NUCLEAR SPACE FACTS, STRANGE AND PLAIN

We present a scenic but practical guide through nuclear spaces and their dual spaces, examining useful, unexpected, and often unfamiliar results both for nuclear spaces and their strong and weak duals.


Introduction
The subject of nuclear spaces saw heavy activity in the 1960s and 1970s but, despite the publication of some deep and voluminous books on the subject, the field largely disappeared from mainstream functional analysis. However, nuclear spaces serve as a practical and useful setting in stochastic analysis. The books on the subject, with the exception of Gel'fand et al., tend to be heavy going on general theory. Our purpose in the present paper is to take a very concrete approach, firmly avoiding generality and focusing on concrete questions; in keeping with this spirit, the terms 'bornological' and 'barrelled' will not appear anywhere outside of this sentence! More regrettably, we do not explore tensor products. The guiding principle is to stay close to questions which arise in applications of the general theory to specific settings closer to hard analysis.

Topological Vector Spaces
By a topological vector space we shall mean a real or complex vector space, equipped with a Hausdorff topology for the which the operations of addition and multiplication are continuous. A general open set in such a space is the union of translates of neighborhoods of 0.
A topological vector space X is said to be locally convex if every neighborhood of 0 contains a neighborhood of 0 which is a convex set. By a locally convex space we shall mean a locally convex topological vector space.
At first sight, a locally convex space appears ethereal, with little structure. Yet, remarkable, any such space can be understood in terms of normed linear spaces, as we explain below.
If U is a convex neighborhood of 0, inside it there is a convex neighborhood of 0 which is also balanced, i.e. mapped into itself under multiplications by scalars of magnitude ≤ 1 (see [3,Theorem 1.14(b)]).
For a convex balanced neighborhood of 0, the function is a semi-norm in the sense that for all a, b ∈ X and scalars t. The function ρ U is uniformly continuous, because , which is less than any given positive if a − b ∈ U .
Each seminorm ρ U generates a topology τ U on X, consisting of all translates of scalar multiples of the unit ball U = {v ∈ X : ρ U (v) < 1}. The topology on X is in fact the smallest topology which contains τ U , for all convex balanced neighborhoods U of 0.
Warning: The topology τ U is Hausdorff if and only if ρ U is actually a norm, i.e. U does not contain any full lines! Suppose X is a locally convex space, and U and V are convex balanced neighborhoods of 0 with U ⊂ V ; using a smaller neighborhood of 0 results in larger magnitudes for vectors: For U ⊂ V , the τ V -open neighborhood rV of 0 contains the τ U -open neighborhood rU of 0, for every r > 0, and so every τ V -open set is the union of translates of suitable balls rU . Thus, using a smaller neighborhood of 0 results in a more refined topology: If the convex balanced neighborhood U of 0 is 'finite' in every direction, i.e. 0 is the only vector for which all multiples lie inside U , then ρ U is a norm, i.e. ρ U is 0 only on the zero vector. Indeed, in this case the topology arises from ρ U in the sense that every open set is the union of translates of scalar multiples of the unit ball U = {v ∈ X : ρ U (v) < 1}.
On the other hand, if U is 'infinite in some direction', i.e. it contains all multiples of some non-zero vector, then ρ U is 0 on any vector in that direction. As we shall see later, in many naturally occurring spaces, open neighborhoods of 0 contain infinite-dimensional subspaces, and so ρ U can be 0 on large subspaces. Specifically, ρ −1 U (0) is the subspace consisting of all vectors all of whose multiples lie inside U . Nullifying this subspace yields the quotient space on which ρ U , of course, induces a genuine norm, not just a semi-norm. Note that ρ U does indeed pass to a well-defined function on the quotient If v is a non-zero vector in the locally convex space then (recalling that X is, by definition, Hausdorff) choosing a convex balanced neighborhood U of 0 for which v / ∈ U , we see that v survives to a non-zero vector in the quotient space X U under the quotient map If for a vector v, the projection p U (v) is 0 then v ∈ ρ −1 U (0) and so every multiple of v lies in U and, in particular, v itself lies in U ; thus, Here we see a phenomenon that is very infinite-dimensional in flavor: every τ U -neighborhood of 0 containing an entire non-zero subspace. One can think of p U (x) as a rough image of x, the roughness resulting from the fact that some information about x lying in ker , and so p V (x) ought to be an even rougher image than p U (x), since more information about x is discarded. The 'full' picture for x is presumably contained in the collection of all the rough renditions p U (x). This is made precise in the notion of projective limit below in Theorems 2.1 and 2.2.
