primer on random processes
TRANSCRIPT
Randomprocesses
E. Savin
ProbabilitySpace
Probabilitytriplet
Randomvariable
Stochasticprocess
Mean-squarestochasticprocess
Stationarity
Powerspectralmeasure
Mean-squarederivation
Filtering
Bibliography
Primer on random processesMG3409 - Lecture #1 part A
E. Savin1,2
1Computational Fluid Dynamics Dept.Onera, France
2Mechanical and Civil Engineering Dept.Ecole Centrale Paris, France
January 25, 2015
Randomprocesses
E. Savin
ProbabilitySpace
Probabilitytriplet
Randomvariable
Stochasticprocess
Mean-squarestochasticprocess
Stationarity
Powerspectralmeasure
Mean-squarederivation
Filtering
Bibliography
Outline
1 Probability SpaceProbability tripletRandom variableStochastic process
2 Mean-square stochastic processStationarityPower spectral measureMean-square derivation
3 Filtering
Randomprocesses
E. Savin
ProbabilitySpace
Probabilitytriplet
Randomvariable
Stochasticprocess
Mean-squarestochasticprocess
Stationarity
Powerspectralmeasure
Mean-squarederivation
Filtering
Bibliography
Outline
1 Probability SpaceProbability tripletRandom variableStochastic process
2 Mean-square stochastic processStationarityPower spectral measureMean-square derivation
3 Filtering
Randomprocesses
E. Savin
ProbabilitySpace
Probabilitytriplet
Randomvariable
Stochasticprocess
Mean-squarestochasticprocess
Stationarity
Powerspectralmeasure
Mean-squarederivation
Filtering
Bibliography
Probability spaceFormal definition
We model a real-world process, or ”experiment”, or”system”, consisting of states occurring randomly.
It is mathematically represented by a triplet pA ,T ,P q ofwhich elements are ultimately specified by the modeler:
A is the set of outcomes, or sample space, constitutedby the states of the causes which influence the states ofthe system;T is a σ-algebra, i.e. a set of all the events the modelerconsiders. An event is a set of zero or more outcomes,thus it is a subset of the sample space;P is a function returning an event’s probability, or aprobability measure.
Randomprocesses
E. Savin
ProbabilitySpace
Probabilitytriplet
Randomvariable
Stochasticprocess
Mean-squarestochasticprocess
Stationarity
Powerspectralmeasure
Mean-squarederivation
Filtering
Bibliography
Probability spaceσ-algebra
Definition
A collection T of subsets of A is a σ-algebra, or σ-field, if:
1 H P T and A P T ;
2 a P T ñ ac “ A za P T ;
3 If tanun“1,2,... is an arbitrary countable family of events
in T , thenď
nPNan P T .
Example #1: a fair coin, A “ t1, 2u. Then the σ-algebraT “ 2A
“ ttu, t1u, t2u, t1, 2uu contains 22“ 4 events.
Example #2: a fair die, A “ t1, 2, 3, 4, 5, 6u. Then the σ-algebraT “ 2A
“ ttu, t1u, t2u, ...t6u, t1, 2u, t3, 4u, t5, 6u, t1, 3u, ...t1, 3, 5u, t2, 4, 6u, t1, 3, 4, 5u, ..., t1, 2, 3, 4, 5, 6uu contains 26
“ 64events.
Randomprocesses
E. Savin
ProbabilitySpace
Probabilitytriplet
Randomvariable
Stochasticprocess
Mean-squarestochasticprocess
Stationarity
Powerspectralmeasure
Mean-squarederivation
Filtering
Bibliography
Probability spaceProbability measure
Definition
A probability measure is a positive, bounded measure thetotal mass of which is 1: P pA q “ 1.
