Circle Method and Sums of Squares

Suppose we want to count the number of ways an integer can be represented as a sum of special kinds of numbers. We might ask how many solutions there are to x_1^2+\cdots+x_s^2=N , or whether every sufficiently large odd integer is a sum of three primes. These are elementary questions to state, but the variables are coupled by an additive equation. Choosing one variable affects what is possible for all the others, and the arithmetic restrictions on squares or primes make direct counting difficult. The circle method begins with a remarkable observation: although the equation couples the variables, there is an exact Fourier identity that separates them. Once this is done, we replace the problem of counting solutions by the problem of estimating exponential sums. The resulting integral is still difficult, but its difficulty is organized in a useful way. Some frequencies correspond to rational numbers with small denominators, where the exponential sums are structured and can be approximated explicitly. At most other frequencies, the phases oscillate sufficiently rapidly to produce cancellation.

The method tries to exploit both phenomena at once. The structured frequencies give the main term, while cancellation elsewhere must be strong enough to make the remaining contribution smaller. The first part involves congruences and complete exponential sums; the second involves Diophantine approximation, differencing, mean values, and increasingly sophisticated estimates. A successful argument must make these two parts fit quantitatively. The central question is therefore not simply why rational numbers appear, or why Fourier analysis is useful. We want to understand how an arithmetic counting problem creates peaks in frequency space, why the peaks are organized by rational denominators, and how much cancellation is required away from them. We will begin with an almost trivial additive problem, where everything can be calculated exactly. We will then study quadratic exponential sums and see rational peaks emerge. After that, we will prove a nontrivial minor-arc estimate using Weyl differencing and a fourth-moment calculation. Finally, we will combine these ingredients to obtain an asymptotic formula for sums of five squares, including the singular series, the singular integral, and an explicit power-saving error.

The discussion will also explain why essentially the same argument fails for four squares, what changes when we replace squares by primes, and how Kloosterman’s refinement and the later delta-symbol methods reorganize the analysis.

Counting Solutions

Let e(t)=e^{2\pi it} . The essential identity is

\displaystyle \int_0^1 e(\alpha n) d\alpha=\begin{cases}1,&n=0,\\ 0,&n\neq0,\end{cases}\qquad n\in\mathbb Z.

There is nothing mysterious in this formula. If n=0 , the integrand is one. Otherwise, integrating the exponential gives (e(n)-1)/(2\pi in)=0 , since e(n)=1 for integers. We have therefore constructed an exact detector of the equation n=0 . Now suppose we want to count solutions of

\displaystyle x_1^k+\cdots+x_s^k=N

with positive integers x_i . Put P=N^{1/k} and define F_k(\alpha)=\sum_{1\le x\le P}e(\alpha x^k) . When we expand F_k(\alpha)^s , each tuple (x_1,\ldots,x_s) contributes e(\alpha(x_1^k+\cdots+x_s^k)) . Multiplying by e(-\alpha N) and integrating kills every tuple except those satisfying the desired equation. Thus

\displaystyle R_{k,s}(N)=\int_0^1F_k(\alpha)^s e(-N\alpha) d\alpha.

This identity is exact. We have not used an approximation, or any estimate of an exponential sum. We have simply expressed the counting function as a Fourier coefficient. Why is this helpful? The original equation ties together s variables. After inserting the Fourier integral, the variables separate into s copies of the same one-variable sum. We have traded a coupled Diophantine problem for the analysis of a single function F_k(\alpha) and its powers. Of course, this trade is worthwhile only if we can estimate the integral. If we use the trivial bound |F_k(\alpha)|\le P , we obtain R_{k,s}(N)\le P^s=N^{s/k} . But this is much larger than the expected answer. An equation of degree k should ordinarily remove roughly P^k=N possibilities, suggesting a main term of size P^{s-k}=N^{s/k-1} . The circle method must therefore save a factor of approximately P^k relative to this trivial bound. Understanding where that saving comes from is the main analytic problem.

Example:

Before studying squares, consider the number of positive integer solutions to x+y+z=N . By elementary counting, the answer is \binom{N-1}{2}=(N-1)(N-2)/2 . We want to see how the same answer appears in frequency space. Define L_N(\alpha)=\sum_{x=1}^{N}e(\alpha x) . Then the number of solutions is exactly \int_0^1L_N(\alpha)^3e(-N\alpha) d\alpha . The sum L_N is a geometric progression, so we can calculate it explicitly:

\displaystyle L_N(\alpha)=e\left(\frac{N+1}{2}\alpha\right)\frac{\sin(\pi N\alpha)}{\sin(\pi\alpha)}.

This formula already shows the basic phenomenon behind major and minor arcs. When \alpha is close to an integer, the denominator \sin(\pi\alpha) is small, and the sum may have size comparable to N . Away from integers, the denominator is bounded away from zero, and cancellation makes the sum much smaller. Indeed, if |\alpha| denotes the distance from \alpha to the nearest integer, the geometric progression gives

\displaystyle |L_N(\alpha)|\le\min\left(N,\frac1{2|\alpha|}\right).

The point is not just that the sum is sometimes large and sometimes small. The scale on which it changes is also visible. Near \alpha=0 , the sum remains of order N when |\alpha|\lesssim1/N . That neighborhood has length about 1/N . Since the integrand contains the third power of the sum, the contribution from this scale is naturally of order N^3\cdot N^{-1}=N^2 , exactly the order of the true answer.

This gives us our first useful heuristic: the size of the main term is the height of a structured peak multiplied by its width. For a linear problem with three variables, the peak has height roughly N^3 and width roughly N^{-1} , producing N^2 . We can also recover the exact answer directly from the Fourier integral. Expanding the cube counts the triples with x+y+z=N . For each choice of x between one and N-2 , there are N-x-1 possibilities for y,z . Summing gives \sum_{x=1}^{N-2}(N-x-1)=(N-1)(N-2)/2 . Thus Fourier analysis has changed our viewpoint. Instead of regarding the count as a sum over integer triples, we can regard it as the accumulated contribution of oscillatory frequencies. The large contribution near rational frequencies is the beginning of the circle method.