Thus, effectively, the space X is encoded in the family of normed linear spaces X U , through the quotient maps p U .
Observe that if a sequence (x n ) n≥1 in a locally convex space X converges then it converges with respect to the topology τ U , equivalently the semi-norm ρ U , for every balanced convex neighborhood U of 0. This is simply because the topology τ U is contained in the full topology τ of X. For the converse, assume that the topology τ of X has a local base N consisting of convex, balanced neighborhoods U of 0 for which ρ U is a norm, i.e. that every neighborhood of 0 contains as subset some neighborhood in N . Suppose (x n ) n is a sequence which is convergent with respect to τ U , for every U ∈ N , and let x U denote the limit, which is unique since ρ U is a norm. Working with two such neighborhoods U 1 and U 2 of 0, and noting that the topology τ U 1 ∩U 2 contains all the open sets in both τ U 1 and τ U 2 , it follows that x U is, in fact, independent of the choice of U . Let x be the common value of all the x U . Now if V is a neighborhood of 0 in τ then it contains as subset some U ∈ N , and so x n − x ∈ U for large n, and hence x n − x ∈ V for large n. Thus, x n → x with respect to the topology τ .
A sequence (x n ) n≥1 in a topological vector space X is said to be Cauchy with respect to a topology τ if the differences x n − x m are all eventually in any given τ -neighborhood of 0 for large n and m. Any convergent sequence is automatically Cauchy. A topological vector space is said to be complete if every Cauchy sequence converges.
In a locally convex space a sequence is Cauchy if and only if it is Cauchy with respect to the topology τ U , for every convex balanced neighborhood U of 0, and this is equivalent to the sequence being Cauchy with respect to each semi-metric d U , specified by (2.9) If U and V are convex balanced neighborhoods of 0 and U ⊂ V then the identity map X → X induces a natural surjection p V U on the quotients: it is the map for all x ∈ X. This takes a picture p U (x) of x and produces a coarses picture p V (x). Note that p V U is a linear contraction mapping. If a locally convex space X is metrizable, then (Rudin [3, Theorem 1.24]) its topology is induced by a translation-invariant metric d for which open balls are convex, and so every neighborhood of 0 contains a locally convex neighborhood W (an open ball, for instance) for which ρ W is a norm (of course, the topology τ W induced by ρ W is, in general, smaller than τ ); if V is a convex balanced neighborhood of 0, with V ⊂ W , then ρ V is also a norm.
Theorem 2.1. Let X be a complete, metrizable, locally convex space. Let B be a set of all convex, balanced neighborhoods of 0 such that every neighborhood of 0 contains as subset some neighborhood in B, i.e. suppose B is a local base at 0, consisting of convex balanced sets, for the topology of X. Suppose now that associated to each V ∈ B an element for every W ∈ B for which W ⊃ V . Then there exists a unique x ∈ X such that p W (x) = x W for every W ∈ B. Moreover, for such W , the quotient X W is a Banach space.
