Skip to main content

Chapter 22 Bounded gaps between primes: proofs

In this chapter, we fill in the missing steps in Chapter 21 to complete Maynard's proof of bounded gaps between primes [16].

Section 22.1 Review of the setup

We begin by collecting the accumulated notation from Chapter 21.

Definition 22.1.

Fix a \(k\)-tuple \(\calH\text{.}\) Let \(F\colon \RR^k \to \RR\) be a smooth function supported on the standard simplex
\begin{equation*} \Delta_k := \{(x_1,\dots,x_k) \in \RR^k: x_i \geq 0, \sum_i x_i \leq 1\}. \end{equation*}
Define the function
\begin{equation} \rho(d_1,\dots,d_k) := \mu(d_1)\cdots \mu(d_k) \sum_{r_i: d_i|r_i} \mu\left(\prod_i r_i \right)^2 \prod_i \frac{d_i}{\varphi(r_i)} F\left( \frac{\log r_1}{\log R}, \dots, \frac{\log r_k}{\log R} \right)\tag{22.1.1} \end{equation}
where \(R = x^{1/4-\epsilon}\text{;}\) then define the weights
\begin{equation} a(n) := \left( \sum_{d_i|n+h_i} \rho(d_1,\dots,d_k) \right)^2 \qquad (n \equiv v_0 \pmod{W})\tag{22.1.2} \end{equation}
where \(W := \prod_{p \leq D_0} p\) for \(D_0 := \log \log \log x\text{,}\) and \(v_0\) is an integer such that
\begin{equation*} \gcd(\prod_i (n+h_i): n \equiv v_0 \pmod{W}) = 1 \end{equation*}
(whose existence certifies that \(\frakS(\calH) > 0\)).
Define
\begin{align} S_1 \amp:= \sum_{x \lt n \leq 2x} a(n),\tag{22.1.3}\\ S_2^{(j)} \amp := \sum_{x \lt n \leq 2x} \chi_P(x+h_j) a(n) \qquad (j=1,\dots,k),\tag{22.1.4}\\ S_2 \amp:= \sum_{j=1}^k S_2^{(j)}.\tag{22.1.5}\\ I_k(F) \amp= \int_0^1 \cdots \int_0^1 F(t_1,\dots,t_k)^2 dt_1 \cdots dt_k\tag{22.1.6}\\ J_k^{(j)}(F) \amp= \int_0^1 \cdots \int_0^1 \left( \int_0^1 F(t_1,\dots,t_k)dt_j \right)^2 dt_1 \cdots dt_{j-1} \, dt_{j_1} \cdots dt_k \qquad (j=1,\dots,k).\tag{22.1.7} \end{align}
We prove Lemma 21.7.
See [16], Proposition 4.1.

Section 22.2 Computing the main terms

Definition 22.3.

Define
\begin{equation*} S_2^{(j)\prime} \amp := \sum_{x \lt n \leq 2x} \chi_P(x+h_j) a(n) \qquad (j=1,\dots,k), \end{equation*}

