THE REGULARITY THEOREM FOR DISTRIBUTIONS

MATT ZIEMKE

The purpose of this paper is to give a proof of the one-dimensional Regularity Theorem for Distribu-
tions which states that if T is a tempered distribution on R then T is the weak nth derivative of some
polynomially bounded continuous function. We will start by giving the necessary definitions then prove the
N-Representation Theorem for the Schwartz class and for tempered distributions which we will then use to
prove the Regularity Theorem.

Definition 1. A seminorm on a vector space V is a map : V [0, ) such that

(i) (x + y) (x) + (y)

(ii) (x) = ||(x) for all C

A family of seminorms ( )A is said to separate points if (x) = 0 for all A implies x = 0.

Definition 2. A locally convex space is a vector space X with a family of seminorms ( )A which
separates points. The natural topology on a locally convex space is the weakest (or smallest) topology for
which all the are continuous and for which the operation of addition is continuous.

Definition 3. If ( )A and (d )B are two families of seminorms on a vector space X such that the
natural topologies with respect to each family are the same then we say the families ( )A and (d )B
on X are equivalent.

Proposition 1. Let ( )A and (d )B be two families of seminorms. Then the families are equivalent
if and only if, for each A, there exists 1 , 2 , . . . , n B and C > 0 so that for all x X

(x) C (d1 (x) + + dn (x))

and for each B there exists 1 , 2 , . . . , m A and D > 0 so that for all x X

d (x) D (1 (x) + + n (x))

Proof. First, suppose the families are equivalent. Let A. Then {x : (x) < 1} is d -open. So there
exists N = Nd1 ,...,n , such that N {x : (x) < 1}. Suppose x X such that dk 6= 0 for some k = 1, . . . n.
Then for all k = 1, . . . n,  
x
dk <
d1 (x) + + dn (x)
So,  
x
<1
d1 (x) + + dn (x)

Date: May 2013.

1
THE REGULARITY THEOREM FOR DISTRIBUTIONS MATT ZIEMKE

hence,
1
(x) < (d1 (x) + + dn (x))

If dk (x) = 0 for all k = 1, . . . , n then dk (ax) = 0 <  for all k = 1, . . . , n and so (ax) < 1. Hence,
1
(x) < a for all a > 0 and therefore (x) = 0. The second statement is symmetric to the first. Now, for
the other direction. Let (x )D be a net in X such that x x in the d -topology, i.e., d (x x) 0 for
all B. Then,
(x x) C (d1 (x) + + dn (x)) 0
So is continuous with respect to the d -topology for all A, hence, d . A symmetric argument
shows d .


Definition 4. A family ( )A of seminorms on a vector space V is called directed if and only if for all
, A there is a A and a C > 0 so that (x) + (x) C (x) for all x V .

Definition 5. If X is a locally convex space then the topological dual, denoted by X , is the set of
continuous linear functionals on X with respect to the natural topology.

Definition 6. The Schwartz class, denoted by S(R), is the set of infinitely differentiable complex-valued
functions on R for which

kkn,m, := sup |xn Dm (x)| < for all n, m I+

xRn

where I+ = N {0}. It is easy to see that (k kn,m, )n,mI+ is a family of seminorms which separates points.

Definition 7. The Space of Tempered Distributions, denoted by S 0 (R), is the topological dual of
S(R).

Note. S(R) embeds (S 0 , S)-continuously into S 0 (R) where the (S 0 , S)-topology is the smallest topology
on S 0 (R) such that the maps {x : S 0 (R) C|x X} are continuous, where x (`) = `(x) for all ` S 0 (R).
Further, S(R) is dense in S 0 (R).

Definition 8. Let T S 0 (R) and n I+ . The weak nth derivative of T, denoted Dn T , is defined by

(Dn T ) (f ) = (1)n T (Dn f )

We are now ready to prove the N-Representation Theorem for S(R) and S 0 (R) after a couple more
definitions and lemmas.

