Lectures On FR Echet Spaces: Dietmar Vogt

Lectures on Fréchet spaces
Dietmar Vogt
Bergische Universität Wuppertal

Sommersemester 2000
Contents
1 Definition and basic properties 2
2 The dual space of a Fréchet space 6
3 Schwartz spaces and nuclear spaces 14
4 Diametral Dimension 33
5 Power series spaces and their related nuclearities 44
6 Tensor products and nuclearity 60

6.1 Tensor products with function spaces . . . . . . . . . . . . . . 71
6.2 Tensor products with spaces L1 (µ) . . . . . . . . . . . . . . . 73
6.3 The ε-tensor product . . . . . . . . . . . . . . . . . . . . . . . 79
6.4 Tensor products and nuclearity . . . . . . . . . . . . . . . . . 86
6.5 The kernel theorem . . . . . . . . . . . . . . . . . . . . . . . . 92
7 Interpolational Invariants 96
1
1 Definition and basic properties
All linear spaces will be over the scalar field K = C oder R.

Definition: A Fréchet space is a metrizable, complete locally convex vector
space.
We recall that a sequence (xn )n∈N in a topological vector space is a Cauchy

sequence if for every neighborhood of zero U there is n0 so that for n, m ≥ n0
we have xn − xm ∈ U . Of course a metrizable topological vector space is
complete if every Cauchy sequence is convergent.
If E is a vector space and A ⊂ E absolutely convex then the Minkowski

functional ∥ ∥A is defined as
∥x∥A = inf{ t > 0 | x ∈ tA }.
∥ ∥A is a extended
∪ real valued seminorm and ∥x∥A < ∞ if and only if
x ∈ span{A} = t>0 tA.
∩
ker ∥ ∥A = { x | ∥x∥A = 0 } = tA
t>0
is the largest linear space contained in A.

We have
{ x | ∥x∥A < 1 } ⊂ A ⊂ { x | ∥x∥A ≤ 1 }.
If p is a seminorm, A = { x | p(x) ≤ 1 } then ∥ ∥A = p.
We set
EA = ( span(A)/ ker ∥ ∥A , ∥ ∥A )∧ ,
where ”∧ ” denotes the completion. Notice that ∥x + y∥A = ∥x∥A for y ∈
ker ∥ ∥A . So ∥ ∥A defines a norm on span(A)/ ker ∥ ∥A .
If E is a topological vector space and A has an interior point, then ∥ ∥A is

continuous. In this case 0 is an interior point of A and
◦
A = { x | ∥x∥A < 1 }, Ā = { x | ∥x∥A ≤ 1 }.
A space is called locally convex if it has a basis of absolutely convex neigh-

borhoods of zero. Since for such a neighborhood of zero U , which may be
2
assumed open, ∥ ∥U is continuous and x ∈ U if, and only if, ∥x∥U < 1,
we have for a generalized sequence (xτ )τ ∈T that xτ −→ x if, and only if,
p(xτ − x) −→ 0 for every continuous seminorm.
Definition: A set P of continuous seminorms on the locally convex space

E is called fundamental system if for every continuous seminorm q there is
p ∈ P and C > 0 so that q ≤ C · p.
Of course, we have for every generalized sequence (xτ )τ ∈T that xτ −→ x if,

and only if, p(xτ − x) −→ 0 for all p ∈ P.
If a fundamental system of seminorms P is countable then we may assume

that P = { ∥ ∥k | x ∈ N } where ∥ ∥1 ≤ ∥ ∥2 ≤ . . .. This can be achieved
by setting ∥x∥k = max pj (x) where { pj | j ∈ N } is a given countable
j=1,...,k
fundamental system of seminorms.
1.1 Lemma: For a locally convex space E the following are equivalent:
(1) E is metrizable.
(2) E has a countable basis of neighborhoods of zero.
(3) E has a countable fundamental system of seminorms.
(4) The topology of E can be given by a translation invariant metric.
Proof: (1) ⇒ (2) and (4) ⇒ (1) are obvious.

To show (2) ⇒ (3) we assume that U1 ⊃ U2 ⊃ . . . is a basis of absolutely
convex neighborhoods of zero. We set ∥ ∥k := ∥ ∥Uk . Then all ∥ ∥k are
continuous seminorms. If p is a continuous seminorm, then we may choose
k such that Uk ⊂ { x | p(x) ≤ 1 }. This implies p(x) ≤ ∥x∥k for all x.
Finally, for (3) ⇒ (4) we put
∞
∑ 1 ∥x − y∥k
d(x, y) =
2k 1 + ∥x − y∥k .
k=1
It is an elementary exercise to show that d(·, ·) is a translation invariant

metric which gives the topology of E. 2
3
We notice that for a translation invariant metric d on a topological vector
space E is complete if and only if E is complete with respect to d.
For a not translation invariant metric this needs not to be the case as the
example of R with the metric d(x, y) = ∥ arctan x − arctan y∥ shows.
As in the case of Banach spaces one can show:
1.2 Theorem: If E is a Fréchet space and F ⊂ E a closed subspace then

F and E/F are Fréchet spaces.
∪
A subset M of a linear space E is called absorbant if t>0 tM = E. A topo-
logical vector space E is called barrelled if every closed, absolutely convex,
absorbant set (”barrel ”) is a neighborhood of zero.
1.3 Theorem: Every Fréchet space is barrelled.

∪
Proof: Obviously E = n∈N nM . By Baire’s theorem there is n0 so that
n0 M and therefore also M has an interior point. Then also 0 is an interior
point of M , which had to be proved. 2
1.4 Lemma: If E and F are locally convex, F a Fréchet space and A : E −→

F linear and surjective, then A is nearly open , i.e. for every neighborhood
of zero U ⊂ E the set AU is a neighborhood of zero in F .
Proof: We may assume that U is absolutely convex, hence AU is a barrel in

F . Theorem 1.3 yields the result. 2
As in the case of Banach spaces one can prove the following lemma.
1.5 Lemma (Schauder): If E and F are metrizable spaces, E complete,

and A : E −→ F linear, continuous and nearly open, then A is surjective
and open.
These two lemmas yield:
1.6 Theorem (Open Mapping Theorem): If E and F are Fréchet spaces,

A : E −→ F linear, continuous and surjective, then A is open.
1.7 Corollary (Banach’s Isomorphy Theorem): If E and F are Fréchet

spaces, A : E −→ F linear, continuous and bijective, then A−1 is continuous.
4
1.8 Theorem (Closed Graph Theorem): If E and F are Fréchet spaces,
A : E −→ F linear and Graph A := { (x, Ax) | x ∈ E } closed in E × F ,
then A is continuous.
Proof: We consider the maps π1 : (x, y) 7→ x, π2 : (x, y) 7→ y from Graph A

to E, resp. F , we notice that π1 is continuous linear and bijective, hence by
Corollary 1.7 π1−1 is continuous and therefore also A = π2 ◦ π1−1 . 2
If E is a Fréchet space, ∥ ∥1 ≤ ∥ ∥2 ≤ . . . a fundamental system of seminorms,

then the Banach spaces Ek := (E/∥ ∥k , ∥ ∥k )∧ are called ”local Banach
spaces ”.
Since ker ∥ ∥k+1 ⊂ ker ∥ ∥k we have natural quotient maps E/ker ∥ ∥k+1 −→
E/ker ∥ ∥k , which extend to continuous linear maps ıkk+1 : Ek+1 −→ Ek ,
(”linking maps ”). Of course we may define in a similar way ınm : Em −→ En
for m ≥ n and obtain ınm = ınn+1 ◦ . . . ◦ ım−1
m . By ı : E −→ Ek we denote the
k
quotient map. We have for m > n that ınm ◦ ım = ın . We define continuous

linear maps
ı : E −→ Πk Ek by ıx = (ık x)k∈N
σ : Πk Ek −→ Πk Ek by σ((xk )k∈N ) = (xk − ıkk+1 xk+1 )k∈N .
1.9 Theorem: By
ı σ
0 −−−−→ E −−−−→ Πk Ek −−−−→ Πk Ek −−−−→ 0
we obtain a short exact sequence (”canonical resolution”).
Proof: ı is clearly injective since ın = 0 for all k implies ∥x∥k = 0 for all k,
hence x = 0.
im ı ⊂ ker σ because ık x − ıkk+1 ık+1 x = ık x − ık x = 0. To prove ker σ ⊂ im ı
we notice that
ker σ = { (xk )k∈N | xk = ıkk+1 xk+1 for all k }.
This implies xn = ınm xm for all n ≤ m. Let x ∈ ker σ. For every xk we
choose ξk ∈ E so that ∥xk − ık ξk ∥k ≤ 2−k .
For k ≤ n < m we obtain
∥ξn − ξm ∥k = ∥ık ξn − ıkn xn + ıkm xm − ık ξm ∥k
≤ ∥ın ξn − xn ∥n + ∥xm − ım ξm ∥m
≤ 2−k+1 .
5
Therefore (ξn )n is a Cauchy sequence, hence convergent. We put x =
limk→∞ ξk . We have
∥xk − ık x∥k = lim ∥xk − ık ξn ∥k
n→∞
= lim ∥ıkn xn − ıkn ın ξn ∥k
n→∞
≤ lim ∥xn − ın ξn ∥n = 0.
n→∞
So xk = ık x for all k.
To prove surjectivity of σ we consider an element (0, . . . , 0, xk , 0 . . . ), xk ∈
Ek . There exists ξk ∈ E such that ∥xk − ık ξk ∥k < 2−k . We set yk =
xk − ık ξk ∈ Ek , then ∥yk ∥k < 2−k .
We define elements
ak = (0, . . . , 0, −ık+1 ξk , −ık+2 ξk , . . . )
bk = (ı1k yk , . . . , ık−1
k yk , yk , 0, . . . ).
Then we have
σ(ak ) = (0, . . . , 0, ık ξk , 0 . . . )
σ(bk ) = (0, . . . , 0, yk , 0, . . . ).
Now we put uk = ak +bk =: (η1k , η2k , . . . ). We obtain σ(uk ) = (0, . . . , 0, xk , 0 . . . )
and ∥ηjk ∥j ∑ < 2−k for j ≤ k. From there
= ∥ıjk yk ∥j ≤ ∥yk ∥k ∏ ∑ it is easily seen
that u := k uk converges in k Ek . We have σ(u) = k σ(uk ) = x. 2
2 The dual space of a Fréchet space
For any locally convex space E we use the following notation of polar sets:
For M ⊂ E we put
M ◦ = { y ∈ E ′ | |y(x)| ≤ 1 for all x ∈ M }
and for N ⊂ E ′ we put
N ◦ = { x ∈ E | |y(x)| ≤ 1 for all y ∈ N }.
If E ′ is equipped with a locally convex topology (e.g. the strong topology,

see below) then the second notation might be ambiguous. If there is any
danger of confusion we will specify whether we mean N ◦ ⊂ E or N ◦ ⊂ E ′′ .
We recall
6
2.1 Theorem (Bipolar Theorem): If E is locally convex and M ⊂ E
absolutely convex then M ◦ ◦ = M .
If p is a continuous seminorm on the locally convex space E and U = { x ∈

E | p(x) ≤ 1 } then we have
U ◦ = { y ∈ E ′ | sup |y(x)| ≤ 1 }.
x∈U
We put
p∗ (y) = sup |y(x)|
x∈U
and obtain
U ◦ = { y | p∗ (y) ≤ 1 }.
Therefore p∗ is the Minkowski functional of U ◦ .
If E is a Fréchet space and
∥ ∥1 ≤ ∥ ∥2 ≤ . . .
a fundamental system of seminorms, then
∥ ∥∗1 ≥ ∥ ∥∗2 ≥ . . .
is a decreasing sequence of extended real valued norms. For the unit balls
we have
Bk := { y | ∥y∥∗k ≤ 1 } = Uk◦ .
We set ∪
Ek∗ = span Uk◦ = tUk◦ .
t>0
Then Ek∗ is the space of linear form in E which are continuous with respect
to ∥ ∥k . (Ek∗ , ∥ ∥∗k ) is a normed space. The following remark shows that it
is even a Banach space.
′
Remark: ık : Ek′ −→ E ′ is an isometric isomorphism from Ek′ onto Ek∗ .
∗ ∗
∪ an ∗increasing sequence E1 ⊂ E2 ⊂ . . . of Banach spaces,
Therefore we have
′
such that E = k Ek .
7
Definition: E ′ is made into a locally convex space by the seminorms
pB (y) = sup |y(x)|

x∈B
where B runs through the bounded subsets of E.
We recall that a subset of a locally convex space is bounded if it is bounded

with respect to all continuous seminorms on E or, equivalently, with respect
to all seminorms of a fundamental system.
Since for y ∈ E ′ the function x 7→ |y(x)| is a seminorm, pB (y) = supx∈B |y(x)|
is finite for every y ∈ E and every bounded set B ⊂ E.
2.2 Theorem: Let E be metrizable locally convex and F locally convex.

If A : E → F is linear and A maps bounded sets into bounded sets then A
is continuous.
Proof: If xn −→ 0 then ∥xn ∥k −→ 0 for all k, where ∥ ∥1 ≤ ∥ ∥2 ≤ . . . is

a fundamental system of seminorms. There is a sequence (λn )n in K, such
that |λn | −→ +∞ and
∥λn xn ∥k = |λn |∥xn ∥k −→ 0
for all k. In particular B = { λn xn | n ∈ N } is bounded. Therefore A(B)

is bounded. If p is a continuous seminorm on F then
1 1
p(Axn ) = p(A(λn xn )) ≤ sup p(Ax) −→ 0.
|λn | |λn | x∈B n→∞
So Axn −→ 0 and A is continuous. 2
This holds, of course, in particular for F = K. That means, a linear form

on E is continuous if, and only if, it is bounded on bounded subsets of E.
2.3 Theorem: If E is metrizable, then E ′ is complete.
Proof: If (yτ )τ ∈T is a Cauchy net in E ′ , then (yτ (x))τ ∈T is a Cauchy net

for every x ∈ E. Hence for every x there is y(x) so that yτ (x) −→ y(x).
Clearly y is a linear form, which is bounded on bounded subsets of E ′ ,
therefore continuous, and pB (yτ − y) −→ 0. 2
8
As a consequence we see that for a metrizable locally convex space the
space E ′ is metrizable if and only if E is normed. This is because Baire’s
theorem implies that for some k the set kUk◦ has an interior point, hence is
a neighborhood of zero. So Um ◦ ⊂ C U ◦ for all m with suitable C
m k m and
the Bipolar Theorem implies that E is normed.
Definition: If E is locally convex then:
(1) the weak topology σ(E, E ′ ) on E is given by the fundamental system

of seminorms
pe (x) = sup |y(x)|
y∈e
where e runs through the finite subsets of E ′
(2) the weak∗ topology σ(E ′ , E) on E ′ is given by the fundamental system

of seminorms
pe (y) = sup |y(x)|
x∈e
where e runs through the finite subsets of E.
We recall that (E, σ(E, E ′ ))′ = E ′ and (E ′ , σ(E ′ , E))′ = E. And we have
the following theorem.
2.4 Theorem (Alaoglu-Bourbaki): If E is locally convex and U ⊂ E a

neighborhood of zero then U ◦ is σ(E ′ , E)-compact.
For the proof see [1, 23.5].
2.5 Theorem: If E is a Fréchet space then for M ⊂ E ′ the following are

equivalent:
(1) M is weak∗ bounded
(2) M is weak∗ relatively compact
(3) M is bounded
(4) There is k ∈ N, C > 0 such that M ⊂ CUk◦ .
9
Proof: (2) =⇒ (1) and (3) =⇒ (1) are obvious. (4) =⇒ (2) follows from the
Alaoglu - Bourbaki Theorem 2.4.
For (4) =⇒ (3) it is enough to show that Uk◦ is bounded. Let B ⊂ E be
bounded and y ∈ Uk◦ then
pB (y) ≤ ∥y∥∗k sup ∥x∥k ≤ sup ∥x∥k .

x∈B x∈B
So it remains to show that (1) =⇒ (4). Now M ◦ ⊂ E is obviously absolutely

convex and closed. For x ∈ E we have supy∈M |y(x)| = Cx < +∞, hence
x ∈ Cx M ◦ . Therefore M ◦ is absorbant. By Theorem 1.3 U := M ◦ is a
neighborhood of zero in E. Hence there are k ∈ N and C > 0 so that
C Uk ⊂ U . This implies
1
M ⊂ M ◦◦ = U ◦ ⊂ CUk◦ . 2
2.6 Corollary: If E is a Fréchet space then E ′ is a complete locally convex

space which has a countable fundamental system of bounded sets.
A set B of bounded sets in a locally convex space E is called a fundamental

system of bounded sets if for every bounded set B ⊂ E there is B0 ∈ B and
C > 0 so that B ⊂ CB0 .
A subset M ⊂ E, E locally convex, is called bornivorous if for every bounded
set B ⊂ E there is t > 0 so that B ⊂ tM .
Definition: E is called bornological if every absolutely convex bornivorous
set is a neighborhood of zero.
Remark: For a locally convex space the following are equivalent
(1) E is bornological
(2) for every locally convex space G and every linear map A : E −→ G
which maps bounded sets into bounded sets, A is continuous
(3) for every Banach space G and every linear map A : E −→ G which
maps bounded sets into bounded sets, A is continuous.
Proof: (1) =⇒ (2) If U is an absolutely convex neighborhood of zero in G

and A maps bounded sets into bounded sets, then A−1 U is locally convex
and bornivorous. Hence it is a neighborhood of zero.
10
(2) =⇒ (3) is obvious.
(3) =⇒ (1) If M ⊂ E is absolutely convex and bornivorous, in particular
absorbant, the space EM = (E/ker ∥ ∥M , ∥ ∥M )∧ is a Banach space. The
quotient map φ : E −→ EM maps bounded sets into bounded sets so, by
assumption, it is continuous. Therefore M ⊃ { x | ∥φx∥M < 1 } is a
neighborhood of zero. 2
Example: By Theorem∪2.2 every metrizable locally convex space is bornolog-

ical. The dual E ′ = ∗
k Ek of a Fréchet space E is bornological if the
following holds:
A linear map A : E ′ −→ G, G locally convex is continuous if, and only if,

the restriction A|Ek∗ is continuous Ek∗ −→ G for all k.
2.7 Theorem: Let E be metrizable locally convex. Then the following are
equivalent:
(1) E ′ is bornological
(2) E ′ is barrelled.
Proof: (1) =⇒ (2) is obvious.

(2) =⇒ (1) Let M ⊂ E ′ be absolutely convex and bornivorous. We put
Bn = Un◦ for all n ∈ N.
For every k there is εk > 0 such that 2εk Bk ⊂ M . Therefore
∪
n
1
Cn := Γ εk Bk ⊂ M
2
k=1
for all n ∈ N. Here Γ denotes the absolutely convex hull.

Cn is weak∗ -compact. This is because Cn is the image of the weak∗ -continuous
map
{ z ∈ Kn | |z|ℓ1 ≤ 1 } × B1 × . . . × Bn −→ E ′
given by
∑
n
(z, x1 , . . . , xn ) 7→ zk εk xk .
k=1
11
Therefore Cn is weak∗ -closed, hence closed in E ′ . We set
∞
∪ ∪ 1
D := Γ εk Bk = Cn ⊂ M.
n
2
k=1
We want to show that D ⊂ 2D, because if we have shown this, then D ⊂ M .

Clearly D is a barrel (i. e. absolutely convex, closed and absorbant) so, by
assumption, a neighborhood of zero, and we are done.
So let x0 ∈ E \ 2D. Since Cn is closed there is an absolutely convex, weak∗ -
closed neighborhood of zero Vn in E ′ (that is Vn = B ◦ , B ⊂ E bounded),
such that
(x0 + Vn ) ∩ 2Cn = ∅.
Therefore
(x0 + Vn + Cn ) ∩ Cn = ∅.
Since Vn is weak∗ -closed and Cn is weak∗ -compact, the set Vn + Cn is weak∗ -
closed.
Therefore ∩
W = (Vn + Cn )
n
∗
is weak -closed, hence closed.
W is a barrel. It is clearly absolutely convex, closed by the previous and
absorbant by the following:
For x ∈ E ′ we have x ∈ λBk for certain λ and k. Therefore x ∈ ελk Cn
for all n ≥ k. There is µ > 0 such that x ∈ µVn for n = 1, . . . , k − 1, so
x ∈ max( ελk , µ)W .
Since E ′ , by assumption, is barrelled W is a neighborhood of zero. We have
(x0 + W ) ∩ Cn = ∅ for all n, therefore (x0 + W ) ∩ D = ∅ which implies
x ̸∈ D. 2
Definition: A Fréchet space is called distinguished if the equivalent condi-

tions of Theorem 2.5 are satisfied.
We refer to [1, page 270] and recall that we have a canonical imbedding
J : E ,→ E ′ given by (J(x))[y] = y(x) for x ∈ E and y ∈ E ′ . By Theorem
2.5 a fundamental system of seminorms in E ′′ is given by
∥z∥k = sup |z(y)|, k ∈ N.

