Checking Whether M/M/1 Assumptions Fit a Queue

What an M/M/1 Model Assumes

An M/M/1 model is useful only to the extent that its assumptions describe the queue being studied. The notation summarizes several distinct claims: arrivals follow a Poisson process, service times are exponential, there is one server, customers are served first-come, first-served, and the waiting capacity is unlimited. Data can reveal conflicts with these claims, but a finite sample cannot establish that a model is exactly true.

Represent the queue on a probability space \((\Omega,\mathcal{F},\mathbb{P})\). Let \(N_t\) count arrivals by time \(t\), and let \(X_t\) be the number of customers in the system at time \(t\), including any customer in service. Both processes have continuous time \(t\ge0\) and state space \(\{0,1,2,\ldots\}\). Arrival times, service starts and completions, queue size, and service order are observable features that can be compared with the model.

The two letters M refer to memoryless arrival and service mechanisms. In particular, the model requires more than exponential-looking marginal histograms: the arrival process must have the appropriate independent-increment structure, and service durations must be independent with a common exponential distribution. The “1” is an operational fact, not a distributional fit: there is one server. Likewise, FCFS and unlimited capacity describe how the queue is run.

Model fit and stability are different questions. When \(\lambda<\mu\), the usual M/M/1 steady-state distribution exists. If the arrival and service assumptions seem reasonable but the estimated arrival rate is not below the service rate, steady-state formulas still cannot be used as if the queue were stable. Conversely, \(\lambda<\mu\) does not make incorrect distributional or operational assumptions valid.

Observable Implications to Check

For a homogeneous Poisson arrival process, successive interarrival times are independent exponential random variables with rate \(\lambda\). An exponential random variable \(T\) has survival probability \(\mathbb{P}(T>t)=e^{-\lambda t}\) for \(t\ge0\), mean \(1/\lambda\), variance \(1/\lambda^2\), and coefficient of variation \(\sqrt{\operatorname{Var}(T)}/\mathbb{E}[T]=1\). The corresponding service-time diagnostics use rate \(\mu\). These moment identities provide useful checks, not decisive tests: samples fluctuate, and distinct distributions can have the same mean and variance.

Proof. Write \(S(t)=\mathbb{P}(T>t)\). For an exponential lifetime, \(S(t)=e^{-ct}>0\), so \(S(s+t)/S(s)=e^{-ct}=S(t)\), which proves the memoryless identity. Conversely, suppose the identity holds and \(S(s)>0\). It gives \(S(s+t)=S(s)S(t)\). Since \(T\) is positive and has a continuous distribution, \(S\) is continuous at zero with \(S(0)=1\); the identity then implies \(S(t)>0\) for every finite \(t\). Define \(g(t)=-\log S(t)\). The identity gives \(g(s+t)=g(s)+g(t)\). Also, \(S\) is nonincreasing, so \(g\) is nondecreasing. For nonnegative rational \(r\), additivity gives \(g(r)=r g(1)\); continuity extends this equality from rational to real \(t\ge0\). Thus \(S(t)=e^{-ct}\) with \(c=g(1)\). Because a proper lifetime has \(S(t)\to0\) as \(t\to\infty\), \(c>0\). Therefore \(T\) is exponential. \(\square\)

In practice, memorylessness means that the remaining service time should not depend on how long the current service has already lasted. A service-time histogram alone cannot check that condition or independence. If service durations change with workload, time of day, customer type, or the preceding service, an exponential marginal pattern is not enough to justify the model.

Poisson arrivals also make a prediction about counts in fixed time intervals. The next result derives the count distribution from independent exponential interarrival times, so it makes clear what the count diagnostic is checking.

Proof. Let \(S_n\) be the time of the \(n\)th arrival. For \(n=0\), no arrival by time \(t\) means the first exponential gap exceeds \(t\), giving \(\mathbb{P}(N_t=0)=e^{-\lambda t}\). For \(n\ge1\), the sum \(S_n\) has density \(\lambda^n u^{n-1}e^{-\lambda u}/(n-1)!\) for \(u>0\), obtained by repeated convolution of exponential densities. Given \(S_n=u\le t\), there are exactly \(n\) arrivals by \(t\) if the next gap exceeds \(t-u\), an event of probability \(e^{-\lambda(t-u)}\), independent of \(S_n\). Hence

Writing \(a=\lambda t\), summing the probability series gives \(\mathbb{E}[N_t]=a\), since \(n/n!=1/(n-1)!\) for \(n\ge1\). Similarly, \(\mathbb{E}[N_t(N_t-1)]=a^2\), since \(n(n-1)/n!=1/(n-2)!\) for \(n\ge2\). Thus \(\operatorname{Var}(N_t)=\mathbb{E}[N_t(N_t-1)]+\mathbb{E}[N_t]-(\mathbb{E}[N_t])^2=a^2+a-a^2=a\). The formulas also hold at \(t=0\), when \(N_t=0\). \(\square\)

Therefore, counts in equal-length intervals should have the same mean and variance under the model, and counts in disjoint intervals should be independent. Compare intervals of a fixed length; choosing interval boundaries after inspecting the counts can distort the comparison. Excess variation, trends, clustering, or dependence are warning signs. With limited data, however, an observed mismatch may be sampling variation rather than proof of a different arrival mechanism.

Worked Diagnostics

A Practical Audit and Common Mistakes

Before using M/M/1 calculations, audit the mechanism and the data separately. Operational records can often settle the single-server, capacity, and service-order questions directly. Arrival and service data instead provide evidence about stochastic assumptions, which require cautious interpretation.

• Arrivals: record arrival times and examine interarrival gaps, counts over equal-length intervals, possible changes in rate, and dependence between intervals. A changing arrival rate over the observation period conflicts with a single homogeneous Poisson rate.
• Service: record individual service durations and check their distribution, variability, and possible dependence on time, customer type, or earlier services. Include only completed durations as complete observations; a service still in progress at the end of data collection is not a completed service time.
• Queue operation: verify that there is exactly one server, that customers are served FCFS, and that the system can accept every arrival without a finite capacity limit. Abandonment, balking, or blocking changes how customers enter or remain in the queue.
• Dependence and observation: check whether arrival and service mechanisms plausibly operate independently, and whether missing or selectively recorded events could conceal departures from the assumptions.

Common mistake. A coefficient of variation equal to \(1\) is not proof of an exponential distribution. For example, let a nonnegative service duration \(T\) be \(0\) with probability \(1/2\) and \(2\) with probability \(1/2\). Then \(\mathbb{E}[T]=(0+2)/2=1\), \(\mathbb{E}[T^2]=(0+4)/2=2\), and \(\operatorname{Var}(T)=2-1^2=1\). Its coefficient of variation is \(1\), but its distribution has point masses and is not exponential. The correction is to treat the coefficient of variation as one diagnostic summary and also assess distributional shape, dependence, and the actual operating rules.

The M/M/1 model is a useful approximation when its assumptions are reasonable for the question and time period at hand. A diagnostic should identify which assumptions have direct operational support, which have data-based evidence, and which remain uncertain. It should not turn a short sample or a matching summary statistic into a claim that the queue is exactly M/M/1.