PACIFIC JOURNAL OF MATHEMATICS

Vol. 108, No. 1, 1983

ANALYTIC LINEARIZATION OF THE

KORTEWEG-DE VRIES EQUATION
E. TAFLIN

We prove that the KdV equation is linearized by an analytic

function, which is projectively analytically invertible. The Cauchy prob-
lem for the KdV equation is entirely solved by this fact. The non-linear
superposition principle is a trivial consequence of convexity for the
image of the linearization operator.

1. Introduction. Since the discovery [7] of the inverse scattering

formalism for the KdV equation

(1.1) -^u{t, x) + J^u(t, *) ~ 6u(t, x)-^u(tx) = 0,

t, x, u(t, x) E R,
it is known, given a certain class of solution for the linear equation

how to construct solutions of equation (1.1). However, it is not clear how

this reduces the Cauchy problem for the KdV equation, on a given space
of initial conditions, into that of the above linear equation. The Cauchy
problem for (1.1) has been solved by direct functional analysis methods
on the Sobolev space if3, ([11]), and on H2, ([1]). The inverse scattering
formalism has been used to solve the Cauchy problem on S'(R), ([12]) and
for sufficiently rapidly decreasing C3 initial conditions ([3]), where in both
cases, two linear problems are associated with the KdV equation.

To formulate the problem of linearization of the non linear Cauchy

problem, it is convenient to give topological vectorspaces of initial condi-
tions and of solutions for the non linear resp. for the linear problem. In
this context it is possible to give a precise meaning to the concept of
linearization (see [5]). What we want to show in this paper is that the
linearization program defined in [6], entirely goes through and solves the
initial value problem for the KdV equation. We stress the fact that this
approach is straightforward (in contrast to the inverse scattering for-
malism), when the spaces of initial conditions are given. The inverse

203
204 E. TAFLIN

scattering formalism is recovered (c.f. [8]). This is not surprising as, with
the particular choice of spaces we have done, the linearization is unique
(for fixed first term).

The choice of spaces is of course important for the properties of the

linearization mapping (if it exists). It is easy to illustrate this fact for the
KdV equation. Let Ex and E2 be two TVS of functions /: R -» R and F:
Ex -> E2 a C1, one to one map, which linearizes the C° vector fields on Ex:
X0(u) = 3w, Xx{u) = -33w

i.e. F is a translation invariant linearization of the Kdv equation. Denoting

D the Frechet derivative, this means that
DF.X0 = X*oF, DF.XX = X{ o F9
where X\(u) = -33w and -XQ(M) = XQ(u). Let w0 7^ 0 be the initial condi-
tion for a soliton, i.e.
kX0(u0) = ^(WQ) f° r some fc < 0.
Then
0 - DF. (kX0 - Xx)(u0) = (kX* - X}) o F(u0)
i.e. (kd + d3)v0 = 0 for v0 = F(w0)

The non-constant solutions of this equation in ^D are exponentials.

Thus if F is one to one and is defined on the solitons then E2 has to
contain exponentials. Further if F~x exists as a C1 function on some open
set O 3 0 in E2 then DF(0) and Z^O)" 1 are continuous linear mappings.
So Ex and E2 are topologically isomorphic. One can then as well choose
Ex = E2 = £. Suppose that C°°(R) D £ D S(R) (the Schwartz space of
test functions) and that E contains at least the exponentials eax, a > 0. If
F is C00, one finds by calculating (in a very formal way) some of the
derivatives Dn(F~x), (see [8] for explicit formulas) that the operator/ -» 3 /
has to be invertible on E9 i.e. f(x) -> 0 fast as x -> -00.