y∈Uk◦
12
Therefore ∥Jx∥k = ∥x∥k and J imbeds E topologically into E ′′ .
Definition: The Fréchet space E is called reflexive if JE = E ′′ .
If there is no danger of confusion we will omit J in future and consider E in

a natural way as a subspace of E ′′ .
2.8 Theorem: If the Fréchet space E is reflexive then it is distinguished.
Proof: We show that E ′ is barrelled. Let U be a barrel in E ′ . Then, due to

the Bipolar Theorem 2.1, U ◦◦ = U . Here the first polar U ◦ has to be taken
in E ′′ . However, due to reflexivity, we may take it in E. Since U ∩ Ek∗ is a
barrel in the Banach space Ek∗ it is a neighborhood of zero there, so there is
εk > 0 with εk Uk◦ ⊂ U and therefore U ◦ ⊂ ε1k Uk . This holds for all k, so U ◦
is bounded in E and therefore U a neighborhood of zero in E ′ . 2
A criterion for reflexivity is:
2.9 Theorem: The Fréchet space E is reflexive if, and only if, every bounded
set in E is relatively weakly compact.
Proof: Let E be reflexive and B ⊂ E bounded. Then B ⊂ B ◦◦ and B ◦

is a neighborhood of zero in E ′ . So, due to Theorem 2.4, (B ◦ )◦ ⊂ E ′′ is
σ(E ′′ , E ′ )-compact. Due to reflexivity, we may take B ◦◦ ⊂ E and it is
σ(E, E ′ ) = σ(E ′′ , E ′ )-compact.
To prove the reverse take z ∈ E ′′ , then there is a neighborhood of zero
U ⊂ E ′ so that z ∈ U ◦ . We may assume U = B ◦ where B ⊂ E is bounded,
absolutely convex and closed, hence, due to B ◦◦ = B, weakly closed. So B
is σ(E, E ′ )-compact.
Now, considering polars (except B ◦ ) with respect to the duality E ′ , E ′′ we
have
σ(E ′′ , E ′ )
U ◦ = (B ◦ )◦ = JB = JB.
This is because J is a topological imbedding, also with respect to the topolo-
gies σ(E, E ′ ) and σ(E ′′ , E ′ ) , hence JB is σ(E ′′ , E ′ )-compact, and therefore
closed. 2
2.10 Corollary: If E is a reflexive Fréchet space and F ⊂ E a closed

subspace, then F is reflexive.
13
Proof: Due to Theorem 2.1, which implies F ◦◦ = F , the space F is weakly
σ(E, E ′ )
closed. If B ⊂ F is bounded, then M := B ⊂ F . Because of Theo-
rem 2.9 M is σ(E, E ′ )- compact and therefore, due to the Hahn - Banach
theorem, also σ(F, F ′ )-compact. 2
∏
Example: Let Ek⊕ , k ∈ N, be Banach spaces and E = k Ek .∑It is easily
verified, that E ′ = k Ek′ , equipped with the seminorms p(x) = ∞ ∗
k=1 ck ∥xk ∥k
for x = ⊕∞k=1 xk , where (ck )k runs through all positive sequences.
∏
Again, it is easily verified that E ′′ = k Ek′′ . Therefore E is reflexive if, and
only if, Ek is reflexive for all k.
2.11 Lemma: If the Fréchet space E has a fundamental system of semi-

norms such that all local Banach spaces Ek are reflexive, then E is reflexive.
∏
can be considered as a closed subspace of k Ek and, due to the
Proof: E ∏
previous, k Ek is reflexive. So Corollary 2.10 gives the result. 2
Definition: The Fréchet space E is called Fréchet-Hilbert space if there

exists a fundamental system of seminorms, such that for all k and x ∈ E we
have ∥x∥2k = ⟨x, x⟩k where ⟨·, ·⟩k is a semiscalar product on E.
Remark: In this case all local Banach spaces Ek are Hilbert spaces.
2.12 Theorem: A Fréchet-Hilbert space is reflexive hence distinguished.
2.13 Lemma: If E is a Fréchet-Hilbert space, F ⊂ E a closed subspace,

then F and E/F are Fréchet-Hilbert spaces.
Proof: This is easily verified, since subspaces and quotient spaces of Hilbert
spaces are Hilbert spaces. 2
2.14 Corollary: If E is a Fréchet-Hilbert space, F ⊂ E a closed subspace,

then F and E/F are reflexive, hence distinguished.
3 Schwartz spaces and nuclear spaces
Let X be a linear space over K and V, U absolutely convex. We use the

following notation: V ≺ U if there is t > 0 so that V ⊂ tU .
14
If V ≺ U then V is called U -precompact if∪for every ε > 0 there are finitely
many elements x1 , . . . , xm such that V ⊂ m
j=1 (xj + εU ).
If we denote by EV , EU the local Banach spaces with respect to ∥ ∥V , ∥ ∥U ,

respectively, that is
EV = (E/ker ∥ ∥V , ∥ ∥V )∧ , EU = (E/ker ∥ ∥U , ∥ ∥U )∧ ,
and denote by ıU V : EV −→ EU the canonical map, then V is U -precompact

if, and only i, ıU
V is compact.
Definition: A locally convex space E is called Schwartz space if for every

absolutely convex neighborhood of zero U there is an absolutely convex
neighborhood of zero V ≺ U , which is U -precompact.
By use of the previous remarks we can state:

3.1 Proposition: The locally convex space E is a Schwartz space if, and
only if, for every absolutely convex neighborhood of zero U there is an
absolutely convex neighborhood of zero V ⊂ U , so that the canonical map
V : EV −→ EU is compact.
ıU
3.2 Lemma: In a complete Schwartz space E all bounded sets are relatively
compact.
Proof: Let U be the set of all absolutely convex neighborhoods of zero in

E. For U ∈ U let EU be the local Banach space, ıU : E −→ EU the quotient
map. Then ∏
ı : E −→ EU , x 7→ (ıU x)U ∈U
U ∈U
∏
is a topological imbedding. Since E is complete, ıE is closed in U ∈U EU .
Let B ⊂ E be bounded. For U ∈ U we find V ∈ U , so that ıU V : EV −→
U U V
EU is compact. Hence ı B = ıV (ı B) is relatively compact in EU . By
∏
Tychonoff’s theorem ıB is relatively compact in U ∈U EU and therefore in
ıE. So B is relatively compact in E. 2
This yields immediately:

3.3 Theorem: Every Fréchet-Schwartz space is reflexive, hence distinguished.
Proof: The bounded sets are relatively compact, hence relatively weakly
compact, so by 2.9 the space is reflexive and by 2.8 it is distinguished. 2
15
We will now try to understand better the structure of compact sets in a
Fréchet space. In the following Γ denotes the absolutely convex hull of a
set.
3.4 Lemma: If E is locally convex and complete and (xk )k∈N a weak null
sequence in E, then
{ ∞ ∞
}
∑ ∑
Γ{ xk | k ∈ N } = ξk xk | |ξk | ≤ 1
k=1 k=1
and this set is weakly compact.
Proof: We define a continuous linear map φ : ℓ1 −→ E by

∑
φ(ξ) := ξj xj , ξ = (ξj )j∈N ∈ ℓ1 .
j
Existence follows from the completeness and continuity from the bounded-
ness of { xk | k ∈ N } and the triangle inequality.
For y ∈ E ′ and ξ ∈ ℓ1 we have
∑
y(φ(ξ)) = ξj y(xj )
j
and (y(xj ))j∈N is, by assumption, a null sequence. This shows that
φ : (ℓ1 , σ(ℓ1 , c0 )) −→ (E, σ(E, E ′ ))
is continuous. Since, by Theorem 2.4, the closed unit ball Uℓ1 of ℓ1 is

σ(ℓ1 , c0 )-compact, the set
{ }
∑ ∑
φ(Uℓ1 ) = ξk xk | |ξk | ≤ 1
k k
is weakly compact, hence weakly closed, hence closed in E. Since Γ{ xk |

k ∈ N } ⊂ φ(Uℓ1 ) this implies
{ ∞ ∞
}
∑ ∑
Γ{ xk | k ∈ N } ⊂ ξk xk | |ξk | ≤ 1 .
k=1 k=1
The reverse inclusion is obvious. 2
16
3.5 Corollary: If E is locally convex and complete and (xk )k∈N is a null
sequence then
{ ∞ ∞
}
∑ ∑
Γ{ xk | k ∈ N } = ξk xk | |ξk | ≤ 1
k=1 k=1
is compact.
Proof: We assume first that E is a Banach space. We consider the map

φ : ℓ1 −→ E defined by
∞
∑
φ(ξ) = ξk xk , ξ = (ξk )k ∈ ℓ1
k=1
and for n ∈ N
∑
n
φn (ξ) = ξk xk , ξ = (ξk )k ∈ ℓ1 .
k=1
Obviously
∥φ − φn ∥ ≤ sup ∥xk ∥ −→ 0.
k>n n→∞
Therefore φ is compact, hence φ({ ξ | ∥ξ∥ℓ1 ≤ 1 }) relatively compact and,

because it is closed by Lemma 3.4, compact.
In the general case we conclude from ∏ the previous that, in the notation of
the proof of Lemma 3.2, ı ◦ φ(Uℓ1 ) ⊂ U ∈U EU is compact. Since, due to
the completeness of E, the space ı(E) is closed and ı is an isomorphism onto
its range, we obtain the result. 2
3.6 Lemma: If E is a Fréchet space and K ⊂ E compact, then there is a

null sequence (xk )k∈N so that
{ ∞ ∞
}
∑ ∑
K⊂ ξk xk | |ξk | ≤ 1 .
k=1 k=1
Proof: Let U1 ⊃ U2 ⊃ . . . be a basis of closed absolutely convex neighbor-

hoods of zero. We set up an inductive procedure and put K1 = K. If the
(k) (k)
compact set Kk is determined, we find x1 , . . . , xmk ∈ Kk , such that
mk (
∪ )
(k) 1
Kk ⊂ xj + Uk
4k+1
j=1
17
and put
mk (
∪ )
(k) 1
Kk+1 = Kk − xj ∩ Uk .
4k+1
j=1
∈ 4−k Uk−1 for k > 1, j = 1 . . . mk , the sequence

(k)
Since xj
(1) (2) (k)

(2x1 , . . . , 2x(1) (2) k k (k)
m1 , 4x1 , . . . , 4xm2 , . . . , 2 x1 , . . . , 2 xmk , . . .)
is a null sequence.
(1)
For x ∈ K we find xj1 so that
(1) 1
x = xj1 + u1 , u1 ∈ U1 .
42
Therefore ( ) 1
(1) (1)
u1 = x − xj1 ∈ K − xj1 ∩ 2 U1 ⊂ K2 .
4
Proceeding inductively we find a sequence jk , jk ∈ { 1, . . . , mk }, such that
∞
∑ ∞
∑
(k) 1 k (k)
x= xjk = (2 xjk ).
2k
k=1 k=1
∑∞ ∑∞
In particular x ∈ { k=1 ξk xk | k=1 |ξk | ≤ 1 }. 2
As an immediate consequence of the previous we obtain
3.7 Lemma: If E is a Fréchet space, K ⊂ E compact, then there is an

absolutely convex compact set L ⊂ E, so that, K ⊂ L and K is compact in
the Banach space EL .
Proof: We choose a null sequence (xk )k∈N such that

{ ∞ }
∑ ∑
K⊂ ξk xk | |ξk | ≤ 1 =: K1
k=1 k
and a sequence 1 ≤ λ1 ≤ λ2 ≤ . . ., λk → ∞ such that λk xk → 0 (cf. the

proof of 2.2). We put
{ ∞ }
∑ ∑
L := ξk λk xk | |ξk | ≤ 1 .
k=1 k
18
Then, due to Corollary 3.5, L is absolutely convex and compact in E, clearly
K1 ⊂ L. Since ∥xk ∥L ≤ λ1k we have xk → 0 in EL .
Therefore K1 is compact in EL , hence K relatively compact in EL . Since
K is closed in E and therefore in EL , K is compact in EL . 2
3.8 Theorem: If E is a Fréchet-Schwartz space then E ′ is a Schwartz space.

◦
Proof: Let U = K be a neighborhood of zero in E ′ . K may be assumed
absolutely convex and, due to 3.2, compact. We choose L according to
◦
Lemma 3.7 and put V = L. Since EK ,→ EL is compact also the canonical
map EV′ −→ EU′ is compact. 2
The above conclusion comes from the following consequence of the Arzelá-
Ascoli Theorem.
3.9 Theorem (Schauder): Let X, Y be Banach spaces and A : X −→ Y

continuous and linear. Then A is compact if and only if A′ is compact.
Definition: A locally convex space E is nuclear if, and only if, for every
absolutely convex neighborhood of zero U there is a neighborhood of zero
V and a finite positive Radon measure on the weak∗ -compact set V ◦ , such
that ∫
∥x∥U ≤ |y(x)|dµ(y)
V◦
for all x ∈ E.
Example: Let Ω ⊂ Rn be open, ω1 ⊂⊂ ω2 ⊂⊂ . . . an exhaustion by open

sets. Then C ∞ (Ω) is a Fréchet space with the seminorms
∥f ∥k = sup |f (α) (x)|.

x∈ωk
|α|≤k
To show that C ∞ (Ω) is nuclear we notice that for every y ∈ D(Rn ) we have
∫
⃗
sup |g(x)| ≤ |g (1) (x)|dx
x∈Rn Rn
where ⃗1 = (1, . . . , 1).
19
For any k we choose φ ∈ D(ωk+1 ), φ ≡ 1 on ωk and obtain
∫
⃗
∥f ∥k ≤ sup |(φf )(α+1) (x)|dx
|α|≤k ωk+1
∑ ∫ ⃗
≤ |(φf )(α+1) (x)|dx
|α|≤k ωk+1
∑ ∫
≤ |f (β) (x)|hβ (x)dx
|β|≤k+n ωk+n
with suitable nonnegative hβ ∈ C (Rn ) with compact support in ωn+1 . We

set
Mβ = { δx(β) | x ∈ ω k+n }
and ∪
◦
M= Mβ ⊂ Uk+n
|β|≤k+n
and define a measure on ◦

Uk+n by
∑ ∫
µφ = φ(δx(β) )hβ (x)dx
|β|≤k+n ωk+n
◦ ). Since the summands define measures on the disjoint

for φ ∈ C (Uk+n
compact sets Mβ we may rewrite the above estimate for f ∈ C ∞ (Ω) as
∫
∥f ∥k ≤ |y(f )|dµ(y).
◦
Uk+n
3.10 Lemma: If E is nuclear then it admits a fundamental system of

Hilbert seminorms.
Proof: Let U be an absolutely convex neighborhood of zero. Then we find V

and µ according to the definition of nuclearity, V may be assumed absolutely
convex. We obtain:
∫
∥x∥U ≤ |y(x)|dµ(x)
V◦
(∫ )1
2
◦ 1
≤ µ(V ) 2 |y(x)| dµ(y)
2
V◦
◦
≤ µ(V )∥x∥V .
20
Therefore the seminorms
(∫ )1
2
pV (x) := |y(x)| dµ(y)
2
V◦
which come from the semi-scalar product

∫
⟨x1 , x2 ⟩ = y(x1 )y(x2 )dµ(y)
V◦
are a fundamental system of Hilbert seminorms. 2
3.11 Corollary: Every nuclear Fréchet space is a Fréchet-Hilbert space.
3.12 Lemma: If X, Y are Hilbert spaces, U = { x ∈ X | ∥x∥ ≤ 1 },

A : X −→ Y linear, such that
∫
∥Ax∥ ≤ |⟨x, y⟩|dµ(y)
U
for a finite positive measure y, then A is a Hilbert-Schmidt operator.
Proof: Let (ei )i∈I be an orthonormal basis in X and M ⊂ I finite. Then
∑ ∑ (∫ )2
∥Aei ∥2
≤ |⟨ei , y⟩|dµ(y)
i∈M i∈M U
∑∫
≤ µ(U ) |⟨ei , y⟩|2 dµ(y)
i∈M U
≤ µ(U ) . 2
Therefore ∑
∥Aei ∥2 < ∞
i∈I
and A is Hilbert-Schmidt. 2
3.13 Corollary: Every nuclear space is a Schwartz space.
3.14 Theorem: If E is nuclear and F ⊂ E a closed subspace, then F and

E/F are nuclear.
21
Proof: Let U be an absolutely convex neighborhood of zero in E. We choose
V and µ according to the definition of nuclearity. To show the nuclearity of
F we consider the quotient map q : E ′ −→ E ′ /F ◦ = F ′ and set W ◦ = q V ◦ .
We notice that it corresponds to (V ∩ F )◦ in F ′ . By
∫ ∫
f (y)dν(y) := f (q(y))dµ(y)
W◦ V◦
for f ∈ C (W ◦ ) we get a positive finite measure ν on W ◦ , such that for x ∈ F

∫ ∫
∥x∥U ≤ |y(x)|dµ(y) = |η(x)|dν(η).
V◦ W◦
To show the nuclearity of E/F we choose V as given by a Hilbert seminorm.

Then EV is a Hilbert space. We notice that the inequality in the definition
of nuclearity extends to
∫
∥ıU
V x∥U ≤ |⟨x, y⟩|dµ(y)
Vb
for all x ∈ EV where Vb is the unit ball of EV and ⟨·, ·⟩ the scalar product of
EV . Let π be the orthogonal projection onto (ıV F )⊥ ⊂ EV . For x ∈ E we
set x̂ = x + F . By ∥ ∥eU we denote the quotient seminorm on E/F of ∥ ∥U .
We obtain
∫ ∫
e
∥x̂∥U ≤ ∥ıV πı x∥U ≤
U V
|⟨πı x, y⟩|dµ(y) =
V
|⟨ıV x, πy⟩|dµ(y).
Vb Vb
∥ ∥V
The first inequality holds because ıV x − πıV x ∈ ıV F , and therefore
∥ ∥U
ıU x − ıU
V πı x ∈ ı F
V U . Clearly the last integral depends only on x̂.
For φ ∈ C (Vb ∩ (ıV F )⊥ ) we set
∫
ν(φ) := φ(πy)dµ(y).
Vb
Then ν is a positive finite measure on Vb ∩ (ıV F )⊥ . The elements in this set

are the Riesz representations under ⟨·, ·⟩ of the elements in V ◦ ∩F ◦ = Q(V )◦ ,
Q : E −→ E/F the quotient map. We have
∫
e
∥x̂∥U ≤ |y(x̂)|dν(y).
Q(V )◦ 2
22
Example: Since C ∞ (Ω) is nuclear for any open Ω ⊂ Rn , so is every closed
subspace. Therefore the space h(Ω) of harmonic functions with the compact
open topology (which coincides with the topology inherited from C ∞ (Ω)) is
nuclear.
Definition: Let X, Y be Banach spaces, A : X −→ Y a linear map. Then

A is called nuclear if there are sequences ξj ∈ X ′ , ηj ∈ Y such that
∑
(1) ∥ξj ∥∗ ∥ηj ∥ < +∞
j
∑
(2) Ax = j ξj (x)ηj for all x ∈ E.
Clearly a nuclear map is continuous and, as a norm limit of finite dimensional

maps, compact. We recall (see e.g. [1, Lemma 16.21]):
3.15 Theorem: If X, Y, Z are Hilbert spaces, A : X −→ Y, B : Y −→ Z

Hilbert-Schmidt operators then BA is nuclear.
We obtain the following characterization of nuclear spaces.
3.16 Theorem: For a locally convex space E the following are equivalent:
(1) E is nuclear.
(2) E has a fundamental system U of neighborhoods of zero given by

Hilbert seminorms and for every U ∈ U there is V ∈ U , V ⊂ U , such
that the canonical map EV −→ EU is Hilbert-Schmidt.
(3) For every absolutely convex neighborhood of zero U there is an abso-

lutely convex neighborhood of zero V ⊂ U , such that the canonical
map EV −→ EU is nuclear.
Proof: (1) =⇒ (2) follows from 3.10 and 3.12.

(2) =⇒ (3) It is obviously sufficient to show (3) for U ∈ U . We choose
V ∈ U , so that EV −→ EU is Hilbert-Schmidt and then W ∈ U so that
EW −→ EV is Hilbert-Schmidt. Due to 3.15 EW −→ EU is nuclear.
V : EV −→ EU be nuclear, i.e.
(3) =⇒ (1) Let ıU
∞
∑
ıU
Vx= ξj (x)ηj
j=1
23
∑
with ∥ξj ∥∗V ≤ 1 for all j and j ∥ηj ∥U < +∞. We define a measure on V ◦
by setting for φ ∈ C (V ◦ )
∞
∑
µ(φ) = φ(ξj )∥ηj ∥U .
j=1
We obtain ∫
∞
∑
∥x∥U ≤ |ξj (x)|∥yj ∥U = |y(x)|dµ(y).
j=1 V◦
Therefore (3) implies that E is nuclear. 2
For a further description of nuclear Fréchet spaces we need the following

lemma.
3.17 Lemma: A locally convex space E is nuclear if, and only if, for every
neighborhood of zero U there exists a sequence (yj )j∈N in E ′ with following
properties:
(1) U ◦ ⊂ Γ{ yj | j ∈ N }
(2) For every k there is a neighborhood of zero Uk ⊂ E, such that the set
{ j k yj | j ∈ N } is contained in Uk◦ .
Proof: We first assume that E is nuclear and we may assume that U is an

absolutely convex, closed Hilbert disk. We find inductively a sequence of
absolutely convex, closed neighborhoods of zero such that
(a) U0 = U
(b) Uk is a Hilbert disk
(c) ıU
Uk+1 : Uuk+1 −→ Uuk is nuclear.
k
We set Ek := EUk , ıkk+1 := ıUk k+1

Uk+1 and jk
∗ . Then
:= (ıkk+1 )′ : Ek∗ −→ Ek+1
∗ ) (see [1,
also maps jkk+1 are nuclear, i.e. in the Schatten 1-class S1 (Ek∗ , Ek+1
Lemma 16.7, (2.)]). So for the singular numbers sn (jkk+1 ) we have
∑
sn (jkk+1 ) < +∞.
n
24
Since the sequence of singular numbers is decreasing we have
lim nsn (jkk+1 ) = 0

n→∞
′
for all k. From [1, Lemma 16.6, (7.)] we conclude that for jk := ı0k =
k
jk−1 ◦ . . . ◦ j01 : E0∗ −→ Ek∗ we obtain
lim nk sn (jk ) = 0.
n→∞
Let
∞
∑
jk (x) = sn (jk )⟨x, ekn ⟩0 fnk
n=0
be the Schmidt-representation of jk , (ekn )n∈N and (fnk )n∈N orthonormal sys-

tems in E0∗ and Ek∗ respectively. (ekn )n is complete since jk is injective. We
arrange the double indexed sequence (ekn )n∈N to a sequence (ẽn )n∈N in the
following way
(ẽ0 , ẽ1 , . . .) = (e00 , e10 , e01 , e20 , e11 , e02 , e30 , e21 , e12 , e03 , e40 , . . .).
We apply the Gram-Schmidt orthogonalization procedure to the sequence

(ẽn )n∈N , throwing away linearly dependent vectors in the order they appear,
and obtain an orthonormal basis (en )n∈N0 of E0∗ .
Obviously we have em ⊥ekn for
1
m > 1 + 2 + . . . + (k + n) = (k + n)(k + n + 1). (1)
2
√ √
We fix k. Then ⟨em , ekn ⟩0 = 0 for n < 2m − k − 1, therefore for n < [ m]
if m is large enough.
So we obtain for m large enough
∑
∥em ∥∗k 2 = s2n (jk )|⟨em , ekn ⟩0 |2
√
n≥[ m]
∑
≤ s2[√m] (jk ) |⟨em , ekn ⟩0 |2
n
= s2[√m] (jk ).
Consequently
√
lim sup ∥mk em ∥∗2k ≤ lim [ m]2k s[√m] (j2k ) = 0.
m m
25
This means that for suitable Ck > 0
◦
{ mk em | m ∈ N0 } ⊂ Ck U2k .
For y ∈ U ◦ we have
∞
∑
y= ⟨y, em ⟩0 em
m=0
where the series converges in E0∗ , hence in E ′ .

