Abstract
We explain how lattice-path counting techniques can be used in conjunction with the random-product representations from Buckingham and Fralix (Markov Process Related Field 21:339–368 2015) to study both the stationary and time-dependent behavior of Markovian queueing systems, and continuous-time Markov chains in general. We illustrate how the approach works by applying it to both the Er/M/1 queue, and the M/Er/1 queue. Interestingly, through this approach we show that the stationary distributions, as well as the Laplace transforms of the transition functions associated with both the Er/M/1 queue and the M/Er/1 queue, can be expressed explicitly in terms of generalized binomial series from Chapter 5 of the text Concrete Mathematics of Graham, Knuth, and Patashnik.
Similar content being viewed by others
References
Abate J, Whitt W (1994) Transient behavior of the M/G/1 workload process. Oper Res 42:750–764
Abate J, Whitt W (1995) Numerical inversion of Laplace transforms of probability distributions. ORSA J Comput 7:36–43
Adan I, Van de Waarsenburg WA, Wessels J (1996) Analyzing e k/e r/c queues. Eur J Oper Res 92:112–124
Adan I, Van leeuwaarden J, Selen J (2017) Analysis of Structured Markov Processes. Draft accessible through the arXiv:1709.09060
Baccelli F, Massey WA (1989) A sample path analysis of the M/M/1 queue. J Appl Probab 26:418–422
Baek JW, Moon SK, Lee HW (2014) A time-dependent busy period queue length formula for the M/Ek/1 queue. Stat Probab Lett 87:98–104
Böhm W (2010) Lattice path counting and the theory of queues. J Stat Plan Inf 140:2168–2183
Böhm W, Krinik A, Mohanty SG (1997) The combinatorics of birth-death processes and applications to queues. Queue Syst 26:255–267
Buckingham P, Fralix B (2015) Some new insights into Kolmogorov’s criterion, with applications to hysteretic queues. Markov Process Related Field 21:339–368
Champernowne DG (1956) An elementary method of solution of the queueing problem with a single server and constant parameters. J Royal Stat Soc B 18:125–128
Flajolet P, Guillemin F (2000) The formal theory of birth-and-death processes, lattice path combinatorics and continued fractions. Adv Appl Probab 32:750–778
Fralix B (2015) When are two Markov chains similar?. Stat Probab Lett 107:199–203
Fralix B, Zwart B (2011) Time-dependent properties of symmetric queues. Queue Syst 67:33–45
Graham RL, Knuth DE, Patashnik O (1994) Concrete mathematics. Addison-Wesley Publishing Company, Boston
Grassmann WK (2011) A new method for finding the characteristic roots of e n/e m/1 queues. Methodol Comput Appl Probab 13:873–886
Griffiths JD, Leonenko GM, Williams JE (2006) The transient solution to m/e k/1 queue. Oper Res Lett 34:349–354
Griffiths JD, Leonenko GM, Williams JE (2008) Approximation to the transient solution of the m/e k/1 queue. INFORMS J Comput 20:510–515
He Q-M (2014) Fundamentals of matrix-analytic methods. Springer, New York
Jain JL, Mohanty SG, Böhm W (2007) A course on queueing models. Chapman and Hall/CRC Press, Bota Raton
Joyner J, Fralix B (2016) A new look at Markov processes of G/M/1-type. Stoch Model 32:253–274
Joyner J, Fralix B (2018) A new look at block-structured Markov processes. Currently under revision: available via bfralix.people.clemson.edu/preprints.htm
Krinik A, Rubino G, Marcus D, Swift RJ, Kasfy H, Lam H (2005) Dual processes to solve single server systems. J Stat Plan Inf 135:121–147
Kyprianou AE (2006) Introductory lectures on fluctuations of lévy processes with applications. Springer, Berlin
Latouche G, Ramaswami V (1999) Introduction to matrix analytic methods in stochastic modeling. ASA-SIAM Publications, Philadelphia
Van Houdt B, Van Leeuwaarden JSH (2011) Triangular M/G/1-type and tree-like quasi-birth-death Markov chains. INFORMS J Comput 23:165–171
