19501

Upload: tyas-agustina

Post on 04-Jun-2018

216 views

Category:

Documents


0 download

TRANSCRIPT

  • 8/13/2019 19501

    1/59

    4. Poisson Processes

    4.1 Definition

    4.2 Derivation of exponential distribution

    4.3 Properties of exponential distribution

    a. Normalized spacings

    b. Campbells Theorem

    c. Minimum of several exponential random variables

    d. Relation to Erlang and Gamma Distribution

    e. Guarantee Time

    f. Random Sums of Exponential Random Variables

    4.4 Counting processes and the Poisson distribution

    4.5 Superposition of Counting Processes4.6 Splitting of Poisson Processes

    4.7 Non-homogeneous Poisson Processes

    4.8 Compound Poisson Processes

    155

  • 8/13/2019 19501

    2/59

    4 Poisson Processes

    4.1 Definition

    Consider a series of events occurring over time, i.e.

    > Time0 X X X X

    DefineTi as the time between the(i 1)st andith event. Then

    Sn =T1+ T2+ . . . + Tn = Time tonth event.

    DefineN(t) =no. of events in(0, t].

    Then

    P{Sn > t}=P{N(t)< n}

    If the time for thenth event exceedst, then the number of events in(0, t]

    must be less thann

    .

    156

  • 8/13/2019 19501

    3/59

    P{Sn > t}=P{N(t)< n}

    pt(n) =P{N(t) =n} =P{N(t)< n + 1} P{N(t)< n}

    =P{Sn+1> t} P{Sn > t}

    where Sn =T1+ T2+ . . . + Tn.

    DefineQn+1(t) =P{Sn+1> t}, Qn(t) =P{Sn > t}

    Then we can write

    pt(n) =Qn+1(t) Qn(t)

    and taking LaPlace transforms

    ps(n) =Qn+1(s) Q

    n(s)

    157

  • 8/13/2019 19501

    4/59

    If qn+1(t)andqn(t)are respective pdfs.

    Qn+1(s) = 1 qn+1(s)

    s , Qn(s) =

    1 qn(s)

    s

    and

    ps(n) = 1 qn+1(s)s

    1 qn(s)s

    = qn(s) q

    n+1(s)

    s

    RecallT1 is time between0 and first event, T2 is time between first and

    second event, etc.

    158

  • 8/13/2019 19501

    5/59

    Assume {Ti} i= 1, 2, . . . are independent and with the exception of

    i= 1, are identically distributed with pdf q(t). Also assumeT1 has pdf

    q1(t). Then

    qn+1(s) =q1(s)[q

    (s)]n, qn(s) =q1(s)[q

    (s)]n1

    and

    ps(n) =qn(s) q

    n+1(s)

    s =q1(s)q

    (s)n1

    1 q(s)

    s

    Note thatq1(t)is a forward recurrence time. Hence

    q1(t) = Q(t)

    m and q1(s) =

    1 q(s)

    sm

    ps(n) = q(s)n1

    m

    1 q(s)

    s

    2

    159

  • 8/13/2019 19501

    6/59

    Assume q(t) =et

    fort >0 (m= 1/)= 0 otherwise.

    Then q(s) = / + s, 1 q(s)

    s = 1/( + s)

    and ps(n) =

    + s

    n1 1

    + s

    2 =

    1

    ( / + s)

    n+1

    .

    ps(n) = 1(/ + s)n+1

    However(/ + s)n+1 is the LaPlace transform of a gamma distribution

    with parameters(, n + 1)

    i.e.

    f(t) = et(t)n+11

    (n + 1) fort >0

    160

  • 8/13/2019 19501

    7/59

    .. pt(n) = L1{ps(n)} =

    et(t)n

    n!

    which is the Poisson Distribution. HenceN(t)follows a Poissondistribution and

    P{N(t)< n}=n1

    r=0

    pt(r) =P{Sn > t}

    P{Sn > t}=n1

    r=0et(t)r

    r! .

    We have shown that if the times between events are iid following an

    exponential distribution theN(t)is Poisson withE[N(t)] =t.

