Transcript Signal Spaces

Signal Spaces
Much can be done with signals by drawing an
analogy between signals and vectors.
A vector is a quantity with magnitude and
direction.
magnitude
q
(NE Direction)
There are two fundamental operations associated
with vectors: scalar multiplication and vector
addition.
Scalar multiplication is scaling the magnitude of a
vector by a value.
a
2a
Vector addition is accomplished by placing the tail
of one vector to the tip of another vector.
a
b
b
a
a+b
A vector can be described by a magnitude and an
angle, or it can be described in terms of coordinates.
Rather than use x-y coordinates we can describe the
coordinates using unit vectors.
The unit vector for the “x” coordinate is i. The unit
vector for the “y” coordinate is j.
Thus we can describe vector a as
(4,3)
a
j
i
We can also describe a as 4i + 3j.
Suppose we had a second vector b = 4i + j.
The sum of the vectors a and b could be described
easily in terms of unit vectors:
a + b = 8i + 4j.
In general, if a = ax i + ay j and b = bx i + by j , we
have
a + b = (ax+bx )i + (ay+by )j .
In other words, the x-component of the sum is the
sum of the x-components of the terms, and the ycomponent of the sum is the sum of the ycomponents of the terms.
At this point we draw an analogy from vectors to
signals.
Let a(t) and b(t) be sampled functions
a(t)
b(t)
When we add two functions together, we add their
respective samples together as we would add the xcomponents, y-components and other components
together.
a(t)
b(t)
a(t) + b(t)
We can think of the different sample times as
different dimensions.
In MATLAB, we could create two vectors (onedimensional matrices), and add them together:
>> a = [3 4 1 2];
>> b = [2 3 4 2];
>> a + b
ans =
5
7
5
4
You can think of the four values in each vector as,
say, w-components, x-components, y-components
and z-components. We can add additional
components as well.
We will now examine another vector operation and
show an analogous operation to signals.
This operation is the dot product.
Given two vectors, a and b, the dot product of the
two vectors is defined to be the product of their
magnitudes times the cosine of the angle between
them:
a•b  |a| |b| cosqab.
If the two vectors are in the same direction, the dot
product is merely the ordinary product of the
magnitudes.
b
a
a•b = |a| |b|.
If the two vectors are perpendicular, then the dot
product is zero.
a
b
a•b = 0.
The dot product of the unit vector i with itself is one.
So is the dot product of the unit vector j with itself.
i•i = 1.
j•j = 1.
The dot product of the unit vector i the unit vector j
is zero.
i•j = 0.
Suppose a = ax i + ay j and b = bx i + by j , Their dot
product is
a • b = (ax i + ay j )• (bxi + by j ) .
Using the dot products of unit vectors from the
previous slide, we have
a • b = axbx + ayby .
As with vector addition, we can draw an analogy for
the dot product to signals.
Let a(t) and b(t) be sampled functions (as before).
We define the inner product of the two signals to be
the sum of the products of the samples from a(t)
and b(t).
The notation for the inner product between two
signals a(t) and b(t) is
a(t ), b(t ) .
The inner product is a generalization of the dot
product.
If we had, say, four sample times, t1, t2, t3, t4, the
inner product would be
a(t ), b(t )  a(t1 )b(t1 )  a(t2 )b(t2 )  a(t3 )b(t3 )  a(t4 )b(t4 ).
Let us take the inner product of our previous
sampled signals a(t) and b(t):
a(t)
4
3
2
1
4
b(t)
3
2
2
a(t ), b(t )  (3)( 2)  (4)(3)  (1)( 4)  (2)( 2)
 26.
In MATLAB, we would take the inner product as
follows:
>> a = [3 4 1 2];
>> b = [2 3 4 2];
>> a * b’
ans =
26
In general, the inner product of a pair of sampled
signals would be
a(t ), b(t )   a(t n )b(t n ).
n 1
N
Now, what happens as the time between samples
decreases and the number of samples increases?
Eventually, we approach the inner product of a pair
of continuous signals.
a(t ), b(t )


0
a(t )b(t ) dt.
T
Again, the inner product can be thought of as the
sum of products of two signals.
Example: Find the inner product of the following two
functions:
t (0  t  1),
a (t )  
0 (elsewhere ).

2

t
b(t )  
0
(0  t  1),
(elsewhere).
Solution:
0
4
4
0
  t 3 dt 
 .
t4
1
1

 t t dt
  a(t )b(t ) dt
1
0
a(t ), b(t )
1
0
T
2
Example: Find the inner product of the following two
functions:
1 (0  t  1),
a (t )  
0 (elsewhere ).

