In the previous post, we have seen how the Hardy–Littlewood circle method can give an asymptotic formula for representations of a large integer as a sum of five squares. The essential analytic step was to show that the minor arcs contribute less than the expected main term. We obtained a pointwise estimate for the quadratic exponential sum using Weyl differencing, combined it with a fourth-moment estimate, and gained a power of the main parameter. The method worked because the fifth variable gave us an extra factor of the exponential sum on which we could use the pointwise saving. What happens if we try the same argument with four variables? This is a natural question because Lagrange’s theorem tells us that every positive integer is a sum of four squares. More generally, we might want to count representations of by a positive-definite quadratic form
. We would expect an asymptotic formula involving a real density and a product of congruence densities, just as in the five-square problem. But the elementary circle-method estimates seem to stop working at exactly four variables.
Kloosterman’s achievement in 1926 was to recognize that this failure did not necessarily mean that the exponential sums were too large. It could mean that we were taking absolute values too early and therefore discarding cancellation between contributions associated with different rational approximations. This is the idea we will investigate. Rather than beginning with the definition of a Kloosterman sum, we will start with the failed estimate and ask where a saving might still be hiding. We will examine the structure of the Farey dissection, open up the quadratic exponential sums by Poisson summation, and calculate the size of the individual contributions. We will see that the missing cancellation has a very specific form: it involves both a residue class and its multiplicative inverse
. We will then prove Kloosterman’s original nontrivial estimate for these sums using an elementary fourth-moment calculation. Finally, we will see quantitatively why this estimate is strong enough to cross the four-variable barrier, how Weil’s later square-root bound improves it, and where the remaining analytic difficulties lie.
The aim is not simply to introduce another exponential sum. We want to understand why the original estimate failed, what kind of cancellation would repair it, and why the arithmetic structure of Farey fractions leads us to precisely the sum that Kloosterman studied. The original paper is H. D. Kloosterman, On the Representation of Numbers in the Form .
Circle Method
Let us begin with the simplest quadratic form, . Write
and
,
where . The number of representations by four positive squares is given by the Fourier integral
.
From dimension counting, we expect the main term to have size . Indeed, we are choosing four variables, each of size roughly
, and imposing one quadratic equation, which should remove approximately two powers of
. In the previous post we proved the fourth-moment estimate
.
For five squares this was enough, because we could combine it with the minor-arc supremum bound
and obtain a contribution smaller than . For four squares, however, there is no additional factor. Our estimate is merely
But the main term itself should be of order . We have no power saving. It is important to identify the precise loss. The fourth-moment estimate is not necessarily poor. In fact, its order of magnitude is essentially natural: the diagonal solutions of
alone contribute about
. We cannot expect the integral of the nonnegative function
to be substantially smaller than this. The problem is that the original circle integral does not contain
. It contains
, a complex oscillatory expression. By taking its absolute value, we discard the possibility that contributions from different frequencies cancel.
This gives us an important clue. Perhaps the missing saving does not come from proving that is smaller at each frequency. Perhaps it comes from allowing the different frequency contributions to interfere with one another before we estimate them. There are several places where we could look for such cancellation. We might try to exploit oscillation in the variable
inside each arc. We might sum over rational numerators
at fixed denominator
. We might average over the denominators themselves. Or we might return to the original lattice-point problem and use Poisson summation to introduce new oscillatory variables. These are genuinely different possibilities. The first reorganizes the real oscillation. The second and third reorganize the arithmetic oscillation. The fourth uses Fourier analysis in the original integer variables. Kloosterman’s refinement combines these ideas in a particularly effective way.
Farey Dissection
In the ordinary major/minor arc approach, we chose a relatively small parameter and treated neighborhoods of rational numbers
with
as major arcs. The complement was estimated by Weyl’s method. But if the minor-arc estimate is not strong enough, perhaps we should not have such a large undifferentiated minor-arc region. We could instead approximate every frequency by a rational number and analyze its contribution through the corresponding denominator. Dirichlet’s approximation theorem allows this. Given
and a parameter of size
, we can find a reduced fraction
with
and
. The rational numbers with denominators at most
, arranged in increasing order, form the Farey sequence of order
. They divide the circle into intervals on which a particular rational approximation is naturally distinguished.
