0

I have a function $$ z(x) = (Kx)^{x/(1-x)}, x \in (0,1)\text{ and }K>1 $$ When I plot the function it has a U-shape. However when I take the second derivative wrt $x$ I have the following expression for it: $$ z\cdot \left[\left[\frac{1}{1-x} + \frac{\log(Kx)}{(1-x)^2}\right]^2+\frac{1}{(1-x)^2}+\frac{\frac{(1-x)^2}{x}+2\cdot(1-x)\log(Kx)}{(1-x)^4}\right] $$ The last term makes it problematic. Since the logarithm can be negative (even though rest all terms are positive). This is really strange as the graph of the function is strictly convex.

I am thinking for hours but do not figure out the mistake

martini
  • 84,101
  • I'd try to prove that this function is log-convex, and, therefore, convex. By the way, $K=1$ seems to work, too. – TZakrevskiy Jan 09 '14 at 11:52

1 Answers1

1

Let's prove that your function is log-convex. The second derivative of the logarithm of your function is (easy to check) $$\frac{ 2 x \ln (K x)-x^2+1}{(1-x)^3 x}.$$

The denominator is always positive, therefore no problem there. Let's study the positivity of the numerator:

$$2 x \log (K x)-x^2+1 = 2x\ln K + 1-x^2+2x\ln x.$$ The first term is nonnegative for $K\ge 1$, so it's ok, we omit it. So we have $$1-x^2+2x\ln x.$$The value of this function in $x=1$ is zero. It's derivative on $x\in (0,1)$ is $$ 2-2x+2\ln x = 2(1-x+\ln x).$$ To prove that this expression is negative is an easy exercise, therefore we conclude that $1-x^2+2x\ln x+2x\ln K$ is always nonegative on $[0,1]$ - in fact, it's zero in $x=1$ for $K=1$ and strictly greater than zero everywhere if $K>1$. If $K>1$, then $$\frac{ 2 x \ln (K x)-x^2+1}{(1-x)^3 x} $$is strictly positive on $[0,1]$. If $K=1$ then we need to study the following limit:$$\lim_{x\to 1}\frac{ 2 x \ln ( x)-x^2+1}{(1-x)^3 x}. $$ It's possible to show that in is, in fact, $+\infty$ (that what I called an exercise), thus for $K\ge 1 $ $$\frac{ 2 x \ln (K x)-x^2+1}{(1-x)^3 x}>0,\quad 0< x<1. $$ Hence, the function $(Kx)^{x/(1-x)}$ is strictly log-convex, and, in particular, strictly convex.

As to your problem with that second derivative, there's no error there. The logarithm can be indeed negative, but not "sufficiently negative" to counter all positive terms that you had before.

TZakrevskiy
  • 22,980
  • Thanks very much for the kind explanation. Just one small thing, could you please explain the part when you say 'simple reasoning with limits'? Would be really helpful – Abhimanyu Arora Jan 09 '14 at 13:27
  • 1
    @AbhimanyuArora you are welcome! I'll edit the post to clarify this detail. – TZakrevskiy Jan 09 '14 at 14:07