the matrix geometric analytic methods for structured
play

The Matrix Geometric/Analytic Methods for Structured Markov Chains - PowerPoint PPT Presentation

William J. Stewart Department of Computer Science N.C. State University The Matrix Geometric/Analytic Methods for Structured Markov Chains Markov chains whose transition matrices have a special block structure. Example: B 00 B 01 0


  1. William J. Stewart Department of Computer Science N.C. State University 2. We check that the system is stable by verifying Equation (8). The infinitesimal generator matrix   − 5 5 0   A = A 0 + A 1 + A 2 = 3 − 8 5     0 3 − 3 has stationary probability vector π A = ( . 1837 , . 3061 , . 5102) and . 4388 = π A A 2 e < π A A 0 e = 1 . 2245 17 SMF-07:PE Bertinoro, Italy

  2. William J. Stewart Department of Computer Science N.C. State University 3. We now initiate the iterative procedure to compute the rate matrix R . The inverse of A 1 is   − . 2466 − . 1598 − . 2283 A − 1   = − . 0959 − . 1918 − . 2740   1   − . 0822 − . 1644 − . 5205 which allows us to compute   − . 2466 − . 1598 − . 2283 V = A 2 A − 1   = 0 0 0   1   − . 0411 − . 0822 − . 2603   0 0 0 W = A 0 A − 1   =  . − . 3836 − . 7671 − 1 . 0959   1  0 0 0 18 SMF-07:PE Bertinoro, Italy

  3. William J. Stewart Department of Computer Science N.C. State University Equation (5) becomes     . 2466 . 1598 . 2283 0 0 0    + R 2   R ( k +1) = 0 0 0 . 3836 . 7671 1 . 0959     ( k )    . 0411 . 0822 . 2603 0 0 0 and iterating successively, beginning with R (0) = 0 , we find     . 2466 . 1598 . 2283 . 2689 . 2044 . 2921     R (1) =  , R (2) =  , 0 0 0 0 0 0       . 0411 . 0822 . 2603 . 0518 . 1036 . 2909   . 2793 . 2252 . 2921   R (3) =  , · · · 0 0 0    . 0567 . 1134 . 3049 Observe that the elements are non-decreasing. 19 SMF-07:PE Bertinoro, Italy

  4. William J. Stewart Department of Computer Science N.C. State University After 48 iterations, successive differences are less than 10 − 12 , at which point   . 2917 . 2500 . 3571   R (48) =  . 0 0 0    . 0625 . 1250 . 3214 4. Proceeding to the boundary conditions: 0 1 − 6 5 . 0 1 0 0 B C 3 − 3 . 5 0 0 . 5 0 1 B C @ B 00 B 01 B C A = ( π 0 , π 1 ) B C ( π 0 , π 1 ) = (0 , 0) 0 0 − 6 6 . 0 0 B C B 10 A 1 + RA 0 B C B C 2 2 3 − 12 . 0 5 . 0 B C @ A 0 0 0 3 . 5 − 3 . 5 20 SMF-07:PE Bertinoro, Italy

  5. William J. Stewart Department of Computer Science N.C. State University Solve this by replacing the last equation with π 0 1 = 1 , i.e., set the first component of the subvector π 0 to 1. 0 1 − 6 5 . 0 1 0 1 B C 3 − 3 . 5 0 0 0 B C B C B C ( π 0 , π 1 ) = (0 , 0 | 0 , 0 , 1) 0 0 − 6 6 . 0 0 B C B C B C 2 2 3 − 12 . 0 0 B C @ A 0 0 0 3 . 5 0 with solution ( π 0 , π 1 ) = (1 . 0 , 1 . 6923 , | . 3974 , . 4615 , . 9011) Now on to the normalization stage. 21 SMF-07:PE Bertinoro, Italy

  6. William J. Stewart Department of Computer Science N.C. State University 5. The normalization constant is π 0 e + π 1 ( I − R ) − 1 e α =   1 . 4805 . 4675 . 7792   = (1 . 0 , 1 . 6923) e + ( . 3974 , . 4615 , . 9011)  e 0 1 0    . 1364 . 2273 . 15455 = 2 . 6923 + 3 . 2657 = 5 . 9580 which allows us to compute π 0 /α = ( . 1678 , . 2840) and π 1 /α = ( . 0667 , . 0775 , . 1512) 22 SMF-07:PE Bertinoro, Italy

  7. William J. Stewart Department of Computer Science N.C. State University 6. Subcomponents of the stationary distribution: — computed from π k = π k − 1 R .   . 2917 . 2500 . 3571   π 2 = π 1 R = ( . 0667 , . 0775 , . 1512) 0 0 0     . 0625 . 1250 . 3214 = ( . 0289 , . 0356 , . 0724) and   . 2917 . 2500 . 3571   π 3 = π 2 R = ( . 0289 , . 0356 , . 0724) 0 0 0     . 0625 . 1250 . 3214 = ( . 0130 , . 0356 , . 0336) and so on. 23 SMF-07:PE Bertinoro, Italy

  8. William J. Stewart Department of Computer Science N.C. State University Block Lower-Hessenberg Markov Chains   B 00 B 01 0 0 0 0 0 · · ·   B 10 B 11 A 0 0 0 0 0 · · ·       B 20 B 21 A 1 A 0 0 0 0 · · ·     B 30 B 31 A 2 A 1 A 0 0 0 · · ·   Q =     B 40 B 41 A 3 A 2 A 1 A 0 0 · · ·     . ... ... ... ... ...   .   . · · ·    . . . . . . .  . . . . . . . . . . . . . . · · · Transitions are now permitted from any level to any lower level. Objective:compute the stationary probability vector π from πQ = 0 . π is partitioned conformally with Q , i.e. π = ( π 0 , π 1 , π 2 , · · · ) — π i = ( π ( i, 1) , π ( i, 2) , · · · π ( i, K )) . 24 SMF-07:PE Bertinoro, Italy

  9. William J. Stewart Department of Computer Science N.C. State University A matrix geometric solution exists which mirrors that of a QBD process,. There exists a positive matrix R such that π i = π i − 1 R, for i = 2 , 3 , . . . i.e., that π i = π 1 R i − 1 , for i = 2 , 3 , . . . From πQ = 0 ∞ � π k + j A k = 0 , j = 1 , 2 , . . . k =0 and in particular, ∞ � π 1 A 0 + π 2 A 1 + π k +1 A k = 0 k =2 25 SMF-07:PE Bertinoro, Italy

  10. William J. Stewart Department of Computer Science N.C. State University Substituting π i = π 1 R i − 1 into ∞ � π 1 R k A k = 0 π 1 A 0 + π 1 RA 1 + k =2 gives � ∞ � � R k A k π 1 A 0 + RA 1 + = 0 k =2 So find R from ∞ � R k A k = 0 A 0 + RA 1 + (9) k =2 Equation (9) reduces to Equation (4) when A k = 0 for k > 2 . 26 SMF-07:PE Bertinoro, Italy

  11. William J. Stewart Department of Computer Science N.C. State University Rearranging Equation (9), we find ∞ � R = − A 0 A − 1 R k A k A − 1 − 1 1 k =2 ∞ � R ( l +1) = − A 0 A − 1 ( l ) A k A − 1 R k R (0) = 0; − 1 , l = 1 , 2 , . . . 1 k =2 In many cases, the structure of the infinitesimal generator is such that the blocks A i are zero for relatively small values of i , which limits the computational effort needed in each iteration. 27 SMF-07:PE Bertinoro, Italy

  12. William J. Stewart Department of Computer Science N.C. State University Derivation of the initial subvectors π 0 and π 1 . From the first equation of πQ = 0 , ∞ � π i B i 0 = 0 i =0 and we may write � ∞ ∞ ∞ � � � � π 1 R i − 1 B i 0 = π 0 B 00 + π 1 R i − 1 B i 0 π 0 B 00 + π i B i 0 = π 0 B 00 + = 0 , i =1 i =1 i =1 (10) From the second equation of πQ = 0 , ∞ ∞ � � R i − 1 B i 1 = 0 . π 0 B 01 + π i B i 1 = 0 , i . e ., π 0 B 01 + π 1 (11) i =1 i =1 28 SMF-07:PE Bertinoro, Italy

