van der Corput's Method for Exponential Sums

Intuition Beginner

Suppose you want to add up a long list of unit arrows, one for each whole number $n$ from $1$ to $N$, where the direction of the $n$-th arrow is set by a smoothly bending rule $f(n)$ — not a straight line, not a polynomial, but some gently curving function like $\sqrt{n}$ or $n\log n$. If the arrows all pointed the same way the total would be huge, about $N$. The whole game is to show that because the directions keep turning, the arrows largely cancel and the total is far smaller. This is the same cancellation you met for straight-line and polynomial rules; the new difficulty is that a curving rule does not telescope and offers no obvious foothold.

Van der Corput found two moves that, in alternation, grind almost any smooth rule down to size. The first, the $A$-process, is a shifting trick: compare each arrow with the one a few steps ahead, and the difference of the directions is a flatter, slower-turning rule that is easier to handle. It is the square-and-shift idea behind Weyl differencing, tuned so the shift length is a free dial you optimise. The second, the $B$-process, is a change of viewpoint: a sum of a fast-turning wave can be rewritten as a new sum indexed by the speeds that appear, and the new sum is often much shorter. You trade a sum over positions for a sum over slopes.

Why care? Because two of the oldest questions in number theory — how the Riemann zeta function grows up the critical line, and how accurately you can count lattice points inside a circle or under a hyperbola — both come down to bounding exactly these sums. Each clever alternation of the two moves shaves the exponent in the answer, and chasing the best alternation is a whole industry.

Visual Beginner

Think of the sum as a walk in the plane: start at the origin, and for each $n$ take a unit step in the direction set by $f(n)$. The end point of the walk is the value of the sum. A straight-line rule sends you around a tight circle, returning near the start. A curving rule sends you on a more complicated path, and the question is how far from the origin you end up.

   A-process: shift and compare        B-process: change index

   step n   step n+h                   sum over positions n
     \        /                              |  |  |  |  |   (fast wiggle)
      \      /     difference of                 ||
       \    /      directions turns        becomes
        \  /       more slowly                sum over slopes m
         \/        (flatter rule)             |    |    |     (fewer terms)
       compare                                  short

The left panel is the shifting move: differencing replaces a turning rule with a flatter one. The right panel is the viewpoint move: a sum that wiggles fast over many positions becomes a sum over the handful of slopes that actually occur. Alternating the two is van der Corput's method, and each alternation is a fresh chance for cancellation.

Worked example Beginner

We estimate a model sum and watch the $B$-process shrink it. Take $f(n)=Tg(n/N)$ where the rule is gently curving, and look at how the sum over $n$ becomes a sum over slopes. To keep arithmetic concrete, take the linear-slope model $f(n)=n^2/(2N)$ for $n$ from $1$ to $N$, with $N=100$, and compare the number of terms before and after.

Step 1. Count the terms in the original sum. There are $N=100$ arrows, one per integer $n$ from $1$ to $100$.

Step 2. Find the range of slopes. The slope of the rule at position $n$ is the rate of change of $f$, which for $f(n)=n^2/(2N)$ is $n/N$. As $n$ runs from $1$ to $100$, the slope $n/N$ runs from $0.01$ up to $1.0$.

Step 3. Count the terms in the new sum. The $B$-process re-indexes by whole-number slopes $m$ in that range. The slopes covered run from about $0$ to $1$, so there is roughly one whole-number slope, $m=1$ (together with the endpoint behaviour). The dual sum has on the order of $1$ term instead of $100$.

Step 4. Read off the saving. Each dual term carries a weight of about $1/\sqrt{|f^{\prime \prime }|}$, and here $f^{\prime \prime }=1/N=0.01$, so $1/\sqrt{0.01}=10$. The dual sum is therefore of size about $10$, against the plain size $100$ of the original. The square root of $N$ has appeared: $\sqrt{100}=10$.