    161

  • 8/13/2019 19501

    8/59

    Alternatively ifN(t) follows a Poisson distribution, thenSn has a

    gamma distribution with pdff(t) = et(t)n1

    (n) fort >0.

    This implies time between events are exponential.

    SinceP{Sn > t}=P{N(t)< n} we have proved the identity

    P{Sn > t}= t

    et

    (t)n1

    (n)

    dx=

    n1r=0

    et

    (t)r

    r! .

    This identity is usually proved by using integration by parts.

    WhenN(t)follows a Poisson distribution withE[N(t)] =t, the

    set {N(t), t >0} is called a Poisson Process.

    162

  • 8/13/2019 19501

    9/59

    4.2 Derivation of Exponential Distribution

    Define Pn(h) =Prob. ofn events in a time intervalh

    Assume

    P0(h) = 1 h + o(h); P1(h) =h + o(h); Pn(h) =o(h) forn >1

    whereo(h)means a term(h)so that limh0

    (h)

    h = 0. Consider a finite

    time interval(0, t). Divide the interval inton sub-intervals of lengthh.

    Thent =nh.

    0 t

    h h h h

    t=nh

    The probability of no events in(0, t)is equivalent to no events in each

    sub-interval; i.e.

    Pn{T > t}=P{no events in(0, t)}

    T =Time for

    1

    st event

    163

  • 8/13/2019 19501

    10/59

    Suppose the probability of events in any sub interval are independent of

    each other. (Assumption of independent increments.) Then

    Pn{T > t}= [1 h + o(h)]n = [1

    t

    n + o(h)]n

    = (1 t

    n)n + n o(h)(1

    tn

    )n1 + . . .

    Since

    limn(1

    t

    n)n

    =et

    and

    limn

    n o(h) = limh0

    t

    ho(h) = 0

    We have P{T > t}= limh0

    Pn{T > t}=et.

    .. The pdf ofT is d

    dtP{T > t}=et. (Exponential Distribution)

    164

  • 8/13/2019 19501

    11/59

    4.3 Properties of Exponential Distribution

    q(t) =et t >0

    E(T) = 1/=m, V(t) = 1/2 =m2

    q(s) =/ + s

    Considerr < t.

    Then

    P{T > r+ t|T > r}=Conditional distribution

    = Q(r+ t)

    Q(r) =

    e(r+t)

    er =et

    i.e. P{T > r+ t|T > r}=P{T > t} for allr andt.

    Also P{T > r+ t}=e(r+t) =Q(r)Q(t) =Q(r+ t)

    Exponential distribution is only function satisfying Q(r + t) =Q(r)Q(t)

    165

  • 8/13/2019 19501

    12/59

    Proof:

    Q

    2

    n

    =Q

    1

    n

    2and in generalQ

    mn

    =Q

    1

    n

    m

    Q(1) =Q 1

    n + 1

    n + . . . +1

    n

    =

    Q 1

    nn

    , m=n

    .. Q

    m

    n =

    Q

    1

    n

    nm/n=Q(1)m/n.

    IfQ()is continuous or left or right continuous we can write

    Q(t) =Q(1)t.

    SinceQ(1)t =et logQ(1) we havelog Q(1)is the negative of the rateparameter. Hence

    Q(t) =et where = log Q(1).

    166

  • 8/13/2019 19501

    13/59

    a. Normalized Spacings

    Let {Ti} i= 1, 2, . . . , n be iid following an exponential distribution with

    E(Ti) = 1/.

    Define T(1)T(2). . . T(n) Order statistics

    Then the joint distribution of the order statistics is

    f(t(1), t(2), . . . , t(n))dt(1), t(2), . . . t(n) =P{t(1) < T(1), t(1)+dt(1), . . . }

    = n!

    1! 1! . . . 1!

    = n!et(1) et(2) . . . et(n)dt(1) . . . d t(n)

    f(t(1), . . . , t(n)) =n!nen

    1 t(i) =n!neS

    whereS=n

    1t(i) =

    n

    1ti, 0t(1) . . . t(n)

    167

  • 8/13/2019 19501

    14/59

    f(t(1), . . . , t(n)) =n!n

    eS

    , 0 t(1) . . . t(n)

    S =n1

    t(i)

    Consider

    Z1=nT(1), Z2 = (n 1)(T(2) T(1)), ,

    Z(i)

    = (n i + 1)(T(i)

    T(i1)

    ), , Z(n)

    =T(n)

    T(n1)

    .

