encs 6161 - ch5 and 6

Upload: dania-alashari

Post on 30-May-2018

229 views

Category:

Documents


0 download

TRANSCRIPT

  • 8/9/2019 ENCS 6161 - Ch5 and 6

    1/48

    Chapter 5, 6

    Multiple Random VariablesENCS6161 - Probability and Stochastic

    ProcessesConcordia University

  • 8/9/2019 ENCS 6161 - Ch5 and 6

    2/48

    Vector Random VariablesA vector r.v. X is a function X : S Rn, where S isthe sample space of a random experiment.

    Example: randomly pick up a student name from alist. S = {all student names on the list}. Let be agiven outcome, e.g. Tom

    H() : height of student W() : weight of student

    A() : age of student

    H,W,A are r.v.s.

    Let X = (H,W,A), then X is a vector r.v.

    ENCS6161 p.1/47

  • 8/9/2019 ENCS 6161 - Ch5 and 6

    3/48

    EventsEach event involving X = (X1, X2, , Xn) has acorresponding region in Rn.

    Example: X = (X1, X2) is a two-dimensional r.v.

    A = {X1 + X2 10}

    B = {min(X1, X2) 5}

    C = {X21 + X

    22 100}

    ENCS6161 p.2/47

  • 8/9/2019 ENCS 6161 - Ch5 and 6

    4/48

    Pairs of Random VariablesPairs of discrete random variables

    Joint probability mass function

    PX,Y(xj , yk) = P{X = xj

    Y = yk} = P{X = xj , Y = yk}

    Obviously jk PX,Y(xj , yk) = 1.Marginal Probability Mass FunctionPX(xj) = P{X = xj} = P{X = xj , Y = anything} =

    k=1PX,Y(xj , yk)

    Similarly PY(yk) =

    j=1 PX,Y(xj, yk).

    ENCS6161 p.3/47

  • 8/9/2019 ENCS 6161 - Ch5 and 6

    5/48

    Pairs of Random VariablesThe joint CDF of X and Y (for both discrete andcontinuous r.v.s)

    FX,Y(x, y) = P{X x, Y y}

    -

    6

    x

    (x,y)

    y

    ENCS6161 p.4/47

  • 8/9/2019 ENCS 6161 - Ch5 and 6

    6/48

    Pairs of Random VariablesProperties of the joint CDF:

    1. FX,Y(x1, y1) FX,Y(x2, y2), if x1 x2, y1 y2.2. FX,Y(, y) = FX,Y(x, ) = 03. FX,Y(, ) = 14. FX(x) = P

    {X

    x

    }= P

    {X

    x, Y = anything

    }= P{X x, Y } = FX,Y(X, )FY(y) = FX,Y(, y)FX(x), FY(y) : Marginal cdf

    ENCS6161 p.5/47

  • 8/9/2019 ENCS 6161 - Ch5 and 6

    7/48

    Pairs of Random VariablesThe joint pdf of two jointly continuous r.v.s.

    fX,Y(x, y) =

    2

    xy FX,Y(x, y)

    Obviously,

    fX,Y(x, y)dxdy = 1

    and

    FX,Y(x, y) =

    x

    y

    fX,Y(x, y)dydx

    ENCS6161 p.6/47

  • 8/9/2019 ENCS 6161 - Ch5 and 6

    8/48

    Pairs of Random VariablesThe probability

    P

    {a

    X

    b, c

    Y

    d

    }=

    b

    a d

    c

    fX,Y(x, y)dydx

    In general,

    P{(X, Y) A} = AfX,Y(x, y)dxdy

    Example:

