we have $ g: R^n \rightarrow R$ is a concave function and $S$={$x :g(x)> 0$} and $f:S \rightarrow R$ and $f(x)$=$1/g(x)$ so we must show that $f$ is a convex function

  • $\begingroup$ Are you getting this from Boyd & Vandenberghe's book? Because if so you can use their composition rules, if you carefully define what the outer function is. $\endgroup$ – Michael Grant Dec 9 '14 at 3:58
  • $\begingroup$ sorry,my master get me this exercise .I don't know which book has this $\endgroup$ – user157745 Dec 9 '14 at 4:15
  • $\begingroup$ That's fine. I just wanted to know what principles you are permitted to use. $\endgroup$ – Michael Grant Dec 9 '14 at 12:33
  • $\begingroup$ thanks .but the question didn't say any thing about domain.so we must assume the domain is true $\endgroup$ – user157745 Dec 9 '14 at 17:50
  • $\begingroup$ Ah, but it did: it's $S$. What you do is create a function that equals $1/x$ on $S$, but has the domain you need to make the proof work. $\endgroup$ – Michael Grant Dec 9 '14 at 17:51

Since $g$ is concave and positive on $S$, for all $x,y \in S$ and $\lambda \in [0,1]$,

$$g[\lambda x + (1- \lambda)y] \geqslant \lambda g(x) + (1-\lambda)g(y)> 0.$$

For every $x,y \in S$, we have $f(x),f(y) > 0$ and

$$f[\lambda x + (1- \lambda)y] = \frac{1}{g[\lambda x + (1- \lambda)y]} \\\leqslant \frac{1}{\lambda g(x) + (1-\lambda)g(y)}= \frac{1}{\lambda [f(x)]^{-1} + (1-\lambda)[f(y)]^{-1}} \\\leqslant \frac{1}{[f(x)]^{-\lambda}[f(y)]^{-(1-\lambda)}}= [f(x)]^{\lambda}[f(y)]^{(1-\lambda)} \\ \leq \lambda f(x) + (1-\lambda)f(y).$$

Therefore, $f$ is convex.

The following inequality -- for nonnegative $a,b$ and $0 \leqslant \lambda \leqslant 1$ -- is used above:

$$a^\lambda b^{1-\lambda} \leqslant \lambda a + ( 1 - \lambda) b.\tag{*}$$

Proof of (*).

Note that $h(t) = \lambda t - t^\lambda$ has a minimum at $t = 1$. Hence, $t^\lambda \leqslant \lambda t + 1 - \lambda$. Let $t = a/b$. Then $(a/b)^\lambda\leqslant \lambda(a/b) + 1 - \lambda.$ Multiplying both sides by $b$, we get $a^\lambda b^{1-\lambda}\leqslant \lambda a + (1 - \lambda)b.$

  • $\begingroup$ I can't understand this part in the proof of * $t^\lambda \leqslant \lambda t + 1 - \lambda$ . $\endgroup$ – user157745 Dec 13 '14 at 9:03
  • $\begingroup$ @user157745: $h(t)$ is minimized at $t=1$. So $h(t) \geqslant h(1) \implies \lambda t - t^{\lambda} \geqslant \lambda -1 \implies t^{\lambda} \leqslant \lambda t +1 - \lambda$. $\endgroup$ – RRL Dec 13 '14 at 9:09
  • $\begingroup$ yes ,thank you for all things $\endgroup$ – user157745 Dec 16 '14 at 16:24

Read Convex Optimization by Boyd & Vandenberghe---Chapter 3 specifically, but you need to be comfortable with the first two chapters to fully digest that. There you will find a set of composition rules that govern when convexity/concavity is preserved in the composition of nonlinear functions.

You can see a proof of one of those rules here on Math.SE, but that's not the one we want here. The rule we need to use is: the composition of a convex, decreasing (or nonincreasing) function with a concave function is convex.

That rule is satisfied here by $f(x)=h(g(x))$, where $$h(x)=1/x \qquad \mathop{\textrm{dom}} h=(0,+\infty)\supseteq S$$ Note that the restricted domain is important: without it, $1/x$ is neither convex nor concave. But with it, $h$ is a convex, decreasing function. So its composition with a concave function produces a convex result.

EDIT: the OP commented: "but the question didn't say any thing about domain.so we must assume the domain is true". In a sense, this is correct. But note that the function $g(x)$ is known to be strictly positive. Therefore, my function $h(g(x))$ is exactly equivalent to $1/g(x)$, and can serve as the basis for the proof. I would say that it was an error on my part not to explicitly point this out at first. Obviously, if $g(x)$ were not known to be positive, you could not do this.


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.