The classical Weyl bound for Riemann Zeta function estimate says that, for ,
The interest of this bound is not confined to the numerical exponent. It is an unusually good place to learn how analytic number theorists think about oscillatory sums: how a finite expression for a function becomes an exponential-sum problem, why one should resist taking absolute values too soon, why the first plausible estimate sometimes detects no cancellation at all, and how different transformations can become effective in different ranges. It is common to prove the estimate after writing down an approximate functional equation of length about , and then invoking a standard estimate for exponential sums. That is an efficient proof if the analytic machinery is already familiar, but it can obscure the actual choices that make the argument work. We shall take another route. Starting from an elementary finite partial sum for zeta, we shall retain a Dirichlet polynomial of length comparable to
and examine its dyadic portions directly. Poisson summation will control the long portions by turning them into shorter dual sums; differencing will control the intermediate portions by replacing the original phase with correlations of much smaller curvature; and the very short portions will not require cancellation at all. No functional equation, gamma factor, or contour integral is needed.
Throughout we write , put
, and assume
is large. Constants implicit in
may depend on a fixed smooth cutoff when one is used, but never on
. The central phase will be
Finite Sum Approximation
For the identity
is absolutely convergent. On the critical line the same individual terms have the form
, but one cannot simply declare that their infinite sum equals zeta: the series no longer converges. We first need a finite approximation with a rigorously controlled error. A very elementary comparison with an integral gives, for
and
,
This formula is often obtained by comparing with
. Since
, the discrepancy on one unit interval is
, and the sum of all discrepancies beyond
is
. Yet this is a striking example of a perfectly valid estimate that is analytically much too weak. When
and
, its error is
, whereas the very sum we want to estimate has trivial size
. The comparison has thrown away the oscillation in the discrepancy between an integer sum and a continuous integral. Before choosing a more sophisticated representation of zeta, we should ask whether we can recover that information within the same elementary setup. The key observation is that the discrepancy between integer points and continuous mass is periodic. Let
for nonintegers
. For a positive integer
, Euler summation gives the exact formula, initially for
and then by continuation for
,
The last integral is absolutely convergent when . Bounding
by
leads back to the crude error, so we must use the oscillation of
. Its Fourier expansion is
where, at discontinuities, the usual midpoint or Abel interpretation is understood. Inserting an Abel-regularized version into the integral reduces the problem to estimates for
Here we have separated the nonoscillatory amplitude from the phase, and that distinction is important. The phase derivative is
. If we choose
, then for every
and every nonzero integer
this derivative has magnitude at least a fixed multiple of
. Thus no stationary point is present in the tail. Integrating by parts, dividing by the phase derivative and differentiating the remaining amplitude, gives
To check the estimate rather than merely quote it, note that differentiation of the amplitude contributes an integral of size , while differentiation of the reciprocal phase derivative contributes
, which is no larger because
. The Fourier coefficients of
supply an additional factor
, and
converges. Therefore
We obtain the much sharper elementary truncation formula
with a positive integer. On the critical line, choose
. The integral correction itself then has size
, so the expression simplifies to
This is the only representation of zeta that we shall need. The cutoff at the scale has not been chosen to imitate a functional equation. It arises from a simple fact about integer sampling: beyond
, the oscillatory correction to the integral has no resonances at nonzero integer frequencies. The omitted tail is consequently under excellent control. We have transferred the entire difficulty to a concrete finite sum.
Exponential Sum
Write . Its trivial estimate is
. The Weyl bound asks for a saving of size
relative to that estimate, but it would be misleading to imagine that this saving must be found uniformly inside one enormous sum. Divide the interval into dyadic ranges
, with the last interval stopped at
. On such an interval the coefficient
changes only by a constant factor. Let
If we can prove the uniform estimate
then ordinary summation by parts gives the same bound for the weighted block after multiplying by . More explicitly, if
, then
Both terms are bounded by . Thus the unweighted target is exactly the right one. It is also important that the estimate be uniform over subintervals, because partial summation exposes all intermediate endpoints, not only the full dyadic sum. Once each block is
, the
blocks contribute the announced logarithm. There is no necessity to save an additional logarithm or to find cancellation between distinct dyadic scales.