    -

    6

    x

    y

    1

    1

    0

    A

    10

    x0

    fX,Y(x, y)dydx

    ENCS6161 p.7/47

  • 8/9/2019 ENCS 6161 - Ch5 and 6

    9/48

    Pairs of Random VariablesMarginal pdf:

    fX(x) =

    d

    dx FX(x) =

    d

    dxFX,Y(x, )=

    d

    dx

    x

    fX,Y(x, y)dydx

    =

    fX,Y(x, y)dy

    fY(y) =

    fX,Y(x, y)dx

    ENCS6161 p.8/47

  • 8/9/2019 ENCS 6161 - Ch5 and 6

    10/48

    Pairs of Random VariablesExample:

    fX,Y(x, y) =

    1 if 0 x 1, 0 y 1

    0 otherwise.Find FX,Y(x, y)

    1) x 0 or y 0, FX,Y(x, y) = 02) 0

    x

    1, and 0

    y

    1

    FX,Y(x, y) =x0

    y0

    1 dydx = xy

    3) 0 x 1, and y > 1FX,Y(x, y) =

    x

    0 1

    0

    1 dydx = x

    4) x > 1 and 0 y < 1FX,Y(x, y) = y

    5) x > 1 and y > 1

    FX,Y(x, y) = 1

    ENCS6161 p.9/47

  • 8/9/2019 ENCS 6161 - Ch5 and 6

    11/48

    IndependencePX,Y(xj , yk) = PX(xj)PY(yk), for all xj and yk(discrete r.v.s)or FX,Y(x, y) = FX(x)FY(y) for all x and y

    or fX,Y(x, y) = fX(x)fY(y) for all x and y

    Example:a)

    fX,Y =

    1 0 x 1, 0 y 10 otherwise -

    6

    b)

    fX,Y = 1 0 x 2, 0 y x0 otherwise -

    6

    ENCS6161 p.10/47

  • 8/9/2019 ENCS 6161 - Ch5 and 6

    12/48

    Conditional ProbabilityIf X is discrete,

    FY(y | x) = P{Y y, X = x}P{X = x} for P{X = x} > 0

    fY(y | x) = ddy

    FY(y | x)

    If X is continuous, P{X = x} = 0

    FY(y | x) = limh0 FY(y | x < X x + h) = limh0P

    {Y

    y, x < X

    x + h

    }P{x < X x + h}

    = limh0

    y

    x+hx

    fX,Y(x, y)dxdy

    x+h

    xfX(x)dx

    = limh0

    y fX,Y(x, y

    )dy hfX(x) h =

    y fX,Y(x, y

    )dy

    fX(x)

    fY(y

    |x) =

    d

    dy

    FY(y

    |x) =

    fX,Y(x, y)

    fX(x)

    ENCS6161 p.11/47

  • 8/9/2019 ENCS 6161 - Ch5 and 6

    13/48

    Conditional ProbabilityIf X, Y independent,

    fY(y

    |x) = fY(y)

    Similarly,

    fX(x

    |y) =

    fX,Y(x, y)

    fY(y)So,

    fX,Y(x, y) = fX(x)

    fY(y

    |x) = fY(y)

    fX(x

    |y)

    Bayes Rule:

    fX(x|y) =fX(x)

    fY(y

    |x)

    fY(y)

    ENCS6161 p.12/47

  • 8/9/2019 ENCS 6161 - Ch5 and 6

    14/48

    Conditional ProbabilityExample: A r.v. X is uniformly selected in [0, 1], andthen Y is selected uniformly in [0, x]. Find fY(y)

    Solution:fX,Y(x, y) = fX(x)fY(y|x) = 1 1

    x=

    1

    xfor 0 x 1, 0 y x and is 0 elsewhere.

    -

    6

    x1

    y

    0

    1 fY(y) =

    fX,Y(x, y)dx

    = 1

    y

    1

    xdx = ln yfor 0 y 1 and fY(y) = 0 elsewhere.

    ENCS6161 p.13/47

  • 8/9/2019 ENCS 6161 - Ch5 and 6

    15/48

    Conditional ProbabilityExample:

    - - -x {0, 1}

    Detector

    Channel

    fY(y | x) Y R X

    X = 0 A (volts), we assume P{X = 0} = P0X = 1 +A (volts), P{X = 1} = P1 = 1 P0Decide X = 0, if P{X = 0 | y} P{X = 1 | y}Decide X = 1, if P{X = 0 | y} < P{X = 1 | y}This is called the Maximum a posterior probability

    (MAP) detection.