These qualitative remarks show that it is a quite natural choice to

treat the linearization problem for the KdV equation on a space like the
TVS of all functions/ e C°°(R) for which the seminorms

= sup
 KORTEWEG-DE VRIES EQUATION 205

are finite. On this space, denoted Sb, we prove (Corollary 3.3 and
Proposition 3.4) that there is a unique translation invariant formal lineari-
zation A (taking the non linear equation into the linear) of the KdV
equation (in the sense of [5]), i.e. the linearization operator exists as a
formal power series where each term is finite and commute with the space
translations. A converges to an entire analytic function A: Sb -> Sb
(Corollary 4.2). The power series A'x defines a C°° mapping A'1: A[Sh] ->
Sb, which is "projectively analytic" (Corollary 4.4 and Proposition 4.6).
The set A[Sb] is convex, which gives the non-linear superposition principle
for solutions of the KdV equation. The Cauchy problem for the KdV
equation can now be solved entirely by linearization (Proposition 5.2 and
Remark 5.3).
2. Formal linearization. Let us first introduce some necessary nota-
tion. Given two TVS, X and Y. Denote F(X9 Y) (resp. FS(X, Y)) the space
of formal power series from X to Y of the form / = 2 n > i / n , where
/ " E &n(X, Y) (resp. £sn(X, Y)), the space of n-linear continuous (resp.
continuous symmetric) mappings from X to Y. fs E FS(X, Y) will denote
the symmetrization of f E F(X, Y). If X, Y and Z are TVS then the
product
FS(Y, Z) X FS(X9 Y) 3 (A, B)^A*B E FS(X9 Z)
is defined by (see [5]):

A * B = 2 [ 2 A'I 2 Iq® B - ^ ® I p \ o \
y
n>\ \\<p<n 0<q<p~\ ' I

where Iq is the identity mapping on X ®s • • • ®sX (q-times) and on is the

normalized symmetrization mapping on ® n X. ®s is the symmetric
projective tensor product. F(X, X) is denoted F(X), etc. If A, B E FS{X)
then the bracket [A, B]* = A * B — B * A is defined.

REMARK 2.1. If A, B E Fs{X) define entire functions A, B: X -> X,

where

A(u) = 1 An(u), An{u)=An{u,...,u),

then
J [ i , B] =[A, B] + .

Here D is the Frechet derivative and [A, B] = DA.B - DB.A the usual
vector field bracket.
206 E. TAFLIN

We denote by Sb the Frechet space of all C°° functions / from R to R

for which II / ((N < oo, N = 0,1, The seminorms in Sb are given by

= sup

Sh is the projective limit of the spaces S(n), n = 0,1,... where £(/?) is the
subset of a l l / e C°°(]-oo, /*]) such that

sup |(l+M)*aa/(x)|<oo, £ = 0,1,...,

xE]-oo,«]
0fSa<A:

PROPOSITION 2.1. 77*e Frechet space Sb is a nuclear Montel space.

Proof. Sb is barreled as it is Frechet. Let B C Sh be a bounded closed

subset, i.e. || 51| N = sup /eB || / 1 | ^ < oo for each TV > 0. It follows from the
Ascoli-Arzela theorem that if | | 5 | | ^ + 1 < o o , then for each sequence
{fn}™=o i n & there exists a subsequence {/• }^ =0 such that/ converges in
the || ||N norm. By a diagonalisation argument there is then a subsequence
{fjH}™=o of {/JJLo which converges in the norm || ||N for each N >0. This
proves that Sh is a Montel space.
That Sh is a nuclear space from the fact that S is nuclear. In fact since
S is nuclear, the space S(n) = S/2sn, « > 0, is nuclear (c.f. [9] Theorem
III.7.4), where En is the closed subspace of functions / G S such that
supp/ C [«, oo[. S^, being the projective limit limS(n), is then nuclear (cf.
[9] corollary of Theorem III.7.4). *" •

Let EN be the closed subspace of S" of distributions with support in

]-oo, N].