For squares, however, the important frequencies are not confined to neighborhoods of integers. Rational numbers such as 1/3,2/5,3/7 also produce unusually large exponential sums. We need to understand why.

Squares

Let F(\alpha)=\sum_{1\le x\le P}e(\alpha x^2) , where P is large. Suppose first that \alpha=a/q , with (a,q)=1 . Then the phase e(ax^2/q) is periodic in x modulo q . If P is divisible by q , we can divide the sum into complete blocks of length q . Every block contributes the same quantity,

\displaystyle S(q,a)=\sum_{r\bmod q}e(ar^2/q),

which is called a complete quadratic Gauss sum. Thus F(a/q)=(P/q)S(q,a) when q\mid P . For general P , there is an error of size at most q from the incomplete final block. This is why rational numbers with small denominators matter. At such a frequency, the exponential phases repeat periodically, so cancellation is governed by a fixed complete sum rather than by all P terms individually. Take q=3 . Since the squares modulo three are zero and one, with one occurring twice, we have S(3,1)=1+2e(1/3)=i\sqrt3 . Thus, when 3\mid P , F(1/3)=iP/\sqrt3. This is a striking calculation. The exponential sum has P terms, each of absolute value one. Even though the phases are not all equal, their periodic arrangement leaves a sum of size proportional to P . The frequency 1/3 is therefore a genuine peak.

Compare this with a frequency that is well approximated only by rationals of large denominator. There the phases do not repeat in short blocks, and we should expect more cancellation. To quantify the rational peaks, we need an estimate for S(q,a) . There is a simple argument that gives essentially the best possible bound for quadratic sums. Expand the square of its absolute value:

\displaystyle  |S(q,a)|^2=\sum_{x,y\bmod q}e(a(x^2-y^2)/q).

Write x=y+h . Then x^2-y^2=2hy+h^2 , so the inner sum over y becomes \sum_{y\bmod q}e(2ahy/q) . This is zero unless q\mid2ah , in which case it equals q . Because (a,q)=1 , the condition reduces to q\mid2h . There are at most two such residue classes h\bmod q . Therefore

\displaystyle |S(q,a)|\le\sqrt{2q},\qquad (a,q)=1.

When q is odd, only h=0 contributes, and in fact |S(q,a)|=\sqrt q . This calculation is our first genuinely nontrivial exponential-sum estimate. A trivial bound would give |S(q,a)|\le q . The argument saves a square root, and that saving will be crucial when we sum over many rational denominators. There is also a useful principle behind the proof. The original phase was quadratic. Squaring the sum and introducing the difference h=x-y transformed the phase into something linear in y . Linear exponential sums can be evaluated exactly. Thus differencing lowers the degree of the phase, making cancellation accessible. This is the fundamental idea behind Weyl’s method. We will use it again for incomplete sums.

Major Arcs

Write a frequency as \alpha=a/q+\beta , where (a,q)=1 and \beta is small. Then

\displaystyle F(\alpha)=\sum_{x\le P}e(ax^2/q)e(\beta x^2).

The first factor is periodic modulo q . The second varies slowly when |\beta|P^2 is small or moderately bounded. This suggests splitting the integers x into residue classes modulo q . Within each residue class, the periodic factor is constant, while the slowly varying factor can be approximated by an integral. Define I(\beta)=\int_0^P e(\beta t^2) dt . A summation-by-parts calculation gives

\displaystyle F(a/q+\beta)=\frac{S(q,a)}q I(\beta)+O\bigl(q(1+|\beta|P^2)\bigr).

Let us derive the error term. For a residue class r\bmod q , the number of integers x\le t satisfying x\equiv r\pmod q differs from t/q by at most a constant. Hence, for a differentiable function g , summation by parts shows that the sum of g(x) over that residue class differs from q^{-1}\int_0^P g(t) dt by at most a constant times 1+\int_0^P|g'(t)| dt . Take g(t)=e(\beta t^2) . Then |g'(t)|=4\pi|\beta|t , so the integral of its absolute derivative is O(|\beta|P^2) . We get an error O(1+|\beta|P^2) for each residue class. Summing over the q classes gives the stated estimate.

The approximation has a clear interpretation. The factor S(q,a)/q records the finite arithmetic behavior modulo q . The integral I(\beta) records how the phase varies continuously as we move away from the rational frequency a/q . These are two different scales of the problem: a finite congruence scale and a real oscillatory scale. Their product describes the shape of the peak near a/q . We also need to know how rapidly I(\beta) decays. For |\beta|P^2\le1 , the trivial bound |I(\beta)|\le P is sufficient. For larger \beta , substitute u=\sqrt{|\beta|} t . Then

\displaystyle I(\beta)=|\beta|^{-1/2}\int_0^{P\sqrt{|\beta|}}e(\pm u^2)du .

The last integral is uniformly bounded. To see this, split it at u=1 . The part below one is bounded trivially. Beyond one, integrate by parts using (e(u^2))'=4\pi iue(u^2) . The resulting boundary terms and the integral of u^{-2} are bounded. Hence

\displaystyle |I(\beta)|\ll P(1+P^2|\beta|)^{-1/2}.

The natural width of a quadratic peak is therefore approximately P^{-2} . This agrees with the heuristic from the first example: for a degree-two equation, the main contribution should have width P^{-2} . At this point we have enough information to describe the major arcs. But we still need to know what happens elsewhere.

Minor Arcs

A reasonable first guess is that F(\alpha) should be small whenever \alpha is not close to a rational number with a small denominator. But this needs proof. The phases e(\alpha x^2) oscillate in a complicated way, and ordinary geometric-series estimates do not apply directly because the phase is quadratic. We can, however, use the same differencing idea that worked for complete Gauss sums.

