Can anyone explain to me how the "smoothing argument for inequalities" works?

I know that basically it can be used to prove an inequality $f(a_1,a_2,\cdots,a_n)\geq C$ subject to the constraint $a_1+a_2+\cdots+a_n=k$ (for constants $C,k$) if equality holds when all $a_i$ are equal. I've heard that the basic idea is to push the $a_i$ closer together, showing that this is pushing $f(a_1,a_2,\cdots,a_n)$ to an extrema.

After doing some research, I saw that this can be rigorously defined by showing inequalities of the form $f(a_1,a_2,\cdots,a_n)\geq f(\frac{k}{n},x_1+x_2-\frac{k}{n},\cdots,x_n)$, but I don't really see a pattern here (why do you replace $a_1$ with $\frac{k}{n}$, $a_2$ with $x_1+x_2-\frac{k}{n}$, etc.) Can someone explain this?

Finally, could someone show me how to apply this to an actual inequality problem? For example, how would one solve the following inequality using smoothing (yes, I am aware that there is a quite simple solution using substitutions and AM-GM)?

Let $a_1,\cdots a_n$ be real numbers in the interval $(0,\frac{\pi}{2})$ such that $\sum_{i=0}^n \tan(a_i-\frac{\pi}{4})\geq n-1$. Prove that $\prod_{i=0}^n\tan(a_i)\geq n^{n+1}$

  • $\begingroup$ My edit consisted of replacing "numbners" in the last sentence with "numbers". My most common typo is "sunset" for "subset". $\endgroup$ – DanielWainfleet Aug 8 '16 at 18:58

The general idea is that,once you know that a min of some function F exists,it can only occur when all the variables are equal,by showing that if they are not, then a small change in one or two of them will lower the value of F. For the example you give, first some simplification: Let $b_i= \tan {(a_i- \pi /4)}$ .Then $ -1 < b_i < 1$ and $\tan a_i = (1+b_i)/(1-b_i)$. Now the logarithm of the product in the example is $$F= \sum ^n_{i=o} \log [(1+b_i)/(1-b_i)].$$ Let us try to minimize this, subject to $$n-1 \le S= \sum ^n_{i=0} b_i$$ and each $|b_i| < 1 $..... Observe that the second derivative of $ g(x)= \log [(1+x)/(1-x)]$ is continuous and positive for $|x| <1$ which implies that $ 2g([x+y]/2) < g(x)+g(y)$ when $-1 < x <y < 1$. THEREFORE if $a_i \ne a_j$ for some $i,j$ then $b_i \ne b_j$ so $$2g([b_i+b_j]/2)<g(b_i)+g(b_j)$$. SO,... if we replace each of these unequal $b_i,b_j$ with their average then $F$ is lowered while $S$ is unchanged.So F cannot have a minimum except possibly when all the $a_i$ are equal. In this problem,if we do this replacement repeatedly,choosing the largest and smallest $b_i$ and $b_j$, they will all converge to the average of the original set of $b_i$'s, which is in the allowable domain and F is continuous, so the minimum (for a given $S$, at least $n-1$ ) occurs there.(Note that for a given S ,it doesn't matter what the original set is.) For some problems,proving that a minimum exists is easier, for example if the domain is closed and bounded in $R^n$ and $F$ is continuous, it will.

  • $\begingroup$ I think my last paragraph could use some $\endgroup$ – DanielWainfleet Aug 17 '15 at 8:21
  • 1
    $\begingroup$ I think the last paragraph of my answer could be written better and my cat should get off the keyboard $\endgroup$ – DanielWainfleet Aug 17 '15 at 8:23

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.