    We shall show that {Zi} are iid exponential.

    f(Z1, Z2, . . . , Z n) =f(t(1), . . . , t(n))(t(1), . . . , t(n))(Z1, . . . , Z n)

    where (t(1), . . . , t(n))

    (Z1, . . . , Z n) is the determinant of the Jacobian.168

  • 8/13/2019 19501

    15/59

    We shall find the Jacobian by making use of the relation

    (t(1), . . . , t(n))(Z1, . . . , Z n)

    = (Z1, Z2, . . . Z n)t(1), . . . , t(n))

    1

    Zi= (n i + 1)(T(i) T(i1)), T(0) = 0

    ZiT

    (j)

    =

    n i + 1 j =i

    (n i + 1) j =i 1

    0 Otherwise

    169

  • 8/13/2019 19501

    16/59

    (Z1, . . . , Z n)

    (t(1), . . . , t(n)=

    n 0 0 0 . . . 0

    (n 1) (n 1) 0 0 . . . 0

    0 (n 2) (n 2) 0 . . . 0

    ...

    0 0 . . . . . . 1 1

    170

  • 8/13/2019 19501

    17/59

    Note: The determinant of a trangular matrix is the product of the maindiagonal terms

    .. (Z1, . . . , Z n)

    (t(1), . . . , t(n))

    =n(n 1)(n 2) . . . 2 1 =n!

    and

    f(z1, z2, . . . , zn) =n!n eS

    1

    n!=ne

    n1

    zi=neS

    asS=n

    i=1

    t(i) =z1+ . . . + zn.

    The spacings Zi = (n i + 1)(T(i) T(i1))are sometimes called

    normalized spacings.

    171

  • 8/13/2019 19501

    18/59

    Homework:

    1. Suppose there aren observations which are iid

    exponential(Ti = 1/). However there arer non-censored observations

    and(n r) censored observations all censored att(r).

    ShowZi = (n i + 1)(T(i) T(i1)) fori = 1, 2, . . . , r are iidexponential.

    2. Show that

    T(i) = Z1

    n +

    Z2n 1

    + . . . + Zi

    n i + 1

    and prove

    E(T(i)) = 1

    ij=1

    1

    n j+ 1

    Find variances and covariances of{T(i)}.

    172

  • 8/13/2019 19501

    19/59

    b. Campbells Theorem

    Let {N(t), t >0} be a Poisson Process. Assumen events occur in the

    interval(0, t]. Note thatN(t) =nis the realization of a random variable

    and has probabilityP{N(t) =n}=et(t)n

    n!DefineWn =Waiting time forn

    th event.

    If{Ti} i= 1, 2, . . . , n are the random variables representing the time

    between events

    f(t1, . . . , tn) = n1e

    ti =ne

    n

    1 ti

    But

    n1

    ti =Wn, hence

    f(t1, . . . , tn) =neWn

    173

  • 8/13/2019 19501

    20/59

    f(t1, . . . , tn) =neWn

    Now consider the transformation

    W1 =t1, W2=t1+ t2, . . . , W n =t1+ t2+ . . . + tn

    The distribution ofW= (W1, W2, . . . , W n)is

    f(W) =f(t)

    (t)

    W

    where(t)W is the determinant of the Jacobian.

    174

  • 8/13/2019 19501

    21/59

    Note: (W)

    t=

    1 0 0 0 . . . 0

    1 1 0 0 . . . 0

    1 1 1 0 . . . 0

    1 1 1 1 . . . 0...

    1 1 1 1 . . . 1

    and

    (t)

    W

    =

    (W)

    t

    1

    = 1

    .. f(w1, . . . , wn) =newn 0< w1. . . wn < t.