    ENCS6161 p.14/47

  • 8/9/2019 ENCS 6161 - Ch5 and 6

    16/48

    Conditional ProbabilityBinary communication over Additive White GaussianNoise (AWGN) channel

    fY(y|0) = 12

    e(y+A)222

    fY(y|1) =1

    2 e(yA)2

    22

    Apply the MAP detection, we need to findP

    {X = 0

    |y}

    and P{

    X = 1|y}

    . Note here X is discrete,Y is continuous.

    ENCS6161 p.15/47

  • 8/9/2019 ENCS 6161 - Ch5 and 6

    17/48

    Conditional ProbabilityUse the similar approach (considering x < X x + h, and leth 0), we have

    P{

    X = 0|y}

    =P{X = 0}fY(y|0)

    fY(y)

    P{X = 1|y} = P{X = 1}fY(y|0)fY(y)

    Decide X = 0, if

    P{X = 0|y} P{X = 1|y} y 22A

    lnp0

    p1Decide X = 1, if

    P{X = 0|y} < P{X = 1|y} y >2

    2A ln

    p0

    p1

    When p0 = p1 =12 : Decide X = 0, if y 0

    Decide X = 1, if y > 0

    ENCS6161 p.16/47

  • 8/9/2019 ENCS 6161 - Ch5 and 6

    18/48

    Conditional ProbabilityProb of error:considering the special case p0 = p1 =

    12

    P = P0P{

    X = 1|X = 0} + P1P{

    X = 0|X = 1}= P0P{Y > 0|X = 0} + P1P{Y 0|X = 1}

    P{Y > 0|X = 0} = 12

    0

    e(y+A)2

    22 dy = QA

    Similarly,

    P{Y 0|X = 1} = Q

    A

    P = Q

    A

    A , P , P

    ENCS6161 p.17/47

  • 8/9/2019 ENCS 6161 - Ch5 and 6

    19/48

    Conditional ExpectationThe conditional expectation

    E[Y|x] =

    yfY(y|x)dy

    In discrete case,E[Y|x] =

    yi

    yiPY(yi|x)

    An important fact:

    E[Y] = E[E[Y|X]]Proof:

    E[E[Y|X]] =

    E[Y|x]fX(x)dx =

    yfY(y|x)fX(x)dydx

    =

    yfX,Y(x, y)dydx = E[Y]

    In general:

    E[h(Y)] = E[E[h(Y)|X]]

    ENCS6161 p.18/47

  • 8/9/2019 ENCS 6161 - Ch5 and 6

    20/48

    Multiple Random VariablesJoint cdf

    FX1,Xn(x1, xn) = P[X1 x1, , X1 xn]

    Joint pdffX1,Xn(x1, xn) =

    n

    x1, xnFX1,Xn(x1, xn)If discrete, joint pmf

    PX1,Xn(x1, xn) = P[X1 = x1, Xn = xn]Marginal pdf

    fXi(xi) =

    all x1xn except xi

    f(x1, xn)dx1 dxn

    ENCS6161 p.19/47

  • 8/9/2019 ENCS 6161 - Ch5 and 6

    21/48

    IndependenceX1, Xn are independent iff

    FX1,Xn(x1,

    xn) = FX1(x1)

    FXn(xn)

    for all x1, , xnIf we use pdf,

    fX1,Xn(x1, xn) = fX1(x1) fXn(xn)

    for all x1, , xn

    ENCS6161 p.20/47

  • 8/9/2019 ENCS 6161 - Ch5 and 6

    22/48

    Functions of Several r.v.sOne function of several r.v.s

    Z = g(X1,

    Xn)

