Sunteți pe pagina 1din 35

ECNG 6700 - Stochastic Processes, Detection and Estimation

Random Processes - Part I Sean Rocke

September 26th & October 3rd , 2013

ECNG 6700 - Stochastic Processes, Detection and Estimation

1 / 35

Outline

Probability & Random Variables Random Vectors Random (Stochastic) Processes Stationarity Power Spectral Densities for Real WSS Processes Conclusion

ECNG 6700 - Stochastic Processes, Detection and Estimation

2 / 35

Probability & Random Variables

Bounds and Approximations


In many applications where we calculate probabilities we can potentially run into two problems:
1

We do not know the underlying distributions completely . . . all we have are sample moments such as E [X ], var (X ) and higherorder moments E [(X )k ], k > 2. We know the distributions, but integration in closed form is not possible (e.g., Gaussian pdf)

Question: How do we solve this? We use approximation techniques to establish upper and\or lower bounds on probabilities. . .

ECNG 6700 - Stochastic Processes, Detection and Estimation

3 / 35

Probability & Random Variables

Bounds and Approximations

AtHome Activity: Read up on the following inequalities\bounds: Markov inequality Tchebycheff (Chebyshev) inequality Chernoff inequality Strong Law of Large Numbers (SLLN) Weak Law of Large Numbers (WLLN) Central Limit Theorem (CLT)

ECNG 6700 - Stochastic Processes, Detection and Estimation

4 / 35

Probability & Random Variables

Bounds and Approximations


Food for thought:
1 2 3 4 5

When would you use one approximation technique over another? Which gives a tighter bound: Markov, Tchebycheff or Chernoff? What is the difference between the SLLN and WLLN? Express the CLT in your own words. SLLN, WLLN, and CLT assume IID RVs. What if the RVS are dependent or not identical? Can you recognize when to use the approximation techniques if given a problem to solve? Can you apply the appropriate technique(s) when you recognize that one is necesary?

ECNG 6700 - Stochastic Processes, Detection and Estimation

5 / 35

Random Vectors

Linear Algebra Review

AtHome Activity: Review the following matrix operations:


1 2 3 4 5 6 7

Transpose Matrix sum Matrix product Trace of a matrix Norm of a vector Vector inner & outer products Block matrices & operations

ECNG 6700 - Stochastic Processes, Detection and Estimation

6 / 35

Random Vectors

Random Vectors & Matrices


Random Vector: A vector whose entries are RVs Random Matrix: A matrix whose entries are RVs Important Parameters: Expectation of a vector\matrix Correlation matrix for a random vector Covariance matrix for a random vector Crosscorrelation matrix for two random vectors Crosscovariance matrix for two random vectors

ECNG 6700 - Stochastic Processes, Detection and Estimation

7 / 35

Random Vectors

Random Vectors & Matrices

Example 1: Write out the correlation, covariance, crosscorrelation & crosscovariance matrices for the ndimensional random vectors = [X1 , . . . , Xn ]T and Y = [Y1 , . . . , Yn ]T . X

ECNG 6700 - Stochastic Processes, Detection and Estimation

8 / 35

Random (Stochastic) Processes

Modelling Uncertainty in Random Signals


A random signal has some element of uncertainty, and thus we can never determine its exact value at any given time. Can describe signal probabilistically (e.g., in terms of average properties, or probability that signal exceeds a given value) Random Process: The probabilistic model used to describe such random signals

Conceptual Denition of Random (Stochastic) Processes: Mathematical model of an empirical process whose development is governed by probability laws . . . Lets look at some examples. . .

ECNG 6700 - Stochastic Processes, Detection and Estimation

9 / 35

Random (Stochastic) Processes

RP Example: Modelling Temperature Anomalies

ECNG 6700 - Stochastic Processes, Detection and Estimation

10 / 35

Random (Stochastic) Processes

RP Example: Modelling Global Precipitation

ECNG 6700 - Stochastic Processes, Detection and Estimation

11 / 35

Random (Stochastic) Processes

RP Example: Modelling Stock Prices

ECNG 6700 - Stochastic Processes, Detection and Estimation

12 / 35

Random (Stochastic) Processes

RP Example: Modelling Annual Rainfall

ECNG 6700 - Stochastic Processes, Detection and Estimation

13 / 35

Random (Stochastic) Processes

RP Example: Modelling Network Trafc

ECNG 6700 - Stochastic Processes, Detection and Estimation

14 / 35

Random (Stochastic) Processes

RP Example: Modelling Network Trafc

ECNG 6700 - Stochastic Processes, Detection and Estimation

15 / 35

Random (Stochastic) Processes

So formally, what is a Random Process?


