Você está na página 1de 6

REFERENCES

1. G. Birkoff and S. MacLane, A Survey of Modern Algebra, A K Peters, Wellesley, MA, 1996.
FaireldUniversity, Faireld, CT 06824
cbernhardt@mail.faireld.edu
A New Constructive Proof of the
Malgrange-Ehrenpreis Theorem
Peter Wagner
1. INTRODUCTIONANDNOTATION. The Malgrange-Ehrenpreis theoremstates
that every (not identically vanishing) partial differential operator with constant coef-
cients possesses a fundamental solution in the space of distributions, i.e.,
P() C[
1
, . . . ,
n
] \ {0} : E D

(R
n
) : P()E = ,
see [1, Thm. 6, p. 892], [6, Thm. 1, p. 288].
This theorem was a rst impressive piece of evidence of the impact of distribution
theory in its application to linear partial differential equations, and several different
proofs of it were therefore devised in the subsequent years, see [3, Ch. X, and the
historic notes in pp. 58, 59]. An elementary proof based on L
2
-theory was presented in
[9]. For more details, we refer to the surveys in [7], [8], [13].
Already very early, constructive proofs were also found, i.e., proofs that represent
a fundamental solution E by a formula, instead of inferring the existence of E from
the HahnBanach theorem as did the original proofs. The prototype of such a formula
is H ormanders staircase, which employs partitions of unity, [11], [7, (1), p. 345],
[8, Thm. 2.3, p. 107]. While these partitions of unity are based on the location of
the zeroes of P(z), z C
n
, and the formula is not completely explicit in this sense,
H. K onig published in 1994 a proof using a representation of a fundamental solution
by an n-fold integral with respect to of the inverse Fourier transforms F
1
( f

) of the
modulus-one functions f

() = P(i + )/P(i + ), see [5], [8, Thm. 2.5, p. 111].


A simpler formula involving a one-fold integral was given in [7, (2), p. 346].
In this paper, we shall further simplify this procedure by constructing a fundamental
solution as a sum over nitely many distributions F
1
( f

j
), j = 0, . . . , m = deg P,
see Proposition 1. In a second part, we shall then generalize this formula so as to
obtain a constructive proof of the existence of a fundamental solution for linear par-
tial differential-difference equations, i.e., convolution equations with kernels of nite
support, see Proposition 2.
Let us introduce next some notation. The letter n is reserved for the dimension
of the underlying vector space R
n
. The inner product on R
n
is indicated by jux-
taposition, i.e., x =
1
x
1
+ +
n
x
n
. In this paper, all partial differential opera-
tors are linear and contain constant complex coefcients, i.e., they belong to C[] =
C[
1
, . . . ,
n
], and we adopt for them the usual multi-index notation. In particular, we
May 2009] NOTES 457
set
i
= /x
i
,

1
1
. . .

n
n
, and || =
1
+ +
n
for a multi-index N
n
0
. If
P() =

||m
c

is an operator of degree m, i.e., deg P = m, then P


m
denotes its
principal part: P
m
() =

||=m
c

.
For our notation concerning distributions, we refer to [4] and [10]. We remind
the reader that E(R
n
) = C

(R
n
) is the space of C

-functions on R
n
, and D(R
n
)
and S(R
n
) denote the subspaces of functions which have compact support, or are,
together with all derivatives, rapidly decreasing at innity, respectively. The inclu-
sions D(R
n
) S(R
n
) E(R
n
) imply D

(R
n
) S

(R
n
) E

(R
n
) for their duals, the
spaces of distributions, temperate distributions, and distributions with compact sup-
port, respectively. Furthermore, denotes the Dirac measure at the origin,
x
0
that at
the point x
0
, and

T the reection at the origin of a distribution T. For the Fourier
transformation, we adopt the following notation:
F : S(R
n
) S(R
n
) :
_
R
n
e
i x
() d.
Either by duality or by density, this yields the isomorphism
F : S

(R
n
) S

(R
n
) : T ( F, T).
(Note that the angle brackets represent the evaluation of a distribution (here T) on a test
function (here F).) The Fourier inversion theorem states that F
1
T = (2)
n
(FT)
for T S