    Let Rz = {x = (x1, , xn) s.t. g(x) z} thenFz(z) = P

    {X

    Rz

    }=

    xRz

    fX1, ,Xn(x1, , xn)dx1 dxn

    ENCS6161 p.21/47

  • 8/9/2019 ENCS 6161 - Ch5 and 6

    23/48

    Functions of Several r.v.sExample: Z = X+ Y, find Fz(z) and fz(z) in terms offX,Y(x, y)

    x

    y

    y=z-x

    Z = X+ Y z Y z XFz(z) =

    zx

    fX,Y(x, y)dydx

    fz(z) =d

    dzFz(z) =

    fX,Y(x, z x)dx

    If X and Y are independent

    fz(z) =

    fX(x)fY(z x)dx

    ENCS6161 p.22/47

  • 8/9/2019 ENCS 6161 - Ch5 and 6

    24/48

    Functions of Several r.v.sExample: let Z = X/Y. Find the pdf of Z if X and Yare independent and both exponentially distributedwith mean one.

    Can use the similar approach as previous example,but complicated. Fix Y = y , then Z = X/y andfZ(z

    |y) =

    |y

    |fX(yz

    |y). So

    fZ(z) =

    fZ,Y(z, y)dy =

    fZ(z|y)fY(y)dy

    =

    |y| fX(yz|y)fY(y)dy =

    |y| fX,Y(yz,y)dy

    ENCS6161 p.23/47

  • 8/9/2019 ENCS 6161 - Ch5 and 6

    25/48

    Functions of Several r.v.sSince X, Y are indep. exponentially distributed

    fZ(z) =

    0 y fX(yz)fY(y)dy =

    0 yeyz

    ey

    dy

    =1

    (1 + z)2for z > 0

    ENCS6161 p.24/47

  • 8/9/2019 ENCS 6161 - Ch5 and 6

    26/48

    Transformation of Random VectorsTransformation of Random Vectors

    Z1 = g1(X1

    Xn) Z2 = g2(X1

    Xn)

    Zn = gn(X1

    Xn)

    The joint CDF of Z is

    FZ1Zn(z1

    zn) = P

    {Z1

    z1,

    , Zn

    zn

    }= x:gk(x)zk

    fX1Xn(x1 xn)dx1 dxn

    ENCS6161 p.25/47

  • 8/9/2019 ENCS 6161 - Ch5 and 6

    27/48

    pdf of Linear TransformationsIf Z = AX, where A is a n n invertible matrix.

    fZ(z) = fZ1Zn(z1,

    , zn)

    =fX1Xn(x1, , xn)

    |A|x=A1z

    =fX(A1z)

    |A|

    |A| is the absolute value of the determinant of A.e.g if A =

    a b

    c d

    , then |A| = |ad bc|

    ENCS6161 p.26/47

  • 8/9/2019 ENCS 6161 - Ch5 and 6

    28/48

    pdf of General TransformationsZ1 = g1(X), Z2 = g2(X), , Zn = gn(X) whereX = (X1, , Xn)

    We assume that the set of equations:z1 = g1(x), , zn = gn(x)has a unique solution given by

    x1 = h1(z),

    , xn = hn(z)

    The joint pdf of Z is given by

    fZ1Zn(z1 zn) =fX1Xn(h1(z), , hn(z))

    |J(x1,

    , xn)

    |= fX1Xn(h1(z), , hn(z)) |J(z1, , zn)| ()where J(x1, , xn) is called the Jacobian of thetransformation.

