In quantum mechanics one routinely speaks of eigenstates of quantum observables with continuous spectrum, notably of eigenstates of position and of momentum, written in bra-ket notation as and and “normalized”, following Dirac, in terms of the Dirac delta distribution, . But these self-adjoint unbounded operators have no actual eigenvectors in the Hilbert space of states (see below). The spectral theorem in its Hilbert space form makes rigorous sense of such operators by trading their eigenvectors for spectral measures (projection-valued measures), but in doing so it gives up the transparent picture of expanding any state in a “basis” of eigenstates, which is what makes Dirac’s formalism so effective in practice.
The notion of rigged Hilbert spaces (also: equipped Hilbert spaces, closer to the Russian original, or Gelfand triples) restores this picture rigorously, by enlarging the given Hilbert space to a space of generalized vectors in which the missing eigenvectors do exist — much as distributions, such as the Dirac distribution, enlarge the space of ordinary functions. To this end one chooses a dense subspace (notation following Gel’fand & Vilenkin 1961 §4.2–3) of particularly well-behaved vectors, equipped with its own topology, finer than the one induced from (typically making a nuclear Fréchet space, such as a Schwartz space of test functions). Since continuity with respect to this finer topology is a weaker condition, there are more continuous antilinear functionals on than on — the latter being, by the Riesz representation theorem, just the vectors of themselves. This yields a triple of continuous inclusions
where denotes the space of continuous antilinear functionals on , into which embeds as .
If now a self-adjoint operator on restricts to a continuous linear operator , then it extends by duality to an operator on generalized vectors, given by , and a generalized eigenvector of with generalized eigenvalue is an element with . The nuclear spectral theorem (due to Gel’fand & Kostyuchenko and to Maurin, hence also called the Gel’fand-Maurin theorem, see Gel’fand & Vilenkin 1961, Ch. I, Blanchard & Brüning 2015, Ch. 29) says that when is nuclear, then has a complete system of generalized eigenvectors , indexed by (almost all) points of its spectrum (and in general also by a multiplicity label, here suppressed), in that there is a measure on the spectrum such that
This is the rigorous form of Dirac’s completeness relation “”.
In this way, rigged Hilbert spaces provide a rigorous home for Dirac’s bra-ket formalism: Kets are elements of , the corresponding bras are their complex conjugates, which are elements of the dual space of continuous linear functionals, and a bracket
is well-defined as long as at least one of its two entries is a test vector in . (Brackets between two generalized vectors, such as , are not defined in general; see below for what they stand for.)
The archetypical example is the quantum mechanics of a particle on the line, with the eigenstates of position and of momentum as generalized vectors. Throughout, let
be the Hilbert space of square-integrable functions on the real line, with respect to Lebesgue measure, with inner product (as usual in physics) antilinear in its first argument.
The position operator and the momentum operator (in units with )
are unbounded self-adjoint operators on , defined not on all of but on dense domains (for instance ). Both have spectrum all of , but:
(no eigenvectors in the Hilbert space)
Neither nor has a single eigenvector in :
If , then for almost all , whence .
If , then in the sense of distributions, whence , which is square-integrable only for .
The would-be eigenfunctions, namely the Dirac distribution and the plane wave , do make sense, but as distributions instead of as elements of . The rigged Hilbert space makes this precise.
As the space of test vectors take the Schwartz space of smooth functions all of whose derivatives are rapidly decreasing, with its Fréchet space-topology given by the seminorms
and write
for the Fourier transform.
(Schwartz space as test space for position and momentum)
The inclusion is continuous with dense image.
and restrict to continuous linear operators .
The Fourier transform restricts to a linear automorphism of which preserves the inner product (by the Plancherel theorem) and turns into multiplication by the momentum variable, (and into ).
is a nuclear space.
Using we have
which gives the continuity in the first item, while density holds since already the compactly supported smooth functions are dense in . For the second item, the Leibniz rule gives
(with ). The last two items are standard, see at Schwartz space and at nuclear space.