  13. William J. Stewart Department of Computer Science N.C. State University In matrix form, we can compute π 0 and π 1 from   B 00 B 01   ( π 0 , π 1 )  = (0 , 0) .    � ∞ � ∞ i =1 R i − 1 B i 0 i =1 R i − 1 B i 1 Once found, normalize by dividing by � ∞ � � R k − 1 e = π 0 e + π 1 ( I − R ) − 1 e. α = π 0 e + π 1 i =1 For discrete-time Markov chains, replace − A − 1 with ( I − A 1 ) − 1 . 1 29 SMF-07:PE Bertinoro, Italy

  14. William J. Stewart Department of Computer Science N.C. State University Same example as before, but with additional transitions ( ξ 1 = . 25 and ξ 2 = . 75 ) to lower non-neighboring states. ξ ξ ξ ξ ξ 1 1 1 1 1 ξ λ 1 λ λ λ λ 1 1 1 1 1 5,1 1,1 3,1 4,1 2,1 λ 1 γ γ γ γ γ γ γ γ γ 0,1 γ 1 1 1 1 2 1 2 2 2 2 µ µ µ µ µ γ γ 1,2 2,2 3,2 4,2 5,2 1 2 γ γ γ γ γ γ γ γ γ 0,2 1 1 γ 1 1 1 2 2 2 2 2 λ 2 1,3 3,3 4,3 2,3 5,3 λ λ λ λ λ 2 2 ξ 2 2 2 2 ξ ξ ξ ξ ξ 2 2 2 2 2 Figure 2: State transition diagram for a GI/M/1-type process. 30 SMF-07:PE Bertinoro, Italy

  15. William J. Stewart Department of Computer Science N.C. State University Q = 0 − 6 5 . 0 1 B 3 − 3 . 5 . 5 B B B − 6 5 1 B B B 2 . 00 2 . 00 3 − 12 5 B B B 3 − 3 . 5 . 5 B B B . 25 − 6 . 25 5 1 B B 4 3 . 00 − 12 5 . 00 B B B . 75 3 − 4 . 25 . 5 B B B . 25 − 6 . 25 5 1 B B B 4 3 . 00 − 12 5 . 00 B B B . 75 3 − 4 . 25 B B @ ... ... ... 31 SMF-07:PE Bertinoro, Italy

  16. William J. Stewart Department of Computer Science N.C. State University The computation of the matrix R proceeds as previously: 0 1 − . 2233 − . 1318 − . 1550 A − 1 B C = − . 0791 − . 1647 − . 1938 1 B C @ A − . 0558 − . 1163 − . 3721 which allows us to compute 0 1 0 1 − . 2233 − . 1318 − . 1550 0 0 0 A 0 A − 1 A 2 A − 1 B C B C = A , = 0 0 0 − . 3163 − . 6589 − . 7752 1 B C 1 B C @ @ A − . 0279 − . 0581 − . 1860 0 0 0 0 1 − . 0558 − . 0329 − . 0388 A 3 A − 1 B C = A , 0 0 0 B C 1 @ − . 0419 − . 0872 − . 2791 32 SMF-07:PE Bertinoro, Italy

  17. William J. Stewart Department of Computer Science N.C. State University The iterative process is 0 1 0 1 . 2233 . 1318 . 1550 0 0 0 A + R 2 B C B C R ( k +1) = 0 0 0 . 3163 . 6589 . 7752 B C ( k ) B C @ @ A . 0279 . 0581 . 1860 0 0 0 0 1 . 0558 . 0329 . 0388 + R 3 B C 0 0 0 ( k ) B C @ A . 0419 . 0872 . 2791 Iterating successively, beginning with R (0) = 0 , we find 0 1 0 1 . 2233 . 1318 . 1550 . 2370 . 1593 . 1910 B C B C R (1) = A , R (2) = A , 0 0 0 0 0 0 B C B C @ @ . 0279 . 0581 . 1860 . 0331 . 0686 . 1999 33 SMF-07:PE Bertinoro, Italy

  18. William J. Stewart Department of Computer Science N.C. State University 0 1 . 2415 . 1684 . 2031 B C R (3) = A , · · · 0 0 0 B C @ . 0347 . 0719 . 2043 34 SMF-07:PE Bertinoro, Italy

  19. William J. Stewart Department of Computer Science N.C. State University After 27 iterations, successive differences are less than 10 − 12 , at which point   . 2440 . 1734 . 2100   R (27) =  . 0 0 0    . 0356 . 0736 . 1669 The boundary conditions are now   B 00 B 01  = (0 , 0) ( π 0 , π 1 )  B 11 + RB 21 + R 2 B 31 B 10 + RB 20 35 SMF-07:PE Bertinoro, Italy

  20. William J. Stewart Department of Computer Science N.C. State University 0 1 − 6 . 0 5 . 0 1 0 0 B C 3 . 0 − 3 . 5 0 0 . 5 B C B C B C = ( π 0 , π 1 ) = (0 , 0) . . 0610 . 1575 − 5 . 9832 5 . 6938 . 0710 B C B C B C 2 . 0000 2 . 000 3 . 000 − 12 . 0000 5 . 0000 B C @ A . 0089 . 1555 . 0040 3 . 2945 − 3 . 4624 Solve this by replacing the last equation with π 0 1 = 1 . 0 1 − 6 . 0 5 . 0 1 0 1 B C 3 . 0 − 3 . 5 0 0 0 B C B C B C ( π 0 , π 1 ) = (0 , 0 | 0 , 0 , 1) . 0610 . 1575 − 5 . 9832 5 . 6938 0 B C B C B C 2 . 0000 2 . 000 3 . 000 − 12 . 0000 0 B C @ A . 0089 . 1555 . 0040 3 . 2945 0 Solution ( π 0 , π 1 ) = (1 . 0 , 1 . 7169 , | . 3730 , . 4095 , . 8470) 36 SMF-07:PE Bertinoro, Italy

  21. William J. Stewart Department of Computer Science N.C. State University The normalization constant is π 0 e + π 1 ( I − R ) − 1 e α =   1 . 3395 . 2584 . 3546   = (1 . 0 , 1 . 7169) e + ( . 3730 , . 4095 , . 8470)  e 0 1 0    . 0600 . 1044 1 . 2764 = 2 . 7169 + 2 . 3582 = 5 . 0751 Thus: π 0 /α = ( . 1970 , . 3383) , and π 1 /α = ( . 0735 , . 0807 , . 1669) . Successive subcomponents are now computed from π k = π k − 1 R . 37 SMF-07:PE Bertinoro, Italy

  22. William J. Stewart Department of Computer Science N.C. State University   . 2440 . 1734 . 2100   π 2 = π 1 R = ( . 0735 , . 0807 , . 1669) 0 0 0     . 0356 . 0736 . 1669 = ( . 0239 , . 0250 , . 0499) and   . 2440 . 1734 . 2100   π 3 = π 2 R = ( . 0239 , . 0250 , . 0499) 0 0 0     . 0356 . 0736 . 1669 = ( . 0076 , . 0078 , . 0135) and so on. 38 SMF-07:PE Bertinoro, Italy

  23. William J. Stewart Department of Computer Science N.C. State University Simplifications occur when the initial B blocks have the same dimensions as the A blocks and when 0 1 B 00 A 0 0 0 0 0 0 · · · B C B 10 A 1 A 0 0 0 0 0 · · · B C B C B C B 20 A 2 A 1 A 0 0 0 0 · · · B C B C B C B 30 A 3 A 2 A 1 A 0 0 0 · · · Q = B C B C B C B 40 A 4 A 3 A 2 A 1 A 0 0 · · · B C B C . ... ... ... ... ... . B C . · · · B C B C . . . . . . . @ A . . . . . . . . . . . . . . · · · In this case π i = π 0 R i , for i = 1 , 2 , . . . , � ∞ i =0 R i B i 0 is an infinitesimal generator matrix π 0 is the stationary probability vector of � ∞ i =0 R i B i 0 — normalized so that π 0 ( I − R ) − 1 e = 1 . 39 SMF-07:PE Bertinoro, Italy