∑
We put ym = λ(m + 1)em , λ2 = ∞ 1
m=0 (m+1)2 and ξm =
1
λ(m+1) ⟨y, em ⟩0 .
Then for y ∈ U ◦ we have
∑∞
y= ξm ym
m=0
with
∞
∑ ∞
∑ 1
|ξm | = |⟨y, em ⟩0 | ≤ 1.
λ(m + 1)
m=0 m=0
and
◦
mk ym ∈ Dk U2k+2
for all m and k with suitable Dk .
For the reverse direction we choose (yj )j∈N according to the assumption,
and obtain
∥x∥U = sup |y(x)| ≤ sup |yj (x)|

y∈U ◦ j∈N
∑ ∑
≤ |yj (x)| = j −2 |(j 2 yj (x)|.
j∈N j∈N
We put V = U2 , then j 2 yj ∈ V ◦ for all j. We define a measure on V ◦ by

setting ∑
µ(φ) = j −2 φ(j 2 yj )
j∈N
for φ ∈ C (V ◦ ), and obtain

∫
∥x∥U ≤ |y(x)|dµ(y).
V◦ 2
26
Remark: The previous argument shows that the following is sufficient for
the nuclearity of E: for every U there are V , a sequence (yj )j∈N in E ′ and a
∑
positive sequence (λj )j∈N with j λ1j < ∞, such that U ◦ ⊂ Γ{ yj | j ∈ N }
and λj yj ∈ V ◦ for all j.
Definition: A sequence (yj )j∈N in E ′ is called rapidly decreasing if for every

k there is a neighborhood of zero U so that { j k yj | j ∈ N } ⊂ U ◦ .
Remark: From Theorem 2.5 we easily derive that a sequence (yj )j∈N in
the dual space E ′ of a Fréchet space E is rapidly decreasing if, and only if,
j k yj −→ 0 for all k.
If E is locally convex then we may define a locally convex topology tN

as given by the seminorms py (x) = supj |yj (x)| where y runs through the
rapidly decreasing sequences in E ′ . It is easily seen to be nuclear and it
is the finest nuclear locally convex topology on E coarser than the given
topology. It is called the associated nuclear topology and Lemma 3.17 says
that E is nuclear if and only if its topology coincides with the associated
nuclear topology.
We use Lemma 3.17 to show that every nuclear Fréchet space is isomorphic
to a subspace of sN or likewise of C ∞ (R). So these spaces are universal spaces
for the nuclear Fréchet space, which had been conjectured by Grothendieck.
Definition: We set
∑
s = { ξ = (ξj )j∈N | |ξ|2k = j 2k |ξj |2 < ∞ for all k }
j
= { ξ = (ξj )j∈N | ∥ξ∥k = sup j k |ξj | < ∞ for all k }.

j
The norm systems (| |k )k∈N0 and (∥ ∥k )k∈N0 are easily seen to be equivalent.
s equipped with the topology given by these systems is a nuclear space. This
is because the local Banach space for | |k is
∑
sk = { ξ = (ξj )j∈N | |ξ|2k = j 2k |ξj |2 < ∞ }
j
and the canonical map sk+1 −→ sk is the identical imbedding which is

Hilbert-Schmidt, since
∞
∑ 1 ( )
ıkk+1 x =x= ⟨x, (n + 1)−k−1 en+1 ⟩k+1 (n + 1)−k en+1
n+1
n=0
27
is the Schmidt representation where en := (0, . . . , 0, 1, 0, . . .) and ⟨ , ⟩k+1 is
the scalar product for | |k+1 .
We obtain the following theorem, which solves a conjecture of Grothendieck:
3.18 Theorem (T. and Y. Komura): A Fréchet space E is nuclear if,

and only if, it is isomorphic to a subspace of sN .
Proof: Since s is nuclear, so is sN and therefore every subspace. This proves

one implication.
For the reverse implication we choose a fundamental system of seminorms
∥ ∥1 ≤ ∥ ∥2 ≤ . . . and for Uk = { x ∈ E | ∥x∥k ≤ 1 } a rapidly decreasing
sequence y (k) in E ′ , so that
Uk◦ ⊂ Γ{ yj ◦
(k) (k)
| j ∈ N }, { j ν yj | j ∈ N } ⊂ Ck,ν Uµ(k,ν) .
We define a map φk : E −→ s by
(k)
φk (x) = (yj (x))j∈N .
We have to explain that this is a continuous map into s.

∑ (k) 1
|φk (x)|m = ( j 2m |yj (x)|2 ) 2
j
∑
j −2 ) 2 sup j m+1 |yj (x)|
1 (k)
≤ (
j j
∑
j −2 ) 2 Ck,m+1 ∥x∥µ(k,m+1) .
1
≤ (
j
Moreover we have
(k)
∥x∥k ≤ sup |yj (x)| ≤ |φj (x)|0 .
j
From these estimates we conclude that
φ(x) := (φk (x))k∈N
defines a topological imbedding into sN . 2
We shall prove that sN ∼

= C ∞ (Ω) for every open Ω ⊂ Rn . For that we need
the following lemmas.
28
3.19 Lemma: If X is complemented subspace of sN then X × sN ∼
= sN .
Proof: Let X × Y ∼
= sN then we have
sN ∼
= (sN )N ∼
= (X × Y )N ∼
= X × XN × Y N ∼
= X × (X × Y )N ∼
= X × sN .
The isomorphims are self explaining. 2
3.20 Lemma: If E is isomorphic to a complemented subspace of sN and sN

is isomorphic to a complemented subspace of E, then E ∼
= sN .
Proof: Let X be a complement for sN in E. Then E ∼ = X × sN . Let Y

be a complement of E in s , then Y × E = s . Therefore sN ∼
N ∼ N
= Y ×E ∼ =
X × (Y × s ). So X is isomorphic to a complemented subspace of sN and
N
therefore E ∼
= X × sN ∼
= sN by Lemma 3.19. 2
For r = (r1 , . . . , rn ), rj > 0 for all j, we set

Cr∞ = { f ∈ C ∞ (Rn ) | f (x + r) = f (x) for all x }.
This is a closed subspace of C ∞ (Rn ) hence a nuclear Fréchet space. For
∞ = C∞
r > 0 we put C(r) (r,...,r) .
3.21 Theorem: Cr∞ ∼

= s.
Proof: Obviously we may assume r = (2π, . . . , 2π). The topology induced

by C ∞ (Rn ) can be expressed by the seminorms
∥f ∥k = sup{ |f (α) (x)| | |α| ≤ k, 0 ≤ xj ≤ 2π, j = 1, . . . , n }
or, equivalently, by the seminorms
∫ ∫
1 ∑ 2π 2π (α)
|f |k =
2
|f (x)|2 dx1 . . . dxn ,
(2π)n 0 0
|α|≤k
which, in terms of the Fourier expansion

∑
f (x) = cν eiνx , νx = ν1 x1 + . . . + νn xn ,
ν∈Zn
can be expressed as
∑ ∑
|f |2k = |cν |2 ν 2α
|α|≤k ν∈Zn
∑ ∑
= |cν |2 ( ν 2α ).
ν∈Zn |α|≤k
29
Here ν α = ν1α1 . . . νnαn .
Since
∑ k!
(1 + ν12 + . . . + νn2 )k = ν 2α
α1 ! . . . αn !(k − |α|)!
|α|≤k
we obtain for ν ̸= 0
∑ ∑
ν 2α ≤ (1 + ν12 + . . . + νn2 )k ≤ (n + 1)k sup ν 2α ≤ (n + 1)k ν 2α .
|α|≤k |α|≤k |α|≤k
Moreover, we have with |ν| = |ν1 | + · · · + |νn |

(1 + |ν|)k ≤ (1 + ν12 + . . . + νn2 )k ≤ (1 + |ν|)2k .
Therefore norm system |f |k , k ∈ N0 is equivalent to the norm system
∑
|f |∼
2
k := |cν |2 (1 + |ν|)2k
ν∈Zn
 
∞
∑ ∑
=  |cν |2  (1 + m)2k , k ∈ N0 .
m=0 |ν|=m
If Mm = { ν ∈ Z | |ν| ≤ m } then we have for all m ∈ N0

( [m] )n
2 + 1 ≤ #Mm ≤ (2m + 1)n .
n
We count the ν ∈ Z with increasing |ν|. Let ν ∈ Zn carry the number ℓ(ν).
We obtain for |ν| = m ≥ 1
( [ ] )n
m−1
2 +1 ≤ #Mm−1 < ℓ(ν) ≤ #Mm ≤ (2m + 1)n .
n
Therefore we have
1
(m + 1)n ≤ ℓ(ν) ≤ 2n (m + 1)n .
nn
where the left inequality holds for m ≥ n + 1, the right one for all m. We
set bℓ = cℓ(ν) for ℓ = ℓ(ν). Then
 
∞
∑ ∑
1 
|f |∼ |cν |2 ℓ(ν)2 n 
2 k
k ≥ k
4
m=0 |ν|=m
∞
∑
1 k
= |bℓ |2 ℓ2 n
4k
ℓ=1
30
and therefore
 
∞
∑ ∑
∼2  |cν |2 ℓ(ν) 
2k
|f |k ≤ n2k n
m=0 |ν|=m
∞
∑ 2k
≤ n2k |bℓ |2 ℓ n .
ℓ=1
Therefore ( )1
∞
∑ 2
n−k |f |∼
k ≤ |bℓ |2 ℓ2k ≤ 4nk |f |∼
nk
ℓ=1
which proves that f 7→ (bℓ )ℓ∈N defines an isomorphism of Cr∞ into s, which
is clearly surjective since the image contains the finite sequences. 2
3.22 Theorem: If Ω ∈ Rn is open then C ∞ (Ω) ∼

= sN .
Proof: We use Lemma 3.20 and show first that sN is isomorphic to a com-
plemented subspace of C ∞ (Ω). ∑ We choose a function φ ∈ D(R), with φ ≥ 0
and supp φ ⊂ [0, 3], such that k∈Z φ2 (x − k) ≡ 1. This can be achieved
∫2 1
by setting φ(x) = ( 1 χ(x − t)dt) 2 where χ ∈ D(R), supp χ ⊂ [−1, +1],
∫ +∞
−∞ χ(t)dt = 1. For r > 0 and x = (x2 , . . . , xn ) ∈ R we put φr (x) =
n
x1 xn
φ( r ) . . . φ( r ). Then we obtain
∑
φ2r (x − rk) ≡ 1.
k∈Zn
We choose a discrete sequence x1 , x2 , . . . ∈ Ω and functions ψj (x) := φrj (x−

xj ) such that supp ψj ⊂ Ω for all j and supp ψj ∩ supp ψk = ∅ for j ̸= k.
We define a continuous linear map
∏
Φ: C(r∞j ) −→ C ∞ (Ω)
j
by ∑
Φ((fj )j∈N ) = ψj fj
j
and we define a continuous linear map

∏
Ψ : C ∞ (Ω) −→ C(r∞j )
j
31
by ( )
∑
Ψ(f ) = ψj (x − rj k)f (x − rj k) .
k∈Zn j∈N
∏ ∞
For (fj )j∈N ∈ j C(rj ) we obtain for the i-th component gi of Ψ(Φ(fj )j∈N ))
∑ ∑
gi (x) = ψi (x − ri k) ψj (x − ri k)fj (x − ri k)
k∈Zn j
∑
= ψi2 (x − ri k)fi (x − ri k)
k∈Zn
= fi (x).
Therefore Ψ ◦ Φ = id. Notice that by Theorem 3.21 C(r∞j ) ∼

= s for all j.
Now we show that C ∞ (Ω) is isomorphic to a complemented subspace of sN .

∑
We choose functions φj ∈ D(Ω) such that j φ2j ≡ 1 and each x ∈ Ω has
a neighborhood which meets only finitely many of the supp φj . We choose
rj > 0 such that supp φj is contained in the interior of a period cube. We
define a map ∏
Φ : C ∞ (Ω) −→ C(r∞j )
j
by ( )
∑
Φ(f ) = φj (x − rj k)f (x − rj k) .
k∈Zn j∈N
Notice that the summands above have disjoint supports. And we define a
continuous linear map
∏
Ψ: C(r∞j ) −→ C ∞ (Ω)
j
by ∑
Ψ((fj )j∈N ) = φj fj .
j
32
For f ∈ C ∞ (Ω) we obtain
∑ ∑
Ψ(Φf )(x) = φj (x) φj (x − rj k)f (x − rj k)
j k∈Zn
∑∑
= φj (x)φj (x − rj k)f (x − rj k)
j k∈Zn
∑
= φ2j (x)f (x)
j
= f (x).
∞
∏ ∞ Ψ∼◦ ΦN= id and C (Ω) is isomorphic to a complemented subspace
Therefore
of j C(rj ) = s . 2
For both sides of the proof we used the following fact:
3.23 Lemma: Let X, Y be topological vector spaces, Φ : X −→ Y, Ψ : Y −→

X continuous linear and Ψ ◦ Φ = idX . Then X is isomorphic to a comple-
mented subspace of Y .
Proof: Φ : X −→ im Φ is an isomorphism since Ψ|im Φ is its continuous linear

inverse map. Φ◦Ψ is a projection in Y , since Φ◦Ψ◦Φ◦Ψ = Φ◦id ◦Ψ = Φ◦Ψ.
Because Ψ is surjective we have im(Φ ◦ Ψ) = im Φ. 2
Finally we have proved
3.24 Theorem: E is a nuclear Fréchet space if, and only if, it is isomorphic
to a closed subspace of C ∞ (R) (resp. C ∞ (Ω), Ω ⊂ Rn open).
4 Diametral Dimension
Let E be a linear space over K. We consider absolutely convex subsets

U, V ⊂ E.
We recall the following notation:
V ≺ U if there is t > 0 such that V ⊂ tU . For V ≺ U and any linear
subspace F ⊂ E we set
δ(V, U ; F ) = inf { δ > 0 | V ⊂ δU + F }.
33
Definition: For V ≺ U and n ∈ N0 the number
δn (V, U ) = inf { δ(V, U ; F ) | dim F ≤ n }
is called the n-th Kolmogoroff diameter of V with respect to U .
4.1 Lemma: We have for any V ≺ U :
(1) δn (V, U ) ≥ δn+1 (V, U ) for all n.
(2) δn (V, U ) → 0 if and only if V is precompact with respect to U .
Proof: (1) is obvious. To show (2) we first assume that V is precompact

with respect to∪U . Then we have for every ε > 0 elements x1 , . . . , xm
such that V ⊂ m j=1 (xj + εU ). Putting F = span { x1 , . . . , xm } we obtain
δ(V, U ; F ) ≤ ε and therefore δn (V, U ) ≤ ε for n ≥ m.
To prove the reverse implication we assume that δn (V, U ) < ε. Then there
exists a subspace F ⊂ E, dim F ≤ n, such that
V ⊂ εU + F.
We may assume F ⊂ span U . Namely if v = εu + f we have f = εu − v ∈

span U , so if necessary we can replace F by F ∩ span U .
We may assume that ∥ ∥U is a norm on F . For otherwise we set F0 =
F ∩ ker ∥ ∥U and F = F0 ⊕ F1 . Then V ⊂ εU + F1 , and we can replace F
by F1 .
Since V ≺ U there is t > 0 such that V ⊂ tU hence, arguing as above, we
obtain
V ⊂ εU + F ∩ (t + ε)U
= εU + M.
Since M = F ∩ (t + ε)U is bounded in the finite∪dimensional space (F, ∥ ∥U )

we may choose x1 , . . . , xm ∈ F such that M ⊂ m
j=1 (xj + εU ) and therefore
∪
m
V ⊂ (xj + 2εU ).
j=1 2
34
4.2 Corollary: A locally convex space E is a Schwartz space if and only
if for every neighborhood of zero U there is a neighborhood of zero V ≺ U
such that δn (V, U ) → 0.
For the special case of Hilbert discs the Kolmogoroff diameters coincide with
the singular numbers of the canonical imbedding. This is contained in the
following lemma.
4.3 Lemma: Let X, Y be Hilbert spaces, UX , UY their unit balls, T : X −→

Y continuous and linear. Then
(1) T is compact if, and only if, δn (T UX , UY ) → 0.
(2) If T is compact and

∞
∑
Tx = sn ⟨x, φn ⟩en
n=0
its Schmidt representation then sn = δn (T UX , UY ).
Proof: (1) is a consequence of Lemma 4.1 so it remains to show (2).

We set F = span { e0 , . . . , em−1 }. Then dim F = m and for x ∈ UX we have
∞
∑
Tx ∈ sn ⟨x, φn ⟩en + F
n=m
and
∞
∑ ∞
∑ ∞
∑
∥ sn ⟨x, φn ⟩en ∥ =
2
s2n |⟨x, φn ⟩| ≤
2
s2m |⟨x, φn ⟩|2 ≤ s2m .
n=m n=m n=m
Therefore δm (T UX , UY ) ≤ sm .
The reverse inequality is trivial if sm = 0, so we assume that sm > 0 and
T UX ⊂ δUY + F, dim F ≤ m. (2)
Let h1 , . . . , hµ , µ ≤ m, be a basis of F . There is a nontrivial solution

y ∈ span { e0 , . . . , em } of the equations ⟨y, hj ⟩ = 0, j = 1, . . . , µ. Hence
y ∈ F ⊥.
35
We may choose ξ0 , . . . , ξm such that:
∑
m ∑
m
y= sn ξn en = T ( ξn en ) = T x. (3)
n=0 n=0
Since x ̸= 0 we may set x′ = tx and y ′ = ty where t = 1/∥x∥ and we still

have
y ′ = T x′ and y ′ ⊥F.
So we may assume (3) with ∥x∥ = 1. For f ∈ F we obtain
∥T x − f ∥2 = ∥y − f ∥2 = ∥y∥2 + ∥f ∥2
∑m ∑
m
≥ s2n |ξn |2 ≥ s2m |ξn |2
n=0 n=0
= s2m .
Therefore we have δ ≥ sm . Since δ and F , fulfilling (2), were arbitrarily

chosen we have shown that
δm (T UX , UY ) ≥ sm .
This proves assertion (2). 2
An easy consequence of the definition is the following:
4.4 Lemma: For any V, U we have:
(1) If Vo ⊂ V ≺ U ⊂ Uo then δn (Vo , Uo ) ≤ δn (V, U ) for all n.
(2) If W ≺ V ≺ U then δn+m (W, U ) ≤ δn (W, V ) δm (V, U ) for all n, m.
Proof: (1) is obvious. To show (2) we assume that
W ⊂ δV + F, dim F ≤ n
′
V ⊂ δ U + G, dim G ≤ m.
Then we have
W ⊆ δδ ′ U + F + G.
Since dim (F + G) ≤ n + m we obtain the assertion. 2
36
We want now to express nuclearity in terms of the Kolmogoroff diameters.
One implication of the intended characterization we have almost done.
4.5 Proposition: If E is nuclear then for every neighborhood of zero U

and k ∈ N there is a neighborhood of zero V ⊂ U , such that
lim nk δn (V, U ) = 0.
n→∞
Proof: We prove it by induction over k.

For k = 1 we choose a hilbertian zero neighborhood U1 ⊂ U and then a
hilbertian zero neighborhood V ⊂ U1 such that jVU1 is nuclear. Then, due to
Lemma 4.3 and Lemma 4.4, we have sn (jVU1 ) = δn (V, U1 ) ≥ δn (V, U ) and
therefore ∑
δn (V, U ) < +∞.
n
Since the sequence of the δn (V, U ) is decreasing, this implies
lim n δn (V, U ) = 0.
n
Assume that we found V1 ⊂ U , such that limn nk δn (V1 , U ) = 0, then we

find V ⊂ V1 , such that limn n δn (V, V1 ) = 0 and therefore, due to Lemma
4.4,
lim nk+1 δ2n (V, U ) = 0
n
which implies
lim nk+1 δ2n+1 (V, U ) = 0.
n
This implies the assertion. 2
To prove the converse of Proposition 4.5 we recall some facts about nuclear
operators between Banach spaces.
Let X, Y Banach spaces then we denote by N (X, Y ) the linear space of
nuclear operators. With the nuclear norm
∑
ν(A) := inf { ∥ξk ∥∗ ∥ηk ∥ | ξk ∈ X ′ , ηk ∈ Y for all k,
k
∑
Ax = ξk (x)yk for all x }
k
N (X, Y ) is a Banach space (see [1]).
37
Every operator A : X −→ Y with finite dimensional range is nuclear and
∥A∥ ≤ ν(A) ≤ n ∥A∥
where n = dim R(A).

While the left estimate follows from the triangular inequality, the right one
follows from Auerbach’s lemma [1, 10.5]. Let e1 , . . . , en be an Auerbach
basis of Y0 = R(A), and f1 , . . . , fn ∈ Y ′ such that
fj (ek ) = δj,k , ∥e1 ∥ = . . . = ∥en ∥ = ∥f1 ∥∗ = . . . = ∥fn ∥∗ = 1.
Then
∑
n
Ax = fk (Ax)ek
k=1
and
∑
n
ν(A) ≤ ∥fk ◦ A∥ ∥ek ∥ ≤ n ∥A∥. (4)
k=1
We obtain:
4.6 Lemma: Let X, Y be Banach spaces, UX , UY their unit balls, A : X −→

Y continuous and linear with
∑
n2 δn (AUX , UY ) < +∞.
n
Then A is nuclear.
∑ 2ε
Proof: We choose a decreasing sequence εn > δn (AUX , UY ) such that nn n <
+∞.
For every n there is a subspace Fn ⊂ Y , dim Fn = mn ≤ n, such that
AUX ⊂ εn UY + Fn . (5)
Let en1 , . . . , enmn , f1n , . . . , fm

n be an Auerbach basis of F . Then
n n
∑
mn
Pn x := fkn (x)enk
k=1
is a projection onto Fn with ∥Pn ∥ ≤ n,
38
If we apply id −Pn to the inclusion (5) and notice that ∥ id −Pn ∥ ≤ n + 1
we see that
(id −Pn )AUX ⊂ (n + 1)εn UY
or
∥A − Pn A∥ ≤ (n + 1)εn . (6)
We put P0 = 0 and Un = Pn+1 A − Pn A. Then dim R(Un ) ≤ 2n + 1 and,
because of (6),
∥Un ∥ = ∥(A − Pn A) − (A − Pn+1 A)∥ ≤ (2n + 1)εn .
Therefore, due to (4),

ν(Un ) ≤ (2n + 1)2 εn .
and therefore ∑
ν(Un ) < +∞.
k
Hence there is U ∈ N (X, Y ) such that

∑
U= Un
n
in N (X, Y ) and therefore also in L(X, Y ). Since, because of (6),
∑
m−1
Un = Pm A −→ A
m→∞
n=0
in L(X, Y ) we have A = U ∈ N (X, Y ). 2
4.7 Theorem: A locally convex space E is nuclear if, and only if, for every
neighborhood of zero U and every k ∈ N there is a neighborhood of zero V ,
such that
lim nk δn (V, U ) = 0.
n→∞
Proof: One implication is Proposition 4.5, to prove the other one we choose
V for k = 4 and obtain
∑
n2 δn (V, U ) < +∞.
n
We apply Lemma 4.6 to ıU V : EV −→ EU taking into account that obviously

