• Signal representation in terms of basis function

– Representing signals as a function of time is just one way, the

most obvious perhaps, but not the only way.

– Another technique is to express a signal as a linear combination of

other signals M
xt    am m t 
m 0
– Whilst this may seem a bizarre thing to want to do, it is often easier
to deal with the response to these individual basis function
separately.

– Dealing with the linear system we can add up all the individual
m t 
amthrough
results, i.e. what happens to each of the the
system, to find the result we want.

– To play this game we need the system to be linear. If the system is

non-linear the answers would be different (superposition does not
hold).
– Essentially we are aiming to solve a lot of easy problems instead
of one hard one. In the initial stages it will appear that we are
solving easy problems a very long-winded way, but this hopefully
will no last.

– Now consider 
xt    a  t 
m m
m  
M
xt    a  t 
m m
m N

xt   a N   N t     aM  M t 

– To represent arbitrary function exactly we need an infinite number

of basis functions. This is impractical so we often make do with a
finite number, although it is important to know how much error we
are introducing by this.

– Note that sometimes we use


xt    am m t 
m 0
– as there are still an infinite number of basis functions the
approximation is still exact.
• How to choose basis functions

– Choosing basis functions at this stage will seem essentially like a

black art.

– For LTI functions sine, cosine an complex exponentials work well.

– The reason for this is they form the eigenfunctions of LTI. Follow
the following rules in the choice of basis functions :

1. Choose the basis functions so that they can represent all time
functions, i.e. with an infinite number of basic functions
approximation should be perfect. In practice, we make do with
a finite number of basis functions and some error.

2. Choose the basis functions to be orthogonal or better still

orthonormal.

3. Choose them to make it as easy as possible for us to get the

answer.
– The following is the rule for orthogonality on interval .


t
1 2
  t 
 n t dt 
* 0, k  n
k , k  n
t1  t 2 t1
k

– where * means complex conjugate. Note : the signals and the

basis functions may be complex.

– If  k* t  is real the complex conjugate has no effect.


– If the above equation is true we say that  n t form an orthogonal
set of basis functions, where k = n, the equation reduces to
t
1 2
  t 
 n t dt
*

t1  t 2 t1
n

which should look familiar as the formula for the average power in
the interval ( t1 , t2 ), note the requirement of complex conjugatation
for calculating the power of a complex sequence.

– Furthermore if k  1k then n t  form an orthonormal set of

basis function.
• Example of testing functions for orthogonality
1. Sinwave
– This basis function n t   sin 2nt  is orthogonal over the
interval -0.5 to +0.5, n > 0 ?
0.5
sin 2kt  sin 2nt dt
1
0.5   0.5 0.5
I

1
cos2 k  n t   cos2 k  n t dt
2
I

1  sin 2 k  n t  sin 2 k  n t  

0.5

I  
2  2 k  n  2 k  n   0.5
I  0, k  0
I  0.5, k  n
where the fact that sin x  / x   (for small ) has been used.
– A point to note about using sine waves as a set of basis functions
is that no amount of sine functions (without phase shift that is) can
ever be added to make a cosine function. The reason for this is
that sine is an odd function and cosine is an even function.
– Also note that when k = n the equation gives the power in a
sinewave, i.e. 0.5 (this is signal not circuit theory so we have left
out the resistance).
2. Polynomial
A set of basis functions defined by
m t   t m m  0
These can be used to represent a well-behaved function over a finite
interval by the use of a Taylor series and hence there is no argument that
these are potentially useful functions.
The signal is represented by
N
Thus xt    amt m
m 0

0 t   1
1 t   t
 2 t   on
The question is are they orthogonal
2
t the interval (-0.5,+0.5) ?
 I  0 if (n+m+1) is
odd
0.5 I = 0 if (n+m+1) is
even I
1
0.5   0.5 0.5
t n m *
t dt  
 0. 5
 t n  m 1 
I  
 n  m  1   0. 5
Thus the set is not orthogonal as it can be non-zero for n  m, e.g.
 n t   t 0  m and
t  t2
Remember it is not good enough to be zero for just some n  m, it has to
be for all values.

The obvious question to this is so what ? We can still represent the

function using the equation (1). The answer is that unless the functions
are orthogonal there is often no simple way of determining the value of

If they are orthogonal, there isa(A

m side Legendre Polynomial are a set of
polynomials are a set of polynomial with desirable property
orthogonality).
3. Complex Exponential
Consider the functions m t   e j 2mt over the interval (-0.5,+0.5).
Thus we will be aiming to represent a signal f(t) by

f t   a m e j 2mt
m  
Using the orthogonality test where
0.5
I
1
0.5   0.5 0.5
e j 2mt
e 
j 2nt *
dt 
 0.5

I = 0 for m  n I   e j 2 m  n t dt
I = 1 for m = n  0.5

Therefore the complex exponentials are orthogonal. They are also

orthonormal, since the power of the basis function is always equal to 1.