Expand

\displaystyle |F(\alpha)|^2=\sum_{x,y\le P}e(\alpha(x^2-y^2)).

Separate the diagonal terms x=y , which contribute P+O(1) . For the off-diagonal terms, write x=y+h . The phase becomes \alpha(2hy+h^2) . For each fixed h , the dependence on y is linear. Using the geometric-series bound from Section 2, we obtain

\displaystyle |F(\alpha)|^2\ll P+\sum_{1\le h\le P}\min\left(P,\frac1{|2h\alpha|}\right).

This inequality is one of the basic forms of Weyl differencing. It replaces an incomplete quadratic sum by a collection of incomplete linear sums. The price is that we must understand how often 2h\alpha comes close to an integer. Suppose \alpha is approximated by a reduced fraction a/q with |\alpha-a/q|\le q^{-2} . The values 2ah/q lie on a rational grid. On a block of h values of length at most a small constant multiple of q , distinct grid points are separated by at least a constant times 1/q . The perturbation from \alpha-a/q is small enough that this separation remains valid, after shortening the blocks by a fixed factor if necessary. Thus, within one block of O(q) consecutive integers, the fractional parts 2h\alpha cannot cluster arbitrarily closely. There are at most a bounded number in each interval of length about 1/q . Summing \min(P,|2h\alpha|^{-1}) over the block gives O(P+q\log(2q)) : one or two points may lie very close to an integer and contribute as much as P , while the remaining points contribute a harmonic sum of size q\log q . There are O(P/q+1) such blocks. Substituting the resulting estimate into the differencing inequality yields

\displaystyle |F(\alpha)|^2\ll\left(\frac{P^2}{q}+P+q\right)\log(2q),\qquad |\alpha-a/q|\le q^{-2}.

This is the quadratic Weyl estimate in a form convenient for the circle method. We should understand what it says. If q is small, the term P^2/q may be large, reflecting the fact that \alpha lies near a major rational frequency. If q is extremely large, the term q may also be large. But when q lies in a substantial intermediate range, all three terms are smaller than P^2 , and we obtain a nontrivial saving over the bound |F(\alpha)|\le P . This explains why rational approximation enters the method. We do not merely approximate real numbers by rationals because Dirichlet’s theorem is available. We do so because the denominator of the approximation controls the amount of cancellation in the exponential sum.

Decomposition into Major and Minor Arcs

How do we choose major and minor arcs? We now need a systematic decomposition of the interval [0,1] . Choose a parameter Q satisfying 1\ll Q\ll P . Later we will take Q=P^{1/10} . Around every reduced fraction a/q with q\le Q , define a major arc by the condition |\alpha-a/q|\le Q/P^2 , interpreting the interval modulo one. The union of these neighborhoods is the set of major arcs, denoted \mathfrak M . The complementary set is the set of minor arcs, \mathfrak m .

Why choose this particular width? The rational approximation from before is accurate when q and |\beta|P^2 are both reasonably small. On our major arcs, we have q\le Q and |\beta|P^2\le Q , so the error in that approximation is at most O(Q^2) . If Q is a sufficiently small power of P , this error is much smaller than the natural size P of the leading term.

The major arcs are also disjoint when Q^3\ll P^2 . Indeed, distinct reduced fractions a/q and a'/q' with denominators at most Q differ by at least 1/(qq')\ge Q^{-2} . Our arcs have radius Q/P^2 , so they cannot overlap if 2Q/P^2<Q^{-2} . The total measure of the major arcs is at most a constant times Q^3/P^2 , because there are at most O(Q^2) reduced fractions with denominator at most Q , and each arc has length 2Q/P^2 .

What about the minor arcs? We want to show that every minor-arc frequency has a rational approximation with denominator neither too small nor too large, so that the Weyl bound gives cancellation. Dirichlet’s approximation theorem supplies exactly this. For any real \alpha and any positive integer R , there exist coprime integers a,q with 1\le q\le R and |\alpha-a/q|\le1/(qR) . This follows by considering the R+1 fractional parts 0,\alpha,\ldots,R\alpha and placing them into R intervals of equal length. Two lie in the same interval, so their difference gives an integer q with |q\alpha|\le1/R . Take R=P^2/Q , with harmless integer rounding. If the resulting denominator satisfied q\le Q , then |\alpha-a/q|\le Q/(qP^2)\le Q/P^2 , and \alpha would belong to a major arc. Therefore every minor-arc frequency has an approximation satisfying Q<q\le P^2/Q . It also satisfies |\alpha-a/q|\le q^{-2} , since q\le R . Substituting this range into the Weyl estimate gives

\displaystyle \sup_{\alpha\in\mathfrak m}|F(\alpha)|\ll P Q^{-1/2}(\log P)^{1/2}.

This is the minor-arc saving we wanted. Compared with the trivial bound P , we gain approximately a factor of Q^{1/2} . But is a pointwise saving enough? It depends on how many variables we have.

Suppose we want to count representations of N as a sum of five squares. The Fourier integral contains F(\alpha)^5 . On the minor arcs, the pointwise Weyl bound alone would give |F(\alpha)|^5\ll P^5Q^{-5/2}(\log P)^{5/2} . Since the minor arcs have measure at most one, this is also a bound for their integral. But the expected main term has size P^3 . For our eventual choice Q=P^{1/10} , the pointwise estimate would give roughly P^{5-1/4} , far larger than P^3 . So the pointwise bound is not enough. This is a fundamental issue in the circle method. Even a substantial pointwise saving may be inadequate when the integrand is a high power of an exponential sum. We also need to know that the sum is rarely large. That information is encoded by its moments.

Consider the fourth moment

\displaystyle \int_0^1|F(\alpha)|^4 d\alpha.