Denition: Family of RVs, {X (t ), t T } dened on a given probability space, S , indexed by the parameter, t , where t varies over the index set, T Function of two arguments, {X (t , ), t T , S} Questions:
1 2 3 4

For each xed t = tk , what is X (tk , )? For a xed sample point, = i , what is X (t , i )? For xed t = tk & = i , what is X (tk , i )? Be sure that you know the denitions of the following terms for random processes: ensemble, member function, sample function, realization
ECNG 6700 - Stochastic Processes, Detection and Estimation 16 / 35

Random (Stochastic) Processes

Random Processes: Classication


Time Discrete Continuous

Continuous

Continuous Random Sequence

Continuous Random Process

Value Discrete

Discrete Random Sequence

Discrete Random Process

ECNG 6700 - Stochastic Processes, Detection and Estimation

17 / 35

Random (Stochastic) Processes

Random Processes: Classication

Stationarity: Stationary (Strictly, Wide Sense) Cyclostationary Nonstationary Real vs Complexvalued: Realvalued bandpass RP - Z (t ) = A(t )cos[2 fc t + (t )] Z (t ) = {A(t )e
j (t ) ej 2 fc t }

= {W (t )e j 2fc t }

Complex envelope, W (t ) = A(t )cos(t ) + jA(t )sin(t ) = X (t ) + jY (t )

ECNG 6700 - Stochastic Processes, Detection and Estimation

18 / 35

Random (Stochastic) Processes

Random Processes: Methods of Description


Joint Distribution: FX (t1 ),...,X (tn ) (x1 , . . . , xn ) = P (X (t1 ) x1 , . . . , X (tn ) xn ) Analytical Description using RVs: Realvalued bandpass RP - Z (t ) = A(t )cos[2 fc t + (t )] Average Values: Mean - X (t ) = E [X (t )] Autocorrelation - RXX (t1 , t2 ) = E [X (t1 )X (t2 )] Autocovariance - CXX (t1 , t2 ) = RXX (t1 , t2 ) X (t1 )X (t2 ) Correlation coefcient - rXX (t1 , t2 ) =
CXX (t1 ,t2 ) CXX (t1 ,t1 )CXX (t2 ,t2 )

ECNG 6700 - Stochastic Processes, Detection and Estimation

19 / 35

Random (Stochastic) Processes

Random Processes: 2 or more RPs


Joint Distribution: FX (t1 ),...,X (tn ),Y (t 1 ),...,Y (t m ) (x1 , . . . , xn , y1 , . . . , ym ) = P (X (t1 ) x1 , . . . , X (tn ) xn , Y (t 1 ) y1 , . . . , Y (t m ) ym ) Analytical Description using RVs: Realvalued bandpass RP - Z (t ) = A(t )cos[2 fc t + (t )] Average Values: Autocorrelation - RXY (t1 , t2 ) = E [X (t1 )Y (t2 )] Autocovariance - CXY (t1 , t2 ) = RXY (t1 , t2 ) X (t1 )Y (t2 ) Correlation coefcient - rXY (t1 , t2 ) =
CXY (t1 ,t2 ) CXX (t1 ,t1 )CYY (t2 ,t2 )

ECNG 6700 - Stochastic Processes, Detection and Estimation

20 / 35

Random (Stochastic) Processes

Random Processes: 2 or more RPs


Properties: For 2 uncorrelated RPs, CXY (t1 , t2 ) = 0, t1 , t2 T For 2 orthogonal RPs, RXY (t1 , t2 ) = 0, t1 , t2 T Independent P [X (t1 ) x1 , . . . , X (tn ) xn , Y (t 1 ) y1 , . . . , Y (t m ) ym ] =
n i =1 {P [X (ti )

xi ]}

m j =1 {P [Y (t j )

