In this work we show that standard results in quantum tomography, which are typically proven in the context of Hilbert spaces, can analogously be established in the theory of Schwartz spaces. In particular, we will see that quantum states are equivalently characterized by their Wigner function, the characteristic of their Wigner function, and the probability distributions for a family of linear combinations of position and momentum. Physically, the latter can be understood as the evolution of its probability distribution in position over time under, for example, free or harmonic evolution. The result is obtained with techniques that are more accessible to a physicist audience, and shows that, at least for tomography results, the full structure of Hilbert spaces is not necessary.
Introduction
In previous work [2], some of us raised awareness of the difficulty of providing a physical justification for Hilbert spaces as a mathematical foundation for quantum mechanics. In particular, the completeness of the space with respect to the topology induced by the inner product necessarily introduces states with pathological physical properties. Efforts to reconstruct the theory of quantum mechanics based on physical requirements might therefore look for alternative targets. Here we follow up on our investigation of Schwartz spaces as a source of promising insights for alternative definitions and see whether we can establish analogue results completely within that theory.
Schwartz functions are smooth functions that decay rapidly at infinity, are dense in L2 and have finite expectation values for all polynomials of position and momentum, which depend continuously on the state. These nice mathematical properties considerably simplify their analysis by avoiding many of the domain issues common for (unbounded) operators on Hilbert spaces and other subtleties like functions only being defined almost everywhere. In this paper, we focus on the use of tomography as a powerful tool to analyze the physical content encoded in Schwartz wavefunctions.
We will see that the state of a particle described by a Schwartz wavefunction can be uniquely identified (up to a global phase) by measuring the probability distribution of a family of linear combinations of position and momentum. This can also be interpreted as the evolution over time of the probability distribution in position, either freely or in the presence of a harmonic potential. While the latter can be achieved in finite time, the former would require infinite time. The probability densities in question are Schwartz functions as well, i.e., rapidly decaying at infinity. This result is analogous to the reconstruction of a probability distribution on classical phase space [6, 11].
While the general claim already exists (in different forms) in the literature (see e.g., Refs. [1, 12, 9, 5, 10]), we provide here a short and straightforward proof, simplified by the properties of Schwartz spaces. In particular, we do not require any reference to the completeness of L2. We hope that this makes the discussion more accessible to a broad physics community.
The paper starts by reviewing some mathematical prerequisites, including Wigner functions. The reconstruction then proceeds in the following steps.
Figure 1. Sequence of reconstructions.
The standard definition of a complex (wave-)function ψ with certain regularity/decay properties.
The well-known Wigner function Wψ, i.e., a quasiprobability distribution in phase space.
The characteristic function χψ, i.e., expectation values of the unitary continuous operator ei(aX+bP) implementing translations in phase space.
Probability distributions ρα for a family of linear combinations of the position and momentum operators cosαX+sinαP, which are all continuous operators.
The probability distribution ρt of the position operator evolving freely (or harmonically) in time.
Mathematical prerequisites
In this section we provide the necessary prerequisites for what follows. First we introduce the central object of our analysis, namely Schwartz functions, which have a number of nice properties. Then we define Wigner functions as an alternative representation of quantum mechanical states and show the well-known result of how to recover the original wavefunction from its Wigner function, thereby showing that there exists a (constructive) one-to-one correspondence. Since the Wigner function is a special case of a more general transformation between operators and functions on phase space (see § 2.3), it is more convenient to show how to reconstruct the kernel of the wavefunction, which is closer to the density matrix/operator of the state. Lastly we introduce a family of operators that will be used in the reconstruction later.
Schwartz space
Seminorms The Schwartz space S(Rn,C) of smooth functions f:Rn→C and its topology are defined in terms of the following family of seminorms, indexed by multi-indices α,β∈Nn
∥f∥α,β=x∈RnsupxαDβf(x)(1)
where xα=x1α1x2α2⋯xnαn are position operators and Dβf=∂1β1∂2β2⋯∂nβnf are (except for a factor of −iℏ) momentum operators.
Topological vector space The Schwartz space is then
S(Rn,C)={f∈C∞(Rn,C)∣∀α,β∈Nn,∥f∥α,β<∞}(2)
with the locally convex topology generated by all these semi-norms.
