Motivation
A periodic function has discrete Fourier coefficients, while a function on the real line has a Fourier transform at continuously varying frequencies. Periodization connects the two: add copies of a function at equally spaced shifts and calculate the Fourier coefficients of the resulting periodic function. The Poisson summation formula says that the same periodic function can be reconstructed from equally spaced samples of the transform.
A formal interchange of infinitely many sums and integrals is not a proof. We start with a Schwartz function, whose derivatives decrease faster than any inverse power. That decay controls the spatial sum, the frequency sum and differentiation. Later examples show why a formula written for a much rougher function needs its own hypotheses and a specified interpretation at discontinuities.
Definitions and normalization
Use the symmetric Fourier-transform convention
A Schwartz function is smooth and, for all nonnegative integers , the function is bounded. In particular and its derivatives are integrable. We use angular frequency ; a convention with places the factors of elsewhere.
Definition 2.2.1
(Periodization)
For a period and , define
For in a bounded interval, rapid decay provides a summable bound on the shifted terms, so the series converges absolutely and uniformly there. Reindexing shows that . The same argument applied to every derivative permits termwise differentiation, making a smooth -periodic function.
The periodization sums translates in physical space. Its Fourier-series frequencies are integer multiples of , so the transform will be sampled at precisely those angular frequencies. This is a consequence of the chosen period, not an arbitrary change of variables.
The summation formula
Theorem 2.2.2
(Poisson summation for Schwartz functions)
If and , then for every ,
Both series converge absolutely; the frequency series and its derivatives converge uniformly on . In particular,
At the identity compares two infinite sums of samples; the general identity is stronger because it reconstructs the entire periodized function. The factor is required by the symmetric normalization above. It disappears from some other common conventions because those conventions define a different transform.
Quick check
(For , at which angular frequencies does the theorem sample , and which factor multiplies the frequency sum?)
Read both off the general formula with .
Solution · Answer
The samples are taken at for , and the factor is . Sampling at the integers themselves belongs to a transform with phase , not to the convention fixed in this section.
Proof by periodization
Proof
(Calculate the coefficients before reconstructing)
For , the Schwartz bounds imply, for every integer and all sufficiently large ,
The finitely many remaining terms are bounded on . The Weierstrass test gives uniform convergence of the series for every derivative. The Fourier-series coefficient of at frequency is therefore
Uniform convergence justifies the interchange. In the th integral set . Since , the phase becomes . The shifted intervals tile the real line up to endpoints, and absolute integrability permits their integrals to be combined. Thus
To reconstruct the function, we need convergence as well as coefficient formulas. Repeated integration by parts, with vanishing boundary terms, shows that decreases faster than any inverse power: for , , and the right side is bounded. The same estimate with sufficiently large makes finite for every . Hence the frequency series and each differentiated series converge uniformly.
Let . The uniform limit and the periodization are both continuous; uniform convergence permits termwise integration, so they have the same Fourier coefficients. To justify uniqueness without assuming pointwise convergence of arbitrary Fourier series, apply Fejér's theorem to their continuous periodic difference: all its Fejér means are zero, while those means converge uniformly to the difference. Thus everywhere. The proof separates the coefficient calculation from the reconstruction theorem.
Proof X-ray
Why neither infinite sum is merely formal
Three separate facts carry the proof, and each rests on a different input. First, the decay of and its derivatives, through the bound with , makes the translate series converge uniformly on a period; this is what permits integrating it term by term. Second, the coefficient calculation closes because each shift is a whole number of periods: the phase factor equals one, so the th integral is the integral of one fixed function over , and these intervals tile .
Third, the decay of is a separate input. Knowing the coefficients does not make converge; the identity supplies that convergence. The closing move avoids every pointwise convergence theorem for Fourier series: two continuous periodic functions with the same coefficients coincide, by Fejér. The endpoint counterexample below shows where the argument stops for an indicator function. The coefficient formula still holds there, but equals off the integers and at them, so it is not continuous and its values at the integers are not determined by its coefficients.
One useful extension beyond Schwartz functions
The same reconstruction works under weaker, explicit conditions. Suppose and there are such that
Then the Poisson formula above still holds pointwise. Indeed, the decay of makes its translate series uniformly convergent on and ensures . The same coefficient calculation applies, and the decay of makes the sampled frequency series absolutely and uniformly convergent. Both sides are continuous periodic functions with the same coefficients; Fejér uniqueness proves equality. Derivative convergence is not asserted under these weaker assumptions.
For instance, with is not smooth at zero, but it is continuous, decays exponentially, and its transform has the required quadratic decay. Taking gives the convergent identity
The spatial side is a geometric sum; the frequency side follows by pairing with . This example shows that Schwartz regularity is sufficient, not necessary, while also showing how a valid extension needs its own convergence proof.
