let $a$ is give postive real number,and $a_{1},a_{2},\cdots,a_{n}$ be postive real numbers,and such $$a^2_{1}+a^2_{2}+\cdots+a^2_{n}=1,a_{1}+a_{2}+\cdots+a_{n}=a$$

show that: there exist $\mu_{i}\in\{-1,1\},i=1,2,\cdots,n$ such $$|\mu_{1}a_{1}+\mu_{2}a_{2}+\cdots+\mu_{n}a_{n}|\le\dfrac{1}{a}\tag{1}$$

I think use this identity: $$\sum_{\mu_{i}\in\{-1,1\}}|\mu_{1}a_{1}+\mu_{2}a_{2}+\cdots+\mu_{n}a_{n}|^2=2^n\sum_{k=1}^{n}a^2_{i}=2^n$$ By the Pigeonhole Principle we obatain: $$|\mu_{1}a_{1}+\mu_{2}a_{2}+\cdots+\mu_{n}a_{n}|^2\le\dfrac{2^n}{2^n}= 1$$ or $$|\mu_{1}a_{1}+\mu_{2}a_{2}+\cdots+\mu_{n}a_{n}|\le 1$$ or but I can't prove the stronger inequality $(1)$,thanks


Let us show the following more general proposition.

Let $a_{1},a_{2},\cdots,a_{n}$ be positive real numbers. Then there exist $\mu_{i}\in\{-1,1\},i=1,2,\cdots,n$ such $$0\le(\mu_{1}a_{1}+\mu_{2}a_{2}+\cdots+\mu_{n}a_{n})(a_{1}+a_{2}+\cdots+a_{n})\le a_{1}^2+a_{2}^2+\cdots+a_{n}^2\tag{1}$$

Proof by induction on $n$.

The proposition is trivially true for the base case, $n=1$.

Assume it is true for $n$. Consider the case of $n+1$. WLOG, suppose $a_1\ge a_2\ge\cdots\ge a_n\ge a_{n+1}\gt0$. By assumption, there exist $\mu_{i}\in\{-1,1\},i=1,2,\cdots,n$ such that inequality $(1)$ holds.

There are two cases.

  • $\mu_{1}a_{1}+\mu_{2}a_{2}+\cdots+\mu_{n}a_{n} \ge a_{n+1}$. Let $\mu_{n+1}=-1$. Then, $$0\le\mu_{1}a_{1}+\mu_{2}a_{2}+\cdots+\mu_{n}a_{n}+\mu_{n+1}a_{n+1}.$$ Moreover, $$\begin{aligned} &\qquad(\mu_{1}a_{1}+\mu_{2}a_{2}+\cdots+\mu_{n}a_{n}+\mu_{n+1}a_{n+1}) (a_{1}+a_{2}+\cdots+a_{n}+a_{n+1}) \\ &=(\mu_{1}a_{1}+\mu_{2}a_{2}+\cdots+\mu_{n}a_{n}- a_{n+1})(a_{1}+a_{2}+\cdots+a_{n}+a_{n+1}) \\ &=(\mu_{1}a_{1}+\mu_{2}a_{2}+\cdots+\mu_{n}a_{n})(a_{1}+a_{2}+\cdots+a_{n})\\ &\qquad-a_{n+1}(a_{1}(1-\mu_1)+a_{2}(1-\mu_2)+\cdots+a_{n}(1-\mu_{n}))-a_{n+1}^2\\ &\lt (\mu_{1}a_{1}+\mu_{2}a_{2}+\cdots+\mu_{n}a_{n})(a_{1}+a_{2}+\cdots+a_{n})\\ &\le a_{1}^2+a_{2}^2+\cdots+a_{n}^2\\ &\lt a_{1}^2+a_{2}^2+\cdots+a_{n}^2+a_{n+1}^2.\\ \end{aligned}$$

  • $\mu_{1}a_{1}+\mu_{2}a_{2}+\cdots+\mu_{n}a_{n} \lt a_{n+1}$. Let $\nu_i=-\mu_i$ for $1\le i\le n$ and $\nu_{n+1}=1$. Then, $$0\le \nu_{1}a_{1}+\nu_{2}a_{2}+\cdots+\nu_{n}a_{n}+\nu_{n+1}a_{n+1}.$$ Moreover, $$\begin{aligned} &\qquad(\nu_{1}a_{1}+\nu_{2}a_{2}+\cdots+\nu_{n}a_{n}+\nu_{n+1}a_{n+1}) (a_{1}+a_{2}+\cdots+a_{n}+a_{n+1}) \\ &=(a_{n+1}-(\mu_{1}a_{1}+\mu_{2}a_{2}+\cdots+\mu_{n}a_{n}))(a_{1}+a_{2}+\cdots+a_{n}+a_{n+1}) \\ &\le a_{n+1}(a_{1}+a_{2}+\cdots+a_{n}+a_{n+1}) \\ &\le a_{1}^2+a_{2}^2+\cdots+a_{n}^2+a_{n+1}^2.\\ \end{aligned}$$


There are two general ways one might try to tackle something like this. One is the probabilistic method which you noted does not work. Another is an extremal approach as follows.

Consider the choice of signs that minimize $ |\sum \mu_i a_i|^2$ and denote $y = \sum \mu_i a_i$ for this choice. By flipping the signs of $a_i$ if necessary, we can assume that $y = \sum a_i$. Then by flipping the sign of $a_1$, we get a strictly larger square so we have $y^2 \le (y-2a_1)^2$ which implies that $a_1y \le a_1^2$. Of course we can replace $a_1$ with any other $a_j$'s and summing the resulting inequalities give us $a y \le 1$ as desired.

This was inspired by the answer to this problem: https://mathoverflow.net/questions/352720/reference-to-a-conjecture-on-unit-vectors-in-euclidean-space/352817#352817

  • 2
    $\begingroup$ I am not sure that you can really assume that all signs are $+1$. This requires you flipping the signs of the $a_i$, so that you don't know $a_i\geq 0$ anymore, which you seem to be using in the last step when you sum over $j$. I might be missing something obvious though, just woke up :) $\endgroup$
    – PhoemueX
    Apr 19 '20 at 5:58
  • $\begingroup$ Oh hmm thats true, I will think about it $\endgroup$ Apr 19 '20 at 17:47

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.