The Inner Product

The inner product (or ``dot product'', or ``scalar product'') is an operation on two vectors which produces a scalar. Defining an inner product for a Banach space specializes it to a Hilbert space (or ``inner product space''). There are many examples of Hilbert spaces, but we will only need $ \{{\bf C}^N,{\bf C}\}$ for this book (complex length $ N$ vectors, and complex scalars).

The inner product between (complex) $ N$-vectors $ \underline{u}$ and $ \underline{v}$ is defined by5.9

$\displaystyle \zbox {\left<\underline{u},\underline{v}\right> \isdef \sum_{n=0}^{N-1}u(n)\overline{v(n)}.}
$

The complex conjugation of the second vector is done in order that a norm will be induced by the inner product:5.10

$\displaystyle \left<\underline{u},\underline{u}\right> = \sum_{n=0}^{N-1}u(n)\o...
...sum_{n=0}^{N-1}\left\vert u(n)\right\vert^2 \isdef {\cal E}_u = \Vert u\Vert^2
$

As a result, the inner product is conjugate symmetric:

$\displaystyle \left<\underline{v},\underline{u}\right> = \overline{\left<\underline{u},\underline{v}\right>}
$

Note that the inner product takes $ {\bf C}^N\times{\bf C}^N$ to $ {\bf C}$. That is, two length $ N$ complex vectors are mapped to a complex scalar.

Linearity of the Inner Product

Any function $ f(\underline{u})$ of a vector $ \underline{u}\in{\bf C}^N$ (which we may call an operator on $ {\bf C}^N$) is said to be linear if for all $ \underline{u}\in{\bf C}^N$ and $ \underline{v}\in{\bf C}^N$, and for all scalars $ \alpha$ and $ \beta $ in $ {\bf C}$,

$\displaystyle f(\alpha \underline{u}+ \beta \underline{v}) = \alpha f(\underline{u}) + \beta f(\underline{v}).
$

A linear operator thus ``commutes with mixing.'' Linearity consists of two component properties:
  • additivity: $ f(\underline{u}+\underline{v}) = f(\underline{u}) + f(\underline{v})$
  • homogeneity: $ f(\alpha \underline{u}) = \alpha f(\underline{u})$
A function of multiple vectors, e.g., $ f(\underline{u},\underline{v},\underline{w})$ can be linear or not with respect to each of its arguments.

The inner product $ \left<\underline{u},\underline{v}\right>$ is linear in its first argument, i.e., for all $ \underline{u},\underline{v},\underline{w}\in{\bf C}^N$, and for all $ \alpha, \beta\in{\bf C}^N$,

$\displaystyle \left<\alpha \underline{u}+ \beta \underline{v},\underline{w}\rig...
...line{u},\underline{w}\right> + \beta \left<\underline{v},\underline{w}\right>.
$

This is easy to show from the definition:

\begin{eqnarray*}
\left<\alpha \underline{u}+ \beta \underline{v},\underline{w}\...
...rline{w}\right> + \beta \left<\underline{v},\underline{w}\right>
\end{eqnarray*}

The inner product is also additive in its second argument, i.e.,

$\displaystyle \left<\underline{u},\underline{v}+ \underline{w}\right> = \left<\underline{u},\underline{v}\right> + \left<\underline{u},\underline{w}\right>,
$

but it is only conjugate homogeneous (or antilinear) in its second argument, since

$\displaystyle \left<\underline{u},\alpha \underline{v}\right> = \overline{\alph...
...{u},\underline{v}\right> \neq \alpha \left<\underline{u},\underline{v}\right>.
$

The inner product is strictly linear in its second argument with respect to real scalars $ a$ and $ b$:

$\displaystyle \left<\underline{u},a \underline{v}+ b \underline{w}\right> = a \...
...ne{v}\right> + b \left<\underline{u},\underline{w}\right>, \quad a,b\in{\bf R}
$

where $ \underline{u},\underline{v},\underline{w}\in{\bf C}^N$.