δn (V, U ) = δn (ıU V b ) where Vb , U
b, U b are the closed unit balls in EV and EU ,
V
respectively. 2
39
Since in the proof of Proposition
∑ 4.5 we used from nuclearity only that
for every U there is V with n δn (V, U ) < +∞ we have also proved the
following theorem.
4.8 Theorem: A locally convex space E is nuclear if, and only if, for every
neighborhood of zero U there is a neighborhood of zero V ⊂ U , such that
∑
δn (V, U ) < +∞.
n
The description of Schwartz spaces and nuclear spaces in 4.2 and 4.5 by
means of Kolmogoroff diameters motivate the following definition.
Definition: For a locally convex space E the set
∆(E) = { t = (t0 , t1 , . . .) ∈ KN
0 | ∀U ∃V : lim tn δn (V, U ) = 0 }
n
is called the diametral dimension of E.
Remark: In 4.2 and 4.5 there was shown:
(1) E is a Schwartz space if, and only if, (1, 1, 1, . . .) ∈ ∆(E).
(2) E is nuclear if, and only if, (nk )n ∈ ∆(E) for all k.
4.9 Theorem: The diametral dimension is a linear topological invariant,

that is, if E ∼
= F then ∆(E) = ∆(F ).
Proof: This follows from the obvious fact that for a linear bijective T
δn (T V, T U ) = δn (V, U ). 2
Remark: If U is a fundamental system of absolutely convex neighborhoods

of zero then
∆(E) = { t = (t0 , t1 , . . .) | ∀U ∈ U ∃V ∈ U : tn δn (V, U ) → 0 }.
Here U is called fundamental system of neighborhoods of zero if for every

neighborhood of zero W there is U ∈ U and ε > 0 such that εU ⊂ W , that
is, if the ∥ ∥U are a fundamental system of seminorms.
If V ≺ U are absolutely convex subsets of E and T : E −→ G a linear map

then clearly T V ≺ T U and δn (T V, T U ) ≤ δn (V, U ). This yields:
40
4.10 Theorem: If F ⊂ E is a closed subspace then ∆(E/F ) ⊃ ∆(E).
Proof: Let (t0 , t1 , . . .) ∈ ∆(E) and Û = q U an absolutely convex neighbor-

hood of zero in E/F , then we find a neighborhood V of zero in E, V ≺ U ,
such that tn δn (V, U ) → 0 and therefore tn δn (qV, qU ) → 0. 2
To handle the case of subspaces we introduce another notion of relative

diameters.
Let V ≺ U be absolutely convex subsets of E, and G ⊂ E a linear subspace.
We set
γ(V, U ; G) := inf{ γ > 0 | V ∩ G ⊂ γU }.
Definition: For V ≺ U and n ∈ N0 the number
γn (V, U ) = inf{ γ(V, U ; G) | codim G ≤ n }
is called the n-th Gelfand diameter of V with respect to U .
A comparison of the Kolmogoroff and Gelfand diameters is given in the

following lemma.
4.11 Lemma: Let V ≺ U be absolutely convex closed neighborhoods of

zero in E.
(a) We have
δn (V, U ) ≤ (n + 1)γn (V, U )

γn (V, U ) ≤ (n + 1)δn (V, U ).
for all n ∈ N0 .
(b) If V and U are Hilbert discs then
δn (V, U ) = γn (V, U )
for all n ∈ N0 .
Proof: Let V ⊂ δU + F , dim F ≤ n. As in the proof of Lemma 4.1 we may

assume that ∥ ∥U is a norm on F . Let P be a projection from E onto F
with
∥P ∥U = sup ∥P x∥U < +∞.
x∈U
41
We set G = ker P , then codim G = dim F ≤ n. For x ∈ V ∩ G we obtain
x = (id −P )x ∈ δ∥ id −P ∥U U,
hence
V ∩ G ⊂ δ∥ id −P ∥U U.
This proves that
γn (V, U ) ≤ ∥ id −P ∥U δn (V, U ).
In the general case we may use Auerbach’s lemma to get P with ∥P ∥U ≤ n
hence ∥ id −P ∥U ≤ 1 + n. If U is a Hilbert disc we may choose for P the
orthogonal projection, so ∥ id −P ∥U ≤ 1.
In a similar way we prove the reverse estimate. Let V ∩ G ⊂ γU , codimG ≤
n. Without restriction of generality we may assume G closed with respect
to ∥ ∥V . Let P be a ∥ ∥V - continuous projection with ker P = G. We set
F := im P . Then we have for x ∈ V
x = (id −P )x + P x ∈ ∥ id −P ∥V (V ∩ G) + F ⊂ ∥ id −P ∥V γU + F
hence
V ⊂ γ∥ id −P ∥V U + F.
This proves that
δn (V, U ) ≤ ∥ id −P ∥V γn (V, U ).
Arguing as previously we complete the proof. 2
Definition: For a locally convex space E we put
Γ(E) := { t = (t0 , t1 , . . .) ∈ KN
0 | ∀U ∃V lim tn γn (V, U ) = 0 }.
n
From Lemma 4.11 we obtain:
4.12 Proposition: If E has a fundamental system of Hilbert seminorms,

then ∆(E) = Γ(E).
If V ≺ U are absolutely convex subsets of E and F ⊂ E a linear subspace

then clearly V ∩ F ≺ U ∩ F and γn (V ∩ F, U ∩ F ) ≤ γn (V, U ). This yields
quite easily:
4.13 Lemma: If F ⊂ E is a linear subspace then Γ(F ) ⊃ Γ(E).
42
Proof: Let (t0 , t1 , . . . ) ∈ Γ(E) and U an absolute convex neighborhood of
zero in F . We may find an absolutely convex neighborhood of zero U0 ⊂ E,
such that U0 ∩F ⊂ U and V0 ≺ U0 , such that limn tn γn (V0 , U0 ) = 0. Putting
V = V0 ∩ F we have V ≺ U and limn tn γn (V, U ) = 0. 2
4.14 Theorem: If E has a fundamental system of Hilbert seminorms and

F ⊂ E is a subspace then ∆(F ) ⊃ ∆(E).
Proof: This follows from proposition 4.12 and lemma 4.13. 2
Examples: (1) If dim E = m < +∞ and V ≺ U in E then δn (V, U ) = 0

for n ≥ m. In particular ∆(E) = ω.
(2) If E = ω and U a neighborhood of zero then there is m ∈ N so that

for all δ > 0
E = δU + span{e1 , . . . , em }
where ej are the canonical unit vectors. This means that for every
V ≺ U we have δm (V, U ) = 0. Therefore ∆(ω) = ω.
(3) If E is infinitely dimensional and a non Schwartzian locally convex

space then there is U so that no V ≺ U is U -precompact . That
means inf n δn (V, U ) > 0. Therefore ∆(E) = c0 .
(4) If E = s and
∞
∑
Uk = {ξ | |ξ|2k = |ξj |2 j 2k < +∞}
j=1
then the local Banach space is sk (see ...?...). Let ıkm : sm ,→ sk be

the canonical imbedding. Then
∞
∑ 1 1 1
ıkm ξ = ξ = m−k
⟨ξ, m
ej+1 ⟩m ej+1
(j + 1) (j + 1) (j + 1)k
j=0
is the Schmidt representation. Therefore

1
δn (Um , Uk ) = sn (ıkm ) = .
(n + 1)m−k
From this we derive easily that
∆(s) = {t = (t0 , t1 , . . .) | ∃ m ∈ N, C > 0 such that |tn | ≤ Cnm for all n}.
43
Theorem 4.7 we may then read as follows:
4.15 Theorem: E is nuclear if, and only if, ∆(E) ⊃ ∆(s).
Now we can continue the series of examples by:
(5) ∆(sN ) = ∆(s). Since sN is nuclear we have ∆(sN ) ⊃ ∆(s) (see 4.15).
Since s is a quotient space of sN we have ∆(s) ⊃ ∆(sN ) (see 4.10).
(6) If Ω ⊂ Rn is open then ∆(C ∞ (Ω)) = s. This is because C ∞ (Ω) ∼

= sN
(see 3.22).
At this point it seems that the diametral dimensional is not very significant
for the isomorphy type of a nuclear Fréchet space. We will try to get a
better understanding in the next section.
5 Power series spaces and their related nucleari-

ties
We start with a short introduction to the theory of Köthe sequence spaces.

We concentrate on the Fréchet Hilbert version of these.
Let A = (aj,k )j∈N0 ,k∈N0 be a matrix such that
(1) 0 ≤ aj,k ≤ aj,k+1 for all j, k
(2) ∀j ∃ k : aj,k > 0.
A is called a Köthe-matrix.
We put
∞
∑
λ(A) := {x = (x0 , x1 , . . .) | |x|2k := |xj |2 a2j,k < +∞ for all k}.
j=0
It is easily seen that λ(A) is a Fréchet-Hilbert space with the semiscalar

products ∑
⟨x, y⟩k = xj ȳj a2j,k .
j
44
To determine the local Banach spaces we put
Ik = {j | aj,k > 0}.
Then ∑
λk (A) := {x = (xj )j∈Ik | |x|2k = |xj |2 a2j,k < +∞}
j∈Ik
is the local Banach space with respect to | |k . For m > k we have Im ⊃ Ik

and the canonical map ıkm : λm (A) −→ λk (A) is given by ıkm : (xj )j∈Im 7→
(xj )j∈Ik . Similar as in the case of s we may write ıkm as
∑ aj,k 1 ( 1 )
ıkm x = ⟨x, ej ⟩m ej .
aj,m aj,m aj,k
j∈Ik
( ) ( )
1 1
Here ej is the canonical j-th basis vector, aj,m ej , and aj,k ej
j∈Ik j∈Ik
are orthonormal sequences in λm (A) and λk (A), respectively.
a
The map ıkm is compact if and only if ( aj,m
j,k
)j∈Im is a null sequence and for
Uk = {x | |x|k ≤ 1} we obtain
δn (Um , Uk ) = sn (ıkm ) = bn
where aπ(n),k
bn := , n ∈ N0
aπ(n),m
is a decreasing rearrangement, π : N0 → Ik bijective (if Ik is infinite, the
other case is trivial).
From the previous considerations, Corollary 4.2 and Theorem 4.8 we derive
5.1 Theorem: λ(A) is a Schwartz space if, and only if, for every k there is
m > k, such that
aj,k
lim = 0.
j∈Ik aj,m
5.2 Theorem: λ(A) is nuclear if, and only if, for every k there is m > k,
such that ∑ aj,k
< +∞.
aj,m
j∈Ik
( )
aj,k
Definition: λ(A) is called regular if aj,k+1 is decreasing for all k.
j
45
In this case either λ(A) = ω or there is k0 such that aj,k0 > 0 for all j.
Hence for a regular Köthe space we will assume that aj,0 > 0 for all j.
5.3 Lemma: If λ(A) is regular, then it is either a Schwartz space or a
Banach space.
Proof: If λ(A) is not a Schwartz space then there exists k such that for every
m > k we have inf j aj,k /aj,m =: Cm > 0. This implies aj,m ≤ Cm −1 a
j,k for
all j and m. Therefore we obtain λ(A) = λk (A). 2
If λ(A) is regular then for m > k

an,k
δn (Um , Uk ) = γn (Um , Uk ) = sn (ıkm ) = .
an,m
If λ is a sequence space then we set
M (λ) := {t = (t0 , t1 , . . .) | (tj xj )j ∈ λ for all x ∈ λ}.
M (λ) is called the space of multipliers on λ. Clearly M (λ) is an algebra

with 1 with respect to pointwise addition and multiplication.
5.4 Theorem: If λ(A) is a regular Schwartz space then ∆(λ(A)) = M (λ(A)).
Proof: If t = (t0 , t1 , . . .) ∈ ∆(λ(A)), then for every k there is m such that

aj,k
tj → 0.
aj,m
In particular this sequence is bounded and for suitable C > 0 we have
|tj |aj,k ≤ Caj,m .
For x ∈ λ(A) this gives

∞
∑ ∞
∑
|tj xj |2 a2j,k ≤C 2
|xj |2 a2j,m < ∞.
j=0 j=0
Therefore (tj xj )j ∈ λ(A) and t ∈ M (λ(A)).

To prove the reverse implication we assume that t ∈ M (λ(A)). We define a
linear map Mt : λ(A) −→ λ(A) by
Mt : x 7→ (tj xj )j .
46
The graph of Mt is closed because, assuming x(n) → 0 and Mt x(n) → y, we
(n) (n)
obtain xj → 0 for all j, tj xj → yj for all j, which implies yj = 0 for all j.
Due to the Closed Graph Theorem 1.8 Mt is continuous. Hence for every k
there is µ ∈ N and C > 0 such that ∥M x∥k ≤ C∥x∥µ .
For x = ej this gives |ej |aj,k ≤ Caj,µ for all j. Since λ(A) is a Schwartz
aj,µ
space we can find m > µ, such that aj,m → 0. Therefore for k we find m,
such that
aj,k aj,k aj,µ
|tj |δj (Um , Uk ) = |tj | = |tj | −→ 0.
aj,m aj,µ aj,m
This means that t ∈ ∆(λ(A)) which completes the proof. 2
5.5 Corollary: If the regular Schwartz space λ(A) is an algebra with 1

with respect to the coordinatewise multiplication then ∆(λ(A)) = λ(A).
Proof: Since λ(A) is an algebra we have λ(A) ⊂ M (λ(A)) = ∆(λ(A)).

Since 1 ∈ λ(A) we get for t ∈ ∆(λ(A)) = M (λ(A)) that t = (tj 1)j ∈ λ(A).
Therefore ∆(λ(A)) ⊂ λ(A). 2
It is easily verified that for y ∈ λ′ (A)

∞
∑
∥y∥∗k = |yj |2 a−2
2
j,k
j=0
where a−2
j,k = +∞ for aj,k = 0, and yj = y(ej ). With this identification
λ∗k (A) = {y = (y0 , y1 , . . . ) | ∥y∥∗k < +∞}
and the imbedding λ∗k+1 (A) ,→ λ∗k (A) corresponds to the identical imbedding
of sequences. So
λ′ (A) = {y = (y0 , y1 , . . .) | ∃ k such that ∥y∥∗k < +∞}.
In particular λ′ (A) is identified with a sequence space. For x = (x0 , x1 , . . .) ∈

λ(A) and y = (y0 , y1 , . . .) ∈ λ′ (A) we have
∞
∑
y(x) = yj xj .
j=0
From this we derive immediately the
47
Remark: M (λ(A)) = M (λ′ (A)).
and then the
5.6 Theorem: If λ(A) is a regular space then ∆(λ(A)) = M (λ′ (A)).
And exactly as previously we obtain:
5.7 Corollary: If λ(A) is a regular Schwartz space and λ′ (A) is an algebra

with 1 with respect to the coordinatewise addition and multiplication of
sequences then ∆(λ(A)) = λ′ (A).
Example: We consider s = λ(A) with aj,k = (j + 1)k . Clearly for 1 =

(1, 1, 1, . . .) we have ∥1∥∗1 < +∞ hence 1 ∈ s′ . Moreover the Cauchy-
Schwartz inequalitity implies for y, z ∈ λ(A)′ and yz = (yj zj )j
∥yz∥∗k+m ≤ ∥y∥∗k ∥z∥∗m .

2
Therefore s′ is an algebra with 1 with respect to coordinatewise multiplica-

tion and we obtain
∆(s) = s′ .
This kind of behaviour extends to a much wider class of Fréchet spaces.

Let α : 0 ≤ α0 ≤ α1 ≤ . . . be a numerical sequence, limk αk = +∞. We call
it exponent sequence. Let r ∈ R ∪ {∞}. We set
∞
∑
Λr (α) = {x = (x0 , x1 , . . .) | |x|2t = |xj |2 e2tαj < +∞ for all t < r}.
j=0
Definition: Λr (α) is called power series space of infinite type if r = +∞,

of finite type if r < +∞.
Equipped with the fundamental system of seminorms (| |t )t<r the space

Λr (a) is a Fréchet space, since for every sequence tk ↗ r the norms ∥ ∥k :=
| |tk are a fundamental system of seminorms. In fact, Λr (α) = λ(A) with
aj,k = etk αj .
Let r < +∞. The map x 7→ (xj erαj )j defines a linear topological isomor-
phism from Λr (α) onto Λ0 (α). So we have
48
Remark: Let α be fixed. Then all spaces Λr (α) , r < +∞ are isomorphic.
Therefore we can, without restriction of generality, confine ourselves to r ∈

{0, ∞}.
Example: s = Λ∞ (log(j + 1)).
5.8 Theorem: (1) Λr (α) is a regular Fréchet-Schwartz space.

log n
(2) Λ∞ (α) is nuclear if, and only if, lim supn αn < +∞.
log n
(3) Λ0 (α) is nuclear if, and only if, limn αn = 0.
Proof: (1) is obvious.

(2) follows from the following equivalences, the first of which is a consequence
of Theorem 5.2
∑
Λ∞ (α) is nuclear ⇐⇒ ∃ k : e−kαn < +∞
n
C
⇐⇒ ∃ k, C > 0 ∀ n ≥ 1 : e−kαn ≤
n
⇐⇒ ∃ k, c ∈ R ∀ n ≥ 1 : log n ≤ c + kαn
⇐⇒ ∃ k, n0 ∀ n ≥ n0 : log n ≤ kαn .
(3) In a similar way we have

∑
Λ0 (α) nuclear ⇐⇒ ∀ ε > 0 : e−εan < +∞
n
⇐⇒ ∀ ε > 0 ∃ nε ∀ n ≥ nε : log n ≤ εαn . 2
The calculation of ∆(s) = s′ generalizes now to:
5.9 Theorem: (1) If Λ∞ (α) is nuclear, then Λ′∞ (α) is an algebra with
1 with respect to coordinatewise multiplication, hence ∆(Λ∞ (α)) =
Λ′∞ (α).
(2) If Λ0 (α) is nuclear, then Λ0 (α) is an algebra with 1 with respect to

coordinatewise multiplication, hence ∆(Λ0 (α)) = Λ0 (α).
49
Proof: (1) The nuclearity condition in Theorem 5.8, (2) guarantees that 1 =
(1, 1, . . .) ∈ Λ′∞ (α). This space is an algebra with respect to coordinatewise
multiplication, since for y, z ∈ Λ∞ (α)′ and yz := (yj zj )j we have, on account
of the Cauchy-Schwarz inequality,
|yz|∗t+s ≤ |y|∗t |z|∗s .
(2) Again the nuclearity condition in Theorem 5.8, (3) guarantees that
1 = (1, 1, . . .) ∈ Λ0 (α). An analogous application of the Cauchy-Schwarz
inequalitygives for x, ξ ∈ λ0 (α)
|xξ|2t ≤ |x|t/2 |ξ|t/2 . 2
An immediate consequence is, that the nuclear power series spaces Λr (α)
and Λρ (β), r, ρ ∈ {0, +∞} are isomorphic if, and only if, they are equal.
This, however, is true also without the assumption of nuclearity. We will
show this now.
5.10 Lemma: If λ(A) is a regular Schwartz space then
aj,k
∆(λ(A)) = {t = (t0 , t1 , . . .) | ∀ k ∃ m : sup |tj | < +∞}.
j aj,m
This is shown like in the proof of Theorem 5.4.

5.11 Corollary: We have
(1) ∆(Λ∞ (α)) = {ξ = (ξ0 , ξ1 , . . .) | supj |ξj |e−kαj < +∞ for some k},
(2) ∆(Λ0 (α)) = {ξ = (ξ0 , ξ1 , . . .) | supj |ξj |etαj < +∞ for all t < 0}.
5.12 Theorem: If r, ρ ∈ {0, ∞} and α, β are exponent sequences the fol-

lowing are equivalent:
(1) Λr (α) ∼
= Λρ (β)
(2) ∆(Λr (α)) = ∆(Λρ (β))
(3) r = ρ and there exist C > 0 and j0 so that

1
αj ≤ βj ≤ Cαj
C
for all j ≥ j0 .
50
(4) Λr (α) = Λρ (β) as sets.
(5) Λr (α) = Λρ (β) as topological vector spaces.
Proof: (1)⇒(2) follows from Theorem 4.9.

(3)⇒(1) is obvious.
(2)⇒(3). Assume ∆(Λ∞ (α)) = C. Then (eαj )j ∈ ∆(Λ0 (β)), hence for every
ε > 0 we have
sup eαj −εβj = Cε < +∞
j
and therefore with cε = log Cε
αj ≤ εβj + cε .
So we obtain
αj
lim sup ≤ε
j βj
for all ε > 0, that is
αj
lim = 0.
j βj
On the other hand for any nonnegative sequence tj → 0 we have ξ :=
(etj βj )j ∈ ∆(Λ0 (β)) = ∆(Λ∞ (α)). Therefore there id k such that
sup etj βj −kαj =: C < ∞,

j
hence
tj βj ≤ kαj + c
with c = log C. So we obtain
βj
lim sup tj <∞
j αj
for all nonnegative sequences tj → 0. This implies
βj
lim sup <∞
j αj
which yields a contradiction.

Therefore we have proved r = ρ.
51
Assume first that r = ρ = +∞ and ∆(Λ∞ (α)) = ∆(Λ∞ (β)). Then (eαj )j ∈
∆(Λ∞ (α)) = ∆(Λ∞ (β)). Therefore we have k so that
sup eαj −kβj = C < +∞
j
hence with c = log C

αj ≤ kβj + c
for all j, hence
αj ≤ (k + 1)βj
for all j ≥ j1 .
Analogously we get l and j2 , so that
βj ≤ (l + 1)αj
for all j ≥ j2 , which completes the proof for r = ρ = +∞.
For r = ρ = 0 we have ∆(Λ0 (α)) = ∆(Λ0 (β)), both being Fréchet spaces.
Since the graph of the identical map is closed, as both topologies imply
coordinatewise convergence, their topologies coincide. So we can find C > 0
and ε > 0, such that
sup |xj |e−αj ≤ C sup |xj |e−εβj
j j
for all x ∈ ∆(λ0 (α)) = ∆(λ0 (β)).

This implies with c = log C
−αj ≤ −εβj + c
and therefore
1 C
βj ≤ αj + .
ε ε
Analogously we get
1 C′
αj ≤ βj +
ε′ ε′
with suitable ε′ > 0, C ′ , and the proof is completed as in the infinite type
case.
(3)=⇒(5) is immediate, (5)=⇒(1) and (5)=⇒(4) trivial.
(4)=⇒(5) follows from the Closed Graph Theorem like previously, since the
graph of the identical map is closed, as both topologies imply coordinatewise
convergence. 2
52
Remark: For r, ρ ∈ R ∪ {+∞} assertion (1), (2), (3), of Theorem 5.12 are
still equivalent. This does, of course, not hold for (4).
A consequence of Theorem 5.12 is:
5.13 Theorem: Let α, r, be given. The following are equivalent:
(1) Λr (α) ∼
= K × Λr (α)
(2) ∆(Λr (α)) = ∆(K × Λr (α))
an+1
(3) lim supn an < +∞
Proof: We apply Theorem 5.12 and the Remark after it to Λr (α) and Λr (β)
with β = (0, α0 , α1 , α2 , . . .). 2
A little bit more of a proof is required by the following:
5.14 Theorem: Let α, r be given. The following are equivalent:
(1) Λr (α) × Λr (α) ∼

= Λr (α)
(2) ∆(Λr (α) × Λr (α)) = ∆(Λr (α))
( )
(3) ∆ Λr (α)N = ∆(Λr (α))
α2n
(4) lim supn αn < +∞.
Proof: We apply Theorem 5.12 and the Remark after it to Λr (α) and Λr (β)
with β = (α0 , α0 , α1 , α1 , α2 , α2 , . . .). Then Λr (β) ∼
= Λr (α) × Λr (α) and we
obtain the equivalence of (1), (2), (4).
(3)=⇒(2) follows from Theorem 4.10 by
( )
∆(Λr (α)) ⊃ ∆(Λr (α) × Λr (α)) ⊃ ∆ Λr (α)N
since Λr (α) is a quotient of Λr (α) × Λr (α) and Λr (α) × Λr (α) a quotient of

Λr (α)N .
( )
So it remains to show that (4) implies (3). Since again ∆(Λr (α) ⊃ ∆ Λr (α)N
follows from Theorem 4.10 we have to show that (4) implies
( )
∆(Λr (α)) ⊂ ∆ Λr (α)N .
53
Let t ∈ ∆(Λr (α)) then we have
lim tn eτ αn = 0
n
for some τ < 0 if r = +∞, or for every τ < 0 if r = 0, respectively.