What this tells us: the $B$-process turned a sum of $100$ unit arrows into a sum of size about $10$. The cancellation is real and quantitative — the total is the square root of the length, not the length. This $\sqrt{N}$ saving is the simplest face of van der Corput's method, and iterating the two processes pushes the saving further.

Check your understanding Beginner

Formal definition Intermediate+

Throughout write $e(t)=e^{2\pi it}$ and $\Vert t\Vert$ for the distance from $t$ to the nearest integer; the additive character $e(\cdot )$ is the one used on $\mathbb{R}/\mathbb{Z}$ in 21.15.01 and 21.15.02. Let $f:[a,b]\to \mathbb{R}$ be smooth (typically real-analytic on the relevant range), and consider the exponential sum $$ S = \sum_{a < n \le b} e\big(f(n)\big). $$ Unlike the polynomial phases of 21.15.02, $f$ is now a general smooth phase, for example $f(n)=(t/2\pi )\log n$, $f(n)=\alpha n^c$ with $c\notin \mathbb{Z}$, or $f(n)=\sqrt{Tn}$. The elementary bound is $|S|\le b-a$; the object is power-saving cancellation.

Definition ($A$-process — Weyl-van der Corput inequality). For an integer $H\ge 1$ the $A$-process is the differencing inequality $$ \Big| \sum_{a < n \le b} e(f(n)) \Big|^2 \le \frac{N + H}{H}\sum_{|h| < H}\Big(1 - \frac{|h|}{H}\Big)\sum_{n \in I_h} e\big(f(n+h) - f(n)\big), \qquad N = b - a, $$ where $I_h$ is the range of $n$ with both $n$ and $n+h$ in $(a,b]$. It reduces the sum of $f$ to averaged sums of the differenced phases $f(\cdot +h)-f(\cdot )$, whose first derivatives are smaller by a factor $\approx h{f}^{\prime \prime }$. The shift bound $H$ is a free parameter, optimised at the end.

Definition ($B$-process — Poisson summation plus stationary phase). Suppose $f^{\prime }$ is monotone on $[a,b]$ with $f^{\prime \prime }$ of constant sign and $|f^{\prime \prime }|\asymp \lambda$. Poisson summation 21.15.01 writes ${\sum }_{a<n\le b}e(f(n))={\sum }_{m\in \mathbb{Z}}{\int }_a^{b}e(f(x)-mx)dx$, and evaluating each integral by stationary phase 02.20.05 — the critical point $x_m$ solves $f^{\prime }(x_m)=m$ — gives the $B$-process transformation $$ \sum_{a < n \le b} e(f(n)) = \sum_{f'(a) \le m \le f'(b)} \frac{e\big(f(x_m) - m x_m + \tfrac18 \operatorname{sgn} f''\big)}{\sqrt{|f''(x_m)|}} + O\big(\log(2 + |f'(b) - f'(a)|)\big) + O\big(\text{endpoint terms}\big). $$ The new sum is the dual sum: its phase is the Legendre transform $g(m)=f(x_m)-m{x}_m$ of $f$, and its length is $|f^{\prime }(b)-f^{\prime }(a)|$, the range of slopes.

Definition (exponent pair). A pair of real numbers $(\kappa ,\lambda )$ with $0\le \kappa \le \frac{1}{2}\le \lambda \le 1$ is an exponent pair if, for every smooth phase $f$ on $[N,2N]$ obeying the standard derivative bounds $f^{(j)}(x)\asymp F/N^j$ (the model being $f(x)=Tg(x/N)$ with $F=T$), one has $$ \sum_{N < n \le 2N} e(f(n)) \ll \Big(\frac{F}{N}\Big)^{\kappa} N^{\lambda}, $$ with the implied constant uniform over the class of phases. The amplitude $F/N\asymp f^{\prime }$ measures how fast $f$ turns; $\kappa$ is the exponent of that amplitude and $\lambda$ the exponent of the length.

