JOSE SEBASTIAO E SILVA
TEXTOS DIDAcTICOS
Volume III
SERVI<;O DE EDUCA<;AO E BOLSAS
FUNDA<;AO CALOUSTE GULBENKIAN
I
LISBOA
Reservados todos os direitos de acordo corn a lei Edic;ao da FVNDA<;AO CALOUSTE GULBENKIAN Av. de Berna - Lisboa
1999 ISBN 972-31-0971-9 Dep6sito Legal
n.O
148805100
111.1 THEORY OF DISTRIBUTIONS路
*
Este texto tern por base apontamentos coligidos por diversos alunos de Jose SebastHio e
Silva na sequencia de urn curso que realizou em
1958 na Universidade de Maryland, e que
posteriormente foram utilizados, e por ele revistos, na Faculdade de Ciencias de Lisboa.
CHAPTE R
VI I I
PA RTIAL I NTEG RA LS AN D M U LTI P L E I NTEG RALS. CO NVO LUTI O N . 8. 1 . Partial limits for distributions of two variables.
Let I and ] be two intervals in IR , and suppose that ] is un bounded on the right. Given two functions f (x, y) and g (x) respec tively on I x ] and I, f(x, y) is said to converge uniformly on I to g (x) as y � + 00 , if and only if for every £> 0 , there exists a 7] E ] (inde pendent of x) , such that: I f (x, y) g (x) 1 < £ for all y > 7] and x E I. On the other hand, if a is any real, we write f(x, y) E o(y a ) -
uniformly on I as
on I as
y � + oo .
y � + 00 , if and only if (iff)
f(x, y )
ya
� 0 uniformly
Put for every f E C(I x ]) :
where x o ' (respectively Yo ) is a fixed point, arbitrarily chosen in I (respectively ]) . The following lemma is easily proved (cf. 6. 1 . 3 . ) : 8 . 1 . 1 . LEMMA. If a > - l
y � + oo, then
and f(x, y) E o(y a ) uniformly on
I
as
1 56
�x f E o(ya) and �y f E o(ya+l )
uniformly on I as y � + 00 . This lemma leads to the following: 8. 1 .2. DEFINITION. If f E !0'(I x J) and a > - I , we write f(x, y) E o ( y a ) on I as y � + 00, iff there exist m, n E 1No and F E C(I x J) such that:
( I ) f(x, y) = DxmD;F (x, y) ; (2) F (x, y) E o (ya+ n ) uniformly on each compact interval I* C I as y � + 00.(7) Applying this definition and the lemma, the following properties are easily shown: 8.1.3. If f(x, y) E o (ya) and g (x, y) E o (y a) on I as y � + 00, all A , !l E e :
then, for
If f(x, y) E o(y a) on I as y � + 00, then Dx f(x, y) E o (ya ) on I as y � + oo.
8.1.4.
If f (x, y) E o (y a ) on I as y � + 00 and (jJ (x) is multipliable by f(x, y) , then (jJ (x)f(x, y) E o (y a ) on I as y � + 00. 8.1.5.
Consider now g E !0'(I ) and f E !0'(I x J) ; then: 8.1.6. DEFINITION. We say that f(x, y) converges on I to g (x) as y � + 00, if and only if f (x, y) - g (x) E o( l ) on I as y � + 00. The distribution f (x, y) is said to be convergent on I as y � + 00, iff there exists a distribution g on I satisfying the preceding condition . (7) A more general condition could be required instead of (2) , but this definition is quite sufficient for applications.
1 57
The uniqueness property as well as the linearity property of convergence are in this case immediate consequences of 8 . 1 . 3 . Then we can write:
g (x) = lim f (x, y) or g (x) = f (x, + 00 ) on I, Y � + OO
to express that f (x, y) � g (x) on I as y � + 00 .
On the other hand, the following important property, which do es not hold in classical analysis, is an immediate consequence of 8 . 1 .4 . : y � + 00, then Dx f(x, y) � Dx g (x) on I as y � + 00 , that is :
S.1.7. DIFFERENTIATION PROPERTY. Iff(x, y) � g (x) on I as
Dx Um f(x, y) Y -' + OO
=
lim Dx f(x, y) on I.
