Locally Convex, Nuclear, and Rigged Hilbert Spaces

A Hilbert norm controls size, but it need not control point values, derivatives, decay, or support. Giving a dense test space a finer locally convex topology makes those operations continuous and makes convergence more demanding. Its antidual contains the restricted Hilbert functionals and can be strictly larger, as the Schwartz-space example below shows. Nuclearity then supplies a summability condition strong enough to remove the usual ambiguity between completed tensor products. That is the mechanism behind the kernel theorem.

A rigged Hilbert space organizes the topology and duality in

$$\Phi\subset\mathcal H\subset\Phi^\times,$$

where $\Phi^\times$ is a declared continuous antidual. It provides a mathematically meaningful home for objects such as momentum kets that are not vectors in $\mathcal H$. When $\Phi$ is nuclear, the same test space also supports the tensor-product results below. The rigging alone does not produce a complete generalized eigenvector expansion: self-adjointness, invariance and continuity on $\Phi$, nuclearity, and a nuclear spectral theorem are separate hypotheses.

Required background. Banach and Hilbert Spaces, Completion, and Riesz Representation supplies Hilbert completion and duality; Test-Function Spaces, Distributions, Support, and Convergence supplies test-space topology and distributional pairings.

Seminorm topologies and nuclear riggings

This page develops the reusable mathematics of seminorm topologies, nuclear spaces, completed tensor products, distributional kernels, and nuclear riggings. It does not develop the physical axioms of a quantum field, a general spectral-theorem framework, or production software.

A topology made from seminorms

Let $E$ be a complex vector space. A seminorm $p$ obeys $p(\lambda x)=|\lambda|p(x)$ and the triangle inequality, but it may vanish on a nonzero vector. A family $\mathcal P=\{p_i\}_{i\in I}$ defines neighborhoods of zero by finite intersections

$$U(i_1,\ldots,i_r;\epsilon) = \left\{ x\in E: \max_{1\leq j\leq r}p_{i_j}(x)<\epsilon \right\}.$$

The resulting vector topology is locally convex. It is Hausdorff exactly when the family separates points: for every $x\ne0$, some $p_i(x)\ne0$. The structure is $(E,\tau_{\mathcal P})$, rather than the underlying vector space alone; different seminorm families may define the same topology.

A net $x_\alpha$ converges to $x$ precisely when $p_i(x_\alpha-x)\to0$ for every $i$. If a countable family defines the topology, the space is pseudometrizable. If that family also separates points, the space is metrizable; a complete metrizable locally convex space is a Fréchet space. For locally convex spaces $E$ and $F$, a linear map $T:E\to F$ is continuous exactly when, for every defining seminorm $q$ of $F$, there are $p_{i_1},\ldots,p_{i_r}$ and $C>0$ such that

$$q(Tx) \leq C\max_{1\leq j\leq r}p_{i_j}(x).$$

This criterion turns continuity into an estimate. It is also why the direction of refinement matters: a finer topology has fewer convergent nets, but more linear functionals may be continuous on it.

Three standard spaces must be kept distinct.

• $C^\infty(\Omega)$, with uniform control of every derivative on each compact subset of $\Omega$, is a nuclear Fréchet space.

• The Schwartz space $\mathcal S(\mathbb R^d)$ is a nuclear Fréchet space with seminorms

  $$p_{\alpha\beta}(f) = \sup_{x\in\mathbb R^d} \left|x^\alpha\partial^\beta f(x)\right|.$$

• $\mathcal D(\Omega)=C_c^\infty(\Omega)$ has its nuclear LF topology, obtained as an inductive limit of the Fréchet spaces of test functions supported in fixed compact sets. It is not one global Fréchet space.

These classification and continuity facts are developed in Trèves 1967, Chapters 7, 10, 13, 50, and 51.

