21.15.01 · number-theory / exponential-sums

Poisson and Voronoi Summation

shipped3 tiersLean: none

Anchor (Master): Iwaniec-Kowalski 2004 *Analytic Number Theory* (AMS Colloquium 53) Ch. 4 (Poisson summation, the Voronoi summation formula for $d(n)$ and for Fourier coefficients of modular forms, the Bessel kernel); Voronoi 1904 *Ann. Sci. École Norm. Sup.* (originator — the summation formula for $d(n)$ with the $Y_0, K_0$ kernel); Iwaniec 2002 *Spectral Methods of Automorphic Forms* (AMS GSM 53, 2nd ed.) Ch. 4 (Voronoi for cusp forms via the functional equation of the additively-twisted $L$-function); Huxley 1996 *Area, Lattice Points, and Exponential Sums* (Oxford) Ch. 1-2 (the lattice-point and circle problems)

Intuition Beginner

Suppose you want to add up the values of a smooth bump-shaped function at every whole number: its value at $0$ , plus its value at $1$ and at $- 1$ , plus at $2$ and $- 2$ , and so on. If the bump is wide and gentle, this is a long, slowly changing sum that is hard to estimate term by term. Poisson summation says something surprising: this sum equals a second sum of exactly the same kind, but taken over the frequencies of the function rather than over its values. The two sums are equal, term for term they look completely different, yet they add up to the same number.

Why is this useful? Because one of the two sums is almost always easier than the other. A bump that is wide and gentle in position is narrow and concentrated in frequency, so the position sum has many large terms while the frequency sum has only a few. You add up whichever side has fewer significant pieces. This trade is the engine behind a long list of results: it turns the count of lattice points inside a circle into a manageable sum, it produces the hidden symmetry of the theta function, and through that symmetry it gives the reflection law of the Riemann zeta function from 21.03.01.

Voronoi summation is the same idea pushed one level further. Instead of adding a smooth function over the integers, you add it weighted by an arithmetic quantity — for instance the number of divisors of each integer. The weighted sum again equals a transformed sum, but now the transform is built from a special oscillating function (a Bessel function) instead of a plain Fourier transform. The arithmetic of the weights is carried into the kernel of the new sum.

Visual Beginner

Picture two number lines stacked vertically. On the top line, place a smooth bell-shaped bump and mark its heights at every integer point $\dots, - 2, - 1, 0, 1, 2, \dots$ ; these heights are the terms of the position sum. On the bottom line, draw the same bump after it has been run through the frequency machine — for a wide bump on top, this is a narrow spike on the bottom — and mark its heights at every integer frequency $\dots, - 2, - 1, 0, 1, 2, \dots$ ; these are the terms of the frequency sum. An equals sign joins the two stacks: the total of the top marks equals the total of the bottom marks.

The picture shows the central trade: width on one line becomes narrowness on the other, and the formula lets you compute a hard sum on the wide side by reading off an easy sum on the narrow side.

Worked example Beginner

We check Poisson summation numerically on a Gaussian bump, where both sides can be computed by hand.

Step 1. Take the function $f (x) = e^{- π x^{2}}$ . Its frequency version (its Fourier transform) is the same Gaussian, $e^{- π y^{2}}$ — a fact special to this bump. So the position sum and the frequency sum are numerically the same sum, which is a clean self-check rather than a coincidence to chase.

Step 2. Compute the position sum, which is $f (0) + 2 f (1) + 2 f (2) + \dots$ (the factor of two pairs each positive integer with its negative), using $f (0) = 1$ , $f (1) = e^{- π} = 0.0432$ , $f (2) = e^{- 4 π} = 0.00000349$ . The total is $1 + 2 (0.0432) + 2 (0.00000349) + \dots = 1 + 0.0864 + 0.0000070 = 1.08641$ to five places. Terms past $n = 2$ are far too small to matter.

Step 3. Now make the bump wider: replace $f$ by $g (x) = e^{- π x^{2} /100}$ , ten times as wide. Its frequency version is $10 e^{- 100 π y^{2}}$ , ten times as tall and one hundred times as narrow. The position sum of $g$ over the integers now has many comparable terms (each near $1$ for small $n$ ), but the frequency sum, namely $10 (1 + 2 e^{- 100 π} + \dots) = 10.0000$ , collapses almost entirely to its single $m = 0$ term.

Step 4. So Poisson summation predicts that the position sum of $g$ over the integers equals $10.0000$ . Adding the position side directly would require dozens of terms; the frequency side needs one.

What this tells us: the formula is a calculator's friend. Whenever a sum over a wide, slow function resists direct addition, transform it and add the few large terms on the other side. The wider and slower the original, the more dramatic the saving.

Check your understanding Beginner

Exercise (easy, multiple choice).

Poisson summation relates a sum of a smooth function over the integers to:

A. The integral of the function over the whole real line
B. A sum of the function's Fourier transform over the integers
C. The Taylor series of the function at the origin
D. The sum of the function's derivatives at the integers

Hint

The formula trades a sum in the position picture for a sum in the frequency picture; the frequency picture is given by the Fourier transform.

Answer

B. Poisson summation states that the sum of $f$ over the integers equals the sum of its Fourier transform $\hat{f}$ over the integers. Option A is only the single $m = 0$ term of the right-hand side (the Fourier transform at frequency zero is the integral of $f$ ), not the whole sum. Option C confuses the Fourier expansion with the Taylor expansion — different bases. Option D is unrelated; derivatives at integers do not enter the formula.

Formal definition Intermediate+

Throughout, the Fourier transform of $f \in S (R)$ is taken in the convention of 02.10.04, $$ \hat f(\xi) = \int_{\mathbb{R}} f(x), e^{-2\pi i \xi x}, dx, $$ under which the Gaussian $e^{- π x^{2}}$ is its own transform and Plancherel holds with no $2 π$ factor.