Lemma 1. For n, m I+ define a seminorm k kn,m,2 on S(R) by

Z 1/2
kf kn,m,2 = |xn Dm f (x)|2 dx
R

Then the families of seminorms (k kn,m, )n,mI+ and (k kn,m,2 )n,mI+ on S(R) are equivalent.
2
MATT ZIEMKE THE REGULARITY THEOREM FOR DISTRIBUTIONS

Proof. Let f S(R) and let g(x) = (1 + x2 )1 . Then g L2 (R) and, for n, m I+ ,

kf kn,m,2 = kxn Dm f (x)k2

(1 + x2 )xn Dm f (x)
=k k2
1 + x2
!1/2
 (1 + x2 )xn Dm f (x) 2
Z  
= 
 dx

R
 1 + x2
1/2
k(1 + x2 )xn Dm f (x)k2
Z
dx
R |1 + x2 |2
= k(1 + x2 )xn Dm f (x)k kgk2
kgk2 kxn Dm f (x)k + kxn+2 Dm f (x)k

C (kf kn,m, + kf kn+2,m, )

Further, for any f S(R),

Z x  Z
f 0 (x)dx |f 0 (x)|dx = kf 0 k1 k(1 + x2 )f 0 k2 k(1 + x2 )1 k2
 
kf k = sup 
xR R

So we have that
d n m
kf kn,m, k(1 + x2 ) (x D f (x))k2 k(1 + x2 )1 k2
dx
= Ck(1 + x2 )(nxn1 Dm f (x) + xn Dm+1 f (x)k2
C knxn1 Dm f (x)k2 + kxn Dm+1 f (x)k2 + knxn+1 Dm f (x)k2 + kxn+2 Dm+1 f (x)k2

= C 0 (kf kn1,m,2 + kf kn,m+1,2 + kf kn+1,m,2 + kf kn+2,m+1,2 )

So, by Proposition 1, the families of seminorms are equivalent. 

Definition 9. Let A : S(R) S(R) and A : S(R) S(R) where

   
1 d 1 d
A= x+ and A = x
2 dx 2 dx
Let N = A A and define a seminorm k kn on S(R) by kf kn = k(N + 1)n f k2 . Further, define
 n
1/2 n 1/4 x2 /2 d 2
n
n (x) = (2 n!) (1) e ex
dx
The functions (n )nI+ are called the Hermite functions.

Lemma 2. The set (n )nI+ is an orthonormal basis for L2 (R).

2 2 2
d n x
Proof. Let Hn (x) = (1)n ex ( dx ) e and let w(x, t) = e2xtt . Then,
 n   n    n 
d  d x2 (xt)2  n x2 d u2 

w(x, t)
 = e e = (1) e e = Hn (x)
dt t=0 dt 
t=0 du 
u=x
3
THE REGULARITY THEOREM FOR DISTRIBUTIONS MATT ZIEMKE

So we have that

2 X Hn (x) n
(1) w(x, t) = e2xtt = t
n=0
n!

Further,
d
(w(x, t)) = (2x 2t)w(x, t)
dt
So, by substituting (1), we have

!
d X Hn (x) n X Hn (x) n X Hn (x) n
0= t 2x t + 2t t
dt n=0
n! n=0
n! n=0
n!

X Hn (x) n1 X Hn (x) n X Hn (x) n+1
= t 2x t +2 t
n=1
(n 1)! n=0
n! n=0
n!

X Hn (x) n X Hn (x) n X Hn1 (x) n
= t 2x t +2 t
n=0
n! n=0
n! n=1
(n 1)!

Then, by equating coefficients, we have

Hn+1 (x) 2xHn (x) 2Hn1 (x)
+ =0
n! n! (n 1)!
for all n 1 and hence,

(2) Hn+1 (x) 2xHn (x) + 2nHn1 (x) = 0

Similarly,
d
(w(x, t)) = 2tw(x, t)
dx
So, by substituting (1), we have

X Hn0 (x) n X Hn1 (x) n
0= t 2 t
n=0
n! n=1
(n 1)!

Then, by equating coefficients we get

(3) Hn0 (x) = 2nHn1 (x)

Now, substituting (3) into (2) we have

Hn+1 (x) 2xHn (x) + Hn0 (x) = 0

0
Hn+1 (x) 2Hn (x) 2xHn0 (x) + Hn00 (x) = 0
And, therefore,

(4) Hn00 (x) 2xHn0 (x) + 2nHn (x) = 0

4
MATT ZIEMKE THE REGULARITY THEOREM FOR DISTRIBUTIONS

2
Let un (x) = ex /2
Hn (x). Then,
2 2
u00n (x) = un (x) + x2 un (x) 2xex /2
Hn0 (x) + ex /2
Hn00 (x)
2
= (x2 1)un (x) + ex /2
(Hn00 (x) 2xHn0 (x))
2
= (x2 1)un (x) + ex /2
(2nHn (x)) by (4)
= (x2 2n 1)un (x)

So, we have that

(5) u00n (x) + (2n + 1 x2 )un (x) = 0

Now, let n, m I+ such that n 6= m. Then, by (5),

(6) um (x)u00n (x) + (2n + 1)um (x)un (x) = 0 and un (x)u00m (x) + (2m + 1)un (x)um (x) = 0

