View on GitHub

memo

Hilbert Space

Hilbert Space

Proposition 1

\[\langle \alpha_{1}, \beta_{1} \rangle \pm \langle \alpha_{2}, \beta_{2} \rangle = \langle \alpha_{1}, \beta_{1} \pm \beta_{2} \rangle \pm \langle \alpha_{2} - \alpha_{1}, \beta_{2} \rangle\]

proof.

\[\begin{eqnarray} \langle \alpha_{1}, \beta_{1} \rangle \pm \langle \alpha_{2}, \beta_{2} \rangle & = & \langle \alpha_{1}, \beta_{1} \rangle \pm \langle \alpha_{1} - \alpha_{1} + \alpha_{2}, \beta_{2} \rangle \nonumber \\ & = & \langle \alpha_{1}, \beta_{1} \pm \beta_{2} \rangle \pm \langle \alpha_{2} - \alpha_{1}, \beta_{2} \rangle \nonumber \end{eqnarray}\]
$\Box$

Theorem 2

Then $\forall x \in X$, \(\exists ! (y, z)\), s.t. \(y \in L\), \(z \in L^{\perp}\),

\[\begin{equation} x = y + z \label{hilbert_space_orthogonal_projection} \end{equation}\]

proof.

For uniquness, suppose that both \(y_{1}, z_{1}\) and \(y_{2}, z_{2}\) satisfy \(\eqref{hilbert_space_orthogonal_projection}\).

\[\begin{eqnarray} & & y_{1} - z_{1} = y_{2} - z_{2} \\ & \Leftrightarrow & y_{1} - y_{2} = z_{1} - z_{2} \end{eqnarray}\]

Hence \(y_{1} - y_{2} \in L\), and \(z_{2} - z_{2} \in L^{\perp}\). \(L \cap L^{\perp} = \{0\}\) so that \(y_{1} - y_{2} = z_{1} - z_{2} = 0\).

For existence, Let $\delta$ be

\[\begin{equation} \delta := \mathrm{dist}(x, L) = \inf_{h \in L} \| x - h\|. \label{inf} \end{equation}\]

There is a sequence \(\{h_{n}\}_{n \in \mathbb{N}} \subset L\) such that

\[\|x - h_{n}\| \rightarrow \delta \ (n \rightarrow \infty).\]

By parallelogram law,

\[\begin{eqnarray} & & \|(x - h_{n}) + (x - h_{m})\|^{2} + \|(x - h_{n}) - (x - h_{m})\|^{2} = 2 \|x -h_{n}\|^{2} + 2 \|x -h_{m}\|^{2} \nonumber \\ & \Leftrightarrow & \|2x - (h_{n} + h_{m})\|^{2} + \|h_{m} - h_{n}\|^{2} = 2 \|x -h_{n}\|^{2} + 2 \|x -h_{m}\|^{2}. \label{inequality_from_parallelogram_law} \end{eqnarray}\]

Since

\[\frac{ h_{n} + h_{m} }{ 2 } \in L,\]

\(2\|x - (h_{n} + h_{m})/2\| \ge 2 \delta\). From \(\eqref{inequality_from_parallelogram_law}\),

\[\begin{eqnarray} 0 & \le & \|h_{m} - h_{n}\|^{2} \nonumber \\ & = & 2 \|x -h_{n}\|^{2} + 2 \|x -h_{m}\|^{2} - \|2x - (h_{n} + h_{m})\|^{2} \nonumber \\ & \le & 2 \|x -h_{n}\|^{2} + 2 \|x -h_{m}\|^{2} - 4 \delta^{2} \nonumber \end{eqnarray}\]

\(h_{n}\) is cauchy sequence. Since $H$ is complete, the sequence converges to a point, say

\[y := \lim_{n \rightarrow \infty} h_{n}.\]

\(h_{n} \in L\) and $L$ is closed so that $y \in L$. By \(\eqref{inf}\),

\[\begin{equation} \delta = \|x - y\| = \min_{h \in L} \|x - h\|. \label{minumum_value} \end{equation}\]

Finally we will show \(z := x - y \perp L\). Suppose $\theta \in \mathbb{R}$, then we define for all $h \in L$

\[\eta(\theta) := \|x - y - \theta h\|^{2} = \| x - (y + \theta h)\|^{2}.\]

From \(\eqref{minumum_value}\), it is easy to show that $\eta$ takes minimum value $\delta^{2}$ when $\theta = 0$. Therefore \(\eta^{\prime}(0) = 0\). On the other hand,

\[\eta(\theta) = \|x - y\|^{2} - 2\theta \mathrm{Re}(x - y, h) + \theta^{2} \|h\|^{2}\]

\(\eta^{\prime}(0) = -2 \mathrm{Re}(x - y, h)\). Hence

\[\begin{equation} \mathrm{Re}(x - y, h) = 0. \label{equation_perp_for_real_part} \end{equation}\]

Substituting $ih$ into \(\eqref{equation_perp_for_real_part}\) instead of $h$,

\[0 = \mathrm{Re}(x - y, ih) = \mathrm{Re}(-i(x - y, h)) = \mathrm{Im}(x - y, h).\]

By \(\eqref{equation_perp_for_real_part}\), we obtain

\[(x - y, h) = 0 \ \forall h \in L.\]
$\Box$

Theorem 3

Then $\forall x \in X$, \(\exists ! y \in K\), s.t.

\[\begin{equation} \|x - y\| = \min_{h \in K} \|x - h\|. \label{projection_to_closed_convex_set} \end{equation}\]

Moreover, it is equivalent to existence of $y \in K$ satisfying

\[\begin{equation} \forall v \in K, \ \mathrm{Re}(x - y, v - y) \le 0 \label{projection_to_closed_convex_set_equivalent_condition} \end{equation}\]

proof.