Since the inner product is linear in both of its arguments for real scalars, it may be called a bilinear operator in that context.


Norm Induced by the Inner Product

We may define a norm on $ \underline{u}\in{\bf C}^N$ using the inner product:

$\displaystyle \zbox {\Vert\underline{u}\Vert \isdef \sqrt{\left<\underline{u},\underline{u}\right>}}
$

It is straightforward to show that properties 1 and 3 of a norm hold (see §5.8.2). Property 2 follows easily from the Schwarz Inequality which is derived in the following subsection. Alternatively, we can simply observe that the inner product induces the well known $ L2$ norm on $ {\bf C}^N$.


Cauchy-Schwarz Inequality

The Cauchy-Schwarz Inequality (or ``Schwarz Inequality'') states that for all $ \underline{u}\in{\bf C}^N$ and $ \underline{v}\in{\bf C}^N$, we have

$\displaystyle \zbox {\left\vert\left<\underline{u},\underline{v}\right>\right\vert \leq \Vert\underline{u}\Vert\cdot\Vert\underline{v}\Vert}
$

with equality if and only if $ \underline{u}=c\underline{v}$ for some scalar $ c$.

We can quickly show this for real vectors $ \underline{u}\in{\bf R}^N$, $ \underline{v}\in{\bf R}^N$, as follows: If either $ \underline{u}$ or $ \underline{v}$ is zero, the inequality holds (as equality). Assuming both are nonzero, let's scale them to unit-length by defining the normalized vectors $ \underline{\tilde{u}}\isdeftext
\underline{u}/\Vert\underline{u}\Vert$, $ \underline{\tilde{v}}\isdeftext \underline{v}/\Vert\underline{v}\Vert$, which are unit-length vectors lying on the ``unit ball'' in $ {\bf R}^N$ (a hypersphere of radius $ 1$). We have

\begin{eqnarray*}
0 \leq \Vert\underline{\tilde{u}}-\underline{\tilde{v}}\Vert^2...
...=& 2 - 2\left<\underline{\tilde{u}},\underline{\tilde{v}}\right>
\end{eqnarray*}

which implies

$\displaystyle \left<\underline{\tilde{u}},\underline{\tilde{v}}\right> \leq 1
$

or, removing the normalization,

$\displaystyle \left<\underline{u},\underline{v}\right> \leq \Vert\underline{u}\Vert\cdot\Vert\underline{v}\Vert.
$

The same derivation holds if $ \underline{u}$ is replaced by $ -\underline{u}$ yielding

$\displaystyle -\left<\underline{u},\underline{v}\right> \leq \Vert\underline{u}\Vert\cdot\Vert\underline{v}\Vert.
$

The last two equations imply

$\displaystyle \left\vert\left<\underline{u},\underline{v}\right>\right\vert \leq \Vert\underline{u}\Vert\cdot\Vert\underline{v}\Vert.
$

In the complex case, let $ \left<\underline{u},\underline{v}\right>=R e^{j\theta}$, and define $ \underline{\tilde{v}}=\underline{v}e^{j\theta}$. Then $ \left<\underline{u},\underline{\tilde{v}}\right>$ is real and equal to $ \vert\left<\underline{u},\underline{\tilde{v}}\right>\vert=R>0$. By the same derivation as above,

$\displaystyle \left<\underline{u},\underline{\tilde{v}}\right>\leq\Vert\underli...
...derline{\tilde{v}}\Vert = \Vert\underline{u}\Vert\cdot\Vert\underline{v}\Vert.
$

Since $ \left<\underline{u},\underline{\tilde{v}}\right>=R=\left\vert\left<\underline{...
...right>\right\vert=\left\vert\left<\underline{u},\underline{v}\right>\right\vert$, the result is established also in the complex case.


