Optimizing seemingly a simple exponential function

103 Views Asked by At

I want to find the optimal solution to the following seemingly simple function:

$$\max_{x \in [0, 1]}~f(x) = x\mathrm{e}^{-a\left(bx^t+1\right)^\frac{1}{1-t}},$$ where $a, b > 0$ and $t \in (0, 1)$.

My attempt: I tried obtaining the optimal solution $x^*$ by exploring its properties. For example, I took the first derivative of $f(x)$ with respect to $x$, which is $$\frac{\mathrm{d} f(x)}{\mathrm{d} x} = \dfrac{\left(\left(t-1\right)\left(bx^t+1\right)^\frac{t}{t-1}+abtx^t\right)\exp\left({-\frac{a}{\left(bx^t+1\right)^{1/(t-1)}}}\right)}{\left(t-1\right)\left(bx^t+1\right)^\frac{t}{t-1}}.$$ The above derivative seems to not yield $x^*$ when the derivative is equated to $0$. How can one get the optimal solution or at least some properties regarding the optimal solution such as uniqueness and reasonable bounds on the optimal solution?

2

There are 2 best solutions below

6
On BEST ANSWER

First, note that it is true that:

$$ x^* = \text{argmax}_{x\in[0,1]} f(x) = \text{argmax}_{x\in[0,1]} \log(f(x)) $$ since $\log(\bullet)$ is an increasing function. Note that: $$ \log(f(x)) = \log(x) - a(bx^t+1)^{\frac{1}{1-t}} $$ Hence, $$ \frac{d}{dx}\log(f(x)) = \frac{1}{x} - \frac{abt}{1-t} (bx^t+1)^{\frac{1}{1-t} - 1}\frac{x^{t}}{x} $$ if we set $\frac{d}{dx}\log(f(x))=0$ we obtain: $$ \frac{abt}{1-t} (bx^t+1)^{\frac{1}{1-t} - 1}\frac{x^t}{x} = \frac{1}{x} $$ Assume, $x\neq 0$. Then,$\frac{abt}{1-t} (bx^t+1)^{\frac{1}{1-t} - 1}x^{t} = 1$.

Now, note that $b(\bullet)^t+1$ is an increasing and positive function, and the same for $(\bullet)^{\frac{1}{1-t}-1}$ and $(\bullet)^t$ since $\frac{1}{1-t}-1>0$. Hence, $g(x;a,b,t) = \frac{abt}{1-t} (bx^t+1)^{\frac{1}{1-t} - 1}x^t $ is an increasing function.

Now, to see if there is a solution to $g(x^*;a,b,t)=1$ in $(0,1)$ we use the fact that $g(0;a,b,t)=0$ and $g(1;a,b,t) = \frac{abt}{1-t} (b+1)^{\frac{1}{1-t} - 1}$. Since, $g(x;a,b,t)$ is continuous and increasing, there exists a solution of $g(x^*;a,b,t)=1$ with $0< x^*< 1$ only if $g(1;a,b,t) = \frac{abt}{1-t} (b+1)^{\frac{1}{1-t} - 1}> 1$.

Once you find this unique $x^*$, the only thing left is to compare the cost function $f(x^*)$ with the value at the boundaries of the interval, this is with $f(0)$ and $f(1)$. Note that $f(0)=0$ and $f(1) = \exp(-a(b+1)^\frac{1}{1-t})\geq f(0)=0$. If $x^*$ inside $(0,1)$ doesn't exists (if $\frac{abt}{1-t} (b+1)^{\frac{1}{1-t} - 1}<1$) or $f(x^*)\leq f(1)$ for such $x^*\in(0,1)$, then the optimal is certainly $x=1$.

Hope this helps!

0
On

I just did some plotting. For some values of $a$ and $b$ there is a max around $t=0.5$ and $x=0.1$ and for other values there is a max at the boundary at $t=0$ and $x=1$.

For finding the max over $x$ and $t$, you can also set $df/dt=0$ to get another equation. But I'm not sure if that helps.

Good luck :)