2t  1 (0  t  1),
b(t )  
0
(elsewhere ).

 t 2  t  0.
0


 2t dt  1 dt
  12t  1 dt
1
a(t ), b(t )
0
1
0
0
1
0
1
T
When the inner product of two signals is equal to
zero, we say that the two signals are orthogonal.
When two vectors are perpendicular, their dot
product is zero. When two signals are orthogonal,
their inner product is zero.
Just as the inner product is a generalization of the
dot product, we generalize the idea of two vectors
being perpendicular if their dot product is zero to the
idea of two signals being orthogonal if their inner
product is zero.
Example: Find the inner product of the following two
functions:
sin ct
a(t )  
0
(0  t  T ),
(elsewhere).
cos ct (0  t  T ),
b(t )  
0
(elsewhere).

Let T be an integral multiple of periods of sin ct or
cos ct.
 12 (0)  0.
a(t ), b(t )


0 2
1
T
0
sin 2 c tdt
sin  c t cos  c t dt
T
The functions sine and cosine are orthogonal to
each other.
Example: Find the inner product of
cos ct
a(t )  
0

(0  t  T ),
(elsewhere)
with itself. Again, let T be an integral multiple of
periods of sin ct or cos ct.

 12 (0)  T2 .
2 2
1 T
0 2

c t  dt
1

cos
2


1
  cos 2  c t dt
T
0
a (t ), a (t )   cos  c t cos  c t dt
T
0
T
The inner product of a signal with itself is equal to its
energy.
The dot product of a signal with itself is equal to its
magnitude-squared (exercise).
Exercise: Find the inner product of a(t) with itself,
b(t) with itself and a(t) with b(t), where
 T2 sin ct (0  t  T ),
a(t )  
(elsewhere).
0
 T2 cosct (0  t  T ),
b(t )  
(elsewhere).
0
As before, let T be an integral multiple of periods of
sin ct or cos ct.
Now, back to ordinary vectors. One of the most
famous theorems in vectors is something called the
Cauchy-Schwarz inequality. It shows how dot
products of two vectors compare with their
magnitudes. It also applies to inner products.
Let us introduce a scalar g. Using this scalar along
with our two vectors a and b, let us take the inner
product of a+ gb with itself.
a  gb, a  gb  a, a  2g a, b  g 2 b, b .
(We have exploited some properties of the inner
product which should not be too hard to verify,
namely distributivity and scalar multiplication.)
The expression on the right-hand side of this
equation is a quadratic in g. If we were to graph this
expression versus g, we would get a parabola. The
graph would take one of the following three forms:
g
Two Roots
g
One Root
g
No (Real) Roots
We know, however, that since this expression is
equal to the inner product of something with itself
<a+ gb a+ gb>, that the expression must be greater
than or equal to zero. Thus only the last two
graphs pertain to this expression.
a, a  2g a, b  g 2 b, b  0.
If this is true, then the quadratic expression must
have at most one root.
If there is at most one root, then the discriminant of
the quadratic must be negative or zero:
2 a, b 
2
 4 b, b a, a  0.
Simplifying, we have
a, b  a, a b, b .
2
Thus we have the statement of the CauchySchwarz inequality.
This expression is a non-strict inequality. In some
cases, we have equality.
Suppose a and b are orthogonal (qab = 90°).
a, b  0.
In this case, the Cauchy-Schwarz inequality is met
easily (zero is less than or equal to anything
positive).
Suppose a and b are in the same direction (qab = 0°).
a, b  a b .
In this case, the Cauchy-Schwarz inequality is an
equality: the upper-bound on <a,b> is met.
Thus, the maximum value of <a,b> is achieved
when a and b are collinear (in the same direction).
The Cauchy-Schwarz inequality as an upper bound
on <a,b> is the basis for digital demodulation. If we
wished to detect a signal a by taking its inner
product with some signal b, the optimal value of b is
some scalar multiple of a.
a
<a,b>
Detector
If we use the inner product of signals, the inner
product detector becomes
a(t)
X
b(t)