Definition ($A$ and $B$ operators on pairs). The $AB$-process acts on exponent pairs by $$ A(\kappa, \lambda) = \Big( \frac{\kappa}{2\kappa + 2},\ \frac{\kappa + \lambda + 1}{2\kappa + 2} \Big), \qquad B(\kappa, \lambda) = \Big( \lambda - \tfrac12,\ \kappa + \tfrac12 \Big). $$ Starting from the base pair $(0,1)$ — the elementary bound $|S|\le N$ — and from $B(0,1)=(\frac{1}{2},\frac{1}{2})$, repeated application of $A$ and $B$ generates the classical exponent pairs. The operator $B$ is an involution ($B^2=\mathrm{id}$ on pairs), corresponding to applying the $B$-process twice and returning to the original sum.

Counterexamples to common slips

• "The $B$-process always shortens the sum." It shortens the sum precisely when the slope range $|f^{\prime }(b)-f^{\prime }(a)|=F$ is smaller than the length $N$, i.e. when $F<N$. When $F>N$ the dual sum is longer, and one applies $A$ first to reduce $F$. The art of the method is sequencing $A$ and $B$ so each is applied in its favourable regime.

• "An exponent pair must have $\kappa +\lambda =1$." Only the base pair $(0,1)$ and the first $B$-image $(1/2,1/2)$ satisfy $\kappa +\lambda =1$. General pairs satisfy $\kappa +\lambda \le$ varying bounds; the quantity minimised for $\zeta (1/2+it)$ is $\kappa +\lambda$ over the convex hull of attainable pairs, but it is not constant.

• "Stationary phase needs a polynomial phase." The $B$-process applies to any smooth $f$ with $f^{\prime \prime }$ of one sign and bounded away from $0$ and $\infty$ on the range; the stationary-phase lemma of 02.20.05 requires only a non-degenerate critical point, not a polynomial. The phases of interest ($\log$, fractional powers) are emphatically non-polynomial.

Key theorem with proof Intermediate+

The signature result is the $B$-process exponent pair: a single application of Poisson summation and stationary phase converts the base pair into a genuine power-saving bound ^{[Graham Kolesnik 1991]}.

Theorem (the $B$-process pair; van der Corput). Let $f$ be smooth on $[N,2N]$ with $f^{\prime \prime }(x)\asymp \lambda >0$ of constant sign, $f^{\prime \prime \prime }$ and $f^{(4)}$ controlled by $f^{\prime \prime }/N$ and $f^{\prime \prime }/N^2$ respectively. Then $$ \sum_{N < n \le 2N} e(f(n)) \ll N, \lambda^{1/2} + \lambda^{-1/2}. $$ Equivalently, in exponent-pair language with $\lambda \asymp F/N^2$ this is the pair $(\frac{1}{2},\frac{1}{2})$: ${\sum }_{N<n\le 2N}e(f(n))\ll (F/N{)}^{1/2}{N}^{1/2}=F^{1/2}$.

Proof. By Poisson summation 21.15.01, for a smooth cutoff $w$ adapted to $[N,2N]$, $$ \sum_{n} w(n), e(f(n)) = \sum_{m \in \mathbb{Z}} J_m, \qquad J_m = \int_{\mathbb{R}} w(x), e\big(f(x) - m x\big), dx. $$ The phase of $J_m$ is ${\varphi }_m(x)=f(x)-mx$ with ${\varphi }_m^{\prime }(x)=f^{\prime }(x)-m$. On $[N,2N]$ the derivative $f^{\prime }$ ranges over an interval of length $\asymp N\lambda$; call its image $[f^{\prime }(N),f^{\prime }(2N)]$.