(R
n
). Sometimes we write F
1

to indicate the variable (here ) with re-


spect to which the inverse Fourier transform is performed. For a distribution T with
compact support, i.e., T E

(R
n
), we dene the Laplace transformLT by (LT)(z) =
e
zx
, T
x
, z C
n
. Of course, in this case, LT and FT(z) = LT(iz) are entire func-
tions.
Let us nally employ the above notation to collect some elementary formulas we
shall take for granted below. They follow from the classical analogues either by duality
or by density arguments. For all C
n
, T D

(R
n
), S S

(R
n
), U E

(R
n
), the
following equations hold in D

(R
n
) :
P()(e
x
T) = e
x
_
P( + )T
_
;
P()F
1
S = F
1

_
P(i)S
_
;
(e
x
U) (e
x
T) = e
x
(U T).
2. THE MALGRANGE-EHRENPREIS THEOREM.
Lemma 1. If
0
, . . . ,
m
C are pairwise different, then the unique solution of the
linear system of equations
m

j =0
a
j

k
j
=
_
0, if k = 0, . . . , m 1,
1, if k = m,
is given by a
j
=

m
k=0,k=j
(
j

k
)
1
.
Proof. Since Vandermondes determinant does not vanish for pairwise different
j
,
the vector (a
0
, . . . , a
m
) C
m+1
is in fact uniquely determined. Furthermore, if p(z) =

m
j =0
(z
j
), then the residue theorem implies that
458 c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 116
m

j =0
p

(
j
)
1

k
j
=
1
2i
lim
N
_
|z|=N
z
k
p(z)
dz =
_
0, if k = 0, . . . , m 1,
1, if k = m,
On the other hand, p

(
j
) =

m
k=0,k=j
(
j

k
), and this completes the proof.
Proposition 1. Let P() =

||m
c

C[] \ {0} be a not identically vanishing


polynomial on R
n
of degree m. If R
n
with P
m
() = 0, the real numbers
0
, . . . ,
m
are pairwise different, and a
j
=

m
k=0,k=j
(
j

k
)
1
, then
E =
1
P
m
(2)
m

j =0
a
j
e

j
x
F
1

_
P(i +
j
)
P(i +
j
)
_
is a fundamental solution of P(), i.e., P()E = .
Proof. (a) Let us rst observe that, for R xed, N = { R
n
: P(i + ) =
0} is a set of Lebesgue measure zero. In fact, after a linear change of coordinates,
we can assume that P
m
(1, 0, . . . , 0) = 0, and then
_
N
d =
_
R
n1
__
N

d
1
_
d

= 0 by
Fubinis theorem and since the sets N

= {
1
R : P
_
i(
1
,

) +
_
= 0} are nite
for

= (
2
, . . . ,
n
) R
n1
. Hence
S() =
P(i + )
P(i + )
L

(R
n
) S

(R
n
)
and the formula in the proposition is meaningful.
(b) For S S

(R
n
) and C
n
, we have
P()(e
x
F
1
S) = e
x
P( + )F
1
S = e
x
F
1

_
P(i + )S
_
.
Taking S = P(i + )/P(i + ) with R, this implies
P()
_
e
x
F
1
_
P(i + )
P(i + )
_
_
= e
x
F
1

_
P(i + )
_
.
Furthermore,
F
1

_
P(i + )
_
= F
1

_
P(i + )
_
= P( + ),
and hence
P()
_
e
x
F
1
_
P(i + )
P(i + )
_
_
= e
x
P( + ) = P( + 2)(e
x
)
= P( + 2) =
m
P
m
(2) +
m1

k=0

k
T
k
,
for certain distributions T
k
E

(R
n
). (Note that e
x
= .) By our choice of the a
j
,
we thus infer from Lemma 1 that P()E = , i.e., that E is a fundamental solution of
the operator P(). This completes the proof.
May 2009] NOTES 459
Remark. As mentioned in the introduction, we gave a more involved formula for a
fundamental solution E in [7, p. 346], viz.
E(x) =
1
P
m
()
_
C,||=1