    ENCS6161 p.27/47

  • 8/9/2019 ENCS 6161 - Ch5 and 6

    29/48

    pdf of General TransformationsThe Jacobian of the transformation

    J(x1,

    , xn) = detg1x1

    g1xn gnx1

    gnxn

    and

    J(z1, , zn) = deth1z1

    h1zn

    hnz1

    hnzn

    = 1J(x1 xn)

    x=h(z)

    Linear transformation is a special case of ()

    ENCS6161 p.28/47

  • 8/9/2019 ENCS 6161 - Ch5 and 6

    30/48

    pdf of General TransformationsExample: let X and Y be zero-mean unit-varianceindependent Gaussian r.vs. Find the joint pdf of Vand W defined by:

    V = (X2 + Y2) 12

    W = (X, Y) = arctan(Y /X) W [0, 2)

    This is a transformation from Cartesian to Polarcoordinates. The inverse transformation is:

    x

    y

    v

    wx

    y x = v cos(w)y = v sin(w)

    ENCS6161 p.29/47

  • 8/9/2019 ENCS 6161 - Ch5 and 6

    31/48

    pdf of General TransformationsThe Jacobian

    J(v, w) = cos w v sin wsin w v cos w = v cos

    2 w + v sin2 w = v

    Since X and Y are zero-mean unit-varianceindependent Gaussian r.v.s,

    fX,Y(x, y) =1

    2 ex2

    2 1

    2 ey2

    2 =1

    2 ex2+y2

    2

    The joint pdf of V, W is then

    fV,W(v, w) = v 1

    2 e

    (v2 cos2 w+v2 sin2 w)

    2 =

    v

    2 ev2

    2

    for v 0 and 0 w < 2

    ENCS6161 p.30/47

  • 8/9/2019 ENCS 6161 - Ch5 and 6

    32/48

    pdf of General TransformationsThe marginal pdf of V and W

    fV(v) =

    fV,W(v, w)dw =

    2

    0

    v

    2e

    v2

    2 dw = vev2

    2

    for v 0. This is called the Rayleigh Distribution.fW(w) =

    fV,W(v, w)dv =1

    2

    0ve

    v2

    2 dv =1

    2

    for 0 w < 2Since

    fV,W(v, w) = fV(v)fW(w)

    V, W are independent.

    ENCS6161 p.31/47

  • 8/9/2019 ENCS 6161 - Ch5 and 6

    33/48

    Expected Value of Functions of r.v.sLet Z = g(X1, X2, , Xn) thenE[Z] =

    g(x1, , xn)fX1Xn(x1, , xn)dx1 dxnFor discrete case,

    E[Z] =

    all possible x

    g(x1, , xn)PX1Xn(x1, , xn)

    Example: Z = X1 + X2 + + XnE[Z] = E[X1 + X2 + + Xn]

    =

    (x1 + + xn)fX1Xn(x1 xn)dx1 dxn= E[X1] + + E[Xn]

    ENCS6161 p.32/47

  • 8/9/2019 ENCS 6161 - Ch5 and 6

    34/48

    Expected Value of Functions of r.v.sExample: Z = X1X2 XnE[Z] =

    x1 xnfX1Xn(x1 xn)dx1 dxnIf X1, X2, , Xn are indep.

    E[Z] = E[X1X2 Xn] = E[X1]E[X2] E[Xn]The (j,k)-th moment of two r.v.s X&Y is

    E[XjYk] =

    xjykfX,Y(x, y)dxdy

    If j = k = 1 , it is called the correlation.

    E[XY] =

    xyfX,Y(x, y)dxdy

    If E[XY] = 0, we call X&Y are orthogonal.

    ENCS6161 p.33/47

  • 8/9/2019 ENCS 6161 - Ch5 and 6

    35/48

    Expected Value of Functions of r.v.sThe (j,k)-th central moment of X, Y is

    E

    (X E(X))j (Y E(Y))k

    when j = 2, k = 0, V ar(X)j = 0, k = 2, V ar(Y)

    When j = k = 1, it is called the covariance of X, Y

    Cov(X, Y) = E[(X E(X))(Y E(Y))]= E[XY] E[X]E[Y] = Cov(Y, X)

    The correlation coefficient of X and Y is defined as

    X,Y = Cov(X, Y)XYwhere X =

    V ar(X) and Y =

    V ar(Y).