We want to show that t ∈ ∆(Λr (α)N ), so let U ⊂ Λr (α)N be a neighborhood
of zero. We may assume that for some m and s < r
∑ ∞
m ∑
U = { x = (x1 , x2 , . . .) | |x|U = |xk,j |2 e2sαj ≤ 1 }.
k=0 j=0
Here xk ∈ Λr (α) for every k and xk = (xk,0 , xk,1 , . . .).

We choose
{ ∞
m ∑
∑ }
V = x = (x1 , x2 , . . .) | |x|V = |xk,j |2 e2σαj ≤ 1
k=1 j=0
where s < σ < r.

It can easily be seen that the sequence of Kolmogoroff diameters δn (V, U )
V ) where ıV : EV −→ EU is the
or, equivalently, of singular numbers sn (ıU U
canonical map, has the form
(e(s−σ)α0 , . . . , e(s−σ)α0 , e(s−σ)α1 , . . . , e(s−σ)α1 , . . .),
where the term e(s−σ)αn is repeated m times. Therefore for ν ∈ N0 , 0 ≤ µ <

m:
δνm+µ (V, U ) = e(s−σ)αν .
By assumption we have
α2n
lim sup =: d < +∞
n αn
hence
α2m+1 ν
lim sup ≤ dm+1 .
ν αν
Therefore we have for large ν and 0 ≤ µ < m
αmν+µ ≤ α(m+1)ν ≤ α2m+1 ν ≤ dm+1 αν ,
54
and therefore with suitable c > 0 for all ν
αmν+µ ≤ dm+1 (αν + c) .
This implies
−m−1 α
δνm+µ (V, U ) ≤ e(σ−s)c · e(s−σ)d νm+µ
.
Therefore
−m−1 α
δn (V, U ) ≤ e(σ−s)c · e(s−σ)d n
for all n ∈ N0 .
We obtain
−m−1 α
|tn |δn (V, U ) ≤ e(σ−s)c · |tn |e(s−σ)d n
.
For r = +∞ we choose σ so large that (s − σ)d−m−1 ≤ τ , and for r < +∞
we may choose any σ with s < σ < r. In both cases we have
lim tn δn (V, U ) = 0. 2
n
Definition: (1) An exponent sequence α = (α0 , α1 , . . .) is called shift-

stable if
αn+1
lim sup < +∞.
n αn
(2) It is called stable if

α2n
lim sup < +∞.
n αn
The names are self-explaining in view of Theorem 5.13 and 5.14. Using the
description of Theorem 4.7 we generalize now the concept of nuclearity in a
twofold way.
Definition: A Fréchet-Hilbert space E is called
(1) (α, ∞)-nuclear if the following holds:
∀U, k ∃V : ekαn δn (V, U ) −→ 0
(2) (α, 0)-nuclear if the following holds:
∀U ∃V, ε > 0 : eεαn δn (V, U ) −→ 0.
55
Here V and U denote absolutely convex neighborhoods of zero. As E is
assumed to be a Fréchet-Hilbert space the definition does not depend on
whether we use Kolmogoroff or Gelfand diameters, or singular numbers of
connecting maps.
Remark: If α is so that Λ∞ (α), resp. Λ0 (α), is nuclear, then, due to

Theorems 5.8 and 4.7, conditions (1), resp. (2), in the definition imply that
E is nuclear. In particular it is automatically a Fréchet-Hilbert space.
Example: By Theorem 4.7 E is nuclear if, and only if, it is (α, ∞)-nuclear
for αn = log(n + 1).
5.15 Theorem: A Fréchet-Hilbert space E is (α, r)-nuclear if, and only if,
∆(E) ⊃ ∆(Λr (α)).
Proof: This is immediate from Corollary 5.11. 2
By means of Theorems 4.10 and 4.14 this implies:
5.16 Theorem: If the Fréchet-Hilbert space E is (α, r)-nuclear and F ⊂ E

a closed subspace then F and E/F are (α, r)-nuclear.
Since, due to Corollary 5.11, ∆(Λ∞ (α)) % ∆(Λ0 (α)) we obtain
5.17 Proposition: (1) Λ∞ (α) is (α, r)-nuclear, r = 0, +∞.
(2) Λ0 (α) is (α, 0)-nuclear, however not (α, ∞)-nuclear.
(3) Every (α, ∞)-nuclear space is also (α, 0)-nuclear.
More generally:
5.18 Proposition: For any α and β we have
αn
(1) Λ∞ (α) is (β, r)-nuclear if, and only if, lim supn βn < +∞.
αn
(2) Λ0 (α) is (β, 0)-nuclear if, and only if, lim supn βn < +∞.
(3) Λ0 (α) is (β, ∞)-nuclear if, and only if, limn αn

βn = 0.
56
Proof: Sufficiency of the condition is easily checked. For necessity we have,
in case of (β, ∞)-nuclearity, to draw the consequences of (e−βn )n ∈ ∆(Λr (α))
and in case of (β, 0)-nuclearity of the continuity of the imbedding ∆(Λ0 (α))
,→ ∆(Λr (α)), the first one being a Fréchet space (cf. the proof of Theorem
5.12). 2
5.19 Lemma: Λr (α)N is (α, r)-nuclear if, and only if, α is stable.
( )
Proof: As ∆ Λr (α)N ⊂ ∆(Λr (α))( (see 4.10), ) Theorem 5.15 implies that
(α, r)-nuclearity is equivalent to ∆ Λr (α)N = ∆(Λr (α)). Due to Theorem
5.14 this is equivalent to the stability of α. 2
So we have:
5.20 Corollary: If α is stable then every subspace of Λr (α)N is (α, r)-

nuclear.
This is one direction of a generalization of Theorem 3.18. The full general-

ization can be given under the additional assumption of nuclearity.
5.21 Theorem: If α is stable and Λr (α) nuclear then the following are
equivalent
(1) E is (α, r)-nuclear,
(2) E is isomorphic to a subspace of Λr (α)N .
Proof: (2)⇒(1) follows from the previous Corollary and the fact that the
diametral dimension and therefore (α, r)-nuclearity is a linear topological
invariant.
(1)⇒(2) is shown for r = 0, +∞ separately. First we assume r = +∞. We
proceed as in the proof of T. and Y. Komura’s Theorem 3.18.
We fix K and denote by
∗ ∗
jk : EK ,→ EK+k
the canonical imbedding. Let (A(k))k∈N be a sequence which will be deter-
mined later. We may assume
sn (jk )eA(k)αn −→ 0
57
for every k. The Schmidt representation is
∞
∑
jk (x) = sn (jk )⟨x, ekn ⟩K fnk . (7)
n=0
We put sn,k := sn (jk ) and set for k ∈ N and n ∈ N0
l(n, k) := 2k−1 (1 + 2n) − 1.
We apply the Gram-Schmidt orthogonalization method with respect to ⟨·, ·⟩K

to the vectors ekn , l(k, n) = 0, 1, 2, . . . and obtain an orthonormal sequence
e0 , e1 , . . ..
By construction we have em ⊥ekn for m > l(n, k) or, equivalently, m ≥
2k−1 (1 + 2n) which is equivalent to n ≤ 12 (2−k+1 m − 1). We set
[ ]
1 −k+1
n(m) = (2 m − 1)
2
and obtain by use of (7)
∑
∥em ∥∗K+k =
2
s2n,k |⟨em , ekn ⟩K |2 ≤ s2n(m)+1
n>n(m)
≤ Ck2 e−2A(k)αn(m)+1
with suitable Ck . We have

[ ]
1 −k+1
n(m) + 1 = (2 m + 1) ≥ [2−k m].
2
Putting s := supn≥n0 α2n
αn , where αn > 0 for n ≥ n0 , we have for large m
αm ≤ α2k+1 [2−k m] ≤ sk+1 α[2−k m] ≤ sk+1 αn(m)+1 .
Therefore we obtain
αm ≤ sk+1 αn(m)+1 + dk
for all m. We choose A(k) = ksk+1 and have with possibly increased Ck
∥em ∥∗K+k ≤ Ck e−2kαm .
For x ∈ E this gives

∑ ∑
|em (x)|2 = |⟨x̃, em ⟩K |2 = ∥x̃∥2K = ∥x∥2K .
m m
58
Here x̃ ∈ EK ∗ is chosen by the Riesz representation theorem to fulfill y(x) =
⟨y, x̃⟩K for y ∈ EK∗ .
Moreover
∑ ∑
|em (x)|2 e2kαm ≤ ∥x∥2K+L ∗2
m ∥em ∥K+L e
2kαm
m
∑
≤ ∥x∥2K+L CL me
2(k−L)αm
where the series in the last term converges for L large enough.
Therefore φK : x 7→ (em (x))m∈N0 defines a continuous linear map
φK : E −→ Λ∞ (α)
such that
|φK x|0 = ∥x∥K .
Consequently φ : x 7→ (φK (x))K∈N defines a continuous linear map,
φ : E −→ Λ∞ (α)N
such that
max |φK (x)|0 = ∥x∥L .
K=1,...,L
Hence φ is a topological linear imbedding.

Next we show (1)=⇒(2) for r = 0. We may assume again that all ∥ ∥k are
Hilbertian seminorms, that ıkk+1 is compact and
sn (ıkk+1 )eεk αn −→ 0
for all k with suitable εk > 0.

The Schmidt representation gives with sn = sn (ıkk+1 )
∞
∑
ıkk+1 (x) = sn ⟨x, en ⟩k+1 fk .
n=0
Here (en )n∈N0 and (fn )n∈N0 are orthonormal systems in the local Banach
spaces Ek+1 and Ek , respectively. The first one needs not to be an or-
thonormal basis, the second one is an orthonormal basis.
We put for x ∈ E, omitting the maps ık and ık+1 ,
φk (x) = (⟨x, fn ⟩k )k∈N0
59
and obtain, using ⟨x, fn ⟩k = sn ⟨x, en ⟩k+1 ,
∑ ∑
e2εk αn |⟨x, fn ⟩k |2 = e2εk αn s2n |⟨x, en ⟩k+1 |2
n n
∑
≤ Ck2 |⟨x, en ⟩k+1 |2
n
≤ Ck ∥x∥2k+1 ,
2
with suitable Ck > 0. Therefore φk defines a continuous (even compact)

linear map E −→ Λεk (α), such that
( )1
∑ 2
|φk (x)|0 = |⟨x, fn ⟩|2k = ∥x∥k . (8)

n
By setting φ(x) := (φk (x))k∈N we define a continuous linear map

∏
φ : E −→ Λεk (x) ∼
= Λ0 (α)N ,
k
which is, due to (8), an isomorphic imbedding. 2
6 Tensor products and nuclearity
We recall the definition of the algebraic tensor product. Let E and F be

linear spaces. A linear space E ⊗ F together with a bilinear map B0 : E ×
F −→ E ⊗ F is called tensor product of E and F if the following holds: for
every bilinear map B : E × F −→ G, G linear space, there is exactly one
linear map Be : E ⊗ F −→ G with B e ◦ B0 = B.
e
E ⊗O F
B /G
xx;
x
xx
B0
xxx B
x
E×F
If (E ⊗ F )1 with B1 : E × F −→ (E ⊗ F )1 is a second tensor product then
we have
e1
B
o / (E ⊗ F )
E⊗F
_? ? 1
??
?? B0
e

??
B0 ?? B1

E × F.
60
Since Be0 ◦ B
e1 ◦ B0 = Be0 ◦ B1 = B0 = id ◦B0 we have B e0 ◦ B
e1 = id and
e e
likewise B1 ◦ B0 = id. So E ⊗ F and (E ⊗ F )1 are isomorphic, with an
isomorphism compatible with B0 and B1 . In this sense the tensor product
is uniquely determined up to isomorphism.
We set x ⊗ y := B0 (x, y). Since B e is uniquely determined for any G and
B we have clearly span{x ⊗ y | x ∈ E, y ∈ F } = E ⊗ F . So taking into
account the bilinearity of (x, y)∑7→ x ⊗ y we see that every u ∈ E ⊗ F can
be written as a finite sum u = j xj ⊗ yj .
This yields also a proof of the existence of a tensor product. We set
H = {(ξx,y )x∈E, y∈F ∈ K | only finitely many ξx,y ̸= 0}
H0 = span{ξλx1 +µx2 ,y − λξx1 ,y − µξx2 ,y , ξx,λy1 +µy2 − λξx,y1 − µξx,y2

| x ∈ E, y ∈ F, λ, µ ∈ K}
and set E ⊗ F := H/H0 and B0 (x, y) = êx,y , where ex,y is the natural unit
basis vector in H and êx,y its residue class.
We turn now to locally convex spaces.
Definition: The π-tensor product of two locally convex spaces E and F

is their tensor product equipped with the uniquely defined locally convex
topology on E ⊗ F such that B0 is continuous and for any continuous linear
map B : E × F −→ G, G locally convex, the map B e is continuous. It is
denoted by E ⊗π F .
The uniqueness of the topology follows from the argument proving the
”uniqueness” of E ⊗ F . To see the existence we put for any continuous
seminorms p1 on E, p2 on F and u ∈ E ⊗ F
{∑ ∑ }
p1 ⊗π p2 (u) = inf p1 (xj )p2 (yj ) | u = xj ⊗ yj .
j j
It is easy to see that p1 ⊗π p2 is a seminorm on E ⊗ F . We further note

that for locally convex spaces E, F, G a bilinear map B : E × F −→ G is
continuous if and only if for any seminorm q on G there are seminorms p1
on E and p2 on F such that
q(B(x, y)) ≤ p1 (x)p2 (y)
61
for all x ∈ E, y ∈ F .
∑
Hence for any representation u = j xj ⊗ yj we have
∑
e
q(B(u)) ≤ e j ⊗ yj ))
q(B(x
j
∑
= q(B(xj , yj ))
j
∑
≤ p1 (xj )q2 (yj )
j
e
and therefore q(B(u)) ≤ p1 (x) ⊗π p2 (u).
Obviously p1 ⊗π p2 (x ⊗ y) ≤ p1 (x)p2 (y) hence B0 is continuous. This proves
the existence of the π-tensor product topology.
Moreover we have:
6.1 Lemma: For x ∈ E, y ∈ F we have p1 ⊗π p2 (x ⊗ y) = p1 (x)p2 (y).
Proof: p1 ⊗π p2 (x⊗y) ≤ p1 (x)p2 (y) is clear. To prove the opposite inequality

we choose linear forms φ ∈ E ′ , ψ ∈ F ′ such that |φ(ξ)| ≤ p1 (ξ) for all ξ ∈ E,
|ψ(η)| ≤ p2 (η) for all η ∈ F and φ(x) = p1 (x), ψ(y) = p2 (y). For the bilinear
form B(ξ, η) = φ(ξ)ψ(η) we have |B(ξ, η)| ≤ p1 (ξ)p2 (η) hence, due to the
previous calculation |B(u)| ≤ p1 ⊗π p2 (u) for all u ∈ E ⊗π F . This gives
p1 (x)p2 (y) = |B(x ⊗ y)| ≤ p1 ⊗π p2 (x ⊗ y). 2
From the previous considerations we obtain:
6.2 Theorem: If P1 (resp. P2 ) is a fundamental system of seminorms on

E (resp. F ) then {p1 ⊗π p2 | p1 ∈ P1 , p2 ∈ P2 } is a fundamental system
of seminorms on E ⊗π F .
As an immediate consequence of this theorem we have the
6.3 Corollary: If E and F are normed then E ⊗π F is normed, if E and

F are metrizable locally convex, then also E ⊗π F .
Definition: The completion E ⊗

ˆ π F is called the complete π-tensor product
of E and F .
62
Remark: If E and F are complete then E ⊗ ˆ π F fulfills the defining proper-
ties of a tensor product in the class of complete locally convex spaces.
We denote by p1 ⊗
ˆ π p2 the continuous extension of p1 ⊗π p2 onto E ⊗
ˆ π F . Then
clearly the seminorms p1 ⊗p2 , p1 ∈ P1 , p2 ∈ P2 , where P1 , P2 are funda-
ˆ
mental systems of seminorms on E and F , respectively, are a fundamental
system of seminorms for E ⊗ ˆ π F . This proves:
6.4 Theorem: If E and F are Banach spaces then E ⊗ ˆ π F is a Banach

space, if E and F are Fréchet spaces then E ⊗
ˆ π F is a Fréchet space.
If E and F are Fréchet spaces with fundamental system of seminorms || ||1 ≤

|| ||2 ≤ . . . then we set || ||k := || ||k ⊗
ˆ π || ||k on the Fréchet space E ⊗
ˆ π F , and
|| ||1 ≤ || ||2 ≤ . . . is a fundamental systems of seminorms there.
∑∞ Then every u ∈ E ⊗π F has

6.5 Theorem: Let∑E and F be Fréchet spaces. ˆ
∞
an expansion u = ∑ x
j=1 j ⊗ y j such that j=1 ||x ||
j k ||y ||
j k < 1 for all k.
∞
We have ||u||k = inf{ j=1 ||xj ||k ||yj ||k } where the infimum runs through all
such representations.
Proof:
∑∞ We fix ε > 0. Then it is easy to see that u can be written as
u = k=1 uk where uk ∈ E ⊗ F and ||uk ||k < 2−k−1 ε for k = 2, 3, . . ..
We choose for every k a representation
∑
mk
uk = xkj ⊗ yjk
j=1
such that
∑
mk
||xkj ||k |yjk ||k ≤ 2−k−1 ε + ||uk ||k .
j=1
Then we have
∞ ∑
∑ mk
u= xkj ⊗ yjk
k=1 j=1
and for any m ∈ N

∞ ∑
∑ mk ∑
m ∑
mk ∞
∑
||xkj ||m ||yjk ||m ≤ ||xkj ||m ||yjk ||m + 2−k ε < +∞.
k=1 j=1 k=1 j=1 k=m+1
63
For m = 1 we have
∞ ∑
∑ mk
3
||xkj ||1 ||yjk ||1 ≤ ||u1 ||1 + ε
4
k=1 j=1
and by choice of the uk

1
||u1 ||1 ≤ ||u||1 + ε.
2
This proves
∞
∑
||u||1 ≥ inf{ ||xj ||1 ||yj ||1 | . . .}.
j=1
The reverse inequality follows from the triangular inequality and Lemma 6.1.
Replacing 1 with k in the whole argument proves the general assertion. 2
Remark: In the previous proof we may assume only that E and F are
locally convex and metrizable and we get the same representation for the
elements of E ⊗
ˆ π F = Ê ⊗
ˆ π F̂ which implies that in the case of Fréchet spaces
the representations can be chosen from any given dense subspaces.
We study now tensor products of continuous linear maps. For locally convex
spaces F1 and F2 and A1 ∈ L(E1 , F1 ), A2 ∈ L(E2 , F2 ) we consider the
diagram
A1 ⊗A2
E1 ⊗O π E2 / F1 ⊗π F2
t: O
t
t
t
B tbilinear B
t
t
t A1 ×A2
E1 × E2 / F1 × F2
(x1 , x2 ) / (A1 x1 , A2 x2 )
which defines the map A1 ⊗A2 . Clearly A1 ⊗A2 ∈ L(E1 ⊗E2 , F1 ⊗F2 ) and it
extends to a map A1 ⊗A
ˆ 2 ∈ L(E1 ⊗
ˆ π E 2 , F1 ⊗
ˆ π F2 ). We have A1 ⊗A2 (x⊗y) =
A1 x ⊗ A2 y.
If E1 , E2 are locally convex and A1 : E1 −→ F1 , A2 : E2 −→ F2 are surjective
and open then clearly
A1 ⊗ A2 : E1 ⊗π E2 −→ F1 ⊗π F2
is surjective and open. In general it does not follow that the extension to
the completion is again surjective. However, in the case of Fréchet spaces
the situation is nice:
64
6.6 Theorem: If A1 and A2 are surjective, F1 and F2 Fréchet spaces, then
A1 ⊗A
ˆ 2 : E1 ⊗
ˆ π E2 −→ F1 ⊗
ˆ π F2 is surjective.
Proof: Let || ||1 ≤ || ||2 ≤ . . . be fundamental systems of seminorms in the

respective spaces, moreover || ||k in Fj the quotient seminorm of || ||k in Ej
under Aj . Let
∞
∑ ∞
∑
u= x j ⊗ yj , ||xj ||k ||yj ||k < +∞ for all k
j=1 j=1
be a representation of u ∈ F1 ⊗ˆ π F2 . By assumption we may find ξj ∈ E1 ,

ηj ∈ E2 such that A1 ξj = xj , A2 ηj = yj and
||ξj ||n ||ηj ||n ≤ ||xj ||n ||yj ||n + 2−j
for all j.
∑∞
Therefore v := j=1 ξj ⊗ ηj ∈ E 1 ⊗
ˆ π E2 and A ⊗ B(v) = u. 2
For the kernels we can not do much more at present, than to describe the
algebraic situation.
6.7 Lemma: ker(A1 ⊗ A2 ) = ker A1 ⊗ E2 + E1 ⊗ ker A2 for all linear maps

A1 : E1 −→ F1 and A2 : E2 −→ F2 .
∑
Proof: Assume u = m j=1 xj ⊗ yj ∈ E1 ⊗ E2 and A1 ⊗ A2 u = 0. We may
assume additionally that the vectors y1 , . . . , ym are linearly independent and
that y1 , . . . , yk span the vector space ker A2 ∩ span{y1 , . . . , ym }.
Then dim span{A2 yk+1 , . . . , A2 ym } = m − k. Therefore
{(φA2 yk+1 , . . . , φA2 ym ) | φ ∈ F2′ } = Rm−k .