Proof. Let d be a translation-invariant metric on X which induced the topology τ of X and for which all open balls are convex. Choose W n ∈ B lying inside the open ball of d-radius 1/n centered at 0, and then let for all n ∈ {1, 2, ...}. As before, let p U : X → X U = X/ρ −1 (0) be the canonical projection. Given x V ∈ X V , for each U ∈ B, satisfying the consistency condition (2.11), choose x n ∈ X for which p Vn (x n ) = x Vn , for every n ∈ {1, 2, 3, ...}. We claim that (x n ) n≥1 is a Cauchy sequence: if W is any neighborhood of 0, choose k ∈ {1, 2, 3, ...} such that V k ⊂ W ; then for any n, m ≥ k we have V k ⊃ V n , V m , and so (2.12) Hence x n − x m is in V k , and therefore also in W . Thus, (x n ) n≥1 is Cauchy in X. Let x be the limit of this sequence. Then (2.13) Now, for any convex balanced neighborhood W ∈ B, choose k such that V k ⊂ W ; then showing that x has the desired properties. Let us check that x is unique. To this end suppose x ∈ X satisfies p W (x) = 0 for every W ∈ B, i.e. x lies in every such neighborhood. But then x lies in every open ball centered at 0 relative to the metric d, and so x must be 0.
The completeness of X W follows from (a somewhat overlooked consequence of) the open mapping theorem, see Rudin [3, Theorem 2.11(iii)], which guarantees that the continuous linear image of a complete metrizable space inside a topological vector space is complete.
Finally, we can prove the fundamental projective limit nature of a complete, metrizable, locally convex space.
be the usual projection. Let B be a set of all convex, balanced neighborhoods of 0 such that every neighborhood of 0 contains as subset some neighborhood in B, i.e. suppose B is a local base at 0, consisting of convex balanced sets, for the topology of X. Assume Y is a topological vector space and, for each V ∈ B there is a continuous Then there is a unique continuous linear map f : Y → X such that Proof. From the previous theorem, we have, for any y ∈ Y , the existence of a unique f (y) ∈ X for which p V f (y) = f V (y) for every V ∈ B.
The uniqueness of f (y) for each y ∈ Y also implies that f is linear. It remains to show that f is continuous.
contains an open neighborhood of 0. Thus, f is continuous.
Thus, a complete, metrizable, locally convex space is obtainable as a projective limit of Banach spaces. In the next section we turn to nuclear spaces, where this specialization is carried just a bit further to Hilbert spaces.
For a topological vector space X over the field F (which, for us, is always R or C), the dual space X is vector space of all continuous linear functionals on X. If X is locally convex then the Hahn-Banach theorem guarantees that X = {0}. There are several topologies of interest on X . For now let us note the two extreme ones: • The weak topology on X is the smallest topology on X for which the evaluation map is continuous for all x ∈ X. This topology consists of all unions of translates of sets of the form with D running over finite subsets of X, and over (0, ∞). • The strong topology on X consists of all unions of translates of sets of the form with D running over all bounded subsets of X and over (0, ∞).

The nuclear space structure
We work with a real complex topological vector space H equipped with the following structure. There is a sequence of inner-products ·, · p , for The completion of H in the norm || · || 0 is denoted H 0 , and inside this Hilbert space the completion of H with respect to || · || p is a dense subspace denoted H p . We assume that H 0 is separable, and that H is the intersection of all the spaces H p . Thus, Furthermore, we assume that each inclusion H p+1 → H p is a Hilbert-Schmidt operator, i.e. there is an orthonormal basis The topology on H is the projective limit topology from the inclusions H → H p , i.e. it is induced by the norms || · || p . Thus an open set in this topology is the union of || · || p -balls with p running over {0, 1, 2, 3, ...}.
All these assumptions make H a nuclear space. An infinite-dimensional topological vector space is never locally compact, but a nuclear space is an excellent substitute in the infinite-dimensional case: the definition ensures, in particular, that an open ball in H p+1 has compact closure in H p .