Why is this useful analytically? Suppose the generating function is of quadratic type. Frequencies within approximately of a rational number have correlated phases. When
is comparable with
, the Farey approximation error is of size
, precisely the scale at which the quadratic phase changes significantly. Thus the Farey dissection of order
is adapted to the natural oscillation scale of the problem. It assigns a rational approximation to essentially every frequency without requiring us to treat a large complement by a separate minor-arc bound. But there is an additional feature that will turn out to be decisive: the lengths of Farey intervals depend on multiplicative inverses modulo the denominator.
Let be a reduced fraction in the Farey sequence, and suppose its immediate left neighbor is
. Consecutive Farey fractions satisfy
. This is an elementary consequence of their reduced denominators and the absence of another Farey fraction between them. It implies
, or
,
where is the multiplicative inverse of
modulo
. The denominator
is not an arbitrary representative of this residue class. It is the unique representative satisfying
. Therefore, once
and
are fixed, the left neighboring denominator is determined by
. Similarly, if
is the immediate right neighbor, then
, giving
. Instead of cutting the interval at its neighboring fractions, it is convenient to use the mediants. The left endpoint is
, and the right endpoint is
. The resulting arc around
has the form
Because are of order
when
, the arc has length comparable with
. Now notice something important. For a fixed denominator
, the arcs corresponding to different numerators
do not have identical endpoints. Their endpoints depend on
. If we eventually sum the contributions over all reduced numerators
, this dependence will introduce phases involving the inverse
. We have not yet defined a Kloosterman sum, but the main arithmetic feature that creates one has already appeared.
We should not immediately introduce complicated exponential sums. First, let us estimate how large an individual Farey-arc contribution might be. For this purpose it is convenient to count integer solutions with a smooth weight. Let be a fixed smooth function of compact support on
. Define
where initially . Then the weighted number of representations of
is
.
The smooth weight is not merely decorative. It allows us to use Poisson summation and integrate by parts repeatedly without generating large boundary terms. If we choose to equal one in a neighborhood of the unit sphere, then for
, every solution of
receives weight one. In that case the weighted count is the ordinary number of four-square representations, including signs and zeros.
Write , where
and
. We want to understand the size of
. The sum contains roughly
lattice points, so the trivial estimate is
. But the phase is periodic modulo
. We expect cancellation among the residue classes, governed by complete quadratic Gauss sums. For four variables, each one-dimensional Gauss sum has size roughly
, so the complete four-dimensional sum has size roughly
rather than
. Dividing by the
residue classes suggests a cancellation factor
. Thus the natural size of the transformed sum should be about
, which is
when
. Each Farey arc has length about
. Therefore its contribution to the circle integral may be of order
How many such arcs are there? There are about
reduced fractions with denominator at most
. Consequently, summing their absolute contributions could give something of order
. But that is exactly the expected size of the main term.
We have arrived at the four-variable barrier again, now in a more informative form. The difficulty is not merely that a fourth-moment estimate fails. The individual Farey-arc contributions are of the right size to produce the main term even when their arithmetic parts are oscillatory. If we take their absolute values and add them, we have no saving. Where can we gain something? The estimate above has treated the roughly reduced numerators at each denominator as unrelated contributions. Perhaps their phases cancel when we sum over
before taking absolute values. If we could save even a fixed positive power of
in that sum, the total over denominators
would become smaller than
. This is the quantitative opportunity that Kloosterman’s refinement exploits.
Poisson Summation
Let us derive the arithmetic expression that appears when we treat a Farey arc more carefully. Write . Separate the lattice points
according to their residue classes
. On each residue class, the factor
is constant, since
has integer coefficients. The remaining sum involves
. Apply Poisson summation on the lattice
. This introduces a dual lattice variable
. To make the formula explicit, define
and the complete exponential sum
The Poisson formula gives the exact identity
This is one of the main transformations in the refined circle method. It separates the original lattice-point sum into complete arithmetic sums and continuous oscillatory integrals
.
