196 7 Gaussian Distributions
                   Xn                         σ2X ¼  Xn    σ2i   ð7:46Þ
            μX ¼ μi                                              ð7:47Þ
                     i¼1 (  Xn Xn                    )i¼1
                  X$N             μi; σi2
                            i¼1 i¼1
Proof Consider the case of n ¼ 3 as follows:
                     X ¼ X1 þ X2 þ X3                            ð7:48Þ
Let
                        Y ¼ X1 þ X2                              ð7:49Þ
By (7.34),
                  Y  $    À    þ  μ2;     σ21  þ  σ22Á
                        N μ1
Substituting (7.49) into (7.48), we have
                        X ¼ Y þ X3
By (7.34),
                  X  $    À    þ  μ3;     σ2Y  þ  σ23Á           ð7:50Þ
                        N μY
By (5.21),
            μY ¼ EðYÞ ¼ EðX1 þ X2Þ ¼ EðX1Þ þ EðX2Þ ¼ μ1 þ μ2     ð7:51Þ
   By (5.52), and, since Cov(X1, X2) ¼ 0 because X1 and X2 are independent, we
have
            σY2 ¼ VarðYÞ ¼ VarðX1Þ þ VarðX2Þ ¼ σ12 þ σ22         ð7:52Þ
Substituting (7.51) and (7.52) into (7.50), we have
            X  $    À   þ  μ2  þ  μ3;     σ12  þ  σ22  þ   σ32Á
                  N μ1
   This proves (7.47) for n ¼ 3. By mathematical induction, (7.48) is proven for an
arbitrary n.
                                                                                            Q.E.D.
7.3 Two Jointly Gaussian RVs                                                                                                      197
7.3 Two Jointly Gaussian RVs
Two RVs X and Y are said to be jointly normal, if its joint pdf is given by the
following equation:
                                                                       &   2           À2ρðxÀμσXXÞσðYyÀμY Þþ     2'
      f      ðx;  yÞ  ¼                p1 ffiffiffiffiffiffiffiffiffiffiffiffiffi  eÀ2ð1À1 ρ2 Þ       xÀμX                              yÀμY          ð7:53Þ
                                          1 À ρ2                             σX                                σY
         XY                2π σ X σ Y
where μX, μY, σ2X, σ2Y, and À 1 ρ 1 are constants. When X and Y are jointly
normal, we write
                                            ðX;   YÞ     $    À          μY  ;  σX2 ;  σY2  ;   Á                           ð7:54Þ
                                                            N μX;                              ρ
   The above pdf is called the bivariate normal pdf. The bivariate normal pdf
satisfies the following conditions.
Theorem 7.3.1 A bivariate normal pdf given by (7.53) satisfies the following
condition:
                                         Z þ1 Z þ1                                                                          ð7:55Þ
                                                        f XYðx; yÞdxdy ¼ 1
                                           À1 À1
   This is a condition that any pdf must satisfy and is not a special condition for a
bivariate normal pdf. In other words, by the definition of a pdf, the integral of a pdf
over the entire real line(s) must be one.
Proof Rewrite (7.53) as follows:
                                      f  XY  ðx;  yÞ  ¼     2π σ X σ Y   p1 ffiffiffiffiffiffiffiffiffiffiffiffiffi   eEðx;yÞ                         ð7:56Þ
                                                                            1 À ρ2
where the exponent is given by the following equation:
                      1     (       À μX 2         2ρðx     À  μX Þðy    À  μY  Þ           À μY    2)
                      À          x  σX                         σXσY                      y  σY
Eðx;  yÞ  ¼  À2ð1                               À                                  þ
                         ρ2Þ
             1     (       À μX 2           ðx  À μXÞ    ρðy  À   μY  Þ            À μY 2                À μY   2           À μY   2)
             À          x  σX                   σX            σY            ρ2 y   σY              ρ2 y  σY              y  σY
¼  À2ð1                               À  2                               þ                     À                    þ
                ρ2Þ
¼  À2ð1      1     "&                 À         À μY  '2       þ  À      À  ρ2Á y      À μY  2#
             À            x  À μX        ρy     σY                 1                   σY
                             σX
                ρ2Þ
                                                                                                                            ð7:57Þ
   Substituting (7.57) into (7.56), we obtain the following equation:
198 7 Gaussian Distributions
                                                    n                                   o2                                  2  !