Before doing anything clever, inspect where the trivial bound is already sufficient. Since , the target is automatic when
, because
. Thus no oscillatory argument is needed for the short blocks. Our task is confined to
. We shall eventually find that one method controls
, while a different method controls
. The precise matching of these ranges will explain the exponent, rather than being something imposed on the proof in advance.
At the derivative of our phase is
. In particular, around
, it is very large. One may be tempted to say that the phase is oscillating quickly, so adjacent terms ought to cancel. For a sum over integers this reasoning is unsafe. If
, then
is large, but
at every integer, giving no cancellation. What matters is not the magnitude of the derivative as a real number but its value modulo the integers, since adding an integer to the phase does not change the exponential. For a linear phase, the geometric-series identity gives
,
where denotes distance to the nearest integer. For a nonlinear phase the local increment
plays the analogous role. The difficulty is that this increment varies with
, potentially passing through many integers. A first-derivative estimate works well while the increment stays uniformly separated from the integers, but it becomes ineffective near the places where the increment approaches an integer. We must therefore understand both how often these resonances occur and how large the contributions around them may be. Poisson summation gives a particularly clean geometric description of those resonances, and it is worth working through it before turning to differencing.
Poisson Summation
Take a smooth function supported inside
, and consider
. Poisson summation gives the exact identity
The smooth weight is not essential to the phenomenon, but it eliminates artificial discontinuities at the endpoints and allows repeated integration by parts. Write , where
. The derivative is
. When that derivative is bounded away from zero, integration by parts makes the integral small. It follows that the important frequencies are those for which
somewhere inside the support. Such a frequency must be negative. Writing
, the stationary point is
Because the original variable satisfies , the relevant frequencies satisfy
. This identifies the natural length of the dual sum. Poisson summation is not intrinsically a machine that makes every sum shorter; it turns the variation of the original derivative into a new summation range. For this logarithmic phase, length
becomes length roughly
. If
, the new sum is shorter; if
, the two lengths are comparable; and if
, the transformed sum is longer. The usefulness of the transformation depends entirely on the scale at which it is applied.
Let us also calculate the size of a typical integral rather than simply invoking stationary phase. At , the first derivative of
vanishes and the second derivative is
. Taylor’s formula begins with a quadratic term. The phase changes by order one across a neighborhood of width
. That width predicts the size of the integral, since the oscillations outside the stationary region largely cancel. A standard stationary-phase calculation, obtained by rescaling the quadratic term to a Fresnel integral, gives the leading term
uniformly for frequencies in a fixed multiple of the stationary range; the leading term is interpreted as zero when the stationary point lies outside the support. Frequencies sufficiently far from that range contribute negligibly by repeated integration by parts. The precise form of the remainder is less important than the scale: there are about relevant frequencies, while each leading term has size about
. Estimating all those terms by absolute values therefore gives a bound of size
, which is the origin of the classical second-derivative scale in this example.
A more revealing calculation concerns the new phase. Since , substitution of
yields
The second term does not depend on . The varying part of the transformed phase is therefore
, the same logarithmic phase with the opposite sign. Schematically, after extracting an overall constant of absolute value one, the transformation takes the form
where is a smooth amplitude. This is the concrete content of the van der Corput
-process for our phase. It reorganizes the oscillation rather than eliminating it. In particular, when
, the transformed expression is essentially as long as the original one, and taking absolute values on the dual side gains nothing. But for long intervals the same operation becomes powerful even without exploiting any cancellation between the dual frequencies.
On a dyadic block of length , the stationary integrals for the unweighted sum have total absolute size bounded by
. The additional
term allows for the possibility of only a bounded number of stationary frequencies. The corresponding estimate for the actual weighted Dirichlet block, whose amplitude is
, is
This bound can also be obtained for sharp subintervals, with harmless endpoint terms, from the second derivative estimate we shall derive below; hence smoothness is not an essential restriction in the final proof. In the range , the first term is at most
, while the second is
. Thus Poisson summation controls every long dyadic block at the strength required for Weyl’s bound. Notice the elementary arithmetic that determines the threshold: asking
is exactly the same as asking
. The exponent
has not come from a mysterious auxiliary theorem. It is the point where the dual sum becomes short enough that even its absolute-value bound is sufficient.