Inner product and Completeness We equip S(Rn,C) with the L2 inner product (and consequent norm)
⟨ψ∣ϕ⟩=∫Rnψ∗(x)ϕ(x)dnx(3)
which also induces another topology (which defines L2), but is not used here. Note that S(Rn,C) is complete w.r.t. to its own topology, but not w.r.t. to that of L2. In fact, the unique completion of S(Rn,C) in that topology is L2. This makes L2 a Hilbert space, while S(Rn,C) is a pre-Hilbert space.
Physical seminorms Alternatively, one can use the following families of semi-norms closely related to the moments of position and momentum
∥f∥α,βL1=∫RnxαDβf(x)dnx(4)
∥f∥α,βL2=(∫RnxαDβf(x)2dnx)1/2(5)
which will analogously define the same set and topology [8].
Fourier transform The Schwartz space is closed under the Fourier transform F:S(Rn,C)→S(Rn,C)
(Fψ)(p)=(2πℏ)n1∫Rne−ℏipxψ(x)dx(6)
and its inverse
(F−1ψ)(x)=∫Rneℏipxψ(p)dp.(7)
The same holds for partial Fourier transforms like
Operator algebra All operators in this paper are everywhere defined on S(Rn,C), linear and continuous. These form the algebra L(S) w.r.t. composition. In this context symmetric and self-adjoint are equivalent, namely for all ψ,ϕ∈S(Rn,C)
⟨Oψ∣ϕ⟩=⟨ψ∣Oϕ⟩.(9)
Wigner functions
Definition 10 (Wigner function). For a functionψ∈S(R,C), its Wigner functionWψ:R2→Ris given by
Wψ(x,p)=2πℏ1∫Re−ℏipyψ(x+2y)ψ(x−2y)dy.(10)
The Wigner function is a real Schwartz function inS(R2,C) (see Prop. 1.92 in [7]).
Definition 12 (Kernel). The kernelKψ:R2→Cof a functionψ∈S(R,C)is given by
Kψ(x,y)=ψ(x)ψ(y).(11)
Proposition 14.The kernelKψ:R2→Cof a Schwartz functionψ∈S(R,C)is again a Schwartz function, i.e.,Kψ∈S(R2,C).
Proof. Since ψ∈C∞(R,C) also ψ∈C∞(R,C) and their product. The norm factors
Alternatively we can prove this using a generalized notion of orthogonality for the de Broglie family eℏipx. See appendix § 5.§ 2 for more.
Theorem 25 (Wavefunction reconstruction). Given the kernel Kψ:R2→C of some (normalized) function ψ∈S(R,C), we can find a function φ∈S(R,C) with the same kernel. For that, find a point y0 where Kψ does not vanish and fix the phase convention
φ(y0)=Kψ(y0,y0)>0.(21)
Then set
φ(x)=Kψ(y0,y0)Kψ(x,y0).(22)
The (re)constructed function φ only differs from ψ by a phase, i.e., φ=eiθψ.
Proof. Notice that diagonal entries in the kernel are necessarily real and the square root is therefore well-defined, real and positive since we picked a point y0 where the kernel does not vanish, which must exist for the kernel of anything but the zero function.
Theorem 29 (Reconstruction). From the previous two theorems it immediately follows that we can reconstruct any (normalized) function ψ∈S(R,C) from its Wigner function Wψ, up to a global phase eiθ.
Weyl–Wigner Correspondence
Definition 30.For a function w:R2→C∈OM(R2) its Weyl quantization Ow:S(R,C)→S(R,C) is given by
Theorem 32 (Weyl–Wigner identity). The expectation values are preserved when transforming operators to functions on phase space and wave functions to Wigner functions, i.e.,
We consider the family of unitary operators W(a,b):S(R,C)→S(R,C) defined for any a,b∈R as
W(a,b)=exp(ℏi(aX+bP))(26)
where the position operator X:S(R,C)→S(R,C) and the momentum/derivative operator P:S(R,C)→S(R,C) are defined respectively as
(Xψ)(x)=xψ(x)(27)
(Pψ)(x)=−iℏdxdψ(x).(28)
Both of these operators are defined everywhere, continuous in the Schwartz topology, and symmetric, which is the same as self-adjoint in this case on S(R,C). The exponential function can be defined in terms of its power series, which would not be possible on L2. Alternatively, they can be defined via translations and modulations.