                       p1 ffiffiffiffiffiffiffiffiffiffiffiffiffieÀ2ð1À1 ρ2Þ       xÀμX      À     ρ     yÀμY           þ   ð1  À    ρ2  Þ     yÀμY
                                                           σX                   σY                                    σY
f XYðx; yÞ   ¼
                2πσXσY 1 À ρ2
                                             n                                     o2                                 2
                                                      xÀμX                  yÀμY                               yÀμY
             ¼ pffiffiffiffiffi  p1 ffiffiffiffiffiffiffiffiffiffiffiffiffieÀ2ð1À1 ρ2Þ     σX      À      ρ      σY               p1 ffiffiffiffiffieÀ21      σY
                2πσX 1 À ρ2                                                                σY 2π                                     2
                                                      n                                          o2
             ¼ pffiffiffiffiffi  p1 ffiffiffiffiffiffiffiffiffiffiffiffiffieÀ2ð1À1ρ2Þσ2X   ðx   À  μX     Þ  À    ρσσXY ðy  À   μY  Þ          p1 ffiffiffiffiffi  eÀ12     yÀμY
                                                                                                                               σY
                2πσX 1 À ρ2                                                                           σY 2π
                                                n                                         ρσσXY μY  o2                                   2
                pffiffiffiffiffiffiffiffiffiffip1 ffiffiffiffiffiffiffiffiffiffiffiffiffieÀ2ð1À1ρ2Þσ2X x      ρσσXY y                                                           yÀμY
             ¼                                              À               þ  μX      À                       p1 ffiffiffiffiffi  eÀ12      σY
                2πσX 1 À ρ2                                                                          2     σY 2π
             ¼ pffiffiffiffiffi  p1 ffiffiffiffiffiffiffiffiffiffiffiffiffieÀ2ð1À1ρ2Þσ2Xfx À Ag2             p1 ffiffiffiffiffi   eÀ12      yÀμY
                                                                                               σY
                2πσX 1 À ρ2         & '2                              σY 2π
                       p1 ffiffiffiffiffiffiffiffiffiffiffiffiffieÀ21    pxÀffiAffiffiffiffiffiffi                             yÀμY    2
                                                                                        σY
             ¼ pffiffiffiffiffi                        σX 1Àρ2              p1 ffiffiffiffiffi    eÀ12
                2πσX 1 À ρ2                                    σY 2π
                                                                                                                                  ð7:58Þ
where
                               A ¼ ρσσXY y þ μX À ρσσXY μY                                                                        ð7:59Þ
Substituting (7.58) into (7.55), we obtain the following equation:
                               Z þ1 8<Z þ1                                                  n o2             9                2
                                                                                                             =
Z þ1   Z þ1                                                           p1 ffiffiffiffiffiffiffiffiffiffiffiffiffieÀ21     pxÀffiAffiffiffiffiffi  dx;       eÀ12   yÀμY
                                                                                                                              σY
  À1      À1 f XY ðx; yÞdxdy ¼ À1 : À1                pffiffiffiffiffi                                  σX 1Àρ2                               dy
                                                                                                                           pffiffiffiffiffi
                                                            2πσX 1 À ρ2                                              σY 2π
                                                                                                                                  ð7:60Þ
   By (7.14), the integral in braces with respect to x is equal to 1 and the outer
integral with respect to y is also equal to 1 to yield the right-hand side of the above
to be equal to 1.
                                                                                            Q.E.D.