What remains is the range . Poisson summation alone gives the same generic second-derivative scale, but it is too expensive near
. There, the original sum and its dual have comparable lengths, and the operation does not make the underlying problem simpler. We need to change something about the phase itself. This is where differencing enters, not as an arbitrary additional trick, but as an attempt to construct new sums on which the second derivative is better behaved.
Second Derivative Estimate
Suppose is twice differentiable on an interval containing at most
integers, and suppose
has constant sign and satisfies
. We claim that
To see why the two terms arise, assume first that and choose a small parameter
. Separate the interval into portions where
is within distance
of an integer, and portions where it is not. Since
has one sign, the derivative is monotone and passes through at most
relevant integers. Around each crossing, the region with
has length
, so estimating those dangerous portions trivially costs at most
. On each of the complementary portions, the phase increments stay separated from integers; a first-derivative estimate gives
. One can prove that estimate by writing the difference
, dividing by
, and applying summation by parts. The monotonicity of the increments makes the total variation of the reciprocal denominator
. Summing over the complementary portions gives
The meaningful choice is , which balances the two types of loss and produces
. If
, the trivial bound
suffices to obtain the same displayed inequality. This derivation is useful because it makes the competing phenomena visible. Curvature that is too small can leave a long, nearly linear, potentially resonant stretch; curvature that is too large can cause the derivative to cross many integer frequencies. The second-derivative estimate balances these difficulties. The Poisson interpretation tells the same story differently: there are about
potentially stationary dual frequencies, each carrying an integral of natural size
.
For our original phase, , so on
we have
. The estimate becomes
At this is no better than the trivial bound
. The failure is real, not merely a weakness of our presentation. A quadratic phase such as
has
and yet
at every integer. Curvature of size one need not imply any cancellation. Although the logarithmic phase is not such a quadratic, a generic bound depending only on the size of the second derivative cannot exploit that distinction. To make progress, we must form a different phase whose curvature lies in a more favorable range.
Shifts and correlations
Put on an interval
and zero elsewhere. If
, then expansion of the square gives
. Reorganizing the pairs according to their displacement
shows that the sum is related to correlations
. This is not a mysterious maneuver: a sequence whose terms reinforce one another strongly should exhibit substantial correlations with many of its translates. The important choice is to restrict attention to small shifts. When
, the difference between nearby phases can be far simpler than either phase taken separately, and the derivatives of that difference can be smaller by a power of
.
For the logarithmic phase the correlation has phase
The Taylor expansion already suggests that the large original logarithmic phase has been replaced by a smaller reciprocal-type phase. We need not rely on that expansion, however, because the relevant derivatives can be calculated exactly. Twice differentiating gives
Therefore, for and
,
The original curvature was ; the curvature of the difference has acquired the additional factor
. At
, the old curvature is of order one, while the new curvature is
, which can be much smaller. We should not say that smaller curvature automatically means stronger cancellation: an almost linear phase can also resonate. What differencing actually does is create a new, adjustable curvature scale. We can choose a range of shifts in which the two terms of the second-derivative estimate are favorable, while paying a separate cost for the averaging that introduced the shifts.
Van der Corput differencing
Let be an integer. Because
has been extended by zero outside its support, the identity
is exact: every summand occurs
times. There are at most
relevant values of
, so Cauchy–Schwarz gives
Expanding the square separates the diagonal pairs from the off-diagonal pairs. The diagonal contributes at most
. A positive displacement
occurs
times, and its contribution is the correlation
. We therefore arrive at
and consequently
It is worth understanding this inequality before substituting any estimates. Cauchy–Schwarz has not itself produced cancellation; it has exchanged the original sum for a diagonal contribution and a collection of nearby correlations. Using more shifts makes the diagonal term smaller, which is beneficial. But larger shifts also change the correlation phases and can make them more expensive to estimate. The parameter
must balance these effects. An averaging device often creates exactly this type of tradeoff: it reduces a trivial or diagonal cost while increasing the complexity of the new objects that must be controlled. The success of the argument depends on finding the range in which the correlations are sufficiently easier than the original sum.