Then, by (6),
d
(u0 (x)um (x) u0m (x)un (x)) + 2(n m)um (x)un (x) = u00n (x)um (x) u00m (x)un (x) + 2(n m)um (x)un (x)
dx n
= (2n 1 + 2m + 1 + 2n 2m)un (x)um (x)
=0

So,
d
2(n m)um (x)un (x) = (u0 (x)um (x) u0m un (x))
dx n
And therefore, Z

2(n m) um (x)un (x)dx = (u0n (x)um (x) u0m (x)un (x)| = 0
R
2
Since u0n (x)um (x) u0m (x)un (x) = p(x)ex /2
0 as |x| where p is a polynomial of degree n + m + 1.
Therefore um and un are orthogonal. For n = m, first substitute n 1 forn in (2) and then multiply through
by Hn (x) to get

(7) Hn2 (x) 2xHn (x)Hn1 (x) + 2(n 1)Hn (x)Hn2 (x) = 0

for n 2. Similarly,
2
(8) Hn1 (x)Hn+1 (x) 2xHn1 (x)Hn (x) + 2nHn1 (x) = 0

Then, subtracting (8) from (7) we have

0 = Hn2 (x) + 2(n 1)Hn (x)Hn2 (x) Hn1 (x)Hn+1 (x) 2nHn1
2
(x)
2
Therefore, by multiplying through by ex /2 and integrating, we have
Z
u2n (x) + 2(n 1)un (x)un2 (x) un1 (x)un+1 (x) 2nu2n1 (x) dx


0=
ZR
u2N (x) 2nu2n1 (x) dx


= by orthogonality
R

5
THE REGULARITY THEOREM FOR DISTRIBUTIONS MATT ZIEMKE

Therefore,
Z Z
u2n (x)dx = 2n u2n1 (x)dx
R R
Z
= 4n(n 1) u2n2 (x)dx
R

Z
2
= 2n n! ex /2
dx
R

= 2n n!

Therefore, (n )nI+ is orthonormal. To prove the Hermite functions are an orthonormal basis for L2 (R) it
suffices to prove that if
Z
2
ex /2
Hn (x)f (x)dx = 0 for all n I+ then f = 0
R
2
ex /2
R
So, suppose R
Hn (x)f (x)dx = 0 for all n I+ . Then, for any t R,
it n Z


2
2
ex /2
Hn (x)f (x)dx = 0
n! R

and hence,
it n

Z
X 2
0= 2
ex /2
Hn (x)f (x)dx
n=0
n! R
 n
Hn (x) it
Z X
2
= ex /2 f (x)dx
R n=0 n! 2
Z
2 2
= etx+t /4 ex /2 f (x)dx
R

Therefore, Z
 2
 2
F ex /2 f (t) = ex /2 f (x)etx dx = 0
R
2
And since the Fourier Transform is an isometry on L2 (R) we have that ex /2
f (x) = 0 and therefore f = 0.


Lemma 3. The family of seminorms (k kn )nI+ is a directed family which is equivalent to the
(k kn,m,2 )n,mI+ family of seminorms on S(R).

Proof. Let Ao denote A or A . Our first goal is to prove the inequality

kAo(1) Ao(2) Ao(m) f k2 k(N + m)m/2 f k2

2
d n x2 /2
Let cn = (1)n (2n n! )1/2 so that n (x) = cn ex /2 ( dx ) e . Then
  n h  n h i
cn 2 d 2
i 2 d 2
An (x) = 2xex /2 ex /2 + ex /2 2xex /2
2 dx dx
6
MATT ZIEMKE THE REGULARITY THEOREM FOR DISTRIBUTIONS

Further,
 n h n  
d x2 /2
i X n 2
 2 2

2xe = 2 (x)() (ex )(n) = 2 x(ex )(n) + n(ex )(n1)
dx =0

So, we have that, !

n1 h


cn 2 d 2
i
An (x) = 2nex ex = nn1 (x)
2 dx
Also,
 n h  n+1 h
1 cn 2 d 2
i cn 2 d 2
i
A n (x) = xn (x) xex /2 ex ex /2 ex
2 2 dx 2 dx
n+1 h
(1)n

1 2 d 2
i
= 1/2
ex /2 ex
2 (2n n! ) dx

= n + 1n+1 (x)

Therefore, N n (x) = A An (x) = nn (x). Now, let f S(R). Then, by lemma 2, there exists (an )nI+ so
P
that f = n=0 an n . Then,