    But there are no events in the interval(wn, t]. This carries probability

    e(twn). Hence the joint distribution of theWis

    f(W) =newn e(twn) =net

    175

  • 8/13/2019 19501

    22/59

    f(W) =net 0w1 w2. . .wn < t

    Consider

    f(W|N(t) =n) = net

    et(t)n/n! =n!/tn.

    This is the joint distribution of the order statistics from a uniform(0, t)distribution; i.e., f(x) =

    1

    t 0< x t.

    Hence E(Wi|N(t) =n) = it

    n + 1

    i= 1, 2, . . . , n

    We can consider the unordered waiting times, conditional onN(t) =n,

    as following a uniform (0, t)distribution.

    176

  • 8/13/2019 19501

    23/59

    Sincew1 =t1, w2=t1+ t2, . . . , wn =t1+ t2+ . . . + tn

    ti =wi wi1 (w0= 0)

    The difference between the waiting times are the original timesti. Thesetimes follow the distribution conditional onN(t) =n; i.e.

    f(t1, . . . , tn|N(t) =n) =n!/tn

    Note that iff(ti) = 1/t 0< ti < t, the joint distribution

    fori = 1, 2, . . . , n ofn independent uniform(0, t) random variables

    isf(t) = 1/tn. If0 < t(1)t(2) . . . t(n) < tthe distribution of the

    order statistics isf(t(1), . . . , t(n)) =n!/t

    n

    which is the same asf(t1, . . . , tn|N(t)).

    177

  • 8/13/2019 19501

    24/59

    c. Minimum of Several Exponential Random Variables

    LetTi(i= 1, . . . , n)be ind. exponential r.v. with parameteri and let

    T =min(T1, . . . , T n)

    P{T > t}=P{T1> t, T2> t, . . . , T n > t}=n

    i=1P{Ti> t}

    =ni=1eit =et, =

    n1

    i

    T is exponential with parameter

    P{T > t}=et =n

    i=1i

    T= min(T1, . . . , T n)

    If alli =0, =n0, P{T > t}=en0t

    178

  • 8/13/2019 19501

    25/59

    DefineNas the index of the random variable which is the smallest failure

    time.

    For example ifTr Ti for alli, thenN=r.

    ConsiderP{T > t, Tr Ti alli}=P{N=r, T > t}

    P{N=r, T > t}=P{T > t, Ti Tr, i=r}

    = t P{T > t

    r, i=r| tr}f(tr)dtr

    =

    t

    e(r)trrertrdtr

    =r t

    etrdtr

    =r

    et

    179

  • 8/13/2019 19501

    26/59

    P{N=r, T > t}= r

    et

    P{N=r, T >0}=P{N=r}= r

    , =n1

    i

    P{N=r, T > t}=P{N=r}P{T > t}

    N (index of smallest) andTare independent

    If i =0

    P{N=r}= 0

    n0=

    1

    n(All populations have the same prob. of being the smallest.)

    180

  • 8/13/2019 19501

    27/59

    D. Relation to Erlang and Gamma Distribution

    ConsiderT =T1+ . . . + Tn

    Sinceqi(s) =

    ii+ s , q

    T(s) =

    ni=1

    ii+ s

    which is L.T. of Erlang distribution. Ifi are all distinct

    q(t) =

    ni=1

    Aieit ,Ai =j=i

    ij i

    Ifi =, qT(s) =

    + sn

    q(t) = (t)n1et

    (n) Gamma Distribution

    181

  • 8/13/2019 19501

    28/59

    E. Guarantee Time

    Consider the r.v. following the distribution having pdf

    q(t) =e

    (tG)

    fort > G= 0 fort G

    The parameterG is called a guarantee time

    If the transformation Y =T Gis made then f(y) =ey fory >0.

    .. E(Y) = 1/, V (Y) = 1/2, . . .

    Since T =Y + G, E(T) = 1 + G

    and central moments ifT andYare the same.

    182

  • 8/13/2019 19501

    29/59

    F. Random Sums of Exponential Random Variables

    Let {Ti} i= 1, 2, . . . , N be iid withf(t) =et and consider

    SN =T1+ T2+ . . . + TN

    with P{N=n}=pn.