Definition (periodization). For $f \in S (R)$ the periodization of $f$ is $$ P f(x) := \sum_{n \in \mathbb{Z}} f(x + n). $$ The Schwartz decay of $f$ makes the series converge absolutely and uniformly on compacta, so $P f$ is a smooth function on $R$ of period $1$ . Its Fourier coefficients are $P f (m) = \int_{0}^{1} P f (x) e^{- 2 π im x} d x = \hat{f} (m)$ , obtained by unfolding the integral over $[0, 1)$ against the sum into a single integral over $R$ .

Definition (Poisson summation formula). For $f \in S (R)$ , $$ \sum_{n \in \mathbb{Z}} f(n) = \sum_{m \in \mathbb{Z}} \hat f(m). $$ More generally, evaluating the periodization at an arbitrary point $x$ gives the shifted (or pointwise) form $$ \sum_{n \in \mathbb{Z}} f(x + n) = \sum_{m \in \mathbb{Z}} \hat f(m), e^{2\pi i m x}, $$ and rescaling $f (x) \mapsto f (t x)$ for $t > 0$ gives the weighted form $\sum_{n} f (t n) = t^{- 1} \sum_{m} \hat{f} (m / t)$ .

Definition (lattice form). Let $Λ \subset R^{d}$ be a full-rank lattice with covolume $covol (Λ) = ∣ det Λ∣$ , and let $Λ^{*} = {y \in R^{d} : ⟨ y, λ ⟩ \in Z for all λ \in Λ}$ be its dual lattice. For $f \in S (R^{d})$ , $$ \sum_{\lambda \in \Lambda} f(\lambda) = \frac{1}{\operatorname{covol}(\Lambda)} \sum_{\mu \in \Lambda^} \hat f(\mu). $$ The one-dimensional formula is the case $\Lambda = \mathbb{Z} = \Lambda^ $, co v o l u m e$ 1$.

Definition (divisor function and its Dirichlet series). The divisor function $d (n) = \sum_{ab = n} 1$ counts the positive divisors of $n$ ; its generating Dirichlet series is $\sum_{n \geq 1} d (n) n^{- s} = ζ (s)^{2}$ for $Re (s) > 1$ , where $ζ$ is the Riemann zeta function of 21.03.01. The Voronoi summation formula for $d (n)$ expresses a smoothed sum $\sum_{n} d (n) g (n)$ in terms of a dual sum $\sum_{n} d (n) \tilde{g} (n)$ whose kernel $\tilde{g}$ is a Hankel-type Bessel transform of $g$ ; the precise statement is the Key result below.

Counterexamples to common slips

*"Poisson summation holds for any continuous $f$ with $\sum ∣ f (n) ∣ < \infty$ ."* Absolute convergence of the left side alone is not enough. The standard sufficient hypothesis (Stein-Shakarchi) is that $f$ is continuous and both $f$ and $\hat{f}$ satisfy a decay bound $∣ f (x) ∣, ∣ \hat{f} (ξ) ∣ \leq C (1 + ∣ x ∣)^{- 1 - δ}$ for some $δ > 0$ ; Schwartz functions satisfy this comfortably. Without control on $\hat{f}$ , the periodization can fail to equal its Fourier series pointwise, and the two sides need not agree.
"$\Lambda^ = \Lambda $a l w a y s ." * T h e d u a l l a tt i cee q u a l s t h e l a tt i ceo n l y f or se l f - d u a l l a tt i ces (e . g .$ \mathbb{Z}^d $, or$ E_8 $) . F or$ \Lambda = t\mathbb{Z} $t h e d u a l i s$ t^{-1}\mathbb{Z} $, an d t h eco v o l u m e f a c t or$ 1/\operatorname{covol}(\Lambda) = t^{-1} $i se x a c tl y w ha t mak es t h e w e i g h t e df or m$ \sum_n f(tn) = t^{-1}\sum_m \hat f(m/t)$ correct. Dropping the covolume factor is the most common error.
"Voronoi summation is just Poisson summation with $d (n)$ inserted." It is not a special case of Poisson. The divisor weights $d (n)$ do not come from a lattice; the formula instead descends from the functional equation of $ζ (s)^{2}$ , and the self-dual Fourier kernel of Poisson is replaced by a Bessel-function kernel reflecting the $Γ$ -factors of that functional equation. The structural parent is the functional equation, not periodization.

Key result with proof Intermediate+

The signature theorem is Poisson summation; its first arithmetic consequence is the theta-function transformation, from which the functional equation of $ζ$ of 21.03.01 follows.

Theorem (Poisson summation; Poisson 1823). Let $f \in S (R)$ . Then $$ \sum_{n \in \mathbb{Z}} f(n) = \sum_{m \in \mathbb{Z}} \hat f(m), \qquad \hat f(\xi) = \int_{\mathbb{R}} f(x) e^{-2\pi i \xi x}, dx. $$

Proof. Form the periodization $F (x) = \sum_{n \in Z} f (x + n)$ . Because $f$ is Schwartz, for each $N$ there is $C_{N}$ with $∣ f (x) ∣ \leq C_{N} (1 + ∣ x ∣)^{- N}$ ; taking $N = 2$ makes $\sum_{n} sup_{x \in [0, 1]} ∣ f (x + n) ∣ < \infty$ , so the series defining $F$ converges absolutely and uniformly on $[0, 1]$ , and the same bound applied to derivatives shows $F \in C^{\infty}$ . By construction $F (x + 1) = F (x)$ , so $F$ is a smooth $1$ -periodic function and equals its own Fourier series.

