Page 459 - A First Course In Stochastic Models
P. 459

454                           APPENDICES

                that P (z) can be represented as
                                                r 0
                                        P (z) =      + U(z)                    (C.9)
                                               z − z 0
                in |z| < R 0 , z 	= z 0 . Here U(z) is an analytic function in the domain |z| < R 0 and
                the residue r 0 = H(z 0 ) is given by

                                                             ′
                                 r 0 = lim (z − z 0 )P (z) = N(z 0 )/D (z 0 ).
                                     z→z 0
                The remainder of the proof is simple. Since U(z) is analytic for |z| < R 0 we have
                                                 
 ∞     j
                                                      u
                the power series representation U(z) =  j=0 j z for |z| < R 0 . Let R 1 be any
                                                             ! !     −j
                number with z 0 < R 1 < R 0 . Then, for some constant b, u j ≤ bR  for all j ≥ 0.
                                                             ! !
                                                                     1
                                                           j
                This follows from the fact that the series  
 j=0 j z is convergent for z = R 1 .
                                                     ∞
                                                        u
                Using the power series representation of U(z) and the fact that the power series
                                  
       j
                                    ∞
                representation P (z) =  p j z extends to |z| < z 0 , it follows from (C.9) that
                                    j=0
                             ∞             ∞          ∞
                                   j              j        j
                                      −r 0
                                p j z =      (z/z 0 ) +  u j z ,  |z| < z 0 .
                                       z 0
                            j=0           j=0        j=0
                Equating coefficients yields
                                             −j−1
                                     p j = −r 0 z  + u j ,  j ≥ 0.
                                             0
                     ! !     −j
                Since u j ≤ bR  for some constant b and R 1 > z 0 , the coefficient u j tends to
                             1
                     ! !
                              −j
                zero faster than z  . Hence we can conclude the asymptotic expansion (C.7).
                              0
                  It is noted that Theorem C.1 does not require that {p j } is a probability distri-
                bution. The theorem applies to any sequence {p j , j = 0, 1, . . . } with p j ≥ 0 for
                all j and  
 j=0 j < ∞. The asymptotic expansion (C.7) is very useful for both
                          ∞
                             p
                theoretical and computational purposes. It appears that in many applications the
                asymptotic expansion for p j can be used for relatively small values of j. To illus-
                trate this, consider the generating function (C.5) for the problem of success runs.
                                                                                s s
                This generating function P (z) is the ratio of the two analytic functions N(z) = p z
                             
 s   k−1       k
                and D(z) = 1−     p   (1−p)z whose domains of definition can be extended
                               k=1
                to the whole complex plane (R = ∞). It is readily verified that the equation
                                          s
                                              k−1       k
                                      1 −    p   (1 − p)x = 0
                                         k=1
                has a unique root z 0 on the interval (1, ∞). Hence condition C1 is satisfied. The
                verification of the technical conditions C2 and C3 is omitted and is left to the
                interested reader. The unique root z 0 of the above equation must be numerically
   454   455   456   457   458   459   460   461   462   463   464