Translation For b∈R, define Tb:S(R,C)→S(R,C) as
(Tbφ)(x)=φ(x+b).(29)
Modulation For a∈R, define Ma:S(R,C)→S(R,C) by
(Maφ)(x)=eℏiaxφ(x).(30)
Definition 39. For (a,b)∈R2, define the Weyl–Heisenberg operator on S(R,C) by
W(a,b)=e2ℏiabMaTb.(31)
Equivalently, for φ∈S(R,C),
(W(a,b)φ)(x)=eℏia(x+2b)φ(x+b).(32)
Reconstructions
We now proceed with the next steps in our reconstruction.
Characteristic function
Definition 42. The characteristic function χψ:R2→C of a function ψ∈S(R,C) is
The inverse relationship follows immediately, once we recognize the first one to be the Fourier transform.
Corollary 54. Since the Fourier transform preserves the Schwartz space we conclude that the characteristic function χψ of a Schwartz function ψ is again a Schwartz function.
Theorem 55. The expectation values of W(a,b)
⟨ψ∣W(a,b)ψ⟩(43)
uniquely define the state ψ∈S(R,C) up to a global phase.
Proof. The expectation values, i.e., values of the characteristic function, are the Fourier transform of the Wigner function. Since the Fourier transform is invertible and the Wigner function uniquely determines the state, we can also recover the state from these expectation values.
Probability distributions
We will define the probability distribution of linear combinations of X and P like2
X(α)=cosαX+sinαP.(44)
Proposition 58. This parametrized curve can be generated by the operator A:S(R,C)→S(R,C), defined as3
A=21(X2+P2),Uα=e−ℏiαA(45)
X(α)=UαXUα†=cosαX+sinαP.(46)
Note that there are no domain problems for A since X and P are already defined everywhere and map back to S(R,C).
which is solved by the Ansatz: X(α)=a(α)X+b(α)P. Then a and b satisfy
dα2d2a=−a(α)dα2d2b=−b(α)(48)
dαdaα=0=0dαdbα=0=1(49)
a(0)=1b(0)=0(50)
and have well-known solutions:
a(α)=cos(α)b(α)=sin(α)(51)
⟹X(α)=cos(α)X+sin(α)P(52)
See appendix § 5.§ 4 for more detailed calculations.
Similarly, we can transform the initial probability distribution in X=X(0) for a given state ψ∈S(R,C) as
ρ0(q)=∣ψ(q)∣2⟶ρα(q)=∣(Uαψ)(q)∣2(53)
Theorem 68.The Fourier transform ρ^α of the probability distribution ρα of X(α) (as defined above) is related to the characteristic function χψ as follows
ρ^α(s):=∫Reℏisqρα(q)dq=χψ(scosα,ssinα).(54)
Proof. Let us call ψα=Uαψ.
ρ^α(s)=∫Rψα(q)eℏisqψα(q)dq(55)
=⟨Uαψ∣eℏisXUαψ⟩(56)
=⟨ψ∣Uα†eℏisXUαψ⟩(57)
=⟨ψ∣eℏisUα†XUαψ⟩(58)
=⟨ψ∣eℏisX(α)ψ⟩(59)
=⟨ψ∣eℏis(cosαX+sinαP)ψ⟩(60)
=⟨ψ∣W(scosα,ssinα)ψ⟩(61)
=χψ(scosα,ssinα)(62)
Note that ρ^α is again a Schwartz function for every α.
Theorem 78.Given the probability distributions ρα with α∈[0,π) of the observables X(α)=cosαX+sinαP for a state ψ∈S(R,C), as defined above, we can reconstruct the original state.
Proof. As shown in theorem 68, we can reconstruct the characteristic function χψ from ρ^α, which we obtain from ρα using the inverse Fourier transform. Then theorem 55 tells us that we can reconstruct the original state from χψ.