Note that if τ p is the topology on H given by || · || p , then by (3.1), the identity map (H, τ p+1 ) → (H, τ p ) is continuous, for p ∈ {0, 1, 2, ...}, and so The inclusions here are strict if H is infinite-dimensional, because of the Hilbert-Schmidt assumption made above. A linear functional on H is continuous with respect to || · || p if and only if it extends to a (unique) continuous linear functional on H p . Thus we may and will make the identification where the left side is the vector space of all linear functionals on H which are continuous with respect to the norm ||·|| p . Virtually always, the space H p will be equipped with the strong topology, i.e. the one which arises from identification of H p with H p by associating to each x ∈ H p the linear functional ·, x p . Then H p is a Hilbert space, isometrically isomorphic to H p . We use the notation for p ∈ {0, 1, 2, ...}, and, of course, this continues to be meaningful as an identification even for negative integers p. We have the norm The topological dual of H is In addition to the weak and strong topologies, there is also the inductive limit topology on H , which is the largest locally convex topology for which the inclusions H p → H are all continuous. For a proof see [1].
3.1. Hilbert-Schmidt facts. We will use Hilbert-Schmidt operators, and recall here some standard facts about them. For a linear operator T : H → K between Hilbert spaces, define the Hilbert-Schmidt norm to be where B is some orthonormal basis of H; if H = 0 the Hilbert-Schmidt norm of T is, by definition, 0. If ||T || HS < ∞ then T is said to be a Hilbert-Schmidt operator. Proof. . These results are standard, but we prove (iii) for entertainment and because we will need the ideaa in the proof later. Let D(R) be the closed ball of radius R ∈ (0, ∞) in H. We will show that T maps D(R) to a compact subset of K. Since K is a complete metric space, compactness of a set C is equivalent to it being closed and totally bounded (which means that for any > 0, the set C can be covered by a finite collection of balls of radius ). Let > 0. Choose an orthonormal basis B in H (if H = 0 the result is trivial, so we assume H = 0). Since u∈B ||T u|| 2 K < ∞, there is a non-empty finite set B f ⊂ B such that Let L f be the linear span of B f , and D(R) f the orthogonal projection of D(R) onto L f . Then D(R) f is compact, being the closed ball of finite radius R with respect to the norm on the finite-dimensional space L f induced by || · || H , and so D(R) f is also totally bounded. Therefore, there is a non-empty finite subset F of D(R) f such that for any x ∈ D(R), there is some y F ∈ F whose distance from (3.10) Hence the image of D(R) under T is totally bounded. Next we check that T maps D(R) onto a closed set in K. Consider a sequence x 1 , x 2 , ... ∈ D(R) such that (T x n ) n≥1 converges to some y ∈ K. The image of D(R) under T is bounded and hence contained inside some closed ball of radius R ∈ (0, ∞) in K, and since a closed ball in a normed linear space is weakly compact (by the Banach-Alaoglu theorem) there is a subsequence (x n k ) such that (T x n k ) converges weakly to some y ∈ K. Then, for any w ∈ K, we have (3.11) Thus, the weak limit y of (T x n k ) is T x. Since x necessarily lies in D(R), we conclude that y is in the image of D(R) under T , and so T D(R) is a closed subset of K. Being closed and totally bounded, it is compact.
Some unexpected facts about nuclear spaces are connected with the following observation: Then there is a sequence of points in H which does not converge in H but whose image by j in K is convergent.
Proof. An orthonormal sequence e 1 , e 2 , ... in H is not Cauchy and hence not convergent, but the sequence j(e n ) n≥1 converges to 0 in K because ∞ n=1 ||j(e n )|| 2 K is convergent.

Balls and Cubes
In this section, as usual, p denotes an element of {0, 1, 2, 3, ...}. We work, as before, with the space H, equipped with inner-products ·, · p , with the Hilbert-Schmidt condition explained in the context of (3.2).   Proof. Let v 1 , v 2 , ... be an orthonormal basis of H p+1 lying inside H (choose a maximal ·, · p -orthonormal set in H; this is necessarily also maximal in H p+1 , and hence an orthonormal basis of H p ). Then, since the inclusion H p+1 → H p is Hilbert-Schmidt, the sum ∞ n=1 ||v n || 2 p is finite, and so the lengths ||v n || p tend to 0. Thus the vectors which are all in B p (R), have ·, · p+1 -norm going to ∞. In particular, B p (R)∩H is not bounded in H p+1 , and hence also not bounded in H.
