Let
$\unicode[STIX]{x1D706}$
denote the Liouville function. The Chowla conjecture, in the two-point correlation case, asserts that
$$\begin{eqnarray}\mathop{\sum }_{n\leqslant x}\unicode[STIX]{x1D706}(a_{1}n+b_{1})\unicode[STIX]{x1D706}(a_{2}n+b_{2})=o(x)\end{eqnarray}$$
as
$x\rightarrow \infty$
, for any fixed natural numbers
$a_{1},a_{2}$
and nonnegative integer
$b_{1},b_{2}$
with
$a_{1}b_{2}-a_{2}b_{1}\neq 0$
. In this paper we establish the logarithmically averaged version
$$\begin{eqnarray}\mathop{\sum }_{x/\unicode[STIX]{x1D714}(x)
of the Chowla conjecture as
$x\rightarrow \infty$
, where
$1\leqslant \unicode[STIX]{x1D714}(x)\leqslant x$
is an arbitrary function of
$x$
that goes to infinity as
$x\rightarrow \infty$
, thus breaking the ‘parity barrier’ for this problem. Our main tools are the multiplicativity of the Liouville function at small primes, a recent result of Matomäki, Radziwiłł, and the author on the averages of modulated multiplicative functions in short intervals, concentration of measure inequalities, the Hardy–Littlewood circle method combined with a restriction theorem for the primes, and a novel ‘entropy decrement argument’. Most of these ingredients are also available (in principle, at least) for the higher order correlations, with the main missing ingredient being the need to control short sums of multiplicative functions modulated by local nilsequences. Our arguments also extend to more general bounded multiplicative functions than the Liouville function
$\unicode[STIX]{x1D706}$
, leading to a logarithmically averaged version of the Elliott conjecture in the two-point case. In a subsequent paper we will use this version of the Elliott conjecture to affirmatively settle the Erdős discrepancy problem.