Is the sequence of partition numbers log-concave?

The first two terms of the Hardy-Ramanujan formula give $$p(n) = \frac{1}{4 \sqrt{3} n} \exp(\pi \sqrt{2n/3}) + O \left(\exp(\pi \sqrt{n/6} ) \right)$$ so $$\log p(n) = \pi \sqrt{2/3} \sqrt{n} - \log n - \log (4 \sqrt{3}) + O(\exp(-\pi \sqrt{n/6} ) ).$$ So $$\log p(n+2) - 2 \log p(n+1) + \log p(n) = $$ $$ \pi \sqrt{2/3} \left( \sqrt{n+2} - 2\sqrt{n+1} + \sqrt{n} \right) - \left( \log(n+2) - 2 \log(n+1) + \log n \right) + O(\exp(-\pi \sqrt{n/6} ) )$$ $$= \left[ \left( \frac{- \pi \sqrt{2/3}}{4} \right) n^{-3/2} + O(n^{-5/2}) \right] + O(n^{-2}) + O(\exp(-\pi \sqrt{n/6} ) ).$$ So this quantity is negative for $n$ sufficiently large.

The larger determinants seem harder; there is probably a smarter way to do this.


With the help of Mathematica, I set $q(n) = a \exp(c \sqrt{n})/n$ and computed that $$\det \begin{pmatrix} q(n) & q(n+1) & q(n+2) \\ q(n-1) & q(n) & q(n+1) \\ q(n-2) & q(n-1) & q(n) \end{pmatrix} = q(n)^3 \left( \frac{c^3}{32 n^{9/2}} + O(n^{-10/2}) \right).$$ The error in approximating $p(n)$ by $q(n)$ (for $a = 1/(4 \sqrt{3})$ and $c = \pi \sqrt{2/3}$) will be exponentially smaller than $n^{-9/2}$, so the $3 \times 3$ determinant is positive for $n$ large.

The $4 \times 4$ determinant vanishes to order at least $n^{-12/2}$, and I gave up waiting for the computation to finish when I asked for more terms.


The statement referenced by Igor Rivin http://www.math.clemson.edu/~janoski/ResearchStatement.pdf uses the phrase

Computationally looking at p(n) we see that for n ≥ 26 the partition function is log-concave [2].

I had seen this reference before probably about the same time this research statement was first released, and I am skeptical for two reasons.

  1. The phrasing "Computationally..." would seem to indicate some type of calculation. This cannot involve a computer since it would have to hold for all n larger than 26, and I am not aware of any simplification that allows one to only consider a finite number of cases. It would have been helpful to at least expound on the type of computations involved.

  2. I checked for the promised reference, and indeed I found it on the CV of the author, http://www.math.clemson.edu/~janoski/VitaTex.pdf, but it refers to the quote below. I did a quick google search and I could find no reference or anything pointing to a publication.

    Brian Bowers, Neil Calkin, Kerry Gannon, Janine E. Janoski, Katie Joes, Anna Kirkpatrick, The Log Concavity of the Partition Function, (in preparation)

  3. Asymptotics will not provide the answer here, since n sufficiently large doesn't hold up unless you can provide a concrete n and test everything less than it, and I don't believe the Hardy-Ramanujan asymptotic expansion yields any guaranteed error estimates.

  4. It may be possible to use DH Lehmer's estimates to obtain a proof. In two papers (1937 and 1939) he investigated the coefficients of both the Hardy-Ramanujan asymptotic expansion and the Hardy-Ramanujan-Rademacher expansion. He provided guaranteed error bounds on the remainder terms in the asymptotic expansions so that, for example, his Theorem 13 says that for n>600, only $2/3 \sqrt n$ terms of the Hardy-Ramanujan asymptotic series are needed to estimate p(n) to the nearest integer.

At present, I don't believe the matter is completely settled, despite the overwhelming computational evidence.

UPDATE 11-1-13:

Igor Pak and I have just uploaded a preprint to the ArXiv: http://arxiv.org/abs/1310.7982 . In it we prove the log-concavity of the partition numbers for all $n>25$, and Section 6.3 addresses Janoski's thesis.

UPDATE 11-23-15:

Igor and I were recently informed of work by Jean-Louis Nicolas which also contains a proof of the log-concavity of the partition numbers:

Sur les entiers N pour lesquels il y a beaucoup de groupes abéliens d’ordre N, Annales de l’institut Fourier, tome 28, no 4 (1978), p. 1-16.

http://www.numdam.org/item?id=AIF_1978__28_4_1_0


This paper from a J. Janoski at Clemson seems to indicate that despite the fact that partitions have been studied half-to-death, the log concavity is still somewhat open (AND the asymptotic way of doing it is the only way known). Note that a related unimodality theorem of Szekeres (for partitions into $k$ parts) is only proved using asymptotics, and not a bijective correspondence, so the "book proofs" of both facts still elude us.