PROPOSITION2.2. 5*^ (the strong dual of Sh) is a nuclear Montel space,

and is isomorphic to the strict inductive limit of

Proof. As S is dense in Sb9 S'h C 5" (set theoretically). Further if

T G 5" has a continuous extension to Sh then supp T C ]-oo, N] for some
N > 0. Each semi norm in S'b has the form Sb 3 F -» qB(F) =
su
P«e51 ^( w ) I > where 5 is a bounded set in 5^. The inductive limit of EN
is then identic with Sb as topological vectorspace. Hence Sb is nuclear as
EN is nuclear ([9] Theorem III.7.4 and corollary). Further Sb is Montel as
Sh is Montel (cf. [2], IV, §3, Prop. 7). •
 KORTEWEG-DE VRIES EQUATION 207

3. Formal linearization of the XDF-equation. Let t2 be the two-di-

mensional commutative Lie algebra. The ATdF-equation is defined in the
representation

(3.1) t2 3 (a, b)^aT0 + bTx = T(a,b) G Fs(Sb),

where
T0=T0\ T^T' + T2, Tox(u) = du,
r/(w) = -33w and T?(ul9 u2) = 3(11,8112 + 11,3M,),

for u, ul9 u2 G Sb. The representation is formally linearizable [5] on Sh

into Sb if there exists an element C E F(Sb)9 with C 1 continuously
invertible and
(3.2) TxC=C*Tk, VXGt2.
(C takes then the space where the linear equation is defined into the space
where the non-linear equation is defined.) The «th order of equation (3.2)
reads

(3.3) TxxCn-
\ -I

For X = (1,0) we get explicitly

(3.4) dCn(cp2 9 - -

LEMMA 3.1. / / Cn G £*(Sft) ««J C" satisfies (3.4)

Fn G 5XR"), symmetric with supp Fw C AT"]-oo, A:] for some k > 0

(3.5) C n ( 9 , ® • • • ®<pj = [ F n ® ( 9 ,

where in(x) = (x,...,x) (n-times), denotes space inversion and ® con-

volution. Conversely given any Fn G S'(Rn) with suppF" C AT"]-oo, k],
k > 0, /Ae« Cw ^ constructed satisfies equation (3.4).

Proof. If Cn G £*(£*), then

208 E. TAFLIN

defines a symmetric element Fn E S\W) and s u p p F " C X w ]-oo, k] for

some k > 0 (Proposition 2.2). Let T^ denote the operator of translation by
a on Sb. Then (3.4) gives
raC"(<px ® • • • ®%) = / " • ( T ^ , ® • • • ®ra<pn) Va E R, <pt E 5 6 .
Hence

The converse is obvious. D

We next prove that a cohomology space H°(t2, tsn(Sb)) = 0.

PROPOSITION 3.2. For n>2, the equation

T\Cn - Cn * Txx = 0, V x G r2
/*as rAe wn/^we solution Cn = 0 w C ^ ^ ) .

Proof. By Lemma 3.1 each solution Cw has the form

2
C ^ ® ••• ®9 n ) = [ J F » ® ( V l ® . . . ®(pj] o | n (Fsymmetric)

Take X = (0,1). Then, we get

- B W ^ ® • • • ®yn) + 2 Cn(<px ® • • • ®33<Pi ® • • • ®v«) = 0.

Thus

After Fourier transformation1

PnF" = 0 , P B ( * , , . . . , * J = ( ^ + • • • +knf - { k f + --- +k3n),

where F" is analytic in the domain TT+ = {& | Re &,- > 0,1 < i < w}.
(Proposition 2.2). But Pn ^ 0 for n > 2 and the ring of analytic functions
on TT" is an integral domain so Fn = 0 is the only solution. •

COROLLARY 3.3. Equation (3.2) /JOS a£ most one solution for a given C 1 .