Why the Hilbert topology is too coarse

On $\mathcal S(\mathbb R)$, evaluation at the origin is continuous because

$$|f(0)|\leq p_{00}(f).$$

It therefore defines the tempered distribution $\delta_0(f)=f(0)$. The same rule is not a bounded functional for the $L^2$ norm. Indeed,

$$f_N(x)=e^{-Nx^2}, \qquad \|f_N\|_2 = \left(\frac{\pi}{2N}\right)^{1/4} \longrightarrow0, \qquad f_N(0)=1.$$

Moreover, an $L^2$ vector is an equivalence class modulo sets of measure zero, so point evaluation is not intrinsically defined on all of $L^2$. The finer Schwartz topology has not turned $\delta_0$ into a Hilbert vector; it has made evaluation continuous on the chosen test space.

Nuclearity is a summability property of the topology

First consider Banach spaces $X$ and $Y$. A linear map $T:X\to Y$ is nuclear if it has a rank-one expansion

$$Tx = \sum_{n=1}^{\infty}\xi_n(x)y_n, \qquad \sum_{n=1}^{\infty} \|\xi_n\|_{X'}\|y_n\|_Y<\infty.$$

Such a map is a norm limit of finite-rank maps and is therefore compact. For a locally convex space $E$ and a continuous seminorm $p$, let $E_p$ be the Banach completion of $E/\ker p$. One equivalent definition says that $E$ is a nuclear locally convex space when, for every continuous $p$, there is a stronger continuous seminorm $q\geq p$ for which the canonical linking map

$$E_q\longrightarrow E_p$$

is nuclear.

The rank-one definition, compactness consequence, and local-completion criterion are stated in Trèves 1967, §§47 and 50 and Vogt 2000, Theorems 3.15–3.16, PDF.

For a space whose topology is defined by an increasing sequence of Hilbert norms $\|\cdot\|_k$, there is a particularly usable criterion: for every $k$, some stronger level $\ell>k$ must give a Hilbert–Schmidt canonical linking map between the corresponding Hilbert completions. Two successive Hilbert–Schmidt links compose to a nuclear map. This criterion, including the hypotheses on the local Hilbert completions, is Vogt 2000, Theorem 3.16, PDF.

A transparent model is the rapidly decreasing sequence space

$$s = \left\{ a=(a_n)_{n\geq0}: \|a\|_k^2 = \sum_{n=0}^{\infty}(1+n)^{2k}|a_n|^2<\infty \ \text{for every }k \right\}.$$

The inclusion from level $k+r$ to level $k$ has singular values $(1+n)^{-r}$. It is Hilbert–Schmidt precisely when

$$\sum_{n=0}^{\infty}(1+n)^{-2r}<\infty, \qquad r>\frac12.$$

Thus a sufficiently strong level always maps summably into a weaker one. For this weighted $\ell^2$ scale the completion maps are genuine inclusions; that property was not assumed in the general criterion. The Hermite-coefficient description of $\mathcal S(\mathbb R^d)$ gives the analogous result there, with the stronger level chosen far enough above the weaker one.

This also supplies an important nonexample. An infinite-dimensional Banach space is not nuclear as a locally convex space: otherwise its identity map would be nuclear, hence compact, but a compact identity forces the closed unit ball, and therefore the space, to be finite-dimensional. In particular, an infinite-dimensional Hilbert space is not nuclear in its norm topology even though it contains individual trace-class or nuclear operators.

The word “nuclear” therefore has three distinct uses here:

1. a nuclear locally convex test space;
2. a nuclear map, which in a Hilbert setting is a trace-class operator; and
3. phase-space nuclearity conditions in algebraic QFT.

Only the first two are related by the linking-map definition above. Phase-space nuclearity is a separate physical condition and is not implied by choosing a nuclear test-function space.

Completed tensor products and distributional kernels

For locally convex spaces $E$ and $F$, the algebraic tensor product $E\otimes F$ contains only finite sums of simple tensors. It does not yet encode the limiting operations needed for distributions. The projective topology $\pi$ is generated by seminorms

$$(p\otimes_\pi q)(u) = \inf_{u=\sum_jx_j\otimes y_j} \sum_jp(x_j)q(y_j);$$

it is the strongest locally convex tensor topology that makes the canonical bilinear map $E\times F\to E\otimes F$ continuous. The injective topology $\epsilon$ is defined by testing tensors against equicontinuous families of dual functionals. In general, $E\widehat\otimes_\pi F$ and $E\widehat\otimes_\epsilon F$ differ.