  24. William J. Stewart Department of Computer Science N.C. State University Also, in some applications more than two boundary columns can occur. 0 1 B 00 B 01 B 02 A 0 B C B 10 B 11 B 12 A 1 A 0 B C B C B C B 20 B 21 B 22 A 2 A 1 A 0 B C B C B C B 30 B 31 B 32 A 3 A 2 A 1 A 0 B C B C B C B 40 B 41 B 42 A 4 A 3 A 2 A 1 A 0 B C Q = B C B B 50 B 51 B 52 A 5 A 4 A 3 A 2 A 1 A 0 C B C B C B 60 B 61 B 62 A 6 A 5 A 4 A 3 A 2 A 1 A 0 B C B C B C B 70 B 71 B 72 A 7 A 6 A 5 A 4 A 3 A 2 A 1 A 0 B C B C B C B 80 B 81 B 82 A 8 A 7 A 6 A 5 A 4 0 0 A 1 A 0 B C B C . . . ... ... ... ... ... ... ... ... ... @ A . . . . . . At present, this matrix is not block lower Hessenberg. 40 SMF-07:PE Bertinoro, Italy

  25. William J. Stewart Department of Computer Science N.C. State University Restructured into the form 0 1 B 00 B 01 B 02 A 0 B C B 10 B 11 B 12 A 1 A 0 B C B C B C B 20 B 21 B 22 A 2 A 1 A 0 B C B C B C B 30 B 31 B 32 A 3 A 2 A 1 A 0 B C B C B 40 B 41 B 42 A 4 A 3 A 2 A 1 A 0 B C B C B C B 50 B 51 B 52 A 5 A 4 A 3 A 2 A 1 A 0 B C B C B C B 60 B 61 B 62 A 6 A 5 A 4 A 3 A 2 A 1 A 0 B C B C B C B 70 B 71 B 72 A 7 A 6 A 5 A 4 A 3 A 2 A 1 A 0 B C B C B C B 80 B 81 B 82 A 8 A 7 A 6 A 5 A 4 0 0 A 1 A 0 B C B C @ A . . . . . . ... ... ... ... ... ... . . . . . . . . . . . . 0 1 0 1 0 1 A 0 A 3 A 2 A 1 B 00 B 01 B 02 B C B C B C A 0 = A , A 1 = A , B 00 = A , A 1 A 0 A 4 A 3 A 2 B 10 B 11 B 12 · · · B C B C B C @ @ @ A 2 A 1 A 0 A 5 A 4 A 3 B 20 B 21 B 22 41 SMF-07:PE Bertinoro, Italy

  26. William J. Stewart Department of Computer Science N.C. State University Block Upper-Hessenberg Markov Chains For QBD and GI/M/1-type processes, we posed the problem in terms of continuous-time Markov chains. Discrete-time Markov chains can be treated if the matrix inverse A − 1 is 1 replaced with the inverse ( I − A 1 ) − 1 . This time we shall consider the discrete-time case.   B 00 B 01 B 02 B 03 · · · B 0 j · · ·   B 10 A 1 A 2 A 3 · · · A j · · ·       0 A 0 A 1 A 2 · · · A j − 1 · · ·   P =     0 0 A 0 A 1 · · · A j − 2 · · ·     0 0 0 A 0 · · · A j − 3 · · ·     . . . . . . .   . . . . . . . . . . . . . . 42 SMF-07:PE Bertinoro, Italy

  27. William J. Stewart Department of Computer Science N.C. State University A = � ∞ i =0 A i is a stochastic matrix assumed to be irreducible. π A A = π A , and π A e = 1 . P is known to be positive-recurrent if � ∞ � � π A iA i e ≡ π A b < 1 . (12) i =1 We seek to compute π from πP = π . As before, we partition π conformally with P , i.e. π = ( π 0 , π 1 , π 2 , · · · ) where π i = ( π ( i, 1) , π ( i, 2) , · · · π ( i, K )) The analysis of M/G/1-type processes is more complicated than that of QBD or GI/M/1-type processes because the subvectors π i no longer have a matrix geometric relationship with one another. 43 SMF-07:PE Bertinoro, Italy

  28. William J. Stewart Department of Computer Science N.C. State University The key to solving upper block-Hessenberg structured Markov chains is the computation of a certain stochastic matrix G . G ij is the conditional probability that starting in state i of any level n ≥ 2 , the process enters level n − 1 for the first time by arriving at state j of that level. This matrix satisfies the fixed point equation ∞ � A i G i G = i =0 and is indeed is the minimal non-negative solution of ∞ � A i X i . X = i =0 It can be found by means of the iteration ∞ � A i G i G (0) = 0; G ( k +1) = ( k ) = 0 , k = 0 , 1 , . . . i =0 44 SMF-07:PE Bertinoro, Italy

  29. William J. Stewart Department of Computer Science N.C. State University Once the matrix G has been computed, then successive components of π can be found. From πP = π π ( I − P ) = 0 , 0 1 I − B 00 − B 01 − B 02 − B 03 − B 0 j · · · · · · B C − B 10 I − A 1 − A 2 − A 3 − A j · · · · · · B C B C B 0 − A 0 I − A 1 − A 2 − A j − 1 C · · · · · · B C ( π 0 , π 1 , · · · , π j , · · · ) B C 0 0 − A 0 I − A 1 − A j − 2 B C · · · · · · B C B C 0 0 0 − A 0 − A j − 3 B C · · · · · · B C @ A . . . . . . . . . . . . . . . . . . . . . (13) = (0 , 0 , · · · 0 , · · · ) . The submatrix in the lower right block is block Toeplitz. There is a decomposition of this Toeplitz matrix into a block upper triangular matrix U and block lower triangular matrix L . 45 SMF-07:PE Bertinoro, Italy

  30. William J. Stewart Department of Computer Science N.C. State University 0 1 0 1 A ∗ A ∗ A ∗ A ∗ I 0 0 0 · · · · · · 1 2 3 4 B C B C 0 A ∗ A ∗ A ∗ − G I 0 0 B C B C · · · · · · 1 2 3 B C B C B C B C A ∗ A ∗ 0 0 0 − G I 0 U = · · · and L = · · · . B C B C 1 2 B C B C B C B C A ∗ 0 0 0 0 0 − G I · · · · · · B 1 C B C B C B C @ A @ A . . . . . ... ... ... . . . . . . . . . . Once the matrix G has been formed then L is known. The inverse of L can be written in terms of the powers of G . 0 1 0 1 0 1 I 0 0 0 I 0 0 0 1 0 0 0 · · · · · · · · · B C B C B C − G I 0 0 G I 0 0 0 1 0 0 B · · · C B · · · C B · · · C B C B C B C B C B C B C G 2 0 − G I 0 G I 0 0 0 1 0 · · · · · · = · · · . B C B C B C B C B C B C G 3 G 2 B C B C B C 0 0 − G I G I 0 0 0 1 · · · · · · · · · B C B C B C B C B C B C @ A @ A @ A . . . . . . . ... ... ... ... ... . . . . . . . . . . . . . . 46 SMF-07:PE Bertinoro, Italy

  31. William J. Stewart Department of Computer Science N.C. State University From Equation (13), 0 1 I − B 00 − B 01 − B 02 − B 03 − B 0 j · · · · · · B C − B 10 B C B C B 0 C B C ( π 0 , π 1 , · · · , π j , · · · ) B C 0 UL B C B C B C 0 B C B C @ A . . . = (0 , 0 , · · · 0 , · · · ) which allows us to write π 0 ( − B 01 , − B 02 , · · · ) + ( π 1 , π 2 , · · · ) UL = 0 or π 0 ( B 01 , B 02 , · · · ) L − 1 = ( π 1 , π 2 , · · · ) U, 47 SMF-07:PE Bertinoro, Italy

  32. William J. Stewart Department of Computer Science N.C. State University 0 1 I 0 0 0 · · · B C G I 0 0 · · · B C B C B C G 2 G I 0 · · · π 0 ( B 01 , B 02 , · · · ) = ( π 1 , π 2 , · · · ) U B C B C B C G 3 G 2 G I · · · B C B C . . ... ... @ A . . . . π 0 ( B ∗ 01 , B ∗ 02 , · · · ) = ( π 1 , π 2 , · · · ) U (14) ∞ B 01 + B 02 G + B 03 G 2 + · · · = X B 0 k G k − 1 B ∗ = 01 k =1 ∞ B 02 + B 03 G + B 04 G 2 + · · · = X B 0 k G k − 2 B ∗ = 02 k =2 . . . ∞ B 0 i + B 0 ,i +1 G + B 0 ,i +2 G 2 + · · · = X B 0 k G k − i B ∗ = 0 i k = i 48 SMF-07:PE Bertinoro, Italy

