Risk, Reliability and Societal Safety – Aven & Vinnem (eds)

© 2007 Taylor & Francis Group, London, ISBN 978-0-415-44786-7

A conceptual interpretation of the renewal theorem with applications

J.A.M. van der Weide

Delft University of Technology, Delft, The Netherlands

M.D. Pandey
Department of Civil Engineering, University of Waterloo, Waterloo, Canada

J.M. van Noortwijk

HKV Consultants, Lelystad, The Netherlands
Delft University of Technology, Delft, The Netherlands

ABSTRACT: The risk management of degrading engineering systems can be optimised through inspections
and renewals. In the probabilistic modelling of life-cycle management, the renewal theorem plays a key role in
the computation of the expected number of renewals and the cost rate associated with a management strategy.
The renewal theorem is well known and its rigourous mathematical proof is presented in the literature, though
probabilistic arguments associated with its derivation are not well understood by the engineering community.
The central objective of this paper is to present a more lucid and intuitive interpretation of the renewal theorem
and to derive asymptotic expansions for the first and second moment of the number of renewals. As far as the
authors know, the latter expansion is new in the sense that it also contains a constant term.

1 INTRODUCTION distributed (iid) sequence of 0–1 random variables,

where 1 means occurrence of failure. This leads to
In industrialised nations, the infrastructure elements a discrete version of the Poisson process where the
critical to economy, such as bridges, roads, power times between failures are independent, geometrically
plants and transmission lines, are experiencing aging distributed random variables. However, in cases where
related degradation, which makes them more vul- the geometric distribution of the inter-occurrence time
nerable to failure. To minimise the risk associated cannot be justified, a natural way to generalise the anal-
with failure of an infrastructure system, inspection ysis is to model it as a renewal process. The renewal
and replacements are routinely carried out. Because theorem provides asymptotic results for the first and
of uncertainty associated with degradation mecha- second moment of the number of failures. For discrete-
nisms, operational loads and limited inspection data, time and continuous-time renewal processes, see Feller
probabilistic methods play a key role in develop- (1949, 1950) and Smith (1954, 1958), respectively. In
ing cost effective management models. The theory this paper, we extend Feller’s asymptotic expression
of stochastic renewal processes and the renewal the- for the second moment of the number of renewals.
orem have been fundamental to the development We propose to model the renewal processes with a
of risk-based asset management models (Rackwitz, sequence of 0–1 variables which are in general neither
2001; van Noortwijk, 2003). For an example in bridge independent nor identically distributed. The probabil-
management, see van Noortwijk and Klatter (2004). ities of occurrence of failure are given by the renewal
Although the renewal processes have been discussed in sequence associated with the distribution between fail-
many mathematical treaties (Feller, 1950, 1966; Karlin ures. This approach gives insight in the dependence
and Taylor, 1975; Tijms, 2003), the concepts are not structure of the indicator variables of the event that
amenable to the engineering community. The objec- failure occurs. The proposed approach allows to cal-
tive of this paper is to present a conceptually simple culate quite easily the mean and variance of the number
and intuitive interpretation of renewal processes with of failures in a finite time horizon.
applications. The proposed model is applied to develop a risk-
For sake of conceptual simplicity, we consider a based asset management framework for utility wood
discrete time scale to model the process of times at poles in an electrical transmission line network. Using
which failure occurs with an independent, identically the actual inspection data, we model the lifetime

477
 The counting process N = (Nt ; t = 0, 1, 2, . . . ) asso-
ciated with the sequence (Sn ) is called the renewal
process with renewal distribution (pk )

where S0 ≡ 0. The simplest example of a renewal pro-

cess is the discrete Poisson process with a shifted
geometric renewal distribution
Figure 1. Terminology associated with the renewal process
where Nt = n.

