1 Bombieri–Vinogradov
1.1 Statement
The main result is the Bombieri–Vinogradov theorem, which gives a strong estimate for primes in arithmetic progressions on average over moduli.
For each fixed \(A \geq 0\),
uniformly for all \(x \ge 2\) and \(1 \le Q \le \frac{\sqrt{x}}{(\log x)^{A+3}}\).
1.2 Assumptions
We take Siegel–Walfisz and the large sieve inequality as axioms.
Let \(A, C {\gt} 0\). If \(1 \leq q \leq (\log x)^C\) and \(a \in (\mathbb {Z}/q\mathbb {Z})^*\), then
Let \(Q \ge 1\), \(H \in \mathbb {Z}\), \(N \in \mathbb {Z}_{\ge 1}\) and \(c = (c_{H+1}, \dots , c_{H+N}) \in \mathbb {C}^N\) We then have
1.3 Definitions
for \(x \ge 1\), \(q \in \mathbb {N}\)
For \(f : \mathbb {N}\rightarrow \mathbb {R}\) and \(r : \mathbb {N}\) we use \(f_r\) to denote \(n \mapsto f(n) 1_{(n, r) = 1}\)
1.4 Reduction to \(\Delta _\Lambda \)
Bombieri–Vinogradov for \(\psi \) is reduced to a bound for \(\Delta _\Lambda \) via the following steps.
For each fixed \(A \ge 0\) we have
uniformly for \(x \ge 2\) and \(1 \le Q \le \sqrt{x}/(\log (x))^{A+3}\)
For all \(A \in \mathbb {N}\),
1.5 Vaughan’s Identity
The von Mangoldt function decomposes as \(\Lambda = \Lambda ^\sharp + \Lambda ^\flat + \Lambda _{\le U}\), where \(UV \le \sqrt{x}\) and \(U, V \ge e^{\sqrt{\log x}}\).
\(\Lambda ^\sharp = \mu _{\le V} * \log - (\Lambda _{\le U} * \mu _{\le V}) * 1\)
\(\Lambda ^\flat = (\Lambda _{{\gt}U} * 1) * \mu _{{\gt}V}\)
\(\Lambda _{\le U} = 1_{≤ U} \cdot \Lambda \)
Decompose \(\Lambda = \Lambda ^\sharp + \Lambda ^\flat + \Lambda _{\le U}\)
1.6 Small Primes (\(\Lambda _{\le U}\))
The contribution of the small part \(\Lambda _{\le U}\) is bounded trivially.
For \(y, U {\gt} 0\), \(q \in \mathbb {N}\) and \(a \in \mathbb {Z}/q\mathbb {Z}\),
For each fixed \(A \ge 0\), \(x \ge 2\) and \(1 \le Q \le \sqrt{x}/(\log x)^{A+3}\),
1.7 Type I Sums (\(\Lambda ^\sharp \))
The sharp part \(\Lambda ^\sharp \) is a type I sum, bounded via Abel summation.
If \(f\) is an arithmetic function supported on \([1, y]\) then
For \(x \ge 1\), \(q \in \mathbb {N}\) and \(a \in (\mathbb {Z}/q\mathbb {Z})^*\),
Carefully consider length \(q\) intervals. Alternatively, write
and note each inner difference is bounded by \(1\) in absolute value.
If \(g\) is continuously differentiable on \([1, x]\) then
By Abel summation.
If \(g\) is continuously differentiable and monotone on \([1, x]\) with \(g(1) = 0\), then for all \(t \ge 1\) and \(a \in (\mathbb {Z}/q\mathbb {Z})^*\),
Let \(v \ge 0\) and let \(f\) be an arithmetic function supported on \([1, x]\). For \(x \ge 2\), \(q \in \mathbb {N}\) and \(a \in (\mathbb {Z}/q\mathbb {Z})^*\),
Straightforward application of the previous lemmas.
For \(U, V \ge 1\), \(x \ge 2\), \(q \in \mathbb {N}\), \(r \le x\) and \(a \in (\mathbb {Z}/q\mathbb {Z})^*\),
Apply 21 twice: once with \(f = \mu _{\le V,\, r}\), \(v = 1\), and once with \(f = \Lambda _{\le U} * \mu _{\le V,\, r}\), \(v = 0\), \(y = UV\).
For each fixed \(A \ge 0\), \(x \ge 2\) and \(1 \le Q \le \sqrt{x}/(\log x)^{A+3}\),
1.8 Type II Sums (\(\Lambda ^\flat \))
1.8.1 Definitions
1.8.2 Reduction to character sums
This is a standard result. Let \(f\) be a function from Dirichlet characters. Then
Let \(f\) be an arithmetic function. For \(r \le x\), \(q {\gt} 1\) and \((a, q) = 1\),
Fix \(P \in \mathbb {N}\) with \(q \mid P\). Define \(F_P\) and \(G_P\) on divisors of \(P\) by
Since every non-principal character mod \(q\) factors through a unique primitive character, \(F_P(q) = \sum _{d \mid q} G_P(d)\). By Möbius inversion,
Set \(P = q\) to conclude.
Push the absolute values inside, then
Sum the error from 29 over \(q \le Q\) using \(\sum _{n \le x} 1/\varphi (n) \ll \log x\), then regroup the main sum by \(r = q/d\).
1.8.3 Large sieve estimates
Let \(f\) and \(g\) be arithmetic functions supported on \([1, M]\) and \([1, N]\) respectively. For \(x, Q \ge 1\),
Uses Cauchy–Schwarz and the large sieve inequality (3). The proof in the book uses the classical version of Perron’s integral formula as \(1_{n \le x} = \int _{-T}^{T}\frac{(x/n)^{\alpha +it}}{\alpha +it} dt/(2\pi ) + O(...)\) But we have a different version in PNT+. I haven’t worked out how this changes the proof yet.
where \(f_j(k) = (\Lambda _{{\gt}U} * 1)(k)\, 1_{2^{j-1} {\lt} k \le 2^j}\) and \(g_j(\ell ) = \mu (\ell )\, 1_{V {\lt} \ell \le x/2^{j-1}}\).
For \(x, Q \ge 2\), \(U, V \in [1, x]\) and \(r \in \mathbb {N}\),
Divide the sum defining \(T_r\) into dyadic intervals in \(d\) and apply 33.
For each fixed \(A \ge 0\), \(x \ge 2\) and \(1 \le Q \le \sqrt{x}/(\log x)^{A+3}\),