k
4. Determining the coefficients of the basis functions
Suppose we have a set of orthogonal basis functions
n t
which meet the orthogonality conditions. That means we can write
 
an arbitrary periodic function of time x(t) as the following,

xt    am m t 
m  
and also the equation of orthogonality is hold.
A simple way of seeing how to do this is to examine the problem of
approximating a function f(t) over an interval [-T/2,+T/2] by a weighted
sum of just two real basis functions and (note since
the basis functions are real we have dropped the conjugate). t 1 
 t 2 
The problem is to choose a1 and a2 optimally. One way is to minimise the
mean square error, f t   a11 t   a22 t 

T /2
E
1
  f t   a 
1 1 t   a 
2 2 t 2
dt
T T / 2
We can thus differentiate partially with respect to a1 to obtain the
following condition for a minimum
E   1 T / 2 
      
a1 a1 T T/ 2
  f t  a 
1 1 t  a 
2 2 t
2
dt 

E 1 T / 2 
      
a1 T T/ 2 a1
 f t  a11 t  a 2 2 t
2
dt

E T /2

It must be equal to zero

2
for a  f t   a11 tThus
minimum.   a22 t 1 t dt
a1 T T / 2

Now if  2  and  last term is

T /2 T /2 T /2

 
equal to zero.
f t 1 t dt  a 
are orthogonal
 
1 1 t 2
dt  a 2 2 t 
on [-T/2,+T/2] the
  1 t dt 0
T T / 2 T / 2 T / 2 
If we had more than just two non-orthogonal basis functions the
1 t  cross
number of these t  rapidly becomes unmanageable and would
2term
mean we have to solve a large set of simultaneous equations to find the
optimal coefficients of the basis functions.
Then we have
T /2 T /2
1
    1
  
T T/ 2 T T/ 2
f t 1 t dt  a1 1 t
2
dt
Therefore
T /2

 f t  t dt
1
where a1 
a1T
1
T / 2

T /2
 
It is important to realise athat
1
  t 2
dta and a are independent
T because
1 1
1 2
T / 2
a1 can be determined independently of a2 and vice versa.

Now there are some conditions where it is not possible to calculate the
expansion given in the signal representation but when dealing with
Fourier series (either sine and cosine form or the complex exponential
form) and a periodic signal obtained from real world measurements, it
can always be done.
Now let the period of the signal x(t) be equal to T. The way to determine
the coefficients an  representation is to multiply by the
in the signal
conjugate of the corresponding basis functions, i.e. and
 n t  over one period.
*
integrate

The starting point t0 of the integration is unimportant as long as it is over

one period T t 0

I
1
 x t  *
n t dt
T t0
T t0
  
  am m t  n* t dt
1
I
T 
t0  m   
 T t0
The reason we can arrange a
I  them above m t  *
n t 
 and then do each part separately.
equation
dt is that integration is
  T
linear, i.e. we can split it up minto parts
t 0
Now since the basis functions m t 
are orthogonal, we know for the
previous equation that
T t0

  m t  *
n t dt
0 t
is only non-zero for m = n, then the above equation can thus reduced
to
T t0

 n n t dt
 
a
I n  t  *

T t0

an n been chosen to be orthogonal on the

I  have
Because the basis functions
period [-T/2,T/2], we can obtain the coefficients an by

Now what is n ? From the aboveT t0

1 equation it is the integral
xt  n t dt
nT t0
an  *

which should be recognised by now as the power of basis function.

T t
1 0
n    n t  n* t dt
T t0
Nothing that  basis
 t the
,xif n
 t  function 
was equal1 to
t  sin(t) it can be
realised immediately that 1 = 0,5 , the power in a sinewave

Now this is the most general form of finding the coefficients of a set of
orthogonal functions. From the above equations we can derive the
formula for Series Fourier series using sin and cos terms, or complex
exponentials, or Walsh function
Parseval’s Theorem
Power in the signal = the sum of powers associated with the
basis functions

The most general formula for the average power in a signal over a
periode of time (i.e. one that can deal with x(t) even when it is a complex
function of time) is

1 t2
P  x t x*  t  dt
t 2  t 1 ttheorem can be stated as
In mathematical terms Parseval’s 1

1 t2 
P  x t x*  t  dt   a a *

t 2  t1 t 1
m  
m m m


 a m
2
P m
m  
We can derive this, for a signal of period , as
2
 follows,
t t 1
 T with

remembering all the time that xby a m m  t 
 t  using
m  


 
*
1 t2 1 t2
P 
t 2  t1 t
x t x*  t  dt 

 x t  a m m
 
t 2 t 1 t 1 m
t dt
1


1 t2 
P  a m  x t   m  t  dt  a a 
* * *

m   t 2  t1 t 1
m  
m m m


 am m
2
P
Note we have used m   the linierity of the summation and integration

operation, and also the useful formula

   j   j       j
2 2
2