Theorem 7.3.2 If RVs X and Y are jointly normal
                               ðX;  YÞ       $        À        μY  ;  σ 2X ;   σY2  ;   Á                                         ð7:61Þ
                                                    N μX;                              ρ
   The marginal distributions of X and Y are normal with means and variances
μX, μY, σX2 and σ2Y, respectively, as follows:
                                                                                                              2
       X  $    À       σ 2X Á                                  f  X ðxÞ       ¼        p1 ffiffiffiffiffi   eÀ12     xÀμX                   ð7:62Þ
             N μX;                                                                  σX 2π                   σX
7.3 Two Jointly Gaussian RVs                                                                                                   199
                                                                                                            2
        Y  $    À   ;  σ  2  Á                                 f       Y ðxÞ     ¼       p1 ffiffiffiffiffi  eÀ21  yÀμY           ð7:63Þ
              N μY        Y                                                           σY 2π               σY
   This theorem states that, if two RVs X and Y are jointly normal, X and Y are
individually normal. However, the converse is not necessarily true. In other words,
two normally distributed RVs X and Y are not necessarily jointly normal.
Proof By (4.60), the marginal pdf of X is given by
           Z þ1
f YðyÞ ¼           f XYðx; yÞdx
               À1
                                                              &               2                                          2  '
           Z þ1
        ¼                       p1 ffiffiffiffiffiffiffiffiffiffiffiffiffi   À2ð1À1 ρ2        xÀμX            À    2ρðxÀμXÞðyÀμY Þ     þ    yÀμY
                                                              Þ σX                              σXσY            σY dx
                                                  e
              À1 2πσXσY 1 À ρ2
   Through the process of (7.58) through (7.60), the bivariate normal pdf in the
above integral can be expressed by (7.58). Substituting (7.58) into the above
equation, we obtain the following equation:
                                                       & '2                                               2  !
           Z1                  p1 ffiffiffiffiffiffiffiffiffiffiffiffiffi   À12     pxÀffiAffiffiffiffiffiffi         p1 ffiffiffiffiffieÀ12     yÀμY
                 pffiffiffiffiffi                                                                            σY dx
f YðyÞ  ¼                                        e      σX 1Àρ2
        ¼
           À1 2πσX 1 À ρ2                                                 σY 2π                         '2
                                                                                 &
                                    2
                             yÀμY      Z  1                                    À21         pxÀffiAffiffiffiffiffiffi
              p1 ffiffiffiffiffieÀ21    σY                        p1 ffiffiffiffiffiffiffiffiffiffiffiffiffi
                                            pffiffiffiffiffi                            e          σX 1Àρ2            dx          ð7:64Þ
           σY 2π                          À1 2πσX 1 À ρ2
   The integrand pinffiffiffiffiffitffihffiffieffiffiffiffiffilast integral is in the form of a normal pdf with
μ ¼ A and σ ¼ σX 1 À ρ2. Therefore, as proven by (7.14), the integral is equal
to 1 and (7.64) becomes
                                                                                2
                                       f  Y ðyÞ      ¼     p1 ffiffiffiffiffi    eÀ12  yÀμY
                                                        σY 2π                 σY
This proves (7.63). Eq. (7.62) is proven similarly.
