Expression of "closest" element in a Hilbert space

48 Views Asked by At

Let $H$ be a Hilbert space and $\{ e_i\}_{i\in \mathbb N}$ be an orthonormal system. Let $B = \{x \in \mathrm{span} \{ e_i\} : \Vert x \Vert \leq 1 \}$.

I was able to show that for any $y\in H$, there exists unique $z \in \overline{B}$ that minimizes $\Vert y-z \Vert$. For existence, we can give a Cauchy sequence $z_n$ whose norm converges to $\inf_{w \in B} \Vert w-y \Vert$; and take its limit as $z$. For uniqueness, if there are two distinct $z_1, z_2$ that minimizes the distance, we can take $z_3 = (z_1+z_2)/2$ to contradict the minimality.

Now what I am curious about is actual representation of $z$ (in terms of $e_i$'s), and whether $z \in B$. I think that $$z = \sum \langle y/\Vert y \Vert, e_i \rangle e_i$$ and $z \in B$. However, I cannot give a proof about it. (I figured out that $z \in B$ using Bessel's inequality, but no more.) Am I correct with the representation of $z$?

Thank you for any form of help, hint, or solution.

1

There are 1 best solutions below

0
On

What you guessed is close, but not exactly what it should be. There are two problems with your guess. One is that when $y\in\overline B$, dividing by the norm worsens the distance because $z=y$ already works. And when $y$ is not in $\overline B$, what you are missing is that there might be a part of $y$ that is orthogonal to $\overline B$ and thus plays no role (this are the $\alpha_k$ below).

Because $\{e_i\}$ is an orthonormal basis for $\def\spann{\operatorname{span}}\overline\spann B$, any element $w\in\overline B$ can be written as $$w=\sum_j\beta_j\,e_j$$ with $\sum_j|\beta_j|^2\leq1$. If we complete $\{e_j\}$ to an orthonormal basis $\{e_j\}\cup\{f_k\}$ of $H$, we can write $$ y=\sum_k\alpha_k\,f_k+\sum_j\gamma_j\,e_j,\qquad\qquad\sum_k|\alpha_k|^2+\sum_j|\gamma_j|^2=\|y\|^2. $$ Then \begin{align} \|y-w\|^2&=\sum_k|\alpha_k|^2+\sum_j|\gamma_j-\beta_j|^2\\[0.2cm] &=\sum_k|\alpha_k|^2+\sum_j|\gamma_j|^2+\sum_j|\beta_j|^2-2\operatorname{Re}\beta_j\overline{\gamma_j}\\[0.2cm] &=\|y\|^2+\|w\|^2-2\operatorname{Re}\beta_j\overline{\gamma_j}\\[0.2cm] &\geq\|y\|^2+\|w\|^2-2\|w\|\,\Big(\sum_j|\gamma_j|^2\Big)^{1/2}. \end{align} If $\sum_j|\gamma_j|^2\leq1$, then $w=\sum_j\gamma_j\,e_j\in\overline B$ and $\|y-w\|^2=\sum_k|\alpha_k|^2$ is minimum. If instead $\sum_j|\gamma_j|^2>1$, we notice that the inequality we used is Cauchy-Schwarz, and equality in Cauchy-Schwarz happens with both vectors are multiples of each other. Hence the minimum happens when $$ w=\lambda\sum_j\gamma_j\,e_j. $$ for some $\lambda>0$. Hence we get $\beta_j=\lambda\gamma_j$ for all $j$, and this forces the condition $$ 1\geq\sum_j|\beta_j|^2=\lambda^2\,\sum_j|\gamma_j|^2. $$ That is, $$ \lambda\leq\frac1{\Big(\sum_j|\gamma_j|^2\Big)^{1/2}}<1. $$ The expression to be minimized is now $$ \|y-w\|^2=\sum_k|\alpha_k|^2+\sum_j|\gamma_j|^2\,(1-\lambda)^2, $$ which will be minimum when $1-\lambda$ is maximum. That is, the minimum is achieved when $$ \lambda=\frac1{\Big(\sum_j|\gamma_j|^2\Big)^{1/2}}. $$ In other words, $$ z=\frac1{\Big(\sum_j|\gamma_j|^2\Big)^{1/2}}\,\sum_j\gamma_j\,e_j $$


So, to summarize, if $\sum_j|\gamma_j|^2\leq1$, then the minimizer is $$z=\sum_j\gamma_j\,e_j.$$ And when $\sum_j|\gamma_j|^2>1$, the minimizer is $$ z=\frac1{\Big(\sum_j|\gamma_j|^2\Big)^{1/2}}\,\sum_j\gamma_j\,e_j $$