∑ ′
Since 0 = (id ⊗φ)(A1 ⊗ A2 )u = m j=k+1 φA2 yj Axj for all φ ∈ F2 , we have
A1 xj = 0 for j = k + 1, . . . , m, that is xk+1 , . . . , xm ∈ ker A1 . 2
Notice that there is no need that ker(A⊗A

ˆ 2 ) = ker A⊗E
ˆ 2 + E1 ⊗
ˆ ker A2 .
Let E and F be linear spaces, G a linear space of linear forms on E which

separates points and let L (G, F ) denote the linear maps from E to F .
Then the bilinear map B : E × F −→ L (G, F ) given by B(e, f )(g) = g(e)f
e : E ⊗ F −→ L (G, F ).
extends to an injective linear map B
65
Here injectivity follows
∑ easily by assuming, as in the previous proof, that
for a given u = m j=1 x j ⊗ yj ∈ E ⊗ F the vectors y1 , . . . , ym are linearly
independent.
We study this map now in the case of Banach spaces. Let X, Y be Banach
spaces, then the continuous bilinear map B : X ′ × Y −→ L(X, Y ) given by
B(x′ , y)(x) = x′ (x)y extends to a continuous linear map B e : X ′ ⊗π Y −→
L(X, Y ). Likewise the continuous bilinear map B : X × Y −→ L(X ′ , Y )
given by B(x, y)(x′ ) = x′ (x)y extends to a continuous linear map B e : X ⊗π
′
Y −→ L(X , Y ). Both are injective. B e again extends to a continuous linear
′
map BL(X,Y ) : X ⊗π Y −→ L(X, Y ) or BL(X ′ ,Y ) : X ⊗
ˆ ˆ π Y −→ L(X ′ , Y ),
respectively. Notice that neither of them needs to be injective.
As an immediate consequence of Theorem 6.5 we obtain.
6.8 Corollary: R(BL(X,Y ) ) = N (X, Y ) topologically, i.e. the quotient
norm under BL(X,Y ) is the nuclear norm, in particular N (X, Y ) is a Banach
space.
If N (BL(X,Y ) ) = {0} then BL(X,Y ) constitutes an isomorphism between

X ′⊗
ˆ π Y and N (X, Y ). We will now study, when this is the case.
Definition: A Banach space X has the approximation property if for every

compact K ⊂ X and ε > 0 there is a finite dimensional map φ ∈ L(X) such
that supx∈K ||x − φ(x)|| < ε.
Example: (1) Every ℓp , 0 ≤ p < +∞ has the approximation property.
(2) Every Hilbert space has the approximation property.
Proof: (1) Set Pn (x) = (x1 , . . . , xn , 0, . . .) for x = (xk )k∈N then ||x −
Pn (x)|| is decreasing and converges to zero for every x, so by Dini’s
theorem it converges uniformly for every compact set K ⊂ X.
(2) We choose an orthonormal

∑ basis (ei )i∈I and for every finite set e ⊂ I
we put Pe x = i∈e ⟨x, ei ⟩ei . Then we proceed as in (1). 2
We denote, as previously, by F (X, Y ) the linear space of finite dimensional

maps in L(X, Y ). Clearly the canonical map X ′ ⊗ Y −→ L(X, Y ) sends
X ′ ⊗ Y bijectively onto F (X, Y ). We have F (X, Y ) ⊂ K(X, Y ), the space
of compact operators in L(X, Y ).
6.9 Theorem: The following are equivalent
66
(1) Y has the approximation property.
(2) F (X, Y ) = K(X, Y ) for every Banach space X.
Proof: (1) ⇒ (2) Let UX be the closed unit ball in X, T ∈ K(X, Y ). We set
K = T UX . For given ε > 0 we find φ ∈ F (Y ) such that ||y − φ(y)|| < ε for
all y ∈ K. Then ||T x − φ ◦ T (x)|| < ε for all x ∈ UX , hence ||T − φ ◦ T || ≤ ε.
Clearly φ ◦ T ∈ F (X, Y ).
(2) ⇒ (1) We choose an absolutely convex compact set L, K ⊂ L ⊂ Y ,
such that K is compact in EL . Let j : YL ,→ Y be the identical imbedding.
Since j ∈ K(YL , Y ) we find for every ε > 0 a map φ ∈ F (YL , Y ) with
supy∈L ||y − φ(y)|| < ε/2. We have
∑
m
φ(y) = ηj (y)yj
j=1
with some m ∈ N, where η1 , . . . , ηm ∈ YL′ , y1 , . . . , ym ∈ Y . We may assume

||ηj ||∗L ≤ 1 for all j.
Let M be the unit ball of YL′ , and
M0 = {η|YL | η ∈ Y ′ , |η(y)| ≤ 1 for all y ∈ L} = {η|YL | η ∈ L◦ }.
Here the polar is taken in Y ′ . From the Bipolar Theorem 2.1 for the duality
of Y and Y ′ we obtain M0◦ = L, independently of whether we think of this
polar being taken in Y or YL . Therefore we have (M0◦ )◦ = L◦ = M , where
the polars are taken in the duality of YL and YL′ . Hence, by the Bipolar
Theorem 2.1 M0 is σ(YL′ , YL )-dense in M . Since M is bounded and an
equicontinuous set of continuous functions on the || ||L -compact set K we
have, due to the Arzelà-Ascoli Theorem, that M0 is dense in M in the C(K)-
topology,∑i.e. we can find ξ1 , . . . , ξn ∈ Y ′ such that
∑sup x∈K |ξj (x)−ηj (x)| < δ
where δ m j=1 ||yj || < ε/2. Therefore ψ(y) = m
j=1 ξj (y)yj ∈ F (Y ) and
supy∈K ||y − ψ(y)|| < ε. 2
We prove now a series of lemmas which altogether will end up in Theorem

6.13 which gives a lot of equivalences for the approximation property among
them the one we are looking for, namely the injectivity of the canonical
map from the complete π-tensor products to the respective spaces of linear
operators.
6.10 Lemma: If Y has the approximation property then the canonical map
ˆ π Y −→ L(X ′ , Y ) is injective for every Banach space X.
BL(X ′ ,Y ) : X ⊗
67
∑∞ ∑∞
k=1 xk ⊗ yk ∈ X ⊗π Y 8 with k=1 ||xk || ||yk || < ∞ we
Proof: For u = ˆ
assume BL(X ′ ,Y ) (u) = 0, that is
∞
∑
ξ(xk )yk = 0, for all ξ ∈ X ′ .
k=1
∑
We may assume that k ||xk || = 1 and limk ||yk || = 0. We set K = {yk |
k ∈ N} and find φ ∈ F (Y ) such that supk ||yk − φ(yk )|| < ε.
Assume that
∑
m
φ(y) = ηj (y)zj ,
j=1
where η1 , . . . , ηm ∈ Y ′ and z1 , . . . , zm ∈ Y . For v = (id ⊗φ)u we obtain

∞
∑
v= xk ⊗ φ(yk )
k=1
∑∞ ∑
m
= xk ⊗ ηj (yk )zj
k=1 j=1
(∞ )
∑m ∑
= ηj (yk )xk ⊗ zj .
j=1 k=1
Therefore v ∈ X ⊗ Y and
(∞ ) ( )
∑
m ∑ ∑
m ∑
(BL(X ′ ,Y ) v)(ξ) = ξ ηj (yk ) xk zj = ηj yk ξ(xk ) zj = 0
j=1 k=1 j=1 k
for all ξ. As BL(X ′ ,Y ) : X ′ ⊗ Y ,→ L(X, Y ) is injective we have v = 0 and

therefore
∞
∑
∥u∥ = ∥u − v∥ ≤ ||xk || ||yk − φ(yk )|| ≤ ε.
k=1
Since this holds for every ε > 0 the proof is complete. 2
ˆ π Y −→ L(X ′ , Y ) is injective for every Ba-

6.11 Lemma: If BL(X ′ ,Y ) : X ⊗
nach space X, then also BL(X,Y ) : X ′ ⊗
ˆ π Y −→ L(X, Y ) is injective for every
Banach space X.
68
∑∞ ′ˆ
∑∞
Proof: For u = k=1 ξk ⊗ yk ∈ X ⊗π Y with k=1 ||ξk || ||yk || < ∞ we
assume BL(X,Y ) (u) = 0, i.e.
∞
∑
ξk (x)yk = 0, for all x ∈ X.
k=1
Then we have for every η ∈ Y ′ and x ∈ X

(∞ ) (∞ )
∑ ∑
η(yk )ξk (x) = η ξk (x)yk = 0.
k=1 k=1
∑∞
Therefore = 0 for every η ∈ Y ′ . Now we take ζ ∈ X ′′ , η ∈ Y ′
k=1 η(yk )ξk
and obtain (∞ ) (∞ )
∑ ∑
η ζ(ξk )yk = ζ η(yk )ξk = 0.
k=1 k=1
This proves that
∞
∑
ζ(ξk )yk = 0
k=1
for every ζ ∈ X ′′ , i.e. BL(X ′′ ,Y ) u = 0. By the assumption, applied to the
Banach space X ′ , we get u = 0. 2
If Y is a Banach space then the continuous bilinear form (η, y) 7→ η(y) on

Y ′ × Y generates a continuous linear form on Y ′ ⊗
ˆ π Y which is called tensor
trace.
If the tensor trace vanishes on N (BL(Y ) ) then, by Corollary 6.8, for a nuclear
map T ∈ N (Y ) with
∞
∑ ∞
∑
Tx = ηk (x)yk , ∥ηk ∥|yk ∥ < ∞
k=1 k=1
the number
∞
∑
ηk (yk )
k=1
does not depend on the representation of T and defines a continuous linear
form on N (Y ) which is called trace.
In this case we say that N (Y ) admits a trace.
If BL(X,Y ) is injective for every Banach space X then, in particular, N (BL(Y ) ) =
0 and therefore N (Y ) admits a trace.
69
6.12 Lemma: If N (Y ) admits a trace, then Y has the approximation prop-
erty.
Proof: Let K ⊂ Y be compact. By Lemma 3.6 we may assume that there

is a null sequence (xn )n∈N in Y such that
{∞ ∞
}
∑ ∑
K= λk xk | |λk | ≤ 1 .
k=1 k=1
We set
H = {A ∈ L(YK , Y ) | A|K is continuous with respect to the Y −topology}.
H is a closed subspace of the Banach space L(YK , Y ), L(Y )|YK ⊂ H and for
every A ∈ H we have limk Axk = 0.
Therefore φ(A) := (Axk )k∈N defines a linear map H −→ c0 (Y ) which, due
to
sup ∥Ax∥ = sup ∥Axk ∥ for all A ∈ H,
x∈K k
is an isometric imbedding. If we identify in canonical way the dual space
c0 (Y )′ with ℓ1 (Y ′ ), then we obtain that φ′ : ℓ1 (Y ′ ) −→ H ′ is surjective.
Hence for every µ ∈ H ′ there is η = (η1 , η2 , . . . ) ∈ ℓ1 (Y ′ ) such that
∞
∑
′
µA = (φ η)A = ⟨η, φA⟩ = ηk (Axk ).
k=1
∑∞
Since k=1 ∥ηk ∥ ∥xk ∥ < ∞ we may set
∞
∑
u := ηk ⊗ xk ∈ Y ′ ⊗
ˆ π Y.
k=1
Let us assume that µ vanishes on F (Y ), i.e. µ(A) = 0 for all Ax = η(x)y

with η ∈ Y ′ , y ∈ Y . Then
∞
(∞ )
∑ ∑
0 = µ(A) = ηk (y)η(xk ) = η ηk (y)xk
k=1 k=1
for all η ∈ Y ′ where we consider y ∈ Y as fixed.

Therefore we have
∞
∑
ηk (y)xk = 0
k=1
70
for all y ∈ Y , i.e. BL(Y ) (u) = 0.
By assumption, now the tensor trace must be zero, i.e.
∞
∑
0= ηk (xk ) = µ(j),
k=1
where j : YK −→ Y denotes the identical imbedding. Notice that j ∈ H.

By the Bipolar Theorem 2.1, applied to H, we have proved that j ∈ F (Y ),
where the closure is taken in H or in L(YK , Y ) which is the same. Therefore
we have for any ε > 0 a map φ ∈ F (Y ), such that
sup ∥x − φ(x)∥ < ε.

x∈K
This completes the proof. 2
Putting together the previous lemmas and Theorem 6.9 we obtain:
6.13 Theorem: For a Banach space Y the following are equivalent:
1. Y has the approximation property.
2. F (X, Y ) = K(X, Y ) for every Banach space X.

ˆ π Y −→ L(X ′ , Y ) is injective for every Banach space X.
3. BL(X ′ ,Y ) : X ⊗
4. BL(X,Y ) : X ′ ⊗
ˆ π Y −→ L(X, Y ) is injective for every Banach space X.
5. N (Y ) admits a trace.
6.1 Tensor products with function spaces
Let M be a set, L a linear space of scalar functions on M and X a linear

space. Then we get from f ⊗ x 7→ xf (·) a linear imbedding L ⊗ X ,→ X M .
In this way we identify from now on L ⊗ X with a linear subspace of X M .
6.14 Lemma: L ⊗ X is the set of all f ∈ X M with finite dimensional

range, such that for some (any) basis e1 , . . . , em of span f (M ) the coordinate
functions are in L.
71
Notice that for any such basis we have a unique expansion
∑
m
f= fj ej
j=1
with scalar valued functions fj . They are called coordinate functions.

The easy proof of the lemma is left to the reader.
Example: (1) Let M be a topological space, X a topological vector space.

Then C (M ) ⊗ X is the space of all continuous X-valued functions on
M with finite dimensional range.
(2) Let Ω ⊂ Rn be open, X again a topological vector space then for 0 ≤

p ≤ +∞ we have: C p (Ω) ⊗ X is the space of all p-times continuously
differentiable X-valued functions on Ω with finite dimensional range.
(3) Let X be a linear space, I an index set, φ(I) the set of all scalar valued
functions on I which are zero outside a finite set, then φ(I) ⊗ X is the
set of all X- valued functions on I which are zero outside a finite set.
If, under this identification, X carries a seminorm p and φ(I) is equipped
with the ℓ1 (I)-norm then for x = (xi )i∈I we have
∑
|| ||1 ⊗π p(x) = p(xi ).
i
Definition: If E is a locally convex vector space and P a fundamental

system of seminorms we define
{ ∑ }
ℓ1 {I, X} = x = (xi )i∈I ∈ | p̂(x) := p(xi ) < ∞, p ∈ P
i∈I
equipped with the seminorms p̂.
We obtain the
6.15 Theorem: If E is a complete locally convex vector space, then ℓ1 (I)⊗

ˆ πE
can be canonically identified with ℓ1 {I, X}.
72
6.2 Tensor products with spaces L1 (µ)
Let (Ω, F , µ) be a σ-finite measure space, i.e. F a σ-algebra on the set Ω

and µ a positive σ-additive set function on Ω.
We denote by T the linear space of step functions on Ω, i.e. T = span{χA |
A ∈ F , µA < +∞} where χA (ω) = 1 for ω ∑ ∈ A, χA (ω) = 0 other-
m
wise. Every f ∈ T has a representation f = j=1 λj χAj with disjoint
sets A1 , . . . , Am , m ∈ N. We call it a disjoint representation.
On T we consider the seminorm
∫ ∑
m
||f ||1 = |f (ω)|dµ(ω) = |λj |µ(Aj ),
Ω j=1
using a disjoint representation.

Let X be a Banach space. Then T ⊗ X is the set of all X-valued step
functions on Ω which admit a representation
∑
m
f (ω) = xj χAj (ω)
j=1
where the sets Aj ∈ F are disjoint and xj ∈ X, j = 1, ∑ . . . , m. We call it

again a disjoint representation. Notice that then ||f (·)|| = m j=1 ||xj ||χAj (·) ∈
T.
6.16 Lemma: For f ∈ T (X) we have
∫ ∑
m
|| ||1 ⊗π || || = ||f (ω)||dµ(ω) = ||xj ||µ(λj ),
Ω j=1
using a disjoint representation.
Proof: We set p = || ||1 ⊗π || ||. For a disjoint representation

∑
m
f (ω) = xj χAj (ω)
j=1
we obtain
∑
m ∑
m ∫
p(f ) ≤ p(xj χAj ) = ||xj ||µ(Aj ) = ||f ||dµ.
j=1 j=1 Ω
73
For any representation
∑
n
f= xk fk (∗)
k=1
where fk ∈ T , k = 1, . . . , n, we now choose disjoint sets A1 , . . . , Am ∈ F

such that for all k
∑m
fk = λk,j χAj .
j=1
Therefore
∑
m ∑ n
f= ( λk,j xk )χAj ,
j=1 k=1
is a disjoint representation. Hence

∫ ∑
m ∑
n ∑
m ∑
n
||f ||dµ = || λk,j xk ||µ(Aj ) ≤ |λk,j | ||xk ||µ(Aj )
Ω j=1 k=1 j=1 k=1
∑n ∑
m ∑n
= ||xk || |λk,j |µ(Aj ) = ||xk || ||fk ||1 .
k=1 j=1 k=1
Since this holds for every representation (∗) we have

∫
||f ||dµ ≤ p(f ),
Ω
which completes the proof. 2
We call an X-valued function on Ω measurable if it is the pointwise limit of

a sequence of X-valued step functions. If f ∈ X Ω is measurable then clearly
||f (·)|| is measurable.
Definition: L1 (µ, X)
∫ is the linear space of all measurable X-valued func-
tions f on Ω with X ||f ||dµ < +∞, where we identify almost everywhere
equal functions. We set ∫
||f ||1 = ||f ||dµ
X
for all f ∈ L1 (µ, X).
Obviously || ||1 is a norm on L1 (µ, X) and T ⊗π X ⊂ L1 (µ, X) as a normed

subspace.
74
6.17 Lemma: T ⊗ X is dense in L1 (µ, X).
is σ-finite we find Ω1 ⊂ Ω2 ⊂ . . . in F , µΩn < +∞ for all n,

Proof: Since Ω∪
such that Ω = n Ωn . We have
∫
||f − χΩn f ||1 = ||f ||dµ
Ω\Ωn
and therefore χΩn f → f in L1 (µ, X). So we may assume without restriction

of generality that µΩ < +∞.
f is measurable, hence pointwise limit of functions which take on only finitely
many values. Therefore f (Ω) is contained in a separable subspace of X. So
we may assume that X is separable.
∪
Let x1 , x2 , . . . be a dense subset of X and ε > 0. Then X = n Uε (xn ). We
set
ω1 = f −1 (Uε (x1 )), ωn+1 = f −1 (Uε (xn )) \ (ω1 ∪ . . . ∪ ωn−1 )
and
∞
∑
fε = xn χωn .
n=1
Obviously fε is measurable and ||fε || ≤ ||f || + ε. Therefore fε ∈ L1 (µ, X).
Since ||fε − f || ≤ ε everywhere we obtain ||fε − f ||1 ≤ εµΩ.
Since
∞
∑
||fε ||1 = ||xn ||µ(ωn ) < +∞
n=1
we have
∑
m ∞
∑
||fε − xn χωn ||1 = ||xn ||µ(ωn ) −→ 0,
n=1 n=m+1
which shows that fε , for every ε > 0, is a limit of step functions. So, finally,
f is a limit of step functions. 2
6.18 Theorem: L1 (µ, X) is a Banach space.
Proof: We have to prove completeness. It suffices to prove that every abso-

lutely convergent series in the dense subspace T ⊗ X converges in L1 (µ, X).
So assume that we have u1 , u2 , . . . ∈ T ⊗ X such that
∞
∑
||un ||1 < +∞.
n=1
75
Then by the Beppo Levi Theorem we have that
∞
∑
||un (ω)|| < +∞
n=1
for almost every ω and

∞
∑
g := ||un || ∈ L1 (µ).
n=1
Therefore the series

∞
∑
f (ω) := un (ω)
n=1
converges for almost every ω, f is measurable, and in L1 (µ, X) since ||f (ω)|| ≤
g(ω). We have
∑m ∞
∑
||f − un ||1 ≤ ||un ||1 ,
n=1 n=m+1
which shows that

∞
∑
f= un
n=1
with convergence in L1 (µ, X). 2
Now T ⊗π X ⊂ L1 (µ, X) is isometrically imbedded as a dense subspace. So

we have
L1 (µ, X) = T ⊗ˆ π X = L1 (µ)⊗
ˆ πX
which shows:
6.19 Theorem: L1 (µ, X) = L1 (µ)⊗

ˆ π X by means of the canonical imbed-
ding generated by f ⊗ x 7→ f .
If (Ω1 , F1 , µ1 ), (Ω2 , F2 , µ2 ) are σ-finite measure spaces and (Ω1 × Ω2 , F , µ)

denotes the product measure space, then the Fubini-Theorem says that
L1 (µ) = L1 (µ1 , L1 (µ2 )). Therefore we have the
6.20 Corollary: L1 (µ) = L1 (µ1 )⊗ ˆ π L1 (µ2 ) by means of the canonical imbed-

ding generated by f (ω1 ) ⊗ g(ω2 ) 7→ f (ω1 )g(ω2 ).
76
Definition: A Banach space X has the bounded approximation property
if there is a constant C and for every finite set e ⊂ X and ε > 0 a map
φ ∈ F (X) such that ||φ|| ≤ C and supx∈e ||x − φ(x)|| < ε.
6.21 Lemma: If X has the bounded approximation property, then it has

the approximation property.
Proof: Let∪K ⊂ X be compact. Given ε > 0 we find x1 , . . . , xm ∈ K such

that K ⊂ m j=1 Uε (xj ).
Using the notation in the above definition, we find φ ∈ F (X) such that,
||φ|| ≤ C and
sup ||xj − φ(xj )|| < ε.
j=1,...,m
For x ∈ K we find xj such that ||x − xj || < ε and therefore
||x − φ(x)|| ≤ ||x − xj || + ||xj − φ(xj )|| + ||φ(xj ) − φ(x)|| ≤ (2 + C)ε. 2
By the same argument as above we can prove:

6.22 Lemma: If X is Banach space, M ⊂ X a dense subset and for every
finite set e ⊂ M and ε > 0 we find φ ∈ F (X) such that ∥φ∥ ≤ C and
sup ||x − φ(x)|| < ε.

x∈e
Then X has the bounded approximation property.
6.23 Corollary: If X is separable then the following are equivalent
(1) X has the bounded approximation property.
(2) There is a sequence (φn )n∈N in F (X) such that limn φn (x) = x for all
x ∈ X.
Proof: (1) =⇒ (2): Let {x1 , x2 , . . .} be a dense subset of X. We find C > 0

and for every n a map φn ∈ F such that ||φn || ≤ C and
1
sup ||xj − φn (xj )|| < .
j=1,...,n n
By means of Lemma 6.22 we get (2).

(2) =⇒ (1): From the Banach-Steinhaus theorem we obtain a constant C
such that ||φn || ≤ C for all n. 2
77
Remark: The implication (2) =⇒ (1) in Corollary 6.23 holds also without
the assumption of separability.
6.24 Theorem: (1) ℓp has the bounded approximation property for all
0 ≤ p < +∞.
(2) Let (Ω, F , µ) be a measure space, then Lp (µ) has the bounded ap-
proximation property for all 0 ≤ p < +∞.
Proof: To prove (1) we use Corollary 6.23 with φn (x) = (x1 , . . . , xn , 0, . . .).
For (2) we use Lemma 6.22 with M = T . Let A = {A1 , . . . , Am } denote a
finite set of disjoint sets in F , 0 < µAj < +∞ for all j. We set for f ∈ Lp (µ)
∑
m ∫
1
PA f = f (ω)dµ(ω) χAj
µAj Aj
j=1
and we obtain
∫ (∫ )p
∑
m
1
|PA f | dµ =
p
|f (ω)|dµ(ω) µAj .
Ω µApj Aj
j=1
Because of
(∫ )p ∫
|f (ω)|dµ(ω) ≤ (µAj ) p/q
|f (ω)|p dµ(ω)
Aj Aj
and 1 + p/q − p = 0 we obtain ||PA f ||p ≤ ||f ||p .