Section 22.3 Estimating the error term

Section 22.4 Application of Bombieri–Vinogradov

Section 22.5 Optimizing the objective function

We conclude by proving Lemma 21.10 and Lemma 21.9.
We follow [16], Proposition 4.3(3). It is sufficient to find a Riemann-integrable function \(F\) and then approximate it by a suitable smooth function. We make a further ansatz of the form
\begin{equation*} F(\rho_1,\dots,\rho_k) = \chi_{\Delta_k} \prod_{i=1}^k g(k\rho_i) \end{equation*}
where \(g: [0, \infty) \to \RR\) is a smooth function supported on some interval \([0,T]\) with \(g(0)=1\) (for normalization purposes). To simplify notation, write \(I_k, J_k\) as shorthand for \(I_k(F), J_k^{(1)}(F)\text{.}\)
We first give an upper bound on \(I_k\text{.}\) Set \(\gamma = \int_{u=0}^\infty g(u)^2\,du > 0\text{;}\) then
\begin{equation*} I_k = \idotsint_{\Delta_k} F(\rho_1,\dots,\rho_k)^2 d\rho_1\cdots d\rho_k \leq \left( \int_0^\infty g(k\rho)^2 \,d\rho \right) = k^{-k} \gamma^k. \end{equation*}
We next work on a lower bound for \(J_k\text{.}\) We first exploit the nonnegativity of squares to restrict the domain of integration for convenience:
\begin{equation*} J_k \geq \idotsint_{\substack{\rho_2,\dots,\rho_k \geq 0}{\sum_{i=2}^k \rho_i \leq 1-T/k}} \left( \int_0^{T/k} \left( \prod_{i=1}^k g(k\rho_i)\right) \,d\rho_1 \right)^2 d\rho_2\cdots d\rho_k. \end{equation*}
We then do the opposite in order to separate variables: we rewrite this as \(J_k \geq J_k' - E_k\) where
\begin{align*} J_k' \amp:= \idotsint_{\rho_2,\dots,\rho_k \geq 0} \left( \int_0^{T/k} \left( \prod_{i=1}^k g(k\rho_i)\right) \,d\rho_1 \right)^2 d\rho_2\cdots d\rho_k\\ \amp = \left( \int_0^\infty g(k\rho_1)d\rho_1 \right)^2 \left(\int_0^\infty g(kt)^2\,dt \right)^{k-1} = k^{-k-1} \gamma^{k-1} \left( \int_0^\infty g(u)\,du\right)^2,\\ E_k \amp:= \idotsint_{\substack{\rho_2,\dots,\rho_k \geq 0}{\sum_{i=2}^k \rho_i > 1-T/k}} \left( \int_0^{T/k} \left( \prod_{i=1}^k g(k\rho_i)\right) \,d\rho_1 \right)^2 d\rho_2\cdots d\rho_k\\ \amp= k^{-k-1} \left( \int_0^\infty g(u)\,du\right)^2 \idotsint_{\substack{u_2,\dots,u_k \geq 0}{\sum_{i=2}^k u_i > k-T}} \left( \prod_{i=2}^k g(u_i)^2 \right) du_2\cdots du_k. \end{align*}
We first show that the error term \(E_k\) is small by comparison with a second moment, assuming that
\begin{equation*} \mu := \frac{\int_0^\infty ug(u)^2\,du}{\int_0^\infty g(u)^2\,du} \lt 1 - \frac{T}{k}. \end{equation*}
Define also
\begin{equation*} \eta := \frac{k-T}{k-1} - \mu > 0. \end{equation*}
We may then give an upper bound on \(E_k\) by multiplying the integrand by \(C := \mu^{-2} \left(\sum_{i=2}^k \frac{u_i}{k-1} - \mu \right)^2\) (this factor being at least 1) and dropping the restriction \(\sum_{i=1}^k u_i > k-T\) from the domain of integration. After expanding the square in \(C\text{,}\) we may further replace \(u_j^2 g(u_j)^2\) with \(T u_j g(u_j)^2\) thanks to the support restriction on \(g\text{.}\) We may then compute that
\begin{align*} E_k \amp\leq \eta^{-2} k^{-k-1} \left( \int_0^\infty g(u)\,du \right)^2 \left( \frac{\mu T \gamma^{k-1}}{k-1} - \frac{\mu^2 \gamma^{k-1}}{k-1} \right)\\ \amp\leq \frac{\eta^{-2} \mu T k^{-k-1} \gamma^{k-1}}{k-1} \left( \int_0^\infty g(u)\,du \right)^2. \end{align*}
Putting everything together and observing that \((k-1)\eta^2 \geq k(1-T/k-\mu)^2\) and \(\mu \leq 1\text{,}\) we get
\begin{equation*} \frac{kJ_k}{I_k} \geq \frac{\left( \int_0^\infty g(u)\,du \right)^2}{\int_0^\infty g(u)^2\,du} \left( 1 - \frac{T}{k(1-T/k - \mu)^2} \right). \end{equation*}
By writing down an Euler–Lagrange equation, we may identify a good candidate for \(g(u)\text{:}\)
\begin{equation*} g(u) = \frac{1}{1+Au} \end{equation*}
for some suitable \(A>0\text{.}\) For this choice,
\begin{align*} \int_0^T g(u)\,du \amp= \frac{\log(1+Au)}{A},\\ \int_0^T g(u)^2\,du \amp= \frac{1}{A} \left( 1 - \frac{1}{1+AT} \right),\\ \int_0^T ug(u)^2\,du \amp= \frac{1}{A^2} \left( \log(1+AT) - 1 + \frac{1}{1+AT} \right). \end{align*}
We now fix \(A := \log k - 2 \log \log k\) and choose \(T\) so that \(1+AT = e^A\text{.}\) We then compute that
\begin{equation*} \frac{kJ_k}{I_k} \geq (\log k - 2 \log \log k) \left( 1 - \frac{\log k}{(\log k)^2 + O(1)} \right) \end{equation*}
which proves the claim.
To extract explicit bounds, we use a slightly different strategy.
We take
\begin{equation*} F(\rho_1,\dots,\rho_k) = \chi_{\Delta_k} P(\rho_1,\dots,\rho_k) \end{equation*}
for some symmetric polynomial \(P\text{,}\) which we write in terms of the power sum polynomials \(P_j(\rho_1,\dots,\rho_k) = \sum_{i=1}^k \rho_i^j\text{.}\) We may then calculate the integrals \(I_k(F)\) and \(J_k^{(1)}(F)\) for \(F\) equal to a monomial in the power sums (see [16], Lemmas 8.1 and 8.2), and then make a numerical calculation on the subspace generated by a suitable collection of monomials to extract the claimed bound. For further details, see [16], section 8 (proof of Proposition 4.3(2)).