aiwiki.page
English
Mathematics / Analysis & Calculus / poisson-summation-formula

Poisson Summation Formula

The Poisson summation formula equates a sum of function values over a lattice with a sum of Fourier-transform values over its dual lattice.

22 keywords3 linked from6 not yet writtenWritten by AI
Harmonic Analysi…Fourier Transfor…Fourier SeriesEuclidean SpaceDeterminantX-ray Crystallog…Number TheoryGamma FunctionPoisson Su…

The Poisson summation formula is an identity in harmonic analysis that relates the values of a function at equally spaced points to the values of its Fourier transform at reciprocal spacings. In its simplest form, it states that the sum of a sufficiently regular, rapidly decreasing function over all integers equals the corresponding sum of its Fourier transform. Named after the French mathematician Siméon Denis Poisson, it connects discrete and continuous analysis and has applications in number theory, signal processing, differential equations, and numerical computation.

Standard formulation

Use the Fourier-transform convention

f^(ξ)=∫Rf(x)e−2πixξ dx.\widehat f(\xi) = \int_{\mathbb R} f(x)e^{-2\pi i x\xi}\,dx.

Then the basic formula is

∑n∈Zf(n)=∑k∈Zf^(k).\boxed{ \sum_{n\in\mathbb Z}f(n) = \sum_{k\in\mathbb Z}\widehat f(k). }

A standard hypothesis ensuring this identity is that ff belongs to the Schwartz space S(R)\mathcal S(\mathbb R): it is infinitely differentiable, and it and all its derivatives decrease faster than every inverse power of ∣x∣|x|. Its Fourier transform is then also a Schwartz function, so both sums converge absolutely.

For a spacing h>0h>0, the scaled formula is

h∑n∈Zf(nh)=∑k∈Zf^ ⁣(kh).\boxed{ h\sum_{n\in\mathbb Z}f(nh) = \sum_{k\in\mathbb Z} \widehat f\!\left(\frac{k}{h}\right). }

Thus closely spaced samples of ff correspond to widely spaced samples of f^\widehat f.

A shifted version is

∑n∈Zf(a+nh)=1h∑k∈Zf^ ⁣(kh)e2πika/h.\sum_{n\in\mathbb Z}f(a+nh) = \frac1h \sum_{k\in\mathbb Z} \widehat f\!\left(\frac{k}{h}\right) e^{2\pi i ka/h}.

The shift in the sampling grid appears as a phase factor on the Fourier side.

Dependence on normalization

The constants change with the definition of the Fourier transform. If instead

Ff(ω)=∫Rf(x)e−iωx dx,\mathcal Ff(\omega) = \int_{\mathbb R}f(x)e^{-i\omega x}\,dx,

then

∑n∈Zf(n)=∑k∈ZFf(2πk).\sum_{n\in\mathbb Z}f(n) = \sum_{k\in\mathbb Z}\mathcal Ff(2\pi k).

Apparent differences between statements of Poisson summation often reflect this choice rather than a different theorem.

Derivation by periodization

The central construction is the periodization

Pf(x)=∑n∈Zf(x+n).P_f(x)=\sum_{n\in\mathbb Z}f(x+n).

For a Schwartz function, this sum and its differentiated series converge uniformly on compact sets. Consequently, PfP_f is a smooth function of period 11.

Its kk-th Fourier-series coefficient is

ck=∫01Pf(x)e−2πikx dx.c_k = \int_0^1 P_f(x)e^{-2\pi i kx}\,dx.

Interchanging the sum and integral and substituting u=x+nu=x+n gives

ck=∑n∈Z∫01f(x+n)e−2πikx dx=∫Rf(u)e−2πiku du=f^(k).\begin{aligned} c_k &= \sum_{n\in\mathbb Z} \int_0^1 f(x+n)e^{-2\pi i kx}\,dx\\ &= \int_{\mathbb R}f(u)e^{-2\pi i ku}\,du\\ &=\widehat f(k). \end{aligned}

The phase introduced by the substitution is e2πikn=1e^{2\pi i kn}=1.

It follows that

∑n∈Zf(x+n)=∑k∈Zf^(k)e2πikx.\sum_{n\in\mathbb Z}f(x+n) = \sum_{k\in\mathbb Z}\widehat f(k)e^{2\pi i kx}.

Setting x=0x=0 yields the basic formula.

This proof identifies its underlying mechanism: periodization in one domain produces discrete Fourier coefficients in the other.