    The Laplace Transform ofSN is (/ + s)n

    for fixedN=n. Hence

    f(s) =E

    + s

    Nresulting in a pdf which is a mixture of gamma

    distributions.

    f(t) =n=1

    (t)n1et(n)

    pn

    183

  • 8/13/2019 19501

    30/59

    Supposepn =pn1q n= 1, 2, . . . (negative exponential distribution)

    f(s) =

    n=1

    + sn

    pn1q=

    n=1

    q

    p p

    + sn

    = q

    pp/ + s

    1 p+s

    = q

    p

    p

    s + (1 p) =

    q

    s + q

    f(s) = q

    s + q

    SN =T1+ T2+ . . . + TN, P{N=n}=pn1q

    has exponential distribution with parameter(q).

    184

  • 8/13/2019 19501

    31/59

    4.4 Counting Processes and the Poisson Distribution

    Definition: A stochastic process {N(t), T >0} is said to be a counting

    process whereN(t)denotes the number of events that have occurred in

    the interval(0, t]. It has the properties.

    (i.) N(t)is integer value

    (ii.) N(t) 0

    (iii.) Ifs < t, N (s)N(t) and N(t) N(s) =number of events

    occurring is(s, t].

    185

  • 8/13/2019 19501

    32/59

    A counting process has independent increments if the events in disjoint

    intervals are independent; i.e.N(s) are N(t) N(s)are independent

    events.

    A counting process has stationary increments if the probability of the

    number of events in any interval depends only on the length of the

    interval; i.e.N(t)andN(s + t) N(s)

    have the same probability distribution for alls. A Poisson process

    is a counting process having independent and stationary

    increments.

    186

  • 8/13/2019 19501

    33/59

    TH. Assume {N(t), t0} is a Poisson Process. Then the dsitribution of

    Ns(t) =N(s + t) N(s) is independent ofs and only depends on the

    length of the interval, i.e.

    P{N(t + s) N(s)|N(s)}=P{N(t + s) N(s)}

    for alls. This implies that knowledge ofN(u)for0 < usis also

    irrelevant.

    P{N(t + s) N(s)|N(u), 0< us}

    =P{N(t + s) N(s)}.

    This feature defines a stationary process.

    187

  • 8/13/2019 19501

    34/59

    TH. A Poisson Process has independent increments.

    Consider0 t1< t2< t3 < t4

    > Time

    0

    X X X X

    t1 t2 t3 t4

    Consider events in(t3, t4]; i.e.

    N(t4) N(t3)

    P{N(t4) N(t3) | N(u), o < u t3}

    =P{N(t4) N(t3)}.

    Distribution is independent of what happened prior tot3. Hence if the

    intervals(t1, t2]and(t2, t4)are non-overlapping

    N(t2) N(t1)andN(t4) N(t3) are independent.

    188

  • 8/13/2019 19501

    35/59

    TH. Cov(N(t), N(s + t)) =t (Poisson Process)

    Proof N(s + t) N(t)is independent ofN(t)Cov(N(s + t) N(t), N(t)) = 0

    = Cov(N(s + t), N(t)) V(N(t)) = 0

    .. Cov(N(s + t), N(t)) =V(N(t)) =t

    as variance ofN(t)ist.

    An alternative statement of theorem is

    Cov(N(s), N(t)) = min(s, t)

    189

  • 8/13/2019 19501

    36/59

    TH. A counting process {N(t), t 0} is a Poisson Process if and only if

    (i) It has stationary and independent increments

    (ii) N(0) = 0 and

    P{N(h) = 0} = 1 h + 0(h)

    P{N(h) = 1} =h + 0(h)

    P{N(h) =j} = 0(h), j >1

    Notes: The notation0(h)little o of h refers to some function(h)for

    which

    limh0

    (h)

    h = 0

    Divide interval(0, t]inton sub-intervals of lengthh; i.e.nh =t

    P{N(kh) N(k 1)h)}=P{N(h)}

    190

  • 8/13/2019 19501

    37/59

    T =Time to event beginning att = 0.

