The autoconvolution supremum problem asks how small the essential supremum of the convolution of a nonnegative function with itself can be when its integral
and the length of a containing interval are fixed (Martin
and O'Bryant 2009, Matolcsi and Vinuesa 2010). Let consist of the nonnegative integrable functions supported on
with integral 1. The
problem is to determine
|
(1)
|
where
denotes the convolution of
with itself and
is the essential
supremum norm. Equivalently,
is the largest constant for which
|
(2)
|
for every nonzero nonnegative integrable function supported on this interval.
With support
instead, the corresponding optimal constant
is
by a change of scale.
In probabilistic terms, is the probability
density function of
for
,
when
and
are independent and identically
distributed with probability density
function
(Matolcsi and Vinuesa 2010).
An explicit step function gives an upper bound for . If
has nonnegative heights
,
,
...,
, not all zero, on
equal subintervals of
,
then
|
(3)
|
where
outside
. The maximum
can be evaluated at the breakpoints because the convolution
is a piecewise linear function (Martin
and O'Bryant 2009, Matolcsi and Vinuesa 2010).
Yuksekgonul et al. (2026) used AI-assisted search to find a 30,000-piece step function giving
|
(4)
|
The autoconvolution inequality concerns a different ratio of norms, without this fixed-support restriction.
Ivanisvili and Xie (2026) refined the discretization estimate of Cloninger and Steinerberger (2017) to prove
|
(5)
|
They report that Grok suggested the refinement. Atris Labs (2026) constructed a 65,536-piece step function with rational heights giving
|
(6)
|
An improved lower bound for also gives an improved upper bound
for the size of a generalized Sidon set.
The exact value remains unknown.