Y � + OO
It turn, from 8 . 1 .5 . , follows S.1.7' . MULTIPLICATION PROPERTY. lf f (x,
y � + 00 and cp (x) is multipliable by f(x, y), then:
y) � g (x) on I as
Um [ tp (x) f (x, y)] = cp (x) lim f(x, y) on I. Moreover, applying 8 . 1 .7 . and the linearity property, it is easily shown: S . 1 . S . S U B S TITUTION PROPERTY.
If f (x, y) � g (x) on I as
y � + 00 , and if h is a mapping of an interval 1* into I, such that f(h (t), y) exists (cf. 6 . 8 . ) , then g (h (t)) exists too and f(h (t), y) � � g (h (t)) on 1* as y � + 00 . This substitution rule concerns the parameter x. Substitution rules concerning the converging variable y can be easily found as generalization of the criteria given in 6 . 6 .
158
The "0 " symbol i s extended to distributions f(x, respect to y, in the following way: 8.1 .9. DEFINITION. If a > - I , we write
y) on I x I, with
f(x, y) E O(y«)
y � + 00 , if there exist rn, n E1No and F E C(I x l) such that: (i) f(x, y) = DxmDynF(x, y) ;
on I as
(ii) for every compact interval I* C /, there exists a number M such that
F(x, y) ( 1 + I y l ) « +n
----
s
M on 1 * x I. (8)
More generally, if (fJ E C OO(I), we write f(x, y ) E O ((fJ ( Y» on I as y � + 00 , if and only if there exists a real Yo and a distribution fo (x, y)E 0 ( 1 ) on I as y � + 00 such that f(x, y)= (fJ ( y)fo (X' y), for y > Yo and x E I. Besides the linearity property, it is easily shown:
is any real and f(x, y) E O(y«) on I as y � + oo, then Dx f(x, y) E O(yIX) and Dy f(x, y) E O(y« - l ) on I as y � + oo. 8. 1 . 1 0 . DIFFERENTIATION PROPERTY. If a
Obviously all preceding considerations extend to the case when I is an interval unbounded on the left, and y � 00 . -
8.2. Partial integrals for distributions of two variables.
Let I and I be any two intervals in IR , f (x, y) a distribution on I x I. A distribution (fJ(x, y) such that Dy (fJ(x, y) f(x, y) will be called a (partial) primitive of f(x, y) with respect to y. On the other hand, a distribution u (x, y) on I x I is said to be independent of y, if =
I* x J.
(8) The choice of ( 1 + I y I)a+n instead of ya + n is only to make the quotient continuous on
1 59
and only if it reduces to a distribution g of the variable x only, i .e., iff it is of the form u (x, y) = �mG (x) with m E /No ' G E C(I).
8.2.1 . LEMMA. A
Dy u = O.
distribution u on I x J is independent of y, iff
u is independent of y, then Dy u = O. Suppose now conversely that Dy u = O and assume u=�mD; U with m, n E /No and UE C(I). Then Dy u=�mD; + l U= O and therefore (cf. 7 .2. axiom 4) U must be of the form PROOF. It is readily seen that, if
U (X, y) = Hence
� x' a. < y) + � y j b/x),
m-I
n
with
a, E C(J) and bj E C(l) ,
u (x, y) = �mD; U(x, y) = n!�mbn (x) . •
That being so, it is easily proved, as in the case of one variable: 8.2.2. THEOREM.
Every distribution f on I x J has infinitely many p rimitives with respect to y and two such primitives differ by a dis tribution independent of y. We are now able to define, in a natural way, the concept of par tial (or parametric) integral of a distribution f (x, y). It will be suf ficient to consider integrals on /R . Let I be any interval in /R and j E f¥(I x /R ); then: 8.2.3. DEFINITION. The integral
if and only if there exists a primitive ({J of f with respect wich is convergent on I as y � + 00 and as y � - 00 . Then, we
gent on
to
y
f f (x, y)dy is said to be conver JIR
write
I,
f f (x, y)dy = qJ (x, + oo) - qJ(x, - 00 ) on I. JIR
From 8 . 2 . 2 . , follows at once the uniqueness of the partial inte gral . From the properties of partial limits we can deduce the linearity property for partial integrals, as well as the following properties :
1 60 8.2.4. DIFFERENTIATION PROPERTY.