1 A positive measure m is a function m : A Ñ r0,`8ss.t. for countably many disjoint events tanun“1,2,... P T ,č
n
an “ H,ď
n
an P T , then
m
˜
ď
n
an
¸
“ÿ
n
mpanq ;
2 It is bounded if mpA q ă `8;
3 It is a probability measure if mpA q “ 1.
Randomprocesses
E. Savin
ProbabilitySpace
Probabilitytriplet
Randomvariable
Stochasticprocess
Mean-squarestochasticprocess
Stationarity
Powerspectralmeasure
Mean-squarederivation
Filtering
Bibliography
Random variableDefinitions
Definition
A random variable is a measurable function X : A Ñ A 1
which maps a combination of outcomes from a sample spaceA to a combination of consequences in a state space A 1.
Definition
A function f : A Ñ A 1 is measurable if @a1 P A 1, f´1pa1q ismeasurable in A : @a1 P T 1, f´1pa1q P T , where T , T 1 areσ-algebras defined on A , A 1 respectively.
Example #1: let A be the sample space describing thereal-world system ”tossing a fair coin”: A “ t”temperature”,”pressure”, ”initial position”, ”initial speed”, etc.u. ThenA 1
“ t1, 2u.
Example #2: let A be the sample space describing a real-worldsystem ”roll of dice”: A “ t”temperature”, ”pressure”, ”initialposition”, ”initial speed”, etc.u. Then A 1
“ t1, 2, 3, 4, 5, 6u.
Randomprocesses
E. Savin
ProbabilitySpace
Probabilitytriplet
Randomvariable
Stochasticprocess
Mean-squarestochasticprocess
Stationarity
Powerspectralmeasure
Mean-squarederivation
Filtering
Bibliography
Random variableCausality principle
Let X : A Ñ A 1 be a random variable, then A 1 isendowed with the σ-algebra T 1 and the probabilitymeasure P 1 inherited from T and P by the causalityprinciple:
@a1 P T 1 , P 1pa1q “ P pX´1pa1qq .
P 1 is now the probability measure of X on A 1.
The original sample space A (the system and its environment inthe foregoing examples) is often suppressed, since it ismathematically hard to describe.
Examples: The possible values of the random variable ”cointoss” or ”dice roll” are then treated as a sample space A 1
constituted by numerical quantities–which is easier to deal with.
Randomprocesses
E. Savin
ProbabilitySpace
Probabilitytriplet
Randomvariable
Stochasticprocess
Mean-squarestochasticprocess
Stationarity
Powerspectralmeasure
Mean-squarederivation
Filtering
Bibliography
Stochastic processDefinition and terminology
Definition
A stochastic processX : T Ñ Rn
t ÞÑXtis a function from T
into the set of random variables defined on pA ,T ,P q withvalues in Rn ” A 1. pXt , t P T q is a collection of randomvariables with values in Rn. Therefore:
If t P T is frozen: Xt is a random variable defined onpA ,T ,P q with values in Rn;
If a P A is frozen: t ÞÑXtpaq “ xptq is a trajectory, ora sample path;
If both t P T and a P A are frozen: a, t ÞÑXtpaq is anoutcome, or a realization, of the random variable Xt.
Randomprocesses
E. Savin
ProbabilitySpace
Probabilitytriplet
Randomvariable
Stochasticprocess
Mean-squarestochasticprocess
Stationarity
Powerspectralmeasure
Mean-squarederivation
Filtering
Bibliography
Stochastic process
Stochastic process indexed on R`.
Randomprocesses
E. Savin
ProbabilitySpace
Probabilitytriplet
Randomvariable
Stochasticprocess
Mean-squarestochasticprocess
Stationarity
Powerspectralmeasure
Mean-squarederivation
Filtering
Bibliography
Stochastic processRemarks
pXt , t P T q is called ”the stochastic process defined onpA ,T ,P q, indexed on T , with values in Rn”.
The indexation set T could be any set, countable ornot, bounded or not, etc.
However we will only consider the real line T “ R in thesubsequent lectures, since t will be a time.
Examples: earthquake, wind, turbulence, swell, ocean surge...can be modeled by stochastic processes, where A “ t”the statesof the environment”u is very difficult to figure out, butA 1
“ t”vectors of R3”u is much easier to handle.