For our logarithmic phase, the correlation is an exponential sum with phase
, supported where
. Applying the second-derivative estimate with
gives
The first term increases with , because the difference phase then crosses more possible resonant frequencies. The second decreases with
, because a shift that is too small makes the phase too nearly linear. Substituting the two terms into the differencing inequality and using
and
yields
Here all the important analytic costs are visible in one expression. The first is the diagonal loss from Cauchy–Schwarz. The second is the accumulation of resonances in the shifted phases. The third is the possible cost of almost linear behavior in those phases. Increasing improves the first and third terms, but worsens the second. Thus the natural candidate is to balance the first two terms, rather than trying to make
as large as possible. The balance
gives
, or
This choice is possible with once
, exactly where the trivial bound ceased to suffice. Substitution makes each of the first two terms
, while the third becomes
. We must not ignore this last term merely because we have balanced the other two. The inequality
is equivalent to
. Hence, for
,
The bound is uniform over subintervals of a fixed dyadic range, because the derivative conditions and the support-length inequality are uniform there. We have proved exactly the exponential-sum estimate needed for partial summation. The exponent has appeared from an explicit optimization, but it is more important to recognize why the optimization was available: differencing produced a new curvature
and allowed us to place the correlation sums where the second-derivative test genuinely detects cancellation.
The mechanism is especially transparent when . The original sum then has about
terms, and its second derivative is of order one. A direct second-derivative estimate permits a sum as large as
, so it proves no saving. The differencing optimization chooses
. For a typical shift
, the curvature of the difference phase has size
. Thus a correlation containing roughly
terms is bounded by
This is the first place where we have a concrete cancellation saving: each typical correlation is bounded by instead of the trivial
. Returning to Cauchy–Schwarz, the diagonal term is
, and the total correlation contribution is of the same order. Consequently
, so
. Multiplication by the natural coefficient
gives
, exactly the desired contribution from the critical dyadic block.
This calculation is useful because it distinguishes the different sources of gain. Cauchy–Schwarz alone does not force the sum to be small; it produces correlations and a diagonal loss. The second-derivative estimate alone does not control the original phase; its curvature is in a potentially resonant range. The gain arises because Cauchy–Schwarz exposes a different phase to which the second-derivative estimate can be applied successfully. In the language of exponent pairs, this is the -process, but the terminology is secondary to the underlying construction. We are arranging for the analytic information that one estimate can detect to become visible after a transformation performed by another estimate.
Weyl Bound
We can now return to the finite sum without introducing any new machinery. The elementary truncation formula gave , with
. Divide that sum into dyadic intervals and use partial summation to reduce each weighted block to the maximum size of an unweighted exponential sum over its subintervals. For
, the trivial bound
gives a weighted contribution of at most
. For
, the differencing estimate gives
, and partial summation makes the weighted contribution
. For
, the second-derivative estimate, whose structure was explained by Poisson summation, gives
and hence a weighted contribution
. Thus every dyadic interval contributes at most
, including the final interval ending at the sharp cutoff
. There are
such intervals. Their absolute contributions therefore sum to
. The
error in the finite-sum formula is negligible on this scale, and we have proved
The short blocks needed no cancellation; the intermediate blocks were handled by differencing; and the long blocks were handled by the second-derivative estimate, which Poisson summation reveals as a dual-frequency estimate. The two nontrivial methods meet at the same boundary : below that scale, the optimized correlation bound is strong enough, while above it, the Poisson dual sum is sufficiently short. This matching is the main structural feature of the argument.
It is instructive to compare Poisson summation and differencing without describing either of them as a formal operation to remember. Poisson summation takes an integer sum and decomposes it into Fourier frequencies. The relevant frequencies are those at which a dual oscillatory integral has a stationary point; their number is controlled by the variation of the original derivative. For the logarithmic phase, the transformation takes length into length about
, while reproducing a logarithmic phase with the opposite sign. Its advantage is thus primarily one of length: it is useful when the new sum is shorter enough that even a crude estimate can finish the job. Differencing, by contrast, does not replace
by a dual frequency. It compares nearby values of the original sequence, introducing a shift
and a new phase
. Its advantage is primarily one of curvature: the new phase has second derivative of size
rather than
. The price is that Cauchy–Schwarz introduces a diagonal term and forces us to control a family of shifted correlations. These descriptions also explain why the methods are related. The second-derivative bound for the correlations can itself be seen as a Poisson-frequency calculation: the number of possible stationary frequencies is approximately the total variation of the derivative, and the size of each stationary contribution is controlled by the reciprocal square root of the curvature. Differencing is therefore not an unrelated trick appended to a Poisson estimate. It constructs a new phase on which the same stationary-frequency analysis operates in a more favorable regime. In traditional terminology, the
-process changes the phase by averaging correlations, while the
-process changes the representation by Fourier duality. Their combination works only when the scale of one transformation is compatible with the losses of the other.