vergent on I, so is Dx
r Ix' (x, y) dy and JIR
If
r f(x, y) dy is con JIR
r f(x, y) dy = r Dx f(x, y) dy on 1. JIR JIR
8.2.5. SUBSTITUTION PROPERTY.
If
r f(x, y) dy = g(x) on JIR
I
and if h is any continuous mapping of an interval 1* into I such that f(h(t), y) exists, then
r f(h(t), y) dy = g(h(t» on 1 *. JIR As for substitutions concerning the integration variable y, the criteria established in 6.6. can be easily extended to partial integrals . In particular, we have, for all h E IR : 8.2.6.
r f(x, y + h) dy = r f(x, y) dy. JIR JIR
8.2.7.
1 r f(x, y) dy. r f(x, hY) dY = _ 1 h 1 JIR JIR
Criterium 6.3 .6. can also be extended to partial integrals :
Iffor any compact interval 1 * C I, there exists a compact interval K such that f(x, y) = O on I * x (lR - K), then the 8.2.8. THEOREM.
integral
r f(x, y) dy is convergent on 1 . JIR
f =�mD;F, with F E C(I x IR ) . The hypothe sis implies that, in a set { x E I * , y < -Yo } , F(x, y) reduces to a pseudo polynomial P of degree < (m, n), which we can assume to be zero, PROOF. Suppose
161
otherwise we could subtract P from
defined by the value
F (remember that P i s uniquely of F (x y) for m values of x in 1* and n values ,
y in IR ). Then there exists a primitive of f with respect to y, s ay qJ, which is zero for x E I* , Y < Y o and reduces to a function
of
If/ independent of y for x E I* , Y > Yo ' Now, it is easily seen that -
qJ � 'If on
I
as
y � + 00 and qJ � 0 on I as y �
f f(x, y) dy = If/(x) . JIR ;
-
00 ,
so that
•
Finally, the following extensions of
6 . 5 . 1 . and 6.5 . 2 . are easily
proved :
8. 2.9. THEOREM.
if
f f (x, y) dy is convergent on I, then f E O (y-l ) JIR
on I as y � 00 . On the other hand, if there exists f E O( y« ) on Remarks. 1
I as y � 00 , then -
If
a< -1
such that
f f(x, y) dy is convergent on l. JIR
f (x, y) is a function, then for the convergence of
f f(x, y) dy on I in distributional sense, it is not sufficient (nor nec JIR essary) that the integral be convergent
for each x E I. Obviously,
a sufficient condition is that the integral be uniformly convergent on each compact subinterval cont ained in be proved that, if
I.
More generally, it can
f is summable on each set
compact interval contained in I, then in distributional sense.
1* x IR, where 1* is
a
f f (x, y) dy is convergent on I JIR
2 - The differentiation property can be associated with the line arity property in a more general property. Let
p (D) be a derivation
polynomial, that i s an operator of the form p (D) ai '
.
..
, an E C. Then we have
� n
=
ak D k , with
1 62
i f (x, Y) dy = i p (Dx ) f(x, y) dy
p (Dx )
IR
IR
on I,
whenever the first integral is convergent on I. Example. The preceding remarks offer a simple justification of for足
mula in 6 . 3 . 5 . 8.2.10.
-
2. Observe that:
i
e i xy dy = n-e- 1 x l for each x E /R . 1 + y2
--
IR
This can be easily found by the meth od are real variables, we have l e i Xy l = 1 and
e i xy 1 + y2
--
-
1
1 + y2
of residues .
for all
Besides, as
x, y
x, y E/R .
Thus the integral 8 .2. 1 0. is dominated, for all x E /R , by the integral
i ( 1 +y2 )-ldy, IR
which is obviously convergent . Hence, according to
uniformly convergent on /R and th erefore convergent on /R in distributional sense. Consequently,
the Weierstrass test, the first integral is
On the other hand,
so that
( 1 - D; )e- 1 x l = 2 o (x).
1 63
Hence, from . 8 .2. 1 0. follows :
L eixydy =
8.2. 1 1.
IR
8. 3 . Multiple integrals (on
2 1C8 (x) on IR .
IRn ).