Randomprocesses
E. Savin
ProbabilitySpace
Probabilitytriplet
Randomvariable
Stochasticprocess
Mean-squarestochasticprocess
Stationarity
Powerspectralmeasure
Mean-squarederivation
Filtering
Bibliography
Stochastic processProbability law
For t P R fixed, Xt is a Rn-valued random variabledefined on pA ,T ,P q, of which probability law PX isparameterized by t and defined by:
PXpB; tq “ P 1pXt P Bq
“ P pX´1t pBq P T q , @B P T 1 .
The most common choice for the σ-algebra T 1 is the Borel
σ-algebra BpRnq–the one generated by the collection of all open
sets in Rn.
The distribution function FXpx; tq : Rn Ñ r0, 1s:
PXpdx; tq “ dFXpx; tq ,
FXpx; tq “ P 1pXt Ps ´8, x1s ˆ ¨ ¨ ¨ ˆs ´8, xnsq .
Randomprocesses
E. Savin
ProbabilitySpace
Probabilitytriplet
Randomvariable
Stochasticprocess
Mean-squarestochasticprocess
Stationarity
Powerspectralmeasure
Mean-squarederivation
Filtering
Bibliography
Stochastic processJoint and marginal probability laws
Joint probability laws: The two-fold probabilitymeasure PXY pdxdy; t, t1q “ dFXY px,y; t, t1q with:
FXY px,y; t, t1q “ P 1pXt Ps ´8, x1s ˆ ¨ ¨ ¨ ˆs ´8, xns
and Y t1 Ps ´8, y1s ˆ ¨ ¨ ¨ ˆs ´8, ynsq .
Marginal probability laws: Let I be the set of all finite,unordered parts of R of the form I Q i “ tt1, t2, . . . tku.Let Xpiq “ pXt1 ,Xt2 , . . .Xtkq be the Rnˆk-valuedrandom variable constructed from the part i, then thecollection of marginal probability laws of the processpXt , t P Rq is tPXpiqpdx1dx2 ¨ ¨ ¨ dxkq, i P Iu.
Randomprocesses
E. Savin
ProbabilitySpace
Probabilitytriplet
Randomvariable
Stochasticprocess
Mean-squarestochasticprocess
Stationarity
Powerspectralmeasure
Mean-squarederivation
Filtering
Bibliography
Stochastic processSecond-order stochastic process
Definition
pXt , t P Rq is a second-order stochastic process if:
Et}Xt}2u ă `8 , @t P R ,
where }x} “b
řnj“1 x
2j and EtfpXtqu “
ş
Rn fpxqPXpdx; tq.
The set L2pA ,Rnq of Rn-valued second-order randomvariables defined on pA ,T ,P q is an Hilbert space, withthe scalar product ppX,Y qq2 “ EtpX,Y qu and thenorm |||X||| “
a
ppX,Xqq2 “a
Et}X}2u.Interests: the same as for L2pRnq! (separability,uniform convexity).
Randomprocesses
E. Savin
ProbabilitySpace
Probabilitytriplet
Randomvariable
Stochasticprocess
Mean-squarestochasticprocess
Stationarity
Powerspectralmeasure
Mean-squarederivation
Filtering
Bibliography
Stochastic processMean and auto-correlation functions
Definition
Let pXt , t P Rq be a Rn-valued, second-order stochasticprocess defined on pA ,T ,P q and indexed on R.
Its mean function mX : RÑ Rn is:
mXptq “
ż
Rn
xPXpdx; tq “ EtXtu .
The process is centered if mXptq “ 0, @t P R.
Its auto-correlation function RX : Rˆ RÑ Rnˆn is:
RXpt, t1q “
ij
RnˆRn
xbyPXXpdxdy; t, t1q “ EtXtbXt1u .