For $m$ outside a neighbourhood of this image, $|{\varphi }_m^{\prime }|=|f^{\prime }-m|\gtrsim |f^{\prime }-m|$ is bounded below, and ${\varphi }_m^{\prime }$ is monotone since ${\varphi }_m^{\prime \prime }=f^{\prime \prime }\asymp \lambda$ has constant sign. The first van der Corput lemma for integrals (the $k=1$ case of 02.20.05) and repeated integration by parts give $|J_m|\ll (1+\mathrm{dist}(m,[f^{\prime }(N),f^{\prime }(2N)]){)}^{-1}\cdot$ rapidly decaying, so ${\sum }_{m\notin [f^{\prime }(N),f^{\prime }(2N)]}|J_m|\ll \log (2+N\lambda )+1$.

For each $m$ with $f^{\prime }(N)\le m\le f^{\prime }(2N)$ the phase ${\varphi }_m$ has a unique stationary point $x_m$ solving $f^{\prime }(x_m)=m$, non-degenerate because ${\varphi }_m^{\prime \prime }(x_m)=f^{\prime \prime }(x_m)\asymp \lambda \ne 0$. The stationary-phase formula of 02.20.05 gives $$ J_m = \frac{e\big(f(x_m) - m x_m + \tfrac18\big)}{\sqrt{f''(x_m)}}, w(x_m) + O\big(\lambda^{-1/2}\cdot (N^2\lambda)^{-1}\big), $$ the error being the second-order stationary-phase term, controlled by $f^{\prime \prime \prime }$ and $f^{(4)}$. Each main term has size $\asymp {\lambda }^{-1/2}$, and there are $\asymp N\lambda$ values of $m$ in the range. Summing the main terms with the triangle inequality bounds them by $(N\lambda )\cdot {\lambda }^{-1/2}=N{\lambda }^{1/2}$; the accumulated stationary-phase errors are $\ll (N\lambda )\cdot {\lambda }^{-1/2}(N^2\lambda {)}^{-1}={\lambda }^{-1/2}/N\le {\lambda }^{-1/2}$; and the off-range tail contributes $O(\log )$. Collecting, $$ \Big| \sum_{N < n \le 2N} e(f(n)) \Big| \ll N\lambda^{1/2} + \lambda^{-1/2} + \log(2 + N\lambda) \ll N\lambda^{1/2} + \lambda^{-1/2}, $$ the logarithm absorbed when $\lambda$ is in the nondegenerate range $N^{-2}\ll \lambda \ll 1$. Writing $\lambda \asymp F/N^2$ gives $N{\lambda }^{1/2}\asymp F^{1/2}$, i.e. $\ll (F/N{)}^{1/2}{N}^{1/2}$, the pair $(\frac{1}{2},\frac{1}{2})$. $\square$

Bridge. The $B$-process pair builds toward the entire exponent-pair calculus, and it appears again in the application to $\zeta (1/2+it)$ below, where $(\frac{1}{2},\frac{1}{2})$ already yields the convexity bound. The foundational reason the method works is that Poisson summation is dual to stationary phase: summation over positions becomes summation over slopes, and the weight ${\lambda }^{-1/2}$ is exactly the stationary-phase normalisation $|f^{\prime \prime }{|}^{-1/2}$ of 02.20.05. This is exactly the sum-side incarnation of the integral van der Corput lemma — where the integral lemma gives ${\lambda }^{-1/2}$ decay for one oscillatory integral, the $B$-process gives a dual sum of terms each of that size. The $A$-process generalises Weyl differencing from 21.15.02 to non-polynomial phases, and the central insight is that the two operators $A$ and $B$ on the $(\kappa ,\lambda )$ region are dual to each other: $A$ trades length for amplitude by differencing, $B$ swaps amplitude and length by the Legendre transform. Putting these together, every word in the language of exponent pairs is a sequence of $A$'s and $B$'s applied to $(0,1)$, and the best bound for any particular phase is the optimal such word.