yj ]}

for all n, m & t1 , . . . , tn , t 1 , . . . , t m T


ECNG 6700 - Stochastic Processes, Detection and Estimation 21 / 35

Random (Stochastic) Processes

Random Processes: Worked Examples


Example 2: Lets use MATLAB to dene and investigate a Bernoulli (p) random process. . . Example 3: Consider the amplier of a radio receiver. Because all ampliers internally generate thermal noise, even if the radio is not receiving any signal, the voltage at the output of the amplier is not zero but is well modeled as a Gaussian random variable each time it is measured. Suppose we measure this voltage once per second and denote the nth measurement by Zn . Assume that the amplier gain is 5 and an input signal, x (t ) = sin(2 ft ), is applied. Lets investigate this further in MATLAB . . .
ECNG 6700 - Stochastic Processes, Detection and Estimation 22 / 35

Random (Stochastic) Processes

Special Random Processes

Markov Processes Gaussian (Normal) Processes Independent increments (e.g., Wiener Process, Poisson Process) You should really just get a better idea of the key characteristics of these processes!!!

ECNG 6700 - Stochastic Processes, Detection and Estimation

23 / 35

Stationarity

Strictsense Stationarity (SSS)


Denition:
1

The distribution function describing the process is invariant under a translation of time/space For all t1 , . . . , tk , t1 + , . . . , tk + T & all k = 1, 2, . . ., P [X (t1 ) x1 , . . . , X (tk ) xk ] = P [X (t1 + ) x1 , . . . , X (tk + ) xk ] If stationary for k N but not k > N then X (t ) is a N th order stationary process

SSS Properties: Constant mean over index: E [X (t )] = X = constant Autocorrelation only depends on time difference, not actual time samples: E [X (t1 )X (t2 )] = RXX (t2 t1 ) Jointly SSS: Joint distributions of X (t ) and Y (t ) are invariant to shifts in time
ECNG 6700 - Stochastic Processes, Detection and Estimation 24 / 35

Stationarity

Widesense Stationarity (WSS)


WSS: Less restrictive form of stationarity than SSS Only based upon the mean and autocorrelation functions, X (t ) and RXX (t1 , t2 ). WSS Properties: Constant mean over index: E [X (t )] = X = constant Autocorrelation only depends on time difference, not actual time samples: E [X (t1 )X (t2 )] = RXX (t2 t1 ) or alternatively, E [X (t )X (t + )] = RXX ( ) Jointly WSS: E [X (t )Y (t + )] = RXY ( ) SSS implies WSS, but WSS does not imply SSS
ECNG 6700 - Stochastic Processes, Detection and Estimation 25 / 35

Stationarity

Other Forms of Stationarity


Asymptotical stationarity: Distribution of X (t1 + ),. . . , X (tn + ) does not depend on when is large Stationary on an interval: X (t ) is SSS for all for which t1 + , . . . , tk + lie in an interval T Stationary increments: Increments X (t + ) X (t ) form a stationary process for every Cyclostationarity/Periodic stationarity: X (t ) is stationary for a constant shift T0 , or integer multiples of T0 Food for Thought: When a complete description of X (t ) is unavailable, but some process samples are available, how would you test for stationarity?
ECNG 6700 - Stochastic Processes, Detection and Estimation 26 / 35

Power Spectral Densities for Real WSS Processes

Autocorrelation Function (ACF) of a WSS Process

Properties:
1 2 3 4 5 6

Average power, RXX (0) = E [X 2 (t )] 0 RXX ( ) is an even function of : RXX ( ) = RXX ( ) |RXX ( )| RXX (0) If X (t ) contains a periodic component, then RXX ( ) will also If lim RXX ( ) = C , then C = 2 X If RXX (T0 ) = RXX (0) for some T0 = 0, then RXX ( ) is periodic with period T0 If RXX (0) < and RXX ( ) is continuous at = 0, then RXX ( ) is continuous for every

ECNG 6700 - Stochastic Processes, Detection and Estimation

27 / 35

Power Spectral Densities for Real WSS Processes

Crosscorrelation Function of a WSS Process

Properties:
1 2 3 4 5

RXY ( ) = RYX ( ) |RXY ( )| |RXY ( )| RXX (0)RYY (0) 1 2 [RXX (0) + RYY (0)]