    P{T > t} =P{N(t) = 0}=P{No events in each sub-interval}

    P{N(t) = 0} =P{T > t}= [1 h + o(h)]n

    = (1 h)n + n(1 h)n1o(h) + o(h2)

    = (1 h)n{1 + n o(h)

    1 h+ . . . }

    =

    1 tnn

    1 + t1 tn

    o(h)h + . . .

    et as n , h 0

    P{T > t}=et

    HenceT is exponential; i.e. Time between events is exponential.

    {N(t), t 0} is Poisson Process

    191

  • 8/13/2019 19501

    38/59

    4.5 Superposition of Counting Processes

    Suppose there arek counting processes which merge into a single

    counting process; e.g.k = 3.

    X X

    X

    X

    X XX X X

    X

    Process 1:

    Process 2:

    Process 3:

    Merged Process:

    The merged process is called the superposition of the individual counting

    processes

    N(t) =N1(t) + N2(t) + . . . + Nk(t)

    192

  • 8/13/2019 19501

    39/59

    A. Superposition of Poisson Processes

    N(t) =N1(t) + . . . + Nk(t)

    Suppose {Ni(t), t0} i= 1, 2, . . . , k are Poisson Processes with

    E[Ni(t)] =it.

    Note that each of the counting processes has stationary and independent

    increments.

    AlsoN(t)is Poisson with parameter

    E(N(t)) =k

    i=1

    (it) =t, =k

    i=1

    i

    N(t)is a Poisson Process

    Hence {N(t), t 0} has stationary and independent increments.

    193

  • 8/13/2019 19501

    40/59

    B. General Case of Merged Process

    Consider the merged process fromk individual processes

    XX

    XX XX X X

    X Vk> v}=Gk(v) =P{Tf(1)> v, Tf(2)> v, . . . , T f(k)> v}

    =Qf(v)k

    195

  • 8/13/2019 19501

    42/59

    Gk(v) =P{Vk > v}=Qf(v)k where

    Qf(v) =

    v

    qf(x)dx, q f(x) =Q(x)/m

    Let gk(x) =pdf of merged process

    Gk(v) =

    v

    gk(x)dx=Qf(v)k

    d

    dvGk(v) =gk(v) =kQf(v)

    k1qf(v)

    gk(v) =kQf(v)k1Q(v)

    m

    196

  • 8/13/2019 19501

    43/59

    Consider transformation z= Vkm/k

    = kVk

    m ,

    dz

    dv =

    k

    m

    gk(z) =gk(v) Vz = k

    mQfmz

    kk1 Qmz

    k m

    k

    gk(z) =Qmz

    k

    1 mz

    k

    o

    Q(x)

    m dx

    k1

    as Qf

    mzk

    =

    mz

    k

    Q(x)

    m dx= 1

    mzk

    0

    Q(x)

    m dx

    For fixedz,ask ,

    zm

    k 0 and Q

    mzk

    1

    197

  • 8/13/2019 19501

    44/59

    Also mzk

    0

    Q(x)

    m dx

    Qmzk

    m

    mz

    k =Qmz

    k

    zk

    z

    k

    .. ask

    gk(z)

    1 z

    k

    k1ez

    Thus ask , the forward recurrence time (multiplied bymk)z =

    mkVk is distributed as a unit exponential distribution. Hence for

    largek,Vk = kmz has an asymptotic exponential distribution with

    parameter =k/m. Since the asymptotic forward recurrence time isexponential, the time between events (of the merged process), is

    asymptotically exponential.

    198

  • 8/13/2019 19501

    45/59

    Note: A forward recurrence time is exponential if and only if the time

    between events is exponential; ie.

    qf(x) = Q(x)

    m =ex ifQ(x) =ex

    and ifqf(x) =ex

    Q(x) =ex

    Additional Note: The merged process isN(t) =k

    i=1

    Ni(t). Suppose

    E(Ni(t)) =t. Units ofare no. of events per unit time

    The units ofm are time per event

    Thus E(N(t)) = (k)tand(k)is mean events per unit time. The units

    of 1k or 1

    is mean time per event. Hencem = 1/for an

    individual process and the mean of the merged process is1/k.

    Ex. =6 events per year m= 126 = 2months (mean time between

    events).