distribution as a discrete Weibull distribution. A sec-

ond example concerns the renewal of hydraulic cylin- where 0 < p < 1. However, the simplest way to define
ders for which the lifetimes are distributed according the discrete Poisson process is by introducing an iid
to a Poisson distribution. sequence X , X1 , X2 , . . . of 0–1 random variables
This paper is organised as follows. In Section 2, we
present the basic properties of discrete-time renewal
processes by studying the number of renewals over
a finite time horizon. For an infinite time horizon, and defining N0 = 0 and
asymptotic results are obtained in Section 3. Dis-
counted renewal cost is incorporated to the renewal
model in Section 4. Two illustrations of the pro-
posed discrete-time renewal model are presented in
It follows that Nt is binomially distributed and there-
Section 5. Conclusions are formulated in Section 6.
fore the discrete Poisson process is also called the
binomial process.
2 THEORY OF RENEWAL PROCESSES

2.1 General concepts 2.2 Distribution of number of renewals

The basic premise is that the lifetime (T ) of a com- The number of renewals, Nt , in the interval (0, t] is a
ponent is a random variable. It is assumed that the random variable, and its distribution can be derived
component is replaced with a new and identical com- from probabilistic arguments. In particular Nt = 0 if
ponent as soon as a failure occurs. The probability only if S1 = T1 > t, hence
distribution of T is given as

We will always assume that the probability distribution For n ≥ 1, it follows by conditioning on T1 and (1) that
(pk ) is aperiodic which is certainly the case if p1 > 0.
This process is shown in Figure 1 in which T , T1 , T2 , . . .
denote an iid sequence of positive, integer valued ran-
dom variables corresponding to inter-arrival times of
failure. The probability generating function of (pk ) is
given by It is clear that the time of occurrence of the (n − 1)th
failure necessarily follows an inequality, Sn − 1 ≥ n − 1.
It means that we only have to sum over k such that
t − k ≥ n − 1 or k ≤ t + 1 − n. In summary, it leads
to a recursive formula to calculate the probability
distributions of Nt as
where p0 = 0. The probability distribution (pk ) is
referred to as the renewal distribution. The time of
occurrence of the nth renewal (failure) is given by the
partial sum

where P(Nt = 0) is given by (2).

478
2.3 Renewal function takes place at discrete times Sk , the indicator variable
can be written as
The expected number of renewals up to time t (first
moment) is referred to as the renewal function, i.e.,
M (t) = E(Nt ). This section presents the derivation of
M (t). An alternative way to write Nt is

The mathematical properties of the sequence (Ij ) are

extensively studied in Kingman (1972) under the name
of discrete-time regenerative phenomena. Indicators
of recurrent events as defined in Feller (1949, 1950)
Note that the indicator function 1A = 1 when the con- are examples of regenerative phenomena. The renewal
dition A is true, otherwise it is zero. Actually, the sum rate at time j, denoted as uj , is defined as
in (4) is over the finite range n = 1, . . ., t, since Sk ≥ k
for all k. Now taking expectations of both sides of (4)
results in
and it can be expressed as

where Fn denotes the cumulative distribution function

of Sn . A recursion equation for E(Nt ) can be derived It follows from (8) that uj = E(Ij ) for j = 1, 2, . . . Let
using (4) and (5) as us define u0 = 1. The renewal rate can be alternatively
derived using the generating function

which is related with the generating function of the

renewal distribution as
for k ∈ [1, t]. Recall from (4) that

Consider an example of renewal distribution, p1 = p

and p2 = 1 − p = q. Its generating function is given as

This equation comes from the fact that the pro-

cess starts afresh after the first failure. Since T1 is
Using Equation (10)
independent of other failure times, (6) is rewritten as

By summation over k = 1, . . ., t, and using the law of and partial fractions, the generating function U (z) is
total probability obtained as

Equation (7) is called the renewal equation, and the Thus,

renewal function is its unique solution corresponding
to the renewal distribution (pk ).