The $m$ -th Fourier coefficient of $F$ is $$ c_m = \int_0^1 F(x) e^{-2\pi i m x}, dx = \int_0^1 \sum_{n \in \mathbb{Z}} f(x+n), e^{-2\pi i m x}, dx. $$ Uniform convergence permits interchanging sum and integral. In the $n$ -th term substitute $u = x + n$ , under which $e^{- 2 π im x} = e^{- 2 π im (u - n)} = e^{- 2 π im u}$ since $e^{2 π imn} = 1$ , and the interval $[0, 1)$ maps to $[n, n + 1)$ : $$ c_m = \sum_{n \in \mathbb{Z}} \int_{n}^{n+1} f(u), e^{-2\pi i m u}, du = \int_{\mathbb{R}} f(u), e^{-2\pi i m u}, du = \hat f(m). $$ Smoothness of $F$ makes its Fourier series converge to $F$ pointwise and absolutely [from 02.10.04, applied on the circle], so in particular at $x = 0$ : $$ \sum_{n \in \mathbb{Z}} f(n) = F(0) = \sum_{m \in \mathbb{Z}} c_m = \sum_{m \in \mathbb{Z}} \hat f(m). \qquad \square $$

Corollary (theta transformation). For $t > 0$ let $Θ (t) = \sum_{n \in Z} e^{- π n^{2} t}$ . Then $Θ (1/ t) = t^{1/2} Θ (t)$ .

Proof. Apply Poisson summation to $f (x) = e^{- π x^{2} t}$ . By the Gaussian self-transform and the dilation rule of 02.10.04, $\hat{f} (ξ) = t^{- 1/2} e^{- π ξ^{2} / t}$ . Poisson gives $\sum_{n} e^{- π n^{2} t} = t^{- 1/2} \sum_{m} e^{- π m^{2} / t}$ , i.e. $Θ (t) = t^{- 1/2} Θ (1/ t)$ , which rearranges to the claim. $□$

This corollary is the precise input that drives the symmetric functional equation $ξ (s) = ξ (1 - s)$ of the completed zeta function in 21.03.01: the Mellin transform of $Θ (t) - 1$ produces $π^{- s /2} Γ (s /2) ζ (s)$ , and the $t \mapsto 1/ t$ symmetry of $Θ$ becomes the $s \mapsto 1 - s$ symmetry of $ξ$ .

Bridge. Poisson summation builds toward the analytic theory of $L$ -functions, and it appears again in the Voronoi formula below, where the self-dual Fourier kernel is replaced by a Bessel kernel. The foundational reason a functional equation exists at all is the self-duality of the lattice $Z$ under the Fourier transform on $R$ : Poisson summation is the analytic shadow of Pontryagin duality $R / Z = Z$ , and this is exactly the identity that the proof of the zeta functional equation in 21.03.01 secretly uses when it invokes the theta transformation. The same template generalises from the constant weights of Poisson to the divisor weights of Voronoi and to the Hecke eigenvalues of modular forms: in each case a summation formula is dual to the functional equation of a Dirichlet series, and the kernel of the dual sum records the $Γ$ -factors. Putting these together, the entire subject of summation formulae is the harmonic-analytic face of the functional equations studied in 21.03.01, with the central insight that symmetry of a continuous transform forces an identity between two discrete sums.

Exercises Intermediate+

Exercise 3 (medium, symbolic).

Let $f \in S (R)$ . Prove the shifted Poisson formula $\sum_{n} f (x + n) = \sum_{m} \hat{f} (m) e^{2 π im x}$ for every $x \in R$ , and recover the unweighted formula as the case $x = 0$ .

Hint

The periodization $F (x) = \sum_{n} f (x + n)$ is smooth and $1$ -periodic with Fourier coefficients $\hat{f} (m)$ . Expand $F$ in its Fourier series and evaluate at general $x$ .

Answer

As in the proof of the theorem, $F (x) = \sum_{n} f (x + n)$ is a smooth $1$ -periodic function with Fourier coefficients $c_{m} = \hat{f} (m)$ . Smoothness makes the Fourier series converge to $F$ pointwise and uniformly, so for every $x$ , $$ \sum_n f(x+n) = F(x) = \sum_m c_m e^{2\pi i m x} = \sum_m \hat f(m), e^{2\pi i m x}. $$ Setting $x = 0$ kills the exponential factor and returns $\sum_{n} f (n) = \sum_{m} \hat{f} (m)$ . The shifted form is strictly more information than the unweighted one: it gives the full Fourier expansion of the periodization, not merely its value at the origin, and it is the version needed when an additive character $e (na / c)$ must be carried through the summation (as in the Voronoi derivation).

Exercise 4 (medium, symbolic).

Prove the lattice Poisson formula $\sum_{λ \in Λ} f (λ) = covol (Λ)^{- 1} \sum_{μ \in Λ^{*}} \hat{f} (μ)$ for a full-rank lattice $Λ \subset R^{d}$ and $f \in S (R^{d})$ , by reducing to the standard lattice $Z^{d}$ .

Hint

Write $Λ = A Z^{d}$ for an invertible matrix $A$ with $∣ det A ∣ = covol (Λ)$ , set $g = f \circ A$ , and use $Λ^{*} = (A^{- T}) Z^{d}$ together with the change-of-variables transform rule $\overset{g}{^} (η) = ∣ det A ∣^{- 1} \hat{f} (A^{- T} η)$ .

