5.7 Complete Orthonormal Sets

Definition 5.26. A set of vectors \(\{x_n\}^{\infty }_{n=1}\) in a Hilbert space \(H\) is complete if for all \(n\geq 1\) \[(y,x_n)=0\implies y=0\]

Definition 5.27. An orthoormal basis is a complete orthonormal set.

Theorem 5.28. Let \(\{x_n\}^{\infty }_{n=1}\) be an orthonormal basis in a Hilbert space \(H\). Then for each \(y\in H\), \[y=\sum ^{\infty }_{n=1}(y,x_n)x_n\qquad (1)\] and \[||y||^2=\sum ^{\infty }_{n=1}|(y,x_n)|^2\qquad (2)\]

Proof. For all \(n\leq N\) \begin {align*} \Bigg (y-\sum ^N_{k=1}(y,x_k)x_k,x_n\Bigg ) &=(y,x_n)-\Bigg (\sum ^{N}_{k=1}(y,x_k)x_k,x_n\Bigg )\\ &=(y,x_n)-\sum ^N_{k=1}(y,x_k)(x_k,x_n)\\ &=(y,x_n)-(y,x_n)(x_n,x_n)\\ &=(y,x_n)-(y,x_n)\\ &=0 \end {align*}

This implies that \(\displaystyle {y-\sum ^N_{n=1}(y,x_n)x_n=0}\hspace {0.2cm}\) Since \(\{x_n\}\) is complete. That is,
\(\displaystyle {y=\lim _{N\rightarrow \infty }\sum ^N_{n=1}(y,x_n)x_n\implies y=\sum ^{\infty }_{n=1}(y,x_n)x_n}.\) Which proves (1).

For (2), we have that \begin {align*} ||y||^2 &=\lim _{N\rightarrow \infty }\Bigg |\Bigg |\sum ^N_{n=1}(y,x_n)x_n\Bigg |\Bigg |^2\\ &=\lim _{N\rightarrow \infty }\sum ^N_{n=1}|(y,x_n)|^2(x_n,x_n)\\ &=\sum ^{\infty }_{n=1}|(y,x_n)|^2.\\\\\\ \end {align*} □

Remark.

i
equation (1) implies that sum of the R.H.S converges to y.
ii
equation (2) is known as Parseval’s relation.

Theorem 5.29 (Gram–Schmidt Orthogonalisation Procedure). Every finite dimensional inner product space has an orthonormal basis.

This is done as follows:-

Let \(\{v_1,v_2,\dots ,v_n\}\) be a basis of the inner product \(V\). Then \begin {align*} e_1 &=v_1\\ e_2 &=v_2-\frac {(v_2,e_1)}{||e_1||^2}e_1\\ e_3 &=v_3-\frac {(v_3,e_2)}{||e_2||^2}e_2-\frac {(v_3,e_1)}{||e_1||^2}e_1\\ \vdots \\ e_n &=v_n-\frac {(v_n,e_{n-1})}{||e_{n-1}||^2}e_{n-1}-\frac {(v_n,e_{n-2})}{||e_{n-2}||^2}e_{n-2}-\cdots -\frac {(v_n,e_1)}{||e_1||^2}e_1 \end {align*}

Then \(\{e_1,e_2,\dots ,e_n\}\) is an orthogonal basis for \(V\).

To normalize, divide each vector by its norm that is set
\[P_i=\frac {e_i}{||e_i||},\hspace {2cm} i=1,2,\dots ,n\] Then \(\{P_1,P_2,\dots ,P_n\}\) is an orthonormal basis.

Proof. The construction in the statement is the Gram–Schmidt procedure; what needs checking is that it does what is claimed at every step.

Let \(\{v_1,\dots ,v_n\}\) be a basis and let \(e_1,\dots ,e_n\) be produced as described, each \(e_k\) being \(v_k\) minus its components along the earlier vectors.