The variables have distinct meanings. The denominator describes the rational approximation. The numerator
enters the finite congruence phases. The displacement
enters the real oscillatory integral. The vector
is a Fourier frequency in the original lattice variables. The term
is the continuous approximation one expects from replacing a lattice sum by an integral. The nonzero terms measure the discrepancy between the lattice and that continuous approximation. They are the part that becomes difficult near the four-variable boundary. Let us calculate the complete sums explicitly for the four-square form. Assume temporarily that
is odd. In one variable we encounter
. Since
is invertible modulo
, we can complete the square:
Translating the variable permutes the residues, so the sum becomes
, where
is the quadratic Gauss sum. Taking the product over the four coordinates gives
.
For odd and
, the classical quadratic Gauss-sum evaluation gives
. We can obtain this from the usual quadratic-character evaluation and its multiplicativity over prime-power moduli; the fourth power removes both the quadratic character and the fourth root of unity. Therefore
Substituting this into the Poisson formula, we obtain
Now the multiplicative inverse of has appeared a second time, independently of the Farey endpoints. It arises from completing the square in the quadratic Gauss sum. This is an important structural fact. The inverse is not an accidental consequence of our choice of Farey dissection. It is also built into the arithmetic Fourier transform of a quadratic phase.
The Poisson expression suggests a natural division. First consider . Then the phase involving
disappears, and the contribution is approximately
. After multiplying by
and summing over the reduced numerators, we encounter
This is a Ramanujan sum. It is much simpler than a general exponential sum involving inverses. In particular, when , it equals the Möbius function
, so its absolute value is at most one. The zero dual frequency therefore has strong arithmetic cancellation after the summation over
. Its contributions for small denominators build the singular series, while the integral over
builds the singular integral.
The more difficult terms have . For these, the summation over
produces phases of the form
where denotes the inverse of four modulo the odd modulus
. This leads us naturally to the complete exponential sum
These are the Kloosterman sums. The two terms in the phase have different origins. The term comes from Fourier detection of the original equation
. The term
comes from the arithmetic Fourier transform of the quadratic form. This is why Kloosterman sums appear in the refined circle method. They are not introduced simply because they are interesting objects. They arise when we sum the rational-arc contributions over the numerator
while retaining the oscillation created by the quadratic Gauss sums.
There is still an issue. The Farey arcs have endpoints depending on , so the integral over
is not identical for every numerator. We cannot simply factor out one common integral and immediately apply a complete Kloosterman-sum estimate. We must also understand how the variable endpoints affect the summation. Rather than hide this difficulty, let us examine it directly.
Fix a denominator . Recall that the left and right endpoints of the arc around
involve the neighboring denominators
, with
and
. Suppose we want to sum a contribution over all reduced numerators
. The individual contribution has the form
where represents an integral whose endpoints depend on the neighboring Farey fractions. The essential point is that the weight depends on
, rather than being constant. One possible route would be to replace every Farey arc by a common interval around
. But this creates endpoint errors. At the four-variable boundary, the total size of these errors may be as large as the main term, so we cannot casually discard them. A better route is to retain the dependence on
and expand the weight into finite Fourier modes. Write
where . Substituting this expansion into the sum over
gives
We have transformed the varying endpoint problem into a collection of complete Kloosterman sums. The price is that we must control the total size of the Fourier coefficients of the weight. This can be done when the weight varies in a sufficiently regular way. For example, suppose is obtained by sampling a function of bounded variation around the residue circle. Let
denote the total variation of those samples. Summation by parts gives, for nonzero
,
To see this, write the Fourier sum as a sum of successive differences divided by
. The numerator is bounded by the total variation. The denominator has size comparable to
, and the normalization
cancels the extra factor of
. Summing over the nonzero Fourier modes gives
Thus, provided the endpoint weights are regular enough, their Fourier expansion costs only logarithmic factors. We can still gain a power of from the complete Kloosterman sums. This is a key part of the refinement. We have not assumed that the Farey arcs have equal lengths. We have converted their unequal lengths into an additional arithmetic Fourier sum, which can be analyzed using the same cancellation mechanism.
We should now ask the crucial quantitative question: how large can a Kloosterman sum be?