Let f be a distribution on IRn and A any complex number. We say that f (x) converges to A as x � + 00 if and only if there exist r E IN; n and F E C(lRn) such that f = Dr F and
A
F(x)
Then, we write ..1 = lim f (x) or ..1 = f( + oo ) . x--+ + oo
n
n
The uniqueness of the limit, as well as the linearity property can be proved by an argument similar to the one used in the case n = 1 . The concept of convergence as x � 00 is analogously defined. On the other hand, every distribution qJ such that D qJ = f (where D Dl . . . D ) will be called a pure mixed primitive of f . It is easily n seen that: -
=
8.3. 1. THEOREM. Every f E !?lJ(IR n ) has infinitely many pure mixed
primitives and two such primitives differ necessarily by a distribution
� u where n
of the form
j
Uj
is a distribution independent of Xj (that is,
of the form D ru where u is a continuous function on IR n independent of x) . That being so, we shall write by definition. 8.3.2.
1 64
where qJ is any pure mixed primitive of f and 4h is the mixed dif ference operator A 1h1 • • • A hn n • From 8 . 3 . 1 . follows that formula 8 . 3 .2. defines actually a dis tribution t/J (x, x' ) on IR 2 n independent of the choice of the pure mixed primitive qJ. To see this it is sufficient to observe that Aj h uj 0 j for every distribution uj independent of Xj
=
•
8.3.3. DEFINITION. A
IR n, iff
f't (�) d�
distribution f is said to be integrable on
is convergent as (x, x' ) --+ (- 00 ' + 00 n ) . Then we n
write:
iRJ(X)dx =x'-x�'!!,+ OOn tt (�)df n=
8.3.4.
For example, if
2
f(x1 , x 2 ) dx1 dx 2= '1' (+ 00, + 00 )- qJ (+ 00, - 00 )- qJ (- oo, + oo ) + '1' (- 00 , - 00 ) f JIR2
where qJ is a primitive of f with respect to x. The integral of f on IR n can also be denoted by by
f t.
JIfRn
f or simply
Uniqueness and linearity properties are immediate conse-
quences of the corresponding properties for limits . In order to obtain further criteria it is convenient to introduce a suitable definition of bounded distributions.
A distribution f is said to be bounded if and only if there exists r E /N� and F E C(/Rn) such that: (i) F; (ii) for every regular matrix A of order n, the function x i r1 . . . x;; rn F (A x ) is bounded on IR n. 8.3.5. DEFINITION.
f=D'
on
IR n,
1 65
The linearity property of boundedness is easily proved. 8.3.6. DEFINITION. Given f E �(IR n ) and qJ E c oo(IR n ), we write
or simply f E O(qJ), if and only if there exists a dis tribution fo bounded on IR n and a real e > 0, such that f qJ fo ' for
fE O(qJ) as I x 1 -+
00
=
Ixl>e.
That being so, the following generalization of 6.5. 1 . is easily obtained: 8.3.7. THEOREM.
If there exists a < - n such that f E O( I x I a ), then
f is integrable on IR n.
On the other hand:
Suppose fE O( l x l a ) with a < - n and let h be a C oo one-to-one mapping of IR n onto itself such that (i) the lacobian matrix [D; h) of h is regular on IR n and con verges to a regular matrix as I t 1 -+ 00 , (ii) DTDi hj E o (t T ), for all r E IN; , i , j 1 , . . . , n (9) . Then the classical substitution rule applies:
8.3.8. THEOREM.
=
1
/R n
f (X ) dx =
l
/R n
f(h(t»
l
( ht ) dt.
We shall outline the proof only in the case when h is a non-de
generate affine mapping, that is a mapping of the form h (t) = c +M(t),
where c i s any vector in IR n and M i s a regular matrix of order n. This
case may be taken as a model for the general case since h behaves
asymptotically j ust as an affine mapping according to (i).
(9) As far as functions are concerned is understood that the stated conditions are to be taken in ordinary sense.
1 66
qJ(x) = ( 1 + X12 + . . . + X; ) 1 I2
and suppose fE O( l x l a ), with a < - n. Then, it is readily seen that f E O(qJa ), i.e. there exists r E IN; and F E C, such that f = qJa Dr F, with X1r\ . x;rn F (A x) bounded on IR n for every regular . matrix A of 路 order n. In such conditions it is easily found: Put
ff (X) dx = (- l )" T " f
..
rp(Tl(X) F (x) dx , where
Il r l l = r] + . . . + rn
路
Now :
and it can be seen, without difficulty, that the last integral is just equal to
(- l ) I T "
ff h
( (t 禄 l det M l dt .