Theorem 7.3.3 If RVs X and Y are jointly normal
                                   ðX;           YÞ  $    À    μY      ;  σX2 ;  σ2Y  ;   Á                             ð7:65Þ
                                                        N μX;                            ρ
ρ is the correlation coefficient of X and Y, and the covariance of X and Y is given by
                                                               qffiffiffiffiffiffiffiffiffiffi
                                cXY ¼ CovðX; YÞ ¼ ρ σ2XσY2 ¼ ρσXσY                                                      ð7:66Þ
200 7 Gaussian Distributions
                                                      ρ   ¼    cXY                                                         ð7:67Þ
                                                              σXσY
Proof By (5.33), we have the following equation:                                                                           ð7:68Þ
                             cXY ¼ CovðX; YÞ ¼ EðXYÞ À μXμY
          Z þ1Z þ1
EðXYÞ ¼          xyf XYðx; yÞdxdy
          À1 À1
                                                                 &            2                                    2  '
          Z þ1Z þ1                 p1 ffiffiffiffiffiffiffiffiffiffiffiffiffi   À2ð1À1 ρ2        xÀμX      À  2ρðxÀμXÞðyÀμY Þ  þ    yÀμY
       ¼ xy                                                      Þ σX                      σXσY
                                                     e                                                      σY dxdy
          À1 À1 2πσXσY 1 À ρ2
                                                                                                                           ð7:69Þ
Through the process of (7.58) through (7.60), the last integral can be expressed
as the following where A is given by (7.59)
           Z þ1 <8Z þ1                                           n                o2       9                                2
EðXYÞ ¼ À1 : À1                                    px ffiffiffiffiffiffiffiffiffiffiffiffiffieÀ21 σX   pxÀffiAffiffiffiffiffi    =
                                   pffiffiffiffiffi                                                dx;σ     py ffiffiffiffiffi  eÀ12     yÀμY      dy
                                     2πσ              1 À ρ2                    1Àρ2                2π                σY
                                                X                                              Y
                                                                                                                           ð7:70Þ
   Through the process of (7.21) through (7.24), the integral in braces in the above
integral is the expected value of X given Y fixed at y as follows:
                    Z þ1                                               n o2
                            pffiffiffiffiffi
                                                px ffiffiffiffiffiffiffiffiffiffiffiffiffi eÀ21   pxÀffiAffiffiffiffiffi   dx ¼ A
                                                                       σX 1Àρ2
                    À1 2πσX 1 À ρ2
where A is given by (7.59)                                                                                                 ð7:71Þ
                                  A ¼ ρσσXY y þ μX À ρσσXY μY
Substituting (7.71) into (7.70), we obtain
          Z þ1   &                                 '                   yÀμY   2
                   ρσσXY y                           σY                 σY
EðXYÞ  ¼    À1              þ  μX  À  ρσσXY μY           py ffiffiffiffiffieÀ21            dy
                                                          2π
                                                       2                                                                     2
                 Z þ1                           yÀμY             &                  '    Z þ1                         yÀμY
             σX                py2ffiffiffiffiffieÀ12      σY                μX        ρσσXY μY                py ffiffiffiffiffieÀ12      σY
       ¼  ρ  σY    À1                                     dy  þ        À                   À1                                   dy
                            σY 2π                                                                 σY 2π
                                                                                                                           ð7:72Þ
7.4 Vector Gaussian RV                                                                201
   By (7.32), the first and the second integral in the above equation are, respec-
tively, given by the following:
          Z þ1                             2
            À1             py2ffiffiffiffiffi     yÀμY           EÀY2Á
                        σY 2π            σY
                                     e dyÀ21        ¼          ¼  σY2  þ    μY2       ð7:73Þ
                                                                                      ð7:74Þ
                 Z þ1                            2
                   À1         py ffiffiffiffiffi        yÀμY
                           σY 2π               σY
                                        e dyÀ21        ¼   EðYÞ   ¼  μY
Substituting (7.73) and (7.74) into (7.72), we have
EðXYÞ  ¼  ρσσXY  Àσ2Y   þ  μY2  Á    þ                     μY  ¼  ρσ X σ Y  þ μX  μY  ð7:75Þ
                                          μX  À ρσσXY μY
Substituting (7.75) into (7.68), we have
cXY ¼ CovðX; YÞ ¼ EðXYÞ À μXμY ¼ ρσXσY þ μX μY À μXμY ¼ ρσXσY
The above proof also proves that the coefficient ρ is the correlation coefficient
                                        ρ     ¼   cXY
                                                 σXσY
Hence, both (7.66) and (7.67) are proven.
                                                                                      Q.E.D.