Answer

Write $Λ = A Z^{d}$ with $A \in GL_{d} (R)$ and $∣ det A ∣ = covol (Λ)$ . Put $g (u) = f (A u)$ , so that $\sum_{λ \in Λ} f (λ) = \sum_{n \in Z^{d}} g (n)$ . The multivariate dilation rule gives $\overset{g}{^} (η) = ∣ det A ∣^{- 1} \hat{f} (A^{- T} η)$ , where $A^{- T} = (A^{- 1})^{T}$ . The $d$ -dimensional Poisson formula on $Z^{d}$ (proved by periodizing over the unit cube and expanding in the multivariate Fourier series, exactly as in dimension one) reads $\sum_{n} g (n) = \sum_{k \in Z^{d}} \overset{g}{^} (k)$ . Substituting, $$ \sum_{\lambda\in\Lambda} f(\lambda) = \sum_{k\in\mathbb{Z}^d} |\det A|^{-1}\hat f(A^{-\mathsf{T}}k). $$ As $k$ runs over $Z^{d}$ , the points $μ = A^{- T} k$ run over $(A^{- T}) Z^{d}$ , which is precisely the dual lattice $Λ^{*}$ (since $⟨ A^{- T} k, A n ⟩ = ⟨ k, n ⟩ \in Z$ ). Hence $\sum_{λ} f (λ) = ∣ det A ∣^{- 1} \sum_{μ \in Λ^{*}} \hat{f} (μ) = covol (Λ)^{- 1} \sum_{μ \in Λ^{*}} \hat{f} (μ)$ .

Exercise 5 (medium, numeric).

The Gauss circle problem counts lattice points $N (R) = # {(a, b) \in Z^{2} : a^{2} + b^{2} \leq R^{2}}$ . The main term is $π R^{2}$ . Using the two-dimensional Poisson formula applied to a smooth approximation of the indicator of the disc, the leading correction comes from the $μ = 0$ dual term, which reproduces the area $π R^{2}$ . Compute the predicted main term $N (R) \approx π R^{2}$ for $R = 10$ to the nearest integer, and compare with the true count $N (10) = 317$ .

Hint

Evaluate $π R^{2}$ at $R = 10$ and round. The discrepancy from $317$ is the error term $Δ (R)$ that the nonzero dual frequencies control.

Answer

$314$ . The main term is $π R^{2} = π \cdot 100 = 314.159 \dots \approx 314$ . The exact count is $N (10) = 317$ , so the error is $Δ (10) = 317 - 314.159 = 2.84$ . In the Poisson framework, $N (R)$ is approximated by summing a smoothed disc indicator over $Z^{2}$ ; the dual frequency $μ = 0$ gives $π R^{2}$ (the area), and the nonzero $μ$ contribute oscillating Bessel-type terms whose total is $Δ (R)$ . The Gauss circle problem asks for the true size of $Δ (R)$ : it is known that $Δ (R) = O (R^{2/3})$ (Sierpiński) and conjectured to be $O (R^{1/2 + ε})$ , with the Hardy-Landau $Ω (R^{1/2})$ lower bound showing $1/2$ cannot be improved.

Exercise 6 (medium, multiple choice).

The kernel appearing in the Voronoi summation formula for the divisor function $d (n)$ is built from:

A. The Gaussian, as in the self-dual Poisson kernel
B. Bessel functions $Y_{0}$ and $K_{0}$
C. The sine and cosine functions only
D. The Riemann zeta function evaluated at the dual points

Hint

The Voronoi kernel reflects the $Γ$ -factors in the functional equation of $ζ (s)^{2}$ ; the inverse Mellin transform of a ratio of two $Γ (s /2)$ factors is a Bessel function.

Answer

B. The Voronoi formula for $d (n)$ has a kernel built from the Bessel functions $Y_{0}$ (Bessel function of the second kind, order zero) and $K_{0}$ (modified Bessel function of the second kind, order zero). These arise because the functional equation of $ζ (s)^{2}$ carries the factor $Γ (s /2)^{2}$ (one $Γ (s /2)$ per $ζ$ ), and the inverse Mellin transform of such a squared $Γ$ -ratio is precisely a combination of $Y_{0}$ and $K_{0}$ . The Gaussian (A) is the Poisson kernel and does not occur; pure trigonometric kernels (C) arise for degree-one objects like single $L$ -functions of $ζ$ itself, not the degree-two $ζ^{2}$ ; option D confuses the kernel with the generating series.

Exercise 7 (hard, symbolic).

Derive the theta transformation in the general form $Θ (t) = t^{- 1/2} Θ (1/ t)$ and use it, via the Mellin transform $\int_{0}^{\infty} (Θ (t) - 1) /2 \cdot t^{s /2 - 1} d t = π^{- s /2} Γ (s /2) ζ (s)$ for $Re (s) > 1$ , to sketch how the $t \mapsto 1/ t$ symmetry produces the functional equation of $ζ$ from 21.03.01.

Hint

Split the Mellin integral at $t = 1$ and apply the theta transformation to the integral over $(0, 1]$ after the substitution $t \mapsto 1/ t$ , mirroring the proof in 21.03.01. The leftover elementary integrals produce the poles, and the remaining integral is manifestly $s \mapsto 1 - s$ symmetric.

Answer

The transformation $Θ (t) = t^{- 1/2} Θ (1/ t)$ is the Corollary, from Poisson summation on the Gaussian $e^{- π x^{2} t}$ . Write $ψ (t) = (Θ (t) - 1) /2 = \sum_{n \geq 1} e^{- π n^{2} t}$ . The transformation rearranges to $ψ (1/ t) = t^{1/2} ψ (t) + (t^{1/2} - 1) /2$ .