8.4. Partial and multiple integrals.
Let us consider a distribution f(x, y) on IR m + n , with x E IR m and
..
y E IR n (rn, n = l , 2, . ) . The concept of partial integral
f f(x, y) dy JlRn
can be easily defined as a generalization of preceding concepts of partial and multiple integral, with similar properties . But there is a new property : 8.4.1 . THEOREM.
the integral
If f(x, y) is integrable on IR m+n and in addition
f f (x, y) dy is convergent on IR m, then JlRn
fJ/Rm+n f (x, y) dx dy = Jf/R m ( JflRn f (x, Y) dY) dx .
1 67
as
This is a consequence of a property for limits that we can state follows :
8. 4.2. THEOREM. If f(x, y) is convergent as (x, y) � (+ OO m ' + oo n ) and if in addition f(x, y) is convergent on IR m as y � + oo n ' then
lim f(x, y)
X - + OO m
=
y - + OO n
lim
X - + OO m
(
)
lim f(x, y» ,
y - + OO n
PROOF. It is sufficient to prove this rule in the case m
=
n= 1. Suppose that the hypothesis holds , Then there exist four integers r, s, 2 t, u, two functions Fp F2E C(IR ), a function G E C(lR) and a num ber Il , such that f �rD/ F;. D:D u F2 and y =
=
(i)
(1' 1' )
FI (x,
y) x r ys
�
F; (X, y) yU
�
---
Il
r! s !
G (X ) U!
We can assume that t = r, exist a, b > 0 such that:
as
(x, y)
�
+ 00 ) ;
.
'+ I y on eac h compact set In /' D , UnhOrm
u
=
1\
m
as
s , Take £ > 0, then according to (i) there
<£ -r! s ! Il
8.4.3.
( + 00 ,
for
x > a, y > b,
Take now r additional points Xj > G , s additional points Yk > b and con sider two pseudo-polynomials rg>l (x, y) , rg>/x, y) of degree (m, n) such that F;. - rg>l and F2 - rg>2 vanish on the lines x = xj ' y = Yk ' Then if we put Fo F;. - rg>l ' we have f �rD/ Fo ' Fo F2 - rg>2 and it is easily seen that (i), (ii) are again satisfied with Fo in the place of FI and F2 (t = r, U = s), since the coefficients of the pseudo-polynomials are obtained as linear combinations of the values of FI (x, y) and F2 (x, y) on the lines =
=
=
1 68
X = Xj ' Y = Yk .
Hence from 8 .4 . 3 . follows, with
and taking the limit as
Fo in
the place of
Y � + 00 :
G (x)
s s ! £ for
x > a.
Th e numb er £ b elng · · " Imp · I·les th at arb ltrary, th IS
x�+
00 ,
which means that A = Urn Urn .'
x -+ + oo y -+ + oo
FI ,
f(x,
y) . •
A G (x) � r! Xr
-
as
More generally:
If f(x, y, Z), with x E/Rm, y E /R n , zE/RP is convergent on /R m + n as z� + oop and iff(x, y , z) is convergent on /Rm as (y, z) � (+ oon ' + oop ) ' 8.4.4.
then
8.5. Convolution of two distributions on /R.
Consider two distributions
f = DmF and g = D n G, where
F, G E C(/R). Then we have: f(x - t) = �mF(x - t) = (_l )m �mF(x - t) so that, for every k 0, 1 , . . . =
D/f(x- t) = (- l ) k �kf(x - t). This suggests to write by definition
f(x - t) g (t) = f(x - t) D," G(t) =
i G) D,"- k(G(t)D:f(x- t))
1 69
with
G(t) D}f(x- t) = Dxm + k (F(x - t) G(t» , that is 8.5 . 1.
f(x - t) g(t)
=
i (�) Dxm + kDtn - k (F(x - t) G(t)).
It is easily seen that the "product" f(x - t) g(t) does not depend on the representation of the distributions f and g. We can prove it as we have done for the product of a C n function with a Cn distribution
in 4. 1 . The analogy between these two situations comes from the following proposition, which can be proved without difficulty, but which is not essential for the following subject: The mapping
t � f(x - t) of IR into the space �(lR ) is infinitely differentiable. Consider now the expression sible interpretations :
f(x - t) g(t - y)
=
f(x - t) g (t-y).