Bounds on Kloosterman Sums
By definition, is a sum of at most
complex numbers of absolute value one. The trivial estimate is therefore
. But what saving do we actually need? Return to the calculation for a typical Farey arc with
. We found that its contribution could have size
. There are approximately
reduced numerators for a given denominator. Taking absolute values over these numerators gives
. There are about
denominators of size comparable with
. Thus the total contribution from this range could be
, as large as the main term.
Suppose, however, that the sum over numerators has a bound for some exponent
. Then the contribution at one denominator would be about
, and summing over
denominators would give
. This is smaller than
by a power of
. We therefore do not necessarily need the strongest possible Kloosterman estimate. To cross the four-variable barrier, any fixed power saving over the trivial bound
is potentially valuable.
Kloosterman’s original 1926 estimate for a prime modulus was of the form
when the relevant coefficients are nondegenerate. This already saves a factor .
In 1948, Weil proved the much stronger square-root estimate for
. For general moduli, the corresponding bound takes the form
where is the divisor function. The square-root estimate is essentially optimal in general. But it is important that Kloosterman’s own weaker estimate was already nontrivial enough for the kind of improvement he needed.
Can we prove his estimate directly? Yes. The proof is particularly instructive because it uses the same principles as the other exponential-sum estimates we have studied: take a moment, expand it, apply orthogonality, and count the resulting congruence solutions.
Fourth moment Bound
Let be an odd prime. For integers
modulo
, write
.
We want to bound one particular sum , assuming
. Our first thought might be to square its absolute value, just as we did for quadratic Gauss sums. But that does not immediately linearize both
and
. The difference between two inverses introduces a rational expression, and the resulting sum remains difficult. Perhaps we should not estimate one Kloosterman sum in isolation. Instead, let us average a high moment over both coefficients
. The purpose of averaging is that orthogonality in
and
will impose two congruence conditions, one involving the original variables and one involving their inverses. Consider
Expand the fourth power. We obtain a sum over four nonzero residues and over the two frequency variables
. The phase involving
is
, while the phase involving
is
. Summing over
gives zero unless
in
. Similarly, summing over
gives zero unless
. When both conditions hold, the two sums contribute
. Thus
The analytic problem has become an elementary problem about four residues. Let . If
, the equality of reciprocal sums gives
. We may cancel
, obtaining
. But two unordered pairs with the same sum and product are roots of the same quadratic polynomial. Therefore
as multisets. Once
are chosen, there are at most two possibilities for the ordered pair
. If
, we have
and
. The reciprocal-sum equality is then automatic. There are exactly
choices in this case. The total number of quadruples is consequently at most
. Hence
So far we have bounded the average fourth moment. How does this give a pointwise estimate for one particular Kloosterman sum? The missing observation is a symmetry. If , changing variables
gives
As varies over the
nonzero residues, the pairs
are distinct when
. Therefore the fourth-moment sum contains at least
separate terms, each equal to
.Thus
, giving
We have proved Kloosterman’s original nontrivial bound.
Notice the structure of the argument. We did not find cancellation by examining the phases individually. We enlarged the problem to a family of exponential sums, averaged a fourth moment, and used orthogonality to convert the moment into two simultaneous congruences. Those congruences were rigid enough to force most solutions to be diagonal. Finally, a multiplicative symmetry transferred the average estimate back to one particular exponential sum. This is an important method of discovering estimates. When a sum cannot be bounded directly, it may become accessible after being embedded in a family with enough symmetry. The additional averaging variables can create orthogonality relations that are invisible in the original sum.
Weil’s bound says that, for prime modulus, the maximum size is of square-root order. Kloosterman’s elementary fourth-moment argument gives the weaker but still useful bound. The distinction between the two estimates is worth keeping clear. The fourth-moment argument is essentially elementary. Weil’s sharp estimate is substantially deeper, and its original proof uses the arithmetic geometry of curves over finite fields. We should not regard the square-root estimate as a consequence of the elementary argument above.