Randomprocesses
E. Savin
ProbabilitySpace
Probabilitytriplet
Randomvariable
Stochasticprocess
Mean-squarestochasticprocess
Stationarity
Powerspectralmeasure
Mean-squarederivation
Filtering
Bibliography
Outline
1 Probability SpaceProbability tripletRandom variableStochastic process
2 Mean-square stochastic processStationarityPower spectral measureMean-square derivation
3 Filtering
Randomprocesses
E. Savin
ProbabilitySpace
Probabilitytriplet
Randomvariable
Stochasticprocess
Mean-squarestochasticprocess
Stationarity
Powerspectralmeasure
Mean-squarederivation
Filtering
Bibliography
Mean-square stochastic processDefinitions
Definition
Let pXt , t P Rq be a Rn-valued stochastic process defined onpA ,T ,P q and indexed on R. It is:
stationary if @u P R, pXt`u , t P Rq and pXt , t P Rqhave the same marginal probability laws;
mean-square stationary if it is second-order,mXptq “mX (independent of time), andRXpt, t
1q “ RXpt´ t1q.
If pXt , t P Rq is stationary then PXpdx; tq “ PXpdxq(independent of time) and PXXpdxdy; t, t1q “ PXXpdxdy; t´ t1q.
Stationary ñ mean-square stationary, but the reverse is not true.
Randomprocesses
E. Savin
ProbabilitySpace
Probabilitytriplet
Randomvariable
Stochasticprocess
Mean-squarestochasticprocess
Stationarity
Powerspectralmeasure
Mean-squarederivation
Filtering
Bibliography
Mean-square stochastic processStationarity
Stationary v.s. non-stationary processes.
Randomprocesses
E. Savin
ProbabilitySpace
Probabilitytriplet
Randomvariable
Stochasticprocess
Mean-squarestochasticprocess
Stationarity
Powerspectralmeasure
Mean-squarederivation
Filtering
Bibliography
Mean-square stochastic processSpectral measure
Data: pXt , t P Rq a Rn-valued, second-order stochasticprocess defined on pA ,T ,P q, indexed on R, andmean-square stationary.
Hypothesis: t ÞÑ RXptq is continuous on R.
Theorem (Bochner)
Then there exists an Hermitian, positive, and boundedmeasure MX P Cnˆn, called the spectral measure of theRn-valued stochastic process pXt , t P Rq, s.t.:
RXptq “
ż
ReiωtMX pdωq .
Randomprocesses
E. Savin
ProbabilitySpace
Probabilitytriplet
Randomvariable
Stochasticprocess
Mean-squarestochasticprocess
Stationarity
Powerspectralmeasure
Mean-squarederivation
Filtering
Bibliography
Mean-square stochastic processPower spectral density
Definition
The spectral density SX is the density of the spectralmeasure MX if it exists:
MX pdωq “ SXpωqdω .
It is Hermitian, positive, and integrable (sinceMXpRq ă `8). Then lim
|t|Ñ`8RXptq “ 0 (as the
Fourier transform of an integrable function).
A necessary condition: mX “ 0.
A sufficient condition: lim|t|Ñ`8
RXptq “ 0.
Let B Ď R, B ÞÑ TrMXpBq is the power spectralmeasure on B, and if SX exists, ω ÞÑ TrSXpωq is thepower spectral density (PSD) of the process.
Randomprocesses
E. Savin
ProbabilitySpace
Probabilitytriplet
Randomvariable
Stochasticprocess
Mean-squarestochasticprocess
Stationarity
Powerspectralmeasure
Mean-squarederivation
Filtering
Bibliography
Mean-square stochastic processMean-square derivative
Data: pXt , t P Rq a Rn-valued, second-order stochasticprocess defined on pA ,T ,P q, indexed on R.
Definition
Let t P R, pXt , t P Rq is mean-square derivable at t iff thesequence of random variables Xh “
1hpXt`h ´Xtq
converges in L2pA ,Rnq as hÑ 0. That limit is denoted by9Xt s.t.:
limhÑ0
|||Xh ´9Xt||| “ 0 .
Randomprocesses
E. Savin
ProbabilitySpace
Probabilitytriplet
Randomvariable
Stochasticprocess
Mean-squarestochasticprocess
Stationarity
Powerspectralmeasure
Mean-squarederivation
Filtering
Bibliography
Mean-square stochastic processSecond-order properties of the derivative
A sufficient condition: B2RXBtBt1 exists for all pt, t1q in RˆR.