Exercises Intermediate+

Advanced results Master

The exponent-pair region and optimisation

The set of attainable exponent pairs is the closure under $A$ and $B$ of $\{(0,1)\}$, together with the convexity property that the attainable region is convex ^{[Graham Kolesnik 1991]}. The operators generate a countable dense family of vertices; for the divisor and zeta problems one minimises a linear functional of $(\kappa ,\lambda )$ over this region. For $\zeta (\frac{1}{2}+it)\ll t^{\theta +\epsilon }$ with $\theta =\frac{1}{2}\inf (\kappa +\lambda )$ over attainable pairs, van der Corput's processes alone give $\theta =\frac{1}{6}$ from $(\frac{1}{6},\frac{2}{3})=AB(0,1)$; the longer words $A^{k}B$ and their conjugates yield the van der Corput exponent pairs $(\frac{1}{2^{k+1}-2},1-\frac{k}{2^{k+1}-2})$, whose infimum of $\kappa +\lambda$ still does not beat $\frac{1}{6}$ for the zeta function but does improve the divisor exponent. Rankin computed the optimal van der Corput exponent $\theta =0.3290\dots$ for the divisor problem from this region. The barrier $\frac{1}{6}$ for zeta is broken only by stepping outside the $AB$-process: the Bombieri-Iwaniec method (1986) and Huxley's discrete Hardy-Littlewood refinements reach $\theta \le 131/416$.

The Dirichlet divisor and Gauss circle problems

Let $D(x)={\sum }_{n\le x}d(n)=x\log x+(2\gamma -1)x+\Delta (x)$ and $\#\{(m,n):m^2+n^2\le x\}=\pi x+P(x)$ ^{[Huxley 1996]}. Both error terms reduce, via the hyperbola method and Voronoi summation 21.15.01, to exponential sums ${\sum }_{n\le M}\psi (x/n)$ and $\sum e(2\sqrt{xn})$ with phase $f(n)=2\sqrt{xn}$, $f^{\prime \prime }(n)\asymp \sqrt{x}{n}^{-3/2}$. Applying an exponent pair $(\kappa ,\lambda )$ yields $\Delta (x),P(x)\ll x^{\frac{\kappa +\lambda }{2(1+\kappa )}}$ up to $\epsilon$-powers; the van der Corput pair $(\frac{1}{6},\frac{2}{3})$ gives the classical $x^{1/3}$, sharpening Dirichlet's elementary $x^{1/2}$ and Voronoi's $x^{1/3}$ obtained by his summation formula. The conjectured exponent is $\frac{1}{4}$, equivalent to the optimal cancellation $\Delta (x)\ll x^{1/4+\epsilon }$; the $\Omega$-results of Hardy show $\frac{1}{4}$ cannot be improved.

Iterated processes and the $A$-process gain

The general bound from a word $w=w_1\cdots w_r$ in $\{A,B\}$ applied to $(0,1)$ is sharp for the phase class it was tuned to, but the same phase may admit a better word. The $A$-process improves a pair when differencing the phase produces a flatter sum: applied to $(\kappa ,\lambda )$ it returns $(\frac{\kappa }{2\kappa +2},\frac{\kappa +\lambda +1}{2\kappa +2})$, which decreases $\kappa$ (good for small-amplitude phases) at the cost of raising $\lambda$ toward $1$. Because $\kappa$ and $\lambda$ trade off, the optimal word depends on the regime $F=N^c$: for the zeta-type phase $c=2$ the optimum among $AB$-words is $AB$, while for the divisor phase $c=1/2$ the optimum is a longer alternation. The exponent-pair conjecture asserts that $(\epsilon ,\frac{1}{2}+\epsilon )$ is attainable for every $\epsilon >0$, which would give $\theta =\frac{1}{4}+\epsilon$ for both the zeta and divisor problems; it remains open and is known to require methods beyond $A$ and $B$.

