An $\mathcal{L}^1$ optimization problem in function space

129 Views Asked by At

Let $f: (-\infty,\infty)\to [0,\infty)$ be a smooth map. Then, how to solve the following optimizatin problem w.r.t. $f$?

\begin{align} \mathrm{minimize}_f&~~\int_{-\infty}^{\infty} |f(x) - \exp(-x^2)| dx\\ s.t.&~~\int_{-\infty}^{\infty} x^2 f(x) = 1 \end{align}

Thanks!

1

There are 1 best solutions below

13
On BEST ANSWER

There is no solution. To see this, use the Lagrange multiplier's theorem infinite dimensional in the space of integrable functions to find critical points of $$ \int_{-\infty}^{\infty}\left\vert f(x)-e^{-x^{2}}\right\vert \,dx+\lambda \left( \int_{-\infty}^{\infty}x^{2}f(x)\,dx-1\right) . $$ Now the function $e^{-x^{2}}$ satisfies $\int_{-\infty}^{\infty}x^{2} e^{-x^{2}}dx=\frac{1}{2}\sqrt{\pi}$ so it is not a solution. Hence, the solution $f$ cannot always be equal to $f$. Consider the set $\left\{ x:\,f(x)>e^{-x^{2}}\right\} $. Since this set is open, we can write it as a union of intervals. Let $(a,b)$ be one of these intervals and let $g\in C_{c}((a,b))$. Then for $t>0$ small enough we have that $f(x)+tg(x)>e^{-x^{2} }$ in $(a,b)$. Consider the function$$ F(t)=\int_{a}^{b}(f(x)+tg(x)-e^{-x^{2}})\,dx+\lambda\left( \int_{a}^{b} x^{2}(f(x)+tg(x))\,dx-1\right) . $$ We must have $F^{\prime}(0)=0$, that is,$$ 0=\int_{a}^{b}g(x)\,dx+\lambda\int_{a}^{b}x^{2}g(x)\,dx=\int_{a}^{b}(1+\lambda x^{2})g(x)\,dx. $$ Since this is true for every $g$ it follows that $1+\lambda x^{2}=0$ for all $x\in(a,b)$, which is impossible.

I think that there should be a simpler way to prove this.

Edit To see why we can apply the Lagrange multiplier theorem in $(a,b)$, assume that $f$ is a solution of the original problem, with $\int_{-\infty}^{\infty}x^{2}f(x)\,dx=1$ and let $[a,b]$ be such that $f(x)-e^{-x^2}>$ in $[a,b]$ (so I am taking $(a,b)$ a bit smaller than above). Then $f(x)-e^{-x^2}\ge\delta/(b-a)>0$ in $[a,b]$ for some $\delta$. Let $m=\int_{a}^{b}x^2 f(x)\,dx$. Consider $g\in C_c((a,b))$ with $\Vert g\Vert_{C([a,b])}\le\delta$ and $\int_a^b x^2g(x)\,dx=0$. Extend $g$ to be zero outside $(a,b)$. Then the function $f+g$ satisfies the constraint $\int_{-\infty}^{\infty}x^{2}(f(x)+g(x))\,dx=1$ and so $$\int_{-\infty}^{\infty}\left\vert (f(x)+g(x))-e^{-x^{2}}\right\vert \,dx\ge \int_{-\infty}^{\infty}\left\vert f(x)-e^{-x^{2}}\right\vert \,dx,$$ which implies that $$\int_{a}^{b}\left((f(x)+g(x))-e^{-x^{2}}\right) \,dx\ge \int_{a}^{b}\left( f(x)-e^{-x^{2}}\right) \,dx.$$ This implies that $f$ is a constrained minimizer of the functional $$H(h):=\int_{a}^{b}\left(h(x)-e^{-x^{2}}\right) \,dx$$ over all $h$ in the open ball of $C_0((a,b))$ centered at $f$ and radius $\delta$, subject to the constraint $\int_{a}^{b}x^2 h(x)\,dx=m$. We can now apply the Lagrange multiplier theorem as above to get a contradiction.