7.4 Vector Gaussian RV
In the general discussion of RVs in earlier chapters, the RVs are not a function of
time. However, in this section, we will consider the RVs of the specific time points
of a stochastic process. The vector RVs discussed in this section follow the
definitions given in Sect. 6.3 and, therefore, have the time points as the arguments.
   The vector RV X(t1, t2, . . . , tn) is said to be jointly normal if its n-dimensional
joint pdf is given by the following equation:
   f Xðt1, t2, ..., tnÞðx1, x2, . . . :, xn; t1, t2, . . . , tnÞ ¼ pðffiffi2ffiffiffiπffiffiÞffiffin1ffiffidffiffiffieffiffitffiffiffiffiXffiffiffiXffiffi eÀ12ðxÀμXÞTÀXX1 ðxÀμXÞ
                                                                                             ð7:76Þ
where the vector of specific values x and the vector RV X are as defined by (6.32a)
and (6.35), respectively, and the mean vector μX and the autocovariance matrix XX
202 7 Gaussian Distributions
are given by (6.164a) and (6.168), respectively. By (2.33) and (2.41), we have the
following equations for the terms occurring in (7.76):
                 detXX≜ X ðÀ1ÞtðjÞC1j1 C2j2 : : :Cnjn                                                     ð7:77Þ
                                    j
                             ÀXX1    ¼     1            adjXX                                            ð7:78Þ
                                         detXX
where adj XX is defined by (2.40) and t( j) is the total number of inversions of the
permutation j defined by (2.34).
   Consider the dimensions of the matrices in the following matrix multiplication
term in the exponent in (7.76):
                             ðx À μXÞTXÀX1 ðx À μXÞ                                                       ð7:79Þ
The following are the dimensions of the three matrices in the above multiplication:
ðx À μXÞT 1  n                       XÀX1 n  n                          ðx À μXÞ n  1
   The multiplication of ðn  nÞ matrix XÀX1 and ðn  1Þ matrix ðx À μXÞ yields
a n  1 column vector. Therefore, multiplication of ð1  nÞ row vector ðx À μXÞT
and ðn  1Þ column vector fÀXX1 ðx À μXÞg yields a scalar. The determinant detXX
is a scalar. Therefore, the multivariate joint normal pdf given by (7.76) is a scalar
function.
   By evaluating the matrix multiplication as discussed above, we may rewrite the
multivariate joint pdf given by (7.76) as follows with the exponent in an expanded form:
            f Xðt1, t2, ..., tnÞðx1, x2, . . . :, xn; t1, t2, . . . , tnÞ
                                                         Pn Pn
            ¼ pffiðffi2ffiffiffiπffiffiÞffiffin1ffiffidffiffieffiffiffitffiffiffiffiXffiffiffiXffiffi eÀ12Âdet1XX i¼1      ðxi Àμi ÞAij ðxj Àμj Þ
                                                                                                           ð7:80Þ
                                                                       j¼1
where detXX and cofactor    Aij are  gYiÀvetn10 ;  by   t(0mÀ7Á.7a7re)  and (2.39), respectively.
   Two vector RVs Xðt1; ::;  tnÞ and                ::;                  jton;int10tl;y::G; tam0 uÁssidaenfi,nifedtheb(ym(þ6.3n3))-
dimensional concatenated     vector   RV ZXY t1; ::;
consisting of m and n RVs have the following (m þ n)-dimensional multivariate
normal pdf
            f ZXYðt1, ::, tn, t10 , ::, tm0 Þðx1, ::, xn, y1, ::, ym; t1, ::, tn, t10 , ::, tm0 Þ          ð7:81Þ
               ¼ pðffiffi2ffiffiffiπffiffiÞffiffinffiffidffiffi1ffieffiffitffiffiffiffiZffiffiffiXffiffiYffiffiZffiffiXffiffiYffi eÀ21ðzXYÀμZXY ÞTÀZX1 YZXY ðzXYÀμZXY Þ
where zXY is the vector of specific values of the vector RV ZXY defined by (6.35) and
μZXY and ZXYZXY are defined by (6.165) and (6.191), respectively.