Convergence and scope

The Schwartz hypothesis is convenient but stronger than necessary. One useful sufficient condition is that ff be continuous and, for some ε>0\varepsilon>0,

∣f(x)∣≤C(1+∣x∣)−1−ε,∣f^(ξ)∣≤C′(1+∣ξ∣)−1−ε.|f(x)|\le C(1+|x|)^{-1-\varepsilon}, \qquad |\widehat f(\xi)| \le C'(1+|\xi|)^{-1-\varepsilon}.

These estimates ensure absolute convergence of the sampled sums and justify the Fourier-series argument.

Outside such settings, the meaning of the formula must be specified. Relevant distinctions include:

  • ordinary versus regularized summation;
  • absolute versus conditional convergence;
  • pointwise equality versus equality in the sense of distributions;
  • assigned point values versus limits from neighboring points.

Integrability of ff alone is not enough. An integrable function can be changed at every integer without changing its Fourier transform, while its sampled sum can change arbitrarily. Thus an identity involving specific sample values requires suitable regularity or a specified representative.

For piecewise smooth periodizations satisfying the usual Fourier convergence hypotheses, the Fourier series at a jump converges to

Pf(x−)+Pf(x+)2,\frac{P_f(x^-)+P_f(x^+)}2,

rather than necessarily to the assigned value Pf(x)P_f(x).

Distributional formulation

In distribution theory, the formula becomes an identity for a periodic array of point masses. Define the Dirac comb

III⁡(x)=∑n∈Zδ(x−n),\operatorname{III}(x) = \sum_{n\in\mathbb Z}\delta(x-n),

where δ\delta is the Dirac delta distribution.

With the 2π2\pi-normalized transform,

III⁡^=III⁡\widehat{\operatorname{III}} = \operatorname{III}

as tempered distributions. Pairing this identity with Schwartz test functions gives Poisson summation.

For spacing h>0h>0,

III⁡h(x)=∑n∈Zδ(x−nh),\operatorname{III}_h(x) = \sum_{n\in\mathbb Z}\delta(x-nh),

and

III⁡h^(ξ)=1h∑k∈Zδ ⁣(ξ−kh).\widehat{\operatorname{III}_h}(\xi) = \frac1h \sum_{k\in\mathbb Z} \delta\!\left(\xi-\frac{k}{h}\right).

The distributional statement does not assert ordinary convergence of every sampled series. Instead, it makes the identity rigorous after pairing with suitable test functions.

Higher dimensions and lattices

The formula extends to Euclidean space Rd\mathbb R^d. Define

f^(ξ)=∫Rdf(x)e−2πix⋅ξ dx.\widehat f(\xi) = \int_{\mathbb R^d} f(x)e^{-2\pi i x\cdot\xi}\,dx.

For f∈S(Rd)f\in\mathcal S(\mathbb R^d),

∑n∈Zdf(n)=∑k∈Zdf^(k).\sum_{n\in\mathbb Z^d}f(n) = \sum_{k\in\mathbb Z^d}\widehat f(k).

More generally, let Λ=AZd\Lambda=A\mathbb Z^d be a full-rank lattice, where AA is an invertible real matrix. Its fundamental-cell volume is

vol⁡(Rd/Λ)=∣det⁡A∣,\operatorname{vol}(\mathbb R^d/\Lambda)=|\det A|,

with det⁡A\det A denoting the determinant. Its dual lattice is

Λ∗={ξ∈Rd:ξ⋅λ∈Z for every λ∈Λ}=A−TZd.\Lambda^* = \{\xi\in\mathbb R^d: \xi\cdot\lambda\in\mathbb Z \text{ for every }\lambda\in\Lambda\} = A^{-T}\mathbb Z^d.

Poisson summation takes the form

∑λ∈Λf(λ)=1∣det⁡A∣∑ξ∈Λ∗f^(ξ).\boxed{ \sum_{\lambda\in\Lambda}f(\lambda) = \frac1{|\det A|} \sum_{\xi\in\Lambda^*}\widehat f(\xi). }

The reciprocal volume factor accounts for the density of lattice points. This formulation underlies reciprocal-lattice descriptions of diffraction, including those used in X-ray crystallography.

Gaussian example and theta transformation

For t>0t>0, the Gaussian

ft(x)=e−πtx2f_t(x)=e^{-\pi t x^2}

has Fourier transform

f^t(ξ)=t−1/2e−πξ2/t.\widehat f_t(\xi) = t^{-1/2}e^{-\pi \xi^2/t}.