!1/2
o o o
X
2 2
kA(1) A(2) A(m) f k2 n + 1 n + 2 n + m an
n=0

!1/2
X
(n + m)m a2n
n=0

= k(N + m)m/2 f k2

Now, let n, m I+ and assume n > m. By our claim,

k(N + 1)n f k2 + k(N + 1)m f k2 Ck(N + 2n)2n f k2 + C 0 k(N + 2m)2m f k2

!1/2 !1/2

X
X
C 00 (k + 2n)2n a2k + (k + 2m)2m a2k
k=0 k=0
00 n
2C k(N + 2n) f k2
C 000 k(N + 1)n f k2

So, (k kn )nI+ is a directed family of seminorms. The fact that the seminorms (k kn )nI+ are equivalent
to the seminorms (k kn,m,2 )n,mI+ follows immediately from our claim and the equation
1
xf = (A + A ) f
2
and hence,
 m  k+m
d 1 k m
x k
f= (A + A ) (A A ) f
dx 2


We are now ready to prove the N-Representation Theorem for S(R) and S 0 (R).
7
THE REGULARITY THEOREM FOR DISTRIBUTIONS MATT ZIEMKE

Theorem 1 (The N-Representation Theorem for S(R)). Let s be the set of sequences (an )nI+ in C
with the property
sup |an |nm < for all m I+
nI+

Topologize s by defining the seminorms

X
k(an )nI+ k2m = (n + 1)2m |an |2
n=0
R
Let f S(R). Then the sequence (an )nI+ , where an = R
f (x)n (x)dx, is in s and the map f 7 (an )nI+
is a topological isomorphism.

R
Proof. Define : S(R) s by (f ) = (an )nI+ where an = R
f (x)n (x)dx. Let n I+ . From the proof of
lemma 2, we saw that N n = nn . Now, let f S(R). Since (n )n I+ is an orthonormal basis for L2 (R)
P
there exists (an )n I+ such that f = n=0 an n . Well,

X
X
m
an n n = an N m n = N m f L2 (R)
n=0 n=0

So,

X
|an |2 n2m = kN m f k22 <
n=0
hence,
sup |an |nm <
nI+

and therefore (an )nI+ s. Further,

X
X
kf km = k(N + 1)m f k2 = k an (N + 1)m n k2 = k an (n + 1)m n k2
n=0 n=0

And,

!1/2
X X
k an (n + 1)m n k2 = (n + 1)2 m|an |2 = k(an )nI+ km
n=0 n=0
Therefore, kf km = k(an )nI+ km . Further, since k km is actually a norm on S(R) we have that is injective.
PN
Let (an )nI+ . For N N, let fN = n=0 an n . Then, if N < M ,
M
!1/2
X
kfN fm km = k(N + 1)m (fN fM )k2 = (n + 1)2m |an |2 0
n=N +1

as N, M . Therefore (fN )N I+ is Cauchy in each k km and thus Cauchy in each k kn,m,2 by lemma 3
and hence Cauchy in each k kn,m, by lemma 1, i.e., (fN )N I+ is Cauchy in S(R). Therefore, there exists
P
f S(R) so that fN f in S(R) and hence fN f in L2 (R). Thus, f = n=0 an n and so is onto.
Lastly we want to show is a homeomorphism. If (fn )nI+ S(R) such that fn f with respect to k km
then
k(fn ) (f )km = k(an )nI+ (bn )nI+ km = kfn f km 0
8
MATT ZIEMKE THE REGULARITY THEOREM FOR DISTRIBUTIONS

So is continuous. Further, if ((am m

n )nI+ )mI+ s such that (an )nI+ (bn )nI+ as m with respect
to k km then

k 1 ((an )nI+ ) 1 ((bn )nI+ )km = kf m f km = k(an )nI+ (bn )nI+ km 0

So 1 is continuous and therefore is a topological isomorphism.



Theorem 2 (The N-Representation Theorem for S 0 (R)). Let T S 0 (R) and let bn = T (n ) for all
n I+ . Then for some m I+ , we have |bn | C(n + 1)m for all n I+ . Conversely, if |bn | C(n + 1)m
for all n I+ , there is a unique T S 0 (R) with T (n ) = bn . Further, if T S 0 (R) and bn = T (n ) then
P 0
n=0 bn n converges in the (S (R), S(R))-topology to T .