Expanding the fourth power and using Fourier orthogonality, this integral counts the solutions of x_1^2+x_2^2=x_3^2+x_4^2 with all variables between one and P . We can estimate this number by elementary factorization. Rewrite the equation as x_1^2-x_3^2=x_4^2-x_2^2 . If x_1=x_3 , then x_4=x_2 , giving O(P^2) diagonal solutions. Now fix x_1\neq x_3 , and put M=|x_1^2-x_3^2| . The remaining equation has the form |x_4^2-x_2^2|=M . Factoring the difference of squares gives |x_4-x_2|(x_4+x_2)=M . Each solution therefore determines a factorization of M into two positive integers. Once the factors are chosen, x_4 and x_2 are determined, if they are integers at all. Consequently, for each fixed pair x_1,x_3 , the number of possible pairs x_2,x_4 is at most a constant times the divisor function d(M) . The elementary divisor estimate says that, for every \varepsilon>0 , we have d(M)\ll_\varepsilon M^\varepsilon . This follows from the prime factorization M=\prod p^{a_p} , since d(M)=\prod(a_p+1) . For sufficiently large primes, a+1\le p^{\varepsilon a} for every positive a ; the finitely many smaller primes contribute a constant depending on \varepsilon . Multiplying gives the bound. Since M\le P^2 and there are O(P^2) choices of x_1,x_3 , we obtain

\displaystyle \int_0^1|F(\alpha)|^4 d\alpha\ll_\varepsilon P^{2+\varepsilon}.

This is a powerful estimate. A trivial bound would give P^4 . The fourth moment saves almost two powers of P .

We can now combine the mean-value estimate with the pointwise minor-arc bound. Write |F|^5=|F|\cdot|F|^4 . On the minor arcs, bound the first factor by its supremum and integrate the remaining four factors over the whole unit interval. We obtain

\displaystyle \int_{\mathfrak m}|F(\alpha)|^5 d\alpha\ll_\varepsilon P^{3+\varepsilon}Q^{-1/2}.

This is exactly the kind of estimate we need. The expected main term has size P^3 , and the minor arcs are smaller by a power of Q .

The calculation illustrates a general strategy in analytic number theory. A pointwise bound tells us that a function cannot be too large on a particular set. A moment bound tells us that large values cannot occur too frequently. Combining the two can be much stronger than either estimate alone. For squares, the fourth moment has an elementary Diophantine interpretation and can be bounded using the divisor function. For higher powers, the corresponding mean-value problems become much harder. Their analysis was a central part of the classical development of the circle method, particularly in the work of Hua and Vinogradov.

Sums of Five Squares

We now have enough tools to prove a genuine circle-method theorem. Let R_5(N) denote the number of ordered representations

\displaystyle N=x_1^2+x_2^2+x_3^2+x_4^2+x_5^2

with all five variables positive integers. Put P=\sqrt N and F(\alpha)=\sum_{1\le x\le P}e(\alpha x^2) . Fourier orthogonality gives

\displaystyle R_5(N)=\int_0^1F(\alpha)^5e(-N\alpha) d\alpha .

Take Q=P^{1/10} , ignoring harmless integer rounding. Divide the circle into the major arcs \mathfrak M around reduced fractions a/q with q\le Q and width Q/P^2 , and the complementary minor arcs \mathfrak m . The preceding section already gives the minor-arc estimate

\displaystyle  \int_{\mathfrak m}|F(\alpha)|^5 d\alpha\ll_\varepsilon P^{3+\varepsilon}Q^{-1/2} .

We must now calculate the major-arc contribution. On an arc centered at a/q , write \alpha=a/q+\beta . The approximation from above gives

\displaystyle mF(\alpha)=q^{-1}S(q,a)I(\beta)+O(Q^2) ,

because q\le Q and |\beta|P^2\le Q . Both F(\alpha) and its main approximation have size O(P) . Therefore replacing the fifth power of F by the fifth power of the main approximation introduces an error O(P^4Q^2) at each point of a major arc. The total measure of the major arcs is O(Q^3/P^2) . Integrating the error over all major arcs consequently gives O(P^2Q^5) . Thus the main major-arc contribution is obtained by integrating

\left(\frac{S(q,a)}q\right)^5 I(\beta)^5e(-aN/q)e(-\beta N)

over |\beta|\le Q/P^2 , and summing over all reduced fractions a/q with q\le Q .

Notice that the variables q,a and \beta have separated. The complete Gauss sums depend on the rational part of the frequency, while the integral I(\beta) depends on the small displacement from that rational frequency. This leads naturally to two factors.

Define the singular series

\displaystyle \mathfrak S_5(N)=\sum_{q=1}^{\infty}\ \sum_{\substack{a\bmod q\\(a,q)=1}}\left(\frac{S(q,a)}q\right)^5e(-aN/q)

and the singular integral

\displaystyle \mathfrak J_5(N)=\int_{-\infty}^{\infty}I(\beta)^5e(-N\beta) d\beta.

Why can we extend the finite sums and integrals to these infinite expressions? We need quantitative bounds on their tails. For the singular series, our Gauss-sum estimate gives |S(q,a)/q|\le\sqrt{2/q} . There are at most q reduced residues a\bmod q , so the total contribution from denominator q is at most a constant times q\cdot q^{-5/2}=q^{-3/2} . The series \sum_{q\ge1}q^{-3/2} converges. In particular, the contribution from denominators q>Q is O(Q^{-1/2}) . For the singular integral, we know that |I(\beta)|\ll P(1+P^2|\beta|)^{-1/2} . The integral of its fifth power therefore converges absolutely. The tail outside |\beta|\le Q/P^2 is bounded by

\displaystyle  P^5\int_{|\beta|>Q/P^2}(1+P^2|\beta|)^{-5/2} d\beta\ll P^3Q^{-3/2}.

Combining these tail bounds with the earlier major-arc approximation gives

\displaystyle R_5(N)=\mathfrak S_5(N)\mathfrak J_5(N)+O_\varepsilon\left(P^{3+\varepsilon}Q^{-1/2}+P^2Q^5\right).