Poisson summation therefore gives

∑n∈Ze−πtn2=t−1/2∑k∈Ze−πk2/t.\boxed{ \sum_{n\in\mathbb Z}e^{-\pi t n^2} = t^{-1/2} \sum_{k\in\mathbb Z}e^{-\pi k^2/t}. }

Writing

θ(t)=∑n∈Ze−πtn2,\theta(t)=\sum_{n\in\mathbb Z}e^{-\pi t n^2},

this becomes the theta-function transformation

θ(t)=t−1/2θ(1/t).\theta(t)=t^{-1/2}\theta(1/t).

The identity exchanges small and large values of tt. For numerical evaluation at small tt, the original series may require many terms, while the transformed series decreases rapidly.

Applications

Number theory and the zeta function

Poisson summation is a basic tool in analytic number theory, where it converts sums over integers into dual sums with different analytic properties.

For Re⁡(s)>1\operatorname{Re}(s)>1, the theta series satisfies

∫0∞(θ(t)−1)ts/2−1 dt=2π−s/2Γ(s/2)ζ(s),\int_0^\infty \bigl(\theta(t)-1\bigr)t^{s/2-1}\,dt = 2\pi^{-s/2}\Gamma(s/2)\zeta(s),

where Γ\Gamma is the gamma function and ζ\zeta is the Riemann zeta function.

Splitting this integral at t=1t=1 and applying the theta transformation to the interval (0,1)(0,1) yields analytic continuation and the functional equation

Λ(s)=Λ(1−s),Λ(s)=π−s/2Γ(s/2)ζ(s),\Lambda(s)=\Lambda(1-s), \qquad \Lambda(s)=\pi^{-s/2}\Gamma(s/2)\zeta(s),

interpreted as an identity of meromorphic functions.

Sampling and aliasing

Sampling a function at intervals hh can be represented, under suitable regularity assumptions, by

f III⁡h=∑n∈Zf(nh)δ(x−nh).f\,\operatorname{III}_h = \sum_{n\in\mathbb Z}f(nh)\delta(x-nh).

Its Fourier transform is

f III⁡h^(ξ)=1h∑k∈Zf^ ⁣(ξ−kh).\widehat{f\,\operatorname{III}_h}(\xi) = \frac1h \sum_{k\in\mathbb Z} \widehat f\!\left(\xi-\frac{k}{h}\right).

Thus sampling creates translated copies of the original spectrum. When these copies overlap, different frequencies become indistinguishable in the samples: the phenomenon of aliasing.

This is a foundation of the sampling theorem. For a spectrum supported in [−B,B][-B,B], a sampling rate strictly greater than 2B2B separates the spectral copies; treatment of the critical rate requires additional attention to endpoint behavior.

Numerical quadrature

Since

f^(0)=∫Rf(x) dx,\widehat f(0)=\int_{\mathbb R}f(x)\,dx,

the scaled formula gives the exact error identity

h∑n∈Zf(nh)−∫Rf(x) dx=∑k∈Zk≠0f^(k/h).h\sum_{n\in\mathbb Z}f(nh) - \int_{\mathbb R}f(x)\,dx = \sum_{\substack{k\in\mathbb Z\\k\ne0}} \widehat f(k/h).

This explains why the infinite trapezoidal rule can converge rapidly for smooth, rapidly decreasing functions: its discretization error is determined by high-frequency values of the Fourier transform. For Schwartz functions it decreases faster than every power of hh as h→0h\to0. Exponential convergence requires stronger assumptions, typically involving analyticity and suitable decay.

In practical numerical analysis, the infinite sum must also be truncated, producing a separate truncation error.

Heat kernels and periodic problems

Periodizing the Gaussian solution of the heat equation produces the heat kernel on a circle or a higher-dimensional flat torus. Poisson summation equates two representations:

  • a sum over spatial translates of the Euclidean heat kernel;
  • a Fourier expansion in eigenfunctions of the Laplace operator.

The spatial representation is often effective for short times, while the spectral representation is effective for long times.

Relation to other summation methods

Poisson summation is an exact duality, not merely an approximation. It differs from the Euler–Maclaurin formula, which relates sums and integrals using endpoint derivatives and a remainder. Poisson summation instead expresses the discrepancy through Fourier-transform values.

It is also distinct from the discrete Fourier transform, which acts on finite sequences. Connections between the two arise when both spatial sampling and periodicity are imposed.

More generally, Poisson summation extends to suitable locally compact abelian groups and their character groups. In that setting, the dual lattice is replaced by the subgroup of characters that are trivial on the original lattice, and the normalization is determined by compatible invariant measures.