It is equally important to understand why one cannot expect indefinite improvement by repeating Poisson summation alone. A sum of length with our logarithmic phase is sent to a dual sum of length roughly
, and applying the same duality again returns the length to approximately
. At the self-dual scale
, neither side is shorter. The transformation reveals structure but does not by itself prove that the structure cancels. One needs an additional ingredient—such as differencing, information about the coefficients, or a more refined analysis of the interaction between stationary points. The same caution applies to repeated differencing: every extra Cauchy–Schwarz operation introduces new diagonal and correlation costs, and more complicated phase differences need not be easier to estimate. What matters is not the number of transformations performed, but whether the resulting sums fall into a range where an available estimate sees genuinely new cancellation.
Our proof is successful, but one learns at least as much by noticing exactly where it is wasteful. In the long range we used Poisson summation and then estimated the contributions of the dual stationary points separately. The leading terms in the dual sum have their own oscillatory logarithmic phase, but we ignored possible cancellation among them. Near the self-dual scale this is especially costly: the dual sum remains long, so adding its terms by absolute value reproduces the trivial size of the original sum. In the intermediate range we applied Cauchy–Schwarz, replaced a family of correlations by a sum of their absolute values, and then estimated each correlation only through the size of its second derivative. Each of those operations discards potentially useful arithmetic information. In particular, the correlations for different shifts are not unrelated, and the stationary points of their Fourier transforms are not distributed arbitrarily. A stronger argument may become possible if their mutual spacing, congruence structure, or cancellation across shifts can be retained. This is the direction in which more sophisticated approaches to exponential sums develop. The deeper question is not simply how to improve a numerical exponent by applying a sharper standard lemma; it is how to avoid an absolute-value bound that destroys a correlation one has worked hard to expose. In the Bombieri–Iwaniec method, for instance, the organization of families of oscillatory phases and the spacing of their associated frequency data become central. One should regard these developments as answers to a specific weakness already visible in the Weyl argument: the elementary method counts individual contributions and then adds their sizes, even when the transformed contributions themselves have highly structured oscillations. The classical bound thus illustrates both the power of elementary transformations and the precise limits of what they prove when one ignores interactions between different frequencies or shifts.
The Weyl estimate follows from a small collection of ideas, but their arrangement is decisive. First, the elementary partial-sum formula puts zeta into a finite exponential sum; handling its error correctly requires using the absence of integer-frequency stationary points in the far tail, rather than accumulating local approximation errors by absolute value. Next, dyadic decomposition clarifies the scale of the required saving. The small blocks already satisfy the target without cancellation. The large blocks become tractable after Poisson summation, because their dual frequency range is short. The intermediate blocks are the genuine obstruction: the original phase may resonate too strongly for a direct second-derivative estimate, so we form correlations of nearby translates and change the curvature. Finally, an explicit balance between the diagonal cost and the resonance cost selects the shift length and produces
. The remaining term in that estimate explains why the argument stops at
, just where the Poisson estimate becomes sufficient. There is a useful general principle behind this. A rapidly varying phase does not automatically produce cancellation on an integer lattice, and a formal transformation does not automatically improve a bound. What matters is whether the transformed problem places the oscillation in a form that an available estimate can detect. Poisson summation reveals the frequencies at which the phase locally resonates with integer sampling; differencing modifies the phase so that these resonances can be counted more effectively; and Cauchy–Schwarz relates the improved correlation estimates back to the original sum at a calculable cost. The exponent
is the numerical trace of that arrangement. Understanding why the arrangement works is substantially more valuable than remembering the exponent alone.