Triangle Inequality

The triangle inequality states that the length of any side of a triangle is less than or equal to the sum of the lengths of the other two sides, with equality occurring only when the triangle degenerates to a line. In $ {\bf C}^N$, this becomes

$\displaystyle \zbox {\Vert\underline{u}+\underline{v}\Vert \leq \Vert\underline{u}\Vert + \Vert\underline{v}\Vert.}
$

We can show this quickly using the Schwarz Inequality:

\begin{eqnarray*}
\Vert\underline{u}+\underline{v}\Vert^2 &=& \left<\underline{u...
...v}\Vert &\leq& \Vert\underline{u}\Vert + \Vert\underline{v}\Vert
\end{eqnarray*}


Triangle Difference Inequality

A useful variation on the triangle inequality is that the length of any side of a triangle is greater than the absolute difference of the lengths of the other two sides:

$\displaystyle \zbox {\Vert\underline{u}-\underline{v}\Vert \geq \left\vert\Vert\underline{u}\Vert - \Vert\underline{v}\Vert\right\vert}
$


Proof: By the triangle inequality,

\begin{eqnarray*}
\Vert\underline{v}+ (\underline{u}-\underline{v})\Vert &\leq &...
...}\Vert &\geq& \Vert\underline{u}\Vert - \Vert\underline{v}\Vert.
\end{eqnarray*}

Interchanging $ \underline{u}$ and $ \underline{v}$ establishes the absolute value on the right-hand side.


Vector Cosine

The Cauchy-Schwarz Inequality can be written

$\displaystyle \frac{\left\vert\left<\underline{u},\underline{v}\right>\right\vert}{\Vert\underline{u}\Vert\cdot\Vert\underline{v}\Vert} \leq 1.
$

In the case of real vectors $ \underline{u},\underline{v}$, we can always find a real number $ \theta\in[0,\pi]$ which satisfies

$\displaystyle \zbox {\cos(\theta) \isdef \frac{\left<\underline{u},\underline{v}\right>}{\Vert\underline{u}\Vert\cdot\Vert\underline{v}\Vert}.}
$

We thus interpret $ \theta$ as the angle between two vectors in $ {\bf R}^N$.


Orthogonality

The vectors (signals) $ x$ and $ y$5.11are said to be orthogonal if $ \left<x,y\right>=0$, denoted $ x\perp y$. That is to say

$\displaystyle \zbox {x\perp y \Leftrightarrow \left<x,y\right>=0.}
$

Note that if $ x$ and $ y$ are real and orthogonal, the cosine of the angle between them is zero. In plane geometry ($ N=2$), the angle between two perpendicular lines is $ \pi/2$, and $ \cos(\pi/2)=0$, as expected. More generally, orthogonality corresponds to the fact that two vectors in $ N$-space intersect at a right angle and are thus perpendicular geometrically.

Example ($ N=2$):

Let $ x=[1,1]$ and $ y=[1,-1]$, as shown in Fig.5.8.

Figure 5.8: Example of two orthogonal vectors for $ N=2$.
\includegraphics[scale=0.7]{eps/ip}

The inner product is $ \left<x,y\right>=1\cdot \overline{1} + 1\cdot\overline{(-1)} = 0$. This shows that the vectors are orthogonal. As marked in the figure, the lines intersect at a right angle and are therefore perpendicular.


The Pythagorean Theorem in N-Space

In 2D, the Pythagorean Theorem says that when $ x$ and $ y$ are orthogonal, as in Fig.5.8, (i.e., when the vectors $ x$ and $ y$ intersect at a right angle), then we have

$\displaystyle \Vert x+y\Vert^2 = \Vert x\Vert^2 + \Vert y\Vert^2$   $\displaystyle \mbox{($x\perp y$)}$$\displaystyle . \protect$ (5.1)

This relationship generalizes to $ N$ dimensions, as we can easily show:
$\displaystyle \Vert x+y\Vert^2$ $\displaystyle =$ $\displaystyle \left<x+y,x+y\right>$  
  $\displaystyle =$ $\displaystyle \left<x,x\right>+\left<x,y\right>+\left<y,x\right>+\left<y,y\right>$  
  $\displaystyle =$ $\displaystyle \Vert x\Vert^2 + \left<x,y\right>+\overline{\left<x,y\right>} + \Vert y\Vert^2$  
  $\displaystyle =$ $\displaystyle \Vert x\Vert^2 + \Vert y\Vert^2 + 2$re$\displaystyle \left\{\left<x,y\right>\right\}
\protect$ (5.2)

If $ x\perp y$, then $ \left<x,y\right>=0$ and Eq.$ \,$(5.1) holds in $ N$ dimensions.