RXY ( ) = 0 if X (t ) and Y (t ) are orthogonal RXY ( ) = X Y if X (t ) and Y (t ) are independent

ECNG 6700 - Stochastic Processes, Detection and Estimation

28 / 35

Power Spectral Densities for Real WSS Processes

Power Spectral Density (PSD) of a WSS Process


Denition (Wiener-Khinchine Relation): SXX (f ) = F{RXX ( )} =
j 2 f d RXX ( )e

F{} is the Fourier transform operator Autocorrelation can be retrieved using inverse transform on PSD RXX ( ) = F 1 {SXX (f )} = SXX (f )e j 2f df Properties:
1 2 3

SXX (f ) is real & nonnegative Average power in X (t ), E [X 2 (t )] = RXX (0) =


SXX (f )df

For X (t ) real, RXX ( ) is even and hence SXX (f ) is also even, SXX (f ) = SXX (f ) If X (t ) has periodic impulses, then SXX (f ) will have impulses

ECNG 6700 - Stochastic Processes, Detection and Estimation

29 / 35

Power Spectral Densities for Real WSS Processes

Power & Bandwidth


Sxx(f)

-B

Lowpass Spectrum
Sxx(f)

-fc - B/2

-fc

-fc + B/2

fc - B/2

fc

fc + B/2

Bandpass Spectrum

Power in band [f1 , f2 ], f PX [f1 , f2 ] = f12 SXX (f )df +

f1 f2

SXX (f )df = 2

f2 f1

SXX (f )df

ECNG 6700 - Stochastic Processes, Detection and Estimation

30 / 35

Power Spectral Densities for Real WSS Processes

Crosspower Spectral Density (CPSD)


Denition: SXY (f ) = F{RXY ( )} =
j 2 f d RXY ( )e

Crosscorrelation can be retrieved using inverse transform on CPSD RXY ( ) = F 1 {SXY (f )} = SXY (f )e j 2f df Properties:
1 2 3

SXY (f ) is generally complex valued


(f ) SXY (f ) = SYX

{SXY (f )} is an even function of f and function of f

{SXY (f )} is an odd

4 5

SXY (f ) = 0 if X (t ) and Y (t ) are orthogonal SXY (f ) = X Y (f ) if X (t ) and Y (t ) are independent


ECNG 6700 - Stochastic Processes, Detection and Estimation 31 / 35

Power Spectral Densities for Real WSS Processes

RP Examples
Questions:
1

In a communication system, the carrier signal at the receiver is modeled by Xt = cos(2 ft + ), where Uniform[, ]. Find the mean function and the correlation function of Xt .
1 HINT: cosAcosB = 2 [cos(A + B ) + cos(A B )].

2 3

Determine the stationarity of the above function. What happens if the frequency or amplitude are random, as opposed to the phase? What happens if more than one are random?

ECNG 6700 - Stochastic Processes, Detection and Estimation

32 / 35

Power Spectral Densities for Real WSS Processes

RP Examples
Questions:
1

Find the PSD of the RP with autocorrelation function, | , | | < T 1 |T RXX ( ) = 0, else Find the PSD and effective bandwidth of the RP with autocorrelation function, RXX ( ) = Ae| | , A, > 0 The PSD of a zero mean Gaussian RP is given by, 1, |f | < 500Hz SXX (f ) = 0, else Find RXX ( ) and show that X (t ) and X (t + 1) are uncorrelated and hence independent. Uncorrelated Gaussian RVs are also independent!
ECNG 6700 - Stochastic Processes, Detection and Estimation 33 / 35

Conclusion

Conclusion
We covered: Concluded Random Variable Fundamentals Introduction to Random Processes Recommended Reading: Kay - Sections 9.19.3, 9.8, 11.8, 15.115.5, 16.116.7, 17.117.4, 17.617.8 Your goals for next class: Continue working with MATLAB and Simulink Revise in class exercises based on todays discussions and ask questions in the next class Start HW3 and ask questions in the next class Review notes on RPs Part II in prep for next class
ECNG 6700 - Stochastic Processes, Detection and Estimation 34 / 35

Q&A

Thank You

Questions????

ECNG 6700 - Stochastic Processes, Detection and Estimation

35 / 35

S-ar putea să vă placă și