Synthesis. Van der Corput's method, exponent pairs, the zeta growth problem, and the divisor and circle problems are one circle of ideas, and the bridge is the duality between a sum and its Legendre-transformed dual. The foundational reason the processes terminate in a power saving is that the $B$-process is exactly Poisson summation from 21.15.01 composed with the stationary-phase lemma of 02.20.05: summation over positions becomes summation over slopes, the weight $|f^{\prime \prime }{|}^{-1/2}$ is the stationary-phase normalisation, and this is dual to the $A$-process, which generalises the Weyl differencing of 21.15.02 from polynomial to smooth phases. Putting these together, every classical bound is a word in $A$ and $B$ applied to the base pair $(0,1)$, the central insight being that $A$ trades length for amplitude and $B$ swaps them by the Legendre transform; the convexity bound $\zeta (\frac{1}{2}+it)\ll t^{1/6+\epsilon }$ is exactly $AB(0,1)$, and the $x^{1/3}$ divisor and circle exponents are the same calculus read on the phase $2\sqrt{xn}$. The exponent-pair conjecture $\theta =\frac{1}{4}$ generalises the whole method to its conjectural limit, and the modern Bombieri-Iwaniec and decoupling methods are the central insight pushed past the $A,B$-closure into the geometry of the dual sum's spacing.

Full proof set Master

Proposition 1 (first-derivative test; van der Corput's lemma for sums). Let $f$ be real on $[a,b]$ with $f^{\prime }$ monotone and $\Vert f^{\prime }(x)\Vert \ge \delta >0$ for all $x\in [a,b]$ (the distance to the nearest integer of $f^{\prime }$ stays $\ge \delta$). Then $$ \Big| \sum_{a < n \le b} e(f(n)) \Big| \ll \frac{1}{\delta}. $$

Proof. By Euler-Maclaurin / partial summation against the integral, or directly by the truncated Poisson summation of 21.15.01, ${\sum }_{a<n\le b}e(f(n))={\sum }_m{\int }_a^{b}e(f(x)-mx)dx+O(1)$, where only $m$ with $m\in [f^{\prime }(a),f^{\prime }(b)]$ contribute a main term. The hypothesis $\Vert f^{\prime }\Vert \ge \delta$ means $f^{\prime }$ avoids a $\delta$-neighbourhood of every integer, so for each integer $m$ the integrand phase $f(x)-mx$ has derivative $f^{\prime }(x)-m$ of absolute value $\ge \delta$ and monotone. The first van der Corput integral lemma (02.20.05, case $k=1$) gives $|{\int }_a^{b}e(f(x)-mx)dx|\ll 1/\Vert f^{\prime }-m\Vert$-type control; summing the geometric-like decay over $m$ produces $\ll {\sum }_{j\ge 0}(j+\delta {)}^{-1}$ truncated, which is $\ll {\delta }^{-1}$ once the monotonicity of $f^{\prime }$ bounds the number of relevant $m$ by $O(1+|f^{\prime }(b)-f^{\prime }(a)|)$ and the tails decay. Hence $|S|\ll {\delta }^{-1}$. $\square$

Proposition 2 (second-derivative test). Let $f$ be real on $[a,b]$ with $\lambda \le |f^{\prime \prime }(x)|\le \alpha \lambda$ for constants and $f^{\prime \prime }$ of constant sign. Then $$ \Big| \sum_{a < n \le b} e(f(n)) \Big| \ll (b - a),\lambda^{1/2} + \lambda^{-1/2}. $$