We have two pos
i G) Dxm + kDtn - k (F(x- t) G(t -y))
8.5.2.
f(x - t) g(t-y) =
� (7 ) (_ l )kD,m-'D; + ' (F(x - t) G (t -y)).
Remembering that the functions F and G can be approached by two sequences { Fn } and { GJ of C OO functions converging uniformly on each compact interval, it is readily seen that:
The right members of the formulas 8.5 .2. represent the same distribution. 8.5.3.
A direct proof of this proposition does not seem to be easy.
1 70
8.5.4. DEFINITION. If the integral
on /R , the distribution
h ex) =
L+ oof(x - t) g (t) dt - 00
is convergent
JfIR f (x - t) g(t)dt
is called the convolution of f and g and denoted by f * g. From this definition, taking into account the linearity property of the partial integral, as well as 8 . 5 . 1 . , follows immediately that the convolution is bilinear, that is, we have:
whenever f1 * g and f2 * g exist, and analogously for the right side. Moreover 8.5.6. COMMUTATIVE LAW:
f * g = g * f.
If f * g exists, g * f exists too, and
PROOF. Suppose that f * g exists and put h = f * g, that is
h ex)
=
r f(x - t) g (t) dt. Then for each y E /R , we have: JIR h(X - y) =
i f(x - y - t) g (t)dt IR
and it is obvious that the last integral is still convergent with respect to (x, y) on /R 2 . On the other hand , Jo r each y E /R , we may perform on this integral the substitution t = u -y, which gives :
h(X - y) =
i f(x - u) g (u - y)du . IR
Now, taking 8 . 5 . 3 . into account, it can be seen that the last integral is also convergent with respect to y Jor each x E /R . In particular, for x 0 , we have: =
17 1
h (- y) =
r f (- u) g (u - y) du . JIR =
Hence by the substitutions y -x, u = - t :
h (x) = that is, h = g * f •
r g (x - t) f(t) dt, JIR
.
In the general case the convolution is not associative. But the fo ll owing criterion can be used in several cases : 8.5. 7.
If
r f(x - y) g ( y - t) h (t) dy dt, where f, g, h EliJ, is convergent JIR2
on IR , then
( f * g) * h = f * (g * h) =
r f(x - y) g (y - t) h (t) dy dt. JIR2
This is an immediate consequence of 8 .4.4. In turn, from the differentiation and substitution properties for partial integrals and from 8 . 5 . 6 . , follows immediately, taking defini tion 8 . 5 .4. into account: 8.5.8. DIFFERENTIATION PROPERTY. If f * g exists, then D (f * g)
exists too, and
D (f * g) = (Df) * g = f * (Dg). 8.5.9. TRANSLATION PROPERTY.
h E IR
On the other hand:
If f * g exists, then for every
1 72 8.5. 10.
If f * g and f * (xg) exists, then (xf ) * g exists too and x (f* g) = (xf ) * g + f * (xg) .
PROOF. It is sufficient to observe that (xf ) * g is given by
r (x - t) f(x - t) g (t)dt = x r f(x - t) g (t) dt - r f(x - t) tg (t)dt. Jm Jm Jm
•
This important property shows that multiplication by x, with re spect to convolution, behaves like a derivation operator. Finally, we can analogously prove that 8.5. 1 1 .
If f * g exists, then
8.6. Convolution of distributions whose carrier is bounded on the left and ( or) on the right.
We shall denote by §'* (IR) or simply §'* the vector space of all distributions on IR with bounded carrier. 8.6. 1 . THEOREM.
The convolution f * g exists whenever fE §2J* and
g E � Besides, (i) f * (g * h) = (f* g) * h, whenever f, g E �* , h E §2J ; (ii) 8 * f = f, !or every f E � . PROOF. a) Suppose f E �* , g E §2J . Then there exists a bounded interval J such that g(x - t)f (t) vanishes for (x, t) f£. IR x J. Hence
r g (x - t)f(t) dt Jm
is convergent on IR and gives f * g.
b) Suppose f, g E §2J* , h E §2J. Then by an argument similar to the preceding it is shown that the integral
173
L
/R 2
f(x - y) g ( y - t) h (t) dy dt
is convergent on IR , and this according to 8 . 5 . 7 . implies (i).
c) Consider f= D nF, where F E C(lR ), and put � = FH, F2= F-F) . Now:
Hence 8 * D nF) = D n+) (H * F',) = D nF) . It is seen analogously that 8 * D nF2 = D nF2 ' so that 8* f = f • .