Physical interpretation
As has been noted above the generator of the family is the Hamiltonian of the harmonic oscillator. Physically, this means that the family can be interpreted as the position operator evolving under the harmonic potential in the Heisenberg picture, and the parameter α represents time. However, this is not the only Hamiltonian one can use for tomography.
For free evolution, we know classically x(t)=x(0)+mpt=x(0)+mtp. Setting tan(α)=mt we can see that this curve of operators corresponds to observing the position X over time t during free evolution.
X(t)=cos(α(t))1X(α(t))=X+tan(α)P=X+mtP(63)
Note that while the point α=π/2 is never reached we can get arbitrarily close and since expectation values are continuous the single point does not contain any further data.
To see this explicitly, define the Hamiltonian for free evolution.
H=2m1P2(64)
It generates time evolution via
Ut=e−ℏiHt(65)
and therefore
ψt=Utψ(66)
The probability density for the distribution of position X in the state ϕt at time t is then
In the Heisenberg picture this corresponds to the distribution of the observable UtXUt†. See appendix § 5.5 for a derivation.
Conclusion
We presented a short proof of several tomographic characterizations of Schwartz wavefunctions. The proof requires only basic tools from analysis, which we hope will make it accessible to a broad audience. The characterization in terms of the probability distribution ρα at specific angles α in phase space has a precise analogue in classical mechanics. Under harmonic evolution, probability densities on phase space rotate around the origin (for some choice of units). The marginals for all rotation angles α uniquely determine the full phase-space measure by application of the Cramér–Wold theorem [6, 11].
This result sits within our larger goal of understanding what vector space representation of quantum states can be both mathematically precise and physically meaningful. The result shows that, at least for quantum tomography, the Schwartz spaces provide analogous results that are independent of the Hilbert space theory. While Schwartz spaces are limited to quantum systems describing only positional degrees of freedom of particles, and therefore cannot be the ultimate solution, studying Schwartz analogues may pave the way to find better alternatives to the standard Hilbert space formalism.
Appendix
Further mathematical prerequisites
We provide some further properties of Schwartz spaces that allow us to use different techniques that are much closer to common practice in physics to prove some of the results in this paper.
Multipliers Schwartz spaces are closed under multiplication by any function in OM(Rn), the space of slowly increasing smooth function, i.e., all f∈C∞(Rn,C), s.t.
∀α∈Nn,∃k∈N,C≥0,∀x∈Rn:∣∂αf(x)∣≤C(1+∣x∣2)k,(68)
which are precisely those functions that preserve Schwartz spaces under multiplication.
Tempered distributions In contrast to L2, which is self-dual, the dual4 space of S is much bigger and contains elements like the Dirac distribution5
δx[f]=f(x)=∫Rnδ(y−x)f(y)dny(69)
which cannot be defined on L2 since it does not contain functions but equivalence classes of functions. The elements of this dual space S′ are called tempered distributions.
Integral and Dirac measure Taking the integral above literally requires some further qualifications, since there is no function δ:Rn→R that would satisfy eq. (85). However, δx∈S′ can be multiplied by f∈S(Rn,C) (or in general OM(Rn)) to give another tempered distribution, which can be evaluated at the constant function 1.
δx[f]=δx[f⋅1]=(δx⋅f)[1]=∫Rn1(fδx)(70)
The first multiplication ⋅ is between an element of OM(Rn) and S(Rn,C), while the second multiplication is between S′ and S(Rn,C) and defined in terms of the first, just as written. Note that even though 1∈/S(Rn,C) we can evaluate δx⋅f on any OM(Rn) function and any indicator function 1U of some Borel set U⊆Rn. This defines a measure
μ(U):=(δx⋅f)[1U](71)
which can be integrated over Rn to give the same result.
δx[f]=∫Rndμ(72)
Approximation by integrals Another justification for using the integral notation is that S is dense in S′, i.e., every linear functional in S′ can be approximated by integrals of the form
∫Rnfi(x)g(x)dx→F[g],(73)
where fi∈S. The δx distribution for example can be approximated by Gaussian probability density functions with mean x and standard deviation σ→0.
