Let $(u_1, u_2, u_3, u_4)$ and $(v_1, v_2, v_3, v_4)$ be vectors in $\mathbb R_+^4$. Is the following inequality true?

\begin{align*} \left(\sum_{{[4] \choose 3}} \sqrt{u_i u_j u_k}\right)^{2/3} + \left(\sum_{{[4] \choose 3}} \sqrt{v_i v_j v_k}\right)^{2/3} \leq \left(\sum_{{[4] \choose 3}} \sqrt{(v_i+u_i) (v_j+u_j) (v_k+u_k)}\right)^{2/3} \end{align*} Here $\sum_{{[4] \choose 3}}$ refers to $\sum_{1\leq i < j < k \leq 4}$.

I ran a million Matlab simulations for random vectors and it did not yield any counterexample.

Note: Previously asked on MSE (https://math.stackexchange.com/questions/2647608/a-minkowski-like-inequality-for-symmetric-sums)


2 Answers 2


Rewrite the inequality in question as \begin{equation*} f(u+v)\le f(u)+f(v) \end{equation*} for $u,v$ in $\mathbb R_+^4$, where \begin{equation*} f(u):=-\left(\left(\frac{1}{\sqrt{u_1}}+\frac{1}{\sqrt{u_2}}+\frac{1}{\sqrt{u_3}} +\frac{1}{\sqrt{u_4}}\right) \sqrt{u_1 u_2 u_3 u_4}\right)^{2/3}. \end{equation*} Note that the function $f$ is positive homogeneous: $f(tu)=tf(u)$ for $t\ge0$. So, $f(u+v)=2f(\frac{u+v}2)$.

It remains to notice that $f$ is convex. Indeed, the determinant of the Hessian matrix \begin{equation} M:=\Big(\frac{\partial^2 f}{\partial u_i\partial u_j}\Big)_{i,j=1}^4 \end{equation} is $0$, and the principal minors of $M$ are manifestly positive, after some algebraic simplifications. (That the determinant of $M$ is $0$ can be shown either by direct calculations or by recalling that $f$ is positive homogeneous and hence $\frac{d^2}{dt^2}\,f(tu)=0$ for $t>0$.)

Dealing with the determinants of the matrices $\Big(\frac{\partial^2 f}{\partial u_i\partial u_j}\Big)_{i,j=1}^k$ for $k=1,2,3$, in view of the positive homogeneity, we may assume without loss of generality that $u_4=1$.

Details of the calculations can be seen in the the Mathematica notebook and/or its pdf image .

  • $\begingroup$ I admire the tour de force solution! I will wait a couple of days before accepting, however, since I would very much like to see a proof that can also generalize to an ${n \choose k}$ setting (the current one being ${4 \choose 3}$.) $\endgroup$
    – VSJ
    Feb 14, 2018 at 4:53
  • 1
    $\begingroup$ I think, if you want an answer to a more general or otherwise significantly amended question, it is better to ask it separately. When trying to decide whether to invest time and effort into answering a question, people don't necessarily anticipate that the question may change down the road. $\endgroup$ Feb 14, 2018 at 16:10
  • $\begingroup$ I agree; I have no intention of changing this question. $\endgroup$
    – VSJ
    Feb 14, 2018 at 17:36

The said claim follows from the following general result on elementary symmetric polynomials, denoted $e_k$ below. $\newcommand{\vx}{\mathbf{x}}\newcommand{\vy}{\mathbf{y}}$

Theorem A (S. 2018). $\,$ Let $p\in (0,1)$ and $x \in \mathbb{R}_+^n$. The map \begin{equation*} \phi_{k,n}(x) := x \mapsto \left[\frac{e_k(x_1^p,\ldots,x_n^p)}{e_{k-1}(x_1^p,\ldots,x_n^p)}\right]^{1/p}, \end{equation*} is concave. Recently, I typed up a proof to this inequality (and a few others). Please see this preprint for a proof.

As a corollary, we obtain a concavity result that implies the OP's conjectured inequality as a special case (using positive homogeneity).

Corollary. Let $p\in (0,1)$ and $\vx \in \mathbb{R}_+^n$. Then, $[e_k(\vx^p)]^{1/pk}$ is concave (we write $\vx^p \equiv (x_1^p,\ldots,x_n^p)$. \begin{align*} [e_k((\vx + \vy)^p)]^{1/pk} &= \left[\frac{e_k((\vx+\vy)^p)}{e_{k-1}((\vx+\vy)^p)}\cdot \frac{e_{k-1}((\vx+\vy)^p)}{e_{k-2}((\vx+\vy)^p)}\cdots \frac{e_1((\vx+\vy)^p)}{e_0((\vx+\vy)^p)} \right]^{1/pk}\\ &= \left[\phi_{k,n}(\vx+\vy)\phi_{k-1,n}(\vx+\vy)\cdots\phi_{1,n}(\vx+\vy)\right]^{1/k}\\ &\ge \left[\left(\phi_{k,n}(\vx)+\phi_{k,n}(\vy)\right)\cdots\left(\phi_{1,n}(\vx)+\phi_{1,n}(\vy)\right)\right]^{1/k}\\ &\ge \prod_{j=1}^k[\phi_{j,n}(\vx)]^{1/k} + \prod_{j=1}^k[\phi_{j,n}(\vy)]^{1/k}\\ &= [e_k(\vx^p)]^{1/pk} + [e_k(\vy^p)]^{1/pk}, \end{align*} where the first inequality follows from Theorem A and the second is just Minkowski.

  • 1
    $\begingroup$ A collaborator and I were able to prove the ${n \choose k}$ case of the original question by directly computing the Hessian by hand and showing it is positive semidefinite. Your generalizations and extensions are superb! $\endgroup$
    – VSJ
    Mar 31, 2018 at 19:57
  • 1
    $\begingroup$ Thanks! Impressive that you could work out a direct proof by hand -- I try to avoid computing Hessians whenever I can! $\endgroup$
    – Suvrit
    Mar 31, 2018 at 20:09

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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