9
$\begingroup$

Let $A = (a_0,a_1,\ldots,a_k)$ be a sequence of strictly positive numbers, and let $A^{\ast k}$ denote the $k$-fold repeated convolution (defined by $A^{\ast 1} = A$ and $A^{\ast k+1} = A^{\ast k} \ast A$).

Is it true that for any such sequence $A$, there exists $n$ such that $A^{\ast n}$ is log-concave?

As an example, any sequence of the form $A = (1,x,1)$ is log concave if $x\geq 1$, and $A\ast A$ is log concave if $x\geq \sqrt{2/3}$. In general, I have checked numerically that $A^{\ast k}$ is log concave if $x\geq \sqrt{2/(k+1)}$. I have also checked this statement numerically for many other sequences of longer length, but I have no idea how to go about proving anything.

$\endgroup$
3
  • $\begingroup$ To make sure, how do you define the convolution of finite sequences? By extending them (by zeroes) to functions on $\mathbb Z$? $\endgroup$ Mar 6, 2019 at 18:03
  • $\begingroup$ If so, your conjecture seems somewhat plausible, in view of the central limit theorem of probability theory. Perhaps, the Fourier transform can be used here. $\endgroup$ Mar 6, 2019 at 18:06
  • $\begingroup$ That's right, one can imagine extending the sequences by zeros to functions on $\mathbb{Z}$. I agree that the central limit theorem supports this conjecture, although I don't know of a way to recover log-concavity from the Fourier transform. $\endgroup$
    – felipeh
    Mar 6, 2019 at 18:19

1 Answer 1

5
$\begingroup$

Odlyzko and Richmond proved your conjecture:

http://www.dtc.umn.edu/~odlyzko/doc/arch/unimodal.convolut.pdf

Odlyzko, A. M.(1-BELL); Richmond, L. B.(3-WTRL) On the unimodality of high convolutions of discrete distributions. Ann. Probab. 13 (1985), no. 1, 299–306.

$\endgroup$
1

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.