1
Fourier transformation in S(W) is defined by f(k) = (27r)~"/2/R« dx e"ikxf(x).
 KORTEWEG-DE VRIES EQUATION 209

To solve equation (3.3) we introduce the holomorphic functions G":

(3.6) Gl(k) = ik(2irYV2, G\k2, k2) = (Iv)'1,

G"(kx,...,kn)

J v
(k + k ) •- (k + k r
K K K K
\ \ * 2) \ n-\ ' n)
We introduce also the holomorphic function Fn\ TT" -> C,

where tyn is the group of permutations of n elements. The holomorphic

functions Gn (resp. Fn) define uniquely (cf. [10]), by Fourier-Laplace
transformation, distributions Gn (resp. Fn) E S\W) with suppG" (resp.
Fn) E xn]-oo, k] for some k > 0. Lemma 3.1 can now be applied to
construct an element C E F(Sh) (resp. Cs E Fs(Sb)) by the distributions
Gn (resp. Fn). The algebraic expressions for Gn coincide with them in [8].

PROPOSITION 3.4. Cs E Fs(Sb) constructed by (3.5) and (3.6) is the

unique translation invariant symmetric formal linearization on Sb of the
KdV-equation, i.e. TXCS = Cs * T\for each X E t2, with Cx = 9.

. C, is unique if it exists (Corollary 3.3). By (3.3) and (3.5) the

linearization of T by Cs is equivalent to

= 3 2 (3i + ••• +dtt)(Fp ® F"-') o on, n>2.

This is by Fourier-Laplace transformation equivalent to

+kn)3 - [k] +

2
1 <p<n- 1
i/
w > 2.
w
is a solution of this equation if

> 2.
210 E. TAFLIN

Direct substitution as in [8] proves that Gn defined by (3.6) satisfies this

equation. •

The inverse Fourier transformation G" of Gn, defined by

Gn(<px ® • • • ®(pn) = Gn((pl ® • • • ®<p/f), (?!,.. .,<pn E S(R),
is explicitly

•^-oo -oo

• • • w«-i(j n - 2 + yn-\)un{yn-x)-
Let t/ be the Frechet space of all functions / G C°°(R X R~,R) for
which the seninorms Mn are finite:

(3.8) Mn{f)= sup \d«f(x,y)\+ sup |(l +\x + y\)ndaf(x, y)\,

/i = 0 , 1 , . . . . Here da = 8 fl3£2 and | a | = ax = a 2 . We introduce for M 6 S 6

the commonly used (see [2], cf. [8], [12]) continuous integral operator
Q(II): U->U:

(3.9) [0(«)>](JC, ^) = / ° ^ + J + t)f(x9 t)dt,

(3.5), (3.7) and (3.9) give the following explicit expression for Cn in
Proposition 3.4:

where
(3.11) B"(q>x ® • •
=
[<*(%)](*> y) <Pn(x+y)>
(3.9) and (3.11) give
(3.12) i(<p) + S2((p)5((p) + a(<p) - 0,

The inverse A = C"1 is needed for solving the Cauchy problem for
the 1&/K equation. As is seen directly from (3.12), the power series B is
easily inverted (on its image):

If B(<j>) = ip, then <j> = P ( ^ ) , where

(3.13)
 KORTEWEG-DE VRIES EQUATION 211

and
(3.14) [<$>(f)u](x)=(° f(x,t)u(x + t)dt
J
-oo
defines a continuous integral operator %(f): Sb -» Sh for each/ E (7.

It seems difficult to invert C directly. However there is a unique

power series Qs E Fs(Sh,U) such that <25 ° C5 = Bs. To find the inverse of
C we first construct a <2 E F{Sb,U). The expressions (3.10) and (3.11)
give

Multiplication with 0(<p) gives by (3.11)

— "S1
nx+

Finally by (3.10) and (3.11)

A lengthy but straight forward calculation gives that

Q»(dq>) = &N-l(d<p)a{q>), iV = 1 , 2
where

(3.15) [&(u)A(x,y) =

defines for each u E Sb a continuous operator U -* U. (The operator $ is

well-known, cf. [4].) It follows then that
(3.16) QN(<P) = ^(.pf-ya-H iv= 1,2
212 E. TAFLIN

and that Q(<f>) satisfies the integral equation

(3.17)