Incomplete Sums
The Farey-endpoint problem often gives sums in which the inverse variable runs through only part of a residue system. We therefore need to understand how a complete Kloosterman estimate can control an incomplete sum. Fix a modulus and consider
where is an interval containing at most
consecutive integers. We can complete the range using finite Fourier inversion. Define the periodic sequence
for
, and zero otherwise. Its finite Fourier transform at frequency
is
Thus every Fourier coefficient of the periodic sequence is a Kloosterman sum. Expand by Fourier inversion and sum over the interval
. The zero Fourier mode contributes the average value
. The nonzero modes involve geometric sums of the form
, whose absolute values are bounded by
. Summing these bounds over
gives a harmonic series, and therefore a logarithm. We obtain
This completion argument is elementary once the complete-sum bounds are available. For a prime modulus and
, the zero-frequency sum is
. Combining the completion formula with Kloosterman’s original bound gives
The bound is nontrivial compared with the trivial estimate when the interval has length substantially larger than
. With Weil’s square-root bound, the corresponding estimate improves to approximately
. This calculation explains why the incomplete sums arising from Farey endpoints are manageable. We do not need the endpoint weights to be constant. We expand them into Fourier modes, and each mode produces a complete Kloosterman sum. The additional cost is essentially logarithmic when the weights have bounded variation. It also illustrates a limitation. Completing a sum of length much smaller than
may give an estimate larger than the trivial bound. This is a recurring problem in applications of Kloosterman sums: complete sums may have square-root cancellation while very short incomplete sums remain difficult.
Final Estimates
We are now ready to reconsider the estimate that originally failed. Focus on denominators , where the Farey arcs have length
. For these denominators, the Poisson formula gives
for odd . The smoothness of
allows us to integrate by parts in
. Since
and
, the integral decays faster than any prescribed power when
becomes large. Thus only bounded dual frequencies contribute significantly in this range. For each such frequency, the prefactor has size
. Integrating over an arc of length
gives a contribution of order one, as anticipated earlier.
If we now sum the absolute values over the roughly numerators, we obtain
per denominator. Summing over the
denominators in this range gives
. No saving. But the phase contains
. When we sum over the numerators first, the relevant arithmetic expression is a Kloosterman sum. The varying arc endpoints introduce weights depending on
, but the completion argument from the preceding section allows us to treat these weights with only logarithmic losses.
Suppose for the moment that is prime and large. Since our denominators satisfy
, we have
. Therefore the gcd factor in the general Kloosterman bound cannot become large merely because of
. With Weil’s estimate, the sum over numerators contributes at most
, rather than
. Thus the contribution of a typical denominator
is reduced from
to approximately
. There are
such denominators. Hence this range contributes at most
This is the missing power saving. The calculation is especially transparent if we compare the three possible estimates for the numerator sum. The trivial estimate leads to an error of size
. Kloosterman’s original
estimate leads, in this range, to
. Weil’s square-root estimate leads to
. All three estimates concern exactly the same sum. The improvement comes from how much cancellation we can prove among the reduced numerators
.
We should emphasize one technical qualification. This calculation treats the critical range . It does not by itself prove the full four-square asymptotic. For smaller denominators, the real oscillatory integral
and the possible nonzero dual frequencies require additional analysis. For even denominators, the quadratic Gauss sums have a different evaluation. And if
is not prime, we must track factors such as
in the Kloosterman estimates. Nevertheless, the calculation has identified the decisive improvement at the troublesome scale. We now know where the saving can come from and why the older absolute-value estimates failed to detect it.
The range is the most revealing one for understanding Kloosterman’s refinement, but a complete argument must treat all denominators
. For small
, the prefactor
in the Poisson formula is much larger than
. We cannot simply use the previous estimate and sum over all
. That would introduce a large error from the small denominators. This tells us that additional cancellation must occur in the real oscillatory integrals, or that the corresponding contributions belong to the main term rather than the error.
Let us inspect the integral
.