For $Re (s) > 1$ , the Mellin transform of $ψ$ is $$ \int_0^\infty \psi(t), t^{s/2-1}, dt = \sum_{n\ge 1}\int_0^\infty e^{-\pi n^2 t} t^{s/2-1},dt = \sum_{n\ge 1}\frac{\Gamma(s/2)}{(\pi n^2)^{s/2}} = \pi^{-s/2}\Gamma(s/2)\zeta(s), $$ the substitution $u = π n^{2} t$ giving the $Γ$ -factor. Split at $t = 1$ , and on $(0, 1]$ substitute $t \mapsto 1/ t$ and insert $ψ (1/ t) = t^{1/2} ψ (t) + (t^{1/2} - 1) /2$ : $$ \pi^{-s/2}\Gamma(s/2)\zeta(s) = \int_1^\infty \psi(t)\big(t^{s/2-1} + t^{(1-s)/2-1}\big),dt + \int_1^\infty \frac{t^{1/2}-1}{2}, t^{-s/2-1},dt. $$ The elementary integral evaluates to $\frac{1}{s - 1} - \frac{1}{s}$ , valid by continuation, and the remaining integral over $[1, \infty)$ is invariant under $s \mapsto 1 - s$ because the integrand $t^{s /2 - 1} + t^{(1 - s) /2 - 1}$ is. Absorbing the polar terms into $ξ (s) = \frac{1}{2} s (s - 1) π^{- s /2} Γ (s /2) ζ (s)$ cancels the poles and yields the entire function with $ξ (s) = ξ (1 - s)$ . This is the route of 21.03.01, with the theta transformation — a Poisson-summation identity — supplying the symmetry.

Exercise 8 (hard, symbolic).

State the Voronoi summation formula for $d (n)$ in smoothed form, and explain in two or three sentences why the error term $Δ (x) = \sum_{n \leq x} d (n) - (x lo g x + (2 γ - 1) x)$ in the Dirichlet divisor problem is controlled by it, identifying the exponent Voronoi obtained.

Hint

The Voronoi formula turns $\sum_{n} d (n) g (n)$ into a dual sum over $n$ with a Bessel-kernel weight; truncating against the sharp cutoff $g = 1_{[0, x]}$ and bounding the oscillatory Bessel sum yields a power saving over the elementary $O (x^{1/2})$ .

Answer

Statement. For a test function $g$ smooth and compactly supported in $(0, \infty)$ , $$ \sum_{n\ge 1} d(n), g(n) = \int_0^\infty \big(\log t + 2\gamma\big) g(t),dt + \sum_{n\ge 1} d(n), \tilde g(n), $$ where the dual kernel is the Hankel-type transform $\tilde{g} (n) = \int_{0}^{\infty} g (t) (- 2 π Y_{0} (4 π n t) + 4 K_{0} (4 π n t)) d t$ , with $Y_{0}, K_{0}$ the order-zero Bessel functions. The main integral term reproduces $x lo g x + (2 γ - 1) x$ after integrating against a smoothed indicator.

Why it controls $Δ (x)$ . Replacing $g$ by a smoothed indicator of $[0, x]$ , the dual sum becomes a sum of $d (n)$ against oscillating Bessel kernels $\sim n^{- 1/4} x^{1/4} cos (4 π n x + phase)$ ; the oscillation produces cancellation that a term-by-term bound on $Δ (x)$ cannot see. Voronoi balanced the length of the dual sum against this cancellation to obtain $Δ (x) = O (x^{1/3} lo g x)$ , the first improvement over the elementary $O (x^{1/2})$ from Dirichlet's hyperbola method. The conjectured truth is $Δ (x) = O (x^{1/4 + ε})$ , with Hardy's $Ω (x^{1/4})$ lower bound showing $1/4$ is best possible; the current record is Huxley's $O (x^{131/416 + ε})$ .

Advanced results Master

The Voronoi summation formula as a functional-equation dual

The clean way to see the Voronoi formula for $d (n)$ is through the functional equation of $ζ (s)^{2}$ ^{[Iwaniec-Kowalski 2004]}. The Dirichlet series $\sum_{n} d (n) n^{- s} = ζ (s)^{2}$ inherits a functional equation from $ξ (s) = ξ (1 - s)$ of 21.03.01: writing $Λ_{2} (s) = π^{- s} Γ (s /2)^{2} ζ (s)^{2}$ , one has $Λ_{2} (s) = Λ_{2} (1 - s)$ . The squared $Γ$ -factor $Γ (s /2)^{2}$ is the analytic fingerprint of the divisor convolution $d = 1 * 1$ . Mellin-inverting against a test function, the symmetry $s \mapsto 1 - s$ becomes a summation identity in which the dual kernel is the inverse Mellin transform of the ratio of $Γ$ -factors at $s$ and $1 - s$ . For a single $ζ$ that ratio inverts to elementary trigonometric kernels; for $ζ^{2}$ it inverts to the order-zero Bessel combination $- 2 π Y_{0} (4 π x) + 4 K_{0} (4 π x)$ . This is why divisor sums acquire a Bessel kernel where prime-counting sums acquire only cosines: the kernel reads off the degree and the $Γ$ -factors of the underlying $L$ -function.

Voronoi for modular forms and the additive twist

The same mechanism generalises from $ζ^{2}$ to the standard $L$ -function of a cusp form ^{[Iwaniec 2002]}. Let $f$ be a holomorphic or Maass cusp form with Hecke eigenvalues $λ_{f} (n)$ . The additively-twisted $L$ -function $L (s, f, a / c) = \sum_{n} λ_{f} (n) e (na / c) n^{- s}$ (with $e (z) = e^{2 π i z}$ and $g cd (a, c) = 1$ ) satisfies a functional equation relating $a / c$ to $\overset{a}{ˉ} / c$ , where $a \overset{a}{ˉ} \equiv 1 (mod c)$ . Inverting that functional equation against a test function produces the Voronoi formula $$ \sum_n \lambda_f(n), e!\left(\tfrac{an}{c}\right) g(n) = \frac{1}{c}\sum_n \lambda_f(n), e!\left(-\tfrac{\bar a n}{c}\right) \tilde g!\left(\tfrac{n}{c^2}\right) + (\text{main term}), $$ with $\tilde{g}$ a Hankel transform whose Bessel order is determined by the weight (holomorphic case) or the Laplace eigenvalue (Maass case) of $f$ . The reciprocal residue $\overset{a}{ˉ}$ and the modulus $c^{2}$ in the dual argument are the source of the Kloosterman sums $S (m, n; c) = \sum_{x \overset{x}{ˉ} \equiv 1 (c)} e ((m x + n \overset{x}{ˉ}) / c)$ that appear when one sums these formulae over $c$ ; this is the analytic backbone of the Kuznetsov and Petersson trace formulae and of subconvexity bounds for $L$ -functions.