The next observation provides some compact sets in H. A subset D of a topological vector space X is said to be bounded if for any neighborhood W of 0 in X, the set D is contained in tW for all scalars t with large enough magnitude. Thus, for instance any open or closed ball in a normed linear space is bounded. Proof. . A closed ball D p (r) ∩ H is closed in the topology τ p ⊂ τ , and so it is a closed subset of H. If W is a neighborhood of 0 in H, then there is a q ∈ {0, 1, 2, ..} and an r ∈ (0, ∞) such that B q (r) ⊂ W ; hence, ∩ p≥0 D p (r p ) ⊂ tB q (r) ⊂ tW if |t| > r q /r. Thus, ∩ p≥0 D p (r p ) is closed and bounded and hence, by the Heine-Borel property proved below in Fact 5.3, it is compact in H We turn to balls in the dual spaces H p and H . Unless stated otherwise, we equip each Hilbert space dual H p H p with the strong topology, which is the same as the Hilbert-space topology.  Proof. . This is because the inclusion H → H p is continuous. The inclusion j : H q → H p induces j * : H p → H q which is also continuous and so (j  Proof. . The inclusion map H p+1 → H p being Hilbert-Schimdt, so is the adjoint inclusion H p → H p+1 , and so D −p (R) is compact inside H p+1 with the strong (Hilbert-space) topology. By continuity of the inclusion map H q → H it follows that D −p (R) is compact in H q with respect to the strong topology.
Continuity of the inclusion map H p+1 → H also implies that D −p (R) is compact in H with respect to the strong topology.
Since D −p (R) is compact in the metric space H q it is sequentially compact in H q , i.e. any sequence on D −p (R) has a subsequence which is convergent in H q . From continuity of the inclusion H q → H it follows that such a subsequence also converges in H . Thus, D −p (R) is strongly, and hence also weakly, sequentially compact in H .

Cubes.
Fact 4.8. If H is a separable Hilbert space then the closed unit cube, relative to any orthonormal basis, in the dual H * is not compact. This is in contrast to the closed unit ball in H * .
Proof. . The closed unit cube in H * relative to an orthonormal basis e 1 , e 2 , ... is and ∞ n=1 λ −2 n < ∞. Let C be the closed cube in H given by C = {x ∈ H : | x , e n | ≤ λ n n for all n ∈ {1, 2, 3, ...}} Then C contains strong neighborhood of 0 and hence is not compact. However, C ∩ F is compact for every finite-dimensional subspace of H 0 .
Hence, ||v || 2 0 ≤ R 2 1 + · · · + R 2 d < ∞. This shows that C ∩ F is a bounded subset of F . Since C is weakly closed and so is any finite-dimensional subspace, C ∩ F is weakly closed in H and hence weakly closed in the induced topology on F ; but on the finite-dimensional subspace F there is only one topological vector space structure, and so C ∩ F is closed. Being closed and bounded in F , it is compact.
It remains to prove the algebraic statement assumed earlier. Suppose that the algebraic linear span of the functionalsê n on F is not the entire dual F * of F ; then they span a proper subspace of F * and hence there is a non-zero vector f ∈ F on which they all vanish, but this would then be a vector in H 0 whose inner-product with every e n is 0 and hence f would have to be 0. Thus the algebraic linear span of the functionalŝ e n on F is the entire dual F * , and so, in particular, any element of F * is a finite linear combination of the functionals the functionalsê n . This proves that the assumption made earlier is correct.