The first error term includes the minor-arc contribution and the tail of the singular series. The second comes from approximating the original exponential sum by the product of its complete Gauss sum and real integral on the major arcs. The tail of the singular integral is smaller than these terms.

We now see why the parameter Q was introduced. Increasing Q improves the minor-arc bound, because more frequencies are classified as major and the remaining rational approximations have larger denominators. But increasing Q also makes the major-arc approximation less accurate and enlarges the total length of the major arcs. We must choose a power of P that balances these competing requirements.

With Q=P^{1/10} , the first error term is O_\varepsilon(P^{3-1/20+\varepsilon}) , while the second is O(P^{5/2}) . The latter is smaller. Therefore

\displaystyle R_5(N)=\mathfrak S_5(N)\mathfrak J_5(N)+O_\varepsilon(P^{3-1/20+\varepsilon}).

We have obtained a power-saving error. It remains to calculate the singular integral and understand the singular series.

Singular Integral

The singular integral arose from the continuous approximation to the quadratic exponential sum. We should calculate it explicitly rather than treat it as an unknown constant. Recall that I(\beta)=\int_0^P e(\beta t^2) dt . Expanding its fifth power and interchanging the integrations, in the sense justified by a standard Gaussian cutoff and subsequent limit, shows that \mathfrak J_5(N) measures the density of real points satisfying t_1^2+\cdots+t_5^2=N in the region 0\le t_i\le P . Because P^2=N and all the terms t_i^2 are nonnegative, every nonnegative real solution automatically satisfies t_i\le P . Thus the upper bounds do not affect the calculation. Make the substitutions u_i=t_i^2 . Then dt_i=\frac12u_i^{-1/2}du_i . The singular integral becomes the density of the sum u_1+\cdots+u_5=N , weighted by 2^{-5}\prod_i u_i^{-1/2} . The classical Dirichlet integral evaluates this density. More generally, for positive numbers a_1,\ldots,a_s , the integral over the simplex u_i\ge0 and \sum_i u_i=N , with weight \prod_i u_i^{a_i-1} , equals N^{a_1+\cdots+a_s-1}\prod_i\Gamma(a_i)/\Gamma(a_1+\cdots+a_s) . We can derive this from the defining gamma integral by multiplying the one-variable integrals \Gamma(a_i)=\int_0^\infty u^{a_i-1}e^{-u}du and changing variables from u_1,\ldots,u_s to their sum and their proportions. The integral over the sum gives the denominator gamma factor, while the proportions give the simplex integral. Taking a_i=1/2 for all five variables, and remembering the factor 2^{-5} , yields

\displaystyle \mathfrak J_5(N)=\frac{\Gamma(3/2)^5}{\Gamma(5/2)}N^{3/2}=\frac{\pi^2}{24}N^{3/2}.

The power N^{3/2} is exactly what we expected from dimension counting. There are five variables, each naturally of size N^{1/2} , and one quadratic equation removes one power of N . Thus the natural order is N^{5/2-1}=N^{3/2} . The singular integral gives the precise real geometric constant, including the fact that we count only positive coordinates. But it knows nothing about congruence restrictions. Those are encoded by the singular series.

The Singular Series

The singular series is not merely a technical artifact of summing over rational frequencies. It has a concrete interpretation in terms of solutions to congruences. For a positive integer q , let M_q(N) denote the number of solutions to x_1^2+\cdots+x_5^2\equiv N\pmod q , with each variable taken modulo q . Using the finite Fourier identity q^{-1}\sum_{a\bmod q}e(am/q)=1 when q\mid m and zero otherwise, we obtain

\displaystyle \frac{M_q(N)}{q^4}=q^{-5}\sum_{a\bmod q}S(q,a)^5e(-aN/q).

Now take q=p^r , where p is prime. Every fraction a/p^r reduces uniquely to a fraction b/p^j with 0\le j\le r and (b,p)=1 when j>0 . The complete exponential sum corresponding to the unreduced fraction is just a repeated copy of the sum for the reduced denominator. Sorting the terms by reduced denominator gives

\displaystyle \frac{M_{p^r}(N)}{p^{4r}}=\sum_{j=0}^{r}\ \sum_{\substack{b\bmod p^j\\(b,p)=1}}\left(\frac{S(p^j,b)}{p^j}\right)^5e(-bN/p^j),

where the term with j=0 is interpreted as one. The Gauss-sum estimates show that the expression on the right converges as r\to\infty . Define

\displaystyle \sigma_p(N)=\lim_{r\to\infty}\frac{M_{p^r}(N)}{p^{4r}}.

This is the local density of solutions at the prime p . The denominator p^{4r} is natural: an equation in five variables should leave approximately four independent residue choices modulo p^r . The Chinese remainder theorem implies that these local factors multiply. More precisely, the complete exponential sums factor under coprime moduli, and rearranging the absolutely convergent singular series gives

\displaystyle \mathfrak S_5(N)=\prod_p\sigma_p(N).

This identity reveals the arithmetic meaning of the singular series. It measures how the number of representations is modified by congruence restrictions at every prime. We can make this visible already at p=3 . Earlier we calculated S(3,1)=i\sqrt3 and S(3,2)=-i\sqrt3 . The contribution from denominator three to the singular series is

\displaystyle  A(3,N)=3^{-5}\bigl(S(3,1)^5e(-N/3)+S(3,2)^5e(-2N/3)\bigr).

Substituting the Gauss sums and simplifying gives A(3,N)=0 when N\equiv0\pmod3 , A(3,N)=1/9 when N\equiv1\pmod3 , and A(3,N)=-1/9 when N\equiv2\pmod3 . Thus the first correction arising from modulus three depends on the residue class of N . The real geometry of the five-square equation does not distinguish these residue classes, but the arithmetic does. This example explains why the main term is not simply a universal constant times N^{3/2} . It must contain a factor that responds to the arithmetic of N .