Van Leeuwaarden JSH, Winands EMM (2006) Quasi-birth-and-death processes with an explicit rate matrix. Stoch Model 22:77–98
Van Leeuwaarden JSH, Squillante MS, Winands EMM (2009) Quasi-birth-and-death processes, lattice-path counting, and hypergeometric functions. J Appl Probab 46:507–520
Leonenko GM (2009) A new formula for the transient solution of the Erlang queueing model. Stat Probab Lett 79:400–406
Neuts MF (1981) Matrix-geometric solutions in stochastic models: an algorithmic approach. Johns Hopkins University Press, Baltimore
Le Ny L-M, Rubino G, Sericola B (1991) Calculating the busy period distribution of the M/M/1 queue. [Research Report] RR-1501, INRIA, inria-00075061: available via https://hal.inria.fr/inria-00075061
Renault M (2007) Four proofs of the ballot theorem. Math Mag 80:345–352
Van Vianen LA, Gabor AF, van Ommeren J-K (2016) Waiting times in classical priority queues via elementary lattice path counting. Queue Syst 84:295–307
Acknowledgments
BF gratefully acknowledges the support of the National Science Foundation, via grant NSF-CMMI-1435261.
Author information
Authors and Affiliations
Corresponding author
Appendix
Appendix
We close this paper by including a brief introduction to generalized binomial series. Many of the results contained in this Appendix can also be found in Graham et al. (1994), but some of our proofs are more probabilistic in nature.
For each integer r ≥ 1, the generalized binomial series \(\mathcal {B}_{r}(z)\) is defined (as a formal power series) on page 200 of Graham et al. (1994) as follows:
Our usage of generalized binomial series does not allow us to treat them as formal power series, meaning we need to determine where these series converge absolutely. Using the ratio test, one can show that for each integer r ≥ 1, \(\mathcal {B}_{r}(z)\) absolutely converges on the open disk Ur, defined as
where we interpret 00 as 1.
The following Lemma will be useful towards establishing the absolute convergence of various series that we encounter in the proofs of our main results.
Lemma A.1
Fix an integerr ≥ 1,and consider the functions\(f_{r + 1},g_{r + 1}: [0, \infty ) \rightarrow \mathbb {R}\),defined forρ ≥ 0 as
Bothfr+ 1andgr+ 1are continuous on[0,∞), differentiable on(0,∞), and both functions have aunique global maximum atρ = 1:furthermore,
Proof
It is easy to verify that both fr+ 1 and gr+ 1 are continuous on [0,∞) and differentiable on (0,∞), and that the functions satisfy fr+ 1(1) = gr+ 1(1) = rr/(r + 1)r+ 1. The key to showing that both fr+ 1 and gr+ 1 each have a unique global maximum at ρ = 1 is to show that \(f_{r + 1}^{\prime }(\rho )\) and \(g_{r + 1}^{\prime }(\rho )\) are both positive when ρ ∈ (0, 1), and both negative when ρ > 1. A quick calculation shows that for ρ > 0,
which proves the lemma. □
The next result establishes important generalized binomial series identities found on page 200 and 201 of Graham et al. (1994). Not only do we state these identities in the interest of making our paper self-contained, we also give a probabilistic proof that differs from the proofs given in Graham et al. (1994).
Theorem A.1
Fix an integerr ≥ 1.The generalized binomial series\(\mathcal {B}_{r}(z)\)satisfiesthe following identities for eachz ∈ Ur:(i) for each integera ≥ 1,we have
Furthermore, (ii)
Proof
Formulas (44) and (45) are easy to establish for the case where r = 1: a small amount of algebra shows that for z ∈ U1,
Then
proving (45). Next, note that for each integer a ≥ 1,
where the second equality is well-known, and follows from applying an induction argument to find the coefficients corresponding to the a-fold convolution of the power series \(\mathcal {B}_{1}(z)\). This proves (44).