Its phase is . For the four-square form, the gradient is
. If the gradient is bounded away from zero throughout the support of
, then the integral is small by repeated integration by parts. Indeed, one can use the differential operator
, which satisfies
. Moving
onto the smooth weight by integration by parts repeatedly produces inverse powers of the size of the gradient. Thus nonstationary dual frequencies can be made negligible. The difficult case is when
somewhere in the support, meaning
. For our parameters,
and
. The stationary-point condition becomes
. This condition is very informative. On a Farey arc,
, so the quantity
is bounded. Consequently, a stationary point can occur only for dual vectors
of bounded size, since
lies in the fixed compact support of
.
We have therefore reduced the infinite dual sum to finitely many relevant modes, together with a rapidly decaying tail. But those finitely many modes still require careful estimation, because their stationary phases may contribute nontrivially. For the zero mode, the dominant contributions from small denominators build the singular series and singular integral. For nonzero modes, we need to combine stationary-phase estimates with the cancellation in the arithmetic sums over . The precise treatment is technically involved, and this is where the full Kloosterman-refined circle method requires more than the scale calculation we have given.
The important conceptual point is that the two kinds of cancellation work together. The real integral suppresses most dual frequencies and controls their dependence on
. The Kloosterman estimate then controls the remaining arithmetic oscillation over the numerators
. Neither mechanism is sufficient by itself at the four-variable boundary.
Let be a fixed positive-definite integral quadratic form in four variables. We want to count integer vectors
satisfying
. The natural main term has the form
. Because the equation is quadratic and there are four variables, the singular integral has size proportional to
. The singular series records the local densities of solutions modulo powers of each prime. For the four-square form, the real density is particularly simple. The singular integral is the density of points on the sphere
. The surface area of the unit three-sphere is
. This follows, for example, by evaluating the Gaussian integral
in both Cartesian and polar coordinates. Passing from surface measure on the sphere of radius
to the density for the equation
introduces the factor
. Thus the singular integral is
. We expect
where counts representations by four integer squares, including signs and zeros. The Kloosterman-refined circle method makes this expectation quantitative. A strong form available by 1996, through Heath-Brown’s treatment of quadratic forms, gives for a fixed positive-definite nonsingular quaternary quadratic form
The main term is of order whenever the relevant local densities stay bounded below. The error is then smaller by a power of
. Observe that
. This is precisely the scale predicted by our critical-range calculation using square-root cancellation in Kloosterman sums.
The full theorem requires uniform control over the small denominators, even moduli, stationary phases, and arithmetic gcd factors discussed above. We have derived the principal algebraic identities, proved the original nontrivial Kloosterman bound, and located the main power saving explicitly. The remaining uniform estimates are the substantial analytic part of the complete theorem; they should not be inferred from the critical-range calculation alone.
There is one arithmetic caution. For some integers, the local densities may become very small. In the ordinary four-square problem, for example, Jacobi’s exact formula gives
For an odd prime , this becomes
, which is of order
. But for
with
, the exact count is only twenty-four. Thus we cannot claim that every four-square representation count is uniformly of order
. The singular series must retain the arithmetic dependence on
. A power-saving error of size
is most informative when the singular series is sufficiently large. The exact Jacobi formula is stronger than what the circle method is needed to prove for this particular quadratic form. But Kloosterman’s contribution was to develop an analytic mechanism that also applies to much more general forms, where no comparable elementary divisor formula is available.
Let us compare the estimates once more, but now from the standpoint of trying to discover the proof. We began with an error of size , exactly the size of the expected main term. It was natural to ask whether Weyl’s inequality could be improved, or whether a higher moment might help. But at four variables, the fourth moment is already at the critical scale. Better pointwise estimates alone do not automatically solve the problem, because there is no spare variable on which to apply them. The next possibility was to stop estimating the integral by its absolute value and instead analyze the contributions of individual Farey arcs. That revealed a more concrete obstruction: roughly
rational arcs, each potentially contributing about one. We then asked whether these contributions were genuinely independent in size or whether they had related arithmetic phases. The Poisson formula revealed the phases
. Their sum over the rational numerator
was not an arbitrary collection of complex numbers. It was a Kloosterman sum. At that point the target estimate became clear. We needed a power saving in a complete finite exponential sum. Kloosterman’s
bound already provided one. Weil’s later
bound gave the stronger saving.