Alternative proof using orthogonal de Broglie family
This is an alternative proof of theorem 20 that uses a generalized notion of orthogonality for the de Broglie family eℏipx.
where the inner integral over z can be interpreted as the application of the tempered distributions ηp=⟨eℏipz⋅⟩ parametrized by p to the function a(z)=ψ(2x+y+z)ψ(2x+y−z). The result is a function of p evaluated by the tempered distribution ηx−y. In formulas,
Kψ(x,y)=2πℏ1ηx−y[p↦ηp∗[a]].(75)
By delaying the evaluation at a we can generalize the inner product
Kψ(x,y)=2πℏ1⟨η∣ηx−y⟩[a]=δx−y[a]=a(x−y)(76)
and use the fact that η is an orthogonal family in that generalized sense (see Refs. [3, 4]) to simplify this to
where we have used the fact that exp(iℏ12mtP2) commutes with P.
References
[1]J. Bertrand and P. Bertrand. “A Tomographic Approach to Wigner’s function”. In: Foundations of Physics 17.4 (1987). DOI: 10.1007/BF00733376.DOI
[2]Gabriele Carcassi, Francisco Calderón, and Christine A. Aidala. “The unphysicality of Hilbert spaces”. In: Quantum Studies: Mathematics and Foundations 12.1 (Jan. 2025), p. 13. ISSN: 2196-5617. DOI: 10.1007/s40509-024-00357-0.DOI
[3]David Carfì. “Dirac-orthogonality in the space of tempered distributions”. In: Journal of Computational and Applied Mathematics 153.1 (June 2001). Proceedings of the 6th International Symposium on Orthogonal Polynomials, Special Functions and their Applications, pp. 99–107. ISSN: 0377-0427. DOI: 10.1016/S0377-0427(02)00634-9.
[4]David Carfì and Alessia Donato. Dirac Calculus and Schwartz Linear Algebra: Dirac inner product and Dirac orthonormality. Available as a Researchgate Paper. Oct. 2015. DOI: 10.13140/RG.2.1.3137.2885.DOI
[5]G. Cassinelli et al. “Group theoretical quantum tomography”. In: Journal of Mathematical Physics 41.12 (Dec. 2000), pp. 7940–7951. ISSN: 0022-2488. DOI: 10.1063/1.1323497.DOI
[6]Harald Cramér and Herman Wold. “Some Theorems on Distribution Functions”. In: Journal of the London Mathematical Society s1-11.4 (1936), pp. 290–294. DOI: 10.1112/jlms/s1-11.4.290.DOI
[7]Gerald B. Folland. Harmonic Analysis in Phase Space. (AM-122). Princeton, NJ: Princeton University Press, 2016. DOI: 10.1515/9781400882427.
[8]Gabriele Carcassi for the AoP Collaboration. Equivalence of sup-norm, L¹, and L² seminorms on the Schwartz space. Assumptions of Physics, Technical Brief 010, 2026. URL: https://assumptionsofphysics.org/autogen/briefs/010-SchwartzNorms.pdf.
[9]J Kiukas, P Lahti, and J-P Pellonpää. “A proof for the informational completeness of the rotated quadrature observables”. In: Journal of Physics A: Mathematical and Theoretical 41.17 (Apr. 2008), p. 175206. DOI: 10.1088/1751-8113/41/17/175206.DOI
[10]U. Leonhardt and M.G. Raymer. “Observation of Moving Wave Packets Reveals Their Quantum State”. In: Physical Review Letters 76.12 (1996). DOI: 10.1103/PhysRevLett.76.1985.DOI
[11]Russell Lyons and Kevin Zumbrun. “A calculus proof of the Cramér-Wold theorem”. In: Proc. Amer. Math. Soc. 146 (2018), pp. 1331–1334. DOI: 10.1090/proc/13794.DOI
[12]K. Vogel and H. Risken. “Determination of quasiprobability distributions in terms of probability distributions for the rotated quadrature phase”. In: Physical Review A 40.5 (1989). DOI: 10.1103/PhysRevA.40.2847.DOI
Anuj Kumar Singh, Atanu Patra, Paul Konrad, Subhamoy Sahoo, Ankur Arora, Johannes Düreth, Georgy V. Astakhov, Monika Emmerling, Martin Kamp, Andreas Sperlich et al. ·