  33. William J. Stewart Department of Computer Science N.C. State University Can compute the successive components of π once π 0 and U are known:   A ∗ A ∗ A ∗ A ∗ · · · 1 2 3 4   A ∗ A ∗ A ∗ 0 · · ·   1 2 3   π 0 ( B ∗ 01 , B ∗ A ∗ A ∗   0 0 · · · 02 , · · · ) = ( π 1 , π 2 , · · · ) 1 2     A ∗  0 0 0 · · ·  1   . . .  ...  . . . . . . Observe that − 1 π 0 B ∗ 01 = π 1 A ∗ π 1 = π 0 B ∗ 01 A ∗ = ⇒ 1 1 − 1 − π 1 A ∗ − 1 π 0 B ∗ 02 = π 1 A ∗ 2 + π 2 A ∗ π 2 = π 0 B ∗ 02 A ∗ 2 A ∗ = ⇒ 1 1 1 − 1 − π 1 A ∗ − 1 − π 2 A ∗ − 1 π 0 B ∗ 03 = π 1 A ∗ 3 + π 2 A ∗ 2 + π 3 A ∗ π 3 = π 0 B ∗ 03 A ∗ 3 A ∗ 2 A ∗ = ⇒ 1 1 1 1 . . . 49 SMF-07:PE Bertinoro, Italy

  34. William J. Stewart Department of Computer Science N.C. State University In general: − 1 , π 0 B ∗ 0 i − π 1 A ∗ i − π 2 A ∗ i − 1 − · · · − π i − 1 A ∗ A ∗ � � π i = i = 1 , 2 , . . . 2 1 i − 1 � � − 1 . � π 0 B ∗ π k A ∗ A ∗ = 0 i − i − k +1 1 k =1 π 0 ( B ∗ 01 , B ∗ First subvector π 0 : 02 , · · · ) = ( π 1 , π 2 , · · · ) U 0 1 I − B 00 − B ∗ − B ∗ − B ∗ − B ∗ · · · · · · 01 02 03 0 j B C A ∗ A ∗ A ∗ A ∗ − B 10 B C · · · · · · 1 2 3 j B C B C A ∗ A ∗ 0 0 A j − 1 · · · · · · B C 1 2 B C ( π 0 , π 1 , · · · , π j , · · · ) B C 0 0 0 A ∗ A ∗ · · · · · · B C 1 j − 2 B C B C . B C 0 0 0 0 . · · · B C . B C @ A . . . = (0 , 0 , · · · 0 , · · · ) 50 SMF-07:PE Bertinoro, Italy

  35. William J. Stewart Department of Computer Science N.C. State University First two equations: − π 0 B ∗ 01 + π 1 A ∗ π 0 ( I − B 00 ) − π 1 B 10 = 0 , 1 = 0 . Second gives − 1 . π 1 = π 0 B ∗ 01 A ∗ 1 Substitute into first − 1 B 10 = 0 π 0 ( I − B 00 ) − π 0 B ∗ 01 A ∗ 1 or � � − 1 B 10 I − B 00 − B ∗ 01 A ∗ π 0 = 0 1 Can now compute π 0 to a multiplicative constant. To normalize, enforce the condition: � ∞ � � ∞ � − 1 � � B ∗ A ∗ π 0 e + π 0 e = 1 . (15) 0 i i i =1 i =1 51 SMF-07:PE Bertinoro, Italy

  36. William J. Stewart Department of Computer Science N.C. State University Computation of the matrix U from 0 1 I − A 1 − A 2 − A 3 − A j · · · · · · B C − A 0 I − A 1 − A 2 − A j − 1 B · · · · · · C B C B C 0 − A 0 I − A 1 − A j − 2 UL = . · · · · · · B C B C B C 0 0 − A 0 − A j − 3 · · · · · · B C B C @ A . . . . . . . . . . . . . . . . . . 52 SMF-07:PE Bertinoro, Italy

  37. William J. Stewart Department of Computer Science N.C. State University 0 1 A ∗ A ∗ A ∗ A ∗ · · · 1 2 3 4 B C A ∗ A ∗ A ∗ 0 B C · · · 1 2 3 B C B C A ∗ A ∗ 0 0 · · · B C 1 2 B C B C 0 0 0 A ∗ · · · B 1 C B C @ A . . . ... . . . . . . 0 1 0 1 I − A 1 − A 2 − A 3 − A j I 0 0 0 · · · · · · · · · B C B C − A 0 I − A 1 − A 2 − A j − 1 G I 0 0 B C B C · · · · · · · · · B C B C B C B C G 2 0 − A 0 I − A 1 − A j − 2 G I 0 = · · · · · · · · · B C B C B C B C G 3 G 2 B C B C 0 0 − A 0 − A j − 3 G I · · · · · · · · · B C B C B C B C @ A @ A . . . . . . . . ... ... . . . . . . . . . . . . . . . . ∞ 1 = I − A 1 − A 2 G − A 3 G 2 − A 4 G 3 − · · · = I − X A k G k − 1 A ∗ k =1 ∞ 2 = − A 2 − A 3 G − A 4 G 2 − A 5 G 3 − · · · = − X A k G k − 2 A ∗ k =2 53 SMF-07:PE Bertinoro, Italy

  38. William J. Stewart Department of Computer Science N.C. State University ∞ I − A 1 − A 2 G − A 3 G 2 − A 4 G 3 − · · · = I − X A k G k − 1 A ∗ = 1 k =1 ∞ − A 2 − A 3 G − A 4 G 2 − A 5 G 3 − · · · = − X A k G k − 2 A ∗ = 2 k =2 ∞ − A 3 − A 4 G − A 5 G 2 − A 6 G 3 − · · · = − X A k G k − 3 A ∗ = 3 k =3 . . . ∞ − A i − A i +1 G − A i +2 G 2 − A i +3 G 3 − · · · = − X A k G k − i , A ∗ = i ≥ 2 . i k = i We now have all the results we need. 54 SMF-07:PE Bertinoro, Italy

  39. William J. Stewart Department of Computer Science N.C. State University The basic algorithm is • Construct the matrix G . • Obtain π 0 by solving the system of equations − 1 B 10 � I − B 00 − B ∗ 01 A ∗ � π 0 = 0 , 1 subject to the normalizing condition, Equation (15). − 1 . • Compute π 1 from π 1 = π 0 B ∗ 01 A ∗ 1 • Find all other required π i from � � 0 i − � i − 1 − 1 . π 0 B ∗ k =1 π k A ∗ A ∗ π i = 1 i − k +1 where ∞ ∞ � � A k G k − 1 B ∗ B 0 k G k − i , A ∗ 0 i = i ≥ 1; 1 = I − k = i k =1 ∞ � A ∗ A k G k − i , and i = − i ≥ 2 . k = i 55 SMF-07:PE Bertinoro, Italy

  40. William J. Stewart Department of Computer Science N.C. State University Computational questions: (1) The matrix G . The iterative procedure suggested is very slow: ∞ � A i G i G (0) = 0; G ( k +1) = ( k ) , k = 0 , 1 , . . . i =0 Faster variant from Neuts: � � ∞ � G ( k +1) = ( I − A 1 ) − 1 A i G i G (0) = 0; A 0 + , k = 0 , 1 , . . . ( k ) i =2 Among fixed point iterations, Bini and Meini has the fastest convergence � − 1 � ∞ � A i G i − 1 G (0) = 0; G ( k +1) = I − A 0 , k = 0 , 1 , . . . ( k ) i =1 More advanced techniques based on cyclic reduction have been developed and converge much faster. 56 SMF-07:PE Bertinoro, Italy

