7.4 Alternating renewal processes

Definition of alternating renewal processes

An alternating renewal process describes a system that cycles between two states repeatedly over time. Think of a machine that runs, then breaks down, gets repaired, runs again, breaks down again, and so on. Each full cycle (one "on" period plus one "off" period) constitutes a renewal, and the process restarts probabilistically after each cycle.

More formally, you have a sequence of paired random variables $(X_1, Y_1), (X_2, Y_2), \ldots$ where $X_n$ is the duration of the $n$th on period and $Y_n$ is the duration of the $n$th off period. The pairs are independent and identically distributed, though $X_n$ and $Y_n$ don't need to follow the same distribution as each other.

Key assumptions

• The system alternates between exactly two states (often called "on"/"off" or "up"/"down").
• The cycle pairs $(X_n, Y_n)$ are i.i.d. across cycles. Within a single cycle, $X_n$ and $Y_n$ may or may not be independent of each other, but the pairs themselves are i.i.d.
• Both $X_n$ and $Y_n$ are non-negative random variables with finite means: $\mathbb{E}[X] < \infty$ and $\mathbb{E}[Y] < \infty$.
• The system starts in one of the two states at time zero (typically the "on" state).

Alternating states

On and off periods

The on period $X_n$ is the duration of the $n$th interval where the system is active or functioning. The off period $Y_n$ is the duration of the $n$th interval where the system is inactive. A single cycle has length $X_n + Y_n$.

Concrete examples:

• A factory machine runs for a random time $X_n$ before failing, then takes a random time $Y_n$ to repair.
• A web server is operational for $X_n$ hours, then goes down for maintenance lasting $Y_n$ hours.

Distributions of on/off times

The on and off durations can follow any non-negative distribution. Common choices include:

• Exponential (memoryless property makes analysis tractable)
• Gamma (useful when a state duration is the sum of several exponential phases)
• Weibull (common in reliability, since it can model increasing or decreasing failure rates)

The choice depends on the system you're modeling. Exponential is the simplest analytically, but real systems often exhibit aging or wear-out effects that require Weibull or gamma fits. Parameters are typically estimated from historical failure and repair data.

Long-run properties

Limiting probabilities

The central result for alternating renewal processes comes from the renewal reward theorem (or the key renewal theorem). The long-run fraction of time the system spends in the "on" state is:

$\pi_{\text{on}} = \frac{\mathbb{E}[X]}{\mathbb{E}[X] + \mathbb{E}[Y]}$

and similarly for the "off" state:

$\pi_{\text{off}} = \frac{\mathbb{E}[Y]}{\mathbb{E}[X] + \mathbb{E}[Y]}$

Why does this work? Over many cycles, the law of large numbers tells you the average cycle length converges to $\mathbb{E}[X] + \mathbb{E}[Y]$, and the average on-time per cycle converges to $\mathbb{E}[X]$. The ratio gives the long-run proportion.

Note that these limiting probabilities depend only on the means of $X$ and $Y$, not on the full distributions. This is a powerful feature: you don't need to know the exact shape of the distributions to compute long-run availability.

Mean on/off durations

The quantities $\mathbb{E}[X]$ and $\mathbb{E}[Y]$ are the building blocks for most long-run calculations. In reliability engineering, these have standard names:

• MTBF (Mean Time Between Failures) = $\mathbb{E}[X]$
• MTTR (Mean Time To Repair) = $\mathbb{E}[Y]$

For exponential distributions with rates $\lambda$ (failure rate) and $\mu$ (repair rate):

$\mathbb{E}[X] = \frac{1}{\lambda}, \qquad \mathbb{E}[Y] = \frac{1}{\mu}$

This gives the steady-state availability as $\frac{\mu}{\lambda + \mu}$.