No \(e_k\) is zero

Suppose \(e_k=0\). By construction \(e_k\) is \(v_k\) plus a linear combination of \(e_1,\dots ,e_{k-1}\), each of which is in turn a combination of \(v_1,\dots ,v_{k-1}\). So \(e_k=0\) would express \(v_k\) as a combination of \(v_1,\dots ,v_{k-1}\), contradicting independence of the basis. Hence each \(e_k\) is non-zero and may be normalised.

The \(e_k\) are pairwise orthogonal

By induction. For \(k=2\), \[(e_2,e_1)=\left (v_2-\frac {(v_2,e_1)}{||e_1||^2}e_1,\ e_1\right ) =(v_2,e_1)-\frac {(v_2,e_1)}{||e_1||^2}(e_1,e_1)=0 .\] Assume \(e_1,\dots ,e_{k-1}\) are pairwise orthogonal. For \(j<k\), taking the inner product of \(e_k\) with \(e_j\) kills every subtracted term except the \(j\)-th, and that one cancels exactly as above, giving \((e_k,e_j)=0\).

They span

Each \(v_k\) is recovered from \(e_1,\dots ,e_k\) by reversing the construction, so the span of \(\{e_1,\dots ,e_n\}\) contains every \(v_k\) and hence all of \(V\). Being \(n\) orthogonal non-zero vectors they are linearly independent, so after normalising, \(\{e_1/||e_1||,\dots ,e_n/||e_n||\}\) is an orthonormal basis. □

Definition 5.30. Let \(V\) be an inner product space. The projection of \(x\) onto \(y\) where \(x,y\in V\) is defined as \(\text {Proj}(x,y)=Cy\ni x-Cy\perp y\).

It is easy to show that \(C=\dfrac {(x,y)}{||y||^2}\) . That is \[x-Cy\perp y\implies (x-Cy,y)=0\] \[C(y,y)=(x,y)\] \[C=\frac {(x,y)}{||y||^2}\]

xCxyy− Cy

If \(\{x_1,x_2,\dots ,x_n\}\) is an orthogonal basis for \(A\) then \(x\in V, A\subset V\) \begin {align*} \text {Proj}(x,A) &=C_1x_1+C_2x_2+\cdots +C_nx_n\\ &=\frac {(x,x_1)}{||x_1||^2}x_1+\frac {(x,x_2)}{||x_2||^2}x_2+\cdots +\frac {(x,x_n)}{||x_n||^2}x_n \end {align*}

Make \(||A-x||\) as small as possible.

Example 5.31. For \(f,g\in C[-1,1]\), define \(\displaystyle {(f,g)=\int ^1_{-1}(1-x^2)f(x)g(x)dx}.\) This is an inner product on \([-1,1]\).

1.
Show that \((f,f)=0\implies f=0\).
2.
Produce an orthonormal basis for span \(\{1,x\}\).
3.
Find \(p\in \text {span}\{1,x\}\) that minimizes.

Solution.

1.
\(\displaystyle {(f,f)=\int ^1_{-1}(1-x^2)f^2(x)dx=0}\)

\(\implies (1-x^2)f^2(x)=0,\hspace {0.8cm} \text {on}\hspace {0.3cm} [-1,1]\)

\(\implies f=0,\hspace {1cm} \text {on}\hspace {0.4cm} (-1,1)\)

Since \(1-x^2>0\) on \((-1,1)\) \(\implies f=0\), on \([-1,1]\) since \(f\) is continuous.

2.
\(e_1 =1\)

\(e_2 =x-\dfrac {(x,1)}{||1||^2}1=x\)

\(\{1,x\}\) is orthogonal basis for \(C[-1,1]\).

\(||1||^2=\dfrac {4}{3}\), \(||x||^2=\dfrac {4}{15}\) orthonormal basis \(\Big \{\dfrac {\sqrt {3}}{2},\dfrac {\sqrt {15}}{2}x\Big \}\).

3.
\begin {align*} \text {p}(x) &=\text {proj}(x,\text {span}\{1,x\})\\ &=\frac {(x,1)}{||1||^2}1+\frac {(x,x)}{||x||^2}x\\ &=0+x\\ &=x. \end {align*}

\(\text {p}(x)=x\) minises.

Questions on this section

Stuck on something here? Ask below and it stays attached to this topic.