Facts about the Nuclear Space Topology
We work with an infinite-dimensional nuclear space H with structure as detailed in the context of (3.1). While there are some treacherous features, such as the one in Fact 5.1, a nuclear space has some very convenient properties which make them almost as good as finite-dimensional spaces. One consequence of the preceding observation is that a Banach space if not a nuclear space. For the following, recall that in a locally convex space, a subset B if said to be bounded if for any U , the set B lies inside some multiple tU of U . If f : X → Y is a continuous linear map between topological vector spaces and B is a bounded subset of X then f (B) is a bounded sunbset of Y , for if V is a neighborhood of 0 in Y then B ⊂ tf −1 (V ) = f −1 (tV ) for some scalar t and hence f (B) ⊂ tV . Proof. . Let B be a closed and bounded subset of H. Take any p ∈ {0, 1, 2, ..}. Since the inclusion H → H p+1 is continuous, B is bounded in H p+1 . Since the inclusion H p+1 → H p is Hilbert-Schmidt it follows, as in the proof of Fact 4.7, that B, as a subset of H p , is contained in a compact set. Let (x n ) n≥1 be a sequence of points in B. Then it follows that there is a subsequence (x (p) ) n≥1 which converges in H p . Applying the Cantor process of extracting repeated subsequences, there is a subsequence (x ) n≥1 which is convergent in H p for every p ∈ {0, 1, 2, ...}. By contunity of the inclusions H p+1 → H p it follows that the limit is the same for every p, i.e. there is a point y ∈ ∩ p≥0 H p = H, such that the subsequence (x n ) converges to y in all the H p . If V is a neighborhood of 0 in H then there is a p ∈ {0, 1, 2, ...} and an > 0 such that the ball B H p ( ) lies inside {x ∈ H : ||x|| p < } ⊂ V . Since x n → y in H p we have x n − y ∈ B H p ( ) ⊂ V for large n, and so x n → y in H. Thus, a closed and bounded subset of H is sequentially compact; since H is metrizable, sequential compactness is equivalent to compactness for any subset of H.

Facts about the Dual Topologies
We turn now to the dual of an infinite-dimensional nuclear space H, and continue with the notation explained earlier in the context of (3.1).
Let us keep in mind that thanks to the Hahn-Banach theorem, the dual of a non-trivial locally convex space is again non-trivial. Indeed, the dual of an infinite-dimensional locally convex space is infinite-dimensional.
If a topology is metrizable then it has a countable local base: any neighborhood U of any point p ∈ X contains a neighborhood of the form p + B(0; 1/n) for large n.
The following facts are proved in Rudin [ x ∈ X then f is in X .
Fact 6.1. If X is a complete, metrizable topological vector space, then the dual space X is complete with respect to both strong and weak topologies.
Proof. . Let A be a non-empty collection of subsets of X, closed under finite unions. Let τ A be the topology on X whose open sets are unions of translates of sets of the form with A running over A and over (0, ∞). If A is the set of all finite subsets of X then τ A is the weak topology; if A is the set of all bounded subsets of X then τ A is the strong topology. Suppose (f n ) n≥1 is Cauchy sequence in X for the topology τ A . This means that, for any A ∈ A, the sequence of functions f n |A is uniformly Cauchy, and hence uniformly convergent, for every A ∈ A. Let f (x) = lim n→∞ f n (x). By Banach-Steinhaus, f is continuous. A τ Aneighborhood U of f contains a set of the form f + B(A; ), and the uniform convergence f n → f on A implies that f n − f ∈ B(A; ) for large n, and so f n − f ∈ U for large n. Thus, f n → f in τ A . Proof. . If Y is a vector space over a field F , andx 1 , ...,x N : Y → F are linear functionals, then the mapping has kernel infinite-dimensional if Y is infinite-dimensional. We apply this to Y = X . A weakly open neighborhood of 0 in X contains contains a set of the form for some x 1 , ..., x N ∈ X, and open neighborhoods W 1 , ..., W N of 0 in ths scalars, and so B contains the ker(x 1 , ...,x N ).
Fact 6.3. The weak dual of an infinite-dimensional Banach space is not metrizable.