2.4 Rate of renewal The following recursion can be used to calculate the
The rate of occurrence of renewals can be derived sequence as
through the use of a random indicator variable Ij . It
is an indicator of a renewal at time j, such that Ij = 1 if
there is a renewal at step j, otherwise 0. Since a renewal

479
In general, the renewal indicators Ij are not indepen- Consider the event
dent. It turns out that

It follows from (16) that the event can only occur if

Since i = t − Sk . So Sk = t − i and, since Sk ≥ k, we get that
i + k ≤ t. If RLt = j and Nt = k, then (15) implies that
Sk + 1 = t + j or Tk + 1 = Sk + 1 − Sk = i + j. The joint
distribution of (At , Nt , RLt ) is given by
the mean and the factorial second moment can be
expressed in terms of the sequence uj as

where i + k ≤ t and j ≥ 1. By the law of total

probability,
and since Ij2 = Ij for 0–1 random variables

and (9) implies that the joint distribution of (At , RLt )

From (12) and (13), the second moment of the number is given by
of renewals follows simply

By the law of total probability, the marginal distribu-

So, the basic idea is that the renewal rate can be directly tions of the age and the remaining life are
obtained from the generating function of the renewal
distribution. Then, it can be used to compute moments
of the number of renewals.

2.5 Distribution of remaining lifetime and age

In the context of the renewal process, it is of interest
to obtain the remaining lifetime of a component at a
future time t. Since it is a random variable, its prob-
ability distribution can be derived from the renewal
arguments as follows.
At time t, the number of renewals is denoted as Nt . 3 ASYMPTOTIC RESULTS
The next renewal will occur at time SNt + 1 . Thus the
remaining lifetime (also called the excess of residual This section analyses the generalised renewal equation
lifetime) can be written as

The time of the last renewal that occurred before time where (at ) is some given sequence. When at = F(t), the
t is denoted as SNt . The age (or current lifetime) of the solution is the renewal function. If a0 = 1 and at = 0
current component is therefore given as for all t > 0, the solution is the renewal sequence (uj ).
Using generating functions, we can show that the solu-
tion of (20) is the convolution of the sequence (at ) and
the renewal sequence (uj )
If there was a renewal precisely at time t, then SNt = t
and the age of the current component is 0. The age
of the current component is also referred to as the
backward recurrence time at t.

480
 If the kth moment of the renewal distribution exists, lines but is much more technical. We only give the
it will be denoted by µk . One of the most important result. Let M2 (t) = E(Nt2 ) and assume that the renewal
theorems about renewal processes is the renewal the- distribution has a finite third moment µ3 . Then
orem. If the first moment of the renewal distribution
exists, then

For the discrete renewal theorem, see Feller (1950,

Chapters 12 & 13) and Karlin and Taylor (1975, Chap- The asymptotic expansion for the first moment can
ter 3). It follows for the renewal function M (t) = E(Nt ) be found in Feller (1949) with a derivation using gen-
that erating functions. In Feller’s paper, there is also an
expansion for the second moment but the constant term
is not explicitly given and it is not clear how this term
can be found using generating functions. Note that
the the coefficients in the expansion for the discrete
The discrete version of the key renewal theorem fol- case and the continuous case are different. The results
lows now directly. Let (xt ) be the solution of the for the continuous case can be found in Tijms (2003,
generalised renewal equation (20). If the first Chapter 8).

∞ moment
of the renewal distribution exists and t = 0 |at | is
convergent, then
4 DISCOUNTED COST

In this section, we consider the case of constant

renewal cost c > 0 with exponential discounting

Following Tijms (2003, Chapter 8), we use the discrete

key renewal theorem to derive asymptotic expansions
for the first and the second moment of the renewal
process N . Suppose that the renewal distribution (pk )
has a finite second moment µ2 . From (22), it follows where r > 0 is the discount rate and e−r is the discount
that M (t) ≈ t/µ1 + o(t) for large t, where the term o(t) factor. It follows that
is of lower order than t. Define

and
Z0 is the solution of the generalised renewal equation
(20) with

where U is the generating function of the renewal

sequence (uj ). It follows from the renewal equation
and (10) that

It follows from an application of the key renewal The variance of Kt is given by

theorem that (Feller, 1949)