Grothendieck nuclearity removes that ambiguity: if $E$ is nuclear, the projective and injective tensor topologies agree on $E\otimes F$ for every locally convex $F$. Taking the corresponding Hausdorff completions gives the same completed tensor product. This equivalence is stated precisely in Vogt 2000, Theorem 6.40, PDF.

For Schwartz spaces, the canonical multiplication of simple tensors extends to a topological isomorphism

$$\mathcal S(\mathbb R^m) \widehat\otimes \mathcal S(\mathbb R^n) \cong \mathcal S(\mathbb R^{m+n}), \qquad f\otimes g\longmapsto \bigl[(x,y)\mapsto f(x)g(y)\bigr].$$

Now let

$$B: \mathcal S(\mathbb R^m)\times \mathcal S(\mathbb R^n) \longrightarrow\mathbb C$$

be jointly continuous and bilinear. Projective completion first turns $B$ into a continuous linear functional on the completed tensor product; the isomorphism above then produces a unique $K_B\in\mathcal S'(\mathbb R^{m+n})$ such that

$$B(f,g) = \langle K_B,f\otimes g\rangle.$$

Equivalently, with the appropriate strong-dual topology, a continuous map $A:\mathcal S(\mathbb R^n)\to\mathcal S'(\mathbb R^m)$ has a unique tempered distribution kernel. This concrete tempered form is Melrose 2007, Theorem 1.2, PDF.

The output is a distributional kernel, not necessarily a function. The theorem does not by itself define a diagonal restriction, a product or composition of singular kernels, or a time-ordered product. Coordinate- independent kernels and density conventions on manifolds belong to Distributional Kernels on Manifolds.

Linear duals, antiduals, and the rigged triple

This site takes $\langle u|v\rangle$ to be conjugate-linear in $u$ and linear in $v$. Accordingly:

• $\Phi'$ denotes the continuous linear dual;
• $\Phi^\times$ denotes the continuous conjugate-linear antidual.

When topology on the antidual matters below, $\Phi^\times$ carries the strong topology $\beta(\Phi^\times,\Phi)$ of uniform convergence on bounded subsets of $\Phi$. Weak statements use $\sigma(\Phi^\times,\Phi)$ explicitly. Chiba 2011, Definition 3.1 and Eqs. (3.3)–(3.4) gives the weak/strong dual distinction and rigged embedding; its later analytic-continuation theory is not used here.

A rigged Hilbert space is a dense continuous embedding $i:\Phi\hookrightarrow\mathcal H$, where the locally convex topology of $\Phi$ is finer than the Hilbert norm topology, followed by the Riesz embedding into the antidual:

$$\begin{aligned} \mathcal H&\longrightarrow\Phi^\times,\\ h&\longmapsto F_h, \qquad F_h(\phi)=\langle\phi|h\rangle_{\mathcal H}. \end{aligned}$$

This map is linear with the site’s convention. It is injective because $\Phi$ is dense in $\mathcal H$, and it is continuous for the strong antidual topology because bounded subsets of $\Phi$ have bounded images in $\mathcal H$. By contrast,

$$h\longmapsto \bigl[\phi\mapsto\langle h|\phi\rangle\bigr]\in\Phi'$$

is conjugate-linear in $h$. The linear distribution $\delta_x(\phi)=\phi(x)$ and the antidual position ket $F_x(\phi)=\overline{\phi(x)}$ are therefore related by conjugation, not by a silent identification.

A rigged Hilbert space need not be nuclear. When $\Phi$ is nuclear, one speaks of a nuclear rigging or Gel’fand triple, but even then the topology and antidual are part of the choice; neither is canonical. Gel’fand and Vilenkin 1964, Chapter I gives the historical construction of the kernel theorem, nuclear spaces, and rigged Hilbert spaces in a single setting.