Proof. This is the conclusion of the $B$-process key theorem, here stated for a general interval. Since $f^{\prime \prime }$ has constant sign, $f^{\prime }$ is strictly monotone, so the $B$-process applies: Poisson summation 21.15.01 reduces $S$ to the dual sum ${\sum }_{f^{\prime }(a)\le m\le f^{\prime }(b)}|f^{\prime \prime }(x_m){|}^{-1/2}e(g(m))$ over stationary points $x_m$ solving $f^{\prime }(x_m)=m$, plus $O(\log )$. The number of integers $m$ in $[f^{\prime }(a),f^{\prime }(b)]$ is $\ll 1+|f^{\prime }(b)-f^{\prime }(a)|\ll 1+(b-a)\lambda$, since $f^{\prime \prime }$ ranges in $[\lambda ,\alpha \lambda ]$. Each dual term has modulus $|f^{\prime \prime }(x_m){|}^{-1/2}\asymp {\lambda }^{-1/2}$. Bounding the dual sum directly by its length times the term size, $$ |S| \ll \big(1 + (b-a)\lambda\big)\lambda^{-1/2} + (\text{errors}) \ll (b-a)\lambda^{1/2} + \lambda^{-1/2}, $$ the stationary-phase and endpoint errors being lower order as in the key theorem. $\square$

Proposition 3 (the basic exponent pair from the second-derivative test). The pair $(\frac{1}{2},\frac{1}{2})$ is an exponent pair: for $f$ on $[N,2N]$ with $f^{(j)}\asymp F/N^j$, ${\sum }_{N<n\le 2N}e(f(n))\ll (F/N{)}^{1/2}{N}^{1/2}$.

Proof. Here $f^{\prime \prime }\asymp F/N^2=:\lambda$, of constant sign on the dyadic block (shrinking the block if necessary so $f^{\prime \prime }$ does not vanish). Proposition 2 with $b-a=N$ gives $|S|\ll N{\lambda }^{1/2}+{\lambda }^{-1/2}=N(F/N^2{)}^{1/2}+(F/N^2{)}^{-1/2}=(FN{)}^{1/2}/N^{1/2}\cdot N^{1/2}\cdots$; explicitly $N{\lambda }^{1/2}=N(F/N^2{)}^{1/2}=F^{1/2}=(F/N{)}^{1/2}{N}^{1/2}$, and ${\lambda }^{-1/2}=(N^2/F{)}^{1/2}=N/F^{1/2}\le N^{1/2}(N/F{)}^{1/2}$. In the standard range $1\le F\le N^2$ where the phase genuinely oscillates and is the regime exponent pairs are defined for, the first term dominates and $|S|\ll (F/N{)}^{1/2}{N}^{1/2}$. This is the assertion that $(\kappa ,\lambda )=(\frac{1}{2},\frac{1}{2})$ is an exponent pair, matching $B(0,1)$. $\square$

Proposition 4 ($AB(0,1)=(\frac{1}{6},\frac{2}{3})$ and the $t^{1/6}$ bound are equivalent inputs). Applying the $A$-operator to $(\frac{1}{2},\frac{1}{2})$ gives $(\frac{1}{6},\frac{2}{3})$, and this pair yields $\zeta (\frac{1}{2}+it)\ll t^{1/6+\epsilon }$.

Proof. The first claim is the computation $A(\frac{1}{2},\frac{1}{2})=(\frac{1/2}{3},\frac{2}{3})=(\frac{1}{6},\frac{2}{3})$ of Exercise 3. For the second, the exponent-pair bound on the zeta sum ${\sum }_{N<n\le 2N}{n}^{-1/2-it}$ at the critical length $N\asymp \sqrt{t}$, summed dyadically against $n^{-1/2}$ by partial summation as in Exercise 8, gives $\zeta (\frac{1}{2}+it)\ll t^{(\kappa +\lambda )/2-1/4}\log t$. Substituting $(\kappa ,\lambda )=(\frac{1}{6},\frac{2}{3})$ makes the exponent $(\frac{1}{6}+\frac{2}{3})/2-\frac{1}{4}=\frac{5}{12}-\frac{1}{4}=\frac{1}{6}$, so $\zeta (\frac{1}{2}+it)\ll t^{1/6}\log t\ll t^{1/6+\epsilon }$. $\square$

Connections Master

The $A$-process is the smooth-phase generalisation of the Weyl differencing inequality of 21.15.02: where Weyl differencing reduces a degree-$k$ polynomial sum to degree-$(k-1)$ sums, the $A$-process reduces a smooth sum to averaged difference sums with a free shift length, and both are Cauchy-Schwarz applied to shifted copies of the same character sum.

