LECTURE-4: CORRELATION
FUNCTIONS
SALMAN GHAFOOR
Problem Scenario
A signal g(t) is transmitted and received after reflection from the
target – received signal is denoted as z(t)
1. How to know if the received signal is a reflection of transmitted
signal?
2. How to determine distance of the target?
SIGNALS VERSUS VECTORS
Signals versus Vectors
There is a strong connection between signals and vectors
Signals that are defined for only a finite number of time instants (say N)
can be written as vectors (of dimension N)
Thus, consider a signal g(t) defined over a closed time interval [a, b]
Let us pick N points uniformly on the time interval [a, b] such that:
Signals versus Vectors
Then we can write a signal vector g as an N-dimensional vector
As the number of time instants N increases, the sampled signal
vector g will grow
As 𝑁 → ∞, the signal values will form a vector g of infinitely long
dimension
Because 𝜖 → 0, the signal vector g will transform into the
continuous-time signal g(t) defined over the interval [a, b]:
Signals versus Vectors
This implies, continuous time signals are straightforward generalizations
of finite dimension vectors
Thus, basic definitions and operations in a vector space can be applied
to continuous time signals as well
We shall denote all vectors by boldface type, for example, x is a certain
vector with magnitude or length 𝐱
In a vector space, we can define the inner (dot or scalar) product of two
real-valued vectors g and x and norm 𝐱 of x as:
𝜃 is the angle between vectors g and x
COMPONENT OF A VECTOR ALONG
ANOTHER VECTOR
Projection of Vectors
Consider two vectors g and x
Let the component of g along x be cx
Geometrically, the component of g along x is the projection of g
on x and is obtained by drawing a perpendicular from the tip of g
on the vector x
Projection of Vectors
What is the mathematical significance of a component of a vector
along another vector?
The vector g can be expressed in terms of vector x as:
However, this does not describe a unique way to decompose g in
terms of x and e, two of the infinite other possibilities are:
Projection of Vectors
The question is: Which is the "best" decomposition?
The approximation in the first figure is unique because its error
vector is the shortest (with the smallest magnitude or norm)
We can now define mathematically the component (or projection)
of a vector g along vector x to be cx, where c is chosen to
minimize the magnitude of the error vector 𝐞 = 𝐠 − 𝑐𝐱
Geometrically, the magnitude of the component of g along x is
𝐠 𝑐𝑜𝑠𝜃, which is also equal to 𝑐 𝐱 , therefore:
Projection of Vectors
Based on the definition of inner product between two vectors,
multiplying both sides by 𝐱 yields:
And
It is apparent that when g and x are perpendicular, or orthogonal,
then g has a zero component along x; consequently, c = 0
We therefore define g and x to be orthogonal if the inner (scalar
or dot) product of the two vectors is zero, that is, if:
ORTHOGNALITY OF CONTINUOUS
TIME SIGNALS
Orthogonality of CT Signals
The concepts of vector component and orthogonality can be directly
extended to continuous time signals
Consider the problem of approximating a real signal g(t) in terms of
another real signal x(t) over an interval [t1, t2]:
The error e(t) in this approximation is:
For "best approximation," we need to minimize the error signal, that is,
minimize its norm
Orthogonality of CT Signals
Minimum signal norm corresponds to minimum energy Ee over the
interval [t1, t2] given by:
To minimize Ee, a necessary condition is:
Or:
Orthogonality of CT Signals
Expanding the squared term inside the integral, we obtain:
Or:
So, if a signal g(t) is approximated by another signal x(t) as
𝑔 𝑡 ≈ 𝑐𝑥 𝑡 , then the optimum value of c that minimizes the
energy of the error signal in this approximation is given by:
Orthogonality of CT Signals
Similarly, we say that if the component of a signal g(t) along x(t) is zero
(i.e., c = 0), the signals g(t) and x(t) are orthogonal over the interval
[t1, t2]:
This integral is identical to the standard definition of the inner product
of two N-dimensional vectors g and x
Therefore, the inner product of two (real-valued) signals g(t) and x(t)
defined over a time interval [t1, t2] may be written as:
Orthogonality of CT Signals
From algebraic geometry: the square of a vector length 𝐱 2 is equal to
< 𝐱, 𝐱 >
Continuing our analogy with vector analysis, we define the norm of a
signal g(t) as:
This is the square root of the signal energy in the time interval
Therefore, the norm of a signal is analogous to the length of a finite
dimensional vector
More generally, signals may not be merely defined over a continuous
segment [t1, t2]
Orthogonality of CT Signals
For complex signals that are a function of t, the optimum
coefficient becomes:
Complex functions (signals) x1(t) and x2(t) are orthogonal over an
interval (𝑡1 ≤ 𝑡 ≤ 𝑡2 ) as long as:
Or:
CORRELATION OF SIGNALS
Correlation Coefficient
By defining the inner product and the norm of signals, we paved
the foundation for signal comparison
We could consider c to be a quantitative measure of similarity
between g and x
Such a measure, however, would be defective because it varies
with the norms (or lengths) of g and x
Therefore, similarity between two vectors is indicated by the
angle 𝜃 between the vectors - smaller the 𝜃, the larger the
similarity, and vice versa
Correlation Coefficient
Amount of similarity can therefore be conveniently measured by 𝑐𝑜𝑠𝜃
The larger the 𝑐𝑜𝑠𝜃, the larger the similarity between the two vectors
Thus, a suitable measure would be 𝜌 = 𝑐𝑜𝑠𝜃, also called correlation
coefficient and is given by:
Observe that:
Two vectors aligned in opposite directions have maximum dissimilarity
(𝜌 = −1)
Correlation Coefficient
We use the same argument in defining a similarity index (the
correlation coefficient) for signals
To establish a similarity index independent of energies (sizes) of
g(t) and x(t), we must normalize c by normalizing the two signals
to have unit energies
Mathematically (out of scope), one can show that the magnitude
of 𝜌 is never greater than 1:
Correlation Functions
Let us revisit the application of correlation to signal detection in a radar
unit, where a signal pulse is transmitted to detect a suspected target
By detecting the presence or absence of the reflected pulse, we confirm
the presence or absence of the target
By measuring the time delay between the transmitted and received
(reflected) pulse, we determine the distance of the target
Let the transmitted and the reflected pulses be denoted by g(t) and
z(t), respectively
Correlation Functions
We can directly to measure the correlation coefficient 𝜌 as:
The correlation is zero because the pulses are disjoint (nonoverlapping
in time)
To avoid this difficulty, we compare the received pulse z(t) with the
transmitted pulse g(t) shifted by 𝜏
If for some value of 𝜏, there is a strong correlation, we not only detect
the presence of the pulse, but we also detect the relative time shift of
z(t) with respect to g(t)
Correlation Functions
We use the modified integral 𝜓𝑔𝑧 (𝜏), the cross-correlation function of
two complex signals g(t) and z(t), defined by:
Correlation of a signal with itself is called the autocorrelation, defined
as:
SUMMARY
Signals and vectors
Component of a vector along another vector
Orthogonality of CT signals
Cross-correlation function
Autocorrelation function