Generalized eigenvectors require an operator theorem

Suppose a self-adjoint operator $A$ on $\mathcal H$ satisfies

$$\Phi\subset D(A), \qquad A\Phi\subset\Phi, \qquad A:\Phi\to\Phi \ \text{continuously}.$$

Its antidual extension is defined weakly by

$$(A^\times F)(\phi)=F(A\phi).$$

A nonzero $F_\lambda\in\Phi^\times$ is a generalized eigenfunctional when

$$F_\lambda(A\phi) = \lambda F_\lambda(\phi) \qquad \text{for every }\phi\in\Phi.$$

This equation defines a candidate; it does not prove that enough such functionals exist.

The safe nuclear spectral-theorem statement is conditional. Let $\mathcal H$ be separable and let $\Phi\subset\mathcal H\subset\Phi^\times$ be a nuclear countably Hilbert Fréchet rigging. If $A$ is self-adjoint and satisfies the domain, invariance, and continuity conditions above, then, relative to a spectral representation and with spectral multiplicity retained, continuous generalized eigenfunctionals may be chosen for almost every spectral parameter. Matrix-element expansions for test vectors hold weakly, in the associated direct-integral sense. If one starts only with a symmetric operator on $\Phi$, essential self-adjointness and passage to its closure must be supplied first.

The generalized-eigenvector and nuclear-spectral-theorem step is developed in Bohm, Dollard, and Gadella 1989, Chapter I, § IV, pp. 22–30. The conditional Gel’fand–Maurin statement is also summarized in Antoine and Trapani 2023, §1, pp. 1–2; the spectral-measure construction, almost-everywhere qualification, and multiplicity labels are made explicit in Colbrook, Horning, and Xie 2025, §2.2, Eqs. (3)–(7). These qualifications are essential: a nuclear rigging does not give one normalized ket for every $\lambda$, turn continuous spectrum into Hilbert-space point spectrum, or produce a canonical basis.

A generalized momentum-eigenvector check

Take

$$\Phi=\mathcal S(\mathbb R^n) \subset \mathcal H=L^2(\mathbb R^n,d^np) \subset \mathcal S(\mathbb R^n)^\times.$$

For $p_0\in\mathbb R^n$, define the antidual functional

$$F_{p_0}(f)=\overline{f(p_0)}.$$

Evaluation is continuous in the Schwartz topology, so $F_{p_0}\in\Phi^\times$, while the earlier scaled-Gaussian test shows why it is not represented by an $L^2$ vector. The self-adjoint momentum operator $(P_jf)(p)=p_jf(p)$ preserves $\mathcal S$ continuously, and

$$\begin{aligned} (P_j^\times F_{p_0})(f) &= F_{p_0}(P_jf)\\ &= \overline{p_{0j}f(p_0)} = p_{0j}F_{p_0}(f). \end{aligned}$$

Thus the ket notation records an exact antidual eigenvalue equation. It does not assert that $F_{p_0}$ is normalizable or that the family $\{F_p\}$ is a Hilbert basis.

Controlled QFT application: a two-point kernel

Assume, without trying to construct it here, a scalar field for which $f\mapsto\phi(f)$ is linear on $\mathcal S(\mathbb R^d)$, a common invariant dense domain $\mathscr D$, and a vacuum vector $\Omega\in\mathscr D$. Assume also the adjoint/domain conditions needed below and the standard matrix-element condition that

$$f\longmapsto \langle\psi|\phi(f)\chi\rangle$$

is tempered for every $\psi,\chi\in\mathscr D$. These are among the operator-valued-distribution hypotheses stated in Dybalski 2018, §1.2.1, PDF. Define

$$B(f,g) = \langle\Omega| \phi(f)\phi(g)\Omega\rangle.$$

The assumptions make $B$ bilinear and separately continuous. Because Schwartz space is Fréchet, the Banach–Steinhaus consequence for scalar-valued bilinear forms upgrades separate continuity to joint continuity; see Trèves 1967, §34.2. The nuclear tensor-product identification therefore gives a unique