To conclude that R_5(N) is asymptotic to a positive main term, we must show that the singular series does not vanish. Absolute convergence alone is not enough. We will prove something stronger: the singular series is bounded below by a positive constant independent of N . First consider an odd prime p . Let M_p(N) count the solutions modulo p . The finite Fourier formula and the Gauss-sum estimate give

\displaystyle M_p(N)=p^4+O(p^{5/2}).

Indeed, the zero frequency contributes p^4 , while the p-1 nonzero frequencies each involve a fifth power of a Gauss sum of absolute value \sqrt p , divided by p . For odd p , a solution modulo p is nonsingular as soon as at least one coordinate is nonzero modulo p , because the derivative of x_1^2+\cdots+x_5^2-N with respect to that coordinate is 2x_i . At most one solution, the all-zero vector, is singular in this sense. Therefore there are at least p^4-p^{5/2}-1 nonsingular solutions modulo p . Each nonsingular solution lifts to exactly p^{4(r-1)} solutions modulo p^r . To see why, suppose we have a solution modulo p^j with x_1\not\equiv0\pmod p . Replace each coordinate by x_i+p^jt_i . Modulo p^{j+1} , the change in the sum of squares is 2p^j\sum_i x_it_i . After choosing t_2,\ldots,t_5 arbitrarily modulo p , the congruence determines a unique t_1 , because 2x_1 is invertible modulo p . Thus every nonsingular solution has exactly p^4 lifts at each step. Consequently,

\displaystyle \sigma_p(N)\ge1-p^{-3/2}-p^{-4}

for every odd prime p . These lower bounds are positive. Furthermore, the infinite product of them converges to a positive number, because the sum of p^{-3/2}+p^{-4} over all primes converges. The prime two requires a different argument, because the derivative 2x_i is never invertible modulo two. But here we can use the elementary description of odd squares modulo powers of two. Every odd square is congruent to one modulo eight. Conversely, every integer congruent to one modulo eight has a square root modulo 2^r for every r\ge3 . We can prove this by induction: if an odd number x is a square root modulo 2^r , replacing it by x+2^{r-1} changes its square by 2^rx modulo 2^{r+1} , allowing us to choose the next binary digit. Now every residue class modulo eight is a sum of five squares, with at least one odd square. For residues one through four, use the corresponding number of copies of 1^2 and fill the remaining positions with zeros. For residues five through seven, use 2^2 and one, two, or three copies of 1^2 . For residue zero, use 2^2+1^2+1^2+1^2+1^2=8 . Fix such a representation modulo eight, with the first coordinate odd. At modulus 2^r , choose the other four coordinates arbitrarily subject to their fixed residues modulo eight. There are 2^{4(r-3)} such choices. For each choice, the number N-x_2^2-\cdots-x_5^2 is congruent to one modulo eight, so it has at least four square roots modulo 2^r . Thus M_{2^r}(N)\ge4\cdot2^{4(r-3)} , giving \sigma_2(N)\ge2^{-10} .

We have proved that every local factor is positive, with a uniform lower bound sufficient to make the infinite product positive. Therefore

\displaystyle \mathfrak S_5(N)\gg1\qquad\text{uniformly for positive integers }N.

The singular series is also uniformly bounded above by absolute convergence. We have therefore completely controlled the arithmetic factor in the main term.

The five square theorem

We can now assemble the argument. The circle integral was divided into major and minor arcs. The minor arcs contributed O_\varepsilon(P^{3+\varepsilon}Q^{-1/2}) . The major arcs were approximated by complete Gauss sums and real integrals, with error O(P^2Q^5) . Extending the truncated sums and integrals gave the singular series and singular integral, whose tails were estimated explicitly. With Q=P^{1/10} , we obtained an overall error O_\varepsilon(P^{3-1/20+\varepsilon}) . Since P=N^{1/2} , the singular integral equals (\pi^2/24)N^{3/2} , and the singular series is bounded above and below by positive constants. We have proved

\displaystyle R_5(N)=\frac{\pi^2}{24},\mathfrak S_5(N)N^{3/2}+O_\varepsilon\left(N^{3/2-1/40+\varepsilon}\right),\qquad \mathfrak S_5(N)\asymp1.

This is a genuine asymptotic theorem, with a power-saving error. In particular, every sufficiently large positive integer is a sum of five positive squares.

The proof illustrates the complete classical circle-method mechanism. The Fourier identity made the variables independent inside the integrand. Rational approximation identified the structured frequencies. Complete Gauss sums supplied square-root cancellation in the arithmetic factors. Weyl differencing controlled the minor arcs pointwise. The fourth moment controlled how frequently the exponential sum could be large. The singular integral described the real density of solutions, while the singular series described the local densities at every prime. None of these ingredients alone would have been sufficient. The theorem follows because the estimates fit together at exactly the right scale.

Four Squares

There is a very instructive limitation in the proof. Suppose we try to count representations as a sum of four squares instead of five. The expected main term is now of size P^{4-2}=P^2 . Our fourth-moment estimate gives \int_0^1|F(\alpha)|^4 d\alpha\ll_\varepsilon P^{2+\varepsilon} . But this is already of the same size as the expected main term. There is no extra factor of F(\alpha) on which we can apply the pointwise minor-arc saving. For five squares, we wrote |F|^5=|F|\cdot|F|^4 and bounded the first factor on the minor arcs. For four squares, that opportunity disappears. The straightforward combination of Weyl differencing and the fourth moment no longer proves that the minor arcs are negligible. This does not mean that the four-square problem is harder in every sense. Lagrange’s four-square theorem says that every nonnegative integer is a sum of four squares. Jacobi’s classical formula gives an exact expression for the number of representations by four integer squares, allowing zeros and signs:

\displaystyle r_4(N)=8\sum_{\substack{d\mid N\\ 4\nmid d}}d.