An asymptotic expansion of the second moment of the

number of renewals Nt can be found along the same

481
Figure 2. Lifetime distribution of the component. Figure 3. Renewal rate of the component during the ser-
vice life.
and with α = e−r we get, after some simplification,
II (Stein and Dattero 1984; Ali Khan et al., 1989). Note
that when α = 1 the distribution becomes a geometric
distribution. Useful explicit expressions for the mean
and variance of the discrete Weibull distribution don’t
exist.
When the renewal cost depends on the renewal time T For the component at hand, let the shape parameter
and is denoted by cT , then the mean and variance of the be α = 3 and the upper bound of the support of the
discounted cost over a finite (and infinite) time horizon distribution m = 40. The mean lifetime of this com-
can be determined using the recurrence relations given ponent is about 15 years. Because the discrete hazard
in van Noortwijk (2003). rate function is defined as

5 ILLUSTRATIONS

This section presents illustrations in which the renewal

the probability function can be written as
rate and the first and second moment of the num-
ber of renewals over a finite time horizon as well as
their asymptotic expansions are derived for renewal
inter-occurrence times having a discrete Weibull dis-
tribution and a shifted Poisson distribution.
The probability function of the lifetime is shown in
5.1 Discrete Weibull distribution Figure 2.
The formulas derived in the previous sections can
This section illustrates an application of the theory be applied to compute the renewal rate and associ-
of renewal processes to the life-cycle management of ated statistics. Figure 3 plots the rate of occurrence of
an electrical transmission line network. The lifetime renewal, calculated using (11), along with its asymp-
distribution of electrical components in power plants, totic limit (1/µ1 = 0.0677). The mean and standard
such as switches and relays, can be modelled by a deviation of the number of renewals is plotted in
discrete Weibull distribution, as shown in Figure 2. Figure 4, which were calculated using (12) and (14).
Suppose the lifetime follows a discrete Weibull
distribution with the hazard rate function given as
5.2 Shifted Poisson distribution
As a simplified example, we study the renewal of a
hydraulic cylinder on a swing bridge (adapted from
van Noortwijk and van der Weide, 2006). The dete-
rioration X (t) at time t ≥ 0 is assumed to be dis-
where 0 < β < 1. In the literature, this distribution is tributed according to a stationary gamma process with
also known as the discrete Weibull distribution of type shape function v(t) = (µ/σ)2 t and scale parameter

482
 Figure 5. Rate of cylinder renewal per unit time of length
Figure 4. Mean and standard deviation of number of  = 0.074 year.
renewals.

u = µ/σ 2 for µ, σ > 0. Hence, the cumulative amount

of deterioration X (t) has the gamma distribution

The expected condition R(t) = r0 − X (t) is assumed to

degrade linearly in time from the initial condition of
r0 = 100% down to the failure level of s = 0% for t ≥ 0.
The cumulative distribution function of the hitting time
of the failure level (lifetime) can be rewritten as:
Figure 6. First moment of cumulative number of cylinder
renewals.

∞ for k = 1, 2, 3, . . . The mean and variance of the shifted

where y = r0 − s and (a, x) = t=x t a−1 e−t dt is the Poisson distribution are
incomplete gamma function for x ≥ 0 and a > 0. For
the hydraulic cylinder at hand, the time at which the
expected condition equals the failure level is 15 years
with parameters µ = 6.67 and σ = 1.81.
A useful property of the gamma process with sta- The rate of cylinder renewal per unit time of length
tionary increments is that the gamma density in (27)  = 0.074 year,
transforms into an exponential density if

is determined over an eighty-year design life and dis-

When the unit-time length is chosen to be (σ/µ)2 , played in Figure 5. For a finite time horizon of eighty
the increments of deterioration are exponentially dis- years, the first and second moment of the cumula-
tributed with mean σ 2 /µ and the probability of failure tive number of cylinder replacements are determined
in unit time i reduces to a shifted Poisson distribution using (12) and (14), respectively. The first and second
(see, e.g., van Noortwijk et al., 1995): moment are approximated with the asymptotic expan-
sions (24) and (25). In Figure 6, we can clearly see that
the asymptotic expansion for the first moment includ-
ing the constant term is a much better approximation

483
 second moment of the number of renewals. As far as
the authors know, the latter expansion is new in the
sense that it also contains a constant term.
The proposed concepts are illustrated by two exam-
ples involving the discrete Weibull distribution and
the shifted Poisson distribution (resulting from a dis-
cretised gamma deterioration process). More detailed
practical applications of this approach to nuclear
power plant systems are under investigation.