    199

  • 8/13/2019 19501

    46/59

    5. Splitting of Poisson Processes

    Example: Times between births (in a family) follow an exponential

    distribution. The births are categorized by gender.

    Example: Times between back pain follow an exponential distribution.

    However the degree of pain may be categorized as the required

    medication depends on the degree of pain.

    Consider a Poisson Process {N(t), t 0} where in addition to observingan event, the event can be classified as belonging to one ofr possible

    categories.

    DefineNi(t) = no. of events of type i during(0, t]fori = 1, 2, . . . , r

    N(t) =N1(t) + N2(t) + . . . + Nr(t)

    200

  • 8/13/2019 19501

    47/59

    This process is referred to as splitting the process.

    Bernoulli Splitting Mechanism

    Suppose an event takes place in the interval(t, t + dt]. Define the

    indicator random variableZ(t) =i(i= 1, 2, . . . , r)such that

    P{Z(T) =i|event at(t, t + dt]}=pi.

    Notepi is independent of time.

    Then ifN(t) =r

    i=1

    Ni(t) the counting processes {Ni(t), t0} are

    Poisson process with parameter(pi)for i= 1, 2, . . . , r.

    201

  • 8/13/2019 19501

    48/59

    Proof: Suppose over time(0, t],n events are observed of whichsi are

    classified as of timei withr

    i=1 si =n.

    P{N1(t) =s1, N2(t) =s2, . . . , N r(t) =sr|N(t) =n}

    = n!

    s1!s2! . . . sr!ps1

    1 ps2

    2 . . . psr

    r

    Hence P{Ni(t) =si, i= 1, . . . , r andN(t) =n}

    =

    n!ri=1 si!p

    s11 p

    s22 . . . p

    srr

    et(t)n

    n!

    =r

    i=1(pit)

    siepit

    si! =

    r

    i=1P{Ni(t) =si}

    which shows that the {Ni(t)} are independent and follow Poisson

    distributions with parameters {pi}.

    {Ni(t), t0} are Poisson Processes.

    202

  • 8/13/2019 19501

    49/59

    Example of Nonhomogenous Splitting

    Suppose a person is subject to serious migraine headaches. Some of theseare so serious that medical attention is required. Define

    N(t) = no. of migraine headaches in(0, t]

    Nm(t) = no. of migraine headaches requiring medical attention

    p() = prob. requiring medical attention if

    headache occurs at(, + d).

    Suppose an event occurs at(, + d); then Prob.of requiring

    attention= p()d.

    Note that conditional on a single event taking place in(0, t], is uniform

    over(0, t]; i.e.

    f(|N(t) = 1) = 1/t 0< t and = 1

    t

    t0

    p()d

    203

  • 8/13/2019 19501

    50/59

    = 1

    t

    t0

    p()d

    0x

    t

    No event(, t)Time to event

    .. P{Nm(t) =k|N(t) =n}=

    n

    k

    k(1 )(nk)

    P{Nm(t) =k, N(t) =n}=n

    k

    k

    (1 )nk e

    t(t)n

    n!

    204

  • 8/13/2019 19501

    51/59

    P{Nm(t) =k}=n=k

    n

    k

    k(1 )nk

    et(t)n

    n!

    = k

    k!et

    n=k

    (t)n

    (n k)!(1 )nk

    = k

    k!e

    t(t

    )k

    n=k

    (t)nk(1 )nk

    (n k)!

    = k

    k!(t)ket et(1)

    P{Nm(t) =k}=et (t)

    k

    k!

    205

  • 8/13/2019 19501

    52/59

    4.7 Non-homogeneous Poisson Processes

    Preliminaries

    LetN(t) follow a Poisson distribution; i.e.

    P{N(t) =k}=et(t)k/k!

    Holdingt fixed, the generating function of the distribution is

    N(t)(s) =E[esN(t)] =

    k=0

    et(t)k

    k!

    esk

    =et

    k=0(est)k

    k! =etee

    st

    N(t)(s) =et[es1] =et(z1) if z=es

    The mean isE[N(t)] =t

    206

    {

  • 8/13/2019 19501

    53/59

    Consider the Counting Process {N(t), t0} having the Laplace

    Transform

    (*) N(t)(s) =e(t)[es1] =e(t)[z1]

    E[N(t)] = (t), P{N(t) =k}=e(t) [(t)]k/k!