m
e
x
F
1

_
P(i + )
P(i + )
_
d
2i
. (1)
The verication of P()E = is very similar to the above. (We observe, incidentally,
that the last two formula lines in [7, p. 346] contain a printing error: therein P( + )
should be replaced by P( + ). The same holds for the last line in [13, p. 408].)
Let us emphasize as in [7, p. 347] thatin concrete cases such as considered, e.g.,
in [12]the explicit calculation of a fundamental solution will not start from formula
(1) or the one in Proposition 1, and, what is more, that for the evaluation of these
formulas the knowledge of some fundamental solution of the corresponding operator is
advantageous. This is shown in [7, pp. 347351] for the Cauchy-Riemann, the Laplace,
and the wave operator. We also observe that fundamental solutions are not unique
(except for the trivial case of order m = 0), the difference between two fundamental
solutions being a solution of the homogeneous equation.
3. FUNDAMENTAL SOLUTIONS FOR CONVOLUTION KERNELS WITH
FINITE SUPPORT. The equation P()E = can also be written as T E = for
T = P() E

(R
n
). In general, E D

(R
n
) is called a fundamental solution of the
convolution kernel T E

(R
n
) if T E = .
Lemma 2. Let T E

(R
n
) and be a Radon measure on R
n
with compact support
such that (L)(2x)

T = . Then the distribution
E =
_
e
x
F
1

_
(FT)( i)
(FT)( i)
_
, ()
_
is a fundamental solution of T. (The angle brackets are used here to symbolize the
integration of the continuous distribution-valued function of on the left-hand side
with respect to the measure .)
Proof. (a) The function (FT)(z) = (LT)(iz) = e
izx
, T
x
, z C
n
, is analytic, and
a consideration similar to the rst part of the proof of Proposition 1 shows that N =
{ R
n
: FT( i) = 0} is a null set for xed R
n
. Furthermore, Lebesgues
theorem implies that the mapping
R
n
D

(R
n
) : e
x
F
1

_
(FT)( i)
(FT)( i)
_
is continuous, and hence the formula for E in the lemma is meaningful.
(b) For S S

(R
n
) and R
n
, we have
T (e
x
F
1
S) = e
x

_
(T e
x
) F
1
S
_
= e
x
F
1
_
F(T e
x
) S
_
= e
x
F
1
_
(FT)( i) S
_
.
This implies
T E = e
x
F
1

_
(FT)( i)
_
, ().
460 c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 116
Furthermore,
F
1

_
(FT)( i)
_
= F
1
_
F(e
x


T)
_
= e
x


T
and hence
T E = e
2x
T, () = (L)(2x)

T = .
The proof is complete.
Remark. Of course, Proposition 1 ensues from Lemma 2 if we set T = P() and
=
1
P
m
(2)

m
j =0
a
j

. Similarly, formula (1) results from using a certain with


support on the line segment [1, 1] . (Note that e
x
F
1

_
P(i + )/P(i + )
_
depends on only the real part of for R
n
.)
Proposition 2. For each T E

(R
n
) \ {0} with nite support, there exists a funda-
mental solution E D

(R
n
).
In more detail, if T =

r
k=1
P
k
()
x
k
with pairwise different x
k
R
n
, P
k
()
C[] \ {0} of degree m, x
1
= 0, deg P
1
= m, q = r(m + 1), and R
n
is such
that x
k
are pairwise different for k = 1, . . . , r, then there exist real numbers
j
and
a
j
, j = 1, . . . , q, such that
E =
q

j =1
a
j
e

j
x
F
1

_
(FT)( i
j
)
(FT)( i
j
)
_
is a fundamental solution of T.
Proof. If we set =

q
j =1
a
j

, then
(L)(2x)

T =
r

k=1
q

j =1
a
j
e
2
j
x
P
k
()
x
k
.
Since
e
2
j
x
P
k
()
x
k
= P
k
( + 2
j
)(e
2
j
x

x
k
) = e
2
j
x
k
P
k
( + 2
j
)
x
k
,
we will choose
j
and a
j
such that
q

j =1
a
j
e
2
j
x
k

i
j
=
_
0, if k 2, i m or k = 1, i < m
1
P
1,m
(2)
, if k = 1, i = m.
(2)
Then (L)(2x)