This theorem along with 8 . 5 . 5 . can be expressed by saying : 8.6.2. The space �* is an algebra under convolution and � is a module over that algebra, having 8 as unit element.
Property (ii) in 8 .6 . 1 . can be expressed explicity by the important formula f(X) =
L 8(x - t) f(t) dt IR
(DIRAC ' S FORMULA) .
We shall denote by sz: (respectively � ) the vector space of all distributions vanishing on the left (resp . on the right) of 0 and by .@. (resp. � ) the space of all distributions whose carrier is bounded on the left (resp. on the right) of O. 8.6.3. THEOREM. The space
.@.
(resp.
�)
is an algebra under
convolution and � (resp. � ) is a subalgebra of � (resp. � ). In fact, if f, g E� there exists a real c such that f and g vanish ,
for x < c. Then f(x - t) g (t) vanish for t < c and t >x - c. Hence
L f(x - t) g (t) dt IR
is convergent on IR and vanishes for x < 2c. The
remaining parts of the theorem are easily proved . •
1 74
8.7. Convolution and order of growth, tempered distributions and rapidly decreasing distributions (on IR).
Several criteria can be found, connecting convolution with order of growth of distributions . One of these criteria is the following :
Let a and f3 be two real numbers satisfying one of the following conditions (i) a +f3 < - 1 and a > 0 ; (ii) a +f3 < 3 and f3 s a < O. On the other hand, let f and g be two continuous functions on IR such that f E O (x a ) and g E O (xP )(lO). Then f * g exists and f* g E O(x a ) . PROOF. a) Suppose a +f3 < - 1 with a ďż˝ O. Then as f E O (x a ), there exists a number M such that I f (x) 1 s M( I + I x l ) a for all x E IR . 8 .7. 1 . THEOREM.
-
Hence
since
a > O.
So the integral
f f(x - t) g (t) dt Jm
is dominated by M( I
+ I x l )a
f ( 1 + I t l ) a l g (t) l dt. Jm
Since g E O(xP ) and a +f3 < 1 , the last integral exists . Hence the first integral is uniformly convergent on each compact interval in IR and -
its absolute value is s MK( I + l x l ) a where K = f ( I + l t l ) a l g (t) l dt. Jm Consequentely, f * g E O (x a ). b ) Suppose now a +f3 < 3, with f3 < a < 0 , and consider the integer n such that 0 < a + n < 1 . Then it is easily seen that a +f3 + n < - I so that x kf* xn- kg exists and is O (x a + k ) for k = 0, 1 , . , n according to the previous conclusion. Hence (cf. 8 .5 . 1 0) : -
.
( 1 0) It is understood : "in ordinary sense as x ďż˝ 00 " .
.
1 75
x " ( f* g ) so that
f* g E O(x a ) . •
=
� (:)
8.7.2. COROLLARY. Let
(X kj * x n kg ) E O (x a + ,, ) -
be a real < 2, A a the set of all continuous functions f on IR such that f E O (x a ) as X � OO and Ba the set of all continuous functions g on R such that there exists a real number /3 > 0 (depending on g) satisfying the conditions a +/3 < - 1 and g E o (xf3 ) Then A a is an algebra under convolution and Ba is a module over that algebra. a
-
•
PROOF.