7.5 Characteristic Function of a Gaussian RV                                           203
Example 7.4.1                                                          "#         ð7:82Þ
Consider the case of n ¼ 3.                                               μ1
                        " X1 #                        μX ¼ EðXÞ ¼ μμ23
                  X ¼ X2
                           X3
              XX  ¼   n        À  μXÞðX  À        o  ¼  EÀXXTÁ  À  μXμXT
                      E ðX                    μXÞT
where
                " X1            #     X2          2       X1X2            3
       XXT ¼ X2                  ½X1                X1X1  X2X2      X1X3
                                                                    X2X3 5
                                          X3 ¼ 4 X2X1
                      X3 X3X1 X3X2 X3X3
                                                  2                       3
                      "#                            μ1μ1      μ1μ2  μ1μ3
                      μ1                                      μ2μ2  μ2μ3 5
       μXμXT ¼        μμ23 ½ μ1       μ2  μ3  ¼ 4 μ2μ1       μ3μ2  μ3μ3
                                                    μ3μ1
       82                                                                              39
       < ðX1 À μ1ÞðX1 À μ1Þ ðX1 À μ1ÞðX2 À μ2Þ ðX1 À μ1ÞðX3 À μ3Þ =
       ¼ E:4                                                                           5
XX       2   ðX2 À μ2ÞðX1 À μ1Þ          ðX2 À μ2ÞðX2 À μ2Þ  ðX2   À  μ2ÞðX3  À  μ3Þ   ;
              ðX3 À μ3Þð3X1 À μ1Þ         ðX3 À μ3ÞðX3 À μ3Þ  ðX3   À  μ3ÞðX1  À  μ1Þ
       c11 c12 c13
       ¼ 4 c21 c22 c23 5
       c31 c32 c33
                                                                                  ð7:83Þ
7.5 Characteristic Function of a Gaussian RV
This section derives the characteristic function of a Gaussian RV first for a scalar
RV and then for a vector RV.
7.5.1 Characteristic Function of a Scalar Gaussian RV
We first consider a scalar RV X. The characteristic function of the normal distri-
bution is, by definition, given by
204 7 Gaussian Distributions
          EÈejωX É     Z þ1        ejωx p1ffiffiffiffiffi  e ð ÞÀ21xÀμ2        Z þ1   p1ffiffiffiffiffi  e ð ÞjωxÀ21xÀμ2
                                       σ 2π       σ                        σ 2π        σ
ΨXðωÞ  ¼            ¼    À1                                   dx  ¼    À1                            dx
                                                                                           ð7:84Þ
   To evaluate the above integral, complete the square in the exponent in the
integral as follows:
       ¼¼¼¼¼¼¼jωxÀÀÀÀÀÀÀÀ12222222111111σσσσσσ21x222222 ÂnnnÀnÀxxxÂÂÂx2xxxÀ2ÀðσÀjÀÀÀÀωÀμσjjσ 2ÀÀÀωω2jjjÀ2ωωω2σjþ¼ωσ2σσσ2σþ222μxj2ωÞþþþÀμ!þxÁ2μμμ2ÃÀμþÁÁÁ2μÃÃÃÁx2À222xj1ωσþÀþþþ122μÀωμωωÀÀxÀj22j222ωωÁσσσ12Àσ2σ44¼ω2þ2ÀÀ22þÀþμσj22x2ω2μμjj1σþωωμÁÁ222μμÈμþσσÀx222Á2μÀÀÀÀ2jωo2σμμÀ222jωþþþσμ2μÁμ2þ22oðþμþÁμxμ2Þþo2oμ2É
   Substituting the last expression in the exponent in (7.84), we obtain the follow-
ing equation:
                Z þ1       p1ffiffiffiffiffieÀ12    xÀðjωσ2þμÞ!2        þ   jωμ  À  21ω2σ2
              ¼
       ΨXðωÞ                                     σ                                dx
                    À1 σ 2π                                xÀðjωσ2þμÞ!2
                           ω2 σ 2  Z  þ1  p1ffiffiffiffiffieÀ21             σ
                             2
              ¼     ejωμÀ                                                dx                ð7:85Þ
                                      À1 σ 2π
   By (7.15) through (7.19), we have shown that the last integral, which is the
Gaussian integral, is equal to 1. Therefore, we have the following characteristic
function of the normal scalar RV X:
                                   ψXðωÞ ¼ ejωμÀω22σ2                                      ð7:86Þ
7.5.2 Characteristic Function of a Gaussian Vector RV
In this section, the time points as the arguments of the RV X(t1, .., tn) are not shown
for simplicity. By the definition given by (6.62), the characteristic function of a
Gaussian vector RV X is as follows:
7.5 Characteristic Function of a Gaussian RV                                                                            205
             ΨXðω1, ω2, . . . , ωnÞ≜EfejωTXg ¼ Efejðω1X1þω2X2þÁÁÁþωnXnÞg                                              ð7:87Þ
    Z þ1              Z þ1                                                                                            ð7:88Þ
¼            .  .  .        ejωTxpðffiffi2ffiffiffiπffiffiÞffiffin1ffiffidffiffiffieffiffitffiffiffiffiffiXffiffiXffiffi  eÀ21ðxÀμXÞTXÀX1 ðxÀμXÞdx1dx2  .  .  .  dxn
      À1                À1
   To evaluate the above integral, let us first evaluate the following simpler integral