T = and hence, by Lemma 2, T E = .
The existence of a
j
and
j
fullling equation (2) follows from the linear indepen-
dence of the functions e
c
k
t
t
i
, k = 1, . . . , r, i = 0, . . . , m, for the pairwise differ-
ent constants c
k
= 2x
k
. (In fact, if a linear combination f (t ) =

i,k
d
i,k
e
c
k
t
t
i
vanishes identically, then d
m,r
= lim
t
f (t )e
c
r
t
t
m
= 0 if, say, c
r
= max{c
k
:
k = 1, . . . , r}, and, inductively, all d
i,k
have to vanish. Hence the vectors v(t ) =
(e
c
k
t
t
i
)
i,k
cannot be contained in any xed hyperplane of R
q
, i.e., v(
j
), j = 1, . . . , q,
May 2009] NOTES 461
constitute a basis of R
q
for appropriate numbers
j
, and, indeed, this is the case for all
in R
q
outside a real-analytic subvariety.)
Finally, let us observe that the assumption x
1
= 0 R
n
does not restrict the gener-
ality, since, for general T and a fundamental solution E of T(x x
1
) as derived above,
E(x x
1
) will be a fundamental solution of T. This completes the proof.
Remark. The existence of a fundamental solution for convolution operators of nite
support was proven by functional analytic methods in [2, Thm. 2, p. 290]. In terms
of the general theory of convolution equations in [3, Ch. XVI], it can be expressed
by the statement that distributions with nite support are invertible, [3, Def. 16.3.12,
p. 330], and this fact is a consequence of [3, Cor. 16.3.18, p. 333]. However, the explicit
construction of a fundamental solution given in Proposition 2 seems to be new.
ACKNOWLEDGMENTS. I am indebted to Prof. U. Oberst for pointing out to me the signicance of
differential-difference equations in system theory, which elucidation was the initial motivation for this paper. I
am particularly grateful to Norbert Ortner for our continuous cooperation, which has informed all of my work.
I would also like to thank to the referees for valuable comments.
REFERENCES
1. L. Ehrenpreis, Solution of some problems of division. Part I. Division by a polynomial of derivation, Am.
J. Math. 76 (1954) 883903
2. , Solution of some problems of division. Part II. Division by a punctual distribution, Am. J. Math.
77 (1955) 286292.
3. L. H ormander, The Analysis of Linear Partial Differential Operators II: Differential Operators with Con-
stant Coefcients, Grundlehren der Mathematischen Wissenschaften, vol. 257, Springer-Verlag, Berlin,
1983.
4. J. Horv ath, Topological Vector Spaces and Distributions I, Addison-Wesley, Reading, MA, 1966.
5. H. K onig, An explicit formula for fundamental solutions of linear partial differential equations with
constant coefcients, Proc. AMS 120 (1994) 13151318.
6. B. Malgrange, Existence et approximation des solutions des equations aux d eriv ees partielles et des
equations de convolution, Ann. Inst. Fourier 6 (1955/56) 271355.
7. N. Ortner and P. Wagner, A short proof of the Malgrange-Ehrenpreis theorem, in Functional Analysis
Trier, Germany, 1994, Proc. 1st Int. Workshop, ed. by S. Dierolf, S. Dineen, and P. Doma nski, de Gruyter,
Berlin, 1996, 343352.
8. , A survey on explicit representation formulae for fundamental solutions of linear partial differ-
ential operators, Acta Appl. Math. 47 (1997) 101124.
9. J.-P. Rosay, A very elementary proof of the Malgrange-Ehrenpreis theorem, this MONTHLY 98 (1991)
518523.
10. L. Schwartz, Th eorie des Distributions, Nouv. ed., Hermann, Paris, 1966.
11. F. Tr` eves, Solution el ementaire d equations aux d eriv ees partielles d ependant dun param` etre, C. R. Acad.
Sci. Paris 242 (1956) 12501252.
12. P. Wagner, Fundamental solutions of real homogeneous cubic operators of principal type in three dimen-
sions, Acta Math. 182 (1999) 283300.
13. , On the explicit calculation of fundamental solutions, J. Math. Anal. Appl. 297 (2004) 404418.
University of Innsbruck, Technikerstr. 13, A-6020 Innsbruck, Austria
http://techmath.uibk.ac.at/wagner/
462 c THE MATHEMATICAL ASSOCIATION OF AMERICA [Monthly 116

Você também pode gostar