Lattice-point problems and the discrete Hardy-Littlewood method

For the Gauss circle and Dirichlet divisor problems, Poisson and Voronoi summation convert a lattice-point count into a short sum of exponential sums ^{[Huxley 1996]}. Counting $\sum_{n \leq x} d (n)$ is, by the hyperbola method, a count of lattice points under $uv = x$ ; Voronoi summation replaces the sharp count by a smooth main term plus a Bessel sum, and stationary phase turns each Bessel factor into an exponential sum $\sum_{n} d (n) e (2 n x)$ . The size of $Δ (x)$ is then a question about cancellation in such sums, attacked by van der Corput's method and its modern refinement, the Bombieri-Iwaniec-Huxley discrete Hardy-Littlewood method, giving the current record $Δ (x) = O (x^{131/416 + ε})$ . The exact analogue holds for the circle problem with $d (n)$ replaced by $r_{2} (n)$ , the number of representations as a sum of two squares; the same $Ω (x^{1/4})$ lower bound of Hardy-Landau caps both.

Synthesis. Poisson summation, the theta transformation, and the Voronoi formulae are one mechanism wearing three masks, and the bridge is the functional equation: each summation formula is dual to the functional equation of a Dirichlet series, and the kernel of the dual sum is exactly what the $Γ$ -factors of that functional equation Mellin-invert to. This is the foundational reason the constant weights of Poisson give a self-dual Fourier kernel while the divisor weights of Voronoi give a Bessel kernel — the central insight being that the analytic continuation and functional equation studied in 21.03.01 are precisely the data that a summation formula makes discrete. Putting these together, the subject generalises in two directions at once: upward in degree, from $ζ$ to $ζ^{2}$ to degree- $d$ automorphic $L$ -functions, where the Bessel order grows with the degree; and outward in the lattice, from $Z$ to general $Λ$ with its dual $Λ^{*}$ , where Poisson is dual to the self-duality of Euclidean space. The lattice-point and circle problems are exactly this machinery turned on the indicator of a region, and the appearance again of Bessel-kernel exponential sums in the modular-form Voronoi formula is what ties the elementary divisor problem to the spectral theory of automorphic forms.

Full proof set Master

Proposition 1 (covolume normalization is forced). In the lattice Poisson formula, the constant $C$ in $\sum_{\lambda\in\Lambda} f(\lambda) = C\sum_{\mu\in\Lambda^}\hat f(\mu) $m u s t e q u a l$ \operatorname{covol}(\Lambda)^{-1}$.*

Proof. Test the asserted identity on the family $f = f_{σ}$ , a Gaussian of covariance $σ^{2} I$ scaled so that $\int_{R^{d}} f_{σ} = 1$ for every $σ$ ; then $\hat{f}_{σ} (μ) \to 1$ pointwise as $σ \to \infty$ while $f_{σ} (λ) \to 0$ for $λ \neq = 0$ . As $σ \to \infty$ the left side tends to $f_{σ} (0) \to$ the peak value, but it is cleaner to take $σ \to 0$ : then $f_{σ} \to δ_{0}$ as a measure, the left side tends to $f_{σ} (0)$ summed, dominated by the $λ = 0$ term, and a Riemann-sum comparison shows $covol (Λ) \sum_{λ} f_{σ} (λ) \to \int f_{σ} = 1$ , so $\sum_{λ} f_{σ} (λ) \sim covol (Λ)^{- 1}$ . Meanwhile $\hat{f}_{σ} \to 1$ uniformly on compacta forces $\sum_{μ} \hat{f}_{σ} (μ) \to 1$ after regularization. Matching the two limits forces $C = covol (Λ)^{- 1}$ . The rigorous version is the reduction-to- $Z^{d}$ change of variables of Exercise 4, where the Jacobian $∣ det A ∣ = covol (Λ)$ appears explicitly; the heuristic limit confirms the sign and placement of the factor. $□$

Proposition 2 (theta transformation is equivalent to Gaussian self-duality under Poisson). The identity $Θ (1/ t) = t^{1/2} Θ (t)$ holds for all $t > 0$ if and only if the Gaussian $e^{- π x^{2}}$ is its own Fourier transform.

Proof. ( $\Leftarrow$ ) Given $e^{- π x^{2}} = e^{- π ξ^{2}}$ , the dilation rule gives $e^{- π t x^{2}} (ξ) = t^{- 1/2} e^{- π ξ^{2} / t}$ ; Poisson summation then yields $Θ (t) = t^{- 1/2} Θ (1/ t)$ , which is the claim. ( $\Rightarrow$ ) Conversely, suppose $Θ (1/ t) = t^{1/2} Θ (t)$ for all $t$ . Poisson summation applied to $e^{- π t x^{2}}$ gives $Θ (t) = \sum_{m} e^{- π t x^{2}} (m)$ regardless of what the transform is; equating with the assumed transformation and varying $t$ over a set with a limit point forces $e^{- π t x^{2}} (ξ) = t^{- 1/2} e^{- π ξ^{2} / t}$ at $ξ \in Z$ , and by analyticity of both sides in $ξ$ (each is entire of order two) the equality extends to all $ξ$ ; at $t = 1$ this is the self-duality. $□$

Proposition 3 (Bessel kernel from the squared Gamma factor). The inverse Mellin transform $\frac{1}{2 π i} \int_{(σ)} (π^{- 1} Γ (s /2)^{2} /Γ ((1 - s) /2)^{2}) x^{- s} d s$ produces a kernel expressible through the order-zero Bessel functions $Y_{0}$ and $K_{0}$ , accounting for the form of the Voronoi kernel.