Gaussian and theta transformation
Theorem 2.2.3
(A Gaussian transform pair)
For , let . Every derivative is a polynomial times the Gaussian, so is Schwartz. Moreover,
Consequently the theta sum satisfies
Proof
(Compute the Gaussian transform and substitute)
Let . The real Gaussian integral gives . Differentiation under the integral is allowed by integrable Gaussian decay. Since , integration by parts gives
Solving this scalar equation with its initial value yields . Divide by for our transform convention. Substitution into Poisson summation with and gives the theta identity; in particular, the factors cancel.
Common mistake
(Mixing two transform conventions)
With the phase and no prefactor, Poisson summation reads . Copying that shape with the symmetric transform of this section, as , is wrong. For the left side is . The Gaussian pair gives , and the theta identity with evaluates the wrong right side as . The correct statement samples at and multiplies by , which returns .
Worked example 2.2.4
(A small parameter becomes a rapidly convergent sum)
For , the identity gives
Since for positive integers, the difference from ten is strictly positive and at most
The original sum has many terms close to one; the transformed expression has a very small tail. This is a statement about exact sums and an analytic truncation bound. A numerical program still needs an assessment of floating-point effects.
Worked example 2.2.5
(A shifted Gaussian produces a Fourier series)
For , substitution of the Gaussian transform at general yields
Pairing the terms and makes the right-hand side real:
The equality holds for every real . The left side makes periodicity visible by reindexing; the right side displays the amplitudes of its periodic frequencies.
Worked example 2.2.6
(Spacing changes the frequency grid)
For and a general spacing , the zero-shift formula is
For example, replacing by doubles the spatial spacing but halves the transform-sampling spacing . The prefactor also changes. It is unsafe to rescale the lattice without rescaling both the sampled frequencies and the normalization.
Through a Mellin integral it yields the symmetry of the Riemann zeta function under s going to 1 minus s.
A natural next question
What is the theta identity used for?
Through a Mellin integral it yields the symmetry of the Riemann zeta function under s going to 1 minus s.
For real , the integrals add up, by Tonelli, to
Splitting at and using on gives an expression unchanged under . This functional equation of belongs to analytic number theory and is not developed here.
Energy of a periodized function
The same coefficient calculation gives an energy identity. If , then
Indeed, both Fourier series converge absolutely and uniformly, so their product may be integrated term by term. The integral of over a period is when and zero otherwise. Substituting the two coefficient formulas gives the factor . This is a periodic inner-product identity; it is not the same as the continuous-frequency Plancherel integral on the whole line. Periodization can make different translates overlap, so generally differs from .
For the Gaussian with , the identity reads
The square on the left contains cross terms between translates; replacing it by a sum of their individual squares would discard those terms.
Why hypotheses matter
Counterexample mode
A pointwise endpoint convention changes one side
Take , which is integrable but not Schwartz. At integer arguments its values sum to . Its integral transform is
At , it equals for and zero for every other integer . Thus the right side of the formal formula is , not . Both displayed sample sums are finite here; the mismatch is not caused by rearranging a divergent series.
Changing at the two endpoints to their midpoint values makes the spatial sum equal one, but this example alone does not prove a general endpoint-corrected formula. The integral transform ignores changes on null sets, whereas pointwise samples can detect them. Any extension beyond the Schwartz class must say which representative or limiting convention it uses.
Common mistake
(Assuming integrability controls samples)
Even a continuous integrable function need not have a convergent sum over its integer values. For define disjoint triangular spikes
The sum is locally finite and continuous, its total integral is , but for every positive integer. Therefore diverges. The assumption alone cannot justify the pointwise lattice sum, much less all exchanges used in the proof.
Summary
Periodization turns spatial translates into an -periodic function; its Fourier coefficients are transform samples at . For Schwartz functions rapid decay and Fejér uniqueness make the identity valid pointwise. The Gaussian yields the theta transformation. For rough functions, convergence and representative values must be examined separately rather than inferred from the formal formula.
Exercises
Quick check
(Using the theorem, write Poisson summation for at in the stated transform convention.)
Solution · Answer
The identity is . The frequency spacing is , and the prefactor is .
Quick check
(Verify that . Explain which side has faster Gaussian decay in its summation index.)
Solution · Answer
The theta formula with gives the equality. The terms on the left decay faster than on the right. The factor belongs to the complete right-hand sum.
Quick check
(Let . Compute both sides of the formal identity. Why does agreement in this example not prove the formula for all integrable functions?)
Solution · Answer
On the left, only the integer zero is included, so the sum is one. The transform is the same as for because the functions differ only at one point; its samples are zero except at , giving one on the right. The closed-interval counterexample has the same transform but a different left side; the continuous spike example also shows that the spatial sum can diverge. Thus neither point values nor convergence are automatic for arbitrary integrable functions.
Quick check
(For and real , apply Poisson summation to at . State the resulting shifted-frequency identity.)
Solution
Modulation preserves the Schwartz class: each derivative of is a finite sum of bounded oscillatory factors times derivatives of . It shifts the transform: . Hence
Both sums converge absolutely for ; the result is real even though its left side is written with complex exponentials, as pairing and shows.