REFERENCES
Ali Khan, M.S., Khalique,A., &Abouammoh,A.M. 1989. On
estimating parameters in a discrete Weibull distribution.
IEEE Transactions on Reliability, 38(3): 348–350.
Figure 7. Second moment of cumulative number of cylinder Feller, W., 1949. Fluctuation theory of recurrent events.Trans-
renewals. actions of the American Mathematical Society, 67(1):98–
119.
than the asymptotic expansion excluding the constant Feller, W., 1950. An Introduction to Probability Theory and
its Applications; Volume 1. New York: John Wiley & Sons.
term. For this particular example, the constant term in Feller, W., 1966. An Introduction to Probability Theory and
the asymptotic expansion for the second moment of its Applications; Volume 2. New York: John Wiley & Sons.
the number of renewals doesn’t have much influence Karlin, S. & Taylor, H.M., 1975. A First Course in Stochastic
(see Figure 7). Processes; Second Edition. San Diego: Academic Press.
Kingman, J.F.C., 1972. Regenerative Phenomena. New York:
John Wiley & Sons.
6 CONCLUSIONS Rackwitz, R., 2001. Optimizing systematically renewed
structures. Reliability Engineering and System Safety,
In the probabilistic modelling of life-cycle manage- 73(3):269–279.
Smith, W.L., 1954. Asymptotic renewal theorems. Pro-
ment of engineering systems, the renewal theorem ceedings of the Royal Society of Edinburgh, Section A
plays a key role in the computation of the expected (Mathematical and Physical Sciences), 64:9–48.
number of renewals and the cost rate associated with Smith, W.L., 1958. Renewal theory and its ramifications.
a management strategy. The renewal theorem is well Journal of the Royal Statistical Society, Series B (Method-
known and its rigourous mathematical proof is pre- ological), 20(2):243–302.
sented in the literature for both discrete and continuous Stein, W.E. & Dattero, R., 1984. A new discrete Weibull dis-
random variables. The mathematical details and tech- tribution. IEEE Transactions on Reliability, 33:196–197.
nicalities associated with its derivation are not well Tijms, H.C., 2003. A First Course in Stochastic Models. New
understood by the engineering community. This paper York: John Wiley & Sons.
van Noortwijk, J.M., 2003. Explicit formulas for the variance
presents a more lucid and intuitive interpretation of the of discounted life-cycle cost. Reliability Engineering and
renewal theorem. System Safety, 80(2):185–195.
To simplify the presentation, we have utilised dis- van Noortwijk, J.M., Cooke, R.M., & Kok, M., 1995. A
crete random variables in the formulation of the Bayesian failure model based on isotropic deteriora-
renewal problem. This approach results in analytical tion. European Journal of Operational Research, 82(2):
expressions for the renewal rate and the probability 270–282.
distributions of the age and remaining lifetime. These van Noortwijk, J.M. & Klatter, H.E., 2004. The use of
formulas are very easy to compute, in contrast with lifetime distributions in bridge maintenance and replace-
the traditional continuous random variable formula- ment modelling. Computers and Structures, 82(13–14):
1091–1099.
tion which requires solution of integral equations. In van Noortwijk, J.M. & van derWeide, J.A.M., 2006. Compu-
the proposed approach, the life-cycle cost in a finite tational techniques for discrete-time renewal processes.
time horizon can be computed in a straightforward In Guedes Soares, C. & Zio, E., (eds.), Safety and Relia-
manner using the rate of renewal. In addition to explicit bility for Managing Risk, Proceedings of ESREL 2006 –
formulas for a finite time horizon, the paper derives European Safety and Reliability Conference 2006, Estoril,
elegant asymptotic results for the renewal rate and dis- Portugal, 18–22 September 2006, pages 571–578.
counted life-cycle cost over an infinite time horizon. London: Taylor & Francis Group.
Asymptotic expansions are derived for the first and

484