A note on integrals for birth±death processes
Valeri T. Stefanov
*, Song Wang
Department of Mathematics and Statistics, The University of Western Australia, Nedlands, WA 6907, Australia
Received 21 December 1999; received in revised form 21 August 2000; accepted 30 August 2000
Abstract
A general integral for birth±death Markov processes is considered. A closed form expression is provided for its expected value in terms of the birth and death rates. A simple route for numerical evaluation of its variance is also suggested. Ó 2000 Elsevier Science Inc. All rights reserved.
Keywords:Birth±death process; Hitting time; Integral of Markov process
Hernandez-Suarez and Castillo-Chavez [1] discusses a methodology for evaluating the
expec-tation of the integral under the stochastic path of a birth±death process up to extinction time and the expected time to extinction. Their main result is a closed form expression for the expectation of the aforementioned integral. In this note we demonstrate a more general result, which embraces and substantially extends that in Ref. [1], ¯ows from Stefanov's [2] results on ®nite-state birth± death processes. The relevance of such results to biological sciences is spelled out in Ref. [1] (see also the references therein). Furthermore, we provide a simple route for numerical evaluation of the variance of this integral together with a code for implementing such evaluation. In what
follows, we will use similar notation to that adopted in Ref. [1], letting X t be a birth±death
Markov process on the state spacef0;1;. . .;g, with birth and death rateski andli, respectively
for statei. Further, denote byZ k the waiting time till reaching state 0 from statek, that is
Z k infft>0:X t 0jX 0 kg
and de®ne
Yf k Z Z k
0
f X tdt;
*Corresponding author. Tel.: +61-8 9380 1870; fax: +61-8 9380 1028.
E-mail address:stefanov@maths.uwa.edu.au (V.T. Stefanov).
0025-5564/00/$ - see front matter Ó 2000 Elsevier Science Inc. All rights reserved.
wherefis a non-negative real function on the state space of the chainX; we assume additionally
that there exists a positive real numberx0, such thatfis a non-decreasing function forxPx0. Note
that iff x 1, thenYf k Z k, and iff x x, thenYf kis the integral considered in Ref. [1].
Denote
Hj i
kjkj1. . .kiÿ1
lj1lj2. . .li; i>jP1
with the conventionHi i 1; i1;2;. . ., and Hj i 0 ifi<j. Our result is stated in the
fol-lowing.
Proposition.The expected value ofYf khas the following closed form expression in terms of theki
andli:
E Yf k Xk
j1 1 lj
X1 ij
f iHj i !
: 1
Note that in the particular case whenk 1 andf x x the aforementioned formula reduces
to the main result in Ref. [1] (see (5) on p. 98 of Ref. [1]; there is a conspicuous misprint in (5)).
Also iff x 1 the formula reduces to a well-known result (see (3.58) on p. 67 of Ref. [3]).
Proof of Proposition. LetSi k be the total time spent in statei within the random time interval
0;Z k. Clearly
Yf k X1
i1
f iSi k:
From the monotone convergence theorem we have
E Yf k limE Xm
i1
f iSi k !
2
asm! 1. Denote byZ m k, Y m
f kandS
m
i k, mPk, the counterparts ofZ k,Yf kand
Si k in a m1-state birth±death process with birth and death rates ki andli, respectively for
each statei,i0;1;. . .;m; m<1; km0. It is easy to see that for largem(mPx0, wherex0
has been introduced above)
E Yf kPE Y m
f kPE Xm
i1
f iSi k !
: 3
In view of the results in Ref. [2] (Theorem 2, cf. the second and fourth identities in Theorem 1) we have
E Z m k Xk j1
1 lj
Xm
ij
Hj i !