  41. William J. Stewart Department of Computer Science N.C. State University 2) Computation of infinite summations: Frequently the structure of the matrix is such that A k and B k are zero for relatively small values of k . Since � ∞ k =0 A k and � ∞ k =0 B k are stochastic A k and B k are negligibly small for large values of i and can be set to zero once k exceeds some threshold k M . When k M is not small, finite summations of the type � k M k = i Z k G k − i should be evaluated using Horner’s rule. For example, if k M = 5 5 Z k G k − 1 = Z 1 G 4 + Z 2 G 3 + Z 3 G 2 + Z 4 G + A 5 � Z ∗ 1 = k =1 should be evaluated from the inner-most parenthesis outwards as Z ∗ 1 = ( [ ( Z 1 G + Z 2 ) G + Z 3 ] G + Z 4 ) G + Z 5 . 57 SMF-07:PE Bertinoro, Italy

  42. William J. Stewart Department of Computer Science N.C. State University Example: Same as before but with incorporates additional transitions ( ζ 1 = 1 / 48 and ζ 2 = 1 / 16 ) to higher numbered non-neighboring states. ζ ζ ζ ζ ζ ζ 1 1 1 1 1 1 λ λ λ λ λ 1 1 1 1 1 5,1 1,1 2,1 3,1 4,1 λ 1 γ γ γ γ γ γ γ γ γ 0,1 γ 1 1 1 1 1 2 2 2 2 2 µ µ µ µ µ γ γ 1,2 2,2 3,2 4,2 5,2 1 2 γ γ γ γ γ γ 0,2 γ γ γ 1 γ 1 1 1 1 2 2 2 2 2 λ 2 1,3 3,3 4,3 2,3 5,3 λ λ λ λ λ 2 2 2 2 2 ζ ζ ζ ζ ζ ζ 2 2 2 2 2 2 Figure 3: State transition diagram for an M/G/1-type process. 58 SMF-07:PE Bertinoro, Italy

  43. William J. Stewart Department of Computer Science N.C. State University 0 23 / 48 5 / 12 1 / 12 1 / 48 B 1 / 4 31 / 48 1 / 24 1 / 16 B B B 23 / 48 5 / 12 1 / 12 1 / 48 B B B 1 / 3 1 / 3 1 / 4 1 / 12 B B B 1 / 4 31 / 48 1 / 24 1 / 16 B B B 23 / 48 5 / 12 1 / 12 B P = B 2 / 3 1 / 4 1 / 12 B B B 1 / 4 31 / 48 1 / 24 B B B 1 / 2 5 / 12 B B B 2 / 3 1 / 4 1 / 12 B B B 1 / 4 31 / 48 B B @ ... ... 59 SMF-07:PE Bertinoro, Italy

  44. William J. Stewart Department of Computer Science N.C. State University 0 1 0 1 0 1 0 0 0 23 / 48 5 / 12 0 1 / 12 0 0 B C B C B C A 0 = A , A 1 = A , A 2 = A , 0 2 / 3 0 1 / 4 0 1 / 12 0 0 0 B C B C B C @ @ @ 0 0 0 0 1 / 4 31 / 48 0 0 1 / 24 0 1 1 / 48 0 0 0 1 0 1 @ 23 / 48 5 / 12 @ 1 / 12 0 0 B C A , A , A 3 = A , B 00 = B 01 = 0 0 0 B C 1 / 4 31 / 48 0 0 1 / 24 @ 0 0 1 / 16 0 1 0 0 0 1 @ 1 / 48 0 0 B C B 02 = and B 10 = A . 1 / 3 1 / 3 A B C 0 0 1 / 16 @ 0 0 60 SMF-07:PE Bertinoro, Italy

  45. William J. Stewart Department of Computer Science N.C. State University First, using Equation (12), we verify that the Markov chain with transition probability matrix P is positive-recurrent. 0 1 . 583333 . 416667 0 B C A = A 0 + A 1 + A 2 + A 3 = A . . 250000 . 666667 . 083333 B C @ 0 . 250000 . 750000 π A = ( . 310345 , . 517241 , . 172414) . Also 0 1 0 1 0 1 . 708333 . 416667 0 1 1 . 125000 B C B C B C b = ( A 1 +2 A 2 +3 A 3 ) e = A = A . . 250000 0 . 083333 1 0 . 333333 B C B C B C @ A @ @ 0 . 250000 . 916667 1 1 . 166667 The Markov chain is positive-recurrent since 0 1 1 . 125000 B C π A b = ( . 310345 , . 517241 , . 172414) A = . 722701 < 1 0 . 333333 B C @ 1 . 166667 61 SMF-07:PE Bertinoro, Italy

  46. William J. Stewart Department of Computer Science N.C. State University Computation of the matrix G : The ij element of G is the conditional probability that starting in state i of any level n ≥ 2 , the process enters level n − 1 for the first time by arriving at state j of that level. For this particular example this means that the elements in column 2 of G must all be equal to 1 and all other elements must be zero —- the only transitions from any level n to level n − 1 are from and to the second element. Nevertheless, let see how each of the three different fixed point formula actually perform. 62 SMF-07:PE Bertinoro, Italy

  47. William J. Stewart Department of Computer Science N.C. State University We take the initial value, G (0) , to be zero. G ( k +1) = � ∞ i =0 A i G i Formula #1: ( k ) , k = 0 , 1 , . . . G ( k +1) = A 0 + A 1 G ( k ) + A 2 G 2 ( k ) + A 3 G 3 ( k ) After 10 iterations, the computed matrix is   0 . 867394 0   G (10) =  . 0 . 937152 0    0 . 766886 0 Formula #2: G ( k +1) = ( I − A 1 ) − 1 � � A 0 + � ∞ i =2 A i G i , k = 0 , 1 , . . . ( k ) G ( k +1) = ( I − A 1 ) − 1 � � A 0 + A 2 G 2 ( k ) + A 3 G 3 ( k ) 63 SMF-07:PE Bertinoro, Italy

  48. William J. Stewart Department of Computer Science N.C. State University After 10 iterations:   0 . 999844 0   G (10) =  . 0 . 999934 0    0 . 999677 0 � − 1 � I − � ∞ i =1 A i G i − 1 Formula #3: G ( k +1) = A 0 , k = 0 , 1 , . . . ( k ) � − 1 � I − A 1 − A 2 G ( k ) − A 3 G 2 G ( k +1) = A 0 ( k ) This is the fastest of the three. After 10 iterations:   0 . 999954 0   G (10) =  . 0 . 999979 0    0 . 999889 0 64 SMF-07:PE Bertinoro, Italy

  49. William J. Stewart Department of Computer Science N.C. State University We continue with the algorithm using the exact value of G . In preparation, we compute the following quantities, using the fact that A k = 0 for k > 3 and B 0 k = 0 for k > 2 . 0 1 . 520833 − . 520833 0 ∞ A k G k − 1 = I − A 1 − A 2 G − A 3 G 2 = X B C A ∗ 1 = I − − . 250000 1 − . 083333 B C @ A k =1 0 − . 354167 . 354167 0 1 − . 083333 − . 020833 0 ∞ A k G k − 2 = − ( A 2 + A 3 G ) = X B C A ∗ 2 = − 0 0 0 B C @ A k =2 0 − . 062500 − . 041667 0 1 − . 020833 0 0 ∞ A k G k − 3 = − A 3 = X B C A ∗ 3 = − 0 0 0 B C @ A k =3 0 0 − . 062500 65 SMF-07:PE Bertinoro, Italy

  50. William J. Stewart Department of Computer Science N.C. State University 0 1 ∞ @ . 083333 . 020833 0 B 0 k G k − 1 = B 01 + B 02 G = X B ∗ 01 = A 0 . 062500 . 041667 k =1 0 1 ∞ @ . 020833 0 0 B 0 k G k − 2 = B 02 = X B ∗ 02 = A 0 0 . 062500 k =2 0 1 2 . 640 1 . 50 . 352941 − 1 = B C A ∗ . 720 1 . 50 . 352941 1 B C @ A . 720 1 . 50 3 . 176470 Now compute the initial subvector, π 0 , from 0 1 . 468750 − . 468750 “ ” − 1 B 10 I − B 00 − B ∗ 01 A ∗ 0 = π 0 = π 0 1 @ A − . 302083 . 302083 gives (un-normalized) π 0 = ( . 541701 , . 840571) . 66 SMF-07:PE Bertinoro, Italy