    For the Poisson Process(t) =tand the mean is proportional tot.

    However whenE[N(t)]=twe call the process {N(t), t 0} a

    non-homogenized Poisson Process and E(N(t)] = (t)

    (t)can be assumed to be continuous and differentiable

    d

    dt(t) = (t) =(t).

    The quantity(t)is called intensity function.(t)can be represented by

    (t) =

    t0

    (x)dx

    207

  • 8/13/2019 19501

    54/59

    IfN(t)has the Transform given by()then

    P{N(t) =k}=e(t)(t)k/k!

    SinceP{Sn > t}=P{N(t)< n}

    We have P{S1> t}=P{N(t) t) =e(t)

    Thus pdf of time between events is

    f(t) =(t)e t

    0(x)dx, (t) =

    t0

    (x)dx

    Note that ifH= (t), thenHis a random variable following a unit

    exponential distribution.

    208

    A i d d i i N (t ) N ( ) d N ( )

  • 8/13/2019 19501

    55/59

    Assume independent increments; i.e.N(t + ) N()andN()are

    independent

    L.T. Transform (z, t) =e(t)[z1] z=es

    Generating function =E[esN(t)] =E[zN(t)]

    e(t+u)(z1) =E[zN(t+u)] =E[zN(t+u)N(u)+N(u)]

    =E[zN(t+u)N(u)] E[zN(u)]

    =e

    (u)[z1]

    .. =E[zN(t+u)N(u)]] = e(t+u)(z1)

    e(u)(z1) =e[(t+u)(u)][z1]

    where(t + u) (u) = t+uu

    (x)dx

    .. P{N(t + u) N(u) =k}= e[(t+u)(u)][(t + u) (u)]k

    k!

    209

  • 8/13/2019 19501

    56/59

    Axiomatic Derivation ofNon-Homogenized Poisson Distribution

    Assume counting process {N(t), t 0}(i)N(0) = 0

    (ii) {N(t), t0} has independent increments; i.e.N(t + s) N(s)

    andN(s)are independent

    (iii)P{N(t + h) =k|N(t) =k}= 1 (t)h + 0(h)

    P{N(t + h) =k+ 1|N(t) =k}=(t) + 0(h)

    P{N(t + h) =k+j|N(t) =k}=o(h) j 2

    p{N(t + s) N(s) =k}=e[(t+s)(s)]2[(t + s) (s)]k

    k!

    210

  • 8/13/2019 19501

    57/59

    4.8 Compound Poisson Process

    Example. Consider a single hypodermic needle which is shared. The

    times between use follow a Poisson Process. However at each use, severalpeople use it. What is the distribution of total use?

    Let {N(t), t 0} be a Poisson process and {Zn, n 1} be iid random

    variables which are independent ofN(t). Define

    Z(t) =

    N(t)n=1

    Zn

    The processZ(t)is called a Compound Poisson Process. It will be

    assumed that {Zn} takes on integer values.

    211

    Define A(s) E[eszn ] Then

  • 8/13/2019 19501

    58/59

    Define A(s) =E[e szn ]. Then

    (s|N(t) =r) =E[esZ(t)] =A(s)r

    (s|N(t) =r) =A(s)r

    (s) =r=0

    (s|N(t) =r)P(N(t) =r)

    =

    r=0A

    (s)

    r et(t)r

    r! =e

    tr=0

    (A(s)t)r

    r!

    (s) =eteA(s)t = et(1A

    (s))

    A(s) =E(eszN) = 1 sm1+ s2

    2m2+ . . .

    t(1 A(s)) =t[sm1 s2

    2m2+ . . . ], mi =E(zin)

    212

  • 8/13/2019 19501

    59/59

    Cumulant function =K(s) = log (s)

    K(s) =sm +s2

    22 + . . .

    where(m, 2)refer toZ(t).

    K(s) =t[sm1s2

    2

    m2+ . . . ]

    E[Z(t)] =tm1

    V[Z(t)] =tm2

    mi =E(zin)

    213