Then p 9Xt , t P Rq is a Rn-valued, second-orderstochastic process defined on pA ,T ,P q, indexed on R.In addition:
m 9Xptq “dmX
dtptq ,
R 9Xpt, t1q “
B2RX
BtBt1pt, t1q ,
RX 9Xpt, t1q “ EtXt b 9Xt1u “
BRX
Bt1pt, t1q ,
R 9XXpt, t1q “ Et 9Xt bXt1u “
BRX
Btpt, t1q .
Randomprocesses
E. Savin
ProbabilitySpace
Probabilitytriplet
Randomvariable
Stochasticprocess
Mean-squarestochasticprocess
Stationarity
Powerspectralmeasure
Mean-squarederivation
Filtering
Bibliography
Mean-square stochastic processDerivative of a mean-square stationary process
A necessary and sufficient condition:
ż
Rω2 trMXpdωq ă `8 .
Then p 9Xt , t P Rq is a Rn-valued, second-orderstochastic process defined on pA ,T ,P q, indexed on R,and mean-square stationary. In addition:
m 9Xptq “ 0 ,
R 9Xptq “ ´d2RX
dt2ptq ,
RX 9Xptq “dRX
dtptq ,
M 9Xpdωq “ ω2MXpdωq .
Randomprocesses
E. Savin
ProbabilitySpace
Probabilitytriplet
Randomvariable
Stochasticprocess
Mean-squarestochasticprocess
Stationarity
Powerspectralmeasure
Mean-squarederivation
Filtering
Bibliography
Outline
1 Probability SpaceProbability tripletRandom variableStochastic process
2 Mean-square stochastic processStationarityPower spectral measureMean-square derivation
3 Filtering
Randomprocesses
E. Savin
ProbabilitySpace
Probabilitytriplet
Randomvariable
Stochasticprocess
Mean-squarestochasticprocess
Stationarity
Powerspectralmeasure
Mean-squarederivation
Filtering
Bibliography
Filtering of a stochastic processLinear filter
x * yh
A linear convolution filter:
yptq “
ż
Rhpt´ τqxpτq dτ
where:
t ÞÑ hptq is the impulse response function in Rmˆn;
ω ÞÑ phpωq “ş
R e´iωthptqdt is the frequency response
function;p ÞÑ Hppq “
ş`8
´8e´pt
hptqdt, p P DH Ă C, is thetransfer function.
Randomprocesses
E. Savin
ProbabilitySpace
Probabilitytriplet
Randomvariable
Stochasticprocess
Mean-squarestochasticprocess
Stationarity
Powerspectralmeasure
Mean-squarederivation
Filtering
Bibliography
Filtering of a stochastic processStationary response
Data: pXt , t P Rq a Rm-valued, second-order stochasticprocess defined on pA ,T ,P q, indexed on R, andmean-square stationary.
If t ÞÑ hptq is integrable or square integrable, thenpY t , t P Rq is a Rn-valued, second-order stochasticprocess defined on pA ,T ,P q, indexed on R, andmean-square stationary s.t.:
mY “php0qmX ,
RY ptq “
ij
RˆR
hpτqRXpt` τ1 ´ τqhpτ 1qT dτdτ 1 ,
MY pdωq “ phpωqMXpdωqph˚pωq .
Randomprocesses
E. Savin
ProbabilitySpace
Probabilitytriplet
Randomvariable
Stochasticprocess
Mean-squarestochasticprocess
Stationarity
Powerspectralmeasure
Mean-squarederivation
Filtering
Bibliography
Further reading...
P. Kree & C. Soize: Mathematics of Random Phenomena:Random Vibrations of Mechanical Structures, Reidel PublishingCo., Dordrecht (1986);
C. Soize: Methodes Mathematiques en Analyse du Signal, Masson,Paris (1993);
T.T. Soong & M. Grigoriu: Random Vibration of Mechanical andStructural Systems, Prentice Hall, Upper Saddle River NJ (1993).