The point is that this arithmetic information is not accessible through the simple major/minor arc estimates we have just used. The method needs additional structure. This boundary was historically important. Kloosterman’s 1926 refinement of the circle method made it possible to treat certain quaternary quadratic forms by working with a more delicate Farey dissection and exploiting cancellation between arithmetic sums associated with neighboring rational approximations. In problems where the elementary minor-arc argument is too weak, one can sometimes gain cancellation by reorganizing the entire integral rather than estimating the minor arcs separately. This is a recurring theme in analytic number theory: a failure of an elementary estimate may indicate that the problem contains additional cancellation that the chosen decomposition does not reveal.

Equations over primes

The circle method becomes substantially more difficult when the variables are restricted to primes. Consider the classical problem of representing a sufficiently large odd integer N as a sum of three primes. Instead of the sum over squares, consider the von Mangoldt-weighted exponential sum

\displaystyle T(\alpha)=\sum_{n\le N}\Lambda(n)e(\alpha n) ,

where \Lambda(n)=\log p when n is a positive power of a prime p , and zero otherwise. The integral \int_0^1T(\alpha)^3e(-N\alpha) d\alpha counts representations of N as a sum of three prime powers, with the corresponding logarithmic weights. The expected main term has size N^2 . The real geometric factor is the area of the simplex t_1+t_2+t_3=N with positive coordinates, which gives N^2/2 . But the arithmetic factor is much more delicate because primes are not distributed uniformly across all residue classes. Near a rational frequency a/q with small denominator, one uses the distribution of primes in arithmetic progressions. The main contribution comes from residue classes coprime to q . If primes were perfectly equidistributed among these classes, each would contribute roughly N/\varphi(q) in von Mangoldt weight. Summing the phases e(ar/q) over reduced residues gives a Ramanujan sum, which equals \mu(q) when (a,q)=1 . This leads to the major-arc approximation

T(a/q+\beta)\approx\frac{\mu(q)}{\varphi(q)}\int_0^N e(\beta t) dt.

The structure resembles our quadratic approximation: a finite arithmetic factor multiplies a continuous oscillatory integral. But the proof is now much deeper. For squares, the periodic factor was handled by elementary Gauss sums and summation by parts. For primes, the approximation requires quantitative information about primes in arithmetic progressions, such as the Siegel–Walfisz theorem for sufficiently small moduli.

The minor arcs are harder still. We want cancellation in a sum of the form \sum_{n\le N}\Lambda(n)e(\alpha n) . Unlike a quadratic polynomial phase over an interval, the von Mangoldt function itself contains difficult arithmetic information. Vinogradov’s method attacks this by decomposing the von Mangoldt function into convolutions and then estimating the resulting sums. One typically encounters expressions of the form \sum_{m\le M}a_m\sum_{n\le N/m}e(\alpha mn) , called Type I sums, and bilinear expressions involving nontrivial coefficients in both variables, called Type II sums. For a Type I sum, if the coefficients a_m are bounded, the inner sum is geometric. We therefore obtain the concrete bound

\displaystyle \left|\sum_{m\le M}a_m\sum_{n\le N/m}e(\alpha mn)\right|\le\sum_{m\le M}|a_m|\min\left(\frac Nm,\frac1{2|m\alpha|}\right).

This is already recognizable from our quadratic Weyl estimate. Rational approximation to \alpha controls how often m\alpha is close to an integer. The difference is that the prime problem also produces Type II sums, where both variables carry arithmetic coefficients. There one needs Cauchy–Schwarz, bilinear cancellation, and additional tools.

The classical proof that every sufficiently large odd integer is a sum of three primes was achieved by Vinogradov in 1937. The circle-method identity is elementary, and its major-arc structure follows a familiar pattern, but the uniform minor-arc estimates for prime exponential sums are a substantial theorem of analytic number theory. We should not pretend that the elementary Weyl calculation for squares proves those estimates. The comparison is useful because it reveals two different sources of difficulty. For squares, the phase itself is nonlinear, but the summation variables are otherwise unrestricted. For primes, the phase is linear, but the arithmetic weight is highly irregular. The circle method can accommodate both situations, provided we have suitable exponential-sum estimates.

Delta Symbol

The identity

\displaystyle \int_0^1e(\alpha n) d\alpha=\mathbf1_{n=0}

is sometimes called the Fourier delta symbol. It is an exact detector of an integer equation. But there are several ways to represent or approximate this detector, and each may be useful for a different arithmetic problem. One elementary alternative is a finite rational-frequency identity. Let M be a positive integer larger than |n| . Then

\displaystyle \mathbf1_{n=0}=\frac1M\sum_{q\mid M}\ \sum_{\substack{a\bmod q\\(a,q)=1}}e(an/q).

This follows because the reduced fractions a/q with q\mid M enumerate all fractions r/M modulo one. The sum is therefore M^{-1}\sum_{r\bmod M}e(rn/M) , which equals one if M\mid n and zero otherwise. Since |n|<M , divisibility by M means n=0 . This is a completely finite version of the Fourier delta identity. It replaces integration over the circle by a sum over rational points. It is exact, but it is not necessarily useful analytically, because the denominators are restricted to divisors of a single integer M . The usefulness of a delta representation depends not merely on exactness, but on whether the resulting arithmetic sums can be estimated effectively.

A second possibility is to use a smooth weight. Choose a smooth compactly supported function w on the real line whose integer translates satisfy \sum_{k\in\mathbb Z}w(\alpha+k)=1 . Then

\displaystyle \int_{\mathbb R}w(\alpha)e(\alpha n) d\alpha=\mathbf1_{n=0} .

Indeed, divide the real line into translates of [0,1] . Since n is an integer, e(n(\alpha+k))=e(n\alpha) , and the sum of the translates of w is one. Thus the integral reduces to the ordinary Fourier delta identity. This smooth version becomes useful because smooth functions have rapidly decaying Fourier transforms. Repeated integration by parts can convert smoothness into strong estimates, which is often essential when applying Poisson summation.