$$W_2\in \mathcal S'(\mathbb R^d\times\mathbb R^d)$$

with

$$B(f,g)=\langle W_2,f\otimes g\rangle.$$

This is the promised kernel-theorem application: a jointly continuous matrix-element rule in two test functions is packaged as one tempered distribution on the product spacetime. The common-domain and temperedness assumptions are inputs, not consequences of nuclearity. The operator-valued distribution, covariance, vacuum cyclicity, spectrum condition, locality, adjoint compatibility, positivity, and existence of a model all remain to be proved separately. The mathematical-QFT destination Wightman Fields, Domains, and Axioms develops the field and common-domain axioms. Wightman Functions and Spectral Support develops the distributional hierarchy and its spectral consequences, while The Wightman Reconstruction Theorem develops reconstruction.

A practical hypothesis check

Before invoking a kernel or generalized spectral statement:

1. Name the seminorms and the topology on the test space.
2. Verify separation, completeness, metrizability or LF structure as claimed.
3. State whether the target is the linear dual or antidual, and name its topology when convergence there matters.
4. Establish nuclearity with a linking-map or equivalent theorem.
5. Distinguish algebraic from completed tensor products and separate from joint continuity.
6. For a rigging, prove that $\Phi\hookrightarrow\mathcal H$ is dense and continuous.
7. For generalized spectral claims, check self-adjointness, invariance of $\Phi$, continuity on $\Phi$, spectral multiplicity, and almost-everywhere qualifications.
8. Read an expansion weakly, after pairing with test vectors, unless a stronger mode of convergence has been proved.

Stop if any required condition is missing. An algebraic dual is too large for continuity claims; a dense subspace without a topology is not a rigging; nuclearity does not follow from being Fréchet; and the notation $|\lambda\rangle$ does not establish existence or completeness.

Common pitfalls

Reversing the topology effect. A finer topology makes convergence harder, not easier. It can nevertheless make more linear functionals continuous because their inverse images have more open sets available.

Calling every test space Fréchet. The fixed-support spaces $\mathcal D_K$ are Fréchet, while global $\mathcal D(\Omega)$ carries an LF topology. Support motion is part of its convergence rule.

Confusing the meanings of nuclear. A nuclear locally convex space is not a trace-class operator, and neither notion is the phase-space nuclearity condition used in algebraic QFT.

Treating a kernel as a function. The kernel theorem yields an element of a distribution space. Restriction to a diagonal, multiplication, and composition require additional wavefront-set or regularity hypotheses.

Identifying dual and antidual. With a conjugate-linear bra, $\phi\mapsto\phi(x)$ is linear while $\phi\mapsto\overline{\phi(x)}$ is conjugate-linear. Dropping the conjugation changes scalar linearity.

Inferring a spectral expansion from the triple. A rigging specifies spaces and embeddings. Generalized eigenvectors and their completeness come from an operator theorem with additional hypotheses.

Exercises

Topology check. On $\mathcal S(\mathbb R)$, explain why $f_N(x)=e^{-Nx^2}$ refutes continuity of evaluation in the inherited $L^2$ norm but not in the Schwartz topology.

Solution

The sequence satisfies $\|f_N\|_2\to0$ while $f_N(0)=1$, so evaluation cannot obey $|f(0)|\leq C\|f\|_2$ on the Schwartz subspace and has no continuous extension to $L^2$. In the Schwartz topology, $|f(0)|\leq p_{00}(f)$ is exactly the required continuity estimate.

Nuclearity check. For the sequence-space scale above, determine when $H_\ell\to H_k$ is Hilbert–Schmidt.

Solution

Its singular values are $(1+n)^{k-\ell}$. Their squares are summable exactly when

$$2(\ell-k)>1.$$

Thus every level has a sufficiently stronger Hilbert–Schmidt level, which is the countably Hilbert nuclearity criterion.