Note that the converse is not true in $ {\bf C}^N$. That is, $ \Vert x+y\Vert^2 = \Vert x\Vert^2 + \Vert y\Vert^2$ does not imply $ x\perp y$ in $ {\bf C}^N$. For a counterexample, consider $ x= (j,1)$, $ y=
(1, -j)$, in which case

$\displaystyle \Vert x+y\Vert^2 = \Vert 1+j,1-j\Vert^2 =
4 = \Vert x\Vert^2 + \Vert y\Vert^2
$

while $ \left<x,y\right> = j\cdot 1 + 1 \cdot\overline{-j} = 2j$.

For real vectors $ x,y\in{\bf R}^N$, the Pythagorean theorem Eq.$ \,$(5.1) holds if and only if the vectors are orthogonal. To see this, note that, from Eq.$ \,$(5.2), when the Pythagorean theorem holds, either $ x$ or $ y$ is zero, or $ \left<x,y\right>$ is zero or purely imaginary, by property 1 of norms (see §5.8.2). If the inner product cannot be imaginary, it must be zero.

Note that we also have an alternate version of the Pythagorean theorem:

$\displaystyle x\perp y\,\,\Rightarrow\,\,
\Vert x-y\Vert^2 = \Vert x\Vert^2 + \Vert y\Vert^2
$


Projection

The orthogonal projection (or simply ``projection'') of $ y\in{\bf C}^N$ onto $ x\in{\bf C}^N$ is defined by

$\displaystyle \zbox {{\bf P}_{x}(y) \isdef \frac{\left<y,x\right>}{\Vert x\Vert^2} x.}
$

The complex scalar $ \left<y,x\right>/\Vert x\Vert^2$ is called the coefficient of projection. When projecting $ y$ onto a unit length vector $ x$, the coefficient of projection is simply the inner product of $ y$ with $ x$.

Motivation: The basic idea of orthogonal projection of $ y$ onto $ x$ is to ``drop a perpendicular'' from $ y$ onto $ x$ to define a new vector along $ x$ which we call the ``projection'' of $ y$ onto $ x$. This is illustrated for $ N=2$ in Fig.5.9 for $ x= [4,1]$ and $ y=[2,3]$, in which case

$\displaystyle {\bf P}_{x}(y) \isdef \frac{\left<y,x\right>}{\Vert x\Vert^2} x
...
...{1})}{4^2+1^2} x
= \frac{11}{17} x= \left[\frac{44}{17},\frac{11}{17}\right].
$

Figure: Projection of $ y$ onto $ x$ in 2D space.
\includegraphics[scale=0.7]{eps/proj}

Derivation: (1) Since any projection onto $ x$ must lie along the line collinear with $ x$, write the projection as $ {\bf P}_{x}(y)=\alpha
x$. (2) Since by definition the projection error $ y-{\bf P}_{x}(y)$ is orthogonal to $ x$, we must have

\begin{eqnarray*}
(y-\alpha x) & \perp & x\\
\;\Leftrightarrow\;\left<y-\alpha...
...}{\left<x,x\right>}
= \frac{\left<y,x\right>}{\Vert x\Vert^2}.
\end{eqnarray*}

Thus,

$\displaystyle {\bf P}_{x}(y) = \frac{\left<y,x\right>}{\Vert x\Vert^2} x.
$

See §I.3.3 for illustration of orthogonal projection in matlab.


Next Section:
Signal Reconstruction from Projections
Previous Section:
Signal Metrics