Kloosterman’s refinement reorganizes the circle integral by a Farey dissection. Rather than separating a small collection of major arcs and treating the entire complement by a general bound, the method assigns frequencies to rational approximations and uses more delicate arithmetic cancellation between them. This was particularly important in quadratic-form problems where the elementary minor-arc argument was too weak. A later development, due essentially to Duke, Friedlander, and Iwaniec and developed further by Heath-Brown, uses a smooth delta-symbol expansion. In one formulation, for a parameter Q>1 , there is a smooth weight h and a normalizing factor c_Q=1+O_A(Q^{-A}) such that

\displaystyle \mathbf1_{n=0}=c_QQ^{-2}\sum_{q=1}^{\infty}\ \sum_{\substack{a\bmod q\\(a,q)=1}}e(an/q) h(q/Q,n/Q^2).

This is not an elementary consequence of the finite identity above; constructing the weight h with the necessary support and derivative estimates is a separate analytic theorem. Its importance lies in reorganizing the delta condition into sums over rational denominators accompanied by smooth weights. In applications to quadratic forms, one inserts this expansion into a sum over lattice points satisfying a polynomial equation. Poisson summation then transforms the lattice-point sum into complete exponential sums and oscillatory integrals. For suitable problems, this can avoid a separate minor-arc estimate altogether. The needed saving instead comes from cancellation in complete sums, often Kloosterman-type sums, and from the decay of smooth oscillatory integrals. Heath-Brown’s 1996 paper A New Form of the Circle Method, and Its Application to Quadratic Forms is a classical reference for this approach. The connection with the earlier Duke–Friedlander–Iwaniec delta method is explicit there.

The distinction between these versions is important. In the classical Hardy–Littlewood method, we usually isolate major arcs and prove that minor arcs contribute little. In Kloosterman-type refinements and smooth delta methods, we may analyze the contributions of rational approximations more uniformly, using cancellation between complete arithmetic sums rather than relying on a crude minor-arc supremum bound. The underlying Fourier principle is the same, but the analytic organization is different.

Conclusion:

We can now return to the basic question: why does the circle method work when there are enough variables? For a quadratic equation in s variables, the expected main term has size P^{s-2} . The two powers of P lost relative to the trivial P^s count come from the frequency width P^{-2} associated with a quadratic phase. But merely knowing the expected scale is not enough. To prove an asymptotic formula, we must show that the contribution away from the structured rational peaks is smaller than that scale. For sums of five squares, the fourth moment bound gave \int_0^1|F(\alpha)|^4 d\alpha\ll P^{2+\varepsilon} . This is already close to the expected size for four variables. The fifth variable supplied an extra factor of F(\alpha) , on which the minor-arc pointwise saving could be applied. That produced a power-saving estimate relative to the main term. This is why the number of variables is so important. Additional variables contribute additional powers of the exponential sum, and those powers can be divided between pointwise estimates and mean-value estimates. The success of the method depends on whether this combination yields a saving beyond the expected main-term scale.

For higher powers, the situation becomes more difficult. Weyl differencing reduces the degree of the phase, but repeated applications of Cauchy–Schwarz weaken the exponent in the resulting pointwise bound. Mean values of higher-degree exponential sums encode complicated systems of Diophantine equations. Hua’s lemma and Vinogradov’s methods were developed to obtain sufficiently strong estimates in this setting. By the end of the twentieth century, Vaughan and Wooley had substantially sharpened these techniques, though the fundamental major/minor arc architecture remained recognizable. There is also a useful distinction between arithmetic structure and analytic cancellation. The singular series is arithmetic: it records the solutions to congruences modulo prime powers. The singular integral is geometric: it records the density of real solutions. The minor-arc estimates are analytic: they explain why the remaining oscillatory contributions do not overwhelm the structured main term.

A good circle-method proof must control all three. It is possible to have positive local densities at every prime and a positive singular integral but still be unable to prove an asymptotic formula, because the minor arcs cannot be estimated sharply enough. Conversely, strong minor-arc cancellation does not create solutions if the equation has a congruence obstruction. In our five-square example, we proved the real density explicitly, showed that every local density was uniformly positive, and established a power-saving minor-arc estimate. That is why the proof was complete.

There is a useful way to think about the circle method beyond the particular formulas. The Fourier transform of an arithmetic sequence measures its correlation with additive characters e(\alpha n) . For sequences such as squares or primes, certain frequencies reveal arithmetic organization. Rational numbers with small denominators detect periodic congruence structure. Other frequencies behave more irregularly and are expected to exhibit cancellation. The circle method separates these two behaviors. It extracts the structured contribution explicitly and tries to show that the remaining oscillation is too small to affect the principal term. This is closely related to a broader principle in analytic number theory: if a sum is unusually large, there is often a reason. For polynomial exponential sums, the reason may be that the leading coefficient is close to a rational with small denominator. For sums over primes, it may be a structured major-arc frequency corresponding to a small modulus. The minor-arc estimates establish that outside such structured situations, substantial cancellation must occur. The difficult part is quantitative. It is not enough to say that a sum oscillates. We need to know how much cancellation occurs, uniformly over an entire range of frequencies, and whether the saving survives integration, summation over moduli, and multiplication by other exponential sums.

The later variants of the method do not discard these principles. They reorganize them. Kloosterman’s refinement and the smooth delta method use more elaborate rational decompositions and complete exponential sums. Vinogradov’s prime method uses arithmetic convolution identities and bilinear estimates. Hua’s work uses higher moments to extract cancellation not visible from pointwise bounds alone. The central lesson is that the circle method is not primarily an integral trick. The Fourier identity is easy. The mathematics lies in understanding the size and structure of the resulting exponential sums, choosing the decomposition at the correct scale, and proving that the errors are genuinely smaller than the arithmetic main term. Once these estimates are available, the answer takes a strikingly unified form: a real geometric density, multiplied by a product of local arithmetic densities, with an analytic error term measuring how successfully we have controlled the remaining oscillation.

Leave a Reply