  51. William J. Stewart Department of Computer Science N.C. State University Normalization: � ∞ � � ∞ � − 1 � � B ∗ A ∗ π 0 e + π 0 e = 1 . 0 i i i =1 i =1 i.e., 3 ) − 1 e = 1 . π 0 e + π 0 ( B ∗ 01 + B ∗ 02 ) ( A ∗ 1 + A ∗ 2 + A ∗ Evaluating 3 ) − 1 ( B ∗ 01 + B ∗ 02 ) ( A ∗ 1 + A ∗ 2 + A ∗ − 1 0 1 . 416667 − . 541667 0 0 1 @ . 104167 . 020833 0 B C = − . 250000 1 − . 083333 B C A 0 . 062500 . 104167 @ A 0 − . 416667 . 250000 0 1 @ . 424870 . 291451 . 097150 = A . 264249 . 440415 . 563472 67 SMF-07:PE Bertinoro, Italy

  52. William J. Stewart Department of Computer Science N.C. State University 0 1 1 0 1 0 1 @ 1 @ . 424870 . 291451 . 097150 A + ( . 541701 , . 840571) B C ( . 541701 , . 840571) 1 B C A 1 . 264249 . 440415 . 563472 @ A 1 = 2 . 888888 Finally, initial subvector is π 0 = ( . 541701 , . 840571) / 2 . 888888 = ( . 187512 , . 290967) − 1 = We can now find π 1 from the relationship π 1 = π 0 B ∗ 01 A ∗ 1 0 1 2 . 640 1 . 50 . 352941 0 1 @ . 083333 . 020833 0 B C ( . 187512 , . 290967) . 720 1 . 50 . 352941 B C A 0 . 062500 . 041667 @ A . 720 1 . 50 3 . 176470 = ( . 065888 , . 074762 , . 0518225) . 68 SMF-07:PE Bertinoro, Italy

  53. William J. Stewart Department of Computer Science N.C. State University Finally, all needed remaining subcomponents of π can be found from � i − 1 � � − 1 π 0 B ∗ π k A ∗ A ∗ π i = 0 i − i − k +1 1 k =1 − 1 ( π 0 B ∗ 02 − π 1 A ∗ 2 ) A ∗ π 2 = 1 = ( . 042777 , . 051530 , . 069569) − 1 = ( − π 1 A ∗ − 1 ( π 0 B ∗ 03 − π 1 A ∗ 3 − π 2 A ∗ 2 ) A ∗ 3 − π 2 A ∗ 2 ) A ∗ π 3 = 1 1 = ( . 0212261 , . 024471 , . 023088) − 1 = ( − π 2 A ∗ − 1 ( π 0 B ∗ 04 − π 1 A ∗ 4 − π 2 A ∗ 3 − π 3 A ∗ 2 ) A ∗ 3 − π 3 A ∗ 2 ) A ∗ π 4 = 1 1 = ( . 012203 , . 014783 , . 018471) . . . 69 SMF-07:PE Bertinoro, Italy

  54. William J. Stewart Department of Computer Science N.C. State University The probability that the Markov chain is in any level i is given by � π i � 1 . Thus the probabilities of this Markov chain being in the first 5 levels � π 0 � 1 = . 478479 , � π 1 � 1 = . 192473 , � π 2 � 1 = . 163876 , � π 3 � 1 = . 068785 , � π 4 � 1 = . 045457 The sum of these five probabilities is 0.949070. 70 SMF-07:PE Bertinoro, Italy

  55. William J. Stewart Department of Computer Science N.C. State University Phase Type Distributions Goals: (1) Find ways to model general distributions while maintaining the tractability of the exponential. (2) Find way to form a distribution having some given expectation and variance. Phase-type distributions are represented as the passage through a succession of exponential phases or stages (and hence the name). 71 SMF-07:PE Bertinoro, Italy

  56. William J. Stewart Department of Computer Science N.C. State University The Exponential Distribution — consists of a single exponential phase. Random variable X is exponentially distributed with parameter µ > 0 . µ The diagram represents customers entering the phase from the left, spending an amount of time that is exponentially distributed with parameter µ within the phase and then exiting to the right. Exponential density function: f X ( x ) ≡ dF ( x ) = µe − µx , x ≥ 0 dx σ 2 X = 1 /µ 2 . Expectation and variance, E [ X ] = 1 /µ ; 72 SMF-07:PE Bertinoro, Italy

  57. William J. Stewart Department of Computer Science N.C. State University The Erlang-2 Distribution Service provided to a customer is expressed as one exponential phase followed by a second exponential phase. µ µ 2 1 Although both service phases are exponentially distributed with the same parameter, they are completely independent — the servicing process does not contain two independent servers. Probability density function of each of the phases: f Y ( y ) = µe − µy , y ≥ 0 σ 2 Y = 1 /µ 2 . Expectation and variance, E [ Y ] = 1 /µ ; 73 SMF-07:PE Bertinoro, Italy

  58. William J. Stewart Department of Computer Science N.C. State University Total time in service is the sum of two independent and identically distributed exponential random variables. X = Y + Y � ∞ f X ( x ) = f Y ( y ) f Y ( x − y ) dy −∞ � y µe − µy µe − µ ( x − y ) dy = 0 � x µ 2 e − µx dy = µ 2 xe − µx , = x ≥ 0 , 0 and is equal to zero for x ≤ 0 — the Erlang-2 distribution: E 2 The corresponding cumulative distribution function is given by F X ( x ) = 1 − e − µx − µxe − µx = 1 − e − µx (1 + µx ) , x ≥ 0 . 74 SMF-07:PE Bertinoro, Italy

  59. William J. Stewart Department of Computer Science N.C. State University Laplace transform of the overall service time distribution: � ∞ e − sx f X ( x ) dx L X ( s ) ≡ 0 Laplace transform of each of the exponential phases: � ∞ e − sy f Y ( y ) dy. L Y ( s ) ≡ 0 Then L X ( s ) = E [ e − sx ] = E [ e − s ( y 1 + y 2 ) ] = E [ e − sy 1 ] E [ e − sy 2 ] = L Y ( s ) L Y ( s ) � 2 � µ = , s + µ To invert, look up tables of transform pairs. 75 SMF-07:PE Bertinoro, Italy

  60. William J. Stewart Department of Computer Science N.C. State University x r 1 r ! e − ax . ⇐ ⇒ (16) ( s + a ) r +1 Setting a = µ and r = 1 allows us to invert L X ( s ) to obtain f X ( x ) = µ 2 xe − µx = µ ( µx ) e − µx , x ≥ 0 Moments may be found from the Laplace transform as E [ X k ] = ( − 1) k d k � � ds k L X ( s ) for k = 1 , 2 , . . . � � s =0 � � E [ X ] = − d = − µ 2 d s =0 = 2 = µ 2 2( s + µ ) − 3 � ds ( s + µ ) − 2 � � ds L X ( s ) µ. � � � � � s =0 s =0 Time spent in service is the sum of two iid random variables: E [ X ] = E [ Y ] + E [ Y ] = 1 /µ + 1 /µ = 2 /µ � 1 � 1 � 2 � 2 = 2 σ 2 X = σ 2 Y + σ 2 Y = + µ 2 . µ µ 76 SMF-07:PE Bertinoro, Italy

  61. William J. Stewart Department of Computer Science N.C. State University Example: Exponential random variable with parameter µ ; Two phase Erlang-2 random variable, each phase having parameter 2 µ . Mean Variance 1 /µ 2 Exponential 1 /µ 1 / 2 µ 2 Erlang-2 1 /µ An Erlang-2 random variable has less variability than an exponentially distributed random variable with the same mean. 77 SMF-07:PE Bertinoro, Italy

  62. William J. Stewart Department of Computer Science N.C. State University The Erlang-r Distribution A succession of r identical, but independent, exponential phases with parameter µ . µ µ µ µ r 2 r−1 1 Probability density function at phase i : f Y ( y ) = µe − µy ; y ≥ 0 Expectation and variance per phase: σ 2 Y = 1 /µ 2 E [ Y ] = 1 /µ, and respectively . 78 SMF-07:PE Bertinoro, Italy