X
k
j1 1 lj
X1 ij
Hj i !
where the second equality is due tokm0. Moreover, as a byproduct from the results in Ref. [2]
To see this, start with the last few lines on p. 848 and derive the identities, corresponding to those
at the bottom of p. 847, for the stopping time sn;nÿ1 that has been introduced in Ref. [2]. Thus,
(1) translates into the following closed form expressions for the four models considered in Ref. [1]
(qk=l<1 and the conventionPÿ1
(note that there is a misprint in the expression forY kprovided in Ref. [1]),
(b)ki k; liil(M=M=1 queue),
(note that the suggested expression forEY kin [1] is incorrect),
It seems that a neat closed form expression for the variance of the random quantity Yf k is not
derivable. However, the tools from Ref. [2] provide a simple route for numerical evaluation of this
variance. First note that, in view of our assumptions for the functionf and the fact that the Si
are non-negative, (2) and (3) remain valid if we replace `E' by `Var'. Therefore,
Var Yf k lim
number of one-step transitions from state i to state j within the time interval 0;Z m k (recall
that km0). Observe that the following system of linear equations hold (m and k have been
suppressed for the sake of brevity):
Var Ni;i1
with the convention N0 ;m1 k 0. These equations are derived after dierentiating twice, with
respect to theki andli, the continuous-time counterpart of the integral identity on p. 847 of Ref.
[2], bearing in mind that the dierentiation and integration signs can be interchanged, the easily noticeable relations
N1 m;0 k 1;
Ni ;mi1 k Ni m1;i k ÿ1; i1;. . .;kÿ1; Ni ;mi1 k Ni m1;i k; ik;. . .;mÿ1;
The expected values of theNi ;mi1 kare known coecients, which are derived from (4) and (7). The solution of system (6) produces the Cov Si m k;Sj m kand subsequently the variance of Yf k.
Note that system (6) is sparse (up to four unknown variables per equation), and can therefore be solved numerically even when the order of the associated matrix is very large. Since the system matrix is not diagonally dominant and it is not clear whether it is positive-de®nite or not, iterative algorithms, such as those based on preconditioned conjugate gradient methods, are not appli-cable. Therefore, we use the LU decomposition for the solution of (6). There are a number of FORTRAN 77/90 and C codes for solving sparse linear systems based on LU decomposition.
Such are also available on the Web (cf.http://www.netlib.org). For the example provided
below we choose a reliable FORTRAN 77 sparse linear system solver, Harwell MA28, which is
freely available for non-commercial use (cf.http://www.netlib.org/harwell/). We have
written a Fortran 77 code for assembling the relevant matrix and the right-hand side vector in (6),
and have made it available at http://www.maths.uwa.edu.au/stefanov. Numerical
results for the SIS model, with dierent values of the parameters, are provided in Table 1.
References
[1] C.M. Hernandez-Suarez, C. Castillo-Chavez, A basic result on the integral for birth±death Markov processes, Math. Biosci. 161 (1999) 95.
[2] V.T. Stefanov, Mean passage times for tridiagonal transition matrices, J. Appl. Prob. 32 (1995) 846.
[3] E. Renshaw, Modelling Biological Populations in Space and Time, Cambridge University, Cambridge, UK, 1991. Table 1
E Y kand
Var Y k
p
for dierent values of the parameters in the SIS model
Initial number of infectives
k0:8;l1 k1:2;l1
k Expected cost Standard deviation of cost Expected cost Standard deviation of cost
N150
1 4.4948 11.8384 263.5898 928.6760
2 8.8927 16.4846 483.8832 1216.2008
5 21.5417 24.9267 958.9993 1577.8288
10 41.0037 32.9079 1394.4095 1743.6087
N200
1 4.5989 12.4200 672.1968 2400.7958
2 9.1202 17.3503 1234.3382 3145.9029
5 22.2418 26.4731 2443.6815 4083.3338
10 42.7658 35.4092 3533.4096 4508.8569
N250
1 4.6670 12.8167 1650.1987 5888.8914
2 9.2691 17.9419 3030.0504 7714.3959
5 22.7031 27.5379 5990.3266 10 000.3876
10 43.9425 37.1583 8620.6198 11 018.9165
N300
1 4.7151 13.1057 3960.3590 14 044.2136
2 9.3745 18.3735 7270.8599 18 387.0272
5 23.0308 28.3188 14 355.8171 23 791.1018