4. Convergence properties of the formal power series. We study in

this paragraph the existence and the properties of the functions defined by
the formal power series Qy P, B and by them A and A~\ Q, P, B are
uniquely defined by the integral equations (3.17), (3.13) resp. (3.12):
(4.1) Q(q>) - &(<p)Q(<p) = fl(a», Q G F(Sb9 U), <p G Sb9
(4.2) p(f) + ®(f)p(f) = -/(.,o), peF(u,sb),feu9
(4.3)

PROPOSITION 4.1. The formal power series Q (resp. P) converges to an

entire function Q: Sh-+ U (resp. P: U -* Sb). g(resp. P) is the unique
solution o/(4.1) (resp. (4.2)).

Proof. Using that

[dx&(<p)f\(x9 y) =f°dyl<p(x+y-yl)f(x+y-yl, yx)

and that

[(3, - 92)<£(<p)/](x, y) - f dxMxi)f(xx, y)

one deduces the existence of seminorms />0 < / j j < • • • for each given
K > 0 such for ax > 1:

(4.4) s u p \d"&{V)f\(x9 y) <pN(<p) sup \d*f(xl9 y})\,

\a\<N \a\<N

V(x,j)G]-oo,^] XR-,
and for ax — 0:

(4.5) sup \d?&(<p)f\(x,y)<pN(<p) sup \d"f(xl9 yx)\9

,X] XR-,
 KORTEWEG-DE VRIES EQUATION 213

where E(x, y) = {(x,, yx) E R X R"| xx + yx < x + y). We have here

used the inequality (cf. [12])

(4.6) |fl( 9 )/|(*, y) < */* df(l + ;2)|<p(0| sup \f(xl9 yx)\.

Denote

The explicit form of <£(<£) and (4.4), (4.5) (4.6) give for M > N + 1 and
(x, y) e ]-oo, K\ X R':

(4.7) sup \r&M-\<p)a{r\)\{x, y)

\\N
(hMM~l'N sup |3

Finally we find that (see (3.8), (3.16))

(4.8) MN(Q»(<p)) ^

where n>N+ 1. pN has (for convenience) been chosen sufficiently large

and K>N. (4.8) proves that the series Q(<t>) converges for each <>
| E Sh.
Hence Q: Sh-+ U is entire analytic.

To prove the second statement, we first remark that

(4.9) [®(+)<P](X) = f W(x, t - x)q>(t)

•'-oo

gives
(4.10) [3®(^)<p](x) = +(x,0)<p(x) + ®((9, -
Formula (4.9) and (4.10) give the estimate, for « > ]V + 1,

t<x

where qN is a sufficiently large seminorm on U. Then for pN sufficiently

large
214 E. TAFLIN

Thus, once more if pN is sufficiently large

(4.12) ||P»(^)|U<^(^)"[(«-l-iV)!]- 1 ,
which proves that P is entire analytic. •

COROLLARY 4.2. Denote A = C'x E F(Sh). Then A is an entire ana-

lytic function.

Proof. A = P © Q is entire analytic by Proposition 4.1. D

We now turn to the question of the existence of A~l. We prove the

convergence of C — A~l and find the maximal connected domain of
analyticity (containing u = 0) for C.

Introduce the operator (see [4]) tix(u): L 2 (]-oo,0]) -> L 2 (]-oo,0]) for

(4.13) \Qx(u)h](y) = / u(x +y + t)h(t)dt.

J
-oo

&x(u) is a self-adjoint Hilbert-Schmidt operator with

SU
(4.14) ||^(W)||H.S.- (f) P 10 + t2)u{x + t)\.