If {f1 , . . . , fn } is a finite set of step functions then we may choose A such
that
∑m
fk = λk,j χAj
j=1
for all k = 1, . . . , n. It is easily seen that PA fk = fk for k = 1, . . . , n. 2
From Lemma 6.21 and Theorem 6.24 we obtain:

6.25 Corollary: L1 (µ) has the approximation property.
Then Corollary 6.8, Lemma 6.10 and Theorem 6.19 give:

6.26 Theorem: For any Banach space X we have N (X, L1 (µ)) ∼ = L1 (µ, X ′ )
isometrically. Here f ∈ L1 (µ, X ′ ) corresponds to the map Tf x = ⟨f (·), x⟩.
78
Example:
∑
N (c0 , ℓ1 ) ∼
= ℓ1 ⊗
ˆ π ℓ1 = {a = (ak,j )j,k∈N | ||a|| = |ak,j | < +∞}
k,j
isometrically. The matrix (ak,j )j,k∈N corresponds to the map

∑
ξ = (ξ1 , ξ2 , . . .) 7→ ( ak,j ξj )k∈N .
j
6.3 The ε-tensor product
Let E, F be linear spaces, ξ, η linear forms on E and ∑F , respectively. Then

(x, y) 7→ ξ(x)η(y) is bilinear on E × F , hence for u = mj=1 xj ⊗ yj ∈ E ⊗ F
∑
m
⟨ξ, u, η⟩ := ξ(xj )η(yj )
j=1
does not depend on the representation of u and u 7→ ⟨ξ, u, η⟩ is a linear form

on E ⊗ F .
Definition: Let p on E, q on F be seminorms. Then for u ∈ E ⊗ F we set
p ⊗ε q(u) := sup{ |⟨ξ, u, η⟩| | p∗ (ξ) ≤ 1, q ∗ (η) ≤ 1}.
As usually p∗ (ξ) := sup{ |ξ(x)| | p(x) ≤ 1} and q ∗ (η) analogous.
6.27 Lemma: (1) p ⊗ε q is a seminorm on E ⊗ F
(2) p ⊗ε q(x ⊗ y) = p(x)q(y)
(3) p ⊗ε q ≤ p ⊗π q
∑m
Proof: (1) and (2) are immediate. To prove (3) let u = j=1 xj ⊗ yj . Then,
because of the triangular inequality and (2), we have
∑
m
p ⊗ε q(u) ≤ p(xj )q(yj ).
j=1
Since this holds for all representations of u we obtain (3). 2
79
Let E, F be locally convex, p a continuous seminorm on E, q on F , and
U = {x ∈ E | p(x) ≤ 1}, V = {y ∈ F | q(y) ≤ 1}. We have the canonical
imbeddings

BL(E ′ ,F ) : E ⊗ F / L(E ′ , F )

BL(F ′ ,E) : E ⊗ F / L(F ′ , E) .
Then the following is quite obvious
6.28 Lemma:
p ⊗ε q(u) = sup{ξ ∈ U o | q((BL(E ′ ,F ) u)(ξ))}

= sup{η ∈ V o | p((BL(F ′ ,E) u)(η))}.
From this we derive easily that for u ∈ E ⊗ F , u ̸= 0 we find a continuous

seminorm such that p ⊗ε q(u) ̸= 0. Namely set u e := BL(E ′ ,F ) u, then ũ ̸= 0.
Therefore we have ξ ∈ E ′ , such that ũ(ξ) ̸= 0. We find p so that ξ ∈ U o and
q such that q(ũ(ξ)) ̸= 0 and obtain p ⊗ε q(u) ̸= 0.
Since p ⊗ε q increases when p and q are increased the following definition
makes sense.
Definition: Let E, F be locally convex. Then E ⊗ε F is E ⊗ F equipped

with the locally convex topology given by the seminorms p ⊗ε q where p runs
through the continuous seminorms on E, q those on F . Its completion is
denoted by E ⊗ ˆ εF .
Remark: If P is a fundamental system of seminorms on E, Q a funda-

mental system of seminorms on F , then {p ⊗ε q | p ∈ P, q ∈ Q} is a
fundamental system of seminorms on E ⊗ε F .
In particular, for E, F metrizable the space E ⊗ε F is metrizable, for E, F
normable the space E ⊗ε F is normable.
Let M be a set, L a linear space of bounded functions on M , equipped with

the norm
||f || = sup |f (x)|.
x∈M
80
Let E be a linear space equipped with a seminorm p, then for u ∈ L ⊗ E ⊂
E M we have
|| || ⊗ε p(u) = sup{||η ◦ u(·)|| | p∗ (η) ≤ 1}
= sup{|η(u(x))| | x ∈ M, p∗ (η) ≤ 1}
= sup{p(u(x)) | x ∈ M }
=: p̃(u).
Example: For complete E we obtain immediately that c0 ⊗ ˆ εE ∼
= c0 (E),
where c0 (E) is the space of all null sequences in E with the seminorms
p̃(x) = supn p(xn ), p running through all continuous seminorms on E.
If M is a compact topological space, we set

C (M, E) = { continuous E-valued functions }
equipped with the locally convex topology given by the seminorms
p̃(f ) := sup{p(f (x)) | x ∈ M }
where p runs through all continuous seminorms on E (or, equivalently,
through a fundamental system of seminorms on E). If E is complete, then
C (M, E) is complete.
6.29 Theorem: If E is complete then
C (M, E) = C (M )⊗
ˆ ε E.
Proof: Due to the previous we have only to show that C (M ) ⊗ E is dense

in C (M, E). We choose a seminorm p ∪ on E and ε > 0. For f ∈ C (M, E)
we find a1 , . . . , am ∈ E such that M = m
j=1 {x | p(aj − f (x)) < ε}.
∑m
We put φj = max(0, ε − p(aj − f (x))) and φ = j=1 φj . Notice that
φ(x) > 0 for all x ∈ M . We put
∑
m
φj
u= aj ∈ C (M ) ⊗ E.
φ
j=1
Then we have for x ∈ M

 
∑m
φ (x) ∑
m
φj (x)
p(u(x) − f (x)) = p  (aj − f (x)) ≤
j
p(aj − f (x)) < ε.
φ(x) φ(x)
j=1 j=1
Therefore p̃(u − f ) < ε. 2
81
If X and Y are normed spaces we consider X ⊗ε Y , by means of || ||⊗ε || ||, in
canonical way as a normed space. X ⊗
ˆ ε Y is then in canonical way a Banach
space.
If X is a Banach space then C (M, X) again is a Banach space by means of
the norm
||f || = sup ||f (x)||.
x∈M
and we can show more precisely:
6.30 Theorem: For compact M and a Banach space X we have C (M, X) ∼ =

C (M )⊗
ˆ ε X isometrically, by means of the continuous linear map generated
by f ⊗ x 7→ f (·)x.
6.31 Corollary: If M1 , M2 are compact, then C (M1 ×M2 ) ∼

= C (M1 )⊗
ˆ ε C (M2 )
isometrically.
We will now consider the space ℓ1 (I)⊗ ˆ ε E, where I is an index set, E is

locally convex and complete. Let p be a continuous seminorm on E, U =
{x ∈ E | p(x) ≤ 1}. For x = (xi )i∈I ∈ ℓ1 (I) ⊗ E and p̃ = || ||1 ⊗ε p we have
∑
p̃(x) = sup |η(xi )|.
η∈U o i∈I
We consider the space

{ ∑ }
ℓ1 [I, E] = x = (xi )i∈I ∈ E I | |η(xi )| < +∞ for all η ∈ E ′
i∈I
of weakly summable families over I.
6.32 Theorem: ℓ1 [I, E] equipped with the seminorms

∑
p̃(x) = sup |η(xi )|
η∈U o i∈I
where U runs through all neighborhoods of zero in E is a locally convex

space. If E is complete then also ℓ1 [I, E] is complete.
Proof: First we have to show that for every neighborhood of zero U = {x |

p(x) ≤ 1} in E ∑
p̃(x) := sup |η(xi )|
η∈U o i∈I
82
is finite, hence is a seminorm on ℓ1 [I, E]. We set for r > 0
∑
Vr ={η ∈ E ′ | |η(xi )| ≤ r}
i∈I
∩ ∑
= {η ∈ E ′ | |η(xi )| ≤ r}.
e⊂I i∈e
e finite
∪ ∪
Vr is absolutely convex and σ(E ′ , E)-closed. Since r>0 rV1 = r>0 Vr = E ′
by definition of ℓ1 [I, E] we obtain that V1 is a barrel in E ′ . Consequently
EU′ o ∩ V1 is a barrel in the Banach space EU′ o , and therefore there is ε > 0
such that εU o ⊂ V1 . This implies that U o ⊂ 1ε V1 = V1/ε , i.e.
∑ 1
sup |η(xi )| ≤ .
η∈U o i∈I ε
It remains to show that ℓ1 [I, E] is complete if E is complete. Let (xτ )τ ∈T

be a Cauchy net in ℓ1 [I, E]. For any τ we have xτ = (xτ,i )i∈I . For any i ∈ I
and any continuous seminorm p on E we have for all τ, σ ∈ T
p(xτ,i − xσ,i ) = sup |η(xτ,i − xσ,i )| ≤ p̃(xτ − xσ ).

η∈U o
Therefore (xτ,i )τ ∈T is a Cauchy net in E and, due to completeness, there is

xi ∈ E such that limτ xτ,i = xi .
For any continuous seminorm p and ε > 0 we have τ0 = τ0 (ε) such that for
τ, σ ≻ τ0
p̃ (xτ − xσ ) ≤ ε.
This means that for every η ⊂ U o , finite e ∈ I, τ, σ ≻ τ0 we have
∑
|η(xτ,i − xσ,i )| ≤ ε.
i∈e
Taking the limit with respect to σ we get that for every η ∈ U o , finite e ⊂ I,
τ ≻ τ0 ∑
|η(xτ,i − xi )| ≤ ε. (∗)
i∈e
First we choose ε = 1 find τ = τ0 (1) and obtain

∑
|η(xi )| ≤ 1 + p̃(xτ )
i∈e
83
for all finite e ⊂ I, η ∈ U o , i.e.
∑
|η(xi )| < +∞ (∗∗)
i∈I
for all η ∈ U o . Since we can apply this to every seminorm p, we get (∗∗) for
all η ∈ E ′ , i.e. x ∈ ℓ1 [I, E].
Then we come back to (∗) which implies that for all τ ≻ τ0
∑
p̃(xτ − x) = sup |η(xτ,i − xi )| ≤ ε.
η∈U o i∈I
Therefore xτ → x. 2
By means of the consideration previous to the theorem we have
ℓ1 (I) ⊗ε E ⊂ ℓ1 [I, E]
as a topological linear subspace. To determine ℓ1 (I)⊗

ˆ ε E we have to deter-
mine the closure of ℓ1 (I) ⊗ε E in ℓ1 [I, E].
Let ℓ1 (I, E) be the linear space of all summable families (xi )i∈I in E.
Summable means that there is an x such that for every seminorm p and
ε > 0 we have a finite set e0 ⊂ I such that for all finite sets e with e0 ⊂ e ⊂ I
∑
p(x − xi ) ≤ ε.
i∈e
6.33 Lemma: If E is complete then ℓ1 (I, E) is a closed subspace of ℓ1 [I, E].
Proof: Let (xτ )τ ∈T be a net in ℓ1 (I, E), convergent in ℓ1 [I, E], x := limτ xτ .
For U = {x | p(x) ≤ 1} and ε > 0 we choose τ such that p̃(xτ − x) < ε. We
(∑ set e0)⊂ N such that for any finite set e ⊂ I with e ∩ e0 = ∅
choose a finite
we have p i∈e xτ,i ≤ ε. We obtain for finite e ∈ I, e ∩ e0 = ∅
( ) ( ) ( )
∑ ∑ ∑
p xi ≤ p xτ,i + p (xi − xτ,i )
i∈e i∈e i∈e
∑
≤ ε + sup | η(xi − xτ,i )|
η∈U o i∈e
≤ ε + p̃(x − xτ ) ≤ 2ε.
∑
Therefore i∈I xi fulfills Cauchy’s convergence criterion and, due to the
completeness of E, is convergent. 2
84
6.34 Lemma: ℓ1 (I) ⊗ E ⊂ ℓ1 (I, E) as a dense subspace.
Proof: Of course, we have only to show the density. For x = (xi )i∈I ∈
ℓ1 (I, E) and finite e ⊂ I we define x(e) by x(e)i = xi for i ∈ e, x(e)i = 0
otherwise. Then clearly x(e) ∈ ℓ1 ⊗ E and for any continuous seminorm p
on E we have ∑
p̃(x − x(e)) = sup |η(xi )|.
η∈U o
i∈e
/
We choose a∑finite set e0 ⊂ I so large that for

∑ any finite e ⊂ I with e∩e0 = o∅
we have p( i∈e xi ) ≤ ε and therefore | ∑ i∈e η(xi )| ≤ ε for all η ∈ U .
Therefore we have (see the remark below) i∈I\e0 |η(xi )| ≤ 4ε for all η ∈ U o
which means p̃(x − x(e0 )) ≤ 4ε. 2
Remark: If (ξi )i∈I is a family in R or C and

∑
| ξi | ≤ ρ
i∈e
for all finite subsets e ⊂ I. Then

∑
|ξi | ≤ 2ρ
i∈I
in the real case ∑

|ξi | ≤ 4ρ
i∈I
in the complex case.
Proof: We give the proof for the real case, the complex case is, mutatis
mutandis, the same. We set I1 = {i ∈ I | ξi ≥ 0}, I2 = {i ∈ I | ξi < 0}
and obtain, applying the assumption to finite sets e ⊂ I1 , resp. e ⊂ I2 ,
∑ ∑
ξi ≤ ρ, − ξi ≤ ρ
i∈I1 i∈I2
∑
hence i∈I |ξi | ≤ 2ρ. 2
Finally we have proved:
6.35 Theorem: If E is complete then ℓ1 (I)⊗

ˆ ε E = ℓ1 (I, E).
85
6.4 Tensor products and nuclearity
Let E be nuclear and x = (xi )i∈I ∈ ℓ1 [I, E]. Let p be a continuous seminorm,
U = {x | p(x) ≤ 1}. Then we find a continuous seminorm q, V = {x |
q(x) ≤ 1} and a positive Radon measure µ such that
∫
p(x) ≤ |η(x)|dµ(η).
Vo
For any finite set e ⊂ I we obtain

∑ ∑∫
p(xi ) ≤ |η(xi )|dµ(η)
i∈e i∈e Vo
∫ ∑
= |η(xi )|dµ(η)
V o i∈e
∑
≤ µ(V o ) sup |η(xi )|
η∈V o i∈I
= µ(V o ) q̃(V ).
∑ o
Therefore i∈I p(xi ) < µ(V ) q̃(V ). Since we may do that for any con-
tinuous seminorm p we have x ∈ ℓ1 {I, E}, and we can find for any con-
tinuous seminorm p a continuous seminorm q and a constant C such that
p̃(x) ≤ C q̃(x). We have shown:
6.36 Theorem: If E is nuclear, then ℓ1 [I, E] = ℓ1 {I, E} as topological

vector spaces. In particular we have ℓ1 (I) ⊗ε E = ℓ1 (I) ⊗π E, therefore
ℓ1 (I)⊗
ˆ ε E = ℓ1 (I)⊗
ˆ π E. If E is complete then ℓ1 (I, E) = ℓ1 {I, E}.
For this Theorem also the converse is true and this gives exactly the orig-
inal definition of nuclearity by Grothendieck. To prove it we need some
preparation.
Definition: Let X, Y be Banach spaces. A map A ∈ L(X, Y ) is called abso-

lutely summing if there is C > 0 such that for any finite set {x1 , . . . , xm } ⊂ X
we have
∑m ∑m
||Axj || ≤ C sup |η(xj )|.
j=1 ||η||≤1 j=1
η∈X ′
We denote by π(A) the infimum over all such constants C.
86
6.37 Theorem: A ∈ L(X, Y ) is absolutely summing if and only if there is
a positive Radon measure µ on U o = {y ∈ X ′ | ||y|| ≤ 1}, such that
∫
||Ax|| ≤ |η(x)|dµ(y)
Uo
for all x ∈ X.
Proof: Assume that ∫

||Ax|| ≤ |η(x)|dµ(y) (9)
Uo
Then we have,
∑
m ∫ ∑
m
||Axj || ≤ |η(xj )|dµ(y)
j=1 U o j=1
∑
m
≤µ(U ) supo
|η(xj )|.
η∈U o j=1
Therefore A is absolutely summing and π(A) ≤ inf{µ(U o ) | µ satisfies (9)}.

To prove the converse, we assume that A is absolutely summing. We put
I := U o and for x ∈ ℓ1 (I, E) we set
∑
a(x) := η(Axη ).
η∈U o
We have:
∑ ∑ ∑
|η(Axη )| ≤ ||Axη || = sup ||Axη ||
η∈U o η∈U o e∈U o
e finite η∈e
∑
≤π(A) sup sup |y(xη )|
e∈U o ||y||≤1 η∈e
e finite y∈X ′
∑
=π(A) sup |y(xη )| = π(A) p̃(x)
||y||≤1 η∈U o
y∈X ′
where p̃ is taken from p(·) = || · ||. Therefore the sum converges and |a(x)| ≤
π(A) p̃(x). We have shown that a ∈ ℓ1 (I, E)′ , ||a|| ≤ π(A).
We set K = [0, 1]U × U o , where U o carries the weak∗ -topology. Then K is
o
compact. We define a map
Φ : ℓ1 (I, E) −→ C(K)
87
by ∑
Φ(x)[λ, y] := λη y(xη )
η∈U o
o
for λ = (λη )η∈U o ∈ [0, 1]U and y ∈ U o .
Clearly Φ(x) is a function on K. We have to show that it is continuous.
Let e ⊂ U o be a finite set. Then
∑
fe (λ, y) := λη y(xη )
η∈e
is continuous on K. We have
∑ ∑
|Φ(x)[λ, y] − fe (λ, y)| = λη y(xη ) ≤ |y(xη )|
η∈U o \e η∈U o \e
∑
≤ sup |y(xη )| = p̃(x − x(e)) ≤ ε
y∈U o
η∈U o \e
for e large enough. For the definition of x(e) and the last conclusion see the
proof of Lemma 6.34.
Therefore fe → Φ(x) uniformly on K, hence Φ(x) ∈ C(K). Since
∑ ∑
∥Φ(x)∥ = sup λη y(xη ) = sup |y(xi )| = p̃(x)
(λ,y)∈K η∈U o y∈U o i∈I
Φ is an isometric imbedding ℓ1 (I, E) ,→ C(K). We set b = a ◦ Φ−1 on

Φ(ℓ1 (I, E)) and extend it by the Hahn-Banach Theorem to µ̃ ∈ C(K)′ with
||µ̃|| = ||b|| = ||a|| ≤ π(A). By the Theorem of Riesz there exist a measurable
function φ0 on K, ||φ0 ||∞ = 1 and a positive Radon measure such that
∫
µ̃f = f φ0 dµ0
K
for all f ∈ C(K).

For ξ ∈ E and η ∈ U o we put now
{
ξ : i=η
xi =
0 : otherwise.
This defines x ∈ ℓ1 (I, E). We have

∑
a(x) = i(Axi ) = η(Aξ)
i∈I
88
and
Φ(x)[λ, y] = λη y(ξ).
Therefore
|η(Aξ)| =|a(x)| = |b(Φ(x))| = |µ̃(Φ(x))|
∫
= λη y(ξ)φ0 (λ, y)dµ0 (λ, y)
∫ K
≤ |y(ξ)|dµ0 (λ, y)
∫K
= |y(ξ)|dµ(y)
Uo
where we have defined the measure µ on U o by
∫ ∫
g(y)dµ(y) = g(y)dµ0 (λ, y).
Uo K
Since the estimate holds for all η ∈ U0 the proof is complete. 2
Notice that µ(U o ) = µ0 (K) = ||µ̃|| ≤ π(A). Together with the estimate
which we have seen earlier, we have shown:
∫
π(A) = inf{µ(U ) | ||Ax|| ≤
o
|y(x)|dµ(y)}
Uo
where µ denotes a positive Radon measure on U o .
We are now ready to prove the intended theorem.
6.38 Theorem: E is nuclear if and only if ℓ1 ⊗ε E = ℓ1 ⊗π E.
Proof: The necessity has already be shown. If ℓ1 ⊗ε E = ℓ1 ⊗π E then
for every continuous seminorm p on E there is a continuous seminorm q
on E such that p̂ ≤ q̃ and therefore we have for any finitely many vectors
x1 , . . . , xm ∈ E that p̂(x1 , . . . , xm , 0, . . .) ≤ q̃(x1 , . . . , xm , 0, . . .), i.e.
∑
m ∑
m
p(xj ) ≤ sup |y(xj )|. (10)
j=1 y∈V o j=1
Here V = {x ∈ E | q(x) ≤ 1}. We denote by Ep and Eq the local Banach

spaces of p and q and ıpq : Eq −→ Ep the canonical map. Then (10) means
that ıpq is absolutely summing and therefore Theorem 6.37 implies that
∫
p(x) = p(ıq x̂) ≤
p
|η(x)|dµ(y)
Vo
for some positive Radon measure µ on V o. This shows that E is nuclear. 2
89
We will extend now this investigation to a more general situation. Let E
and F be locally convex spaces, E nuclear.
We assume that U , V are absolutely convex neighborhoods of zero in E,
V ⊂ U , such that the canonical map ıU
V : EV −→ EU is nuclear. Then ıV
U
can be represented in the form

∞
∑
ıU
Vξ = e∗n (ξ) ên
n=1
where e∗n ∈ EV∗ , en ∈ E, ên its residue class in En and

∞
∑
||e∗n ||∗V ||en ||U < ∞.
n=1
∑m
For u = k=1 ak ⊗ bk ∈ E ⊗ F we set
∑
m
fn = e∗n (ak )bk .
k=1
Let W be an absolutely convex neighborhood of zero in F then
||fn ||W = sup |y(fn )| = sup |⟨e∗n , u, y⟩|

y∈W o y∈W o
≤||e∗n ||∗V · || ||V ⊗ε || ||W (u).
For any M ∈ N we obtain
∑
M ∑
M (∑
m )
en ⊗ f n = en ⊗ e∗n (ak )bk
n=1 n=1 k=1
m (∑
∑ M )
= e∗n (ak )en ⊗ bk
k=1 n=1
and therefore
∑
M m (
∑ ∑
M )
u− en ⊗ fn = ak − e∗n (ak )en ⊗ bk .
n=1 k=1 n=1
We can estimate
∑
M ∞
∑ ∞
∑
ak − e∗n (ak )en = e∗n (ak )ên ≤ ||e∗n ||∗V ||en ||U ||ak ||V .
U U
n=1 n=M +1 n=M +1
90
Finally we have
(∑
M )
∥ ∥U ⊗π ∥ ∥W (u) ≤∥ ∥U ⊗π ∥ ∥W en ⊗ fn
n=1
( ∑
M )
+ ∥ ∥U ⊗π ∥ ∥W u − en ⊗ fn
n=1
∑
M m ( ∑
∑ ∞ )
≤ ∥en ∥U ∥fn ∥W + ||e∗n ||∗V ||en ||U ||ak ||V ∥bk ∥W
n=1 k=1 n=M +1
∑
M ∞
∑ ∑
m
= ∥en ∥U ∥fn ∥W + ∥e∗n ∥∗V ∥en ∥U ∥ak ∥V ∥bk ∥W
n=1 n=M +1 k=1
∑∞
≤ ∥en ∥U ∥e∗n ∥∗V · ∥ ∥V ⊗ε ∥ ∥W (u)
n=1
∞
∑ ∑
m
+ ∥e∗n ∥∗V ∥en ∥U · ∥ak ∥V ∥bk ∥W
n=M +1 k=1
Letting M → +∞ and taking the inf over all nuclear representations of ıU

V
we obtain
∥ ∥U ⊗π ∥ ∥W ≤ ν(ıUV )∥ ∥U ⊗ε ∥ ∥W
and we have proved:

6.39 Theorem: If E and F are locally convex spaces and E is nuclear then
E ⊗π F = E ⊗ε F .
Combining this with Theorem 6.38 we get

6.40 Theorem: The following are equivalent:
(1) E is nuclear
(2) E ⊗π F = E ⊗ε F for every locally convex space F
(3) E ⊗π F = E ⊗ε F for every Banach space F
(4) E ⊗π ℓ1 = E ⊗ε ℓ1 .
The equivalence of (1) and (2) is the original definition of nuclearity given
by Grothendieck.
91
6.5 The kernel theorem
We will prove the classical kernel theorem of L. Schwartz. First we calculate

some tensor products.
Using the notation of Theorem 3.21 we have:
6.41 Theorem: C(r) ∞ (Rn1 )⊗ ∞ (Rn2 ) = C ∞ (Rn1 +n2 ) by means of the map
ˆ ε C(r) (r)
f (x1 ) ⊗ g(x2 ) 7→ f (x1 )g(x2 ).
Proof: We use the same argument as in the proof of Theorem 6.30.

