Gaussian Random Vectors and Processes

Saravanan Vijayakumaran
[email protected]

Department of Electrical Engineering

Indian Institute of Technology Bombay

September 19, 2013

1 / 14
Gaussian Random Vectors
 Jointly Gaussian Random Variables
Definition (Jointly Gaussian RVs)
Random variables X1 , X2 , . . . , Xn are jointly Gaussian if any non-trivial linear
combination is a Gaussian random variable.

a1 X1 + · · · + an Xn is Gaussian for all (a1 , . . . , an ) ∈ Rn \ 0

Example (Not Jointly Gaussian)

X ∼ N(0, 1) 
X, if |X | > 1
Y =
−X , if |X | ≤ 1
Y ∼ N(0, 1) and X + Y is not Gaussian.

3 / 14
 Gaussian Random Vector
Definition (Gaussian Random Vector)
A random vector X = (X1 , . . . , Xn )T whose components are jointly Gaussian.

Notation
X ∼ N(m, C) where

m = E[X] is the n × 1 mean vector

h i
C = E (X − m)(X − m)T is the n × n covariance matrix

mi = E[Xi ], Cij = E[(Xi − mi )(Xj − mj )] = cov(Xi , Xj )

Definition (Joint Gaussian Density)

For a Gaussian random vector, C is invertible and the joint density is given by
 
1 1
p(x) = p exp − (x − m)T C−1 (x − m)
(2π)n det(C) 2

For derivation, see Problems 3.31(f) and 3.32 in Madhow’s book.

4 / 14
Uncorrelated Jointly Gaussian RVs are Independent
If X1 , . . . , Xn are jointly Gaussian and pairwise uncorrelated, then they are
independent. For pairwise uncorrelated random variables,

0 if i 6= j
Cij = E[(Xi − mi )(Xj − mj )] =
σi2 otherwise.

The joint probability density function is given by

 
1 1
p(x) = p exp − (x − m)T C−1 (x − m)
(2π)n det(C) 2
n
(xi − mi )2
 
Y 1
= exp −
2σi2
q
i=1 2πσi2

where mi = E[Xi ] and σi2 = var(Xi ).

5 / 14
Uncorrelated Gaussian RVs may not be Independent
Example
• X ∼ N(0, 1)
• W is equally likely to be +1 or -1
• W is independent of X
• Y = WX
• Y ∼ N(0, 1)
• X and Y are uncorrelated
• X and Y are not independent

6 / 14
Gaussian Random Processes
 Gaussian Random Process
Definition
A random process X (t) is Gaussian if its samples X (t1 ), . . . , X (tn ) are jointly
Gaussian for any n ∈ N and distinct sample locations t1 , t2 , . . . , tn .
 T
Let X = X (t1 ) · · · X (tn ) be the vector of samples. The joint density is
given by
 
1 1
p(x) = p exp − (x − m)T C−1 (x − m)
(2π)n det(C) 2

where
h i
m = E[X], C = E (X − m)(X − m)T

8 / 14
 Properties of Gaussian Random Process
• The mean and autocorrelation functions completely characterize a
Gaussian random process.
• Wide-sense stationary Gaussian processes are strictly stationary.
• If the input to a stable linear filter is a Gaussian random process, the
output is also a Gaussian random process.

X (t) h(t) Y (t)

9 / 14
 White Gaussian Noise
Definition
A zero mean WSS Gaussian random process with power spectral density
N0
Sn (f ) = .
2
N0
2
is termed the two-sided PSD and has units Watts per Hertz.

Remarks
• Autocorrelation function Rn (τ ) = N20 δ(τ )
• Infinite Power! Ideal model of Gaussian noise occupying more
bandwidth than the signals of interest.

10 / 14
 White Gaussian Noise through Correlators
• Consider the output of a correlator with WGN input
Z ∞
Z = n(t)u(t) dt = hn, ui
−∞

where u(t) is a deterministic finite-energy signal

• Z is a Gaussian random variable
• The mean of Z is
Z ∞
E[Z ] = E [n(t)] u(t) dt = 0
−∞

• The variance of Z is
h i Z Z 
var(Z ) = E (hn, ui)2 = E n(t)u(t) dt n(s)u(s) ds
Z Z
= u(t)u(s)E [n(t)n(s)] dt ds
Z Z
N0
= u(t)u(s) δ(t − s) dt ds
2
Z
N0 2 N0
= u (t) dt = kuk2
2 2
11 / 14
 White Gaussian Noise through Correlators
Proposition
Let u1 (t) and u2 (t) be linearly independent finite-energy signals and let n(t)
be WGN with PSD Sn (f ) = N20 . Then hn, u1 i and hn, u2 i are jointly Gaussian
with covariance
N0
cov (hn, u1 i, hn, u2 i) = hu1 , u2 i.
2

Proof
To prove that hn, u1 i and hn, u2 i are jointly Gaussian, consider a non-trivial
linear combination ahn, u1 i + bhn, u2 i
Z
ahn, u1 i + bhn, u2 i = n(t) [au1 (t) + bu2 (t)] dt.

This is the result of passing n(t) through a correlator. So it is a Gaussian

random variable.

12 / 14
 White Gaussian Noise through Correlators
Proof (continued)

cov (hn, u1 i, hn, u2 i) = E [hn, u1 ihn, u2 i]

Z Z 
= E n(t)u1 (t) dt n(s)u2 (s) ds
Z Z
= u1 (t)u2 (s)E [n(t)n(s)] dt ds
Z Z
N0
= u1 (t)u2 (s) δ(t − s) dt ds
2
Z
N0
= u1 (t)u2 (t) dt
2
N0
= hu1 , u2 i
2

If u1 (t) and u2 (t) are orthogonal, hn, u1 i and hn, u2 i are independent.

13 / 14
Thanks for your attention

14 / 14