This illustrates an important principle in analytic number theory: when an estimate is exactly at the critical scale, it is useful to ask which summation was discarded by the absolute-value inequality. Sometimes the next step is not to estimate each summand more strongly, but to reorganize the argument so that another summation can be performed before taking absolute values. Kloosterman’s refinement does precisely this. The summation over rational numerators becomes a source of cancellation rather than a source of multiplicity.
Once we recognize the importance of cancellation over the rational numerators, several further questions become natural. Could we also gain cancellation when summing over the denominators ? The Kloosterman sums associated with different moduli are not independent objects. They have multiplicative properties, and they also arise in the Fourier expansions of automorphic forms. This suggests that a sum over moduli might exhibit additional cancellation not visible in an individual Weil bound. Such questions lead toward the spectral theory of automorphic forms and the Kuznetsov trace formula. The spectral method relates averages of Kloosterman sums to sums over automorphic eigenfunctions. It can be particularly useful when estimating a family of Kloosterman sums is easier than estimating each one separately. Another possibility is to reorganize the circle method so that the summation over numerators is incorporated from the beginning. The smooth delta-symbol methods developed by Duke, Friedlander, and Iwaniec, and further by Heath-Brown, do exactly this in a different form. Instead of first writing an ordinary circle integral and then partitioning it into Farey arcs, they express the equation detector through a smooth sum over moduli and reduced residue classes. Poisson summation then leads directly to complete exponential sums and oscillatory integrals. These methods retain the same basic arithmetic ingredients, but they avoid some of the complications created by variable Farey endpoints.
There is also the possibility of using modular forms. For positive-definite quaternary quadratic forms, the theta series is a modular form of weight two. Its Eisenstein part encodes the local-density main term, while its cusp-form part measures the error. Bounds for Fourier coefficients of cusp forms can therefore give stronger error terms than the elementary circle method. In particular, for the appropriate fixed integral forms, theory gives error terms of order through the Ramanujan–Petersson bounds proved by Deligne. For the sum of four squares itself, the cusp-form contribution vanishes in the relevant classical identity, which is why Jacobi’s formula is exact.
These are different ways of accessing the cancellation left over after the principal arithmetic and geometric terms have been separated. The circle method makes the cancellation visible through rational approximations and exponential sums. Modular forms reorganize it through spectral decomposition. The delta method uses a smoother arithmetic decomposition of the equation detector. The methods are related, but they do not make exactly the same information easy to see.
The central difficulty in the four-variable problem was not the lack of a Fourier representation. We already had one. Nor was it the inability to calculate the main term. The real density and the local congruence densities could be identified by the usual major-arc analysis. The problem was that our available error estimates were of the same order as the main term. We therefore asked where a further saving might be hidden. The fourth-moment method had reached its natural limit. The individual rational arcs were not small enough. The crucial change was to retain the arithmetic phases when summing those arcs, rather than replacing every contribution by its absolute value. The Farey dissection made multiplicative inverses appear through the geometry of neighboring rational fractions. Poisson summation produced another inverse through completion of the square in quadratic Gauss sums. Together, these led to exponential sums with phases . The original Kloosterman bound showed that these sums have nontrivial cancellation. Its fourth-moment proof is a particularly instructive example of how one can discover such cancellation: enlarge the family of sums, use orthogonality to produce simultaneous congruences, count their solutions, and then use symmetry to recover a pointwise estimate. The resulting saving is not merely a technical improvement in a bound. It changes the scale of the entire circle-method error. At the critical denominator range, absolute values give
, Kloosterman’s original bound suggests
, and square-root cancellation gives
. The broader lesson is that one should not automatically interpret a failed analytic estimate as evidence that the individual exponential sums must be made smaller. Sometimes the problem is the order in which the sums and integrals are being estimated. Taking absolute values too soon may destroy precisely the arithmetic cancellation that makes the theorem true. Kloosterman’s refinement is a powerful example of this principle. By reorganizing the rational contributions and exploiting their inverse-residue phases, it finds a saving that the elementary major/minor arc argument cannot see. The same search for hidden arithmetic averaging has continued to shape analytic number theory, from classical quadratic forms to automorphic trace formulas and the delta method.