and then return to complete the above integral:
                                    Z þ1 Z þ1                                                                         ð7:89Þ
        ΨXðω1; ω2; . . . ; ωnÞ ¼ . . . ejωTxeÀ12xTxdx1dx2 . . . dxn
                                                  À1 À1
   Consider the second exponent without -½ in the above integral, xTx, in which
 is assumed to be an (n  n) symmetric matrix. Since À1 ¼ , by
pre-multiplying and post-multiplying  by À1, we have the following expression:
                      xTx  ¼  xT  ÀℙℙÀ1  ÁÀℙℙÀ1                        Á   ¼  xTℙÀℙÀ1ℙÁℙÀ1x
                                                                          x
    In  the  above    equation,    we  choose  ℙ                         to  diagonalize    so  that      ÀℙÀ1      Á  ¼  ,
                                                                                                                 
yielding the following expression:
                                       xTx ¼ xTℙÀ1x                                                                ð7:90Þ
where
                                    ¼ ℙÀ1, det ℙ ¼6 0                                                              ð7:91Þ
    is a diagonal matrix with the eigenvalues of  as the diagonal elements as
follows:
                            23                                                                                        ð7:92Þ
                              λ1 Á Á Á 0
                       ¼ 4 ⋮ ⋱ ⋮ 5 , λi ¼ eigenvalue, i ¼ 1, . . . , n:
                               0 Á Á Á λn
    From (7.91), we have
                                           ¼ ℙÀ1                                                                   ð7:93Þ
   Taking the transposition of both sides of the above and using the matrix identity
(2.20d), we have
                      T ¼ ÀℙℙÀ1ÁT ¼ ÀℙÀ1ÁTTℙT ¼ ÀℙÀ1ÁTℙT                                                          ð7:94Þ
   Since  is symmetric, T ¼ , and, so, by equating (7.93) with (7.94), we obtain
the following equation:
                                       ℙÀ1 ¼ ÀℙÀ1ÁTℙT                                                              ð7:95Þ
or
206 7 Gaussian Distributions
                ¼ ℙÀ1ÀℙÀ1ÁTℙTℙ                                  ð7:96Þ
   From the above, we derive the following relationship that holds true with the
diagonalizing matrix ℙ:
                                             ℙTℙ ¼ 
By left multiplying both sides, we obtain                         ð7:97Þ
                                        ℙÀ1 ¼ ℙT
   The above result shows that the matrix ℙ diagonalizing  is an orthogonal matrix
if  is a symmetric matrix. In (7.90), let
                     23
                              z1
                     6466666      7777577
               z  ¼           z2           ¼  À1x                ð7:98Þ
                              :
                              zi
                              :
                              zn
Then, we have
               zT ¼ ÀℙÀ1xÁT ¼ xTÀℙÀ1ÁT                            ð7:99Þ
Substituting (7.97) into the above equation, we have              ð7:100Þ
                 zT ¼ ½ z1 z2 : zi : zn  ¼ xTÀℙTÁT ¼ xTℙ
Substituting (7.98) and (7.100) into (7.90), we obtain the following equation:
                                    Xn                            ð7:101Þ
               xTx ¼ zTz ¼ λizi2
                                              i¼1
   Following the method described in Sect. 2.2.6, we can find the diagonalizing
matrix ℙ as follows. First, find n eigenvalues of  by solving the following
equation:
               det ð À λÞ ¼ 0                                   ð7:102Þ
   Then, obtain the eigenvectors corresponding to the eigenvalues from the fol-
lowing n linearly independent relationships:
               bi ¼ λibi, i ¼ 1, . . . :, n                      ð7:103Þ
where the components of the eigenvectors are denoted as follows:
7.5 Characteristic Function of a Gaussian RV                                                                                                               207
                                       23
                                                b1i
                                       6666664         7777577,
                                bi  ¼           b2i              i  ¼      1,   :      :  :, n                                                             ð7:104Þ
                                                 :
                                                bki
                                                 :
                                                bni
   For the symmetric matrix , its diagonalizing matrix ℙ is orthonormal. The