Proof sketch (the full evaluation is classical). The functional equation of $ζ (s)^{2}$ contributes the factor $γ (s) = π^{- (2 s - 1) /2} Γ (s /2)^{2} /Γ ((1 - s) /2)^{2}$ on inversion. Using the duplication and reflection formulae for $Γ$ to rewrite $γ (s)$ in terms of $Γ (s)$ and $sin / cos$ factors, the contour integral $\frac{1}{2 π i} \int_{(σ)} γ (s) x^{- s} d s$ matches the Mellin-Barnes representation of the Bessel functions: $K_{0} (z) = \frac{1}{2} \cdot \frac{1}{2 π i} \int_{(σ)} Γ (s /2)^{2} (z /2)^{- s} d s$ for $σ > 0$ , and $Y_{0}$ arises from the companion contour picking up the trigonometric residues. The combination $- 2 π Y_{0} (4 π x) + 4 K_{0} (4 π x)$ is exactly the residue-plus-integral decomposition of the inverted $γ (s)$ , with the $4 π x$ argument coming from the $x^{- s}$ scaling against the $π^{- s} n^{- s}$ from $ζ (s)^{2}$ . Thus the Bessel kernel is forced by the squared $Γ$ -factor, which is itself forced by $d = 1 * 1$ . $□$

Connections Master

The theta transformation proved here is the precise harmonic-analytic input to the functional equation of the Riemann zeta function in 21.03.01: the Mellin transform of $Θ (t) - 1$ is $π^{- s /2} Γ (s /2) ζ (s)$ , and the $t \mapsto 1/ t$ symmetry from Poisson summation becomes the $s \mapsto 1 - s$ symmetry of the completed zeta function.

The entire formula rests on the Fourier transform on $R^{n}$ and the Gaussian self-duality established in 02.10.04; Poisson summation is the discretization of Plancherel duality, pairing the lattice $Z$ against its Pontryagin dual, and the Schwartz-space decay hypotheses are exactly those under which the periodization equals its Fourier series.

The Voronoi formula for the divisor function $d (n)$ , with its Dirichlet series $ζ (s)^{2}$ , is the degree-two instance of the functional-equation-dual mechanism whose degree-one instance is the single zeta function of 21.03.01; the Bessel kernel here is the trace of the squared $Γ$ -factor that distinguishes $ζ^{2}$ from $ζ$ .

Historical & philosophical context Master

The summation formula appears in Poisson's 1823 memoir on definite integrals and the summation of series ^{[Poisson 1823]}, where it served his study of the convergence of trigonometric series; the modern statement as an identity between $\sum_{n} f (n)$ and $\sum_{m} \hat{f} (m)$ for Schwartz $f$ is a twentieth-century reformulation in the language of the Fourier transform on $R$ . Jacobi's theta transformation, which the formula yields in one line for the Gaussian, predates the Fourier-analytic proof and was known to Jacobi in the 1820s through the theory of elliptic functions; Riemann's 1859 use of it to derive the zeta functional equation is the source of its centrality in analytic number theory, as recorded in 21.03.01.

The weighted formula for the divisor function is due to Georgy Voronoi, in two long papers in the Annales de l'École Normale Supérieure of 1904 ^{[Voronoi 1904]}, where he introduced the transcendental function (the Bessel combination) carrying his name and used it to obtain $Δ (x) = O (x^{1/3} lo g x)$ , improving Dirichlet's elementary $O (x^{1/2})$ . The lower bounds $Δ (x) = Ω (x^{1/4})$ for both the divisor and circle problems were established by Hardy and Landau in 1915-1916 ^{[Hardy Landau 1916]}, fixing the conjectural exponent at $1/4$ . The reinterpretation of Voronoi's formula as the functional equation of an additively-twisted $L$ -function, which exposes its generalization to automorphic forms and its link to Kloosterman sums, is a development of the spectral theory of automorphic forms in the second half of the twentieth century, consolidated in Iwaniec's treatment ^{[Iwaniec 2002]}.

Bibliography Master

@article{poisson1823,
  author  = {Poisson, Sim{\'e}on Denis},
  title   = {Sur les int{\'e}grales d{\'e}finies et sur la sommation des s{\'e}ries},
  journal = {Journal de l'{\'E}cole Polytechnique},
  volume  = {12},
  number  = {19},
  pages   = {404--509},
  year    = {1823}
}

@article{voronoi1904,
  author  = {Voronoi, Georgy},
  title   = {Sur une fonction transcendante et ses applications {\`a} la sommation de quelques s{\'e}ries},
  journal = {Annales Scientifiques de l'{\'E}cole Normale Sup{\'e}rieure},
  series  = {3},
  volume  = {21},
  pages   = {207--267, 459--533},
  year    = {1904}
}

@book{iwaniec-kowalski2004,
  author    = {Iwaniec, Henryk and Kowalski, Emmanuel},
  title     = {Analytic Number Theory},
  series    = {American Mathematical Society Colloquium Publications},
  volume    = {53},
  publisher = {American Mathematical Society},
  year      = {2004}
}