    ENCS6161 p.34/47

  • 8/9/2019 ENCS 6161 - Ch5 and 6

    36/48

    Expected Value of Functions of r.v.sThe correlation coefficient 1 X,Y 1Proof:

    0 EX E[X]X Y E[Y]Y 2

    = 1 2X,Y + 1 = 2(1 X,Y)

    If X,Y = 0, X, Y are said to be uncorrelated.If X, Y are independent,E[XY] = E[X]E[Y] Cov(X, Y) = 0 X,Y = 0.Hence, X, Y are uncorrelated.The converse is not always true. It is true in the caseof Gaussian r.v.s ( will be discussed later)

    ENCS6161 p.35/47

  • 8/9/2019 ENCS 6161 - Ch5 and 6

    37/48

    Expected Value of Functions of r.v.sExample: is uniform in [0, 2).Let X = cos and Y = sin X and Y are not independent, since X2 + Y2 = 1.

    However

    E[XY] = E[sin cos ] = E[1

    2sin(2)]

    =20

    1

    2

    1

    2sin(2)d = 0

    We can also show E[X] = E[Y] = 0. So

    Cov(X, Y) = E[XY] E[X]E[Y] = 0 X,Y = 0

    X, Y are uncorrelated but not independent.

    ENCS6161 p.36/47

  • 8/9/2019 ENCS 6161 - Ch5 and 6

    38/48

    Joint Characteristic FunctionJoint Characteristic Function

    X1Xn(w1, , wn) = E

    ej(w1X1++wnXn)

    For two variablesX,Y(w1, w2) = E

    ej(w1X+w2Y)

    Marginal characteristic function

    X(w) = X,Y(w, 0) Y(w) = X,Y(0, w)

    If X, Y are independent

    X,Y(w1, w2) = E[ejw1X+jw2Y]

    = E[ejw1X ]E[ejw2Y] = X(w1)Y(w2)

    ENCS6161 p.37/47

  • 8/9/2019 ENCS 6161 - Ch5 and 6

    39/48

    Joint Characteristic FunctionIf Z = aX+ bY

    Z(w) = E[ejw(aX+bY)] = X,Y(aw,bw)

    If Z = X+ Y, X and Y are independent

    Z(w) = X,Y(w, w) = X(w)Y(w)

    ENCS6161 p.38/47

  • 8/9/2019 ENCS 6161 - Ch5 and 6

    40/48

    Jointly Gaussian Random VariablesConsider a vector of random variablesX = (X1, X2, Xn). Each with mean mi = E[Xi], fori = 1,

    , n and the covariance matrix

    K =

    V ar(X1) Cov(X1, X2) Cov(X1, Xn)

    ...... ...

    Cov(Xn, X1)

    V ar(Xn)

    Let m = [m1, , mn]T be the mean vector andx = [x1, , xn]T where ()T denotes transpose. ThenX1, X2,

    Xn are said to be the jointly Gaussian if

    their joint pdf is:

    fX(x) =1

    (2)n2

    |K

    |

    12

    exp

    1

    2(x m)TK1(x m)

    where |K| is the determinant of K.

    ENCS6161 p.39/47

  • 8/9/2019 ENCS 6161 - Ch5 and 6

    41/48

    Jointly Gaussian Random Variables

    For n = 1, let V ar(X1) = 2, then fX(x) =

    12

    e(xm)2

    22

    For n = 2, denote the r.v.s by X and Y. Let

    m =

    mXmY

    K = 2X XYXYXYXY

    2Y

    Then

    |K

    |= 2X

    2Y(1

    2XY)

    K1 =1

    2X2Y(1 2XY)

    2Y XYXYXYXY 2X

    and

    fX,Y(x, y) = 12XY

    1 2XY

    exp 12(1 2XY)

    ( x mxx

    )2

    2XY( x mxx

    )(y my

    y) + (

    y myy

    )2ENCS6161 p.40/47

  • 8/9/2019 ENCS 6161 - Ch5 and 6

    42/48

    Jointly Gaussian Random VariablesThe marginal pdf of X:

    fX(x) =1

    2xe

    (xmx)2

    22x

    The marginal pdf of Y:fY(y) =

    12y

    e

    (ymy)2

    22y

    If XY = 0

    X, Y are independent.