eigenvectors are orthogonal to each other and their norms are unity. Therefore, the
inner products, h., .i, of the eigenvectors are given by
                                    Xn                                                                                                                     ð7:105Þ
                                 bi; bj ¼ bkibkj ¼ δij                                                                                                     ð7:106Þ
                                                                                                                                                           ð7:107Þ
                                                  k¼1
The diagonalizing matrix ℙ is given by
                                                                    23
                                                                           b11 : b1i : b1n
                                                                    64666                                                                           77775
      ℙ ¼ ½b1               b2 : :         :    bn    ¼                    :          :      :                       :       :
                                                                                       :                              :
                                                                           bk1         :     bki                      :      bkn
                                                                            :                 :                               :
                                                                           bn1 : bni : bnn
                                                2                                                                     3
                                                       b11          : bk1 : bn1
                                                66466                                                                 57777
                        ℙT  ¼   ℙÀ1    ¼                :           :           :         :                :
                                                                    :                     :
                                                       b1i          :          bki        :               bni
                                                        :                       :                          :
                                                       b1n : bkn : bnn
Substituting (7.107) into (7.98), the transformation of x by ℙT to z is given by
      23                            2                                                                     32 3
             z1 b11 : bj1 : bn1 x1
      46666      77757              66646                                                                 7777566664         77577
z  ¼         :          ¼  ℙTx  ¼           :          :             :          :          :                          :
                                                       :                        :
             zi                            b1i         :            bji         :         bni                         xj
             :                              :                        :                     :                          :
             zn b1n : bjn : bnn xn
                                                       2 Xn 3
   ¼  2      b11x1      þ   ::  þ   bj1xj       þ  ::  þ            bn1xn       3      ¼  66664666666666      bj1xj                 57777777777777         ð7:108Þ
      66664  b1ix1      þ   ::  þ     :         þ  ::  þ            bnixn       57777
             b1nx1      þ   ::  þ               þ  ::  þ            bnnxn                                 j¼1 :
                                    bjixj
                                      :                                                                       :
                                                                                                          Xn
                                    bjnxj
                                                                                                              bjixj
                                                                                                          j¼1 :
                                                                                                              :
                                                                                                          Xn
                                                                                                              bjnxj
                                                                                                          j¼1
Now, consider the first exponent in (7.89)