Let o(£lx(u)) be the spectrum of &x(u), let U(n), n — 0 , 1 , . . . , be the

factorspace of £/ and the closed subspace of functions / with supp / C
|rc, oo[ XR". B E i ^ , U) defines an element in F(S(n), U{n)) which
also will be denoted B. The seminorms on U(n) induced by the complete
set of seminorms (3.8) on U will be denoted M^, N — 0,1,

PROPOSITION 4.3. For each n E N , B defines an analytic function B:

OnZ) S(n) -> [/(«), where On is the set ofall u E S(n) such that® x(u) > - 7 ,
Vx E ]-oo, w]. Ow w open in S(n) and

(4.15) [B(u)](x, y) = -[(I + Qx(u)r*[a(u)](x, • )](y).

(Analytic means here convergence on a neighbourhood of each point in On).

Proof. Let u E On. Then by estimate (4.14) there exists an neighbour-

hood V of u such that, for each v E V and x G ]-oo, 4 <*(®x(v)) c
[-1 + c, oo[ for some 0 < e. Hence Ow is open in S(n).
 KORTEWEG-DE VRIES EQUATION 215

It follows from Lemma III.9 in [12], (where it is only used that

o(tix(u)) C [-1 + e, oo[ for each x G ]-oo, n]), that B(u) G U(n) and
that [ / + Qx(u)]'lf(x9 •) defines an element in U(n) for each/G U{n).
Let u, u + <>f G On. Then at the point u + <J>, the r.h.s. of (4.15) can be
written

We prove the convergence of the series

(4.16) 2 (-l

in U(n) for all ^» in some S(n) neighbourhood of u. Let <J>,,.. .,<|>i_1 S

S(n), ux...,uk^On, k>2, / £ [ / ( « ) and define the functions Fk:
S(n)k'1 X Ok X t/(n).-» t/(«):
(4.17) -F*(q>,,.-.,9>*-i; "i "*-i; / )

The following estimates are obtained directly from formula (3.9) and the
identity - ( / + Q ( « ) ) - ' / - Q(«K^ + Q(«))"'/ + / = 0, u E On, f E U(n):

(4.18) l n
() |/(x, y)|, V<p G 5 ( « ) , / G £/(«), iV G N,

where/70 < • • • < 77^ < • • • are sufficiently large seminorms on S(n) and

(4.19) | / ( X , ^ ) - [ ( / + Q(«))-I>](JC,^)|

+ Qx{u)Y\t- sup |(1 - tfu(x +y + t)\

0

(4.19) gives:

(4.20) sup \[(l + a(u)YXf](x, y ) \ < C M sup |(1 -

216 E. TAFLIN

where (for given n) Cu is a positive constant depending on M G O r (4.18)

and (4.20) give for some sufficiently large seminorm q on S(n):

(4.21) sup \[

\f(x, y)\, Vw e On, <p G S(n),fE U{n).

By (4.18) (4.21) and then (4.20) an estimation for Fk is obtained:

(4.22)

-tff{x,t)\.

The identity (for a given x G ]-oc, n]) for e < n — x

- (I + Qx+e(u))~l - tix+£(u){l + fi^^w))-1 + 7 = 0

gives

(4.23) ~J~{l + ^x + e(W)) |e = 0 — (^ + ^x( W )) ^JC(9W)(/ + Qx(u))~

(This is well-defined in U(n), cf. [9].)

Explicit expressions for the derivatives of Fk((p; u; f) are obtained

from (4.17), (3.9) and (4.23):

(4.24) d}Fk((pl9...,?>*_,; M,,...,!/^,; /)

and

(4.25)
 KORTEWEG-DE VRIES EQUATION 217

(4.22), (4.24) and (4.25) give the estimate for k > N + 1 {qN is a suffi-
ciently large seminorm):
(4.26) MN{Fk(<p,...,q>;u,...,u;f)

X sup \(l-t)2dV(x,t)\q(q>) k-N-\

t<0
0<a<N
Thus, for a given u E On the series
2 Fk(<p9...,<p; u9...9u; f )

converges for all / E U{n) and all $ <E S(n) with # ( 0 ) C W < 1 . / =
( / + S2(w))~1a(w + <£) is linear (and continuous) in <j>, which proves that
the series (4.16) converges on a neighbourhood of u in S(n). •

COROLLARY 4.4. The mapping On B u H> [ 3 , 5 ( I / ) ] ( - , o ) = £{u) E

A.