Proof. . Let X be a Banach space. Since it is normed, there is a bounded neigborhood D 1 of 0 (say the unit ball). For x ∈ X, letx be the linear functional on X which is evaluation at x. Let D 1 be the unit ball in X , i.e. the set of all f ∈ X for which sup x∈D 1 |f (x)| ≤ 1. Thus, D 1 is the intersection of the setsx −1 (W 1 ), where W 1 is the closed unit ball in the scalars, with x running over B 1 . Hence, B 1 is weakly closed. Observe that ∪ n≥1 nB 1 is all of X , for if f ∈ X then f (D 1 ) is a bounded set of scalars and so sup x∈D 1 |f (x)| ≤ n, i.e. f ∈ nDB 1 , for some n ∈ {1, 2, ...}. Now X , with the weak topology, is complete; if it were metrizable, then the Baire category theorem would imply that X could not be the union of a countable collection of nowhere dense subsets. Thus it will suffice to show that D 1 has empty interior in the weak topology. If, to the contrary, B 1 contains a nonempty weakly open subset, then D 1 −D 1 contains a weakly open neighborhood of 0, and hence an infinite-dimensional subspace of X . Now D 1 − D 1 ⊂ 2D 1 , and so it would follow that 2D 1 , and hence D 1 itself, contains an infinite-dimensional subspace. In particular, there is a non-zero g ∈ X such that all scalar multiples of g belong to D 1 ; but this is impossible from the definition of D 1 .
In a similar vein there is the following negative result: Fact 6.4. The weak topology and the strong topology of the dual of an infinite-dimensional nuclear space are not metrizable.
Proof. Let X be a nuclear space, with topology generated by norms || · || p for p ∈ {0, 1, 2, ...}. Let D p (R) be the subset of X given by the closed ball of radius R, center 0, for the |! · || p -norm, and let D p (R) be the set of all linear functionals f on X which map D p (1) into scalars of magnitude ≤ R. Each set D p (R), being the intersection of weakly closed sets, is a weakly closed set, and hence also strongly closed, in X . The union of the sets D p (n) with p ∈ {0, 1, 2, ...} and n ∈ {1, 2, ...} is all of X . Now D p (n) lies in the proper subspace X p of X , and so has empty interior in any topology on X which makes X a topological vector space. Thus, X , which is a complete topological vector space with respect to both weak and strong topologies, is the countable union of nowhere dense sets, and hence the weak and strong topologies on X are not metrizable.
In a sense, sequences do not detect the difference between the weak and strong topologies on the dual space: Fact 6.5. A sequence in the dual of a nuclear space is weakly convergent if and only if it is strongly convergent.
Proof. Since the weak topology is contained in the strong topology, strong convergence implies weak convergence. For the converse, let (x n ) n≥1 be a sequence in the dual X of a nuclear space X, converging weakly to x ∈ X . By Banach-Steinhaus, {x 1 , x 2 , x 3 , ...} is a uniformly continuous set of functions on X and hence the convergence x n → x is uniform on compact subsets of X. The closure D of the bounded set D is bounded (see Rudin [3,Theorem 1.13(f)]) and hence also compact by the Heine-Borel property for nuclear spaces Fact 5.3. So the sequence (x n ) ≥1 is uniformly convergent on D, and hence also on D, and therefore it is strongly convergent as a sequence in X .

Continuous Functions
Linear functionals automatically stand a better chance of having continuity properties in locally convex spaces because they map convex sets to convex sets. In this section we look at some examples of nonlinear functions.
Fact 7.1. If X is an infinite-dimensional topological vector space then the only continuous function on X having compact support is 0.
Proof. This is because any compact set in X has empty interior, since X, being infinite-dimensional, is not locally compact.
Fact 7.2. There is a weakly continuous function S on H which satisfies 0 < S ≤ 1, and S equals 1 exactly on D −1 (R).
Fact 7.3. For the cube C in Fact 4.9, with assumptions as stated in Fact 4.9, the function f given on H by f (x) = inf n≥1 {λ n n − |x n |} for x ∈ C 0 for x / ∈ C (7.1) is sequentially continuous, where x n = x, e n .
Proof. We can consider this as two functions g : C → R with g(x) = inf λ n n − |x n | and h : D → R with h(x) = 0 where D = {x ∈ H : | x , e n | ≥ λ n n for all n ∈ {1, 2, 3, ...}}. Since g and h agree on the boundary (their intersection) proving the continuity of them separately would give us continuity of f . Thus, f is sequentially continuous.