Proof. Let T S 0 (R). First, we want to show there exists m I+ and C > 0 such that |T ()| Ckkm for
all S(R). Well, T 1 (B(0, 1)) is open, where B(0, 1) is the ball centered at 0 of radius 1, so there exists
an open neighborhood of zero N T 1 (B(0, 1)) such that N nk=1 Umk ,xk ,k where
Umk ,xk ,k = {y S(R) : ky xk kmk < k }. Further, for all k = 1, ..., n we have that 0 Umk ,xk ,k so
there exists k > 0 such that Umk ,0,k Umk ,xk ,k . Then M = nk=1 Umk ,0,k is an open neighborhood of
zero and M if and only if kkmk < k for all k = 1, . . . , n. Also, note that M N T 1 (B(, 0, 1)).
Since (k kn )nI+ is directed, there exists m I+ and C > 0 so that kkm1 + + kkmn Ckkm for all
S(R). Let  = min{1 , . . . , n }. Then, for k = 1, . . . , n,
       
  
    + +   
       

 2Ckkm   2Ckkm   2Ckkm  C  2Ckkm  = 2 < k
 
mk m1 mn m


So, 2Ckkm M T 1 (B(0, 1)) hence
  

T  
1
 2Ckkm 
and therefore
2C
|T ()| kkm

Hence,
|bn | = |T (n )k Ckn km = C(n + 1)m
9
THE REGULARITY THEOREM FOR DISTRIBUTIONS MATT ZIEMKE

Conversely, suppose (bn )nI+ C such that |bn | C(n + 1)m for some m I+ . Let (an )nI+ s. Define
P
B : s C by B((an )nI+ ) = n=0 bn an . Then,

X
|B((an )nI+ )| |bn ||an |
n=0
X
C(n + 1)m |an |
n=0
X
C (n + 1)1 (n + 1)m+1 |an |
n=0

!1/2
!1/2
X
2m+2 2
X 1
(n + 1) |an |
n=0 n=0
(n + 1)2
r
2
=C k(an )nI+ km+1
6
So B is a continuous linear functional on s. Then, if is the topological isomorphism from S(R) into s, we
have that B S 0 (R). Hence, if T = B then

!
X X
T an n = B((an )nI+ ) = an bn
n=0 n=0
P
In particular, T (n ) = bn . Lastly, if bn = T (n ) and f S(R). Then we can write f = n=0 an n and

! Z !
X X
bn n (f ) = bn n (x) f (x)dx
n=0 R n=0

X Z
= bn n (x)f (x)dx
n=0 R

X
= T (n )an
n=0

!
X
=T a n n
n=0

= T (f )
P
Hence n=0 bn n = T in the (S 0 (R), S(R))-topology. 

Theorem 3 (Regularity Theorem for Distributions). Let T S 0 (R). Then T = Dn g for some
polynomially bounded continuous function g and n I+ , that is,
Z
T (f ) = (1)n g(x) (Dn f ) (x)dx for all f S(R)
R
0
Proof. Let T S (R). Let bn = T (n ). Then, by the last theorem, |bn | C(n + 1)m for some m I+ and
C > 0. Let an = bn (n + 1)(m+3) . Note that

kn k k0n k1 ck(1 + x2 )0n k c0 (n + 1)3/2

10
MATT ZIEMKE THE REGULARITY THEOREM FOR DISTRIBUTIONS

and so,

X X |bn |
|an |kk = kk
n=0 n=0
(n + 1)m+3

X C
kk
n=0
(n + 1)3

X c0

n=0
(n + 1)3/2

P
Therefore, n=0 an n converges uniformly to some continuous F on R.
Then we have that

X
(N + 1)m+3 F = (N + 1)m+3 an n
n=0

X bn
= (N + 1)m+3 n
n=0
(n + 1)m+3

X bn
= m+3
(N + 1)m+3 n
n=0
(n + 1)

X bn
= m+3
(n + 1)m+3 n
n=0
(n + 1)

X
= bn n
n=0

=T

where convergence is in the (S 0 (R), S(R))-topology by theorem 2. So, we have that T = (N + 1)m+3 F . It
remains to show that T = Dn g for some polynomially bounded continuous function g which is fairly easy to
convince ourselves of but quite tedious to prove formally. One simply has to do integration by parts many
times.


11
THE REGULARITY THEOREM FOR DISTRIBUTIONS MATT ZIEMKE

References

1. Lebedev, N. (1972), "Special Functions and Their Applications". New York: Dover Publications.

2. M. Reed and B. Simon (1972), "Methods of Modern Mathematical Physics", Volume 1. New York:
Academic Press.

3. Sunder, V.S. (1998), "Functional Analysis: Spectral Theory". Boston: Birkhauser Advanced Texts.

12