Kernel stop check. A bilinear rule on $\mathcal S(\mathbb R^m)\times\mathcal S(\mathbb R^n)$ is algebraically defined, but no continuity estimate is known. What may be concluded?

Solution

Only an algebraic functional on simple finite sums is presently defined. One may not extend it to the completed projective tensor product or claim a tempered distribution kernel until joint continuity, or hypotheses that imply it, have been proved.

References

• Jean-Pierre Antoine and Camillo Trapani, “Operators in Rigged Hilbert Spaces, Gel’fand Bases and Generalized Eigenvalues”, Mathematics 11 (2023) 195, §1, printed pp. 1–2. This open-access specialist anchor states the conditional Gel’fand–Maurin theorem and separates completeness from simple spectrum.
• A. Bohm, J. D. Dollard, and M. Gadella, editors (1989), Dirac Kets, Gamow Vectors and Gel’fand Triplets, §§ II–IV of Chapter I, printed pp. 9–30, especially § IV, “Generalized eigenvectors and the nuclear spectral theorem,” printed pp. 22–30. This is the theorem-specific specialist anchor. Later resonance constructions in the book are outside this page’s scope.
• Hayato Chiba (2011), “A Spectral Theory of Linear Operators on Rigged Hilbert Spaces under Analyticity Conditions”, §3.1, Definition 3.1 and Eqs. (3.3)–(3.4), printed pp. 6–7. This is the specialist anchor for weak and strong dual topologies and the rigged embedding. Its additional analyticity hypotheses and resonance theory are not imported.
• Matthew J. Colbrook, Andrew Horning, and Tianyiwa Xie, “Computing Generalized Eigenfunctions in Rigged Hilbert Spaces”, Pure and Applied Analysis 7 (2025) 413–443, §2.2, Eqs. (3)–(7). This peer-reviewed comparison makes the spectral-measure, almost-everywhere, weak-expansion, and multiplicity qualifications explicit; its numerical method is outside this page’s scope.
• Wojciech Dybalski (2018), Lectures on Mathematical Foundations of Quantum Field Theory, PDF, §§1.2.1–1.2.2 and §2.1, printed pp. 4–6. This supports the QFT handoff: operator-valued distributions, their common domain, and Wightman distributions. None of those physical conditions is inferred from nuclearity here.
• I. M. Gel’fand and N. Ya. Vilenkin (1964), Generalized Functions, Volume 4: Applications of Harmonic Analysis, Chapter I, “The kernel theorem. Nuclear spaces. Rigged Hilbert space.” This is historical specialist reading for the three constructions; its dual convention must be translated to the site’s explicit antidual convention.
• Richard B. Melrose (2007), Introduction to Microlocal Analysis, Chapter 1, PDF, §1.1 and Theorem 1.2 in §1.4, printed pp. 13–18. This is the independent specialist anchor for Schwartz seminorms and the unique tempered kernel of a continuous map $\mathcal S\to\mathcal S'$.
• François Trèves, Topological Vector Spaces, Distributions and Kernels, Pure and Applied Mathematics 25, Academic Press, 1967, §§7, 10, 13, 43, 47, and 50–51. § 50, printed pp. 509–525, treats nuclear spaces; Proposition 50.2 and Corollary 2, printed p. 520, give the normable-space nonexample; Theorem 51.5 and its corollary, printed pp. 529–530, cover nuclear test spaces; and Theorem 51.7, printed pp. 531–532, is the kernel theorem. Together, these results establish the locally convex, nuclear-space, and kernel-theorem claims used on this page.
• Dietmar Vogt (2000), Lectures on Fréchet Spaces, PDF, §§1, 3, and 6.4–6.5, especially Theorems 3.15–3.16, 6.39–6.40, and 6.45–6.46. This is a specialist anchor for the seminorm criteria, Hilbert–Schmidt linking maps, projective–injective equivalence, and the concrete route to the kernel theorem. Its Fréchet-centered statements are not used to misclassify global $\mathcal D(\Omega)$.