To prove existence of $y$ satisfying \(\eqref{projection_to_closed_convex_set}\), we only need to replace $L$ with $K$ in the above proof.

Now we prove \(\eqref{projection_to_closed_convex_set} \Leftrightarrow \eqref{projection_to_closed_convex_set_equivalent_condition}\).

($\Rightarrow$)

Suppose $v \in K$ and \(0 \le \theta \le 1\),

\[y + \theta(v - y) = (1 - \theta)y + \theta v \in K.\]

Hence by taking $h := v - y$ we can define

\[\eta(\theta) := \|x - y - \theta h\|^{2} = \|x - (y + \theta h)\|^{2}.\]

$\eta$ takes the minimum value when $\theta = 0$.

\[\eta(\theta) = \|x - y\|^{2} + 2 \theta \mathrm{Re}(x - y, h) + \theta^{2} \|h\|^{2}.\] \[\begin{eqnarray} & & \eta^{\prime}(0) = 0 \nonumber \\ & \Leftrightarrow & \mathrm{Re}(x - y, h) = 0. \nonumber \end{eqnarray}\]

(\(\Leftarrow\))

Suppose \(y_{1}, y_{2} \in K\) satisfying \(\eqref{projection_to_closed_convex_set_equivalent_condition}\). Then

\[\begin{eqnarray} \mathrm{Re}(x - y_{1}, y_{2} - y_{1}) & \le & 0, \nonumber \\ \mathrm{Re}(x - y_{2}, y_{1} - y_{2}) & \le & 0. \nonumber \end{eqnarray}\]

By adding equations,

\[\mathrm{Re}(y_{2} - y_{1}, y_{2} - y_{1}) = \| y_{2} - y_{1} \|^{2} \le 0.\]

Therefore \(y_{1} = y_{2}\).

$\Box$

Definition.

\(\mathcal{H}_{a, c}\) is called halfspace iff

\[\mathcal{H}_{a, c} := \{x \in H \mid \langle a, x \rangle \le c\}\]

\(P_{\mathcal{H}_{a, c}}\) defined as below is said to be a projection onto hyperplane \(\mathcal{H}_{a, c}\).

\[x \in H, \ P_{\mathcal{H}_{a, c}}(x) := \arg \inf\{\|x - y \| \mid y \in \mathcal{H}_{a, c}\} \in H\]

Remark

\(\mathcal{H}_{a, c}\) is a convex set so that the definition of a projection on to hyperplane is well-defined. (i.e. The point which minimizes \(\|x - y\|\) is unique).

If \(x \in \mathcal{H}_{a, c}\), \(P_{\mathcal{H}_{a, c}}(x) = x\).

Theorem 4

\[\forall x \notin \mathcal{H}_{a, c}, \ P_{\mathcal{H}_{a,c}}(x) = \begin{cases} x - \frac{ \langle a, x \rangle - c }{ \| a \|^{2} } a & (x \notin \mathcal{H}_{a, c}) \\ x & (x \in \mathcal{H}_{a, c}) \end{cases}\]

proof.

For simplicity, we define \(\mathcal{H} := \mathcal{H}_{a, c}\). Suppose $v \in \mathcal{H}$ and \(x \notin \mathcal{H}\),

\[\langle a, x \rangle - c > 0, \ \langle a, v \rangle - c \le 0.\]

Then

\[\begin{eqnarray} \forall v \in \mathcal{H}, \ \left\langle x - \left( x - \frac{ \langle a, x \rangle - c }{ \| a \|^{2} } a \right) , v - \left( x - \frac{ \langle a, x \rangle - c }{ \| a \|^{2} } a \right) \right\rangle & = & \left\langle \frac{ \langle a, x \rangle - c }{ \| a \|^{2} } a , (v - x) + \frac{ \langle a, x \rangle - c }{ \| a \|^{2} } a \right\rangle \nonumber \\ & = & \left\langle \frac{ \langle a, x \rangle - c }{ \| a \|^{2} } a , (v - x) \right\rangle + \left\langle \frac{ \langle a, x \rangle - c }{ \| a \|^{2} } a , \frac{ \langle a, x \rangle - c }{ \| a \|^{2} } a \right\rangle \nonumber \\ & = & \frac{ \langle a, x \rangle - c }{ \| a \|^{2} } \left\langle a, (v - x) \right\rangle + \frac{ (\langle a, x \rangle - c)^{2} }{ \| a \|^{2} } \nonumber \\ & = & \frac{ \langle a, x \rangle - c }{ \| a \|^{2} } \left( \left\langle a, (v - x) \right\rangle + \langle a, x \rangle - c \right) \nonumber \\ & = & \frac{ \langle a, x \rangle - c }{ \| a \|^{2} } \left( \left\langle a, v \right\rangle - c \right) \nonumber \\ & = & \frac{ \langle a, x \rangle - c }{ \| a \|^{2} } \left( \left\langle a, v \right\rangle - c \right) \le 0. \end{eqnarray}\]

By the theorem of projection on to closed convex set, the statement holds.

$\Box$

Reference