  63. William J. Stewart Department of Computer Science N.C. State University Distribution of total time spent is the sum of r iid random variables. � 1 � 1 � 2 � = r = r σ 2 E [ X ] = r µ ; X = r µ 2 . µ µ � r � µ Laplace transform of the service time : L X ( s ) = s + µ x r 1 r ! e − ax Using the transform pair : ⇐ ⇒ with a = µ ( s + a ) r +1 f X ( x ) = µ ( µx ) r − 1 e − µx , x ≥ 0 . (17) ( r − 1)! This is the Erlang-r probability density function. The corresponding cumulative distribution function is given by r − 1 ( µx ) i � F X ( x ) = 1 − e − µx , x ≥ 0 and r = 1 , 2 , . . . (18) i ! i =0 79 SMF-07:PE Bertinoro, Italy

  64. William J. Stewart Department of Computer Science N.C. State University Differentiating F X ( x ) with respect to x shows that (18) is the distribution function with corresponding density function (17). r − 1 r − 1 k ( µx ) k − 1 µ ( µx ) k f X ( x ) = d � � µe − µx − e − µx dxF X ( x ) = k ! k ! k =0 k =0 r − 1 r − 1 ( µx ) k k ( µx ) k − 1 µ µe − µx + µe − µx � � − e − µx = k ! k ! k =1 k =1 r − 1 � k ( µx ) k − 1 − ( µx ) k � µe − µx − µe − µx � = k ! k ! k =1 � r − 1 �� � ( µx ) k − 1 ( k − 1)! − ( µx ) k � µe − µx = 1 − k ! k =1 1 − ( µx ) r − 1 = µ ( µx ) r − 1 � � �� µe − µx ( r − 1)! e − µx . = 1 − ( r − 1)! 80 SMF-07:PE Bertinoro, Italy

  65. William J. Stewart Department of Computer Science N.C. State University The area under this density curve is equal to one. Let � ∞ µ r x r − 1 e − µx I r = dx, r = 1 , 2 , . . . ( r − 1)! 0 I 1 = 1 is the area under the exponential density curve. Using integration by parts: �� � vdu with u = µ r − 1 x r − 1 / ( r − 1)! and dv = µe − µx dx ) udv = uv − � ∞ µ r − 1 x r − 1 µe − µx I r = dx ( r − 1)! 0 � ∞ ∞ µ r − 1 x r − 1 µ r − 1 x r − 2 � ( r − 1)! e − µx � ( r − 2)! e − µx dx = 0 + I r − 1 = + � � 0 0 It follows that I r = 1 for all r ≥ 1 . 81 SMF-07:PE Bertinoro, Italy

  66. William J. Stewart Department of Computer Science N.C. State University Squared coefficient of variation, C 2 X , for the family of Erlang-r distributions. X = r/µ 2 ( r/µ ) 2 = 1 C 2 r < 1 , for r ≥ 2 . “More regular” than exponential random variables. Possible values: 1 2 , 1 3 , 1 4 , · · · Allows us to approximate a constant distribution. 82 SMF-07:PE Bertinoro, Italy

  67. William J. Stewart Department of Computer Science N.C. State University Mixing an Erlang- ( r − 1) distribution with an Erlang- r distribution gives a distribution with 1 /r ≤ C 2 X ≤ 1 / ( r − 1) . r−1 2 1 µ µ µ α 1−α µ µ µ µ r 2 r−1 1 α = 1 ⇒ C 2 α = 0 ⇒ C 2 X = 1 / ( r − 1) ; X = 1 /r . For a given E [ X ] and C 2 X ∈ [1 /r, 1 / ( r − 1)] choose 1 � � µ = r − α � rC 2 r (1 + C 2 X ) − r 2 C 2 α = X − and E [ X ] (19) 1 + C 2 X X 83 SMF-07:PE Bertinoro, Italy

  68. William J. Stewart Department of Computer Science N.C. State University The Hypoexponential Distribution µ µ r−1 µ µ 2 1 r r 2 r−1 1 Two phases: exponentially distributed RVs, Y 1 and Y 2 : X = Y 1 + Y 2 . � ∞ f X ( x ) = f Y 1 ( y ) f Y 2 ( x − y ) dy −∞ � x µ 1 e − µ 1 y µ 2 e − µ 2 ( x − y ) dy = 0 � x µ 1 µ 2 e − µ 2 x e − ( µ 1 − µ 2 ) y dy = 0 µ 1 µ 2 e − µ 2 x − e − µ 1 x � � = ; x ≥ 0 . µ 1 − µ 2 84 SMF-07:PE Bertinoro, Italy

  69. William J. Stewart Department of Computer Science N.C. State University Corresponding cumulative distribution function is given by µ 2 µ 1 e − µ 1 x + e − µ 2 x , F X ( x ) = 1 − x ≥ 0 . µ 2 − µ 1 µ 2 − µ 1 Expectation, variance and squared coefficient of variation: � µ 2 1 + µ 2 E [ X ] = 1 + 1 V ar [ X ] = 1 + 1 C 2 2 , , and X = < 1 , µ 2 µ 2 µ 1 µ 2 µ 1 + µ 2 1 2 Laplace transform � � � � µ 1 µ 2 L X ( s ) = . s + µ 1 s + µ 2 The Laplace transform for an r phase hypoexponential random variable: � � � � � � µ 1 µ 2 µ r L X ( s ) = · · · . s + µ 1 s + µ 2 s + µ r 85 SMF-07:PE Bertinoro, Italy

  70. William J. Stewart Department of Computer Science N.C. State University The density function, f X ( x ) , is the convolution of r exponential densities each with its own parameter µ i and is given by r r µ i � � α i µ i e − µ i x , f X ( x ) = x > 0 where α i = , µ j − µ i i =1 j =1 , j � = i Expectation, variance and squared coefficient of variation: r r i 1 /µ 2 � 1 1 � � C 2 i E [ X ] = , V ar [ X ] = and X = i 1 /µ i ) 2 ≤ 1 . µ 2 µ i ( � i i =1 i =1 Observe that C 2 X cannot exceed 1. 86 SMF-07:PE Bertinoro, Italy

  71. William J. Stewart Department of Computer Science N.C. State University Example: Three exponential phases with parameters µ 1 = 1 , µ 2 = 2 and µ 3 = 3 . 3 1 = 1 1 + 1 2 + 1 3 = 11 � E [ X ] = µ i 6 i =1 3 1 = 1 1 + 1 4 + 1 9 = 49 � V ar [ X ] = µ 2 36 i i =1 121 / 36 = 36 49 / 36 C 2 = 121 = 0 . 2975 . X 87 SMF-07:PE Bertinoro, Italy

  72. William J. Stewart Department of Computer Science N.C. State University Probability density function of X . r r µ i � � α i µ i e − µ i x , f X ( x ) = x > 0 where α i = , µ j − µ i i =1 j =1 , j � = i r µ 1 µ 1 µ 1 = 1 1 × 1 2 = 1 � α 1 = = × µ j − µ 1 µ 2 − µ 1 µ 3 − µ 1 2 j =1 , j � = i r µ 2 µ 2 µ 2 = 2 − 1 × 2 � α 2 = = × 1 = − 4 µ j − µ 2 µ 1 − µ 2 µ 3 − µ 2 j =1 , j � = i r µ 3 µ 3 µ 3 = 3 − 2 × 3 − 1 = 9 � α 3 = = × µ j − µ 3 µ 3 − µ 1 µ 3 − µ 2 2 j =1 , j � = i It follows then that 3 α i µ i e − µ i x = (0 . 5) e − x + 8 e − 2 x + (13 . 5) e − 3 x , � f X ( x ) = x > 0 i =1 88 SMF-07:PE Bertinoro, Italy

  73. William J. Stewart Department of Computer Science N.C. State University The Hyperexponential Distribution Our goal now is to find a phase-type arrangement that gives larger coefficients of variation than the exponential. µ 1 α 1 1−α 1 µ 2 The density function: f X ( x ) = α 1 µ 1 e − µ 1 x + α 2 µ 2 e − µ 2 x , x ≥ 0 Cumulative distribution function: F X ( x ) = α 1 (1 − e − µ 1 x ) + α 2 (1 − e − µ 2 x ) , x ≥ 0 . 89 SMF-07:PE Bertinoro, Italy

