N − 2 ε .

The way is now prepared to contradict the inequality in (21). Suppose indeed that the
flow map U is C 2 as a mapping of H s (R) to C([0, T ]; H s (R)) for some s < 0 and some T > 0.
Then (21) holds for such values of s and T . Fix a non-zero value of t in [−T, T ]. Then using
(26) and the just derived lower bound for kf2 (t, .)kH s , it is seen that
3
1 & ku0 k2H s & ku2 kXTs ≥ ku2 (t, .)kH s & kf2 (t, .)kH s & N −2s N − 2 ε .

Since s < 0, one can make ku2 (t, ·)kH s larger than any fixed positive real number by first
taking ε > 0 small enough and then choosing N large enough. This contradiction disproves
the validity of (21) and the proof of Theorem 5 is complete. 2

4 Conclusion
In the body of the paper, it has been shown that the BBM-equation is globally well posed in
H s for any s ≥ 0. Moreover, the solutions are remarkably smooth in their temporal variable,

11
but there is not even local smoothing in the spatial variable (see [2]). It was also shown that
the method for establishing this result, namely converting to an integral equation, solving the
integral equation locally in time by a Picard iteration, and then showing the local solution can
be continued by a long-wave–short-wave splitting argument, will not be successful if s < 0.
Indeed, what was shown was that the prospect of using a Picard iteration must necessarily
fail if s < 0.
One would be tempted to then assert that the BBM-equation is ill-posed in H s for negative
values of s. However, recent experience with the Korteweg-de Vries equation suggests one
should be cautious about asserting ill-posedness in the face of a result like Theorem 5. As
mentioned in the introduction, a similar result obtains for the Korteweg-de Vries equation
posed on the real line R, in which the cut-off index is s = − 34 . However, the recent work
of Kappeler and Topalov [17] has shown the initial-value problem to be well posed below
3
H − 4 + , at least for the periodic initial-value problem. (The flow map is continuous, but not
even uniformly continuous on bounded sets in this case.)

References
[1] D. Bekiranov, The initial value problem for the generalized Burger’s equation, Diff. Int.
Eq. 9 (1996), 1253-1256.
[2] T. B. Benjamin, J. L. Bona and J. J. Mahony, Model equations for long waves in
nonlinear dispersive systems, Phil. Trans. Royal Soc. London 272 (1972) 47-78.
[3] J.L. Bona and P.J. Bryant, A mathematical model for long wave generated by a wave-
maker in nonlinear dispersive systems, Proc. Cambridge Philos. Soc. 73 (1973) 391-405.
[4] J.L. Bona, H. Chen, S. Sun and B. Y. Zhang, Comparison of quarter-plane and two-point
boundary-value problems: the BBM-equation, Discrete & Cont. Dynamical Systems,
Series A 13 (2005) 921-940.
[5] J.L. Bona, H. Chen, S. Sun and B.-Y. Zhang, Comparison of quarter-plane and two-point
boundary-value problems: the KdV-equation, Discrete & Cont. Dynamical Systems,
Series B 7 (2007) 465-495.
[6] J. L. Bona, M. Chen and J.-C. Saut, Boussinesq equations and other systems for small
amplitude long waves in nonlinear dispersive media I : Derivation and linear theory, J.
Nonlinear Sci. 12 (2002) 283-318.
[7] J. L. Bona, M. Chen and J.-C. Saut, Boussinesq equations and other systems for small
amplitude long waves in nonlinear dispersive media II : Nonlinear theory, Nonlinearity
17 (2004) 925-952.
[8] J. L. Bona, W. G. Pritchard and L. R. Scott, An evaluation of a model equation for
water waves, Philos. Trans. Royal Soc. London Series A 302 (1981) 457-510.
[9] J. L. Bona and L. R. Scott, The Korteweg - de Vries equation in fractional order Sobolev
spaces, Duke Math. J. 43 (1976) 87-99.

12
[10] J. L. Bona, S. Sun and B.-Y. Zhang, A non-homogeneous boundary-value problem for the
Korteweg-de Vries equation in a quarter plane, Trans. American Math. Soc. 354 (2002)
427-490.
[11] J. Bourgain, Fourier transform restriction phenomena for certain lattice subsets and
application to nonlinear evolution equations I. Schrödinger equations, GAFA 3 (1993),
107–156, II. The KdV equation, GAFA 3 (1993), 209–262.
[12] J. Bourgain, Periodic KdV equation with measures as initial data, Sel. Math. New Ser.
3 (1997) 115-159.
[13] J. Bourgain, Refinements of Strichartz’ inequality and applications to 2D-NLS with crit-
ical nonlinearity, IMRN 5 (1998), 253-283.
[14] M. Christ, J. Colliander and T. Tao Asymptotics, frequency modulation, and low reg-
ularity ill-posedness for canonical defocusing equations, American J. Math. 125 (2003)
1235-1293.
[15] J. Colliander, M. Keel, G. Staffilani, H. Takaoka and T. Tao, Sharp global well-posedness
for KdV and modified KdV on R and T, J. American Math. Soc. 16 (2003) 705-749.
[16] J. Hammack and H. Segur, The Kortweg-de Vries equation and water waves. II. Com-
parison with experiments, J. Fluid Mech 65 (1974) 289-313.
[17] T. Kappeler and P. Topalov, Global well-posedness of KdV in H −1 (T), Duke Math. J.
135 (2006) 327-360.
[18] C. E. Kenig, G. Ponce and L. Vega, A bilinear estimate with applications to the KdV
equations, J. American Math. Soc. 9 (1996) 573-603.
[19] C. E. Kenig, G. Ponce and L. Vega, On the ill-posedness of some canonical dispersive
equations, Duke Math. J. 106 (2001) 617-633.
[20] L. Molinet, J.-C. Saut and N. Tzvetkov, Well-posedness and ill-posedness results for the
Kadomtsev-Petviashvili-I equation, Duke Math. J. 115 (2002) 353-384.
[21] G. Pólya and G. Szegö, Problems and theorems in analysis, Volume 1, Springer-Verlag,
1972.
[22] N. Tzvetkov, Remark on the local ill-posedness for KdV equation, C. R. Acad. Sci. Paris
329 (1999) 1043-1047.
[23] B.-Y. Zhang, Taylor series expansion for solutions of the KdV equation with respect to
their initial values, J. Funct. Anal. 129 (1995) 293-324.
[24] B.-Y. Zhang, Analyticity of solutions of the generalized KdV equation with respect to
their initial values, SIAM J. Math. Anal. 26 (1995) 1488-1513.
[25] B. Y. Zhang, A remark on the Cauchy problem for the KdV equation on a periodic
domain, Diff. Int. Eq. 8 (1995) 1191-1204.

13
[26] N. J. Zabusky and C. Galvin, Shallow-water waves, the Korteweg-de Vries equation and
solitons, J. Fluid Mech. 47 (1971) 811-824.

University of Illinois at Chicago

E-mail address : bona@math.uic.edu

Université de Lille
E-mail address : nikolay.tzvetkov@math.univ-lille1.fr

14