The $B$-process is Poisson summation from 21.15.01 composed with the method of stationary phase from 02.20.05; the dual sum it produces is indexed by the slopes $m=f^{\prime }(x_m)$, exactly the transform side of the Poisson formula, and each term carries the stationary-phase weight $|f^{\prime \prime }(x_m){|}^{-1/2}$ that the oscillatory-integral lemma of 02.20.05 supplies.

The $t^{1/6+\epsilon }$ bound for $\zeta (1/2+it)$ proved here is the first convexity-breaking estimate feeding the growth and zero-density theory of the Riemann zeta function in 21.03.01; subconvexity for $\zeta$ and for general $L$-functions continues this line, with the complete exponential sums of 21.15.04 supplying the arithmetic input on the major-arc side that complements the smooth-phase analysis here.

Historical & philosophical context Master

Johannes Gualtherus van der Corput introduced his method in two papers in the Mathematische Annalen: the 1921 Zahlentheoretische Abschätzungen ^{[van der Corput 1921]}, which set out the derivative tests and the differencing and Poisson-summation passages, and the 1922 Verschärfung der Abschätzung beim Teilerproblem ^{[van der Corput 1922]}, which iterated the processes to improve the Dirichlet divisor exponent below Voronoi's $1/3$. The two operations were later christened the $A$- and $B$-processes, and the bookkeeping of their effect on the bound was systematised into the theory of exponent pairs by Phillips, Rankin, and others in the 1930s-1950s; Rankin found the optimal van der Corput exponent for the divisor problem by linear programming over the attainable region.

The method's first triumph was the bound $\zeta (\frac{1}{2}+it)\ll t^{1/6}\log t$, breaking the convexity exponent $\frac{1}{4}$ inherited from the functional equation and the elementary estimate. It stood essentially unchallenged in its conceptual form until the Bombieri-Iwaniec method of 1986 introduced the spacing of the dual sum's terms as a new resource, and Huxley's discrete Hardy-Littlewood method refined this to the current $\theta \le 131/416$ ^{[Huxley 1996]}. Graham and Kolesnik's 1991 monograph remains the systematic account of the classical $A,B$-calculus. The method is the discrete shadow of the stationary-phase analysis of oscillatory integrals: van der Corput himself stated the integral lemmas alongside the sum estimates, and the modern view that exponent pairs are words in two dual operations on the $(\kappa ,\lambda )$ region dates to this systematisation.

Bibliography Master

@article{vandercorput1921,
  author  = {van der Corput, J. G.},
  title   = {Zahlentheoretische Absch{\"a}tzungen},
  journal = {Mathematische Annalen},
  volume  = {84},
  pages   = {53--79},
  year    = {1921}
}

@article{vandercorput1922,
  author  = {van der Corput, J. G.},
  title   = {Versch{\"a}rfung der Absch{\"a}tzung beim Teilerproblem},
  journal = {Mathematische Annalen},
  volume  = {87},
  pages   = {39--65},
  year    = {1922}
}

@book{grahamkolesnik1991,
  author    = {Graham, S. W. and Kolesnik, G.},
  title     = {Van der Corput's Method of Exponential Sums},
  series    = {London Mathematical Society Lecture Note Series},
  volume    = {126},
  publisher = {Cambridge University Press},
  year      = {1991}
}

@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{titchmarsh1986,
  author    = {Titchmarsh, E. C.},
  title     = {The Theory of the Riemann Zeta-Function},
  edition   = {2},
  note      = {Revised by D. R. Heath-Brown},
  publisher = {Oxford University Press},
  year      = {1986}
}

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