^(w) is analytic and this mapping composed with A is i d o .

Proof. The first statement is a trivial consequence of Proposition 4.3.

Secondly denote F: On -* S(n) the composite map A ° C. F is analytic
(Corollary 4.2 and Proposition 4.3) and the formal power series F on S(n)
is the identity (§3). •

REMARK 4.5. Let [ul}%x C On be a convergent sequence in 5(«) with

limit v9 such that - 1 G o(tix(v)) for some x G ]-oo, n\. Then by (4.15)
5(w z ) will likely (if there is no cancellation) develop singularities as
/ -> oo, and u would not be in A[S(n)]. Further if for u E S(n), - 1 £
a(fix(w)) for i E ]-oo, n\ then by the continuity of x -^ fix(w), a(i2x(w))
C [-1 + e, oo[ for each x E ]-oo, n] for some e > 0. Thus by Corollary
4.4. it seems reasonable to try to prove that S2x(w) > - / for each u E
A[S(n)] and x E ]-oo, 4 The result is well-known, (see [4]).

PROPOSITION 4.6. The image A[Sb] is exactly the subset O C Sh, where
O is the projective limit of On, n E N , i.e. u E O iff
(4.27) 2x(u)>I VxER.

Proof. For given x E R choose n > x. Then as A: Sh -^ Sh defines a

mapping^: S(n) -> S(n)
[A(u)](x)=[A(en+lu)](x)9
218 E. TAFLIN

where 0W+1 G C°°(R). dn+Y(t) = 1 for t G ]-oo, H] and 0n+x(t) = 0 for

/ > w + 1. But 0u G S(R), so Qx(A(0u)) > - / , VJC G ]-OO, n], (see [4], cf.
[12]). •

REMARK 4.7. The condition (4.27) resembles to that obtained in

linearizing Burger's equation in [13].

5. Solution of the non-linear initial value problem and superposition

of solutions. Corollaries 4.2 and 4.4 reduce the solution of the KdV
equation in Sb:

(5.1) ftu(t) = -d3u(t) + 6u(t)du(t), "(0) = uo£Sh

to the linear equation

(5.2) ftv(t) =- 3 M 0 , t>(0) = vo = A(u0) G 0

together with the condition in Proposition 4.6. But before stating formally
the result we prove the following lemma:

LEMMA 5.1. Let Vbe a n-dimensional C°° differentiable manifold and F:

V -> O a C°° mapping then
f = A~l o F: V-> ShisC°°.

Proof. The mapping F 9 ^ H [Ftf)] [^^ G On is C°° and A~x\ On ->

S(n) is analytic (Corollary 4.4). Then the mapping V 3 £ H>
[/?„ o i " 1 o /rj(^) e £(„) i s c 0 0 , where /?w: Sh -> 5(/i) is the canonical
projection. Thus K B ^ [ i " 1 o F](£) G 5^ is C°° as 5fr = Iim5(n). •

PROPOSITION 5.2. Let u0 G Sb. Then equation (5.1) has a C 1 solution

for t G Jo,, a 2 [, -oo < «j < 0 < a 2 ^ oo, iff r/ze equation (5.2) wif/i t;0 =
y4(w0) Aa^ ^ C 1 solution v(t) G O /or r G ]al9 a2[. On the interval of
existence u(t) = A~\v(t)) and t\-*u(t) is C 00 .

REMARK. Instead of C 1 solutions one can write C° solution of the

integrated equations.