<a,b>
So the optimal digital detector is simply an
application of the Cauchy-Schwarz inequality.
The optimal “local oscillator” signal b(t) is simply
whatever signal that we wish to detect.
Using our previous notation a(t) is equal to s(t) if
there is no noise, or r(t)=s(t)+n(t) if there is noise.
The “local oscillator signal” b(t) is simply s(t) [we do
not wish to detect the noise component].
r(t)
X

s
s(t)
The resultant filter is called a matched filter. We
“match” the signal that we wish to detect s(t) with a
“local oscillator” signal s(t).
Another way to think of the inner product operation
or matched filter operation is as a vector
projection.
Suppose we have two vectors a and b.
a
b
The projection of a onto b is the “shadow” a casts on
b from “sunlight” directly above.
a



projection
b
The magnitude of the projection is magnitude of a
times the cosine of the angle between a and b.
projection a cosq ab .
The projection can be defined in terms of the inner
product:
projection
a, b
b
.
The actual projection itself is a vector in the
direction of b. To form this vector, we multiply the
magnitude by a unit vector in the direction of b.
b
unit vector  .
b
a, b b
projection
.
b b
The denominator |b||b| can be expressed as the
magnitude squared of b or the inner product of b
with itself.
projection
a, b
b, b
b.
If the magnitude of b is unity, the projection
becomes
projection a, b b.
The signal b has unity magnitude in the following
matched filter:
s(t)
X
ds(t)

s
(2/T) cos ct
This was the detector with the “normalized” local
oscillator.
Let us do an example of projections with signals.
Example: “Project” t2 onto t. Restrict the interval of
the signals to [0,1].
1
projection
a, b
b, b

b
t
0
1
0
t 3 dt
t.
2
dt
Evaluating the integrals, we have
1

projection
t
0
1
0
3
t dt
2
dt
t
1
4
1
3
t  34 t.
Let us plot the original function t2 and its projection
onto t.
t2 and Its Projection onto t
1
0.9
t
0.8
2
t2, (3/4)t
0.7
0.6
0.5
(3/4)t
0.4
0.3
0.2
0.1
0
0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
t
1
The projection (3/4)t can be thought of as the best
linear approximation to t2 over the interval [0,1]. (It is
the best linear approximation in the minimum, meansquared error sense.)
When we project a vector onto another vector, we
take the best approximation of the first vector in the
direction of the second vector.
a



projection
b
We can also project onto multiple vectors:
c


projection 


a



projection
b
If we were to add the two projections (onto b and c),
we would no longer have an approximation to a, but
rather we would have exactly a.
Example: Let us project cos(ct + q) onto cos(ct)
and sin(ct) .
1
cos( t  q ) cos t dt

cosineprojection
cos t.
 cos  t dt
c
0
c
1
c
2
c
0
1
cos( t  q ) sin  t dt

sine projection
sin  t.
 sin  t dt
c
0
1
0
c
c
2
c
Evaluating the integrals, we have
cos(2 t  q )  cosq dt

cosine projection
cos t
 cos2 t  1dt
cosq  dt

cos t
1 dt
1
1
0 2
c
c
1
1
0 2
c
1
1
2
0
1
2
1
0
 cosq cos c t .
c
1
cos( t  q ) sin  t dt

sine projection
sin  t
 sin  t dt
sin(2 t  q )  sin q dt


sin  t
 1  cos2 t dt
 sin q   dt

sin  t
1  dt
c
0
1
c
c
2
c
0
1
1
0 2
c
c
1
1
0 2
c
1
1
2
0
1
2
1
0
  sin q sin  c t .
c
Adding the two projections, we have
cosq cosct  sin q sin ct .
This should not be surprising because
cos(ct  q )  cosct cosq  sin ct sin q .
Example: Let us project an arbitrary function x(t)
onto cos(ct), cos(2ct), cos(3ct), … Restrict the
interval of the signals to [0,T], where T is an integral
multiple of cycles of cos(ct).
We will be projecting x(t) onto cos(nct), for
n=1,2,3, … These projections will then be summed.

x(t )   projectionontocos(nct ).
n 1

projectionontocos(n t ) 
T
x(t ) cos n c t dt
0
c

T
0


T
0
cos2 n c t dt
cos n c t
x(t ) cos n c t dt
 1  cos2n t dt
T
1
0 2


T
0
cos n c t
c
x(t ) cos n c t dt
1
2
T 
cos n c t
T

2
 T  x(t ) cos n c t dt cos n c t
 0

So the summation of the projections is

x(t )    T2  x(t ) cosnct dt cosnct.


0


n 1
Or,
T

x(t )   an cos n ct ,
n 1
where
T
an  T2  x(t ) cos nct dt.
0
We have just derived the trigonometric Fourier
series (for cosine).
Exercise: Project an arbitrary function x(t) onto
sin(nct), n=1,2,3, … (Complete the trigonometric
Fourier series.)
Exercise: Project an arbitrary function x(t) onto
ejnct, for n=0,±1,±2, … (Derive the complex
exponential Fourier series.)
The previous examples an exercises worked
because we were projecting onto orthogonal
signals. If the signals were not orthogonal, we could
not simply sum the projections and expect to
reconstruct the original signal.