It remains to prove both (44) and (45) for the case where r ≥ 2. It helps to rewrite (44) and (45) as follows: for each integer a ≥ 1, and each integer r ≥ 1,
and
Thus, to prove Theorem A.1, it suffices to instead prove (46) and (47).
We begin by proving (47). From Formula (10) found within Theorem 3.1, we have
where we further assume that λ < μ. Using now Theorem 3.2, we observe after some algebra that this equality can alternatively be expressed as
or, equivalently,
This equality shows that Eq. 47 must be true. To see why, observe that from Lemma A.1, we may say that given each z ∈ [0,rr/(r + 1)r+ 1), we can choose a corresponding λ and μ which satisfies z = fr+ 1(ρ), where ρ = λ/μ. However, z = fr+ 1(ρ) = xλ,μ, which proves (47).
It remains to establish (46). Fix an integer a ≥ 1: then given a fixed integer r ≥ 1, there exists a unique pair of integers j,i satisfying j ≥ 1, 0 ≤ i ≤ r − 1, that further satisfy a = (j − 1)r + i + 1. Then from Eq. 19, we find that
We can derive an alternative expression for w(0,r− 1),(j,i)(0) using Theorem 2.1: following a line of reasoning nearly identical to that used to derive w(0,r− 1),(1,0)(α) in Theorem 3.2 gives
Combining Eqs. 48 and 49 and further applying Theorem 3.2 to write β(0) in terms of \(\mathcal {B}_{r + 1}(x_{\lambda , \mu })\) gives
or, equivalently,
From this equality, we can argue similarly as we did to establish (47) to say that for each z ∈ [0,rr/(r + 1)r+ 1),
which proves (46). □
Our last Lemma is used in the proof of Theorem 4.3.
Lemma A.2
For each integera ≥ 1,and each integerr ≥ 1,
Proof
Proceeding as indicated in Graham et al. (1994), the key to proving this lemma is to first show that for z ∈ Ur
Next, use Eq. 47 to further show that
Plugging (52) into the right-hand-side of Eq. 51 and simplifying—while making further use of Eq. 47—proves the claim. □
Remark
It should be noted here that identities (44) and (50) still hold when a is a nonpositive integer and r is a positive integer. Furthermore, given a fixed positive integer r, both series still converge absolutely on Ur (as can be verified with the ratio test).
We show how to properly extend identity (44): given any two positive integers a,b, we can say from Eq. 44 that for any integer k ≥ 1,
A bit of thought shows that when a is fixed, the left and right-hand-sides of this equality are finite-degree polynomials in b. Since these polynomials are equal for each integer b ≥ 1, they must also be equal for each \(b \in \mathbb {R}\). In particular, for each integer k ≥ 1 we have
which proves, for z ∈ Ur,
or equivalently,
This argument shows (44) still holds for each integer a ≤ 0.
The authors of Graham et al. (1994) further claim that these identities hold even when a and r are arbitrary real numbers, but it is not clear to us how/if the radius of convergence of these series change when a and/or r are no longer integers. Fortunately our arguments only require the cases where r and a are integers satisfying r ≥ 2, and a ≥−r.
Lemma 5.2 is used in our proof of Theorems 4.3 and 4.4, but we must ensure that \(z_{\lambda }z_{\mu }^{r} \in U_{r + 1}\) first. Here Lemma A.1 is needed: recall that
First, notice that when \(\alpha \in \mathbb {C}_{+}\) and λ,μ > 0, we have
Using this simple observation, we further find that
which proves \(z_{\lambda }z_{\mu }^{r} \in U_{r + 1}\). Next, notice that when λ,μ satisfy λ < μ, we have ρ = λ/μ < 1, meaning
which, again, shows that \(z_{\lambda }z_{\mu }^{r} \in U_{r + 1}\) for this case as well.
Rights and permissions
About this article
Cite this article
Liu, X., Fralix, B. On Lattice Path Counting and the Random Product Representation, with Applications to the Er/M/1 Queue and the M/Er/1 Queue. Methodol Comput Appl Probab 21, 1119–1149 (2019). https://doi.org/10.1007/s11009-018-9658-8
Received:
Revised:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s11009-018-9658-8