Proof. vQ G O (Proposition 4.6). There exists a maximal interval

]bl9 b2[9 -oo < Z>! < 0 < 62 < oo, where equation (5.2) has a C 1 solution
in S^. Hence the solution u(t) = A~\v(t)) of equation (5.1) exists on a
maximal interval t G ]a}9 a2[, bx < ^ < 0 < a 2 < Z?2? on which t;(^) G O.
 KORTEWEG-DE VRIES EQUATION 219

t\->v(t) is C00 if it is C1. Lemma 5.1 gives now the "if part" of the
proposition. Conversely if u{t) is a C1-solution on a maximal interval of
existence }av a2[ then by Corollary (4.2) and Proposition 4.6 A{u(t)) E O
is a C1 solution of equation 5.2 on ]a l5 a 2 [. Further t H>A(u(t)) is then
C°°, so by Lemma 5.1 u(t) is C00. •

REMARK 5.3.
(a) u{t) can blow up in finite time (t = a) for two reasons:
(i) v(t) E does not converge in Sb as t -» #;
(ii) t;(a) 3 in Sft but t>(a) £ O.
(b) If u0 E S(R) then u(t) E S(R) V/ E R as is seen from space-time
inversion of equation 5.1 on S(R) and linearization by A. A(S(R)) is
invariant under the linear evolution, [12].

The non-linear superposition principle is a trivial consequence of the

following Corollary:

COROLLARY 5.4. O is a convex set.

Proof. See Proposition 4.6.

Acknowledgement. I want to thank Professor M. Flato for helpful

suggestions and constructive criticism.

REFERENCES

[1] J. L. Bona and R. Smith, The initial value problem for the Korteweg-de Vries equation,
Phil. Trans. Royal Soc. London, (A) 278 (1975), 555-604.
[2] N. Bourbaki, Espaces vectoriels topologiques, Elements de mathematique, Fascicule
XVIII, Livre V, Hermann Paris 1973.
[3] A. Cohen, Existence and regularity for solutions of the Korteweg-de Vries equation,
Arch. Rat. Mech. Anal., 71 (1979), 143-175.
[4] L. D. Faddeev, Properties of the S-matrix of the one-dimensional Schrodinger equa-
tion, Trudy. Matem. in-ta im. V.A. Steklova, 73 (1964), 314-336.
[5] M. Flato, G. Pinczon and J. Simon, Non-linear representations of Lie groups, Ann.
Scient. Ecole Norm. Sup. Paris, 10 (1977), 405-418.
[6] M. Flato and J. Simon, On a linearization program of non-linear field equations, Phys.
Lett., 94B( 1980), 518-522.
[7] C. S. Gardner, J. M. Green, M. D. Kurskal and R. M. Miura, Method for solving the
Koretweg-de Vries equation, Phys. Rev. Lett., 19 (1967), 1095-1097.
[8] R. R. Rosales, Exact solutions of some non-linear evolution equations, Stud. Appl.
Math., 59 (1978), 117-151.
[9] H. H. Schaefer, Topological Vector Spaces, Springer-Verlag, New-York, Heidelberg,
Berlin 1980.
220 E. TAFLIN

[10] L. Schwartz, Theorie des Dsitributions, Hermann 1973,

[11] A. Sjoberg, On the Korteweg-de Vries equation, J. Math. Anal. Appl., 29 (1970),
569-579.
[12] E. Taflin, Dynamical symmetries and conservation laws for the Kortweg-de Vries
equation, to appear in Rep. Math. Phys.
[13] , Analytic linearization, Hamiltonian formalism and infinite sequences of con-
stants of motion for Burger's equation, Phys. Rev. Lett., 47 (1981), 1425.

Received January 15, 1982. Supported by the Swiss National Science Foundation.

UNIVERSITY DE GENEVE
1211 GENfevE 4, SWITZERLAND