Now let be the space of continuous antilinear functionals . Since is closed under complex conjugation, is a linear isomorphism onto the usual space of tempered distributions, so that is just the space of tempered distributions in a guise adapted to the inner product. Concretely, every locally integrable function of at most polynomial growth defines an element of by
and so does every , continuously so by the Cauchy-Schwarz inequality and the first item of Prop. . The resulting map is injective, since is dense in , and so we have the rigged Hilbert space
For and we write, as in the Idea section, and , which for reproduces the inner product.
Finally, for define by
By self-adjointness of , this restricts on the domain to the original operator , while under the identification it is the usual action of and of on tempered distributions.
For consider the generalized vectors given by
hence with
(These are continuous on , since and .)
(generalized position and momentum eigenvectors)
For all we have
For all we have, using that is real,
and, by integration by parts (the boundary terms vanishing since is rapidly decreasing while is bounded),
(completeness relations)
For all we have
equivalently, every test vector has the generalized eigenvector expansions
where the integrals converge weakly in , i.e. after pairing with any .
The integrands are and , respectively. Hence the first equality is the definition of the inner product on , and the second is the Plancherel theorem. In the weak form, the second expansion of is thereby Fourier inversion.
(Dirac’s notation)
In Dirac’s bra-ket notation, is the position eigenket and is the momentum eigenket , and the above says precisely that:
and are the wavefunctions of in the position and in the momentum representation, respectively;
— here, exceptionally, a bracket between two generalized vectors makes sense, because is a smooth function, whose value at is defined.
On the other hand, formulas such as and pair two generalized vectors and have no literal meaning in the rigged Hilbert space; they are mnemonics for the completeness relations, obtained by formally inserting (resp. ) into the expansions of Prop. .
(dependence on the choice of test space)
In this example, the generalized eigenvalues of and on are exactly the points of their spectrum : For , multiplication by the function (all of whose derivatives are bounded) is a continuous inverse to on , so that is invertible on and hence has trivial kernel; similarly for , via Fourier transform.
But the generalized eigenvalues do depend on the choice of test space:
A smaller test space has a larger antidual, which may contain spurious generalized eigenvectors. For instance, with the space of compactly supported smooth functions (also nuclear and also preserved by and ) as test space, the antidual (equivalently: all distributions) contains for every complex number , and these are generalized eigenvectors of with non-real eigenvalue , which do not contribute to the completeness relation.
A larger test space with coarser topology may lose the generalized eigenvectors. For instance, on the domain with its graph norm, point evaluations are not even defined (elements of being only -classes), and so its antidual contains no .
The Schwartz space sits well between these extremes, simultaneously for and for , which is no accident:
(intrinsic characterization via the harmonic oscillator)
The Schwartz space is intrinsically determined by and : It is the space of smooth vectors of the Schrödinger representation of the Heisenberg group on , and coincides with the intersection of the domains of all powers of the harmonic oscillator Hamiltonian . Expanding in the orthonormal basis of Hermite functions , which are honest eigenvectors , identifies the above rigged Hilbert space with
where is the space of rapidly decreasing sequences and that of sequences of at most polynomial growth. In this form the nuclearity of is manifest: is the intersection of the weighted Hilbert spaces , and the inclusions are Hilbert-Schmidt.
The notion of rigged Hilbert spaces, together with the nuclear spectral theorem for them, is due to:
Early treatments in the context of Dirac’s bra-ket formalism:
John E. Roberts: The Dirac Bra and Ket Formalism, J. Math. Phys. 7 (1966) 1097–1104 [doi:10.1063/1.1705001]
John E. Roberts: Rigged Hilbert spaces in quantum mechanics, Communications in Mathematical Physics 3 (1966) 98–119 [doi:10.1007/BF01645448]
Textbook account, including the nuclear spectral theorem:
Pedagogical introduction aimed at physicists, via an exactly solvable example:
Claim of a unification of various inequivalent approaches:
See also:
S. Wickramasekara, A. Bohm, Symmetry Representations in the Rigged Hilbert Space Formulation of Quantum Mechanics [arXiv:math-ph/0302018]
MathOverflow: Good references for Rigged Hilbert spaces?
Last revised on October 26, 2026 at 20:11:40. See the history of this page for a list of all contributions to it.