    The conditional pdf.

    fX|Y(x

    |y) =

    fX,Y(x, y)

    fY(y)

    =

    exp

    1

    2(12XY

    )2X

    x XY XY (y my) mx

    222X(1 2XY)

    N(XY XY

    (y my) + mx

    E[X|Y], 2X(1 2XY)

    V ar(X|Y)

    )

    ENCS6161 p.41/47

  • 8/9/2019 ENCS 6161 - Ch5 and 6

    43/48

  • 8/9/2019 ENCS 6161 - Ch5 and 6

    44/48

    Linear Transformation of Gaussian r.v.s

    Since K is symmetric, it is always possible to find a matrix A s.t.

    = AKAT is diagonal. =

    1 0

    2

    . . .

    0 n

    so

    fY(y) =1

    (2)n2 || 12 e

    12 (ym)T1(ym)

    =1

    21e

    (y1m1)2

    21

    1

    2ne

    (ynmn)2

    2n

    That is, we can transform X into n independent Gaussian r.v.s

    Y1 Yn with means mi and variance i.

    ENCS6161 p.43/47

  • 8/9/2019 ENCS 6161 - Ch5 and 6

    45/48

    Mean Square Estimation

    We use g(X) to estimate Y, write as Y = g(X). The

    cost associated with the estimation error is C(Y Y).

    Y X

    e.g.

    C(Y Y) = (Y Y)2= (Y g(X))2

    The mean square error

    E[C] = E[(Y g(X))2]Case 1: if Y = a

    E[(Y

    a)2] = E[Y2]

    2aE[Y] + a2 = f(a)df

    da= 0 a = E[Y] = Y

    The mean square error: E[(Y

    Y)

    2] = V ar[Y]

    ENCS6161 p.44/47

    S i i

  • 8/9/2019 ENCS 6161 - Ch5 and 6

    46/48

    Mean Square Estimation

    Case 2: if Y = aX+ b, then E[(Y aX b)2] = f(a, b)

    fa = 0f

    b = 0 a = XY

    Y

    X

    , b = Y

    aX

    Y = XY YX

    (X X) + YThis is called the MMSE linear estimation.

    The mean square error:

    E[(Y Y)2] = 2Y(1 2XY)If XY = 0, Y = Y, error=

    2

    Y

    , reduces to case 1.

    If XY = 1, Y = YX (X X) + Y, and error= 0

    ENCS6161 p.45/47

    M S E i i

  • 8/9/2019 ENCS 6161 - Ch5 and 6

    47/48

    Mean Square Estimation

    Case 3: Y is a general function of X.

    E[(Y Y)2] = E[(Y g(X))2] = E[E[(Y g(X))2|X]]

    =

    E[(Y g(x))2|x]fX(x)dxFor any x, choose g(x) to minimize E[(Y g(x))2|X = x]

    g(x) = E[Y

    |X = x]

    This is called the MMSE estimation.

    Example: X, Y joint Gaussian.

    E[Y|X] = X,YY

    X (X X) + YThe MMSE estimation is linear for Gaussian.

    ENCS6161 p.46/47

    M S E ti ti

  • 8/9/2019 ENCS 6161 - Ch5 and 6

    48/48

    Mean Square Estimation

    Example: X uniform(1, 1), Y = X2. We haveE[X] = 0

    XY = E[XY]

    E[X]E[Y] = E[XY] = E[X3] = 0So, the MMSE linear estimation:

    Y = Yand the error is 2Y.

    The MMSE estimation:g(x) = E[Y|X = x] = E[X2|X = x] = x2

    So Y = X2 and the error is 0.

    ENCS6161 p.47/47