  74. William J. Stewart Department of Computer Science N.C. State University Laplace transform: µ 1 µ 2 L X ( s ) = α 1 + α 2 . s + µ 1 s + µ 2 First and second moments: E [ X ] = α 1 + α 2 E [ X 2 ] = 2 α 1 + 2 α 2 and . µ 2 µ 2 µ 1 µ 2 1 2 Variance: V ar [ X ] = E [ X 2 ] − ( E [ X ]) 2 . Squared coefficient of variation: X = E [ X 2 ] − ( E [ X ]) 2 = E [ X 2 ] ( E [ X ]) 2 − 1 = 2 α 1 /µ 2 1 + 2 α 2 /µ 2 C 2 2 ( α 1 /µ 1 + α 2 /µ 2 ) 2 − 1 ≥ 1 . ( E [ X ]) 2 90 SMF-07:PE Bertinoro, Italy

  75. William J. Stewart Department of Computer Science N.C. State University Example: Given α 1 = 0 . 4 , µ 1 = 2 and µ 2 = 1 / 2 . 0 . 4 2 + 0 . 6 E [ X 2 ] = 0 . 8 4 + 1 . 2 E [ X ] = 0 . 5 = 1 . 40 0 . 25 = 5 √ � 5 − 1 . 4 2 = σ X = 3 . 04 = 1 . 7436 5 C 2 = 1 . 4 2 − 1 = 2 . 5510 − 1 . 0 = 1 . 5510 X 91 SMF-07:PE Bertinoro, Italy

  76. William J. Stewart Department of Computer Science N.C. State University With r parallel phases and branching probabilities � r i =1 α i = 1 : µ 1 α 1 µ 2 α 2 α r−1 µ r−1 α r µ r r r µ i � � α i µ i e − µ i x , f X ( x ) = x ≥ 0; L X ( s ) = α i . s + µ i i =1 i =1 92 SMF-07:PE Bertinoro, Italy

  77. William J. Stewart Department of Computer Science N.C. State University r r α i α i � � E [ X 2 ] = 2 E [ X ] = and µ 2 µ i i i =1 i =1 ( E [ X ]) 2 − 1 = 2 � r X = E [ X 2 ] i =1 α i /µ 2 C 2 i i =1 α i /µ i ) 2 − 1 . ( � r To show that this squared coefficient of variation is greater than or equal to one, it suffices to show that � r � 2 r � � α i /µ 2 α i /µ i ≤ i . i =1 i =1 Use the Cauchy-Schwartz inequality: for real a i and b i � 2 �� �� � �� � a 2 b 2 a i b i ≤ . i i i i i 93 SMF-07:PE Bertinoro, Italy

  78. William J. Stewart Department of Computer Science N.C. State University Substituting a i = √ α i and b i = √ α i /µ i implies that √ α i � 2 � 2 �� �� √ α i α i = µ i µ i i i � √ α i � 2 √ α i 2 � � ≤ , using Cauchy − Schwartz µ i i i �� � �� � α i α i � � = α i = , since α i = 1 . µ 2 µ 2 i i i i i i Therefore C 2 X ≥ 1 . 94 SMF-07:PE Bertinoro, Italy

  79. William J. Stewart Department of Computer Science N.C. State University The Coxian Distribution α α 2 µ 1 µ 1 µ r 2 1 1−α 1−α 1 2 With probability p 1 = 1 − α 1 , process terminates after phase 1. With probability p 2 = α 1 (1 − α 2 ) , it terminates after phase 2. With probability p k = (1 − α k ) � k − 1 i =1 α i , it terminates after phase k . A Coxian distribution may be represented as a probabilistic choice from among r hypoexponential distributions: 95 SMF-07:PE Bertinoro, Italy

  80. William J. Stewart Department of Computer Science N.C. State University µ 1 p 1 µ µ 1 2 p 2 p r−1 µ 2 µ µ 1 r−1 p r µ r−1 µ µ 1 µ 2 r 96 SMF-07:PE Bertinoro, Italy

  81. William J. Stewart Department of Computer Science N.C. State University Phase 1 is always executed and has expectation E [ X 1 ] = 1 /µ 1 . Phase 2 is executed with probability α 1 and has E [ X 2 ] = 1 /µ 2 . Phase k > 1 is executed with probability � k − 1 j =1 α j and E [ X k ] = 1 /µ k . Since the expectation of a sum is equal to the sum of the expectations: r E [ X ] = 1 + α 1 + α 1 α 2 + · · · + α 1 α 2 · · · α r − 1 A k � = , µ 1 µ 2 µ 3 µ r µ k k =1 where A 1 = 1 and, for k > 1 , A k = � k − 1 j =1 α j . The case of a Cox-2 random variable is especially important. E [ X ] = 1 + α 1 = µ 2 + αµ 1 , (20) µ 1 µ 2 µ 1 µ 2 97 SMF-07:PE Bertinoro, Italy

  82. William J. Stewart Department of Computer Science N.C. State University Laplace transform of a Cox-2 µ 1 µ 1 µ 2 L X ( s ) = (1 − α ) + α s + µ 1 s + µ 1 s + µ 2 ( − 1) 2 d 2 �� � (1 − α ) µ 1 αµ 1 µ 2 E [ X 2 ] � = + � ds 2 s + µ 1 ( s + µ 1 )( s + µ 2 ) � s =0 d � − (1 − α ) µ 1 � − 1 − 1 �� = + α 1 µ 1 µ 2 ( s + µ 1 )( s + µ 2 ) 2 + ( s + µ 1 ) 2 ( s + µ 1 ) 2 ( s + µ 2 ) ds 2(1 − α ) µ 1 2 αµ 1 µ 2 αµ 1 µ 2 = ( s + µ 1 ) 3 + ( s + µ 1 )( s + µ 2 ) 3 + ( s + µ 1 ) 2 ( s + µ 2 ) 2 � 2 αµ 1 µ 2 αµ 1 µ 2 � + ( s + µ 1 ) 3 ( s + µ 2 ) + � ( s + µ 1 ) 2 ( s + µ 2 ) 2 � s =0 98 SMF-07:PE Bertinoro, Italy

  83. William J. Stewart Department of Computer Science N.C. State University 2(1 − α ) + 2 α α + 2 α α = + + µ 2 µ 2 µ 2 µ 1 µ 2 µ 1 µ 2 1 2 1 2 + 2 α 2 α = + µ 2 µ 2 µ 1 µ 2 1 2 � 2 � 1 � 2 + 2 α 2 α � + α V ar [ X ] = + − µ 2 µ 2 µ 1 µ 2 µ 1 µ 2 1 2 2 µ 2 1 + 2 αµ 2 − ( µ 2 + αµ 1 ) 2 1 + 2 αµ 1 µ 2 = µ 2 1 µ 2 µ 2 1 µ 2 2 2 µ 2 2 + 2 αµ 2 1 − α 2 µ 2 = µ 2 2 + αµ 2 1 (2 − α ) 1 = µ 2 1 µ 2 µ 2 1 µ 2 2 2 E [ X ] 2 = µ 2 2 + αµ 2 ( µ 2 + αµ 1 ) 2 = µ 2 µ 2 1 µ 2 2 + αµ 2 X = V ar [ X ] 1 (2 − α ) 1 (2 − α ) 2 C 2 × . µ 2 1 µ 2 ( µ 2 + αµ 1 ) 2 2 (21) 99 SMF-07:PE Bertinoro, Italy

  84. William J. Stewart Department of Computer Science N.C. State University Example: Coxian-2 RV with parameters µ 1 = 2 , µ 2 = 0 . 5 and α = 0 . 25 , 1 + α = 1 2 + 1 / 4 E [ X ] = 1 / 2 = 1 µ 1 µ 2 2 + 2 α 2 α = 2 4 + 1 / 2 1 / 4 + 1 / 2 E [ X 2 ] = + = 3 µ 2 µ 2 µ 1 µ 2 1 1 2 E [ X 2 ] − E [ X ] 2 = 3 − 1 = 2 V ar [ X ] = V ar [ X ] /E [ X ] 2 = 2 C 2 = X 100 SMF-07:PE Bertinoro, Italy

Recommend


More recommend