@book{iwaniec2002,
  author    = {Iwaniec, Henryk},
  title     = {Spectral Methods of Automorphic Forms},
  series    = {Graduate Studies in Mathematics},
  volume    = {53},
  edition   = {2},
  publisher = {American Mathematical Society},
  year      = {2002}
}

@book{huxley1996,
  author    = {Huxley, Martin N.},
  title     = {Area, Lattice Points, and Exponential Sums},
  series    = {London Mathematical Society Monographs (New Series)},
  volume    = {13},
  publisher = {Oxford University Press},
  year      = {1996}
}

@article{hardy1916divisor,
  author  = {Hardy, Godfrey Harold},
  title   = {On Dirichlet's divisor problem},
  journal = {Proceedings of the London Mathematical Society},
  series  = {2},
  volume  = {15},
  pages   = {1--25},
  year    = {1916}
}

Prerequisites

02.10.04
21.03.01

Tier anchors

beginner: Stein-Shakarchi 2003 *Fourier Analysis: An Introduction* (Princeton Lectures in Analysis I) Ch. 5 §3 (Poisson summation and the theta function); Körner 1988 *Fourier Analysis* (Cambridge) Ch. 18 informal opening on the summation formula
intermediate: Iwaniec-Kowalski 2004 *Analytic Number Theory* (AMS Colloquium 53) Ch. 4 (the Poisson and Voronoi summation formulae); Stein-Shakarchi 2003 *Fourier Analysis* Ch. 5 §3 (Poisson summation via periodization); Montgomery 1994 *Ten Lectures on the Interface Between Analytic Number Theory and Harmonic Analysis* (CBMS 84) Ch. 1
master: Iwaniec-Kowalski 2004 *Analytic Number Theory* (AMS Colloquium 53) Ch. 4 (Poisson summation, the Voronoi summation formula for $d(n)$ and for Fourier coefficients of modular forms, the Bessel kernel); Voronoi 1904 *Ann. Sci. École Norm. Sup.* (originator — the summation formula for $d(n)$ with the $Y_0, K_0$ kernel); Iwaniec 2002 *Spectral Methods of Automorphic Forms* (AMS GSM 53, 2nd ed.) Ch. 4 (Voronoi for cusp forms via the functional equation of the additively-twisted $L$-function); Huxley 1996 *Area, Lattice Points, and Exponential Sums* (Oxford) Ch. 1-2 (the lattice-point and circle problems)

References

Poisson, S. D. — Sur les intégrales définies et sur la sommation des séries · *Journal de l'École Polytechnique*, cahier 19, tome 12 (1823), 404-509. The origin of the summation formula relating $\sum_n f(n)$ to $\sum_m \hat f(m)$; the formula appears in Poisson's study of the convergence of trigonometric series and of definite integrals.
Voronoi, G. — Sur une fonction transcendante et ses applications à la sommation de quelques séries · *Annales Scientifiques de l'École Normale Supérieure* (3) 21 (1904), 207-267 and 459-533. The summation formula for the divisor function $d(n)$, $\sum_{n\le x} d(n) = x\log x + (2\gamma-1)x + 1/4 + \Delta(x)$ with $\Delta(x)$ expanded in a series of Bessel functions $Y_0, K_0$; the first non-trivial bound $\Delta(x) = O(x^{1/3}\log x)$ on the error term in the Dirichlet divisor problem.
Iwaniec, H. & Kowalski, E. — Analytic Number Theory · American Mathematical Society Colloquium Publications 53 (2004), Ch. 4. Poisson summation on $\mathbb{R}^n$ and on lattices, the truncated/smoothed Poisson formula, the Voronoi summation formula for $d(n)$ and for Hecke eigenvalues $\lambda_f(n)$ of holomorphic and Maass cusp forms, with the Bessel-function kernels; applications to lattice-point counting and to bounds for moments of $L$-functions.
Hardy, G. H. & Landau, E. — On the Dirichlet divisor problem / Über die Gitterpunkte in einem Kreise · Hardy, *Proc. London Math. Soc.* (2) 15 (1916), 1-25 (the $\Omega$-result $\Delta(x) = \Omega(x^{1/4})$ for the divisor and circle problems); Landau, *Göttinger Nachrichten* (1915), 148-160. The lower bounds matching the conjectured exponent $1/4$ in the Dirichlet divisor and Gauss circle problems.
Iwaniec, H. — Spectral Methods of Automorphic Forms · American Mathematical Society, Graduate Studies in Mathematics 53, 2nd edition (2002), Ch. 4. The Voronoi summation formula for the Fourier coefficients of automorphic forms, derived from the functional equation of the additively-twisted $L$-function $L(s, f, a/c) = \sum_n \lambda_f(n) e(na/c) n^{-s}$; the appearance of Kloosterman sums and Bessel-function transforms.
Huxley, M. N. — Area, Lattice Points, and Exponential Sums · London Mathematical Society Monographs (N.S.) 13, Oxford University Press (1996). The discrete Hardy-Littlewood method and the bound $\Delta(x) = O(x^{131/416 + \varepsilon})$ for the divisor problem (Huxley 2003 refined to $517/1648$); the role of Poisson and Voronoi summation in converting a lattice-point count into a sum of exponential sums.
Stein, E. M. & Shakarchi, R. — Fourier Analysis: An Introduction · Princeton Lectures in Analysis I, Princeton University Press (2003), Ch. 5 §3. Poisson summation via periodization and Fourier series, the theta function $\Theta(t) = \sum_n e^{-\pi n^2 t}$, and the functional equation $\Theta(1/t) = t^{1/2}\Theta(t)$ as the prototypical application.

Estimated time

beginner: 20m
intermediate: 55m
master: 90m