For n ∈ N, k ∈ N0 and the canonical seminorm
∥f ∥k = sup{|f (α) (x)| | x ∈ Rn , |α| ≤ k}
∞ (Rn ) we set U = {f ∈ C ∞ (Rn ) | ∥f ∥ ≤ 1}. Then we have for
on C(r) k (r) k
∞ 1 ∞ 2 ∞
f ∈ C(r) (R ) ⊗ C(r) (R ) ⊂ C(r) (R
n n n 1 +n 2 ) by Lemma 6.28
∥ ∥k ⊗ε ∥ ∥k (f ) = sup ∥µy f (·, y)∥k

µ∈Uko
{ ∂ |α| }
= sup sup µ y f (x, y) | x ∈ R n1
, |α| ≤ k
µ∈Uko ∂xα
{ ∂ |α| }
= sup f (x, ·) | x ∈ R n1
, |α| ≤ k
∂xα k
{ ∂ |α| ∂ |β| }
= sup f (x, y) | x ∈ R n1
, y ∈ R n2
, |α| ≤ k, |β| ≤ k .
∂xα ∂y β
Therefore
∥f ∥k ≤ ∥ ∥k ⊗ε ∥ ∥k (f ) ≤ ∥f ∥2k
for all k, which shows that
∞ ∞ ∞
C(r) (Rn1 ) ⊗ε C(r) (Rn2 ) ⊂ C(r) (Rn1 +n2 )
as a topological linear subspace.
∞ (Rn1 ) ⊗ C ∞ (Rn2 ) is dense in C ∞ (Rn1 +n2 ). For
It remains to show that C(r) (r) (r)
∞ (Rn1 +n2 ) let
f ∈ C(r)
∑ 1 ∑n1 1 ∑n2
f (x, y) = ck,l e2πi r ν=1 kν xν e2πi r ν=1 lν yν
k∈Zn1 ,l∈Zn2
∞ (Rn1 ) ⊗ C ∞ (Rn2 ) and
be the Fourier expansion. The partial sums are in C(r) (r)
∞ (Rn1 +n2 ).
the series converges in C(r) 2
92
For compact K ⊂ Rn we set
D(K) = {φ ∈ C ∞ (Rn ) | suppφ ⊂ K}
with the induced topology of C ∞ (Rn ). A canonical fundamental system of

seminorms are the same norms ∥ ∥k as above in C(r) ∞ (Rn ). D(K) is a closed
subspace of C ∞ (Rn ), hence a Fréchet space. For any set M ⊂ Rn we put
D(M ) = {φ ∈ C ∞ (Rn ) | supp φ compact, supp φ ⊂ M },
without assigning a topology to it.

◦
6.42 Lemma: For any compact K ⊂ Rn the space D(K) is dense in D(K).
∫
Proof: Let φ1 ∈ D(Rn ), supp φ1 ⊂ {x | |x| < 1}, φ1 (x)dx = 1 and set
φε (x) = ε−n φ1 (x/ε). We put Kε = {x ∈ K | dist(x, ∂K) ≥ ε} and
∫
χε (x) = φε (x − ξ)dξ.
Kε
◦
Then χε ∈ D(K ), χε (x) = 1 for x ∈ K2ε and ∥χε ∥k ≤ Ck ε−k for ε ≤ 1.
For φ ∈ D(K) and 0 < ε < 1 we obtain
∑ (α)
∥φ − φχε ∥k =∥φ (1 − χε )∥k ≤ Ck′ sup |φ(β) (x)|ε|β|−|α|
x̸∈K2ε β
|α|≤k
β≤α
≤Ck′′ ∥φ∥k+1 εk+1−|β|+|β|−k = Ck′′ ∥φ∥k+1 ε.
For the second inequality we used that
|φ(β) (x)| ≤ (2ε)k+1−|β| ∥φ∥k+1
on K \ K2ε = {x ∈ K | dist(x, ∂K) < 2ε}.

◦
Since φχε ∈ D(K ) we obtain the result. 2
We use the previous Lemma to show:

6.43 Theorem: For any compact sets K1 ⊂ Rn1 , K2 ⊂ Rn2 we have
D(K1 )⊗
ˆ ε D(K2 ) = D(K1 × K2 )
by canonical identification.
93
Proof: As in the case of periodic functions we have D(K1 ) ⊗ε D(K2 ) ⊂
D(K1 × K2 ) as a linear topological subspace. It remains to show that
D(K1 ) ⊗ D(K2 ) is dense in D(K1 × K2 ).
For that it suffices, due to Lemma 6.42, to show that
◦ ◦
D(K1 ) ⊗ D(K2 ) ⊃ D(K1 × K2 ).
◦ ◦
For φ ∈ D(K1 × K2 ) we choose ψ ∈ D(K1 × K2 ), ψ(x, y) = ψ1 (x)ψ2 (y),
x ∈ Rn1 , y ∈ Rn2 , such that ψ(x, y) = 1 for (x, y) ∈ supp φ.
Since, in the notation of Lemma 6.42, we have supp φ ⊂ K1,2ε × K2,2ε for
suitable ε > 0, we may use ψ1 (x) = χ1ε (x), ψ2 (y) = χ2ε (y) as in the proof of
Lemma 6.42 where χjs is defined on Rnj , j = 1, 2.
We choose r > diam(K1 × K2 ) and set with n = n1 + n2
∑
ϕ(ξ) := φ(ξ + rk)
k∈Zn
∞ (Rn ) and ϕψ = φ. Therefore, for ξ = (x, y) as
for ξ ∈ Rn . Then ϕ ∈ C(r)
before,
∑ 1 ∑n1 1 ∑n2
φ(x, y) = ck,l ψ1 (x)ψ2 (y)e2πi r ν=1 kν xν e2πi r ν=1 kν yν
k∈Zn1 ,l∈Zn2
where ck,l are the Fourier coefficients of ϕ. Arguing as in the proof of The-
orem 6.41 we obtain the result. 2
Since for compact K ⊂ Rn the space D(K) is a closed subspace of C ∞ (Rn )

it is nuclear (see Theorem 3.24). From Theorem 6.40 and Theorem 6.43 we
conclude
6.44 Corollary: For any compact set K1 ⊂ Rn1 , K2 ⊂ Rn2 we have D(K1 ×
K2 ) = D(K1 )⊗
ˆ π D(K2 ) by canonical identification.
To formulate the classical kernel theorem we need the following concept.

For Ω ⊂ Rn and any locally convex space G we define L(D(Ω), G) to be
the linear space of all continuous linear maps A : D(Ω) −→ G such that
A|D(K) ∈ L(D(K), G) for every compact set K ⊂ Ω.
We set D ′ (Ω) := L(D(K), K) and call it the space of Schwartz distributions.
It is a locally convex space with the fundamental system of semi-norms
pB (T ) := sup |T φ|,
φ∈B
94
where B runs through the following sets: there is a compact set K ⊂ Ω such
that B ⊂ D(K) and bounded there.
For any locally convex space a linear map A : G −→ D ′ (Ω) is continuous if,
and only if, x 7→ Ax|D(K) is continuous for every compact K ⊂ Ω.
6.45 Theorem: Let Ω1 ⊂ Rn1 , Ω2 ⊂ Rn2 be open. For every continuous

linear map A ∈ L(D(Ω1 ), D ′ (Ω2 )) there is a distribution T ∈ D ′ (Ω1 × Ω2 )
such that (Aφ)ψ = T (φ(x)ψ(y)) for every φ ∈ D(Ω1 ), ψ ∈ D(Ω2 ).
Proof: The bilinear form B(φ, ψ) := (Aφ)ψ on D(Ω1 ) × D(Ω2 ) extends to a

linear form T0 on D(Ω1 ) ⊗ D(Ω2 ). Let K1 ⊂ Ω1 , K2 ⊂ Ω2 be compact and
call SK1 ,K2 the restriction of T0 to D(K1 ) ⊗ D(K2 ). If we can show that
SK1 ,K2 is continuous on D(K1 ) ⊗π D(K2 ), then it extends ∪ to a continuous
linear form TK1 ,K2 on D(K1 )⊗ ˆ π D(K2 ) = D(K1 × K2 ). K1 ,K2 D(K1 ×
K2 ) = D(Ω1 × Ω2 ) and for L1 ⊂ K1 , L2 ⊂ K2 we have TK1 ,K2 |D(L1 ×L2 ) =
TL1 ,L2 , because the restriction of both to D(L1 )⊗D(L2 ) is SL1 ,L2 . Therefore
we have a linear form T on D(Ω1 × Ω2 ) such that T |D(K1 ×K2 ) = TK1 ,K2 for
all K1 , K2 .
If K ⊂ Ω1 × Ω2 is compact we find compact K1 ⊂ Ω1 , K2 ⊂ Ω2 such that
K ⊂ K1 × K2 and therefore T |D(K) = (T |D(K1 ×K2 ) )|D(K) is continuous.
So it remains to show that SK1 ,K2 is continuous on D(K1 ) ⊗π D(K2 ), or,
that B|D(K1 )×D(K2 ) is a continuous bilinear form. If ψ ∈ D(K2 ) is fixed and
φn → 0 in D(K1 ) then Aφn → 0 in D ′ (Ω), in particular (Aφn )ψ → 0. This
shows that for every ψ ∈ D(K2 ) the linear form φ 7→ B(φ, ψ) is continuous
on D(K1 ).
If φ ∈ D(K1 ) is fixed and ψn → 0 in D(K2 ) then (Aφ)ψn → 0, since
Aφ ∈ D ′ (Ω). So for every φ ∈ D(K1 ) the linear form ψ 7→ B(φ, ψ) is
continuous on D(K2 ).
We have show that B|D(K1 )×D(K2 ) is separately continuous bilinear form on
D(K1 ) × D(K2 ). The following Theorem completes the proof. 2
6.46 Theorem: If E and F are Fréchet spaces then every separately con-
tinuous bilinear form B on E × F is continuous.
Proof: For every y ∈ F and bounded M ⊂ E we have by assumption
sup |B(x, y)| < +∞. (11)

x∈M
95
We fix M and set
Ln ={y ∈ F | sup |B(x, y)| ≤ n}

x∈M
∩
= {y ∈ F | |B(x, y)| ≤ n}.
x∈M
Ln is an intersection of absolutely convex closed sets, hence absolutely con-

vex and closed. Obviously Ln = n · L1 .
∪ ∪
By (11) we have n nL1 = n Ln = F , i.e. L1 is a barrel, hence by Theorem
1.3 a neighborhood of zero in F . We have shown that supx∈M |B(x, ·)| is a
continuous seminorm on F .
Now assume xn → x in E, ym → y in F . Then
|B(xn , yn ) − B(x, y)| ≤|B(xn , yn ) − B(xn , y)| + |B(xn , y) − B(x, y)|

=|B(xn , yn − y)| + |B(xn − x, y)|
≤ sup |B(xν , yn − y)| + |B(xn − x, y)|.
ν
Since, due to the previous, supν |B(xν , ·)| is a continuous seminorm on F

and B(·, y) is continuous, both terms on the right hand side converge to
zero. 2
7 Interpolational Invariants
Throughout this section E, F, . . . will be Fréchet spaces with a fundamental

system ∥ ∥0 ≤ ∥ ∥1 ≤ ∥ ∥2 . . . of seminorms.
Definition: E has property (DN) (or: E ∈ (DN)) if the following holds
∃p ∀k ∃K, C : ∥ ∥2k ≤ C∥ ∥p ∥ ∥K .
∥ ∥p is called a dominating norm.
Remark: A dominating norm is a norm on E.
7.1 Lemma: (1) If E ∼

= F and E ∈ (DN) then F ∈ (DN).
(2) If F ⊂ E is a subspace and E ∈ (DN) then also F ∈ (DN).
96
Examples for (DN) and non (DN) spaces are given in the following propo-
sition.
7.2 Proposition: Λr (α) ∈ (DN) if, and only if, r = +∞.
Proof: If r = +∞ then the norms (| |k )k∈N0 are a fundamental system of

seminorms and we have, due to the Cauchy-Schwarz inequality:
∑
|x|2k = |xj |2 e2kαj
j
(∑ )1 ( ∑ )1
2 2
≤ |xj |2 |xj |2 e4kαj
j j
= |x|0 |x|2k .
So Λ∞ (α) has property (DN) with | |0 as a dominating norm. Now assume

Λr (α) ∈ (DN) and let tk ↗ r. Assume | |t0 is a dominating norm. Then for
every k there is K, C such that
|x|2tk ≤ C|x|t0 |x|tK
for all x ∈ Λr (α). We apply that to ej and obtain
e2tk αj ≤ Ce(t0 +tK ) αj
and therefore for large j

log C 1
tk ≤ + (t0 + tK ).
2αj 2
Letting j → +∞ we get for all k

1
tk ≤ (t0 + r),
2
which is possible only for r = +∞. 2
This can, of course, also be derived from the following theorem.
7.3 Theorem: λ(A) ∈ (DN) if, and only if, the following holds
∃p ∀k ∃K, C ∀j : a2j,k ≤ C aj,p aj,K .
97
Proof: To get the necessity of the condition we apply the inequality in the
definition of (DN) to the canonical basis vectors ej . Sufficiency follows from
the Cauchy-Schwarz inequality:
∑
∥x∥2k = |xj |2 a2j,k
j
∑
≤ C |xj |aj,p |xj |aj,K
j
≤ C∥x∥p ∥x∥K . 2
Remark: In the previous theorem ∥ ∥p is a dominating norm.
If λ(A) ∈ (DN), we find a dominating norm ∥ ∥p and set

aj,k+p
bj,k := .
aj,p
Then y 7→ (xj aj,p )j defines an isomorphism from λ(A) onto λ(B). For B we
have bj,0 = 1 for all j, ∥ ∥0 is a dominating norm on λ(B) and the condition
in Theorem 7.3 takes the form
∀k ∃K, C ∀j : b2j,k ≤ Cbj,K .
Therefore the multipliers of λ(B)′ are easily seen to be the set of all sequences
|tj |
t such that there is k with supj bj,k < +∞. By Theorem 5.6 this yields:
7.4 Theorem: If λ(A) ∈ (DN) is a regular Schwartz space and B as above

then
|tj |
∆(λ(A)) = {t | there is k such that sup < +∞}.
j bj,k
7.5 Corollary: If λ(A) and λ(Ã) are regular Schwartz spaces with property
(DN) then ∆(λ(A)) = ∆(λ(Ã)) implies that λ(A) and λ(Ã) are isomorphic
by a diagonal transformation.
Here diagonal transformation means a transformation whose matrix is diag-

onal, i.e. of the form x = (xj )j∈N 7→ (dj xj )j∈N .
Proof: Let B and B̃ be as above then for every l we have (b̃j,l )j∈N ∈
∆(λ(B̃)) = ∆(λ(B)) which means the existence of k and C such that
b̃j,l ≤ Cbj,k .
98
In the same way we obtain, k̃, C̃ such that
bj,l ≤ C̃bj,k̃ .
Therefore λ(B) = λ(B̃) which implies the assertion, since λ(B) and λ(B̃)
are isomorphic to λ(A) and λ(Ã), respectively, by means of diagonal trans-
formations. 2
References
[1] R. Meise, D. Vogt, Introduction to Functional Analysis, Clarendon

Press, Oxford 1997.
99
Index
ℓ1 (I, E), 84 Hilbert-Schmidt operator, 21
ℓ1 [I, E], 82
ℓ1 {I, X}, 72 Köthe-matrix, 44
Kolmogoroff diameter, 34
absolutely summing, 86
absorbant, 4 linking maps, 5
associated nuclear topology, 27
Minkowski functional, 2
Banach space, 7
nearly open, 4
approximation property, 66
norm
bounded approximation proper-
on L1 (µ, X) and L1 (µ) ⊗ X, 74
ty, 77
nuclear, 19
local, 5
(α, ∞), (α, 0), 55
Banach’s Isomorphy Theorem, 4
map, 23
barrel, 4
barrelled, 4 Open Mapping Theorem, 4
Bipolar theorem, 7
bornivorous, 10 Pel-Dec, 29
bornological, 10 polar set, 6
bounded, 8 power series space of type
finite,infinite, 48
canonical resolution, 5
Closed Graph Theorem, 5 regular, 45
dimension Schauder, 4, 19
diametral, 40 Schwartz distributions, 94
dominating norm, 96 Schwartz space, 15
sequence
Fréchet space, 2 rapidly decreasing, 27
distinguished, 12 space
reflexive, 13 locally convex, 2
Fréchet-Hilbert space, 14 stable, 55
Fréchet-Schwartz space, 15 shift, 55
fundamental system
of bounded sets, 10 tensor product
of neighborhoods of zero, 40 ε-tensor product, 79
completition, 80
Gelfand topology, 80
n-th diameter, 41
100
π-tensor product
complete, 62
topology, 61
algebraic, 60
tensornorm, 62
Theorem of
Alaoglu-Bourbaki, 9
Schauder, 19
T. and Y. Komura, 28
101

Lectures On FR Echet Spaces: Dietmar Vogt

Uploaded by

Document Informationclick to expand document information

Copyright:

Available Formats

Lectures On FR Echet Spaces: Dietmar Vogt

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Lectures On FR Echet Spaces: Dietmar Vogt

Uploaded by

Copyright:

Available Formats

Lectures on Fréchet spaces

Bergische Universität Wuppertal

1 Definition and basic properties 2

2 The dual space of a Fréchet space 6

3 Schwartz spaces and nuclear spaces 14

5 Power series spaces and their related nuclearities 44

6 Tensor products and nuclearity 60

All linear spaces will be over the scalar ﬁeld K = C oder R.

We recall that a sequence (xn )n∈N in a topological vector space is a Cauchy

If E is a vector space and A ⊂ E absolutely convex then the Minkowski

∥x∥A = inf{ t > 0 | x ∈ tA }.

is the largest linear space contained in A.

If E is a topological vector space and A has an interior point, then ∥ ∥A is

A space is called locally convex if it has a basis of absolutely convex neigh-

Definition: A set P of continuous seminorms on the locally convex space

Of course, we have for every generalized sequence (xτ )τ ∈T that xτ −→ x if,

If a fundamental system of seminorms P is countable then we may assume

(2) E has a countable basis of neighborhoods of zero.

(3) E has a countable fundamental system of seminorms.

(4) The topology of E can be given by a translation invariant metric.

Proof: (1) ⇒ (2) and (4) ⇒ (1) are obvious.

It is an elementary exercise to show that d(·, ·) is a translation invariant

1.2 Theorem: If E is a Fréchet space and F ⊂ E a closed subspace then

1.3 Theorem: Every Fréchet space is barrelled.

1.4 Lemma: If E and F are locally convex, F a Fréchet space and A : E −→

Proof: We may assume that U is absolutely convex, hence AU is a barrel in

1.5 Lemma (Schauder): If E and F are metrizable spaces, E complete,

These two lemmas yield:

1.6 Theorem (Open Mapping Theorem): If E and F are Fréchet spaces,

1.7 Corollary (Banach’s Isomorphy Theorem): If E and F are Fréchet

Proof: We consider the maps π1 : (x, y) 7→ x, π2 : (x, y) 7→ y from Graph A

If E is a Fréchet space, ∥ ∥1 ≤ ∥ ∥2 ≤ . . . a fundamental system of seminorms,

quotient map. We have for m > n that ınm ◦ ım = ın . We deﬁne continuous

2 The dual space of a Fréchet space

If E ′ is equipped with a locally convex topology (e.g. the strong topology,

If p is a continuous seminorm on the locally convex space E and U = { x ∈

If E is a Fréchet space and

a fundamental system of seminorms, then

pB (y) = sup |y(x)|

where B runs through the bounded subsets of E.

We recall that a subset of a locally convex space is bounded if it is bounded

2.2 Theorem: Let E be metrizable locally convex and F locally convex.

Proof: If xn −→ 0 then ∥xn ∥k −→ 0 for all k, where ∥ ∥1 ≤ ∥ ∥2 ≤ . . . is

∥λn xn ∥k = |λn |∥xn ∥k −→ 0

for all k. In particular B = { λn xn | n ∈ N } is bounded. Therefore A(B)

So Axn −→ 0 and A is continuous. 2

This holds, of course, in particular for F = K. That means, a linear form

2.3 Theorem: If E is metrizable, then E ′ is complete.

Proof: If (yτ )τ ∈T is a Cauchy net in E ′ , then (yτ (x))τ ∈T is a Cauchy net

Definition: If E is locally convex then:

(1) the weak topology σ(E, E ′ ) on E is given by the fundamental system

where e runs through the ﬁnite subsets of E ′

(2) the weak∗ topology σ(E ′ , E) on E ′ is given by the fundamental system

where e runs through the ﬁnite subsets of E.

2.4 Theorem (Alaoglu-Bourbaki): If E is locally convex and U ⊂ E a

For the proof see [1, 23.5].

2.5 Theorem: If E is a Fréchet space then for M ⊂ E ′ the following are

(1) M is weak∗ bounded