Applications of alternating renewal

Reliability theory

Alternating renewal is the natural model for a single repairable component. The on period is the component's lifetime until failure; the off period is the repair duration. From this model you can derive:

• Steady-state availability: $A = \pi_{\text{on}}$, the long-run fraction of time the system is operational.
• Failure frequency: In the long run, failures occur at rate $\frac{1}{\mathbb{E}[X] + \mathbb{E}[Y]}$ per unit time.

These metrics directly inform maintenance scheduling and spare-parts inventory decisions.

Queueing systems

When a server in a queue is unreliable, alternating renewal models the server's availability. The on period is the time the server works before breaking down; the off period is the repair time. Performance measures like average queue length and waiting time depend on the server's availability, which you can compute from $\pi_{\text{on}}$.

Applications include call centers with agent breaks, manufacturing lines with machine failures, and cloud servers with intermittent outages.

Age and residual life

Forward and backward recurrence times

At any time $t$, two quantities describe where you are within the current cycle:

• Age (backward recurrence time) $A(t)$: the time elapsed since the most recent renewal before $t$.
• Residual life (forward recurrence time) $B(t)$: the time remaining until the next renewal after $t$.

If the most recent renewal occurred at time $S_n$ and the next occurs at $S_{n+1}$, then $A(t) = t - S_n$ and $B(t) = S_{n+1} - t$.

As $t \to \infty$, the distributions of $A(t)$ and $B(t)$ converge to limiting distributions that depend on the full cycle-length distribution (not just the mean). For a cycle length $Z = X + Y$ with CDF $F_Z$, the limiting density of the residual life is:

$f_{B}(x) = \frac{1 - F_Z(x)}{\mathbb{E}[Z]}, \qquad x \geq 0$

This is the equilibrium (or spread) distribution of the cycle length.

Renewal functions and equations

The renewal function $M(t) = \mathbb{E}[N(t)]$ gives the expected number of renewals (completed cycles) by time $t$. It satisfies the renewal equation:

$M(t) = F_Z(t) + \int_0^t M(t - x)\, dF_Z(x)$

where $F_Z$ is the CDF of the cycle length $Z = X + Y$.

Solving this integral equation directly is usually hard. Two common approaches:

1. Laplace transforms: Take the transform of both sides. The convolution becomes a product, giving $\hat{M}(s) = \frac{\hat{F}_Z(s)}{1 - \hat{F}_Z(s)}$, which you can invert (analytically for simple distributions, numerically otherwise).

2. Numerical methods: Discretize time and solve the equation iteratively.

The elementary renewal theorem provides the asymptotic result: $\frac{M(t)}{t} \to \frac{1}{\mathbb{E}[Z]}$ as $t \to \infty$. So in the long run, renewals occur at a constant rate equal to the reciprocal of the mean cycle length.

Generalizations and extensions

Semi-Markov processes

Semi-Markov processes generalize alternating renewal by allowing the next state's identity and holding time to depend on the current state. The transition structure is described by a semi-Markov kernel $Q_{ij}(t)$, which gives the probability of transitioning to state $j$ within time $t$ given the current state is $i$.

When there are exactly two states and the transitions always alternate (on $\to$ off $\to$ on $\to \cdots$), the semi-Markov process reduces to the standard alternating renewal process. The added flexibility of semi-Markov models is useful when transition patterns are more complex, such as multi-state degradation models in reliability.

Multivariate alternating renewal

When a system cycles through more than two states, you get a multivariate alternating renewal process. Each state $i$ has its own holding-time distribution, and a transition probability matrix governs which state comes next.

The limiting probability of being in state $i$ generalizes to:

$\pi_i = \frac{p_i \, \mathbb{E}[Z_i]}{\sum_j p_j \, \mathbb{E}[Z_j]}$

where $p_i$ is the stationary probability of visiting state $i$ (from the embedded Markov chain) and $\mathbb{E}[Z_i]$ is the mean holding time in state $i$. This formula reduces to the two-state version when there are only on and off states with deterministic alternation.