Applying to the theorem (changing the roles of a and
f3 ), it is readily seen that f * g exists and belongs to Bo: whenever f EAo: and g E Bo: ; and that f* g E A whenever f, g EAo: . So we have only to a prove the associative law: f* ( g * h ) ( f * g ) * h , f, g EA a , h E B . But a =
this can be easily seen applying 8 .5 .7 . as we did for 8 . 6 . 1 . • 8.7.3. Remark :
The preceding theorem and corollary can be extended
to locally summable functions according to the following criterium
(FUBINI-TONELLI THEOREM) : If f, g EL(lR), then
r f(x - t) g (t) dt JIR
is convergent almost everywhere in IR and defines
function h EL(lR). It can still be stated that the preceding integral is convergent in the mean on IR , so that f* g exists in the distribu
a
tional sense. Applying 8 . 5 . 1 1 . and taking the Fubini-Tonelli theorem into account, it is
a
simple matter to obtain the following generaliza
tion of 8 .7 . 1 . :
Let a, /3 be two real numbers satisfying the co nditions (i) or (ii) of 8 .7 . 1 . , a ', /3' two real numbers such that a ' + /3' :s; 0 and f, g two locally summable functions such that f E O(x a e a ' lxl ) and g E O(x f3ef3 l x l ). Then f* g exists and f* g E O (x ae r l x l ) , where Y= max ( a ', /3 ' ) . 8.7.4. THEOREM .
'
1 76
For the proof it is convenient to consider I and g in the form 1 = /) + /2 ' g = g ) + g2 ' with II ' g) E C+ , 12 , g2 E C_ , remembering that I) * g 1 E C+ ' 12 * g2 E C . _
From 8 . 7 .4 . is easily deduced a corresponding generalization of 8 . 7 . �. Now, applying the differentiation property, we can derive frop! the preceding criteria corresponding rules for distributions. For example, --...
let us denote by A a for every a < 2, the set of all distributions of the -
fonn f = k�DnkF. ' where p, n" . . . , np are arbitrary integers and F. p
locally summable functions such that �E O(x a ), and by B the set of all distributions of the fonn g =
k�"Gk q
a
where q , r" . . . , rq are
arbitrary integers and Gk locally summable functions such that
Gk E O(xP ) with a + f3 < - 3 and a < f3 ( 13 depending on g ). Then it is easily seen that A a is an algebra under convolution and B a module a over A a . '"'-'
8.7.5. DEFINITION. A distribution I on IR is said to be tempered (slowly increasing or of polynomial type) if there exists a real a
such that IE O(x a ) (in distributional sense) .
An equivalent definition to this is the following: I is tempered if
and only if there exist two integers n , k and a/unction F E C(lR) such that I =DnF and F E O(X k ) in ordinary sense. "-'
"-'
We shall denote by !!lJ(IR) or simply by !?lJ the set of all tempered
distributions. It is readily seen that §!J is a vector space closed under D. "-'
8.7.6. DEFINITION. A distribution f on IR is said to be rapidly de creasing if and only if for every
a < O, I can be represented in the
1 77
fonn f = �lDn.F" where p, p
nl ' . . . , np
(n, "' O,
are arbitrary integers
p � 1 ) and � continuous functions such that � E O(x a ) in ordinary sense. We shall denote by §J the set of all rapidly decreasing distributions on IR . From preceding results it is easily deduced: r--..
8 .7. 7. COROLLARY. §J r--..
r--..
'-"
is an algebra under convolution and §J a
module over §J.
v
A similar result can be obtained concerning the space §J of all
DnF, where a continuous function on IR such that F E O(e a 1 x l ) for some real
distributions of exponential type (that is, of the form f F is
=
/\
a) and the space §J of all exponentially decreasing distributions
(that is, of the form f = DnF, where
F is a continuous function such
that F E O(e a 1 x l ) for all real number a). /\
r--..
'-"
v
Observe that §J* c §J c §J c §J C §! C §!. 8.7.8. Convolution in IR n . The concept of convolution of distribu
tions on IR is readily extended to the case of distributions on IR n , and all preceding properties of convolutions can be generalized to this
case: only we are now concerned with derivation operators, transla tion operators, etc. , corresponding to the different variables. Theorem 8.6. 1 . is readily extended to distributions of several variables. As for theorem 8.6.3 . it gives place to new possibilities in the case of n variables.
Let r be any convex cone in
IRn
whose vertex is at the origin '"'-'
and not reducing to a half space. We shall denote by §!r the set of all
distributions on
IR n vanishing
Then it is easily seen that §r
outside some cone a + r with a EIR n .
is an algebra under convolution and
�r a subalbegra of �r ; besides, there exists a maximal subspace of � distinct from §!r which is a module over �r .
178
Finally, the criteria given i n 8 . 7 . can also b e extended to the
case of
n
variables and combined between them and the preceding
ones, according to the different variables .