UNIT-4 Spectral Representation and Estimations

Spectral Factorization and Innovations

A power spetrum that is also known as power spectral density in frequqncy domail can be expressed
in Z domain in factroziation terms .Actually, PSD is factorized in three terms here, that is the
disccreter factorization of PSD in discretet time random process.

Hence, 𝜎02 = 𝑣𝑎𝑟𝑖𝑎𝑛𝑐𝑒, 𝑄(𝑧) = 𝑖𝑛𝑛𝑜𝑣𝑎𝑡𝑖𝑜𝑛 𝑓𝑖𝑙𝑡𝑒𝑟

UUnit-4 {pg} 1
 Hints can be used to solve it

UUnit-4 {pg} 2
Spectral Representation of random process

Let us consider a WSS process (X(t)) passes through a LTI system, then output process y(t) also
remain WSS process.

Now , let us start the description of random process in term of linear functionals (V1,V2).Which is
the inner product of process. The process is Z of [ (Z(t)] with a deterministic.

Here,
Z(t) is a random process
.g1(t) and g2(t) are the deterministic
functions.So ,when a take the inner
product of random process with
deterministic process then we get
the random variable in term of linear
functional
Here g1(t) and g2(t) are the real function, so its conjugate will be also same.Now , if you remember a
receiver system that have the correlator . so what does this correlator do?

Let us consider noise Z(t) , which

is random process in nature is
passes through the Correlator
with signal g1(t). then it gives the
random variable in terms of
linear functional.

UUnit-4 {pg} 3
Now, lets us evaluate the expected value of V1 &V2.

Sine, g1(t) and g2(t) are the deterministic process , so expectation will be apply only on random part

Again , it is clear that expectation of two random process can be written in term of its auto-
correlation , there for,

Now, Z(t) is WSS random process in nature, hence a autocorrelation function based to two argument
( say, t and λ here) can be converted in to single argument form,. so,

Therefor, above equation can be wriiten as,

Here, u(t) can be also seen as the convolution of Rz(t) and g2(λ) [ u(t)=Rz(t)*g2(t)].

UUnit-4 {pg} 4
Now, by using Parseval theorem, it can be shown that inner product of two time domain signal
preserve same in frequency domain also, means

Here, x1(f) is the FT of X1(t) and so on.

Now, look what is conjugate of U(f).

Now, let us define a very useful property

Hence, depending upon above equation let realize that,

If G1(f) and G2(f) are non-overlapping in frequency, means both are in different band .then Expected
value of v1 and v2 becomes zero. Therefore, in other word it can be said that when G1(f) ang G2(f)
are uncorrelated then expected value of V1 and V2 becomes zero.

UUnit-4 {pg} 5
For WGN, E(V1,V2) will be

∞
∞
𝑁0
𝐸[𝑉1 𝑉2] = ∫ ∫ 𝑔1(𝑡) 𝛿(𝑡 − 𝜆)𝑔2(𝜆)𝑑𝜆 𝑑𝑡 − − − − − − − − − 1
−∞ 2
−∞

Now, it must be recorded that for all value of λ that is not equal to t is zero. So at λ=t , it reduce to

Single integration as,

∞
𝑁0
𝐸[𝑉1 𝑉2] = ∫ 𝑔1(𝑡) 𝑔2(𝑡) 𝑑𝑡 − − − − − − − − − 2
2
−∞
∞
𝑁0
𝐸[𝑉1 𝑉2] = ∫ 𝑔1(𝑡)𝑔2(𝑡) 𝑑𝑡 − − − − − − − − − 3
2
−∞

UUnit-4 {pg} 6
Now if we pass WGN with two Ortho-normal process, we will get two random variable v1, v2.so,
expected value of V1 and V2 will be zero.

ERGODICITY

A random process X(t) which is WSS also is said to be ergodic , if it satisfied two conditions
̿̿̿̿̿̿
1- Ensemble Average is equal to the time average 𝒖𝑿 = 𝑬(𝑿(𝒕) = 𝒙(𝒕)

∞
E{X(t)}= ∫−∞ 𝒙𝑭(𝒙)𝒅𝒙
𝟏 𝑻
̿̿̿̿̿̿ = 𝐥𝐢𝐦
𝒙(𝒕) ∫ 𝑿(𝒕)𝒅𝒕
𝑻→∞ 𝟐𝑻 −𝑻

2- Ensemble ACF is equal to the time ACF

𝟏 𝑻
.𝑹𝑿𝑿 (𝝉) = 𝐥𝐢𝐦 ∫ 𝑿(𝒕)𝑿(𝒕 + 𝝉)𝒅𝒕
𝑻→∞ 𝟐𝑻 −𝑻

Q-Consider a random process 𝒙(𝒕) = 𝑨𝑪𝒐𝒔(𝟐𝝅𝒇𝒄 𝒕 + ∅) .Where A and Fc are the constant
and 𝝋 is the random variable uniformly distributed in 0-2π.Check its ergodicity.

UUnit-4 {pg} 7
UUnit-4 {pg} 8
UUnit-4 {pg} 9